Bugs
Post-ship defect descriptions. One section per bug, keyed by id (BUG-NNN).
This file is the prose home for bugs; the structured metadata (status, severity,
affects, delivery_phase, fix_pr) lives in .forge/tracker.yaml under the
top-level bugs: collection, joined to these sections by id. This mirrors the
feature↔spec and lesson patterns: structured state in tracker.yaml, prose here.
Append-only. fixed / wontfix bugs keep their section as a historical record —
the dashboard filters to open bugs by default. A heavy bug may graduate to its own
.forge/bugs/BUG-NNN.md via the optional doc: field in the tracker; the single
section below is the default so a batch of small bugs doesn’t explode into a file each.
The bug-fix flow is right-sized — NOT forge-deliver: the section below is the
spec, a plan is optional and proportionate, and a regression test + a pre-merge diff
review (/forge-review-pr on the PR, which writes its review record) are mandatory regardless of size.
See .claude/rules/tracker.md → “Bug Tracking” for
the full lifecycle, the phase-seal rule, and the fix flow.
Section format
Each bug is an h3 heading ### BUG-NNN — <short title>, in id order with the highest
on top — a bug that takes a free lower number goes between its neighbours (BUG-011
below BUG-015; .claude/rules/tracker.md § Bug Tracking) — followed by the repro
and fix detail:
### BUG-000 — Sub-org Manage card: last child-link unlink must route to Archive
**Affects:** feature-a · **Found in:** feature-c · **Severity:** high · **Phase:** 1
**Repro:** open a record with exactly one linked child → Unlink is enabled.
**Expected:** last-link unlink is blocked / routed to Archive (FR-5, FR-10).
**Actual:** bare unlink attempted → backend rejects (or orphans the record).
**Fix:** disable Unlink on the sole remaining link; point to the Archive flow.
**Resolved:** 2026-09-30 — fixed in <repo>#<n>; the regression test and its fails-before result.
**Lesson:** none — one-off: the rule applies to this one card only.
A bug specific to one operating system — every flaky test failure, and a defect seen on one
system only — ends its heading, and its tracker.yaml title, with (Windows) or (macOS)
( (Linux CI) for a failure seen only on GitHub’s Linux runner), and a flaky failure’s starts with
Flaky test:, Flaky tests: or Flaky <what>:. A UI, functional or data defect carries no
operating system unless the defect itself exists on one system alone (.claude/rules/tracker.md
§ Bug Tracking).
The **Lesson:** line is the bug’s Reflect step (.claude/rules/tracker.md § Bug Tracking, step 6):
an existing L-NNN, a new L-NNN written in the fixing PR, or none — one-off: <why>.
A bug closed as won’t fix has a Resolved: line that says why it will not be fixed and who
decided (step 5): **Resolved:** <date> — closed as wontfix: <reason>; <who> decided. The same
person is the bug’s decided_by in tracker.yaml, and its owner is "N/A".
Lesson backfill
The Lesson: line was added to the bug-fix flow on 2026-09-30. The 36 bugs already marked fixed at that
point are mapped here instead of in their own sections, which keep what shipped. A bug marked fixed after
that carries the line in its own section (BUG-038 is the first). For a “none — one-off” row, the reason says
why the cause is not expected to recur.
| Bug | Lesson or rule | Why |
|---|---|---|
| BUG-001 | L-009 | L-009’s example of a defect with no runtime signature, which only review finds |
| BUG-002 | none — one-off | Spring falls back to the principal’s toString() for the session name in one place; the principal now names itself by user id, which closes that path |
| BUG-003 | .claude/rules/tracker.md § Bug Tracking, step 4 (see a UI fix in a running browser) |
The same class as BUG-004, BUG-006 and BUG-010: a shell layout defect only a browser shows (here a flex gap left beside a zero-width label) |
| BUG-004 | .claude/rules/tracker.md § Bug Tracking, step 4 |
The same class as BUG-003 (here a 73px header beside a 64px top bar) |
| BUG-005 | none — one-off | A product decision (sign-out lands on the default section), not a defect that can recur |
| BUG-006 | .claude/rules/tracker.md § Bug Tracking, step 4 |
The same class as BUG-003 (here a control with no flex-grow) |
| BUG-007 | L-056 | An FR clause with no matching AC |
| BUG-008 | none — one-off | A cosmetic menu defect; the process lesson L-057 came from renumbering it, not from its cause |
| BUG-009 | none — one-off | /me lacked one field; the process lessons L-057 and L-058 came from renumbering and fixing it, not from its cause |
| BUG-010 | design-system §5, “Build posture: layout-first alignment”; .claude/rules/tracker.md § Bug Tracking, step 4 |
The two rules this bug produced: no pixel nudges for alignment, and every UI fix seen in a running browser |
| BUG-015 | none — one-off | A screen shipped with no link to it; BUG-020 is its follow-up, not a second instance |
| BUG-016 | L-099 | A cache not cleared on re-consent |
| BUG-017 | coding-standards.md §8.1, “No raw object dumps” |
Already covered: a record’s default toString() is a raw object dump |
| BUG-018 | none — one-off | Only re-consent creates a Graph subscription a second time, and it now replaces the first |
| BUG-019 | none — one-off | The token below AA that was being used for normal text (text/muted) is now marked large-text only in the theme file (text/faint was already marked decorative only), and the axe scan checks every screen |
| BUG-020 | none — one-off | A refinement of BUG-015’s link, not a new cause |
| BUG-021 | none — one-off | The interim list rule was replaced by AD #35’s single visibility rule for lists and records |
| BUG-023 | L-065 | A race fixed at one caller is a class |
| BUG-024 | L-098 | A redirect button cleared its busy state too early |
| BUG-025 | none — one-off | A renewal error classification with one catch-all, now split into retry and lapse cases. L-099 cites it only as supporting evidence: its 401 path clears the token cache on one ECS task only |
| BUG-026 | L-133 | A focused control removed by the state change its own press caused, so focus fell to <body>. Mapped “none — one-off” until BUG-056 had the same cause (2026-10-05) |
| BUG-027 | L-098 | A redirect button with no busy state at all |
| BUG-028 | .claude/CLAUDE.md § Quality Gates, Browser QA |
The browser gate checks every supported size |
| BUG-029 | .claude/CLAUDE.md § Quality Gates, Browser QA |
The browser gate checks target sizes |
| BUG-030 | .claude/CLAUDE.md § Quality Gates, Browser QA |
The browser gate checks 320px reflow |
| BUG-031 | .claude/CLAUDE.md § Quality Gates, Browser QA |
The browser gate checks target sizes |
| BUG-032 | L-068 | A test tier no PR lane runs stays red |
| BUG-033 | L-101 | A scan before the dialog settled |
| BUG-034 | L-101 | A key press before the dialog’s close reached React |
| BUG-037 | L-100 | Found on the first real Entra sign-in |
| BUG-039 | L-101 | An empty-inbox check while a back-fill could land |
| BUG-040 | L-100 | Found on the first real Entra sign-in |
| BUG-041 | none — one-off | The value’s shape is now validated at startup, so a bad value stops the backend instead of reaching a redirect |
| BUG-047 | L-100 | The test harness skipped the step a browser goes through |
| BUG-048 | L-100 | Found on the first real Entra connect |
| BUG-049 | L-100 | Found on the first real Entra connect |
The open bugs get their line when they are fixed. BUG-022 (L-100) and BUG-045 (possibly L-101) are already cited as sources by those lessons. The lessons cited above list these bugs among their sources (step 6).
BUG-543 — The capture app’s dialogs and sheets do not lock the page’s scroll on deployed environments: the CSP blocks Radix’s scroll-lock style
Affects: F36-a, F36-c (forge-et-pwa every shadcn Dialog and Sheet, since altrium-io/forge-et-pwa#201, row 4 of the shadcn move) · Found in: frontend-reviewer on altrium-io/forge-et-pwa#213 (row 11), 2026-11-22, by reading the code and the deployed policy · Severity: low · Owner: unclaimed
Repro: on capture.dev.crm.empire.lk, open any sheet (the bell’s, Sort and Filter, the booth drawer) and look in the console: a Content-Security-Policy violation for an inline <style>; then scroll the page behind.
Expected: while a modal Dialog or Sheet is open, the page behind it does not scroll, as Radix’s scroll lock (react-remove-scroll-bar) arranges, and no policy violation is reported.
Actual: the scroll lock adds a <style> element at run time (its CSS includes the scrollbar width, worked out at run time, so a hash cannot cover it), and the capture app’s policy is style-src 'self' (environments-and-identity.md §4), so the browser blocks it and the page behind can scroll. The page behind is still inert (BUG-537’s pageBehind), so nothing behind can be pressed or focused. owed.md’s CSP row says the capture app adds no <style>, which stopped being true at row 4. Sonner’s injected <style> (row 11) is blocked the same way; #213 bundles Sonner’s stylesheet so its rules do not depend on it.
Fix: not yet. A forge-et-iaac PR to dev that adds 'unsafe-inline' to the capture app’s style-src, as the CRM’s policy has (a DevOps decision, Shyamin Ayesh’s approval), then a check on dev that the scroll lock works and no violation is reported.
BUG-542 — Sort and Filter gives focus to nothing on close when Reset removed the halls summary button that opened it
Affects: F36-a, F36-c (forge-et-pwa src/features/plan/exhibitors/PlanList.tsx, the SortFilterSheet it renders with no fallbackFocus) · Found in: frontend-reviewer on altrium-io/forge-et-pwa#210 (row 9 of the shadcn move), 2026-11-22, by reading the code · Severity: low · Owner: unclaimed
Repro: on Plan with a hall filter on, open Sort and Filter from the summary line’s halls button (plan-summary-halls), press Reset, then close the sheet.
Expected: focus goes back to a control in the list, for example Filter.
Actual: Reset removes the halls button, which only shows while a filter is on, so the sheet’s opener has left the page; PlanList passes the sheet no fallbackFocus, so focus falls to <body>. The older native Sheet behaved the same, so this is not row 9’s doing.
Fix: not yet. Pass fallbackFocus={filterRef} to SortFilterSheet in PlanList.tsx, with a test that opens it from the halls button, resets and closes.
BUG-541 — Plan’s Try Again, for an exhibitor list that never loaded, unmounts as it is pressed
Affects: F36-a (forge-et-pwa src/features/plan/exhibitors/PlanList.tsx, the plan-list-error panel’s Try Again) · Found in: frontend-reviewer on altrium-io/forge-et-pwa#210 (row 9 of the shadcn move), 2026-11-22, by reading the code · Severity: low · Owner: unclaimed
Repro: open Plan with the exhibitor list’s first read failing, then press Try Again in the error panel.
Expected: Try Again stays on the screen, busy, while the list is read again, and keeps focus.
Actual: the panel shows only while list.data === undefined && list.isError. TanStack Query v5’s fetch sets status: 'pending' and clears error when there is no data, so the press replaces the panel with the loading skeleton and the focused button leaves the page (L-183’s pattern). This predates row 9.
Fix: not yet. Hold the error panel while a retry it started is in flight (as L-183’s other fixes do), with a test that the button stays and keeps focus.
BUG-540 — Flaky test: Plan’s AC-28 sometimes loses the second tier chip pressed in Sort and Filter (Windows)
Affects: F36-a, F36-c (forge-et-pwa e2e/browser/plan/plan-research.spec.ts:87, “AC-28: tiers and a hall send one tier each, after sort and before hall; …”) · Found in: the Chromium run of e2e/browser/plan for the capture app’s shadcn row 9, Windows 11, 2026-11-22 · Severity: low — test code or a timing gap, not yet known · Owner: unclaimed
Repro: in forge-et-pwa, pnpm exec playwright test e2e/browser/plan/plan-research.spec.ts -g "AC-28" --project=chromium --retries=0 --repeat-each=24 --workers=8.
Expected: the test passes on every run: pressing Hot, Warm and Hall 4 reads the list with tier=hot&tier=warm&hall=Hall+4.
Actual: the last list query is page=1&size=50&status=all&sort=best_fit&tier=hot&hall=Hall+4: Warm is lost. On main at 456bc01 (the older Sheet), 1 of 48 failed; on row 9’s branch (the shadcn Sheet), 2 of 48, and it failed once in each of two whole-folder runs. Run alone, it passed 6 of 6. So it is not row 9’s doing, and load makes it show.
Fix: not yet. Find whether Warm’s press is lost (a chip re-rendered between Playwright’s locate and click) or its change is overwritten by a later one (the latest-params path BUG-431 fixed), with a trace from a failing run.
BUG-539 — Flaky test: the held-lead ready-notice test sometimes sees no fourth read after the fake clock moves 25 s (Windows)
Affects: F36-c (forge-et-pwa e2e/browser/notifications/ready-notice.spec.ts, “a held lead chosen in the sheet opens its page, with no notice for it”, line 324) · Found in: the Chromium browser project run on fix/BUG-537-inert-page-behind-dialog, Windows 11, 2026-11-22 · Severity: low — test code only · Owner: unclaimed
Repro: in forge-et-pwa, pnpm exec playwright test e2e/browser/notifications/ready-notice.spec.ts -g "a held lead chosen in the sheet" --project=chromium --retries=0 --repeat-each=6.
Expected: the test passes on every run.
Actual: expect.poll(() => api.listReads(), { message: 'the 20 s poll read N-1 once' }).toBe(4) received 3. It failed once in the whole Chromium project run (756 tests), once in 3 runs alone, then passed 6 of 6, on the BUG-537 fix branch; on main at 29d7919 it passed 6 of 6. The fix changes nothing on this test’s path (frontend-reviewer on altrium-io/forge-et-pwa#209), so the flake is the test’s own and is on main too. The likely cause: TanStack Query v5 restarts refetchInterval on every query update (queryObserver onQueryUpdate → #updateTimers). listReads() counts a read when its request arrives, so the item’s mark-read re-read counts as the third read before its answer lands. When page.clock.fastForward('00:25') runs before that answer, the interval restarts when the answer arrives, and the next poll is 20 s of real time away, past the poll’s 5 s wait.
Fix: not yet. Start a page.waitForResponse for the mark-read re-read before the item’s click, await it, then move the clock.
BUG-538 — Flaky test: the customer check’s accessibility test releases each held check before the request has arrived (Windows)
Affects: F36-d (forge-et-pwa e2e/browser/leads/lead-customer-check-a11y.spec.ts, “one role=“status” node holds Checking the CRM for …, the answer and Checking again…, and focus does not move”, lines 83 and 94 on main) · Found in: frontend-reviewer’s review of forge-et-pwa#208 (BUG-491’s fix), 2026-11-22, by reading the code; then reproduced on Windows 11 · Severity: low — test code only · Phase: 1
Renumbered: filed as BUG-536 on fix/BUG-491-BUG-495-BUG-534, renumbered BUG-538 on 2026-11-22 because forge-et-harness#1640, opened earlier, holds BUG-536 and another branch holds BUG-537 (.claude/rules/id-numbering.md). The fix commit d1652a3 (fix(BUG-536): …) on forge-et-pwa#208 still says BUG-536.
Repro: a temporary copy of the test that delays every check request 1.5 s before it reaches the double (a page.route registered in openLeads’s before hook that waits, then calls route.fallback()), Chromium, no retries, --repeat-each=3.
Expected: the test passes whatever else the machine is running.
Actual: 3 of 3 failed: toContainText('Palmgate Supermarkets is an account · owner Dana Owner') timed out.
Cause: BUG-491’s cause, in the sibling spec. The test holds the check, sees Checking the CRM for … and calls check.release('account'); later it sees Checking again… after the reconnect and calls check.release('not-found'). Both texts show before the request reaches the double, so under load release() finds nothing held, the request is then held with no release to come, and the next answer never shows.
Fix: before each release, wait until the double has received that request: await expect.poll(() => checksOf(check, BADGE)).toBe(1), then toBe(2), with the same checksOf helper as lead-customer-check.spec.ts.
Regression test: the fixed test itself: with every check request delayed 1.5 s, the old test failed 3 of 3 and the fixed one passed 3 of 3. The whole e2e/browser/leads folder then passed 350 of 350 (Chromium, no retries).
Resolved: 2026-11-22 — fixed in forge-et-pwa#208 (commit d1652a3).
Lesson: L-307 — a second instance in the same feature, found by review after BUG-491’s fix; the reviewer’s check 14 already names the pattern.
BUG-537 — A node mounted after a capture-app shadcn Dialog or Sheet opens, beside a live region, stays reachable behind it
Affects: F36-a, F36-c (forge-et-pwa src/ui/shadcn/dialog.tsx, InertWhileOpen; the Sheet uses the same) · Found in: frontend-reviewer’s delta review of altrium-io/forge-et-pwa#207, 2026-11-22 · Severity: medium — since #207 it fails 9 browser page checks on main · Owner: Nanndish
Repro: open the campaign sheet on Plan, then make Plan mount a new child of its section.screen while the sheet stays open (for example plan-message), and read the new node: it has neither aria-hidden nor inert. The same holds for new content in the lead’s page while a sheet is over it, and for a route that finishes loading in place of RouteSkeleton while the bell’s sheet is open.
Expected: everything behind an open modal dialog is out of the accessibility tree and cannot be pressed, including content that mounts after it opened, as the native modal <dialog> made it (design-system §9).
Actual: Radix’s hideOthers leaves every ancestor of a live region unhidden and hides only their other branches, once, when the dialog opens; InertWhileOpen reads its marks and the live regions (liveRegionsBehind) once, at the same moment. A node mounted later into one of those ancestors is neither hidden nor inert. The window is narrow, and it was there before #207.
Since #207 it also fails the browser page check. On Plan, <main> holds the polite lead region, so Radix leaves <main> unhidden; #207 made the region itself inert, which left <main> a scrolling region with nothing focusable in it, and axe refuses it (scrollable-region-focusable, serious). Row 9’s run of e2e/browser/plan in Chromium found it on 2026-11-22: 9 campaigns-a11y page states on main, and 12 of row 9’s sheet states. #207 skipped the browser run (L-341).
Fix: while a Dialog or Sheet is open, InertWhileOpen makes inert and aria-hidden each child of <body> that does not hold its content, is not one of Radix’s focus guards and is not inert yet (pageBehind, replacing #207’s liveRegionsBehind), and removes only the marks it set. The whole page is then out of reach, including what mounts later, as the native modal <dialog> made it. The toasts and the Ready to review notice behave as before: they carry no aria-live, so Radix already hid them, and an open popover is still hidden by hideOpenPopovers (pwa-shadcn.md D-3).
Regression test: dialog.test.tsx makes the page round a live region behind it inert and hidden … and keeps a node mounted beside a live region after it opened inert. Both failed on main’s dialog.tsx (toHaveAttribute("inert"); expected null not to be null) and pass with the fix. PlanRoute.test.tsx’s two plan-lead tests now read the inert ancestor.
Resolved: 2026-11-22 — fixed in altrium-io/forge-et-pwa#209.
Lesson: L-341 — the regression reached main because #207, which changed the shared Dialog, skipped the browser run; row 9’s own run of the Plan specs found it.
BUG-536 — A live region behind a CRM shadcn Dialog or AlertDialog stays reachable by a screen reader: Radix never hides [aria-live]
Affects: F1-b, F20-a, and every CRM screen that opens a shadcn Dialog or AlertDialog over a page with an aria-live region (forge-et-fe src/ui/shadcn/dialog.tsx, alert-dialog.tsx) · Found in: frontend-reviewer on altrium-io/forge-et-pwa#207, 2026-11-22 · Severity: low · Owner: unclaimed
Repro: put an aria-live region holding a button on a page, open a shadcn Dialog over it, and read the region’s element: it has neither aria-hidden nor inert, while the rest of the page has both. The capture app’s dialog.test.tsx test makes the page round a live region behind it inert and hidden … is this case. The CRM has many aria-live regions (git grep aria-live -- 'src/**/*.tsx'); which of them sit behind a dialog is not yet listed.
Expected: everything behind an open modal dialog is out of the accessibility tree and cannot be pressed (design-system §9; WCAG 2.1 SC 1.3.2, 2.4.3).
Actual: Radix hides the page with the aria-hidden package, whose hideOthers never hides an [aria-live] element or anything inside it (its own comment cites theKashey/aria-hidden#10). InertWhileOpen makes inert only what Radix hid, so a live region behind the dialog, and any control inside it, stays reachable by a screen reader’s virtual cursor. The dialog’s focus trap still holds the keyboard.
Fix: not yet. Copy the capture app’s fix (pageBehind in src/ui/shadcn/dialog.tsx, BUG-537’s fix, which makes inert and hidden each child of <body> that does not hold the open dialog and is not a focus guard, and removes only the marks it set) into the CRM’s dialog.tsx and alert-dialog.tsx, with the same test (pwa-shadcn.md D-2: a fix made in one repo is copied to the other by hand). It is not a plain copy: the CRM’s toasts are Sonner’s, whose <section> carries aria-live, so making it inert would silence a toast raised while a dialog is open. Leave the child of <body> that holds Sonner’s section out, or choose another approach for the CRM.
BUG-535 — The capture app’s lost-convert browser test expects no aria-disabled on Convert to Lead, which the shadcn Button always renders
Affects: F36-a (forge-et-pwa e2e/browser/leads/leads-offline.spec.ts, “a convert whose answer is lost says Not confirmed, then It didn’t go through. with Convert offered again”; the button moved to the shadcn Button in forge-et-pwa#204) · Found in: a run of e2e/browser/leads on main 1b446e0 while fixing BUG-491, Windows 11, 2026-11-22 · Severity: low — test code only, but it fails every run, so main’s browser suite is red · Phase: 1
Repro: in forge-et-pwa on main, pnpm exec playwright test e2e/browser/leads/leads-offline.spec.ts -g "a convert whose answer is lost" --project=chromium --retries=0 --repeat-each=3.
Expected: the test passes: after It didn’t go through., Convert to Lead is offered again, enabled.
Actual: 3 of 3 failed at line 236, not.toHaveAttribute('aria-disabled'): the button has aria-disabled="false". It failed the same way on e0efe1a, the merge before forge-et-pwa#205, so it dates from #204.
Cause: the shadcn Button always renders aria-disabled and aria-busy, "false" when not set (src/ui/shadcn/button.tsx), so that a reader can tell an enabled button from one that is never disabled. Convert to Lead moved to it in forge-et-pwa#204. #205 updated the same file’s aria-busy checks (lines 66 and 69) for this reason, but not this aria-disabled check. The button itself is right; only the test was out of date.
Fix: expect toHaveAttribute('aria-disabled', 'false').
Regression test: the fixed test itself: it failed 3 of 3 on main and passes 5 of 5 with the fix (Chromium, no retries).
Resolved: 2026-11-22 — fixed in forge-et-pwa#208 (commit 6885507).
Lesson: L-341 — a move to shadcn changed what the browser tests see, and the moving PR did not run the screen’s browser specs.
BUG-534 — The capture app’s leads browser tests cannot find the Leads heading or a To Review row behind the open lead drawer
Affects: F36-a (forge-et-pwa e2e/browser/leads/leadsSpecSupport.ts openLeads, used by every leads spec that opens a /?lead= path; e2e/browser/leads/card-read-again.spec.ts, AC-9’s “sends { version } only …”; the drawer moved to the shadcn Sheet in forge-et-pwa#205) · Found in: runs of lead-customer-check.spec.ts and e2e/browser/leads on main 1b446e0 while fixing BUG-491, Windows 11, 2026-11-22 · Severity: medium — test code only, but main’s browser suite fails on every run · Phase: 1
Repro: in forge-et-pwa on main, Chromium, no retries, on an idle machine:
lead-customer-check.spec.ts -g "the re-read answered 503" --repeat-each=5failed 2 of 5.card-read-again.spec.ts -g "sends \{ version \} only" --repeat-each=3failed 3 of 3. Expected: both pass. Actual:openLeadstimed out atexpect(page.getByRole('heading', { level: 1, name: 'Leads' })).toBeVisible(). The failure snapshot shows only the lead’s open drawer in the accessibility tree.getByRole('listitem').filter(…).getByTestId('lead-row-title')found no element. One0efe1a, the merge before forge-et-pwa#205, the same tests passed 8 of 8 and 3 of 3. Cause: since #205 the lead drawer is a modal shadcnSheet, which makes the page behind it inert while it is open, so role queries no longer find what is behind it. With a/?lead=path,openLeadsreads the heading afterpage.goto. It passes when the heading is read before the drawer opens, and fails when the drawer opens first. The card-read-again test reads the To Review row while the drawer is open, so it fails every time. The page is right: content behind a modal sheet should be inert. The same cause made the bell’s accessible name empty behind the notifications sheet after forge-et-pwa#201 (BUG-525, and BUG-523’s fix commitcb55537). Fix: both queries find the element hidden or not (includeHidden), with a comment saying the drawer makes the page behind it inert, aslead-page.spec.ts:137does for the row’s outcome image. InopenLeadsit applies only when the path opens a lead (includeHidden: path.includes('lead='), withexact: true), so a page left inert on/is still caught there (frontend-reviewer’s Nit). The other waits on the Leads heading load/with no drawer open. Regression test: the fixed tests themselves: onmainthey failed 2 of 5 and 3 of 3; with the fix each passes 5 of 5 (Chromium, no retries). The wholechromiumbrowser project then passed 755 of 756 before BUG-535’s fix; the one failure was BUG-525, open onmain. Resolved: 2026-11-22 — fixed in forge-et-pwa#208 (commit5c9710d). Lesson: L-341 — a move to a modal Radix sheet made the page behind it inert, and the moving PR did not run the screen’s browser specs.
BUG-527 — The capture app’s selected tab is a white card on a pale track, unlike the CRM’s filled view tabs
Affects: F36-a (forge-et-pwa src/ui/Tabs.tsx and Tabs.module.css: Leads’ To Review / Completed and Capture’s Business Card / Badge QR) · Found in: Nanndish’s review of the built capture app, 2026-11-20 · Severity: low · Phase: 1
Filed as BUG-522 and renumbered, because older open harness PRs (forge-et-harness#1602, #1604) add that number (id-numbering rule 5); the app PRs’ commits and the merged forge-et-be#499 name it BUG-522.
Repro: open Leads or Capture.
Expected: the tabs look like the CRM’s saved-view tabs (design-system §6, Rev 141): the warm track, and the selected tab filled in the primary button’s colour with white text.
Actual: the selected tab is a white card with an outline on a light track, which is easy to miss in bright light; Capture’s track also has an outline that Leads’ has not (Rev 142).
Fix: three options were drawn (the CRM’s pill tabs, one outlined bar, separate tiles); Nanndish chose the first. The shared Tabs draws the track in --et-accent with no outline and the selected tab in accent/gold (the CRM’s primary, #8a5a12) with accent/gold-on text; a count pill stays surface/card, with accent/gold figures on the selected tab. The bordered option (BUG-486) is removed, since no track has an outline now. themeContrast.test.ts measures the new pairs: the text on a tab not selected on the track, the selected count, and the selected fill against the track at 3:1 (SC 1.4.11). The move of these screens onto shadcn’s Tabs (.forge/design/ui/pwa-shadcn.md rows 5 and 6) keeps this look.
Resolved: forge-et-pwa#200 (Nanndish). Measured at 390 × 844 in Chromium on the leads and capture page states: the selected tab rgb(138, 90, 18) with white text, 44px tall, on an rgb(244, 238, 226) track with no border.
Lesson: none — one-off: a design choice made after seeing the built screens, not a defect that can recur.
BUG-526 — The capture app’s Settings shows a version line the reps do not need
Affects: F36-a (forge-et-pwa src/features/settings/SettingsRoute.tsx, src/app/router.tsx, src/main.tsx) · Found in: Nanndish’s review of the built capture app, 2026-11-20 · Severity: low · Phase: 1
Filed as BUG-520 and renumbered, because older open harness PRs (forge-et-harness#1602, #1604) add that number (id-numbering rule 5); the app PRs’ commits and the merged forge-et-be#499 name it BUG-520.
Repro: open Settings.
Expected: the install card (when there is one) and the account, nothing else.
Actual: Version 0.1.0 under the account card.
Fix: remove the line, its prop and its style. The release rule that a deployed bundle reports its version (.claude/rules/release-and-branch-protection.md § Release Process) is kept as the CRM keeps it since F5-b: main.tsx stamps document.documentElement.dataset.appVersion. The prototype’s Settings screen drops its line too. Regression test: SettingsRoute.test.tsx checks no Version line shows.
Resolved: forge-et-pwa#200 (Nanndish). Checked on the settings page state: no line, and data-app-version reads 0.1.0.
Lesson: none — one-off: a line the screen carried from the prototype, removed on request.
BUG-525 — With the notifications sheet open, the bell’s accessible name is empty, failing the Mark All as Read browser test on main
Affects: F36-c (forge-et-pwa src/features/notifications/NotificationsBell.tsx, NotificationsSheet.tsx, as moved to shadcn/ui by forge-et-pwa#201) · Found in: forge-et-pwa#200’s selected browser run, 2026-11-21, then on main 3bf83c3 · Severity: medium — main’s browser suite fails this test on every run · Phase: 1
Repro: in forge-et-pwa at origin/main 3bf83c3, E2E_GATE=1 pnpm exec playwright test e2e/browser/notifications/notifications.spec.ts:182 --project=chromium --retries=0 (Windows, 2026-11-21; E2E_GATE=1 so the run starts its own server instead of reusing another worktree’s).
Expected: after Mark All as Read, the bell is named Notifications (toHaveAccessibleName('Notifications'), F36-c AC-34).
Actual: Received: "". The locator resolves to the shadcn <button data-variant="ghost" aria-expanded="true" aria-haspopup="dialog">, so the bell keeps its aria-label, but its computed name is empty while the sheet is open, because the shadcn sheet makes everything behind it inert (forge-et-pwa#201). The seam test e2e/seam/notifications-screens.seam.spec.ts:502 (F36-c AC-34 smoke, “the read after Mark All as Read: the bell’s name”, Expected: "Notifications, 42 unread", Received: "") fails the same way against a running backend (forge-et-pwa#199’s selected browser run, 2026-11-21). forge-et-pwa#199’s clear tests read the bell’s aria-label instead, with comments naming this bug.
Fix: not yet. Decide whether the tests should close the sheet before reading the bell’s name, or whether the sheet should leave the header reachable as it did before forge-et-pwa#201; then make the tests and the code agree, and bring back toHaveAccessibleName in #199’s clear tests.
Lesson: none yet.
BUG-524 — The booth note card on Capture keeps an empty line under its box
Affects: F36-a (forge-et-pwa src/ui/Field.tsx, src/features/capture/CaptureRoute.tsx) · Found in: Nanndish’s review of the built capture app, 2026-11-20 · Severity: low · Phase: 1
Filed as BUG-521 and renumbered, because main’s BUG-521 (the nightly run lock) was filed first; forge-et-pwa#200’s first commits name it BUG-521.
Repro: open Capture at phone width and look at the Booth Note card.
Expected: the card ends just under the text box, as the prototype’s other cards do.
Actual: about 26px of blank space sits between the box and the card’s edge. The shared Field always draws the line under its control, empty or not, so an error appearing does not move what is below (design-system §5). The booth note has no helper, and its error (over 2,000 characters, or a control character) is rare, so the line is nearly always empty.
Fix: Field gains reserveMessageLine (default true); with false it draws the line only while there is an error or a counter under the control. The booth note passes false; every other field keeps the line. An error or the count from 1,800 characters now pushes Submit for Research down by one line when it appears. Regression tests: Field.test.tsx and CaptureRoute.test.tsx (“leaves no empty line under the box…”).
Resolved: forge-et-pwa#200 (Nanndish). Measured at 390 × 844 in Chromium on the capture-card-taken page state: 13px from the box to the card’s edge (its 12px padding and 1px border).
Lesson: none — one-off: a field with no helper was given the reserved line meant for fields with one; no other field has that combination.
BUG-523 — Capture-app notifications cannot be cleared, so read ones pile up in the bell for 30 days
Affects: F36-c (forge-et-be capture/notification; forge-et-pwa src/features/notifications/) · Found in: Nanndish’s review of the built capture app, 2026-11-20 · Severity: medium — every rep’s bell fills with read items during a show · Phase: 1
Filed as BUG-523; for a while renumbered to BUG-528 while forge-et-harness#1604 held 523, then back once #1604 moved off it (id-numbering rule 5). forge-et-be#499 names it BUG-523; some of forge-et-pwa#199’s commits name it BUG-528.
Repro: receive a few notifications, open the bell and press Mark All as Read.
Expected: a rep can clear notifications they are done with, one at a time or all at once.
Actual: read notifications move to Earlier and stay there for 30 days (up to 50); the sheet offers no way to remove them.
Fix: F36-c spec Rev 11 (FR-ST2, FR-NT5, FR-NT7, FR-AU1, FR-AU2, AC-61). Backend: a migration adds capture_notification_recipient.cleared_at with CHECK (cleared_at IS NULL OR read_at IS NOT NULL), mirrored in docs/data-model.md §13.5 and docs/proofs/pricing/01-schema.sql; the list’s window leaves cleared rows out; POST /api/v1/capture/notifications/{id}/clear and POST …/clear-all with { upToId } (the same body reader as read-all; it clears every notification of the last 30 days up to upToId, not only the 50 shown) set cleared_at, and read_at where still null, under the same row locks as the mark-read calls, and record one capture.notifications_cleared event (notificationIds, readNotificationIds) in the same transaction when they clear anything. Capture app: N-4 and N-5 in notificationsApi.ts, two mutations beside the mark-read ones (remove from the cached list on 204, re-read on every outcome), Clear All in the sheet’s title row, a clear button on each item, the offline line moved under the title row and reworded, a visible failure line, the focus moves, and page states for the new states. Regression tests: backend integration specs for both endpoints, the window, the 404, idempotence, the audit event and the constraint; Vitest for the sheet and the mutations; a browser spec for both controls, focus and offline.
Resolved: forge-et-be#499 (migration V81, the endpoints and the event; CaptureNotificationClearIT 17 tests) and forge-et-pwa#199 (the sheet; 52 notification browser tests, 240 size-matrix checks on the sheet’s states) (Nanndish). Checked at 390 × 844 in Chromium on the notifications-sheet, notifications-sheet-clear-failed and notifications-sheet-offline-read page states.
Lesson: none — one-off: the bell was specified without a way to remove items (F36-c D-10 kept every row), not a defect in how it was built.
BUG-532 — Every dev.yaml plan shows the SES configuration set’s max_delivery_seconds 0 -> null, so no plan reads No changes
Affects: F38-e (AC-3’s no-change plan), F38-a (forge-et-iaac base/ses/identity.tf, aws_sesv2_configuration_set.this) · Found in: F38-e-WI-3.1, 2026-11-21 · Severity: low · Owner: Tharinda
Repro: run dev.yaml on forge-et-iaac dev (gh workflow run dev.yaml --repo altrium-io/forge-et-iaac --ref dev) and read its terraform::plan step. Run 44945483585 (2026-11-21, manual, fd846ff) planned 0 to add, 1 to change, 0 to destroy: module.ses.aws_sesv2_configuration_set.this updated in place, - max_delivery_seconds = 0 -> null. The push runs 44488259851 (2026-11-19) and 44678328363 (2026-11-20) show the same line.
Expected: a plan on an unchanged dev reads No changes. (F38-e spec AC-3).
Actual: the configuration set has no delivery_options block in Terraform; AWS reports max_delivery_seconds = 0, so every plan removes it and every apply leaves it to come back. The estate is not changed in any way that matters, but no plan can show No changes., so F38-e AC-3 cannot be met.
Fix: not yet. A forge-et-iaac PR to dev that stops the repeated diff — for example lifecycle { ignore_changes = [delivery_options] } on the configuration set, since AWS accepts only 300–50400 for max_delivery_seconds and so 0 cannot be written — then one manual dev.yaml run that reads No changes., recorded as F38-e-WI-3.1 R-2 attempt 2. Needs Shyamin Ayesh’s approval (.claude/rules/forge-et-iaac.md).
BUG-521 — The nightly full run never gets the machine’s run lock while sessions pass it from one run to the next
Affects: FS-4 (.claude/lib/nightly/run-nightly.py acquire_lock, and every other taker of et-heavy-run.lock: .claude/lib/heavy-run.py, forge-et-fe and forge-et-pwa scripts/browser-gate/machine.mjs, forge-et-be scripts/local-build-evidence.py) · Found in: the nightly full run of 2026-11-21 on the lead’s Mac (forge-et-be#493, forge-et-fe#602, forge-et-pwa#197) · Severity: high — the night ran no suite, so main went untested · Phase: 1
Repro: the lead’s Mac, 2026-11-21. The nightly run started at 02:00:08 and found the lock held. pmset -g log shows the wake locks of the runs that held it during its hour: one ended at 02:03:20 and the next began at 02:03:22, then handoffs at 02:15:41, 02:15:44 and 02:16:01–02:16:03 with no gap or two seconds, then one run from 02:16:56 to 03:07:25. The nightly run gave up at 03:00:34 and reported every suite not-run — another heavy run held the machine for an hour.
Expected: the nightly run waits for the run that holds the lock and then takes it, so the night runs.
Actual: the nightly run looked at the lock every 30 seconds, and every other taker every 10 seconds or at its own start. Each time the lock was released, another session’s run (often the next command of the same session) took it within two seconds, so the nightly run never found it free. Nothing gave the nightly run the next turn.
Fix: the nightly run reserves the next turn while it waits: it writes et-heavy-run.next beside the lock (the lock’s JSON form) and removes it when it takes the lock or the wait ends, and it looks every 5 seconds. Every other taker leaves a free lock to a reservation that names a live process other than itself, written since the last boot, less than 90 minutes ago and not more than a minute ahead, and waits as it would for a holder; the reservation never hides a lock the taker already holds, and a run started by the holder (ET_HEAVY_RUN_LOCK_HOLDER) is not affected. Built in the harness (run-nightly.py, heavy-run.py; forge-et-harness#1624, #1626), forge-et-fe and forge-et-pwa (machine.mjs, one shared copy; forge-et-fe#603, forge-et-pwa#198) and forge-et-be (local-build-evidence.py; forge-et-be#495).
Resolved: 2026-11-21 — fixed in forge-et-harness#1624 and #1626, forge-et-fe#603, forge-et-pwa#198 and forge-et-be#495. Each taker’s new tests fail without the guard they cover: forge-et-fe’s “never hides a lock this process already holds” failed on 071460d7, and the build record’s cases each failed with their guard removed (listed on forge-et-be#495). The nightly run’s own case checks it writes the reservation while it waits and removes it when the wait ends. A by-hand nightly run on 2026-11-21 after the merges is the end-to-end check.
Lesson: L-337 — a job that must run on a shared lock gets a reservation every other taker honours.
BUG-519 — Tests that choose a row-menu item through a helper check its result before the menu has run it
Affects: F1-e, F1-d, F1-b, F1-f, F11-f, F26-a, F20-g (forge-et-fe tests; the menu is src/ui/recordRow.tsx RowActionsMenu, on main) · Found in: frontend-reviewer’s delta review of forge-et-fe#598 (BUG-489’s fix), 2026-11-20 · Severity: low, a test-only race; no failure seen yet · Phase: 1
Repro: in forge-et-fe, change RowActionsMenu’s onCloseAutoFocus to run trigger.focus() and action.run() inside setTimeout(…, 50) (a temporary change), then run pnpm exec vitest run. On fix/BUG-489-accounts-view-wait at f6ecdffd, Windows, 2026-11-20, the suite failed 72 of 16,061 tests in 12 files; two node files (src/app/knownFailureMarks.test.ts, src/app/contactPanels.test.ts) read source files as text and fail on the temporary change itself, which leaves the 70 below.
Expected: a test that chooses a menu item waits for what the item does before it checks it.
Actual: BUG-489’s cause. RowActionsMenu runs the chosen item from Radix’s onCloseAutoFocus, on a timer after the menu has gone, and these tests choose the item through a helper of their file (choose, rowAction, chooseRowAction, chooseNoteAction) that returns without waiting, then read the result (a callback, a dialog, an edit box, the router’s path, focus) at once; a few check only that something is absent, which passes before the item has run. The delay run fails: RecordNotesSection.test.tsx 26, LeadNotesSection.test.tsx 23, AccountContactsPanel.writes.test.tsx 11, ContactNotesAndTagsAction.test.tsx 3, PricingRoute.test.tsx 2, and one each in AccountContactsPanel.test.tsx, AccountListScreen.test.tsx, ContactsTable.test.tsx, ScheduledEmailsRoute.test.tsx and AuctionSalesTab.test.tsx. BUG-489 (forge-et-fe#598) fixed the tests in six other files that chose the item inline.
Fix: not yet. Wait at each call site on the visible result (findByRole for the edit box or the dialog, waitFor on a callback or the path), with an absence check placed after a positive wait; not a generic wait for the trigger’s focus inside the helpers, because a note’s Edit moves focus on into its edit box. Prove it with the same 50 ms delay run over the whole suite. As made:
- Where every call of a helper expects the same result, the wait is in the helper, on that result:
RecordNotesSection,LeadNotesSectionandContactNotesAndTagsAction: Edit waits for theEdit Notebox, Delete for the delete question;AccountContactsPanel.writes: Archive waits forArchive Contact?.
- Every other call site waits on its own result:
ContactsTable:waitForon the handler;AccountContactsPanel.testandcontactPanels.test.ts:findByfor the question;PricingRoute:findByTestIdfor each reopened dialog. Its double-Restore test waits for the trigger’s focus, set in the same step as the item runs and the one sign of an item that opens nothing, before checking one PATCH;AccountContactsPanel.writes’ 403 test: waits for the PATCH before moving focus away;ScheduledEmailsRoute’s left-the-screen test: waits for the trigger’saria-busy;AccountListScreen’s Edit: reads the dialog once the account is in it, because the loaded form draws its own dialog frame in place of the loading one. Resolved: 2026-11-21 — fixed in forge-et-fe#604 (mergece2833bc;afdea29f, review fixes641b2621).
- The proof, on macOS: the whole suite with the 50 ms delay failed 69 of 16,061 tests on
main’s test files:RecordNotesSection26,LeadNotesSection23,AccountContactsPanel.writes11,ContactNotesAndTagsAction3,PricingRoute2, and one each incontactPanels.test.ts,AccountContactsPanel.test,AccountListScreenandContactsTable. On the fix it failed 0.ScheduledEmailsRouteandAuctionSalesTab, listed from the Windows run, did not fail on the Mac. - The review:
frontend-reviewerfound one Important. The double-Restore test could pass without testing the once-per-row guard. With the guard removed it now fails (got 2). - On the merged head: the test record passed 3,766 tests, and the selected browser run selected none (a unit-test-only change). Lesson: L-101 — another instance, with BUG-489: the test checked before the page had settled, because the row menu runs its item on a timer after the click.
BUG-518 — The capture ranking answers in prose instead of its answer tool for a subject with little detail, so the subject is left unranked
Affects: F36-c (forge-et-be capture-ranking prompt v2, used by plan-ranking and capture-ranking; BedrockAiProvider’s toolChoice auto) · Found in: the live v2 evaluation on dev, 2026-11-20 (.forge/cloud-gated-validation.md §B F36-c item (k)) · Severity: medium · Phase: 1
Repro: AWS_PROFILE=et-dev ./gradlew liveEvaluation -Pfeature=plan-ranking -PconfirmBedrockCalls=true on forge-et-be main b33e11b8 (costs about US$0.15; ask the lead first). Cases lead-vague-note-no-company and exhibitor-no-research ended INVALID_OUTPUT on both attempts, each logged NO_TOOL_USE stopReason=end_turn.
Expected: every ranking call answers through the ranking tool, so every subject gets a priority, however little the input holds.
Actual: for a subject with no company or no research, Haiku 4.5 answered in plain text; the call ends INVALID_OUTPUT and the subject stays unranked. The same would happen to a real exhibitor or captured lead with as little detail.
Fix: not decided. BedrockAiProvider sends toolChoice auto on purpose (its class comment); the choices are a prompt change that tells the model to always call the tool, or a forced tool choice for these two features, each checked by a new live run. Part of the lead’s decision on the failed v2 run (owed.md).
Lesson: none yet.
BUG-517 — Flaky tests: capture-app page states miss the 5-second ready wait or the page setup on a loaded machine (macOS)
Affects: F40 (forge-et-pwa e2e/browser/size-matrix/capture-layout.spec.ts, AC-14b and AC-15b, and size-matrix.spec.ts; test code, and possibly the page states’ arrangements in e2e/pageStates.ts) · Found in: the weekly flaky-test report of 2026-11-20 (python3 .claude/lib/flaky-report.py), macOS · Severity: low — one occurrence for each state, every one passed on its retry · Phase: 1
Repro: no single command; each state failed once on a Mac and passed on its retry:
- forge-et-pwa#190’s selected browser run on
0ae2aab, 2026-11-20, load 0.91 average, 1.00 peak, 8 cores:plan-ended-view-plan-emptyAC-14b at 390×844 WebKit,Test timeout of 30000ms exceeded while setting up "page";lead-edit-details-offlineAC-15b at 1024×768 Chromium,getByTestId('edit-details-offline')not found after 5 s atcapture-layout.spec.ts:922. - forge-et-pwa#164’s gate, 2026-11-18, load 0.77:
booth-loadingAC-14b and AC-15b at 430×932 Chromium. - The nightly run on
main(load not in the report): 2026-11-16plan-campaigns-errorAC-14b at 360×780 WebKit; 2026-11-17sign-in-failedat 1280×800 WebKit (size matrix); 2026-11-19lead-remove-person-dialogAC-14b at 430×932 WebKit,leads-completed-no-matchAC-14b at 360×780 WebKit,lead-card-nothing-readAC-15b at 768×1024 WebKit,lead-ready-no-inputAC-15b at 1024×768 WebKit andlead-card-photo-unreadableat 430×932 WebKit (size matrix). The nightly logs are in~/.et-nightly/logs/<date>/pwa-gate.logon the lead’s Mac. - The same report lists Windows occurrences of other states (
booth-archived,booth-ranked-from-listat 844×390 Chromium, forge-et-pwa#165, load 0.84). They are single occurrences under load, left in the report; a Windows bug is filed if they recur.
Expected: each state reaches its ready test id within the 5-second ready wait, whatever else the machine runs.
Actual: see the repro: a different state each time, in both engines, mostly on loaded runs.
Fix: not yet. Triage in the order of .forge/test-strategy.md § Flaky Tests: read each failing run’s result.json load and stalls, and check for a sleep in the nightly window (pmset -g log). If every failure is load or sleep, the bug closes wontfix (step 4). If one state fails again on an idle machine, reproduce it under forced load and look for an arrangement that races, as BUG-495 does for the lead-check states on Windows.
Triage (2026-11-21):
- Forced load (step 2) did not reproduce it. The 13 states seen flaky, at the 10 size projects they failed at, ran twice each with
--retries=0: 740 tests, all passed. One busy loop per core ran beside them, started once the run held the heavy-run lock. The run was on capture-appmain66ed96e, macOS, 8 cores. - The one failing run with a recorded process list was overloaded by other work (step 4). That was forge-et-pwa#196’s selected gate on
054debf, 2026-11-21:- load: busy 0.30 at start, 0.89 average, load average 10.83 per core at the peak;
- the busiest process at the peak was Docker’s virtual machine (
com.apple.Virtualization.VirtualMachine, 102%), the containers beside the run; - three states failed once and passed on their retry.
sign-inat 360x780 WebKit stopped withTest timeout of 30000ms exceeded while setting up "page", before any app code ran.lead-refreshingat 844x390 WebKit andbooth-research-failed-offlineat 320x640 WebKit missed the ready test id.
- The same suite at lower load had no flaky tests. It passed twice on the same branch, at busy 0.82 and 0.86 on average.
- The other runs (forge-et-pwa#164 and #190, and the nightly runs) recorded loads of 0.77 to 0.91 but no process list.
Resolved: 2026-11-21 — closed
wontfixas an overloaded machine (.forge/test-strategy.md § Flaky Tests, triage step 4); decided by Tharinda (lead). - Why it is not fixed: forced load did not reproduce a failure in any of the 13 states, and the one failing run with a recorded process list was overloaded by work outside the run. A different state failed each time, and one failure came before the page existed, so nothing in a test or an arrangement could be changed.
- The tests are kept unchanged. A recurrence on a machine within the load limit is a new bug that links this one. Lesson: none — the overload outcome of triage step 4, as for BUG-451. Here the load rose after the start check, from containers that the heavy-run lock does not cover, as in BUG-455.
BUG-516 — The empty plan’s upload states flake in WebKit (macOS), a recurrence of BUG-458
Affects: F36-a (forge-et-pwa e2e/pageStates.ts plan-list-empty-reading and plan-list-empty-refused, in size-matrix.spec.ts and capture-layout.spec.ts), WebKit sizes only · Found in: the weekly flaky-test report of 2026-11-20, the nightly full run on the lead’s Mac · Severity: medium — they flake in the macOS nightly run on main, where a second failed attempt turns the night red · Phase: 1
Repro: the nightly run on main, macOS: on 2026-11-17, plan-list-empty-refused AC-15b at 320×640 WebKit (capture-layout.spec.ts); on 2026-11-20 (forge-et-pwa 9ddaddd), plan-list-empty-reading at 1280×800 WebKit (size-matrix.spec.ts). Each failed once and passed on its retry. The nightly logs are in ~/.et-nightly/logs/<date>/pwa-gate.log on the lead’s Mac. To reproduce, BUG-468’s command run on a Mac, with forced load as .forge/test-strategy.md § Flaky Tests step 2 describes, since BUG-458’s command passed 36 of 36 on an unloaded Mac.
Expected: each state reaches its ready test id (exhibitors-empty-upload-reading, upload-refused).
Actual: each failed once and passed on its retry; the nightly log has the error, BUG-468 the same symptom on Windows.
Fix: not yet. Waits on BUG-468’s cause, which is investigated once (the arrangement gives the file before the panel’s mount effects run, and StrictMode’s double mount aborts the read, BUG-458). This bug then proves that fix on macOS. BUG-458 was closed wontfix because it did not reproduce on a Mac, and its entry asks for a recurrence to be filed as a new bug that links it. The same report’s Windows occurrences of these states (forge-et-pwa#165’s gate, 2026-11-18) are BUG-468’s.
macOS proof of BUG-468’s fix (2026-11-21, Tharinda): BUG-468’s fix is forge-et-pwa#167 (a draft, 096efdc): arrangeUpload gives the file again when the entry is back at rest. Its own change (e2e/support/arrange.ts, e2e/pageStates.ts, its diff from its merge base) was applied to main at fe57499 and compared with main unchanged, on macOS, WebKit, E2E_GATE=1 … --project=size-{320x640,360x780,768x1024,1280x800}-webkit --retries=0 -g "plan-list-empty-(reading|refused)", under the heavy-run lock:
- one busy loop per core (8),
--workers=2 --repeat-each=5:main120 passed; with the fix, 120 passed. It did not reproduce at this load, as BUG-458 did not. - two busy loops per core (16),
--workers=4:mainat--repeat-each=10(load average 32 to 78) 1 failed, 239 passed —plan-list-empty-readingat 360×780,getByTestId('exhibitors-empty-upload-reading')not found after 5 s, its page snapshot showing the entry back at rest on Upload Exhibitor List, which is the cause above; with the fix at--repeat-each=20(load average 45 to 78) 480 passed.
So the fix holds on macOS. This bug closes with forge-et-pwa#167: when it merges, its harness half (forge-et-harness#1488) or a follow-up sets this bug fixed with fix_pr: "forge-et-pwa#167".
Lesson: none yet.
BUG-515 — Flaky test: edit-details.spec.ts’s lost-answer case reads the held PATCH before the request reaches the mock (macOS)
Affects: F36-b (forge-et-pwa e2e/browser/leads/edit-details.spec.ts:277, “a save whose answer is lost: Not confirmed, then the read shows it saved and the page says It went through.”; test code only) · Found in: the weekly flaky-test report of 2026-11-20: the browser gate runs of forge-et-pwa#164 (2026-11-18, load 0.77) and #190 (0ae2aab, 2026-11-20, load 0.91), macOS · Severity: low — a test-only failure; it passed on its retry both times · Phase: 1
Repro: forge-et-pwa#190’s run: TypeError: Cannot read properties of undefined (reading 'contactJobTitle') at edit-details.spec.ts:293, 1.3 s into the case. The case clicks Save, waits for the button’s aria-busy="true", then takes the first recorded PATCH (const [held] = patches(opened)) and reads held?.body. aria-busy is set by the click, before the browser’s request reaches the test’s route, so on a loaded machine patches(opened) can still be empty: held is undefined, and String(body.contactJobTitle) throws.
Expected: the case reads the PATCH only once the mock has recorded it.
Actual: see the repro.
Fix: wait for the request before reading it, with expect.poll(() => patches(opened).length).toBe(1) or page.waitForRequest on the PATCH, before const [held] = …; the same wait for any other case in the file that reads patches() right after a click. Not a longer timeout or a sleep (.forge/test-strategy.md § Flaky Tests).
Resolved: 2026-11-21 — fixed in forge-et-pwa#196 (merge 10edbcf2; the fix 3fd8039, review fixes e178a70).
- The fix: the lost-answer case polls the double’s PATCH list to 1 before reading the held PATCH.
- From
frontend-reviewer’s review:- the same case now waits for the sheet’s held re-read before releasing it;
card-read-again.spec.ts’s “30-second bound” rows wait for the held Read Again POST before aborting it.
- The proof, with temporary 1-second delays in the double:
- the old case failed with the reported
TypeErrorand passed 3 of 3 with the fix; - both “30-second bound” rows failed on the old spec and passed with the fix;
- the re-read race did not reproduce, and its wait follows the reviewer’s reading of the sheet’s effect.
- the old case failed with the reported
- Shared copies: the PR also carried forge-et-fe#600’s and #603’s shared-file changes, without which the capture app’s browser gate refused to run. It merged
mainafter forge-et-pwa#198, re-recording the copies from forge-et-fece2833bc. - On the merged head
022872c: the test record passed 1,180 tests, and the selected browser run passed 11,178. Lesson: L-307 — the second instance (BUG-441 was the first): a test read, released or aborted a held request after only a UI change. L-307’s sources gain BUG-515, andfrontend-reviewerdimension 14 now checks every held-request release.
BUG-514 — Flaky test: MailboxContentLeakIT’s ingest-and-refusals log check passed only on its retry on a lightly loaded machine (Windows)
Affects: F20-b (forge-et-be src/test/groovy/com/empireteas/mailbox/MailboxContentLeakIT.groovy:118, “ingest, retries, the stored detail and the refusals put no mail content in logs, problems or metrics”; test code only, as far as known) · Found in: the weekly flaky-test report of 2026-11-20: forge-et-be#483’s build record on ed9b4f4 (F36-d wave 2), Windows 11, 2026-11-19 · Severity: low — it passed on its retry · Phase: 1
Repro: the build record: 19,037 tests, 0 failed, this one flaky; it took 38.8 s, the third-slowest test of the run. Machine load 0.26 at start, 0.98 peak, 0.20 average, 16 cores. The record does not carry the failure message; it is in that machine’s build/test-results/test/TEST-com.empireteas.mailbox.MailboxContentLeakIT.xml (flakyFailure), if still there. The case waits with PollingConditions(timeout: 30, delay: 0.2) (line 132). The run was after BUG-466’s fix (forge-et-be#471, merged 2026-11-18), so it is not the plain ListAppender race BUG-466 removed.
Expected: the case passes whatever else the machine is running.
Actual: it failed once and passed on its retry, at a low average load.
Fix: not yet. Get the failure message first: from the machine that ran #483’s record if it still has the XML, or by running the spec repeatedly on Windows (./gradlew test --tests 'com.empireteas.mailbox.MailboxContentLeakIT' -Pet.testRetries=0, several times, under load and without). A 30-second polling bound reached at 38.8 s is the first suspect.
Lesson: none yet.
BUG-513 — Flaky test: knownFailureMarks.test.ts’s e2e-tree check reads every e2e file in one case and runs past the 20-second timeout under load (Windows)
Affects: F40 (forge-et-fe src/app/knownFailureMarks.test.ts, “the e2e tree has no stray test-fail call and no mark without a bug id”; test code only) · Found in: the weekly flaky-test report of 2026-11-20: the test records of forge-et-fe#432 and #434 (2026-11-06), #531 (2026-11-13) and #583 (2026-11-18), Windows 11 · Severity: low — a test-only failure; it passed on its retry each time · Phase: 1
Repro: in two of the records the case took 21.5 s (#531, load 0.90 average) and 21.8 s (#583, load 0.89 average), past the 20-second test timeout; the other two ran at 0.58 and 0.65. The case walks e2e/ with readdirSync and readFileSync (knownFailureMarks.test.ts:140–153) and runs regular expressions over every file’s text, all inside the one case.
Expected: the case passes whatever else the machine is running.
Actual: on a loaded Windows machine it runs past its budget.
Fix: not yet. The case does more work than its budget allows (.forge/test-strategy.md § Flaky Tests cause 6). Read the tree once, outside the case (in beforeAll, which can take its own time as the git-fixture suites do, L-257), and leave only the checks inside the case; or narrow the walk to the files that can hold a mark. Measure it on Windows under load before and after.
Resolved: 2026-11-21 — fixed in forge-et-fe#600. Measuring showed the read is cheap (about 20 ms for the 428 files) and the cost is the comment strip: a character-by-character loop over about 12 MB, run twice per file, once for each check. Each file is now stripped once, by one regular-expression pass that gives the same output as the old loop on all 1,262 inputs tried (every .ts/.tsx file under e2e/ and src/, and edge cases). On macOS with one busy loop per core (8), the case took 550–622 ms before and 73–85 ms after, about 7× less. Measured on macOS only: the cut does not depend on the operating system, and the next Windows test records show whether the case stays inside its budget; a recurrence there is a new bug that links this one. No case fails before the fix, since the old code passes on an idle machine; the measurement is the evidence.
Lesson: L-125 — one more case that did more work than the 20-second default allows under load; the work is cut, not the timeout raised.
BUG-512 — Lead ranking retries a lead whose every ranking fails, once a day with no end
Affects: F11-g (forge-et-be lead/service/LeadRanking*, the hourly lead-ranking-sweep, on main) · Found in: the week’s count that .forge/owed.md’s row “Watch lead ranking’s no-tool replies on dev” asked for, read from dev’s /ecs/dev/backend logs for 12–19 Nov 2026 (read-only), 2026-11-20 · Severity: low — one paid call a day (about US$0.006) per stuck lead, and the lead shows no priority · Phase: 1
Repro: on dev, the logs’ Lead Ranking Completed lines for lead 83. Its ranking failed every day: TIMEOUT on 12 Nov, then INVALID_OUTPUT on 13, 14, 15, 16, 17 and 18 Nov. Each AI Call Completed line shows inputTokens=3972 (other leads send about 1,700) and two attempts, and each attempt logs Bedrock Attempt Failed … NO_TOOL_USE: Haiku 4.5 answers in text instead of calling the ranking tool, which toolChoice auto allows (BUG-248).
Expected: a lead whose ranking keeps failing on the same inputs is not called again every day without limit, or the call succeeds.
Actual: the sweep re-queues lead 83 each day and the same call fails the same way. In the same week, the other ranked leads succeeded (97, 99–103), except lead 96, which failed once and succeeded on the next try: about 1 miss in 9 calls, near the 1 in 10 seen on 9 Nov. So the general rate is unchanged; the problem is one lead that never succeeds.
Fix: not decided. Two parts to look at: (1) why lead 83’s input, about twice the usual size, makes Haiku 4.5 answer in text — its research findings or inquiry text, read on dev (a new prompt version, F11-g FR-A2, if the instruction to answer only through the tool needs to be firmer); (2) whether the sweep should stop re-ranking a lead after a set number of INVALID_OUTPUT results on unchanged inputs (its fingerprint, F11-g FR-J6) until the inputs change. Either is an F11-g spec revision. Matters before the Pipedrive first load ranks about 700 leads.
Lesson: none yet.
BUG-511 — Flaky test: LeadsRoute.test.tsx’s “keeps each section search and the chip across a section switch” reads the address before the To Review search is back in it (Windows)
Affects: F36-c (forge-et-pwa src/features/leads/LeadsRoute.test.tsx, Completed, on main) · Found in: a run of 91 Vitest files (src/ui src/features/settings src/lib/install src/features/leads src/app, 45 s) on forge-et-pwa fix/BUG-495-field-error-icon-and-ios-install-line, Windows, 2026-11-19 · Severity: low, a test-only failure · Phase: 1
Repro: run pnpm exec vitest run src/ui src/features/settings src/lib/install src/features/leads src/app on a Windows machine.
Expected: the test passes.
Actual: expected 'section=completed&completedSearch=ma&…' to be 'section=completed&toReviewSearch=leaf&completedSearch=ma&outcome=rejected' — the address was read before the To Review search was written back to it. It passed alone and twice with its whole file on the same tree. The branch does not change the Leads page. No machine load was recorded for the run (it was not a test record).
Cause (2026-11-20): a defect in the page, not the test, and BUG-475’s cause. LeadsRoute’s update() calls React Router’s functional setSearchParams, which starts from the parameters of the last render, and the router applies each write in a transition. The test’s waitFor saw toReviewSearch=leaf on the router before the page had rendered it, and the Completed tab pressed then was handled by the earlier render, whose write left the search out. Reproduced in Vitest by pressing Completed the moment the router takes the search’s write (a router.subscribe callback and a microtask): it fails on main 04dc9c7 with the reported message.
Fix: BUG-475’s fix, forge-et-pwa#180 (a draft): LeadsRoute writes the URL with useLatestSearchParams, moved to src/lib/url/latestSearch.ts, so each write starts from the latest written parameters. The regression test above passes on #180’s head f9c28f2; it is added to #180 with the fix, and this bug closes with #180. Released by Nanndish on 2026-11-20 for BUG-475’s owner to take with it.
Lesson: none yet.
BUG-510 — The capture app’s field errors show an alert icon that the CRM’s do not
Affects: F36-a (forge-et-pwa src/ui/Field.tsx, the error line under every field, on main) · Found in: Nanndish’s review of the capture app, 2026-11-19 · Severity: low · Phase: 1
Repro: in the capture app, leave a required field empty and submit (for example Edit Details on a captured lead with the company name cleared).
Expected: the message under the field is tone/danger text at type/meta with no icon, as in the CRM (design-system §5 Form fields and validation, Errors).
Actual: Field draws AlertIcon before the message ({hasError ? <AlertIcon … /> : null}), a difference from the CRM’s Field, of which it is a port.
Fix: remove the icon and its .icon rule from Field; the message text, and aria-invalid for assistive technology, still mark the error, so colour is not its only signal. Test: Field.test.tsx, an error renders no svg. pwa-current draws a field error without an icon.
Resolved: 2026-11-19 — fixed in forge-et-pwa#191 (1584925), whose branch name calls it BUG-495, the number this bug held until F36-d’s BUG-495 reached main first (renumbered 2026-11-20, id-numbering rule 5). The new Field.test.tsx row finds an svg in the message on the old code, so it cannot pass there. The field-error size-matrix states (lead-edit-details-errors, -server-errors, plan-campaign-create-invalid, plan-add-exhibitor-invalid) pass at every size and engine; pwa-current screenshot 24 is retaken.
Lesson: none — one-off: a port that kept a detail the CRM’s component never had; design-system §5 now says no icon in both apps.
BUG-509 — The Add Tag list is clipped at the window’s bottom edge instead of opening above when it does not fit below
Affects: F1-e (forge-et-fe src/ui/tags/AddTagPicker.module.css, on main; the Manager and sharing pickers of F5-c and F5-g reuse the picker and keep its earlier cap) · Found in: the selected browser run of BUG-505’s fix, 2026-11-20 · Severity: low · Phase: 1
Repro: at 320×640, open a lead whose Add Tag control sits about 250px above the window’s bottom edge (since BUG-505, where the size matrix’s pinned lead puts it), and open Add Tag with more tags than the list shows; or run the size state leads-lead-tag-picker-many at size-320x640-chromium.
Expected: the list opens on the side that holds it whole, with the line “Type to find N more tags” under it in the window (design-system §5 Add Tag; the state’s expectShortTagList check).
Actual: the list opened below with 248px of room for a 250px popover, so the popover scrolled by a few pixels and its last line ended 7px past the window, while 352px were free above. The popover’s max-block-size was min(var(--radix-popover-content-available-height), 12 rows). Radix decides whether to flip from the popover’s measured size, so a popover already capped to the room below always fits there and never flips.
Fix: AddTagPicker takes windowBounded, which only AddTagControl sets: there the cap is the fixed 12 rows only, because the row count is already chosen from the window’s height (useAddTagOptionsMax), so the popover flips above when it does not fit below. Every other caller (SearchSelect, RecordSearchPicker, AddColleagues) keeps the cap to the room on its side, because its list can be longer than either side holds: a first fix that dropped that cap for every caller left the match modal’s 25 account results 12px past the window at 320×640 (lead-detail-layout.spec.ts AC-22). Tests: AC-22 passes at every size (13 of 13; 13 of 13 on main’s cap, and failing at 320×640 with the cap dropped for all); every tag size state at all four sizes passes (353; before the fix the one 320×640 state failed); src/ui/tags, src/ui/fields, src/features/sharing, src/features/notesAndTags and LeadTagsSection.test.tsx pass (466). FollowUpDialog.module.css and customFieldRenderers.module.css cap their popovers the same way and are left for a later check (owed.md).
Resolved: 2026-11-20 — fixed in forge-et-fe#597 (acba35f7, narrowed to Add Tag in ba45745c), whose commits call it BUG-499, the number it held until forge-et-harness#1576, opened first, took it (id-numbering rule 5). Fails before: the size state leads-lead-tag-picker-many at size-320x640-chromium.
Lesson: L-334 — a popover capped to Radix’s available height never flips; cap a list whose rows already fit the window at a fixed size.
BUG-506 — The Lead Research settings switch sits at the far left against its label, above a long paragraph
Affects: F11-i (forge-et-fe src/features/leadResearchSettings/LeadResearchSettingsRoute.tsx, SettingCard, on main) · Found in: Nanndish’s second UI review list of the CRM, 2026-11-19 · Severity: low · Phase: 1
Repro: open Settings › Lead Research at 1440 × 900.
Expected: the page reads as a settings row: what the setting is, then whether it is on, with the control where a reader looks for it.
Actual: the switch is the card’s first item, at its far left and pressed against the start of its label, with two lines of explanation under the label; nothing in words says whether it is on, and the switch reads as a bullet rather than the control.
Fix: design-system Rev 147 (§8 A Settings page holding one setting): the card is a settings row. The setting’s name, in the card-title weight, leads the first column, with the two explanation lines and Last changed by under it; the second column holds On or Off (text/secondary, aria-hidden, because the switch already says its state) and then the switch, centred on the name’s line. The paragraphs stay in the first column, so they never run under the switch. The switch keeps its label, description, test id and every save outcome. Tests: LeadResearchSettingsRoute.test.tsx (the name before the switch, the state word beside it and hidden from assistive technology, Off after a flip, the switch’s name unchanged); the row checked by hand in the browser at 1440px and 320px (no browser test asserts its layout); crm-current redrawn in the same delivery.
Resolved: 2026-11-20 — fixed in forge-et-fe#597 (a3922958).
Lesson: none — one-off: the page was built before design-system §8 had a pattern for a page holding one setting; Rev 147 adds it.
BUG-505 — The lead page puts Attachments last in its side column, where the account page has it in the main column
Affects: F11-f (forge-et-fe src/features/leads/LeadView.tsx, the Overview tab’s RecordColumns, on main), F33-a (the attachments place) · Found in: Nanndish’s second UI review list of the CRM, 2026-11-19 · Severity: low · Phase: 1
Repro: open a lead and an account and compare their Overview columns.
Expected: the two record pages place the same cards the same way, so Shared With ends the side column on both.
Actual: the account page holds Attachments at the end of its main column, after Profile, and ends its side column with Tags then Shared With; the lead page ends its side column with Tags, Shared With, then Attachments (F11-f Rev 10 FR-V1). Both already put Tags before Shared With.
Fix: one rule for both pages, decided with Nanndish on 2026-11-19: the record’s own content goes in the main column and the short facts about it in the side column, with Tags then Shared With last, because sharing controls who sees the record and changes least often. The lead’s attachments place moves to the main column, after customFields (the end of the lead’s own details) and before the participants place, which draws nothing until filled; the account page is unchanged. F11-f gets a spec revision for FR-V1 and design-system Rev 147 records the rule (§6 Record page). Tests: LeadView.test.tsx (both columns’ order) and e2e/browser/leads/lead-detail-layout.spec.ts (the attachments place in the wide column’s gap check, the wide column ending at it); crm-current redrawn.
Resolved: 2026-11-20 — fixed in forge-et-fe#597 (e63fa4ec; the composed lead view’s test in c7127816).
Lesson: none — one-off: a layout choice that differed between two pages, now one rule in design-system §6.
BUG-504 — The account list’s filter row wraps, and Clear Filters pushes Columns onto another line
Affects: F1-b (forge-et-fe src/features/accounts/AccountsFilters.tsx, the data bar, on main) · Found in: Nanndish’s second UI review list of the CRM, 2026-11-19 · Severity: low · Phase: 1
Repro: open Accounts at 1440 × 900, then choose any filter.
Expected: the filters keep to one row, with Columns at its end, as on the Leads list.
Actual: the row holds Search, five multi-choice filters and two switches; at 1440px it already wraps Shared With Me to a second line before any filter is set, and choosing a filter adds Clear Filters, which pushes Columns (and Save View beside it) further down. The Leads list keeps its row short with a More Filters popover (F11-b, MoreFilters).
Fix: the account list uses the shared MoreFilters, as the Leads list does: the row is Search, Customer Type, Country, Owner, Pipedrive Owner (for a hand-out role), More Filters, Clear Filters, Columns; the popover holds Region, Tag, Shared With Me and Show Archived, one per row, each with its URL state and test id unchanged, and the trigger says how many of them are applied (More Filters (2)). The Tag filter’s Try Again is inside the portalled popover, so focus after a successful retry goes to the Tag trigger, as on Leads. More Filters alone was not enough: measured on 2026-11-19 at 1440px the bar is 1144px wide and the row without Pipedrive Owner fits with 10px to spare, so Clear Filters still pushed Columns down, and a hand-out role’s row never fit. So the bar is two groups: the filters wrap inside their own box (basis 384px), and the Columns / Save View cell keeps to the end of the first line; below about 600px of content that cell wraps under the filters. Measured after the fix at 1440px with a filter set, as an Administrator: Columns and Save View at the search box’s top (93px), More Filters (1) and Clear Filters on the filters’ second line; at 768px the cell wraps below, with no page overflow. Design-system Rev 147 records it (§6 List screen). Tests: AccountsFilters.test.tsx, AccountListScreen.test.tsx and AccountListScreen.views.test.tsx reach the moved controls through More Filters, with new rows for the count, the row’s controls and the two groups; accounts-layout.spec.ts checks Columns level with the search box while Clear Filters shows; the browser specs that use the moved filters open the popover first; crm-current redrawn.
Resolved: 2026-11-20 — fixed in forge-et-fe#597 (e2fb390f; browser specs in cf5ad0f7).
Lesson: none — one-off: the account list’s filters grew past one row as features added them; the Leads list already used More Filters, and the account list now does too.
BUG-503 — The lead page’s Edit and Archive buttons are not the size of Add Follow-Up and Approve & Send beside them
Affects: F11-f (forge-et-fe src/features/leads/LeadFormActions.tsx EditLeadAction, src/features/leads/useLeadStatusAction.tsx, on main) · Found in: Nanndish’s second UI review list of the CRM, 2026-11-19 · Severity: low · Phase: 1
Repro: open a lead with lead.manage and measure the header’s buttons.
Expected: the header’s buttons share one height and text size.
Actual: measured in the browser on 2026-11-19: Edit is the shadcn outline button at size="sm" (32px tall, 14px text); Archive is still the older src/ui Button at size="sm" (36px tall, 12px text); Add Follow-Up and Approve & Send are shadcn buttons at the default size (36px, 14px).
Fix: Edit takes the default size, and Archive / Restore moves to the shadcn Button (outline, default size, the registry’s archive / restore icon as its first child), so all four are 36px with 14px text. useLeadStatusAction.tsx keeps its shadcnBackfill.json row for its two refusal-panel buttons, which still use the older Button. Tests: LeadStatusAction.test.tsx (Edit and Archive are data-slot="button", outline, default size); the four buttons measured by hand in the browser (no browser test asserts their size).
Resolved: 2026-11-20 — fixed in forge-et-fe#597 (41f80f61).
Lesson: none — one-off: a header part-way through the AD #39 move to shadcn; the buttons still on the older component are tracked in shadcn-backfill.md.
BUG-508 — A busy shadcn button with a leading icon draws the icon beside the spinner, a recurrence of BUG-274
Affects: F20-d (forge-et-fe src/ui/shadcn/button.tsx, the busy spinner of design-system Rev 83, on main), F32-d (the saved views’ ⋯) · Found in: Nanndish’s second UI review list of the CRM, 2026-11-19 · Severity: low · Phase: 1
Repro: on the Email page press Refresh; or on the account, lead or contact list open a saved view, open its ⋯ menu and press Share.
Expected: design-system §5 Busy: the spinner takes the leading icon’s place, so the button shows one glyph and keeps its width.
Actual: the spinner and the icon both show. ButtonSpinner marks itself data-leading-icon and the button hides the icon after it with the Tailwind class [&>[data-leading-icon]+svg]:hidden, but Tailwind’s utilities sit in a cascade layer (src/ui/styles/shadcn.css) and the registry Icon’s CSS Module rule .icon { display: inline-block } is unlayered, and an unlayered rule beats any layered one. Measured in the browser on 2026-11-19: with the spinner placed before the lead page’s Edit icon, the icon still computed display: block. In the ⋯ (size="icon-sm", 32px square) the two glyphs squeeze into the square, which reads as the loader being “off”. jsdom applies no layers, so button.test.tsx, which checks only the class name, passed throughout. BUG-274’s fix had hidden the icon another way; the Rev 83 rework (06793bc2) replaced it with this rule.
Fix: the rule becomes !important ([&>[data-leading-icon]+svg]:hidden!): an important declaration in a layer beats a normal unlayered one. Tests: button.test.tsx asserts the important class; a new browser test in e2e/browser/accounts/account-views.spec.ts holds the Share PATCH and checks that the busy ⋯ draws the spinner alone (computed display).
Resolved: 2026-11-20 — fixed in forge-et-fe#597 (dcbbb110), whose commits call it BUG-502, the number it held until forge-et-harness#1576, opened first, took it (id-numbering rule 5).
Lesson: L-333 — a CSS Module rule is unlayered and beats a layered Tailwind utility; make the utility important (hidden!). frontend-reviewer dimension 35 checks for it.
BUG-507 — The Email page’s collapsed folder rail draws a two-digit unread count over its icon
Affects: F20-i (forge-et-fe src/features/email/FolderColumn.module.css .tileUnread, FolderColumn.tsx Unread, on main) · Found in: Nanndish’s second UI review list of the CRM, 2026-11-19 · Severity: low · Phase: 1
Repro: on the Email page with 20 or more unread messages in the Inbox, collapse the folder pane to its rail.
Expected: the count sits at the icon’s top-right corner, as a badge, and the icon stays readable.
Actual: the rail is 76px wide (--et-layout-sidebar-width-collapsed). The badge’s end was pinned 22px past the tile’s middle (inset-inline-end: calc(50% - space/2 - space/6)) and it grew toward the start: a one-digit badge (18px) clears the 18px icon, but “20” (about 26px) covers the icon’s right half and “120” all of it. Every rail tile uses the rule: the Mailbox folders, the Folders tile and the Scheduled tile’s waiting count. The main sidebar hides its badges in icon mode, so it is not affected.
Fix: the badge’s start is pinned space/2 past the tile’s middle, so it covers only the icon’s top-right corner and grows into the rail’s free space; its inline padding on a tile is space/2; and a tile draws a count above 99 as 99+ (RAIL_UNREAD_CAP), named as drawn (WCAG 2.5.3), while the open pane keeps the exact count. The widest badge, 99+, ends inside the 76px rail. Design-system Rev 147 records it. Tests: FolderColumn.test.tsx (Inbox 99+ unread on the rail, 99 drawn as it is, Inbox 120 unread in the open pane); a browser measurement of the rail; crm-current redrawn.
Resolved: 2026-11-20 — fixed in forge-et-fe#597 (5f568e3e; review fixes cbc636dd), whose branch, title and commits call it BUG-501, the number it held until forge-et-harness#1576, opened first, took it (id-numbering rule 5).
Lesson: none — one-off: the badge’s place was set for a one-digit count; the main sidebar hides its badges in icon mode, so no other rail shares the rule.
BUG-502 — Two capture-app seam specs still look for a visible In CRM icon on a converted row, which F36-c hides
Affects: F36-c (forge-et-pwa e2e/seam/capture-journey.seam.spec.ts line 543, e2e/seam/review-screens.seam.spec.ts line 449) · Found in: the nightly full run of 2026-11-20 on main 9ddaddd (forge-et-pwa#189) · Severity: medium — a test defect; the product follows its design, but the nightly capture-app gate on main is red · Phase: 1
Repro: with a backend on ET_BACKEND_PORT, in forge-et-pwa on main 9ddaddd: E2E_GATE=1 pnpm exec playwright test e2e/seam/capture-journey.seam.spec.ts e2e/seam/review-screens.seam.spec.ts --project=seam -g 'AC-40, AC-39, AC-35|AC-22 \(smoke\)' --retries=0.
Expected: the converted row under Completed › In CRM is checked the way F36-c draws it.
Actual: both fail at done.getByRole('img', { name: 'In CRM', exact: true }) — element not found, in both attempts. F36-c wave 3 (forge-et-pwa#187, 6560247, 3c03289) hides the text a decided row’s link is named from, its outcome icon included, so a screen reader hears it once, through the link’s name. lead-page.spec.ts was updated to includeHidden: true and a check of the link’s name; the seam specs, which need a backend and were not in the wave’s runs, were not.
Fix: in both specs, find the icon with includeHidden: true and check that the row’s lead-row-link is named with In CRM, as e2e/browser/leads/lead-page.spec.ts does.
Regression test: the two seam tests themselves, which fail on main and pass with the fix.
Resolved: 2026-11-20 — fixed in forge-et-pwa#190. Against the night’s backend (forge-et-be b33e11b8, the night’s ports), the two tests failed alone at the night’s commit 9ddaddd with --retries=0, at the same line as the night (2 failed, the setup project passed), and passed on the fix e2a8039 (3 passed, 15.4 s).
Lesson: L-329 — a UI change that rebuilds an element updates every test that selects it, the seam specs included; this is the case the lesson describes, in the two seam specs F36-c wave 3 did not run.
BUG-501 — The saved-views row check expects a Default tab for a user with no view to choose, where F32-d Rev 6 draws no row
Affects: F32-d (forge-et-fe e2e/browser/accounts/views-support.ts expectRowMatchesRead, used by views-across-users.spec.ts) · Found in: the nightly full run of 2026-11-20 on main c77a7f67 (forge-et-fe#593) · Severity: medium — a test defect; the product follows its design, but the nightly CRM gate on main is red · Phase: 1
Repro: against a backend whose database holds no shared account view, in forge-et-fe on main c77a7f67: E2E_GATE=1 pnpm exec playwright test e2e/browser/accounts/views-across-users.spec.ts --project=chromium -g 'AC-8a, AC-9' --retries=0.
Expected: for OTHER, who has no view of their own and none shared with them, the check finds no views row, as F32-d Rev 6 FR-V7 says (“A caller with neither sees no views row”).
Actual: expectRowMatchesRead always expects the Default tab first, so the test fails at “the default tab is first” — element not found, in both attempts. On a database where an earlier run left a shared account view, OTHER’s row is drawn and the test passes, which is why forge-et-fe#586’s run passed it.
Fix: when the last complete row-1 read holds no view, expectRowMatchesRead checks that no tab list is drawn and returns the empty read; expectRowSettled already allows for no row.
Regression test: the AC-8a, AC-9 test itself, which fails on a fresh database on main and passes with the fix.
Resolved: 2026-11-20 — fixed in forge-et-fe#596. On a fresh database (the triage stack started that morning) the test failed alone at the night’s commit c77a7f67 with --retries=0, at “the default tab is first”, and passed on the fix 9329ef32.
Lesson: L-332 — a test helper that encodes a rule a spec revision changes is run against a fresh database, because data an earlier run left can keep the old rule’s case alive.
BUG-500 — The leads-focus-marker-tooltip page state reaches the marker with Tab, which skips buttons in WebKit
Affects: F36-c (forge-et-fe e2e/browser/size-matrix/arrange.ts, the leads-focus-marker-tooltip state) · Found in: the nightly full run of 2026-11-20 on main c77a7f67 (forge-et-fe#593) · Severity: medium — a test defect; the product follows its design, but the nightly CRM gate on main is red · Phase: 1
Repro: on a Mac with keyboard navigation off (macOS’s default), in forge-et-fe on main c77a7f67: E2E_GATE=1 pnpm exec playwright test e2e/browser/size-matrix/size-matrix.spec.ts --project=size-768x1024-webkit -g 'leads-focus-marker-tooltip$' --retries=0 (also size-1024x768-webkit).
Expected: the arrangement focuses the Focus marker with a key press, so its tooltip opens, in every engine.
Actual: expect(marker).toBeFocused() fails, in both attempts: the arrangement presses Shift+Tab then Tab, and WebKit’s Tab skips buttons on such a machine, so focus never comes back to the marker.
Fix: in WebKit, press the key webkitFocusKey(page) measures (e2e/support/focusKey.ts: Tab, or Alt+Tab where Tab skips buttons), measured before the marker is focused.
Regression test: the page state itself in both WebKit size projects, which fails on main and passes with the fix.
Resolved: 2026-11-20 — fixed in forge-et-fe#596. On this Mac (keyboard navigation off) the state failed alone at the night’s commit c77a7f67 with --retries=0 at both WebKit sizes, and passed on the fix 9329ef32 at both, and at 1280×800 Chromium.
Lesson: L-168 — a keyboard step in a browser test uses the focus key the machine’s WebKit takes (webkitFocusKey), not a bare Tab; this state is another instance.
BUG-499 — The Hot / Cold Tab-after-choose test expects Sort next, but F36-c draws the Campaign Focus filter there
Affects: F36-c (forge-et-fe e2e/browser/leads/selectTabAfterChoose.ts) · Found in: the nightly full run of 2026-11-20 on main c77a7f67 (forge-et-fe#593) · Severity: medium — a test defect; the product follows its design, but the nightly CRM gate on main is red · Phase: 1
Repro: in forge-et-fe on main c77a7f67: E2E_GATE=1 pnpm exec playwright test e2e/browser/leads/select-tab-after-choose.spec.ts --project=chromium --retries=0.
Expected: Tab from the Hot / Cold trigger reaches the next control in the data bar.
Actual: “in Chromium, the next control is Sort” — expected leads-sort, received leads-filter-ranking-focus, in both attempts. F36-c wave 3 (forge-et-fe#589) draws the Campaign Focus filter right after Hot / Cold (LeadRankingFilters.tsx), and the test’s constant still named Sort.
Fix: expect the Campaign Focus filter (leads-filter-ranking-focus) after Hot / Cold, and name it in the helper’s comment on what WebKit’s Tab skips.
Regression test: the test itself, which fails on main and passes with the fix.
Resolved: 2026-11-20 — fixed in forge-et-fe#596. The test failed alone at the night’s commit c77a7f67 with --retries=0 (received leads-filter-ranking-focus) and passed on the fix 9329ef32, with its WebKit twin select-tab-after-choose-webkit.spec.ts.
Lesson: L-329 — a change that puts a new control into the data bar’s tab order updates every test that asserts what comes next (the lesson now says so, with this bug and BUG-498).
BUG-498 — The Columns menu keyboard tests still expect Tab from Reset to reach the table, but Save View now follows Columns
Affects: F32-d (forge-et-fe e2e/browser/accounts/columns-support.ts, the afterColumns value of the account, lead and contact lists) · Found in: forge-et-fe#591’s selected browser run on f2146788 (2026-11-19), then the nightly full run of 2026-11-20 on main c77a7f67 · Severity: medium — a test defect; the product follows its design, but the nightly CRM gate on main is red · Phase: 1
Repro: run pnpm exec playwright test -g 'AC-20b \(keyb' e2e/browser/accounts/account-columns.spec.ts (or columns-keyboard-webkit, contacts-columns, e2e/browser/leads/lead-columns.spec.ts) in forge-et-fe on main after forge-et-fe#586.
Expected: Tab from Reset closes the popover and moves focus to the next control after the Columns trigger, and the test checks that control.
Actual: all four fail at the same step, in both attempts, after 7–14 seconds: “Tab from the last control moves focus to the next control after the trigger, account-list-table-scroll” (leads-table-scroll, contacts-table-scroll) — expected focused, received inactive. forge-et-fe#586 (F32-d saved views, spec Rev 6) draws Save View right after Columns through ColumnsMenu’s new trailing slot, so it is now the next control in the Tab order, while columns-support.ts still names the table’s scroll region as afterColumns. #586 ran only the saved-views specs, not the Columns specs, and its selected browser run was skipped; the last nightly CRM gate before it (cc97f8cd, 2026-11-19) passed these tests.
Fix: set each list’s afterColumns in columns-support.ts to its Save View test id (saveViewTestId(<prefix>), src/ui/filters/ViewActions.tsx), after confirming in a browser that focus lands there, and check the comment at line 1008 and the WebKit walk’s Option+Tab case against the same order.
Regression test: the four AC-20b tests themselves, which fail on main now and pass with the fix.
Resolved: 2026-11-20 — fixed in forge-et-fe#596. The four tests failed alone at the night’s commit c77a7f67 with --retries=0, at the same step as the night, and passed on the fix 9329ef32. Focus lands on Save View because the popover’s Tab out moves to the next tabbable after the trigger in document order (src/ui/filters/popoverTabOut.ts), in Chromium and with WebKit’s Tab and Option+Tab alike.
Lesson: L-329 — a change that puts a new control into the data bar’s tab order updates every test that asserts what comes next (the lesson now says so, with this bug and BUG-499).
BUG-497 — The live inbox refuses the second page of a category, Focused or Other view: Graph writes the next link’s $filter spaces as ‘+’
Affects: F20-i (forge-et-be GraphMailboxProvider.PageCursor, on main 1b2e53e) · Found in: F20-i-WI-5.1’s laptop walk against CRM01’s real mailbox, 2026-11-19 (record R-9, .forge/plans/F20-i/F20-i-WI-5.1-record.md) · Severity: high: the Email page opens on Focused, so a mailbox with more than one page of Focused mail cannot show the rest · Phase: 1
Repro: in laptop mode with a real mailbox connected, read GET /api/v1/mailbox/folders/inbox/messages?category=<a category>&size=1 (or classification=FOCUSED&size=1) where more than one message matches, then read the same request with cursor=<the answer's nextCursor>.
Expected: 200 with the next page (F20-i FR-7, FR-8; WI-3.1 E-20).
Actual: 400 VALIDATION_FAILED on cursor; the backend log shows ApiExceptionHandler Request Rejected … status=400, code=VALIDATION_FAILED and no Graph call. Graph’s @odata.nextLink carries the filter form-encoded — %24filter=receivedDateTime+ge+1900-01-01T00%3a00%3a00Z+and+categories%2fany(c%3ac+eq+%27F20I+Buyer%27), spaces as + — and PageCursor.decode(String) calls URLDecoder.decode(component.replace("+", "%2B"), UTF_8), so each + stays a +, the decoded filter no longer equals GraphFilterQuery.of(...), and bindTo refuses the cursor. Unfiltered paging ($skip) and a one-word search ($skiptoken) page normally; a search of more than one word is likely refused the same way, since bindTo compares $search with the same decoder. ConversationCursor form-decodes and pages correctly on real Graph (R-16).
Fix: decode PageCursor’s query values as form-encoded, as ConversationCursor.formDecode does (URLDecoder.decode(component, UTF_8): + is a space, %2B a plus), with a regression test that follows a next link whose $filter and $search carry + for spaces. forge-et-be#486 does this.
Resolved: 2026-11-19 — fixed in forge-et-be#486 (merge f318e3de), whose title and commits call it BUG-495 (its test names and Javadoc were changed to BUG-497 by forge-et-be#491), the number this bug held until F36-d’s BUG-495 reached main first (renumbered 2026-11-20, id-numbering rule 5). Three of its four new cases failed on the commit before the fix (both filter cases and the search with a space as +; 64 tests, 3 failed) and pass on it; the test record ran the whole suite on 87732fc (18,855 tests, 0 failed). F20-i-WI-5.1 R-9 is re-run against it as Attempt 2.
Lesson: L-325 — decode a URL an outside service returns the way that service encodes it, and test with one it really returned.
BUG-495 — Flaky test: the lead-check page states time out at the 5-second ready wait in WebKit (Windows)
Affects: F36-d (forge-et-pwa e2e/pageStates.ts, the lead-check-* page states added by F36-d-WI-2.3, run by e2e/browser/size-matrix/size-matrix.spec.ts and capture-layout.spec.ts in the WebKit size projects; on feature/F36-d-wave-2) · Found in: the review round of F36-d wave 2’s wave PR (forge-et-pwa#185), Windows 11, 2026-11-19 · Severity: low — test only; every state passes when the machine is not shared · Phase: 1
Repro: on Windows, from forge-et-pwa feature/F36-d-wave-2, run the lead-check-* page states in the 390 × 844 WebKit size project (size-390x844-webkit) with --retries=0, several repeats and two workers. At 08ed7ee, before wave 2’s review fixes, 18 of 27 failed (three repeats, two workers), lead-check-failed-earlier among them, which was already on the branch before the review. Failures were seen at 390 px WebKit and sometimes at 320 px WebKit. The run’s recorded load is not in this record. With nothing else holding the machine’s heavy-run lock, all 15 lead-check-* states passed 15 of 15 at 390 px WebKit.
Expected: each state reaches its ready test id within the 5-second ready wait, whatever else the machine runs.
Actual: the state times out at the 5-second ready wait. The failure snapshots show the lead’s drawer not yet open, or the check still running.
Fix: not yet. Find whether WebKit on Windows is only slow under load here (the drawer opens and the check answers after 5 s) or whether the arrangement races. Compared with BUG-468 and BUG-491, it is a separate defect: BUG-468’s plan-list upload states fail in WebKit on Windows even when run alone (3 of 3 alone on main), from the upload arrangement and StrictMode’s double mount; BUG-491 is a browser test (lead-customer-check.spec.ts) that releases a held check before the request arrives. These page states pass when run alone, and their snapshots show the drawer or the check not yet ready, not a released request.
Triage (2026-11-22, Nanndish): the order of .forge/test-strategy.md § Flaky Tests, on capture-app main 1b446e0, Windows 11, 16 cores, idle before the runs (CPU 6%). Every run held the heavy-run lock and used --retries=0.
- Idle: the 15
lead-check-*states, three repeats, two workers, at 390x844 WebKit: 45 of 45 passed. The ready step (the arrangement, the page load and the ready wait) took a median of 1.6 s and at most 2.3 s. - Forced load (step 2) did not reproduce a defect in these states. One busy loop per core started once the run held the lock, keeping the CPU fully busy, well past the gate’s 0.65 limit:
- every
lead-*state at 390x844 WebKit, once, two workers: the 15lead-check-*states passed 15 of 15, while 8 of the 60 other lead-drawer states missed the 5-second ready wait (lead-convert-several,lead-edit-details-emptyand others). The check states’ ready step (median 6.5 s) was no slower than the others’ (6.9 s). - the 15
lead-check-*states at 390x844 and 320x640 WebKit, three repeats, two workers (the reported shape): 88 of 90 passed.lead-check-failedandlead-check-failed-earliereach missed the wait once, their snapshots showing the check still running.
- every
- No step can land out of order. The check is sent once the lead’s own read has landed (
useLeadCustomerCheck,enabled: !pageRead.reading), withretry: false, so a 503 state waits no longer than an answered one. Theafterarrangements (recheck-*,retry-held,go-offline) start from aMutationObserveronce the first answer is on screen (e2e/support/arrange.ts). Each state’s ready test id appears only at the end of that fixed chain: the list, the drawer, the lead’s read, then the check. A late step makes the state slower; it cannot leave the state stuck. - The reported run was overloaded by other work. 18 of 27 failing needs far more than one busy loop per core, which failed 2 of 90. The same states passed 15 of 15 once nothing else held the machine’s heavy-run lock. The run’s load figure was not recorded.
Resolved: 2026-11-22 — closed
wontfixas an overloaded machine (.forge/test-strategy.md § Flaky Tests, triage step 4); decided by Tharinda (lead). - Why it is not fixed: the lead-check states have no race. Under forced load they fail less often than the other lead-drawer states, at the same 5-second ready wait, and a longer wait is the fix the test strategy rules out.
- The tests are kept unchanged. A recurrence on a machine within the load limit is a new bug that links this one, as BUG-517 says for macOS. Lesson: none — the overload outcome of triage step 4, as for BUG-517 on macOS and BUG-451.
BUG-494 — Flaky test: lead-page.spec.ts’s “a 409 shows the lead as it is now, and so does a 422 a teammate decided” misses the write message once in a wider run (Windows)
Affects: F36-a (forge-et-pwa e2e/browser/leads/lead-page.spec.ts, line 194, on main) · Found in: a run of e2e/browser/settings and e2e/browser/leads (222 tests, Chromium) on forge-et-pwa fix/BUG-486-leads-switch-and-ios-install, Windows, 2026-11-19 · Severity: low, a test-only failure · Phase: 1
Repro: run pnpm exec playwright test e2e/browser/settings e2e/browser/leads on a Windows machine.
Expected: the test passes.
Actual: after Convert to Lead, getByTestId('lead-write-message') in the lead’s drawer was not found within 5 s (line 208). Run alone with --retries=0 --repeat-each=3 on the same tree it passed 3 of 3. The branch changes only the Leads switch’s outline on that page, not the lead drawer or its writes. No machine load was recorded for the run (it was not a test record).
Cause (2026-11-20): a race in the test. The drawer reads the lead when it opens (showing the lead from the list meanwhile, if the list has loaded), as useCapturedLead has no staleTime. The test changed the double’s copy to version 1 as soon as Convert to Lead showed; when the drawer’s read reached the double after that change, as it can under load, the drawer held version 1, the convert was accepted and no message came. Reproduced by holding that read until after the change: the test fails with the reported error.
Fix: before leads.put, wait until the double has recorded the drawer’s read (expect.poll(() => detailReads(leads)).toContain(DEMO_LEAD_IDS.badge)); the double builds each answer in the same step that records the read, so that read carries version 0 (now stated in the double’s requests doc). The other leads.put calls on an open drawer (edit-details.spec.ts, leads-offline.spec.ts) are made inside a write’s handler or after a write has been sent, so they need no change; the other store helpers (setCardRead, updateStoredLead, setResearch) were not walked one by one.
Resolved: 2026-11-20 — fixed in forge-et-pwa#194 (3605759). With the drawer’s read held until after leads.put (a temporary route), the test on main fails as reported; with the read delayed 1.5 s, the fixed test passes, and the spec passes 36 of 36 (--repeat-each=3, Chromium, no retries).
Lesson: L-101 — the test acted before an earlier step had settled, here a read still on its way to the double, as BUG-469’s save came before the holders read had ended.
BUG-492 — Four CRM account page states with a dialog open overflow by 5px at 320x640 in Chromium
Affects: F1-d, F1-b (forge-et-fe e2e/pageStates.ts: account-edit-leave-dialog, account-detail-archive-confirm, account-detail-archive-unconfirmed, account-detail-archive-refused) · Found in: forge-et-fe#585’s selected browser run on b665fbe, 2026-11-19 · Severity: low · Phase: 1
Repro: in forge-et-fe, with a backend on ET_BACKEND_PORT: E2E_GATE=1 pnpm exec playwright test e2e/browser/size-matrix/size-matrix.spec.ts --project=size-320x640-chromium -g "account-edit-leave-dialog|account-detail-archive" --retries=0. On 2026-11-19 at about 13:30, on macOS with Chromium 1255, it failed 4 of 4 on main dbdfdb5 and on #585’s b665fbe, against the same backend (a forge-et-be worktree on an isolated stack).
Expected: the shared page check passes at 320px: no horizontal overflow (WCAG 2.1 SC 1.4.10).
Actual: <main> overflows by 5px while each state’s dialog is open; the same pages without a dialog pass. The same four states passed in that morning’s nightly full run on cc97f8c, and main has not changed the account pages since (git diff cc97f8c dbdfdb5 -- src e2e touches only lead ranking). The Mac’s Playwright browsers were reinstalled during the day (Chromium 1254 at 12:02 and 1255 at 12:29), so the browser build the night used is not known.
Investigation (2026-11-19 and 2026-11-20):
- On
main9d61c303at 21:50, this Mac’s headless Chromium 1255 drew overlay scrollbars (0px) and the four states passed. A second run at 23:00 on the fix branch passed 222 of 222 account and confirmation states at 320x640 Chromium with overlay scrollbars. Overlay scrollbars on this machine do not reproduce the failure. - Injecting
::-webkit-scrollbar{width:15px;height:15px}(a temporary change insize-matrix.spec.ts) reproduces the failure: the four states fail with<main> overflows by 5px. The element past<main>'s inner edge is the header’sprice-email-actionbutton (215.5px wide, its right edge at 309.5px against<main>'s client edge at 305px). - Measured with that CSS: with a question open,
<main>isoverflow: hidden; scrollbar-gutter: stable, set bylockScrollBehindinsrc/ui/shadcn/alert-dialog.tsx. It then has scrollbar width 15px, client width 229px and scroll width 234px. Without a question (account-detail,account-detail-saved),<main>'s ownoverflow: autoscrollbar stays 0px and its client width stays 244px. - In all four states
<main>overflows downwards (scrollHeight3744–3911px against 640px). The first fix tried on forge-et-fe#592 (keep the gutter only when<main>overflows downwards) therefore changed nothing in these states. The run in step 3, made with that fix, still failed 4 of 218. - A plain page in Playwright’s headless Chromium (
chrome-headless-shell1255) settled why the gutter takes room where the scrollbar does not. Playwright launches headless Chromium with--hide-scrollbars. The shell was started through a wrapper that adds-AppleShowScrollBars Always, which is what macOS does while a mouse is connected. With that, anoverflow: autoregion that overflows downwards has a 0px scrollbar, but anoverflow: hidden; scrollbar-gutter: stableregion reserves 15px. Without--hide-scrollbars, both take 15px. With macOS’s default overlay scrollbars, both take 0px. So--hide-scrollbarshides the scrollbar but not the gutter, and the gutter follows the Mac’s scrollbar setting. - The size matrix run through the same wrapper, with no injected CSS, on #592’s earlier fix, reproduced the afternoon failure exactly: 4 failed, 144 passed among the account states, each failure
<main> overflows by 5px. Every state without a question measured a 0px scrollbar and a 244px client width. The four question states measured a 15px gutter and a 229px client width. - The same wrapper run on the fix (#592 at
38a8833b) passed 19 of 19 account states at 320x640 Chromium, the four that failed among them, plus the delete and discard questions and the states without a question. Every state measured a 0px scrollbar, no gutter and a 244px client width. Cause: the AlertDialog’s scroll lock gave<main>scrollbar-gutter: stablebehind a question to keep the room a classic scrollbar took. Under Playwright’s headless Chromium on a Mac with classic scrollbars (a mouse connected), the scrollbar itself took no room but the gutter took 15px, so<main>became 15px narrower only while a question was open, andprice-email-actionreached 4.5px past its inner edge. The night ran with overlay scrollbars, so the gutter was 0px and the states passed. For a person using the CRM the gutter is right: a real classic scrollbar takes the same room the gutter keeps. Not a defect for users: a person with real classic scrollbars has a 15px scrollbar on<main>with or without a question, so in a 320px-wide layout the account header’s widest action would sit 5px past<main>'s edge in both. A 320px layout is reached only by zooming a wider window to 400% (WCAG 2.1 SC 1.4.10), and a browser does not zoom the scrollbar with the page, so at 400% a 15px scrollbar takes about 4 CSS px, and the action fits. Headless Chromium cannot show this case, because it hides the scrollbar. Fix: the scroll lock insrc/ui/shadcn/alert-dialog.tsxmeasures the room each region’s vertical scrollbar takes before it locks any region (offsetWidth - clientWidthless the side borders). It setsscrollbar-gutter: stableonly when that room is at least 2px, so the lock keeps exactly the room that was there; the 2px threshold allows for the two widths each being rounded to whole pixels. This replaces both the old rule (<main>always got a gutter) and #592’s first fix (a gutter whenever<main>overflows downwards). The marked regions use the same measure. Regression tests insrc/ui/shadcn/alert-dialog.test.tsxcover a<main>that overflows downwards with a 15px scrollbar (gutter kept), a 0px one and a 1px rounding remainder (no gutter), side borders counted out, and an earlier inline gutter given back on close. The 0px row failed on #592’s earlier fix, and the border row fails without the border subtraction. Resolved: 2026-11-20 — fixed in forge-et-fe#592 (merge761bcc8d; the fixd969391d, review fixes38a8833b). Under classic scrollbars the size matrix failed 4 of 148 account states on the earlier fix (step 6) and passed 19 of 19 on the fix (step 7). The PR’s selected browser run on85bbd761(the fix withmainmerged in) passed the four states at 320x640 Chromium; its 7 failures had all failed in the nightly full run onmainc77a7f67the night before (forge-et-fe#593; four are BUG-498).frontend-reviewerfound no Blocker; its three Importants (two wrong comments and the owed classic-scrollbar run) and four Nits were fixed. Lesson: L-224 (extended): Playwright’s headless Chromium hides scrollbars but not ascrollbar-gutter; reproduce classic scrollbars on a Mac with a wrapper that adds-AppleShowScrollBars Always, and set a gutter only for measured room.
BUG-490 — The capture app’s leads-not-sent page state does not list To Review’s tier chips as a scroll region, so the size matrix fails at 320px
Affects: F36-c (forge-et-pwa e2e/pageStates.ts, the leads-not-sent entry; the tier chips are F36-c WI-2.4’s, cdda966) · Found in: the hand-run nightly capture app gate on main a52b413, 2026-11-19 (forge-et-pwa#173) · Severity: low · Phase: 1
Repro: pnpm e2e:gate --full in forge-et-pwa on a52b413, or the size matrix alone at 320x640: e2e/browser/size-matrix/size-matrix.spec.ts › leads-not-sent.
Expected: the shared page check passes. To Review’s tier chips scroll sideways inside their own row at 320px by design (F36-c WI-2.4), and every page state that shows To Review names them in allowedScrollRegions.
Actual: leads-not-sent opens /, which shows To Review, but its entry has no allowedScrollRegions, so the check reports div[data-testid="leads-tier-chips"] scrolls sideways by 9px. It failed both attempts on Chromium and WebKit at 320x640. The page itself is right: only the test’s list of allowed scroll regions is incomplete.
Fix: add allowedScrollRegions: LEADS_TIER_CHIPS to the leads-not-sent entry, as every other To Review entry has. Made in forge-et-pwa#183 on its own, so that main could go green before the dependency upgrade (forge-et-pwa#172, which carried the same one-line change and now gets it from main).
Resolved: 2026-11-19 — fixed in forge-et-pwa#183 (merge 2f7bf77). The capture app gate then passed on main 2f7bf77 (9,449 tests, 0 failed), and forge-et-pwa#173 closed.
Lesson: none — one-off: WI-2.4 named the chips on each entry that opens To Review by its section, and this entry, which opens / with no section and was written for the offline banner, was missed. No second instance in the other Lesson: lines.
BUG-489 — Flaky test: AccountsTable.test.tsx’s “does not open the row when its menu is opened, and opens the account page from View” misses the account page under load (Windows)
Affects: F1-b (forge-et-fe src/features/accounts/AccountsTable.test.tsx, “AccountsTable — populated”, on main) · Found in: a run of 453 Vitest files on forge-et-fe feature/F32-d-saved-views-rev-6 (from 039737a4), Windows, 2026-11-19 · Severity: low, a test-only failure · Phase: 1
Repro: run pnpm exec vitest run src/ui src/lib src/features/accounts src/features/leads src/features/pricing src/app (453 files, 373 s) on a Windows machine with a backend and a Vite server running beside it.
Expected: the test passes.
Actual: Unable to find an element by: [data-testid="probe"] — the account page’s probe route did not render after View within the wait. Run alone on the same tree it passed (34 of 34). The branch does not change AccountsTable or the account page. No machine load was recorded for the run (it was not a test record).
Cause (2026-11-20): a race in the test. View, Edit and Archive run from the row menu’s onCloseAutoFocus (src/ui/recordRow.tsx RowActionsMenu), which Radix calls on a timer after the menu unmounts, and the test read the probe straight after the click. Reproduced by delaying the action 50 ms: the test fails with the reported error, and so do the same file’s Archive and Edit tests, which checked their callbacks the same way.
Fix: the tests that choose a RowActionsMenu item inline and read its result now wait for it: the three AccountsTable tests (findByTestId('probe'), waitFor on onArchive and onEdit), and nine more assertions with the same race that frontend-reviewer found in recordRow.test.tsx, EmailTemplatesScreen.test.tsx, AuctionSalesTable.test.tsx, RoundTable.test.tsx and StandardsTable.test.tsx (a callback, the router’s path or the trigger’s focus, each now read in a waitFor). A first search had matched only the *-row-actions-<action> test ids and missed the pricing tables’ -item-<action> ids and the menu items chosen by role. The 70 tests in 10 files that choose the item through a helper of their file have the same race and are BUG-519, found by the review’s second pass and confirmed by the delay run over the whole suite.
Resolved: 2026-11-20 — fixed in forge-et-fe#598 (9c0e92e7, the review’s additions in f6ecdffd). With the action delayed 50 ms (a temporary change in RowActionsMenu), the six files on main fail 10 tests, each at a changed assertion, and the fixed files pass 133 of 133; without the delay they pass 133 of 133.
Lesson: L-101 — the test checked before the page had settled: the row menu’s action lands on a timer after the click, as BUG-135’s focus landed a frame late.
BUG-488 — The selected saved view’s tab on the CRM lists is hard to tell from the others
Affects: F32-d (forge-et-fe src/ui/filters/ViewsRow.tsx, the views row’s TabsTrigger, on main) · Found in: Nanndish’s UI review list of the CRM and the capture app, 2026-11-19 · Severity: low · Phase: 1
Repro: open a saved view on the account, lead or contact list.
Expected: the open view’s tab is plainly the selected one.
Actual: a saved view’s tab is never drawn as selected at all. Its trigger is also a TooltipTrigger, and Radix’s Tooltip writes its own data-state (closed, delayed-open) over the tab’s data-state="active", so the pill’s data-[state=active] look (the surface/card fill, the shadow, the text colour) never applies: the open view’s tab looked like every other (measured in the browser on 2026-11-19: the open view’s tab had data-state="closed" and a transparent background). Only the default tab, which has no tooltip, showed as selected.
Fix: design-system Rev 141 (§6 Saved views): the views row’s selected tab is filled with primary and its text is primary-foreground, keyed on aria-selected, which the Tooltip does not touch, and on the pill’s own data-state for the default tab (ViewsRow.tsx SELECTED_TAB_CLASS). The other tabs are unchanged, and only the views row changes; the §5 pill tabs elsewhere keep their look. Measured after the fix: the selected tab, default or saved, is rgb(138, 90, 18) with white text (about 5.9 : 1). Tests: ViewsRow.test.tsx (the classes on the selected default and saved tab, and a saved tab whose data-state is not active); crm-current updated in the same delivery.
Resolved: 2026-11-19 — fixed in forge-et-fe#586 (b288e3bf). Before the fix the open saved view’s tab measured data-state="closed" and a transparent background in the browser; after it, rgb(138, 90, 18) with white text, for the default and a saved tab. SELECTED_TAB_CLASS did not exist before, so its ViewsRow.test.tsx row cannot pass on the old code. frontend-reviewer checked the badges on the fill (6.4 : 1) and the merge of the pill’s classes.
Lesson: L-324 (new): two Radix parts composed on one element share one data-state; style the inner part on its ARIA state. frontend-reviewer dimension 34 points at it.
BUG-487 — The CRM’s Pricing page title does not line up with the other pages’ titles
Affects: F26-b (forge-et-fe src/features/pricing/PricingRoute.tsx, the page around its PageHeader, on main) · Found in: Nanndish’s UI review list of the CRM and the capture app, 2026-11-19 · Severity: low · Phase: 1
Repro: open Pricing, then Accounts, at the same window size, and compare where the Pricing and Accounts titles sit.
Expected: the Pricing title has the same position, size and spacing as every other page’s title.
Actual: measured in the browser on 2026-11-19: both titles are 24px bold at the same left edge, but the Pricing <h1> is 24px lower (top 42px against 18px). PageHeader centres the title against its actions (BUG-164), and Pricing’s actions hold a status line below Send Prices to Customers, kept at its height by a hidden stand-in button from the first paint (F37-a, PricingActionsPlace, lesson L-208), so the actions are 84px tall and the centred title sits at their middle.
Fix: PageHeader gains titleAlign="top", which lines the title up with the top of the actions; Pricing passes it, and every other screen keeps the centred title. Pricing’s buttons are 36px, the title’s height, so the title lands where it does on every other screen: measured after the fix, the Pricing <h1> is at top 18px and left 94px, as on Accounts. The status line’s reserved space still sits below the buttons, so the description under the title is lower than on other screens. Tests: PageHeader.test.tsx (titleAlign="top" adds .rowTop, align-items: flex-start) and the measurement above.
Resolved: 2026-11-19 — fixed in forge-et-fe#586 (533e5142). Measured in the browser: the Pricing <h1> was at top 42px before and is at 18px after, as on Accounts; titleAlign did not exist before, so its PageHeader.test.tsx row cannot pass on the old code. The Pricing size-matrix states at 768px and 320px pass on the branch. The description’s lower place under the title is left: moving the status line is a change to F37-a’s actions place, not part of this bug.
Lesson: none — one-off: Pricing is the only screen whose header actions keep a line of their own below the buttons, and BUG-164’s centring is right for every other header.
BUG-486 — The capture app’s Leads page draws a border around its To Review and Completed switch
Affects: F36-b (forge-et-pwa src/features/leads/LeadsRoute.tsx, the Lead sections Tabs; src/ui/Tabs.module.css .tablist, on main) · Found in: Nanndish’s UI review list of the CRM and the capture app, 2026-11-19 · Severity: low · Phase: 1
Repro: open Leads in the capture app at phone width.
Expected: the To Review / Completed switch has no outer border: its track is the surface/inset fill alone.
Actual: the track carries a 1px border-frame outline (.tablist). The same Tabs component draws the Capture screen’s mode switch, which keeps its outline.
Fix: give Tabs a way to draw its track without the outline (a bordered prop that defaults to true), and use it on the Leads page only. The selected tab keeps its own frame, so it still differs by more than colour (design-system §5 Tabs). Tests: Tabs.test.tsx for the borderless track, and LeadsRoute.test.tsx that the Leads switch uses it; pwa-current and its Leads screenshots updated in the same delivery.
Resolved: 2026-11-19 — fixed in forge-et-pwa#186 (59aea6e). .borderless sets border-color: transparent, so the track keeps its 1px and nothing moves. bordered and .borderless did not exist before, so the Tabs.test.tsx and LeadsRoute.test.tsx rows cannot pass on the old code. The Leads size-matrix states at 390px, 768px and 320px pass (114 of 114); pwa-current’s Leads screenshots are retaken.
Lesson: none — one-off: a look the design system asked to change on one screen; the shared component keeps its default.
BUG-485 — The CRM lists’ Columns button is taller than the buttons beside it
Affects: F32-d (forge-et-fe src/ui/filters/ColumnsMenu.tsx, the PopoverTrigger’s Button, on main) · Found in: Nanndish’s UI review list of the CRM and the capture app, 2026-11-19 · Severity: low · Phase: 1
Repro: open the account, lead or contact list at desktop width and compare Columns with the controls on its row.
Expected: Columns is as tall as the buttons beside it (and, after F32-d Rev 6, as Save View, which moves onto the same row).
Actual: measured in the browser on 2026-11-19 (1280px, the local stack): every control of the data bar is 40px tall, Columns included (h-10, as the search box, the filters, Sort, More Filters and Clear Filters). The shorter button beside it was Save View (size="sm", 32px), which sat at the views row’s right end directly above Columns.
Fix: Save View moves onto the data bar after Columns (F32-d Rev 6) and takes the data bar’s 40px height (SaveViewAction, h-10, outline); Columns is unchanged. Measured after the fix: Search, the filters, Sort, More Filters, Columns and Save View all 40px on the lead list, and Columns, Save View and the filters 40px on the account list. Regression test: ViewActions.test.tsx (Save View carries h-10) and the measurement above.
Resolved: 2026-11-19 — fixed in forge-et-fe#586 (b288e3bf; Save View placed right after Columns through ColumnsMenu’s trailing slot in 433037e2). SaveViewAction did not exist before, so its ViewActions.test.tsx row cannot pass on the old code; the lead and contact screen tests check Save View is Columns’ next sibling.
Lesson: none — one-off: the 32px Save View sat on another row directly above the 40px Columns; with Save View on the data bar every control there shares the bar’s height.
BUG-496 — Flaky test: CapturePlanStatementCountIT X-1 counts 9 statements at size 5 and 10 at size 50 when the best-fit page of 5 names nobody (Windows)
Affects: F36-c (forge-et-be src/test/groovy/com/empireteas/capture/CapturePlanStatementCountIT.groovy, “X-1 at size 5 and size 50 runs the same statements”, on main; its research seed came with F36-c WI-1.3) · Found in: F36-c WI-3.1’s full build on bc6f3da3, 2026-11-19 · Severity: low, a test-only defect
Repro: run the spec alone on Windows (./gradlew test --tests 'com.empireteas.capture.CapturePlanStatementCountIT') at main (fc11045e): the &sort=best_fit case fails, 11 tests, 1 failed. The nightly full run on the Mac passed it the same night.
Expected: the page of 5 and the page of 50 run the same statements.
Actual: the seed gave each research row a fit score of 70 + (e.id % 31), so which exhibitors make the best-fit page of 5 depends on the id sequence, which moves with every run and every earlier test. When none of those 5 was visited or noted, the page names no person and skips the one app_user read the page of 50 makes (9 statements against 10). The product’s reads are still a fixed set; only the test’s data changed.
Fix: the seed takes the score from the company’s number (split_part(e.company_name, ' ', 2)), so the page of 5 is always the same and names a visitor (Company 27). Fixed in its own commit on F36-c WI-3.1’s branch, which ships with the wave 3 backend PR. Renumbered from BUG-482 on 2026-11-19, after main gave that number to another bug while the wave was open (id-numbering rule 5); the fix commit forge-et-be 8dc053a0 keeps its old subject, fix(BUG-482), and the test’s comment cites BUG-496 (09d40ad4).
Resolved: 2026-11-20 — fixed in forge-et-be#484 (F36-c wave 3, merge 0634a2d8), whose local build record on ce975e9f ran every test with 0 failed (19,360 tests). Closed at F36-c’s Reflect.
Lesson: none — one-off: a statement-count test whose page contents depend on database ids.
BUG-481 — The capture app’s Read Again answer can be overwritten by a lead read that lands after it
Affects: F36-b (forge-et-pwa src/features/leads/leadsQueries.ts:409-421, useReadAgain, F36-b WI-3.4, on main) · Found in: the third delta review of forge-et-pwa#171 (F36-c wave 2), 2026-11-18, while checking the same race F36-c fixed for Research Again (I-12, aee3a44) · Severity: low · Phase: 1
Repro: on a card lead, press Read Again while the lead’s own read is held or starts during the write (the card read’s 5-second re-read, or a reconnect), and let that read answer after the write’s answer.
Expected: the page keeps the write’s answer (Reading the card…).
Actual: useReadAgain cancels the lead read only in onMutate; its onSuccess writes the answer and invalidates only the lists, so a read started during the write lands afterwards and brings back the lead as it was before Read Again (by code reading; not reproduced in a run). useSaveLeadDetails is not affected: it reads the lead again after its answer.
Fix: as aee3a44 did for useResearchAgainLead and useRemoveResearchPerson: make onSuccess async and await client.cancelQueries({ queryKey: capturedLeadKeys.detail(id), exact: true }) before setQueryData, with a hook test that holds the write, starts a read during it and answers that read after the write. Done in forge-et-pwa#179 for useReadAgain and for useReviewWrite (convert, reject, reopen), which had the same gap. Two regression tests in leadsQueries.test.tsx, one for Read Again and one for convert, fail on main (the late read puts back version 2) and pass with the fix.
Resolved: 2026-11-19 — fixed in forge-et-pwa#179.
Lesson: L-323 — new: the third write to show this race (after the two research writes and Read Again, the review writes); a write must cancel the lead’s read again in onSuccess, not only in onMutate.
BUG-480 — Flaky test: F11-c’s LeadResearchConcurrencyIT AC-15 case times out on a loaded or sleeping machine (Windows)
Affects: F11-c (forge-et-be src/test/groovy/com/empireteas/lead/LeadResearchConcurrencyIT.groovy, “AC-15 — two deliveries of one message at once …”, on main) · Found in: F36-c WI-2.2’s full build on forge-et-be 42d723da (forge-et-be#470), 2026-11-18 · Severity: low, a test-only timeout
Repro: run the full ./gradlew build on a machine that is heavily loaded (another session’s build and bootRun beside it) or that sleeps during the run.
Expected: the test passes.
Actual: its 30-second polling condition gave up after 15,374 s of wall time, the machine having slept and been loaded during the run; the other 18,670 tests passed. Re-run alone on the same head it passed (3 tests, 42 s). F36-c did not change the file.
Fix: none — no test change can make the case survive a long host sleep, so the test is kept unchanged. Triaged 2026-11-19 by reading the code; no run was needed:
- The 30-second wait that gave up is the file’s
waitForCalls(token, 2). It is a SpockPollingConditions, whosewithinmeasures wall-clock time (System.currentTimeMillis()). A sleep counts against the timeout, and on waking it makes at most one more attempt before giving up, which matches “15,374 s”. - The second delivery reaches the AI call only after the listener’s next 1-second long poll (LocalStack in Docker). After a host sleep the long poll that was open has to fail or time out first, so it cannot arrive in time for that one attempt.
- A longer or sleep-aware wait would not help either. The held first call releases itself after 60 s (
LeadResearchScriptedAiProvider.groovy:145,await(60, SECONDS)). Once that run completes, the second delivery is skipped byLeadResearchJobHandler.skipReason(“the message’s key is already claimed”,LeadResearchJobHandler.java:124), sowaitForCalls(token, 2)can never pass after the hold has expired. - No earlier test’s run held a concurrency slot:
setup()waits for the queue to be idle and for no AI call in flight. - Nothing shows that load alone, without the sleep, exceeds 30 s. Every other
PollingConditionsin the suite (awaitQueueIdle,awaitRow) measures wall time the same way; making the suite survive a host sleep is not worth a suite-wide change. Resolved: 2026-11-19 — closedwontfixas a machine that slept about 4.3 hours mid-run (.forge/test-strategy.md § Flaky Tests, triage step 4); decided by Tharinda (lead). Not fixed because the case’s wait and its held AI call are both bounded in wall time, so no wait can pass after a long sleep, and it passes alone in 42 s. Lesson: none — one-off: a host that sleeps during a run invalidates any wall-clock wait, and the nightly run already requires mains power and the lid open (.claude/rules/git-conventions.md § Pull Requests).
BUG-477 — The capture app’s Edit Details keyboard test cannot choose a country with ArrowDown (macOS)
Affects: F36-b (forge-et-pwa e2e/browser/leads/edit-details-a11y.spec.ts:130, “by keyboard alone: open, choose a country, save with errors, correct, save”, Chromium) · Found in: forge-et-pwa#164’s selected browser run (a full run, the capture app has no feature map) on e54a260, macOS, 2026-11-18 · Severity: low · Phase: 1
Repro: the run (busy 0.85 average, load average 8.58 per core at peak, 8 cores) failed both attempts at line 142: after page.keyboard.press('ArrowDown') on the focused, closed Country select, expect(country).not.toHaveValue('') received "" for 5 s. #164 does not touch the Edit Details sheet. Alone on main’s version, --repeat-each=3 --retries=0, Chromium, macOS: 3 of 3 failed the same way, so it fails on every Mac run, not only under load.
Expected: the keyboard test chooses a country on every operating system.
Actual: on macOS, Chromium’s closed native <select> does not change on the arrow keys (.forge/test-strategy.md § Flaky Tests, cause 4; BUG-103), so the country stays empty.
Fix: choose the country by type-ahead (page.keyboard.type('Germany')), assert the value DE, and check the PATCH carries country: 'DE'.
Regression test: main’s test alone failed 3 of 3 (above); the fixed test, in the same run, passed 3 of 3.
Resolved: 2026-11-18 — fixed in forge-et-pwa#164 (889636d), the PR whose selected browser run found it. The test came with F36-b wave 3 (forge-et-pwa#159). The final PATCH check is stronger too: it was not.toMatchObject({ country: null }) and is now toMatchObject({ contactName: 'Lena Hartmann', country: 'DE' }). frontend-reviewer checked that “Ger” already picks Germany alone, that the card starts with no country, and that no other capture-app test presses an arrow key on a select.
Lesson: L-143 — the rule (type-ahead for a closed native select on macOS) and frontend-reviewer’s check 33 already name it; the check did not catch it in that wave’s review, and the arrow key works on Windows, so it fails only on a Mac.
BUG-475 — Typing in the capture app’s Leads search right after a section switch can undo the switch
Affects: F36-a (forge-et-pwa src/features/leads/LeadsRoute.tsx, update(), with LeadsSection.tsx’s search box) · Found in: frontend-reviewer’s review of forge-et-pwa#164 (BUG-470’s fix), reproduced 2026-11-18, macOS · Severity: low — it needs a tap and typing within one render, which a slow phone can give · Phase: 1
Repro: on main (forge-et-pwa e2e, Chromium, the CDP CPU throttle at 120x, a temporary spec, --repeat-each=10 --workers=5): open /?section=completed&outcome=rejected&completedSearch=da, click To Review, at once type maria into the search box on the page, wait 3 s. 8 of 10 runs ended on /?section=completed&completedSearch=mariada&outcome=rejected, with Completed selected again.
Expected: the switch to To Review stays; typed text never brings back the section the rep left.
Actual: React Router 7 applies the switch in a transition, so Completed’s box stays on the page for a while after the tap and takes the keys. When its 300 ms debounce fires before the switch has rendered, update() calls React Router’s functional setSearchParams, which starts from the parameters of the last render, still section=completed, and writes them back with the new search.
Fix: not yet. Likely: build the Leads screen’s URL writes with useLatestSearchParams (src/features/plan/latestSearch.ts, BUG-431’s fix), so each write starts from the latest written parameters, and add a test that holds the switch’s render (L-201).
BUG-474 — Flaky test: the capture app’s local-record test that carries a record across a merge of main times out at 60 s under load (Windows)
Affects: F40 (forge-et-pwa src/test/scripts/localTestRecord.test.ts:188, “the local test record › carries a record across a merge of main whose own paths touch no test input (harness design test-run-cost.md W-5)”; the file is a byte copy of forge-et-fe’s, scripts/shared-with-fe.json) · Found in: BUG-473’s checks on branch fix/BUG-473-capture-logo-csp (off forge-et-pwa main at b7a7516), Windows 11, 2026-11-18 · Severity: low — test code only · Phase: 1
Repro: on Windows 11, with other sessions’ work keeping the CPU busy (Win32_Processor LoadPercentage 100% when the run started), pnpm test failed it after 255 s, and pnpm exec vitest --run of six files including this one failed it again after 155 s: Test timed out in 60000ms. The test builds about 40 commits with one git process each (its own comment at line 186). BUG-473 does not touch the record script or this test. A third time, in pnpm test:record on forge-et-pwa#165’s head 9dae0b5 (the full suite, 168 files), it failed both attempts (213 s with its retry); the record’s load line: busy 0.28 at start, 1.00 peak, 0.75 average, 16 cores, with the antivirus (MsMpEng) and System the busiest processes during the earlier runs. Run alone straight after (pnpm exec vitest --run src/test/scripts/localTestRecord.test.ts -t "carries a record across a merge of main" --retry 0), it passed, the whole command taking 42 s. The record that followed, with VITEST_MAX_WORKERS=4 (busy 0.15 at start, 0.86 peak, 0.70 average), passed it in 56.2 s, close to its 60 s limit.
Expected: the test passes under the load a developer’s machine normally carries, or its timeout matches the time its git fixture needs on Windows.
Actual: see the repro.
Fix: the same test as BUG-471, filed the same day from another branch. BUG-471’s fix, merged on 2026-11-18, gives it 180 s (forge-et-fe#580, re-copied by forge-et-pwa#168). Biseni claimed this entry the same day; she confirms whether that fix settles it and closes it.
BUG-473 — The capture app’s Empire Teas logo is a broken image on every deployed environment
Affects: F36-a (forge-et-pwa src/ui/brand/Logo.tsx: the header logo and the sign-in panel’s logo) · Found in: the deployed dev capture app, https://capture.dev.crm.empire.lk/sign-in, 2026-11-18 · Severity: medium — the brand is missing on the first screen every rep sees; nothing else is affected · Phase: 1
Repro: open the dev capture app’s sign-in page. The logo above Empire Teas is a broken image (naturalWidth 0), and the console says: loading the image data:image/webp;base64,… violates the Content-Security-Policy directive img-src 'self'. Logo.tsx imported the logo with ?inline, so it shipped inside the bundle as a data: URL (F36-a WI-1.2 plan, subtask 15 note, chose this to keep .webp out of the service worker’s allow-list). The capture app’s policy, applied on dev by forge-et-iaac#14 (.forge/design/environments-and-identity.md §4), allows img-src 'self' only. pnpm dev, pnpm preview and the browser gate send no policy, so every local check passed. The favicon, the iOS home-screen icon and the manifest icons are separate files and load correctly (curl returns 200 with the right type for each).
Expected: the logo shows on every environment, offline included, under the policy as it is.
Actual: see the repro.
Fix: import the logo as a file (import logoUrl from './empire-logo.webp'), so Vite builds it as /assets/empire-logo-<hash>.webp, which img-src 'self' allows; add .webp to SHELL_EXTENSIONS in src/lib/sw/shouldCache.ts and public/sw.js together, so the worker serves the file it already stores at install and the logo shows offline, with a .webp row in src/lib/sw/swParity.test.ts; set build.assetsInlineLimit: 0 in vite.config.ts, so Vite never inlines a small asset by size either. Regression tests: src/ui/brand/Logo.test.tsx fails while the logo’s src is a data: URL; scripts/validate-pwa.mjs, run by pnpm build, fails a build whose JS or CSS carries an image, font or media file as a data: URL (the older font types included). Rejected: adding data: to the capture app’s img-src, which loosens the policy for every image and needs a forge-et-iaac change, for a problem the app can fix itself.
Resolved: 2026-11-18 — fixed in altrium-io/forge-et-pwa#165, as the Fix above describes. Logo.test.tsx failed before the fix (expected 'data:image/webp;base64,UklGRuYwAABXRU…' not to match /^data:/) and passes after; validate-pwa.mjs failed on a built bundle with a data:image URL added and passes on the fixed build.
Lesson: L-317 (new) — the capture app’s Content-Security-Policy exists only on deployed environments, so nothing that runs locally checks against it; the blob: photo previews (.forge/owed.md, “Allow blob: images in the capture app’s CSP”) are the same cause.
BUG-472 — Flaky test: the Email screen’s retry focus tests check the heading’s focus before the retry has settled (macOS)
Affects: F20-i (forge-et-fe src/features/email/EmailRoute.test.tsx, the block “EmailRoute — focus when a pressed control leaves the page (L-133)”: the heading-focus checks after Try Again on a failed first read, on the stale band and on a failed folder tree) · Found in: the nightly full run on main, 2026-11-18, forge-et-fe 6b53f6dd, macOS (forge-et-fe#571) · Severity: low — 1 failure in 15,662 tests that night · Phase: 1
Repro: the night’s CRM Vitest run (744 files, the lead’s Mac; the night records no load figure for this suite) failed “a press that never focused it (fireEvent.click) still moves focus to the heading” under Try Again on a failed first read, in 97 ms: findByTestId('inbox-list') resolved, then expect(heading()).toHaveFocus() found focus on <body>. Alone at the night’s commit, with retries off, it passed 3 of 3. It did not fail by chance locally on c2add06e either: the focus block alone 15 of 15 and the email folder (2,490 tests) twice, each under one busy loop per core.
Expected: focus is on the Email heading once the retry has settled.
Actual: once, focus on <body> at the check.
Fix: wait for the focus with waitFor in each check that follows a retry, and let the retry settle before each check that focus did not move.
Regression test: with the product’s settle step delayed by 50 ms (setTimeout around setSettledActions in useFocusAfterSettle, a temporary local edit), the old tests failed 6 of the block’s 17; with the fix, 17 of 17 passed.
Resolved: 2026-11-18 — fixed in forge-et-fe#576. No product defect. After a retry, the Email screen moves focus in a separate render, queued when the retry’s promise settles (useFocusAfterSettle in EmailRoute.tsx: a state update follows the settled promise, and a layout effect moves focus when focus is lost). That render can commit with the list or folder tree, or after it. Six checks waited for the list, the stale band to go, or the tree, and then checked focus at once, so a busy machine could run the check between the two renders; the delayed-settle run above showed all six depend on it. Each now waits with await waitFor(() => expect(heading()).toHaveFocus()), and the block’s comment says why. The review (frontend-reviewer) found the reverse gap in the block’s six checks that focus did not move (“does not take focus back”, “a failed retry leaves the same Try Again in place and focused”): they were anchored on the list, the tree or aria-busy, so when the focus render came later they passed even with the “only when focus is lost” condition removed (3 of the 6 failed under that mutation with the move queued one task later). Each now runs settle() (an act around a zero-delay timer) first, and all 6 fail under the same mutation. The Show tests in the same block already waited. Whole file: 182 of 182; the focus block 15 times under one busy loop per core: no failure. waitFor accepts a focus move up to 1 s late, so a regression to a timer-based move (L-271) would no longer fail these checks; nothing in the page marks the settle, so the review accepted that. MailboxRoute.test.tsx’s matching check after a connection re-read (another screen and hook, useRefetchRestoringFocus) was left as is.
Lesson: L-101 — an action whose focus move is deferred is followed by a wait for focus to land, not a check made when an earlier change appears.
BUG-471 — The capture app’s local-record test “carries a record across a merge of main” runs past its 60-second limit, even alone (Windows)
Affects: F40 (forge-et-pwa src/test/scripts/localTestRecord.test.ts, “carries a record across a merge of main whose own paths touch no test input (harness design test-run-cost.md W-5)”, line 188, timeout: 60_000; a byte copy of forge-et-fe’s, last re-copied by 158e1b8 from forge-et-fe cbdfaf75; test code only) · Found in: F36-a-WI-5.1’s /forge-test-verify (corepack pnpm test:selected on forge-et-pwa c1c3583, branch feature/F36-a-WI-5.1-e2e-capture-and-upload), Windows 11, 2026-11-18 · Severity: medium — pnpm test:record records nothing when a test fails, and it selects this file, so on this machine a forge-et-pwa PR whose change selects it cannot get its Test record check; the product is not affected · Phase: 1
Repro: in a forge-et-pwa worktree on Windows 11, corepack pnpm exec vitest run src/test/scripts/localTestRecord.test.ts -t "carries a record across a merge of main". The test builds a git fixture of about 40 commits and runs the record script’s merge-of-main case against it.
Expected: it passes within its limit on every supported developer machine, alone and inside a full or selected run.
Actual: it timed out at 60 s inside the selected run (987 tests, 1 failed), and again alone at 78.9 s, with the CPU at 36% on 16 logical processors, so machine load alone does not explain it. Process start-up for git is slow on Windows (as in BUG-172 and BUG-427, the same file’s earlier timeouts).
Fix: the test gets 180 s, nearly twice the slowest time measured (the first commit gave it 120 s; the review of #580 asked for more room): forge-et-fe#580 (b638b385) changes it in forge-et-fe, and forge-et-pwa#168 (d79961d) re-copies it and records the new hash in src/ui/SHARED-WITH-FE.sha256. Merge #580 first. Building the fixture with fewer git processes was not needed.
Lesson: an instance of L-257 (synchronous git work in a test needs a timeout of its own); its update of 2026-11-18 adds that 60 s is not enough for a fixture of about 40 commits on Windows.
Resolved: 2026-11-18 — fixed in forge-et-fe#580 and forge-et-pwa#168. The test passed alone in about 41 s (retries off, 53 s with set-up), and the record of each PR runs it under the new limit.
BUG-470 — Flaky test: the capture app’s Leads list keeps-the-search browser test sees five rows where it expects one (macOS)
Affects: F36-a (forge-et-pwa e2e/browser/leads/leads-list.spec.ts:72, “each section’s search and the chip are kept across a section switch, a reload and a lead opened and closed”, Chromium) · Found in: forge-et-pwa#160’s selected browser runs (widened to full runs) on 7539c91, macOS, 2026-11-18 · Severity: low · Phase: 1
Repro: two runs on the same head. The first (load: busy 0.72 average, load average 2.13 per core at peak, 8 cores) passed it only on its retry. The second (busy 0.76 average, load average 7.24 per core at peak) failed both attempts: expect(getByTestId('lead-row')).toHaveCount(1) received 5 for 5 s, so the list still showed every lead where the kept search should have narrowed it to one. #160 copies the browser gate’s scripts only and does not touch the Leads screens.
Expected: the test passes whatever else the machine is running.
Actual: see the repro.
Fix: after the switch to To Review, wait for its five rows, type into the To Review box by its accessible name, and check the box holds maria before counting the rows.
Regression test: a temporary copy of the steps under the CDP CPU throttle at 120x, each wait raised to 30 s so that slowness alone cannot fail it, --repeat-each=10 --workers=5, Chromium, macOS: main’s steps failed 6 of 10 (in 5 the focus was on <body> with the To Review box empty or not yet on the page, in 1 the text went into the Completed box as mariada); the fixed steps passed 10 of 10.
Resolved: 2026-11-18 — fixed in forge-et-pwa#164. No product defect, and the failing check was not the one the entry first suspected: the kept search after the reload was fine. React Router 7 applies the tab switch in a startTransition, LeadsSection is keyed on the section, and a section’s loading state shows no search box (the list query keeps no placeholder data across sections). Right after the click, under load, getByRole('searchbox') could still find the Completed box, so pressSequentially('maria') typed into a box that was about to unmount, and the keys were lost with it. Nothing was searched, and To Review kept its five rows. No other capture-app spec types right after a section switch.
Lesson: L-101 — the test acted on an element while the previous click’s change was still being applied; it now waits for that change to show (the new section’s rows) first.
BUG-469 — Flaky test: the roles self-admin AC-24 browser test times out waiting for the refused roles read, and its retry finds the role still changed (macOS)
Affects: F5-d (forge-et-fe e2e/browser/roles/self-admin-next-request.spec.ts:117, “AC-24: after the save, the next request is refused 403, the /me re-read lacks the Users and Roles page keys, and the no-access state shows with no sign-out and no focus event”; test code or a race it exposes) · Found in: forge-et-fe#568’s selected browser run (widened to a full run) on ae291179, macOS, 2026-11-18; earlier passed only on its retry in forge-et-fe#553’s selected run on d4f3352, macOS, 2026-11-17 · Severity: low · Phase: 1
Repro: in the gate run, with nothing else on the machine but the run (result.json load: busy 0.08 at start, 0.97 peak, 0.83 average; load average 2.79 per core at peak, 8 cores), the first attempt timed out after 30 s in page.waitForResponse for the GET /api/v1/roles read the back link starts (spec line 223). Its retry then failed at line 210, roles-view not found, because the first attempt never restored the role it had stripped of the Users and Roles page keys. Alone, pnpm exec playwright test e2e/browser/roles/self-admin-next-request.spec.ts --project=chromium --retries=0 --repeat-each=3 --workers=1 on the same head passed 3 of 3.
Expected: the test passes under the gate’s own load, and an attempt that fails leaves the role as seeded, so its retry starts clean.
Actual: see the repro. The change under test (forge-et-fe#568, the gate’s flaky comment) does not touch the roles screens.
Fix: wait for the role page’s holders list (roles-view-holders) before opening the editor, so the page’s own reads have ended before the save.
Regression test: a temporary copy of the test that holds the role page’s holders read (GET /api/v1/users, a page.route) until the save has answered, backend forge-et-be main at 3eae0925 on the local profile, Chromium, macOS: main’s steps failed 2 of 2 with the reported error, page.waitForResponse: Test timeout of 30000ms exceeded at the list read; the request log showed GET /api/v1/users 403 after the save, then the /me re-read, and no GET /api/v1/roles. The fixed steps with the same hold (released after 3 s when no save comes) passed 3 of 3.
Resolved: 2026-11-18 — fixed in forge-et-fe#577. No product defect: the client did what F5-d FR-12 asks. The role’s page reads the people who hold the role, and the test saved without waiting for that read. Under load the read reached the server after the save had removed the Users and Roles pages, so it was refused 403; the client re-read /me and showed the no-access state, the back link sent no list read, and the test’s waitForResponse waited until the test timed out. With other timing the no-access state replaces the role’s page before the test checks roles-view, which is the retry’s failure at line 210. The entry’s guess that the first attempt left the role changed was wrong: the test heals the role first (healRole), so each attempt starts from the seeded role, and no afterEach is needed. self-admin.spec.ts removes only the Roles page, so its holders read is not refused, and its comment already allows an earlier /me re-read. The CDP CPU throttle at 20x made main’s test fail on its 30 s budget during sign-in and navigation instead, so it was not used as proof.
Lesson: L-101 — before an action whose effect another read can see, let the page’s own reads end and show on screen (the BUG-115 bullet); here a read answered after the save was refused.
BUG-468 — The empty plan’s upload states still fail in WebKit (Windows), a recurrence of BUG-458
Affects: F36-a (forge-et-pwa e2e/pageStates.ts plan-list-empty-reading and plan-list-empty-refused, in size-matrix.spec.ts:134 and capture-layout.spec.ts:884/:969), WebKit sizes only · Found in: F36-b WI-3.6’s full browser gate (R3) on feature/F36-b-WI-3.6-e2e at 8ef9f17, Windows 11, 2026-11-17, then reproduced on main · Severity: medium — every forge-et-pwa gate run on Windows fails or flakes on them · Phase: 1
Repro: on Windows, from a forge-et-pwa checkout: E2E_GATE=1 corepack pnpm exec playwright test --project=size-320x640-webkit --project=size-360x780-webkit --project=size-768x1024-webkit --project=size-1280x800-webkit --retries=0 -g "plan-list-empty-(reading|refused)".
- On
mainat f429b05: 17 of 24 failed. - On F36-b WI-3.6’s branch at 8ef9f17: 21 of 24 failed.
- In the full gate (one retry, machine load averaging 0.86): 3 failed and 10 flaky, every one of them these two states in WebKit.
Expected: each state reaches its ready test id (
exhibitors-empty-upload-reading,upload-refused). Actual:getByTestId('exhibitors-empty-upload-reading')“element(s) not found” after 5 s — the same symptom as BUG-458. Fix: not yet. BUG-458 was closedwontfixbecause it passed 36 of 36 on a Mac, and its entry asks for a recurrence to be filed as a new bug linking it. This one reproduces on Windows, so BUG-458’s likely cause (the arrangement gives the file before the panel’s mount effects run, and StrictMode’s double mount aborts the read) and its tried arrangement change are where to start. Seen again: 2026-11-18, Windows 11, in the full browser runs of forge-et-pwa#168 (d79961d, flaky at every WebKit size) and forge-et-pwa#162 (6d2d791, load 0.92 average): in #162’s runplan-list-empty-refusedat 320×640 WebKit failed both attempts, and the importing and offline upload states (plan-upload-importing,plan-upload-check-offline) failed attempts too. Onmainatfdd93fa,plan-list-empty-refusedat 320×640 WebKit failed 3 of 3 alone (--retries=0 --repeat-each=3).plan-upload-importingat 844×390 WebKit, which failed both attempts in #162’s run, passed alone. Seen again: 2026-11-19, Windows 11, in the full browser run of forge-et-pwa#185 (F36-d wave 2,a1bd8e6, load 0.92 average, 1.00 peak): both states failed both attempts at 320×640 WebKit and were flaky at every other WebKit size; the run’s other 10,632 tests passed. Run alone (--project=size-320x640-webkit --project=size-390x844-webkit -g "plan-list-empty-re" --retries=0), 10 of 12 failed on the branch and 9 of 12 onmainat7f5d2fd, so the branch does not cause it. Lesson: none yet.
BUG-467 — Flaky test: CardReadLogHygieneIT reads its log list while a worker still writes to it (Windows)
Affects: F36-b (forge-et-be src/test/groovy/com/empireteas/capture/CardReadLogHygieneIT.groovy, F36-b wave 2’s code) · Found in: F36-c WI-2.1’s full build on the F36-b WI-3.1 line, Windows 11, 2026-11-17, reported by the F36-c wave 2 session · Severity: low · Phase: 1
Repro: not reproduced on demand. One of WI-2.1’s two full builds threw a ConcurrentModificationException at line 107; the other passed.
Expected: the test passes whatever the card-read worker is doing when it reads the log lines.
Actual: at line 107, lines.list.collect { … } iterates the Logback ListAppender’s plain ArrayList while the background card-read worker can still append a log event to it, which throws ConcurrentModificationException. Lines 90 and 102 read the same list the same way.
Fix: capture the lines with com.empireteas.support.MdcCapturingAppender, as BUG-252 did for the same shape (it is already used by ExportLogHygieneIT and PlanResearchRetryIT), so every read takes a snapshot under the appender’s lock. Made on F36-b wave 3’s forge-et-be wave branch as 5118031a (forge-et-be#468): the retry wait and the then: block each read a copy of the appender’s CopyOnWriteArrayList, and the appender fixes each event’s message and MDC on the logging thread. The test passed 3 of 3 runs alone.
Resolved: 2026-11-18 — fixed in forge-et-be#468 (merged as 3eae0925). The race cannot be made to fail on demand, so there is no fails-before result; MdcCapturingAppenderSpec proves the appender. The same PR’s review moved the new CardReadReviewLogHygieneIT onto it too (b6d718e2), and its build record passed 5,842 tests with 0 failed.
Numbering: filed as BUG-465 on the F36-b wave 3 branch, and renumbered on 2026-11-17 because main had its own BUG-465 and forge-et-harness#1474 claims BUG-466. The fix commit’s message still says BUG-465. #1474’s BUG-466 sweeps the same ListAppender shape across the backend’s log tests and lists this test; this entry is the one instance seen failing, fixed here. BUG-464 is the same shape in CardReadPipelineIT.
Lesson: a test that reads a ListAppender while a background worker logs uses MdcCapturingAppender, as L-228 already says; this test was written after L-228 and missed it.
BUG-466 — Backend log tests read a plain ListAppender’s list on the root logger while other threads may still add to it
Affects: F39, F36-b, F20-b, F26-c and other features’ log tests (forge-et-be src/test/groovy; test code only) · Found in: backend-reviewer’s review of forge-et-be#466 (BUG-464’s fix), 2026-11-18 · Severity: low — no failure seen yet outside BUG-464 · Phase: 1
Repro (from the code): BUG-464’s cause, in other specs. Each attaches a plain Logback ListAppender to the root logger and reads its ArrayList while it is still attached, so another thread logging meanwhile can make the read throw ConcurrentModificationException; where a spec copies the list with new ArrayList<>(captured.list) without holding the appender’s lock, the copy can see an entry not yet written and the check throws NullPointerException.
- Reads the live list:
api/AttachmentPipelineIT,api/ContactControllerIT,api/CustomFieldValueLeakIT,api/ErrorLoggingIT(ERROR only, so less exposed),api/TracePropagationIT,capture/CardReadLogHygieneIT,capture/cardread/CardPhotoPreparerSpec,capture/cardread/CardReadRulesSpec,customfield/CustomFieldLabelLeakIT,emailtemplate/EmailTemplateContentLeakIT,mailbox/AllMailContentLeakIT,mailbox/EmailSendContentLeakIT,mailbox/LiveConversationIT,mailbox/LiveInboxContentLeakIT,mailbox/LiveNewMailIT,mailbox/MailboxContentLeakIT,mailbox/MailboxObservabilityIT,platform/audit/AuditServiceIT,pricebroadcast/PriceBroadcastContentLeakIT,priceemail/PriceEmailContentLeakIT,pricing/PriceImportLogIT. - Copies without the lock:
account/LogHygieneIT,AccountNoteLogHygieneIT,AccountTagLogHygieneIT,AccountTimelineLogHygieneIT,account/repository/ContactAddressLookupIT,capture/CaptureLogHygieneIT,CapturePlanLogHygieneIT,CaptureReviewLogHygieneIT,CustomerCheckLogHygieneIT,ExhibitorListLogHygieneIT,capture/research/CaptureResearchLogHygieneIT,ExhibitorResearchLogHygieneIT,PlanResearchLogHygieneIT, everylead/*LogHygieneIT,migration/MigrationLogHygieneIT,PipedriveRunHygieneIT,mailbox/EmailAttachmentContentLeakIT,LiveSourceContentLeakIT,ScheduledEmailContentLeakIT. - Already safe: the specs that copy under
synchronized (captured)(MailboxRenewalITexplains why) or detach first, and those onMdcCapturingAppender.
Expected: every log test reads a list no other thread changes while it reads.
Actual: see the repro. Listed from the code on forge-et-be main at 6eaea2b2; re-check with git grep -n "new ListAppender" -- src/test before fixing.
Fix: every spec that builds a Logback ListAppender builds a MdcCapturingAppender instead (L-228), the already-safe ones included, so there is one rule for every log test; a guard spec fails a new plain ListAppender.
Regression test: support/LogCaptureGuardSpec reads src/test/groovy and fails when a spec other than MdcCapturingAppenderSpec (which keeps one on purpose, to show the race) builds a plain ListAppender. On main’s test sources it failed, naming the specs; with the sweep it passes. The race itself cannot be made to fail on demand in each spec; MdcCapturingAppenderSpec proves the appender, as for BUG-464 and BUG-467.
Resolved: 2026-11-18 — fixed in forge-et-be#471. git grep "new ListAppender" -- src/test found 126 spec files beside MdcCapturingAppenderSpec, more than the list above. Each now constructs a MdcCapturingAppender; the declared types stay ListAppender<ILoggingEvent> and no read changed, because the appender’s CopyOnWriteArrayList makes every read, a live iteration or an unlocked copy, a consistent snapshot. ExportLogHygieneIT and CardReadReviewLogHygieneIT already captured the root logger at DEBUG with it, so its copy-on-write cost was already in the suite.
Lesson: L-228 — the cause is the one L-228 names, and the guard spec now enforces it for every new log test.
BUG-465 — A selected browser run names its flaky tests nowhere the weekly flaky report reads
Affects: FS-4, the browser gate’s flaky-test reporting (forge-et-fe and forge-et-pwa scripts/browser-gate/run.mjs, where flakyComment is posted only after a --push; harness .claude/lib/flaky-report.py) · Found in: forge-et-fe#553’s selected browser run on d4f3352, 2026-11-17 · Severity: low — no test or product defect; flaky occurrences go uncounted · Phase: 1
Repro: run pnpm e2e:gate --selected on a PR into main and have a test fail and then pass on its retry. The gate prints the flaky names and puts the count in its Browser gate: line, which the developer copies into the PR description as Selected browser run: …, but it posts no <!-- browser-gate-flaky comment: that comment is written only in the --push path, which a PR into main no longer uses (PRD Rev 81, Rev 93). flaky-report.py reads the gate comments, the record comments and the nightly log, so it never sees the names.
Expected: every flaky test a gate run finds is counted by the weekly report (.forge/test-strategy.md § Flaky Tests: “every published flaky name over the last 14 days”), the selected run’s included.
Actual: the selected run, now the browser run of every frontend PR into main, publishes its flaky names only in the PR description, which the report does not read. forge-et-fe#553’s two (the roles AC-24 test in Chromium, and the size-matrix account-detail-tab-standards state at 768×1024 WebKit) were posted by hand in the gate’s comment format so the report counts them.
Fix: after a passing --selected run with flaky tests — whether the selection narrowed the run or widened it to every spec — the gate comments the names on the checked-out branch’s open PR, in the same <!-- browser-gate-flaky sha=… --> form with “(a selected run)” in its sentence, which flaky-report.py already parses; with no open PR, no gh sign-in, a detached HEAD or main/release/*, it says on stderr that the names were not commented, and the run’s exit code does not change. --push posts through the same helper, unchanged. In forge-et-fe, then copied byte for byte into forge-et-pwa (its scripts/shared-with-fe.json) once the frontend PR merges. The harness’s flaky-report.py docstring and self-test and test-strategy.md § Flaky Tests say so.
Regression test: forge-et-fe src/test/scripts/browserGateRun.test.ts § a selected run names its flaky tests on the branch’s PR (BUG-465): the comment on the open PR and the comment after a widened run fail on main’s run.mjs (no comment is posted), as does the no-PR warning.
Resolved: 2026-11-18 — fixed in altrium-io/forge-et-fe#568, as the Fix above describes, and copied into the capture app by altrium-io/forge-et-pwa#160.
Lesson: none — one-off: the comment was tied to --push while --push was every PR’s gate run, and the change that made the selected run a PR’s browser run (PRD Rev 81, Rev 93) did not move it; no other bug has that cause.
BUG-464 — Flaky test: CardReadPipelineIT’s AC-24 log check reads the captured log list while other threads still add to it (Windows)
Affects: F36-b (forge-et-be src/test/groovy/com/empireteas/capture/cardread/CardReadPipelineIT.groovy, “with the root and com.empireteas loggers at DEBUG, the three runs log no card value and no photo (AC-24)”, line 136; test code only) · Found in: forge-et-be#457’s build record (F36-a wave 4) on 44af099e, Windows 11, 2026-11-17 · Severity: low — it passed on its retry · Phase: 1
Repro (from the code and the run): the test attaches a Logback ListAppender to the root logger with DEBUG on, then iterates captured.list with every { … } while it is still attached (it is detached only at line 142). Any other thread that logs at DEBUG during that check (a connection pool, a scheduler, Spring’s own threads) adds to the same ArrayList. The run failed with ConditionFailedWithExceptionError … Caused by: java.util.ConcurrentModificationException at line 136 (TEST-com.empireteas.capture.cardread.CardReadPipelineIT.xml, flakyFailure), and passed on its retry.
Expected: the check reads a list no other thread changes, and passes whatever else is logging.
Actual: see the repro.
Fix: the test captures with MdcCapturingAppender (src/test/groovy/com/empireteas/support/), the project’s appender for a test that reads events while other threads log: it keeps a CopyOnWriteArrayList, so any and every read a stable copy, and it fixes each event’s lazy fields on the logging thread, which this test also reads (L-228).
Regression test: the same feature method now starts a thread that keeps logging at DEBUG from the end of when: to cleanup:, so its checks always read the list while another thread adds to it; on the plain ListAppender (1c79f3de) that throws ConcurrentModificationException.
Resolved: 2026-11-18 — fixed in altrium-io/forge-et-be#466, as the Fix above describes.
Lesson: L-228 — a captured event read on another thread than the one that logged it; the same appender BUG-252 introduced.
BUG-491 — Flaky test: the lead page’s customer-check browser test releases the held check before the request has arrived (Windows)
Affects: F36-d (forge-et-pwa e2e/browser/leads/lead-customer-check.spec.ts, added by F36-d-WI-2.3 in eab880b, on feature/F36-d-wave-2) · Found in: F36-d wave 2’s Stage 11 Pass 1 browser run, Windows 11, 2026-11-19 · Severity: low · Phase: 1
Repro: the test the matched account deactivated: opened again from its row, the lead reads Not found; a browser reload shows loading (line 76) failed once in wave 2’s Pass 1 run, in Chromium with 4 workers, while the machine was under load from that run; the run’s load figure is not in this record. Run alone, it passed 4 of 4.
Expected: the test passes whatever else the machine is running.
Actual: the test sets the lead’s check to hold, reloads the page, sees Checking the CRM for Palmgate Supermarkets LLC…, and then calls check.release('not-found') at line 95 without first waiting for the held E-2 request. The region shows Checking… as soon as the app starts the request, before the request reaches the route, and release() answers only the requests held at that moment (e2e/support/mockCustomerCheck.ts). Under load nothing is held yet, so nothing is released, the request is then held with no release to come, and the wait for Not found times out.
Fix: wait until the double holds the request before releasing it, as the test at line 151 does at line 161: add await expect.poll(() => checksOf(check, BADGE)).toBe(3); before line 95’s release (three requests: the two opens from the row and the reload). The test at line 185 releases at line 194 the same way after only seeing Checking again…, so the fix gives it the same wait (toBe(2)). Another instance of .forge/lessons.md L-307. Not fixed in F36-d wave 2. On main the second test already had its wait (line 196, added after this bug was filed), so only line 95 changed. The test reads the request count just before page.reload() (const sent = checksOf(check, BADGE)) and waits for sent + 1 before check.release('not-found'), so it waits for the reload’s own request whatever came before (a fixed toBe(3) at first; made relative on frontend-reviewer’s Nit). The double records each request and holds it in the same synchronous step (e2e/support/mockCustomerCheck.ts), so the count reaching sent + 1 means the reload’s request is held.
Regression test: the fixed test itself. A temporary copy delayed the reload’s request 1.5 s before it reached the double (a later page.route that waits, then calls route.fallback()): the old test failed 3 of 3 (customer-check-not-found not found within 5 s), and the fixed one passed 3 of 3 (Chromium, no retries, Windows 11).
Resolved: 2026-11-22 — fixed in forge-et-pwa#208 (commit 6093236). The fixed test passed 5 of 5 alone, and the whole spec’s only other failure was BUG-534, fixed in the same PR.
Lesson: L-307 — the test released a held request after seeing only the screen change that comes before the request arrives.
BUG-482 — The customer check’s Not found sentence ends with two full stops when the company name ends in “.”
Affects: F36-d (forge-et-pwa src/lib/customerCheck/customerCheckText.ts, wave 1, forge-et-pwa#139) · Found in: pwa-current screenshot phone/28-lead-converted.png, re-taken by F36-d’s wave-2 drawings on 2026-11-17; the built text adds the same full stop · Severity: low · Phase: 1
Renumbered: filed as BUG-465 on feature/F36-d-wave-2, renumbered BUG-471 on 2026-11-18, and renumbered BUG-482 on 2026-11-19, because main took BUG-471 first (the capture app’s local-record test timeout, forge-et-pwa#168) while this branch was unmerged (.claude/rules/id-numbering.md rule 5). The fix commit 0ef587b (Refs: BUG-471) and two code comments on forge-et-pwa feature/F36-d-wave-2 (src/lib/customerCheck/customerCheckText.ts and its test) still say BUG-471.
Repro (from the code): a check that answers not_found for the company Thornbury & Vale Tea Co. shows No account or contact you can see in the CRM matches Thornbury & Vale Tea Co…; the “several” sentences end the same way.
Expected: one full stop: the sentence adds none when the name already ends in “.”.
Actual: see the repro.
Fix: in F36-d-WI-2.3 (.forge/plans/F36-d/F36-d-WI-2.3-lead-page-check-plan.md, Decision 13), which changes the same file: one private helper in customerCheckText.ts adds a sentence’s full stop only when the value does not already end in “.”, used by the not-found sentence and the four several sentences, on Capture and on the lead’s page.
Resolved: 2026-11-19 — fixed on forge-et-pwa feature/F36-d-wave-2 by commit 0ef587b (fix(F36-d): add no second full stop after a name ending in one), with rows in customerCheckText.test.ts (checkLines — a value that already ends in a full stop) and the browser test a company ending in a full stop gets no second one in lead-customer-check.spec.ts; it ships in F36-d wave 2’s forge-et-pwa wave PR, and the status is true when that PR merges.
Lesson: none — one-off: one sentence helper in one file appended its own full stop to a value from the lead; the fix covers every sentence in that file that ends with a value.
BUG-463 — Flaky test: the capture app’s SearchBox send-once test failed once under a loaded test run (Windows)
Affects: F36-a (forge-et-pwa src/ui/SearchBox.test.tsx, “owns its text and sends it once, delayMs after the last keystroke”; wave 3 code, not changed by F36-a WI-4.2) · Found in: F36-a WI-4.2’s pnpm test:selected run at f0e2354 on feature/F36-a-WI-4.2-leads-review, Windows 11, 2026-11-16 · Severity: low · Phase: 1
Repro: not reproduced alone. It failed once in the selected Vitest run while WI-4.1’s backend build was running on the same machine, and passed 3 of 3 when run alone.
Expected: the test passes whatever else the machine is running.
Actual: one failure under load; the exact assertion is in WI-4.2’s ## Progress (.forge/plans/F36-a/F36-a-WI-4.2-leads-review-plan.md).
Cause: the test typed with userEvent on the real clock against a 50 ms wait, so a loaded machine could leave more than 50 ms between two keystrokes, and the box rightly sent the text typed so far before the whole word: WI-4.2’s Progress records the callback firing twice. The same file’s resetKey test, and the plan list’s “sends the search once, 300 ms after the last keystroke”, had the same exposure.
Fix: the send-once test and the resetKey test run on fake timers with fireEvent.change, each key 30 ms after the last, and check that nothing is sent 49 ms after the last key and the word once at 50 ms, which also proves that each key restarts the wait; the plan list’s test types with delay: null, so no timer can run between keys. No wait was lengthened.
Regression test: forge-et-pwa src/ui/SearchBox.test.tsx “owns its text and sends it once, delayMs after the last keystroke” no longer reads the machine’s clock. Forced load on macOS (one busy loop per core) did not reproduce the failure, so the cause was proved by holding back the late step: main’s test with 60 ms between keystrokes fails with onChange called 4 times (forge-et-pwa#158’s test evidence).
Resolved: 2026-11-18 — fixed in altrium-io/forge-et-pwa#158, as the Fix above describes. The fixed tests run on the test’s own clock, so they cannot fail this way on Windows either.
Lesson: L-125 — a test that assumed typing finished inside a debounce; BUG-367 is the same cause in the CRM.
BUG-462 — A control character in the lead triage’s or the lead research’s model answer makes their write fail
Affects: F11-d and F11-c (forge-et-be LeadTriageAnalyser.textOf, which feeds the stored summary and spam reason; LeadResearchFindings.kept, whose values go to the jsonb columns lead_research.company_findings and person_findings, and capture_research.company_findings through F36-c; and ExaSearchProvider.title, the search-result title stored beside each kept value) · Found in: backend-reviewer’s review of forge-et-be#459 (BUG-457’s fix), 2026-11-17, by reading the code · Severity: low — it needs the model to return a U+0000 · Phase: 1
Repro (from the code): the two lead classes take the model’s text with strip() only, and the provider keeps a page title as Exa returned it. PostgreSQL text refuses a U+0000, and jsonb refuses \u0000 (“unsupported Unicode escape sequence”), so the write fails and the message is retried, then dead-lettered.
Expected: control characters are removed before the text is stored, as PlainText.toMultiLine does in the lead ranking (BUG-457) and the capture ranking.
Actual: see the repro.
Fix: LeadTriageAnalyser.textOf and LeadResearchFindings.kept run PlainText.toMultiLine before strip() and the cut, and ExaSearchProvider.title cleans each page title once with PlainText.toOneLine (null when nothing is left), so every reader of a search hit gets the cleaned title. Capture research reuses LeadResearchFindings, so capture_research.company_findings is covered too.
Regression test: LeadTriageAnalyserSpec “BUG-462 — a summary and a spam reason lose their control characters before the cut…”, LeadResearchFindingsSpec “BUG-462 — a value loses its control characters before it is kept…” (with a 304-character value that pins the cleaning before the 300 cut), ExaSearchProviderSpec “BUG-462 — a title loses its control characters, and one left empty is null”. Each fails on main’s source and passes on the fix.
Resolved: 2026-11-17 — fixed in forge-et-be#459 (merge 60f4948e; ./gradlew build record on 0129e0f), with BUG-457.
Lesson: none — one-off: the same gap as BUG-457, fixed in every place in the lead and search code that stores model or search text.
BUG-461 — A lead the AI cannot rank yet shows the red Hot chip beside “Not enough detail to rank this lead yet.”, which reads as a contradiction
Affects: F11-g (forge-et-fe src/ui/components/LeadRankIndicators/LeadRankIndicators.tsx, the hot / cold pill; src/features/leadRanking/LeadRankingPanel.tsx, the lead page’s Priority card; the Leads table’s priority and hot / cold column) · Found in: F38-g-WI-4.4, the group 4.4 retest on dev, seen in passing during F11-b.1 (.forge/plans/F38-g/F38-g-WI-4.4-record.md R-4), 2026-11-17 · Severity: low — presentation; nothing is computed wrongly · Phase: 1
Repro: on dev (backend 627b29ee, CRM 70fc7709), create a lead with only a contact name (POST /api/v1/leads {leadSource: 'Website', contactName: '…'}, so no triage or ranking call starts) and open it: the Priority card shows a red flame pill “Hot” with “Buyer last wrote RETEST 2026-11-16 4.4 Below lead, on 2026-11-17 (Biseni’s screenshots in the WI-4.4 save folder).
Expected: a reader can tell at a glance that the lead has no priority yet; nothing on the card suggests it is urgent.
Actual: the red “Hot” pill, inside the card titled Priority, reads as a high priority while the same card says the lead could not be ranked. By the F11-g rule the pill is right — Hot means the buyer last wrote within et.leads.cold-after-days (30) days, and a hand-entered lead’s last message is its received date — but the pill sits where a priority would be, with the danger tint.
Fix: not decided; a UI decision for F11-g’s owner with the lead. Options: show hot / cold apart from the Priority card (for example in the meta grid, beside the received date); or label it as engagement (“Buyer engagement: Hot”); or leave it out of the card until a priority exists. The Leads table’s column needs the same decision.
Lesson: none yet — set when fixed.
BUG-456 — Five mailbox integration tests empty a connection’s rate limiter and assume it stays empty, though it refills a token a second
Affects: F20-c and F20-i (forge-et-be src/test/groovy/com/empireteas/mailbox/: EmailSendOutcomeIT.groovy:53 and :72, EmailSendIdempotencyIT.groovy:263, EmailSendFilesTokenIT.groovy:34, LiveInboxRefusalsIT.groovy:148, on main) · Found in: backend-reviewer’s review of forge-et-be#448 (BUG-443’s fix), 2026-11-16, by reading the code · Severity: low — no failure seen yet; the same race as BUG-443 · Phase: 1
Repro (from the code): each test drains the connection’s bucket (EmailSendBaseIT.drain, or its own loop), then sends a request that must find the bucket where the drain left it. MailboxIngestContextSpec pins the bucket at 20 tokens refilling at 60 a minute on real time (System::nanoTime), so a stall between the drain and the request refills a token a second.
Expected: each test passes whatever the machine’s load.
Actual (the window each test has): EmailSendOutcomeIT:72 (the token refresh’s own refusal; asserts wait in (4..6)) and LiveInboxRefusalsIT:148 (drains before each endpoint) need one token, so 1 s; EmailSendOutcomeIT:53 (a send’s four tokens) about 4 s; EmailSendIdempotencyIT:263 asserts availableTokens <= 1, about 2 s; EmailSendFilesTokenIT:34 polls for at least six tokens and then needs the seven-token send refused, so a 1 s stall after the poll turns the refusal into a 201.
Fix: BUG-443’s hold, generalised. MailboxIngestContextSpec.holdAt(long connectionId, double tokens = 0d) sets the bucket to tokens and moves its last refill reading an hour past the ticker, under the bucket’s own lock; the limiter refills nothing from a reading that has not moved forward. It replaces drain and drainAndHold in all six tests (EmailSendOutcomeIT three times, EmailSendIdempotencyIT, EmailSendFilesTokenIT with holdAt(id, 6d) and no polling wait, LiveInboxRefusalsIT, MessageAttachmentReadIT). The limiter and the ticker are autowired once in MailboxIngestContextSpec. ConnectionRateLimiterSpec tests the refill contract the hold relies on, on a ticker the spec moves.
Regression test: a two-second stall forced after the drain (the bucket’s reading moved back two seconds, uncommitted): on b7b2730d (drain) EmailSendOutcomeIT failed 1 of 15 and EmailSendFilesTokenIT 1 of 4; on b4165da1 (holdAt) both passed. The eight affected specs passed, 76 tests.
Resolved: 2026-11-17 — fixed in forge-et-be#448 (merge 589948e5; ./gradlew build record on eca184e), with BUG-443.
Lesson: none — one-off: the same cause as BUG-443, fixed in the same PR across every test that had it.
BUG-460 — Flaky test: PriceEmailDialog’s 51-contacts Also copy case (AC-27) times out at 20 seconds under load (macOS)
Affects: F26-d (forge-et-fe src/features/priceEmail/PriceEmailDialog.test.tsx, on main) · Found in: the F37-a wave-1 frontend review fixes (forge-et-fe#552), 2026-11-17 · Severity: low, a test-only timeout
Repro: run src/features/priceEmail (or the file alone) with Vitest on a loaded machine (load average 47–60 on 8 cores).
Expected: the test passes.
Actual: “51 contacts and 49 ticked” timed out at 20 s twice, in the folder run and alone; at normal load it passed, as did the PR’s test record on 03cd8255. The file was not changed by F37-a. The same case was BUG-220 on Windows, fixed on 2026-10-31 (forge-et-fe#367), so the fix’s margin does not hold at this load on macOS.
Fix: the 48 arranged ticks run in one act(), so the dialog renders once rather than 48 times; the tick handler updates from the previous set (setTicked((previous) => toggleAlsoCopy(…))), so every tick lands, and the test still checks all 48 ticked. The 49th tick and every click after it stay userEvent presses.
Regression test: the case itself, timed alone at load average about 28 on 8 cores: 3.4 s and 4.8 s before, 1.0 to 1.8 s after; src/features/priceEmail passed 26 files, 477 tests.
Resolved: 2026-11-17 — fixed in forge-et-fe#557 (merge 638ad5ce; Test record on 45f4158).
Lesson: none — one-off: the fifth case of the many-interactions-in-one-test cost the frontend reviewer’s dimension 26 already names; L-125 is about code budgets, not test timeouts.
BUG-459 — The capture app’s capture seam specs expect 17 members, but the backend has answered 20 since F36-b’s wave 2
Affects: F36-a (forge-et-pwa e2e/seam/capture-upload.seam.spec.ts:305 and capture-resend.seam.spec.ts:279, the exact member list of CapturedLeadResponse) · Found in: F36-c wave 1’s selected browser run on forge-et-pwa#146 (f40c707) against forge-et-be a22c5933, 2026-11-17; the same tests are on forge-et-pwa main (3bc35ec), and the same response is on forge-et-be main since forge-et-be#455 · Severity: medium — 12 seam tests fail against any backend that contains main, and they turn the nightly browser gate red; the product is not affected · Phase: 1
Repro: a backend on forge-et-be main (627b29ee or later), then ET_BACKEND_PORT=<port> corepack pnpm exec playwright test --project=seam e2e/seam/capture-upload.seam.spec.ts e2e/seam/capture-resend.seam.spec.ts: 9 of the upload spec’s tests and 3 of the resend spec’s fail.
Expected: the specs list the members the backend’s contract now has.
Actual: exactly the contract's members fails with three more members received: cardRead, detailsEditedAt and detailsEditedBy. F36-b’s wave 2 (forge-et-be#455, merged 2026-11-17 01:01 UTC as 627b29ee) added them to CapturedLeadResponse (F36-b WI-2.1 plan Decision 5; WI-2.2), and no capture-app change to these two specs reached main with it. This is the case .forge/lessons.md L-200’s note in BUG-324 describes: a backend PR that changes what a response holds breaks the frontend’s exact-shape specs.
Fix: F36-a’s wave 4 (forge-et-pwa#147) moves the member list into e2e/seam/support/review.ts with the three members, and its diff of both specs removes the old lists.
Resolved: 2026-11-17 — fixed in forge-et-pwa#147 (merge f429b05f). The capture app’s browser gate, run by hand on main at 23:50 (run-nightly.py --now --only pwa-gate, forge-et-pwa 1a79bb7, forge-et-be 6eaea2b), passed 6,714 tests with both specs in it, and closed forge-et-pwa#140.
Lesson: an instance of L-200, as its note in BUG-324 describes: a backend PR that changes what a response holds breaks the frontend’s exact-shape specs once it merges ahead of them.
BUG-458 — The empty plan’s upload states fail in WebKit since BUG-434’s fix: the list file is given but the upload never starts
Affects: F36-a (forge-et-pwa e2e/pageStates.ts plan-list-empty-reading and plan-list-empty-refused, in the size matrix’s size-matrix.spec.ts:134 and capture-layout.spec.ts:706/:788), WebKit sizes only · Found in: F36-c wave 1’s selected browser run on forge-et-pwa#146 (f40c707), 2026-11-17, and reproduced on main (3bc35ec) · Severity: medium — every branch that contains main shows them failed or flaky in the WebKit gate, and they will turn the nightly run red · Phase: 1
Repro: E2E_GATE=1 corepack pnpm exec playwright test --project=size-390x844-webkit --project=size-768x1024-webkit e2e/browser/size-matrix --grep "plan-list-empty-(refused|reading)" --retries=0 --repeat-each=3.
- On
main(3bc35ec): 27 of 36 failed. - On
fe5f3de(mainbefore forge-et-pwa#144), on45a9d9aand on9a79547: 36 of 36 passed. - On
1a473fc(“fix(BUG-434): hold the list’s campaign in state and end the hold on the next navigation”): it fails. - Chromium passes 18 of 18.
Expected: each state reaches its ready test id (exhibitors-empty-upload-reading, upload-refused).
Actual: element(s) not found after 5 s. The snapshot shows the empty panel Add the Show’s Exhibitors with Upload Exhibitor List idle, so the upload never started, or started and was reset.
Fix: not yet. Likely cause (not tested): 1a473fc holds the list’s campaign in state (listCampaign), and its mount effect adds a render after mount. The arrangement (e2e/support/arrange.ts arrangeUpload) gives the file from a MutationObserver as soon as exhibitors-empty-upload appears, which can be before the panel’s mount effects run. Under StrictMode’s double mount, UploadExhibitorListEntry’s unmount cleanup (cancelReading) aborts the read. The fix belongs either in the arrangement (give the file only after the panel’s effects have run) or in BUG-434’s code; BUG-434’s owners decide.
Resolved: 2026-11-17 — closed as wontfix (not reproduced), Tharinda decided. On a Mac, with this entry’s command (E2E_GATE=1 … --project=size-390x844-webkit --project=size-768x1024-webkit … --grep "plan-list-empty-(refused|reading)" --retries=0 --repeat-each=3), 36 of 36 passed on 3bc35ec (the commit the entry names, at load average 77 on 8 cores) and 36 of 36 on main at a23d478. A change to the arrangement that gives the file only after two frames and a task also passed 36 of 36, so with nothing to prove it was not shipped. A recurrence is filed as a new bug that links here, with the machine, its operating system and load, and the failing run’s output.
Lesson: L-162 (a failure is reproduced before it is fixed) — an instance where it was not reproduced, so nothing was changed.
BUG-457 — A control character in the lead-ranking model’s answer makes the ranking write fail
Affects: F11-g (forge-et-be src/main/java/com/empireteas/lead/service/LeadRankingAnalyser.java, textOf, line 90) · Found in: the review of F36-c wave 1 (altrium-io/forge-et-be#456), 2026-11-17, which found the same gap in F36-c’s new CaptureRankingAnalyser and fixed it there · Severity: low — it needs the model to return a NUL character, and the lead is retried, not lost · Phase: 1
Repro: not reproduced. By reading the code: textOf only trims the model’s reason and nextAction. A U+0000 in either makes PostgreSQL refuse the write (invalid byte sequence / character with value 0x00).
Expected: control characters are removed before the text is stored, as PlainText.toMultiLine does.
Actual: the write throws, and the message is retried and then goes to the lead-ranking DLQ.
Fix: LeadRankingAnalyser.textOf runs PlainText.toMultiLine before strip(), so the reason and next action lose their control characters (tab, line feed and carriage return are kept) before the 500/200 cut, as F36-c’s CaptureRankingAnalyser does. A text left empty by the cleaning is INVALID_OUTPUT, as a blank one was.
Regression test: LeadRankingAnalyserSpec “BUG-457 — a reason and a next action lose their control characters, so a U+0000 is never stored” (a next action of 199 characters and five NULs pins the cleaning before the cut), and a where-row of the blank-reason case with '\u0000\u0007'; both fail on main’s analyser.
Resolved: 2026-11-17 — fixed in forge-et-be#459 (merge 60f4948e; ./gradlew build record on 0129e0f).
Lesson: none — one-off: model text stored without cleaning; the review of this fix found the same gap elsewhere (BUG-462), fixed in the same PR.
BUG-453 — Flaky test: ContactsScreen “Clear filters stays cleared when the search debounce fires before the clear has rendered” under load (macOS)
Affects: F1-b (forge-et-fe src/features/accounts/ContactsScreen.test.tsx, on main) · Found in: F37-a-WI-1.6 Phase 1’s selected Vitest run on forge-et-fe 8335799b, 2026-11-16 · Severity: low · Phase: 1
Repro: run VITEST_MAX_WORKERS=4 pnpm test:selected on a loaded machine (load average about 14).
Expected: the test passes.
Actual: it failed once with Unable to find an element by: [data-testid="contacts-table"]; the file alone passed 56 of 56 and the next full selected run passed. The test (BUG-221’s “a write while an earlier navigation is pending”) races a debounce against a render.
Fix: the test waits up to 10 s for the contact table (findByTestId('contacts-table', {}, { timeout: 10_000 })) before it holds the loader and starts the race. The failure was the first render outlasting findBy’s 1 s default under load, before the part of the test that races, so the race itself is unchanged.
Regression test: the case itself; ContactsScreen.test.tsx passed 56 of 56.
Resolved: 2026-11-17 — fixed in forge-et-fe#553 (merge 029affc1; Test record and selected browser run on d4f3352).
Lesson: none — one-off: a first-render wait under load. L-270’s rule covers a held answer released before a 1 s wait, not the screen’s first render.
BUG-452 — forge-et-be’s test compile runs out of heap at 2.5 GB once F36’s and F37-a’s tests are on one branch
Affects: forge-et-be’s build (build.gradle, GroovyCompile memoryMaximumSize), on F37-a’s feature/F37-a-wave-1 · Found in: F37-a wave 1 Stage 11 Pass 1, 2026-11-16 · Severity: medium (no backend test can compile) · Phase: 1
Repro: on feature/F37-a-wave-1 (main at 88309db5 plus F37-a WI-1.1 to WI-1.3), run ./gradlew compileTestGroovy --rerun.
Expected: the test source set compiles.
Actual: Java heap space at memoryMaximumSize = '2560m', three runs in a row on macOS; with 3072m it compiles in 40 s. BUG-243 raised the same setting from 1536m for the same reason.
Fix: memoryMaximumSize raised to 3072m, with the comment updated, on feature/F37-a-wave-1 at 4c9d9144 (it ships with F37-a’s wave PR). Not checked on main alone; no nightly-red issue is open for it as of 2026-11-16.
Resolved: fixed by the 3 GB limit on F37-a’s wave branch, which ships in forge-et-be#446; that PR’s build record on de598998 compiled the whole test source set and ran 16,870 tests (2026-11-17).
Lesson: a second instance of BUG-243’s cause, the test source set outgrowing the forked Groovy compiler’s heap; no new lesson, as the limit is raised when the tests grow and the build fails loudly when it is short.
BUG-443 — Flaky test: MessageAttachmentReadIT “AC-7: with the limiter empty the read is 429” answers 200 in a long loaded full run (macOS)
Affects: F20-i (forge-et-be src/test/groovy/com/empireteas/mailbox/MessageAttachmentReadIT.groovy, on main) · Found in: F37-a-WI-1.1’s full suite on eae99a8b, 2026-11-16 · Severity: low · Phase: 1
Repro: run the full backend suite on a loaded machine (2 h 10 min, load average about 47). The test empties the per-connection rate limiter, then reads an attachment and expects 429.
Expected: 429 MAILBOX_RATE_LIMITED.
Actual: 200 once in 15,584 tests; the same test passed 8 of 8 when run alone. The limiter refills on real time, so a slow run can refill a token between emptying the bucket and the read.
Fix: the test now holds the bucket empty rather than draining it once. MailboxIngestContextSpec pins the bucket at 60 tokens a minute, so a stall of a second between the drain and the read put a token back. MailboxIngestContextSpec.holdAt(connectionId, tokens = 0) sets the bucket’s tokens and moves its last refill reading an hour past the ticker, under the bucket’s own lock, and the limiter refills nothing from a reading that has not moved forward. ConnectionRateLimiterSpec tests that contract on a ticker the spec moves. Five other tests had the same race (BUG-456) and use the same hold.
Regression test: AC-7 with a two-second stall forced after the hold (the bucket’s reading moved back two seconds, uncommitted): with the parent’s drain it fails on result.status.value() == 429 with 200 OK; with the hold, MessageAttachmentReadIT passed 8 of 8 and ConnectionRateLimiterSpec 28 of 28.
Resolved: 2026-11-17 — fixed in forge-et-be#448 (merge 589948e5; ./gradlew build record on eca184e).
Lesson: none — one-off: a test that empties a limiter refilling on real time; every such test in the mailbox suite now uses the hold (BUG-456).
BUG-455 — Flaky test: the account page’s follow-ups panel wait times out again in loaded browser runs, a recurrence of BUG-415 (macOS)
Affects: F1-d (forge-et-fe settleAccountPage in e2e/browser/accounts/detail-support.ts:2085, which waits for follow-ups-panel-loading to leave, and the size-matrix account-detail-follow-ups-* states) · Found in: F46 wave 1’s browser runs, 2026-11-16 · Severity: low — test failures, no product defect seen · Phase: 1
Repro: two runs on 2026-11-16, both against forge-et-be 6ad6440b (forge-et-be#440’s head) on an isolated stack, on an 8-core Mac with other sessions’ backend test JVMs running beside them:
- F46-WI-1.3’s orchestrator run,
E2E_GATE=1 pnpm exec playwright test --project=chromiumwith the fourteen lead-form specs, forge-et-fe9bdf15de:e2e/browser/custom-fields/values-flow.spec.ts:164(F34-c AC-18, the account custom fields) failed atdetail-support.ts:2085,follow-ups-panel-loadingstill present after 5 s; alone on the same head and database it passed (11 passed). - forge-et-fe#548’s
pnpm e2e:gate --selected(a full run on a wave branch), headf87a9823, started under the load limit (busy share 0.42 at the start;.browser-gate/result.jsonload: busy share average 1, load per core average 5.23, peak 15.44): the size-matrix statesaccount-detail-follow-ups-loading,account-detail-follow-ups-staleandaccount-detail-follow-ups-completed-erroratsize-320x640-chromiumfailed and passed on their retry, three of the run’s 26 flaky tests. Expected: the account page settles in a loaded run, or the load check stops the run first. Actual: the same wait BUG-415 recorded times out under load. BUG-415 was closedwontfixon 2026-11-13 because forced load did not reproduce it, and its record says a recurrence on a machine that passed the load check is filed as a new bug that links there; the wave PR’s run started under the limit, and the load rose during it. Fix: the lead chose the longer wait (2026-11-16).settleAccountPage’s own waits — the heading, the page’s loading and refreshing markers, the follow-ups and attachments panels’ first reads, and every tab’s loading marker — use oneexpect.configure({ timeout: 15_000 })in place of the 5 s default. The section helpers it calls (settleContactsand the others) and the lead page’s equivalent wait (e2e/browser/leads/support.ts,e2e/browser/size-matrix/arrange.ts) keep 5 s, so a recurrence there is a new bug that links here. Regression test: none can force the load; the size-matrixaccount-detail-follow-ups-*states and every test that callssettleAccountPagerun through the changed waits. Resolved: 2026-11-17 — fixed in forge-et-fe#553 (merge029affc1; Test record and selected browser run ond4f3352). Lesson: none — BUG-415’s (L-162) applies; this recurrence came from load that rose after the load check let the run start.
BUG-454 — Flaky test: the account list’s “vol” search test fails when a run’s random token contains “vol” (macOS)
Affects: F1-d (forge-et-fe e2e/browser/accounts/accounts-list.spec.ts:334, “AC-4 (smoke): typing vol sends q=vol after the debounce and lists Volga Tea Imports…”) and the Pipedrive handout tests’ account names (e2e/support/pipedriveExport.ts:113, E2E f32c <label> org <index> <token>) · Found in: forge-et-fe#548’s pnpm e2e:gate --selected (F46 wave 1, a full run), head f87a9823, 2026-11-16 · Severity: low — a test failure, no product defect · Phase: 1
Repro: in one full browser run, accounts-pipedrive-handout.spec.ts created its handout accounts named E2E f32c ann org 0001 mwj4volj1t8rtw8r0 onwards, whose token happens to contain vol. AC-4 then searched vol as the executive (account.view + account.all), and page 1 of the result was those accounts, so Volga Tea Imports was not on it. It failed on its run and its retry, and again run alone with retries off on the same head and database (the snapshot shows E2E f32c ann org 0021 mwj4volj1t8rtw8r0 and its neighbours as the list’s rows).
Expected: the test passes whatever other tests created earlier in the run.
Actual: it fails whenever a run’s random token contains vol (a three-letter run inside a random base-36 token, so roughly one run in a few hundred, and every later run against the same database).
Fix: the test opens the list sorted by name, descending (/accounts?sort=name,desc), and checks that the search keeps the sort. Every generated name starts E2E, so it is listed below Volga Tea Imports in that order, and only names that sort above Volga could push it off page 1; no test creates such names. The search still types vol, as F1-b AC-4 states. The handout tests’ token naming (e2e/support/pipedriveExport.ts:113) is left as it is, because the sort makes it harmless.
Regression test: the AC-4 test itself, in the PR’s browser run.
Resolved: 2026-11-17 — fixed in forge-et-fe#553 (merge 029affc1; Test record and selected browser run on d4f3352).
Lesson: L-306 (a browser test that searches seeded data reads a result no other test’s records can crowd out).
BUG-451 — Flaky test: the sharing panel’s failed re-read test times out in a full Vitest run (Windows)
Affects: F5-g (forge-et-fe src/features/sharing/SharingPanel.test.tsx:358, “SharingPanel — the data states (F5-g AC-5) › keeps the list after a failed re-read, marked as shown from its time, with Try Again”; it said Retry until forge-et-fe#554, BUG-447) · Found in: forge-et-fe#549’s first local test record (pnpm test:record, every test, as the PR changes src/lib), Windows, 2026-11-16 · Severity: low · Phase: 1
Repro: pnpm test:record on forge-et-fe bc2042b0: 722 files, 15,223 tests; this test timed out on its run and its retry. The next pnpm test:record on the same commit passed every test, and pnpm exec vitest --run src/features/sharing/SharingPanel.test.tsx passed alone.
Expected: the test passes every run.
Actual: “Test timed out in 20000ms” at line 358, twice (the run and its retry). The test fakes Date only (vi.useFakeTimers({ toFake: ['Date'] })).
Fix: none — not reproduced, so the test is kept unchanged; closing it wontfix as an overloaded machine waits on the lead (.forge/test-strategy.md § Flaky Tests, triage step 4). Triaged 2026-11-18 on the same 16-core Windows machine, retries off, with 5 busy loops on top of about 70% load from other work: 10 of 10 runs of the file passed (40 of 40 tests each, this test 0.35–1.6 s) and 10 of 10 runs of src/features/sharing passed (slowest 4.5 s, against the 20 s limit). Holding back each read in turn shows that only one step can give “Test timed out”: the act around invalidateQueries (and the three user.click calls), which have no limit of their own. Every other wait is a 1 s findBy or waitFor, and holding back the Try Again read gives a 1.5 s assertion failure, not a timeout. The stub answers the re-read at once, so nothing in the test or the product can hold it for 20 s: the Vitest worker got almost no CPU for over 40 s (the run and its retry). BUG-450 timed out at 20 s in the same record, and the next full record on the same commit passed (machine load 0.94 average, 1.00 peak). The failing record kept no load figures, so there is no list of the busiest processes to show other work.
Resolved: 2026-11-18 — closed wontfix as an overloaded machine (.forge/test-strategy.md § Flaky Tests, triage step 4); decided by Tharinda (lead). Not fixed because nothing in the test or the product can produce the error: only a re-read inside act that never finishes gives “Test timed out”, the stub answers it at once, and every reproduction under forced load passed (10 of 10 file runs, 10 of 10 folder runs, retries off). The evidence that other work starved the run is a second, unrelated test (BUG-450) timing out at 20 s in the same record, on its run and its retry; the record kept no busiest-process list.
Lesson: none — one-off: an overloaded machine, which the test records’ load check (0.65 busy) now refuses at the start. The same record’s other timeout taught L-322.
BUG-450 — Flaky test: the account filters’ Customer type tick test times out in a full Vitest run (Windows)
Affects: F1-a (forge-et-fe src/features/accounts/AccountsFilters.test.tsx:238, “AccountsFilters › reports each tick”) · Found in: forge-et-fe#549’s first local test record, Windows, 2026-11-16 · Severity: low · Phase: 1
Repro: as BUG-451: one full pnpm test:record on forge-et-fe bc2042b0 failed it; the next full record on the same commit and pnpm exec vitest --run src/features/accounts/AccountsFilters.test.tsx alone passed.
Expected: the test passes every run.
Actual: the first run “Test timed out in 20000ms” at line 238; the retry failed with “expected last vi.fn() call to have been called with [ { customerTypes: [ ‘Importer’ ] } ]”.
Fix: none — the timeout did not reproduce, so the test is kept unchanged; closing it wontfix as an overloaded machine waits on the lead (.forge/test-strategy.md § Flaky Tests, triage step 4). Triaged 2026-11-18 on the same 16-core Windows machine, retries off: 10 of 10 runs of the file passed with the CPU already 88–100% busy from another session’s Gradle check (this test 2.9–5.3 s), 10 of 10 passed with 5 busy loops added (2.1–7.0 s), and the test passed in 7.1 s in one run of the whole src/features/accounts folder (53 files) with the loops. Most of its time is the Country popover (about 250 boxes, 1.9 s of 2.7 s idle). The retry’s wrong-last-call failure is not a second defect: Vitest does not stop a timed-out attempt, which kept sending keys and clicks to the retry’s screen and closed its popover; holding the first attempt back before its Escape reproduced that failure every time (L-322). If the timeout recurs, “reports each tick” can drop its Country leg, which “reports both countries when two are ticked” already proves.
Resolved: 2026-11-18 — closed wontfix as an overloaded machine (.forge/test-strategy.md § Flaky Tests, triage step 4); decided by Tharinda (lead). Not fixed because the 20 s timeout did not reproduce under any forced load (its slowest run was 7.1 s), and the retry’s failure follows from the timed-out attempt. The evidence that other work starved the run is a second, unrelated test (BUG-451) timing out at 20 s in the same record, on its run and its retry; the record kept no busiest-process list.
Lesson: L-322 — new: a Vitest test that times out keeps running, so its retry’s failure message is not a second defect.
BUG-449 — Flaky test: the Scheduled list’s Edit and Send on a failed forward does not find the To field in a loaded Vitest run (Windows)
Affects: F20-g (forge-et-fe src/features/email/scheduled/ScheduledEmailsRoute.recover.test.tsx:173, “Edit and Send on a FAILED forward (AC-16) › opens the dialog with every stored field, the file ready and the original removed; Send sends, dismisses once and the list loses the row”) · Found in: forge-et-fe#549’s local Vitest run on Windows, 2026-11-16 · Severity: low · Phase: 1
Repro: pnpm exec vitest --run src/features/accounts/AccountListScreen src/features/leads/LeadsScreen src/features/email/scheduled on Windows (16 files in parallel) failed this one test once; pnpm exec vitest --run src/features/email/scheduled/ScheduledEmailsRoute.recover.test.tsx then passed twice (5 of 5 each).
Expected: the test passes every run.
Actual: at line 176, screen.findByRole('textbox', { name: /^To/ }, { timeout: 5000 }) found no To field; the test took 5.8 s. The change under test (forge-et-fe#549) only moved useMountedRef’s import path, and the hook’s body is unchanged.
Fix: in the test. The compose dialog (ComposeEmailLayer, with the TipTap editor) loads through import() the first time ComposeEmailProvider opens it, and in Vitest that first load transforms and evaluates the dialog’s whole module tree. The first case to press Edit and Send paid it inside its own 5 s wait for the To field: timing logs showed the load took 1.4–2.0 s of a 2.6–3.0 s wait on a quieter machine, 2.1–10.3 s under forced load, and both failures came at 6.6 s and 10.3 s. Every later open took 0–11 ms. The test file now runs await loadComposeEmailLayer() once at the top level, which has no time limit (a beforeAll has Vitest’s 10 s), so each case waits only for the row’s read and the dialog’s render. No timeout changed and no product code changed (forge-et-fe commit af784c76). Under forced load (5 busy loops on 16 cores already at 100% from other work, retries off): before, 2 of 10 runs of the file failed and 2 of 3 runs of the original 16-file command failed this test; after, 10 of 10 and 3 of 3 passed (501 tests each), and the press-to-To time fell to 47–139 ms. Not changed: src/app/aiDraftFlow.test.tsx handles the same first load with a 10 s findDialog wait (.forge/owed.md).
Resolved: 2026-11-18 — fixed by forge-et-fe#583 (test only). Reproduced under forced load before the fix (2 of 10 file runs and 2 of 3 runs of the original 16-file command failed it); none failed after (10 of 10 and 3 of 3).
Lesson: L-116 — another instance: a test that opens a lazily loaded dialog pays its module’s first load inside the first case’s wait; added to L-116’s sources.
BUG-448 — The Publish the Round dialog says standards not in the round keep their current price, when those prices end as the sale starts
Affects: F26-f (forge-et-fe PublishRoundDialog.tsx) · Found in: the weekly-prices audit of 2026-11-16 · Severity: medium — the Tea Room may publish without carrying prices forward, believing the missing standards stay priced, and those standards then have no price from the sale’s start · Phase: 1
Repro: open Publish N Prices on a round with standards not in it.
Expected: the dialog says those standards keep their current price only until the sale starts, as the Standards not in this round dialog and the user guide say, and points at Carry Forward All.
Actual: it says they “keep their current price”, with no end.
Fix: the sentence names the end: their current prices end when the sale starts unless they are carried forward. The Remove Draft dialog said the same wrong thing (“keeps its current price until a new one is published”) and gets the same correction. PublishRoundDialog.test.tsx and RemoveDraftDialog.test.tsx pin the sentences; before the fix they fail on main’s wording.
Resolved: 2026-11-16 — fixed in altrium-io/forge-et-fe#546, as the Fix above describes; crm-current draws both sentences.
Lesson: none — one-off: two dialogs said what happens to a price in their own words, and one was wrong; the user guide now states the rule once, and the Standards not in this round dialog already had the right sentence.
BUG-447 — The CRM says both Retry and Try Again for trying a failed action again
Affects: F26-f, and screens across the CRM (forge-et-fe: Retry in AccountDetailPage.tsx, AccountListScreen.tsx and others; Try Again in AccountContactsPanel.tsx, MessageView.tsx, LoginRoute.tsx, the Pricing screens and others) · Found in: the weekly-prices audit of 2026-11-16, which first read it as a Pricing-only difference · Severity: low — wording only · Phase: 1
Repro: make a read fail on the account page (Retry), then on the account’s contacts panel or the Pricing page (Try Again).
Expected: one word for the same action on every screen (design-system §9, One word for one action).
Actual: each screen chose its own; both words are in wide use.
Fix: the lead chose Try Again on every CRM screen on 2026-11-16, because the CRM already used it most (design-system Rev 132, §9 One word for one action). altrium-io/forge-et-fe#554 changes every Retry label, test and test id that reads it, in one PR. Order with F46 (decided 2026-11-17): F46 wave 2 splits six files #554 also edits (RecordAttachmentsPanel, FollowUpEditBody, AccountListScreen, RecordNotesSection, LeadNotesSection, TemplateEditorForm), so wave 2 merges first. #554 stays a draft until then; then main is merged in, the Try Again edits are made again in whichever split file now holds each label (the F46 plans’ ## Notes record where each block moved), and #554 gets its test record, selected browser run and review record before it merges.
Regression test: src/app/tryAgainLabel.test.ts, a scan that fails any code under src/ or e2e/ using the old word outside a comment (it fails on main), and the component and browser tests #554 updates to find the button by the name Try Again.
Resolved: 2026-11-18 — fixed in altrium-io/forge-et-fe#554, as the Fix above describes: after F46 wave 2 merged, main was merged in and Try Again applied again in the split files (AccountListRegion, AttachmentsRegion, useRecordAttachmentsFocus, LeadNotesRegion, NotesRegion) and in the price-broadcast screens F37-a added meanwhile. The one comment F37-a added to crm-current with the old word says Try Again too.
Lesson: none — one-off: the CRM had no rule for this word; design-system Rev 133’s One word for one action adds the rule, and Rev 132 fills in the word.
BUG-446 — Adding an auction sale has three names: Add a Sale, New Sale and Add Sale
Affects: F26-g, F26-c (forge-et-fe AuctionSalesTab.tsx, PricingWarnings.tsx, SaleStep.tsx, NewSaleDialog.tsx) · Found in: the weekly-prices audit of 2026-11-16 · Severity: low — wording only · Phase: 1
Repro: the Auction Sales tab and the no-next-sale warning say Add a Sale; the import page says New Sale; the dialog they open is titled New Sale and its button says Add Sale.
Expected: the CRM’s wording for a create flow, as Leads, Accounts, Roles and Email templates have it: the opening button and the dialog title New Sale, the dialog’s button Create Sale.
Actual: as in Repro.
Fix: New Sale on the tab, the warning and the import page, title New Sale, button Create Sale; the user guide and the prototype say the same.
Resolved: 2026-11-16 — fixed in altrium-io/forge-et-fe#546, as the Fix above describes; the user guide’s Step 1 and crm-current say the same. Regression tests: AuctionSalesTab.test.tsx, PricingWarnings.test.tsx, SaleStep.test.tsx and NewSaleDialog.test.tsx press New Sale and Create Sale by name, which fail on main.
Lesson: none — one-off: the CRM had no rule for naming a create flow; design-system Rev 133 (§9, One word for one action) adds it, and BUG-445 has the same cause.
BUG-445 — The New Sale dialog names its date fields Start Date and End Date, while the table and the Edit Sale dialog say Starts and Ends
Affects: F26-g, F26-c (forge-et-fe newSaleForm.ts) · Found in: the weekly-prices audit of 2026-11-16 · Severity: low — wording only · Phase: 1
Repro: open New Sale, then Edit Sale on a sale, and compare both with the Auction Sales table’s headers.
Expected: one name per field everywhere (design-system §9: a column header and the label of the same field read the same): Starts and Ends.
Actual: Start Date and End Date in the New Sale dialog only.
Fix: newSaleForm.ts labels Starts and Ends; the tests that find the fields by label change with them.
Resolved: 2026-11-16 — fixed in altrium-io/forge-et-fe#546, as the Fix above describes. Regression test: NewSaleDialog.test.tsx finds the fields by /^Starts/ and /^Ends/, which fail on main’s labels.
Lesson: none — one-off: no design-system rule asked a field to keep one label across its forms and its column header; design-system Rev 133 (§9) adds it and frontend-reviewer checks it.
BUG-444 — The Auction Sales table shows Edit Sale and Delete Sale as buttons in each row instead of the row’s ⋯ menu
Affects: F26-g (FR-15; forge-et-fe AuctionSalesTable.tsx) · Found in: the lead’s review of the Pricing page, 2026-11-16, confirmed by that day’s weekly-prices audit · Severity: low — every action works, but the table looks and behaves unlike every other CRM list table · Phase: 1
Repro: open Pricing → Auction Sales as a buying_standard.manage holder.
Expected: each row ends with the ⋯ menu every CRM list table has (RowActionsMenu, design-system §5 Table and §6 List screen, Rev 70), holding Edit Sale and Delete Sale, Delete marked destructive, in the pinned menu column (Rev 94).
Actual: an Actions column with an outline Edit Sale button and a red-outline Delete Sale button in each row.
Why it happened: design-system Rev 94 listed “the auction-sales table … keep plain buttons in their last column” as an exception, written when that column held one Edit dates button. BUG-422 (forge-et-fe#533) added Delete Sale beside it under that exception, and neither the plan nor frontend-reviewer checked the exception against §6’s rule, which has none.
Fix: the row’s actions move into RowActionsMenu (named “Actions for DATA_TABLE_MENU_CELL and DATA_TABLE_HEAD_MENU, each item shown by the same server flags as before; the table joins the pinned ⋯ tables in pinnedFirstColumn.test.tsx. A new source test fails any file outside a short allow-list that uses the plain DATA_TABLE_ACTIONS_CELL, so a list table cannot put buttons in its rows again unnoticed. The design system drops Rev 94’s exception for this table and states the rule; frontend-reviewer gets a check; crm-current draws the menu.
Resolved: 2026-11-16 — fixed in altrium-io/forge-et-fe#546, as the Fix above describes (design-system Rev 133; crm-current checked in the browser). Regression tests: AuctionSalesTable.test.tsx opens the row’s ⋯ menu and chooses Edit Sale and Delete Sale, and pinnedFirstColumn.test.tsx § no list table puts buttons in its rows fails on main’s AuctionSalesTable.tsx, which used DATA_TABLE_ACTIONS_CELL.
Lesson: L-304 — an exception written into a UI rule outlives what it excused; added for this bug.
BUG-442 — Pricing wording slips: the Weekly prices tab in one message, two table names cased differently, and Quoted per deal beside Priced per quote
Affects: F26-f, F26-g (forge-et-fe editSaleForm.ts, RoundTable.tsx, AuctionSalesTable.tsx, roundFormat.ts) · Found in: the weekly-prices audit of 2026-11-16 · Severity: low — wording only · Phase: 1
Repro: (1) edit a sale with a published price and try to change its start: the message says “the Weekly prices tab”. (2) Read the two tables’ accessible names: “Weekly prices table” and “Auction Sales Table”. (3) A private-label draft that cannot be published reads Quoted per deal; the same state reads Priced per quote on the account page, in the price email and in design-system §5’s price-state row.
Expected: the tab’s own name, Weekly Prices; the two table names cased alike; one name, Priced per quote, for a private-label standard’s price state.
Actual: as in Repro.
Fix: “Weekly Prices tab” in ONLY_END_CAN_CHANGE; the round’s blocked reason and the import’s private-label group read Priced per quote. Tests that pin the old strings change with them. Still to do, in forge-et-be: two server messages say the same thing in words the screen shows as they are, “a private-label standard, quoted per deal, which is not in the weekly round” (DraftPriceService.java:76, PriceCorrectionService.java:172); they become “priced per quote”, with the backend tests that pin them. BUG-442 stays in progress until that lands. The table names (2) are left as they are: forge-et-fe 1c090aed recased them on purpose, and the CRM’s tables are named both ways, so they need a CRM-wide choice, not a Pricing fix.
Regression test: the screens’ tests changed in altrium-io/forge-et-fe#546; DraftPriceWriteIT and PriceCorrectionIT’s private-label rows now assert “priced per quote”, which main’s messages do not contain; forge-et-fe’s e2e/seam/price-round.seam.spec.ts expects the new sentence.
Resolved: 2026-11-18 — fixed in altrium-io/forge-et-fe#546 (the screens), altrium-io/forge-et-be#465 (the two server messages and ImportReason’s Javadoc) and altrium-io/forge-et-fe#570 (the seam test that pins one of them), as the Fix above describes; the two last merged back to back, so the nightly seam run never met one without the other. The table names (2) stay as the Fix says.
Lesson: none — one-off: the same cause as BUG-445 and BUG-446, wording chosen screen by screen with no §9 rule for one name per state; design-system Rev 133 adds the rule, so no further lesson.
BUG-441 — Flaky test: the capture app’s lost-import-answer test does not see the Not confirmed message (macOS)
Affects: F36-a (forge-et-pwa e2e/browser/plan/upload.spec.ts:245, “an import whose answer is lost says Not confirmed, offers Import again, and the list is read on reconnect”, Chromium; wave 3 code) · Found in: a hand run of the capture-app browser gate on main at 2037aa9, macOS, 2026-11-16 (failed and was retried there; machine load 0.75 average, 0.96 peak) · Severity: low · Phase: 1
Repro: playwright test e2e/browser/plan/upload.spec.ts:245 --project=chromium --retries=0 --repeat-each=3 --workers=1 on macOS failed 1 of 3, with nothing else heavy running (load average 3.2 on 8 cores).
Expected: the test passes every run.
Actual: at line 260, getByTestId('check-exhibitor-list-problem') was not found within 5 s, so the Not confirmed — check before trying again… text was never read.
Fix: a test defect. The import button turns busy as soon as the click starts the request, before the request reaches Playwright’s route, and the test called abortHeld() as soon as it saw the busy button. Under load the abort ran first, found nothing held and did nothing, so the import waited for ever and the Not confirmed message never came. The test now polls the mock’s own requests list until the import is there before aborting; the mock records and holds a request in the same synchronous step.
Regression test: the spec itself, playwright test e2e/browser/plan/upload.spec.ts:245 --project=chromium --retries=0 --repeat-each=10 --workers=1 at load average 40 to 90 on 8 cores: before the fix 1 of 10 failed at the Not confirmed text; after it 10 of 10 passed.
Resolved: 2026-11-17 — fixed in forge-et-pwa#144 (merge 3bc35ec2; Test record and a full browser gate, 4,624 passed, on b7367e0).
Lesson: L-307 (a test that releases or aborts a held request waits until the mock holds it).
BUG-440 — The capture app’s scrollbar-space test cannot pass on macOS, whose scrollbars take no space (macOS)
Affects: F36-a (forge-et-pwa e2e/browser/size-matrix/capture-scrollbar.spec.ts:19, the 1280 × 800 and 320 × 640 Chromium size projects; added 2026-11-15 in 3db7724) · Found in: a hand run of the capture-app browser gate on main at 2037aa9, macOS, 2026-11-16 · Severity: low — a test defect; it blocks every macOS browser gate · Phase: 1
Repro: playwright test e2e/browser/size-matrix/capture-scrollbar.spec.ts:19 --project=size-1280x800-chromium --project=size-320x640-chromium --retries=0 --repeat-each=2 --workers=1 on macOS: 4 of 4 fail.
Expected: the test checks that the column’s content lines up with the header row when <main> has a scrollbar that takes space, on every machine the gate runs on.
Actual: line 51 asserts measured.scrollbar > 0 (“the scrollbar takes space in <main>”) for desktop Chromium, but macOS Chromium draws overlay scrollbars of 0px, so it fails before checking alignment.
Fix: in the desktop Chromium projects the spec styles <main>'s scrollbar (main::-webkit-scrollbar { width: 15px }, injected with addStyleTag), which Chromium draws as a classic scrollbar on every platform; the alignment check is unchanged.
Regression test: the spec itself: before the fix it failed 4 of 4 alone on macOS; after, 32 of 32 across the size projects.
Resolved: 2026-11-16 — fixed in forge-et-pwa#141 (merge cf2e694f).
Lesson: none — one-off, together with BUG-439: browser tests written and run only on Windows met macOS behaviour for the first time on the Mac nightly run; the nightly run is what catches this.
BUG-439 — The capture app’s column-match keyboard test cannot pass on macOS: ArrowDown opens a select’s list instead of changing it (macOS)
Affects: F36-a (forge-et-pwa e2e/browser/plan/upload-a11y.spec.ts:74, “each column match is changed by the keyboard alone”, Chromium; added 2026-11-15 in 0db275a) · Found in: a hand run of the capture-app browser gate on main at 2037aa9, macOS, 2026-11-16 · Severity: low — a test defect; it blocks every macOS browser gate · Phase: 1
Repro: playwright test e2e/browser/plan/upload-a11y.spec.ts:74 --project=chromium --retries=0 --repeat-each=3 --workers=1 on macOS: 3 of 3 fail.
Expected: the test proves each column match can be changed by keyboard alone, on every machine the gate runs on.
Actual: line 82: after focusing the native <select> (match-companyName-select) and pressing ArrowDown, its value is still "0". On macOS, Chromium’s ArrowDown on a closed <select> opens its list rather than choosing the next option; on Windows it changes the value, which is where the test passed.
Fix: the test types the first letter of an option that starts with another letter (a focused native select’s type-ahead, the same on every platform), then checks that the value changed and that the chosen option starts with that letter. ArrowDown, ArrowDown, Enter also failed on macOS.
Regression test: the spec itself: before the fix it failed 3 of 3 alone on macOS; after, 3 of 3 passed.
Resolved: 2026-11-16 — fixed in forge-et-pwa#141 (merge cf2e694f).
Lesson: none — one-off; the review’s OS-dependent-keys check (type-ahead for a closed native select, L-143) already names this approach.
BUG-438 — The backend on main does not start: two migrations are numbered V72
Affects: F36-a (forge-et-be V72__capture_plan.sql, F36-a wave 3, forge-et-be#435) and the F26-d follow-up forge-et-be#437 (V72__auction_sale_reference_indexes.sql) · Found in: a hand run of the capture app’s browser gate (run-nightly.py --now --only pwa-static,pwa-gate), 2026-11-16 · Severity: high — every backend start on main fails, the dev deploy of F36-a wave 3 failed, and every browser gate that starts the backend fails · Phase: 1
Repro: start forge-et-be at fde90b9a (./gradlew bootRun, or any spec that migrates a database).
Expected: Flyway applies the migrations and the backend starts.
Actual: FlywayException: Found more than one migration with version 72. forge-et-be#437 merged at 00:31 with V72__auction_sale_reference_indexes.sql, and #435 merged at 11:25 with V72__capture_plan.sql. Each PR’s record ran on its own head, the two files have different names, and main does not require a branch to be up to date, so both merged cleanly. Dev’s 11:25 deploy failed four times with the same error, and ECS kept the task started at 00:33 (backend:139, #437’s image), so F36-a wave 3 is not live on dev. Tag v0.1.0-rc1 (f392e8ec) holds only the auction-sale V72.
Fix: rename the one never applied in a shared database: V72__capture_plan.sql becomes V73__capture_plan.sql, with its header comment and the docs and test comment that name it (forge-et-be#440). No SQL changes. docs/migrations.md stops saying a collision “is caught in CI” and adds a check, just before merging, that the number is still free on main. A laptop database that applied the capture plan as V72 needs a reset; flyway repair would mark the auction-sale V72 applied without its indexes. F36-b WI-2.1’s approved plan had claimed V73 for captured_lead_card_read, so it re-chooses its number (owed.md).
Regression test: none added: every spec that migrates a database (GrantSchemaSpec, the migration ITs) fails on fde90b9a with Flyway’s duplicate-version error and passes on the fix.
Resolved: 2026-11-16 — fixed in forge-et-be#440 (merge 0bb55aca; full-suite build record on 74e5747). Dev’s deploy of it (backend:141) applied one migration and started; the rollout completed with no failed task.
Renumbered: filed as BUG-437, which BUG-437 (All mail search) took first on main; forge-et-be#440, its commits and docs/migrations.md still say BUG-437.
Lesson: L-120 — a second instance (#437 took V72 while F36-a wave 3’s branch held it); its new bullet is the check just before merging.
BUG-413 — Flaky test: the plan list’s booth-drawer archive does not find the drawer in a loaded full Vitest run (Windows)
Affects: F36-a (forge-et-pwa src/features/plan/exhibitors/PlanList.test.tsx, “focus when a press removes its control (L-133) › an archive from the booth drawer: when the re-read drops the card, focus moves to Booth Plan”; wave 3 code) · Found in: the F36-d wave 1 merge of main into feature/F36-d-wave-1 (forge-et-pwa a4c709b), its whole Vitest run, Windows 11, 2026-11-16 · Severity: low · Phase: 1
Repro: not reproduced alone. VITEST_MAX_WORKERS=4 pnpm test on a4c709b (140 files, 402 s, after the backend’s build record had held the machine) failed it, and the file took 15.9 s; the file run alone on the same tree passed 42 of 42, three times. The merge changes nothing under src/features/plan/ (git diff origin/main HEAD -- src/features/plan/ is empty).
Expected: the test passes whatever else the machine is running.
Actual: Unable to find role="dialog" and name "Westmoor Tea Importers" — the booth drawer had not opened within the query’s wait.
Cause: PlanList loads BoothDrawer through lazy(), and this test is the first in its file to open the drawer, so it paid the drawer module tree’s first transform inside findByRole’s 1-second wait. A loaded whole-suite run outlasted it.
Fix: PlanList.test.tsx imports ../booth/BoothDrawer once in beforeAll, as CaptureOutboxProvider.test.tsx does for its send pass (BUG-432), so the drawer’s import() resolves from the module cache. No wait was lengthened (.forge/test-strategy.md § Flaky Tests).
Regression test: the test itself. Forced load on macOS did not reproduce the failure, so the cause was proved by holding back the late step: with BoothDrawer’s module load delayed 1.5 s, main’s file fails with the reported Unable to find role="dialog" and name "Westmoor Tea Importers", and this fix’s file passes (forge-et-pwa#158’s test evidence).
Resolved: 2026-11-18 — fixed in altrium-io/forge-et-pwa#158, as the Fix above describes. The preload takes the first load out of every test, so it holds on Windows too.
Lesson: L-116 — a component test that renders a lazy() screen loads its module before the test, never with a longer wait.
BUG-437 — A search in All mail returns nothing on dev, although the same mail is listed in All mail and found by a search of the Inbox
Affects: F20-i (forge-et-be GET /api/v1/mailbox/all-mail/messages?search=, MailboxAllMailController, AllMailFolders) · Found in: F38-g WI-4.2’s second pass on dev (F20-b.4), 2026-11-16 · Severity: medium — the All mail search finds nothing; the All mail list and the Inbox search work · Phase: 1
Repro: on dev, as CRM03 (RETEST 4.2 email user, mailbox CRM03@empire.lk connected), with the messages RETEST 2026-11-16 4.2 inbound (from CRM01, 13:22) and RETEST 2026-11-10 4.2 inbound in the mailbox: open /email, choose All mail (the list shows both), type inbound in the search box, and wait about 30 seconds and press Refresh. No row appears; crm01 gives the same. In Inbox the same two words list the messages. The request GET /api/v1/mailbox/all-mail/messages?search=inbound answers 200 {"items":[],"nextCursor":null,"searchCapped":false,"stillPreparing":false}, so the server reports an empty search that is not still filling.
Expected: a search of All mail lists the mail whose subject or participants match, as the Inbox search does.
Actual: an empty result, with stillPreparing false.
Also seen, 2026-11-12 (the same walk’s first run of F20-b.4, as CRM03; a duplicate filing was closed in its favour): GET /api/v1/mailbox/all-mail/messages?search=inbound → 200, items: [], stillPreparing: false, while GET /api/v1/mailbox/folders/<Inbox id>/messages?search=inbound → 200 and lists “RETEST 2026-11-10 4.2 inbound”. No WARN or ERROR in /ecs/dev/backend for the call. Likely cause, not proved: Graph’s $search returns an empty page for the messages of a mailSearchFolder. Owner to confirm: F20-i’s owner, Tharinda.
Fix: not decided. Start from AllMailFolders and MailboxProvider.searchFolderFill: whether the search folder on this connection was created over the right source folders (the connection was released by its owner’s disconnect and reconnected on 2026-11-10, and its folder recorded earlier), whether its fill had finished, and what /ecs/dev/backend logs for the call. /ecs/dev/backend showed Mailbox Read Throttled … [operation=listFolders] and Mailbox Read Unavailable … [kind=UNREACHABLE] lines for a connection in the same hour, which may be unrelated.
Regression test: a search of All mail over a seeded mailbox lists a message the Inbox search finds.
Lesson: pending — decided with the fix.
BUG-436 — Flaky test: the capture app’s camera-refused browser test does not see the camera-off panel in a loaded full browser run (Windows)
Affects: F36-a (forge-et-pwa e2e/browser/capture/capture.spec.ts line 206, “the camera off (F36-a AC-11, AC-13) › refused: says so, and Choose a Photo sends the photo shrunk to 2,048 × 1,536”, Chromium; wave 2 code) · Found in: forge-et-pwa#116’s selected browser run (a full run) on 1edd349, Windows 11, 2026-11-16 · Severity: low · Phase: 1
Repro: not reproduced alone. ET_E2E_WORKERS=4 pnpm e2e:gate --selected failed it on its first attempt and its retry (8.5 s and 9.4 s); the same test run alone with retries off (playwright test e2e/browser/capture/capture.spec.ts:206 --project=chromium --retries=0 --repeat-each=5 --workers=1) passed 5 of 5 at about 1.4 s each. The same full run passed it on c5ff7e4 an hour earlier (1,389 passed), and 1edd349 changes only a unit test and a CSS comment.
Expected: the test passes whatever else the machine is running.
Actual: the first attempt timed out after 5 s waiting for getByTestId('capture-camera-off') to contain The camera is off for this site.; the retry got past that and timed out waiting for the Choose a Photo button. Neither element was found, so the screen had not reached the refused-camera state within the assertion’s 5 s.
Fix: in the test, not a longer wait. The spec installed choosePictureWhenShown, an init script that gives the hidden file input a picture as soon as the input appears. The input renders only in the camera-off state (src/features/capture/CaptureRoute.tsx), so the photo was chosen the moment that state rendered, and the photo’s preview then replaced the message and Choose a Photo. A late poll found neither. The spec now asserts the message and the button first, then presses Choose a Photo through a new helper, choosePicture (e2e/support/arrange.ts), which answers the file chooser with the same 4,000 × 3,000 picture. The other users of choosePictureWhenShown (capture-timing.spec.ts, the capture-card-taken page state) assert only the state after the photo, so they do not race.
Resolved: 2026-11-18 — fixed by forge-et-pwa#148 (test only). With E2E_GATE=1, Chromium, retries off and 8 busy loops beside the run: before the fix, capture.spec.ts:206 --repeat-each=20 --workers=4 failed 8 of 20, all at capture-camera-off; after it, 20 of 20 and then 40 of 40 passed (--repeat-each=40 --workers=4).
Lesson: L-321 — new: an action that fires as soon as an element appears races any check of the state before that action.
BUG-435 — A price email is sent and then cannot be recorded when its sale is deleted or replaced while the rep writes it
Affects: F26-d (FR-U4; forge-et-fe AccountPriceEmailAction.tsx), through F20-c’s compose dialog (FR-O1) · Found in: forge-et-be#430’s review (BUG-422), 2026-11-14, and accepted until UAT in F26-g FR-12 Rev 5 (.forge/owed.md) · Severity: medium — the email reaches the customer, but the account has no record or timeline row of it, and it can list the wrong week’s prices · Phase: 1
Repro: a rep chooses Send This Week’s Prices and Continue, and the compose dialog opens filled for the current sale. Before the rep sends, the Tea Room deletes that sale (a mistyped one, possible since BUG-422), corrects its dates so another sale is current, or the next sale starts. The rep sends.
Expected: the email is not sent, and the rep is told why and how to send this week’s prices.
Actual: the email is sent; the record call is refused (422 PRICE_EMAIL_NOT_RECORDABLE, SALE_NOT_FOUND) and the action shows Sent, but not marked as sent, which Retry cannot clear. When another sale had become current, the email names the old sale and lists its prices.
Fix: decided by the lead on 2026-11-15: refuse the send before it leaves, rather than record the email against the current sale. The compose opener gains a beforeSend check (F20-c Rev 14); the price email’s check reads the preparation again and refuses when its sale is not the one the email was filled for, or when the read fails (F26-d Rev 5). F26-g Rev 6 drops the accepted window.
Regression test: ComposeEmailDialog.beforeSend.test.tsx (the dialog waits for the check, locked; a refusal shows in the message region with focus and sends nothing; a failed check shows BEFORE_SEND_FAILED with its reference; Send Again checks again, a refusal there keeps the uncertain notice, and its buttons are held while the check runs), ComposeEmailLayer.test.tsx (the layer hands beforeSend to the dialog) and AccountPriceEmailAction.test.tsx § the check before the email leaves (the same sale lets it go; another sale or no sale refuses; a failed read rejects with its reference; the check reads the preparation again).
Resolved: 2026-11-15 — fixed in altrium-io/forge-et-fe#538, as the Fix above describes (F20-c spec Rev 14, F26-d spec Rev 5, F26-g spec Rev 6, the user guide’s Sending prices to a customer). The compose dialog runs the opener’s beforeSend on Send and Send Again. The price email’s check reads the preparation again and refuses unless its sale is the fill’s: The email was not sent: this week’s sale has changed since the prices were filled in. Close this email and choose Send This Week’s Prices again. A failed read shows F20-c’s check-failed sentence with its reference. A sale deleted in the moment between the check and the send is still refused by the record. Before the fix the dialog tests fail (it sends at once) and the action’s tests find no check. With it, altrium-io/forge-et-be#437 indexes price_email(auction_sale_id) and buying_standard_cost(anchor_sale_id), the foreign keys a sale delete checks (the owed.md row the lead approved on 2026-11-15).
Lesson: L-302 — added as its third bullet: a change that makes a record correctable or deletable lists every flow that holds its id while a person works.
BUG-434 — Changing the active campaign in the capture app can send one plan-list read with the previous campaign’s filters
Affects: F36-a (forge-et-pwa src/features/plan/exhibitors/PlanList.tsx, on feature/F36-a-wave-3, not on main) · Found in: the BUG-430 fix (dd91617), 2026-11-15 — what that fix left · Severity: low · Phase: 1
Repro: with a search, chip or hall filter set on one campaign’s plan list, choose another campaign that the app has already loaded, while Plan stays on screen.
Expected: the new campaign’s list is read only without the filters.
Actual: the list sends one read with the old filters in the render before its clear runs, then reads again without them.
Fix: PlanList holds the campaign the URL’s list controls belong to in its own state (listCampaign, initialised from notedListCampaign()), and notes it in sessionStorage best-effort for a remount (noteListCampaign). While that campaign is another one, or while the URL is still the one whose controls are being cleared (clearingKey === location.key), the list reads the defaults and the drawer stays shut. The clear’s navigation, or any the user makes first, has a new location key and ends the hold. A storage that refuses a write cannot lock the list, and the effect acts once per campaign under StrictMode. The bug was on main by the time it was fixed (F36-a wave 3 had merged).
Regression test: PlanList.test.tsx “a change of campaign reads the new campaign’s list once, without the last one’s controls (BUG-434)”, under StrictMode: with the campaign already loaded (fails on main with a second read carrying q=west&status=visited&sort=company&hall=Hall+5) and not yet loaded, both checking the drawer shuts; and a case where the tab’s storage refuses writes (fails on the first fix, 45a9d9a). planListParams.test.ts rows cover the helpers.
Resolved: 2026-11-17 — fixed in forge-et-pwa#144 (merge 3bc35ec2; Test record and a full browser gate, 4,624 passed, on b7367e0).
Lesson: none — one-off: the hold lived in a sessionStorage read during render; the reviewer’s checks (StrictMode, refused storage, two navigations in one render) are frontend-reviewer dimensions already.
BUG-433 — Pressing Add Exhibitor once in the capture app may add the exhibitor twice under load
Affects: F36-a (forge-et-pwa Add Exhibitor sheet, src/features/plan/exhibitors/ExhibitorFormSheet.tsx, WI-3.4, on feature/F36-a-wave-3, not on main) · Found in: the BUG-430 and BUG-431 fix run on feature/F36-a-wave-3, Windows 11, 2026-11-15 · Severity: low · Phase: 1
Repro: not yet reproduced. Under machine load, ExhibitorFormSheet.test.tsx › “Add Exhibitor alone sends one…” failed once, before the BUG-430 and BUG-431 changes, with the toast 2 exhibitors added. instead of one company’s name. The same file failed once on a different test with those changes reverted.
Expected: one press of Add Exhibitor adds one exhibitor.
Actual: once, the sheet reported two added after one press. It is not decided whether the form submitted twice or the test misread a timing.
Fix: not decided. Re-run the test alone under load (.forge/test-strategy.md § Flaky Tests), read how the sheet guards against a second submit while the first X-2 write runs, and if it does not, disable the button until the write settles, with a test that presses twice quickly.
Resolved: 2026-11-16 — closed as wontfix (not reproduced), Tharinda decided. On main at forge-et-pwa fe5f3de, ExhibitorFormSheet.test.tsx passed all 10 tests in each of 8 runs on a Mac whose load average was 65 to 90 on 8 cores. The sheet already refuses a second press while a write runs, twice over: submit returns at once while busy is set, and the Button drawn busy ignores a press. A test that clicked Add Exhibitor twice in one task sent one create both with and without an extra ref-held guard, because React renders the first press’s busy state before the second click runs, so no change was made. A recurrence is filed as a new bug that links here, with the failing run’s output and the machine’s load.
Lesson: L-162 (a failure that appears only under load is reproduced before it is called flaky) — an instance where it was not reproduced, so nothing was changed.
BUG-432 — Flaky tests: two CaptureOutboxProvider accept tests time out in a loaded full Vitest run (Windows)
Affects: F36-a (forge-et-pwa src/lib/capture/CaptureOutboxProvider.test.tsx, “accept” › “resolves once the capture is in the store…” and “online, does not list a capture while its first send runs…”; wave 2 code) · Found in: the F36-a wave 3 merge’s whole Vitest run on feature/F36-a-wave-3 a272eec, Windows 11, 2026-11-15 · Severity: low · Phase: 1
Repro: the whole suite (pnpm test) under machine load: that run took 655 s and both tests failed; the file alone passed 34 of 34, three times.
Expected: the tests pass whatever else the machine is running.
Actual: they failed in two whole-suite runs: the wave 3 merge’s (under load) and WI-3.6’s test-verify (feature/F36-a-WI-3.6-verify-plan 75d90c1, started at 25% CPU), each time expected [] to have a length of 1 but got +0 at await eventually(() => expect(sends).toHaveLength(1)). Neither wave 3 work item touches src/lib/capture. Cause: they are the file’s first tests to send, so they wait on the first transform of the outbox’s dynamic import('./sendPass') (forge-et-pwa#132), and eventually gives up after 200 rounds of setImmediate, which a whole-suite run outlasts.
Fix: fixed on forge-et-pwa feature/F36-a-wave-2 in 78a5ef7 (forge-et-pwa#132): the test file loads ./sendPass once in beforeAll, so the provider’s import resolves from the module cache. The whole wave 2 suite then passed (90 files, 1,626 tests, EXIT=0). One green run proves little for an intermittent failure; the stronger evidence is that the import is the only wait in these tests that takes real time — fake-indexeddb, the mocked upload and React’s commits all settle within the counted rounds (frontend-reviewer’s delta review of #132).
Resolved: 2026-11-15 — fixed in altrium-io/forge-et-pwa#132 (merged as 7de6e638), as the Fix above describes: CaptureOutboxProvider.test.tsx loads ./sendPass once in beforeAll. Wave 2’s whole Vitest suite passed with it, and its test record on #132’s head (78a5ef7) passed.
Lesson: none — one-off: these were the only tests whose first send waited on a dynamic import’s first transform.
BUG-431 — Choosing two halls quickly in the capture app’s plan list can send only the second
Affects: F36-a (forge-et-pwa the plan list’s hall filter, WI-3.4, on feature/F36-a-wave-3, not on main) · Found in: the F36-a wave 3 merge’s Chromium browser run, e2e/browser/plan/plan-list.spec.ts line 82, 2026-11-15 (1 failure in 50 runs on the merged head, none in 30 at WI-3.5’s tip) · Severity: low · Phase: 1
Repro: in Sort and Filter, press Hall 4 and then Hall 5 in quick succession.
Expected: the list read carries both halls.
Actual: once, the read carried only Hall 5: the second press’s URL write likely read the parameters before the first press’s write had landed (a lost update between two setSearchParams calls).
Fix: fixed on forge-et-pwa feature/F36-a-wave-3 in 5f3b701; ships with the wave 3 capture-app PR. The functional form of setSearchParams was not enough: in React Router 7.18.5 its updater receives the parameters of the render that made the setter, not the latest URL, and Sort and Filter built each choice from the current render’s parameters. New src/features/plan/latestSearch.ts (useLatestSearchParams) remembers each write against the location it was made from, so a second write before the next render starts from the first one’s result; every list control’s write goes through it. Tests: two halls in one tick (PlanList.test.tsx, SortFilterSheet.test.tsx), a sort and a hall in one tick, and latestSearch.test.ts; each failed before the fix. Left as is: an exhibitor card’s link is still built from the current render’s URL, so a card clicked in the same moment as a control press could leave that press out.
Resolved: 2026-11-16 — fixed in altrium-io/forge-et-pwa#136 (F36-a wave 3, merged as 2037aa9e), as the Fix above describes. The PR’s review record (approve), Test record and full browser run (4,418 passed) were on its head dd7f75c.
Lesson: L-201 — added to its sources: the capture app’s URL writes started from one render’s parameters, the same lost update in another app.
BUG-430 — Switching campaigns while viewing an ended campaign’s plan can keep that plan’s list filters in the URL
Affects: F36-a (forge-et-pwa the plan list, WI-3.4, forge-et-pwa#134, on feature/F36-a-WI-3.4-plan-list-booth, not on main) · Found in: F36-a-WI-3.4’s build, 2026-11-15 · Severity: low · Phase: 1
Repro: open an ended campaign’s plan with View Plan (/plan?view=plan), set a search, chip or hall filter, then choose another campaign in the picker.
Expected: the new campaign’s plan opens with no list filters, since halls and searches belong to the campaign they were set on.
Actual: the old filters can stay in the URL and apply to the new campaign’s list.
Fix: fixed on forge-et-pwa feature/F36-a-wave-3 in dd91617; ships with the wave 3 capture-app PR. Two causes: when the new campaign was also ended, the list never mounted to clear its own parameters; and when it was already loaded, the list’s clear and Plan’s drop of view ran from the same parameters, so Plan’s write put the filters back. usePlanView().close is replaced by leaveCampaign(), which drops view, q, status, sort, hall and exhibitor in one write on any change of campaign (helper withoutListParams in planListParams.ts). Tests in PlanList.test.tsx and planView.test.ts failed before the fix. What remains is BUG-434: the list can send one read with the old filters in the render before its clear runs, when the new campaign is already loaded.
Resolved: 2026-11-16 — fixed in altrium-io/forge-et-pwa#136 (F36-a wave 3, merged as 2037aa9e), as the Fix above describes. The PR’s review record (approve), Test record and full browser run (4,418 passed) were on its head dd7f75c. BUG-434 stays open for the one read the fix leaves.
Lesson: L-201 — added to its sources: two URL writes made from the same render’s parameters, the second putting back what the first removed.
BUG-429 — The capture app’s campaign reads are not read again when the connection returns
Affects: F36-a (forge-et-pwa src/lib/campaign/ queries, WI-3.3, forge-et-pwa#133, on feature/F36-a-WI-3.3-campaigns, not on main) · Found in: F36-a-WI-3.4’s build, 2026-11-15 (its finding F-2) · Severity: low · Phase: 1
Repro: open Plan or the campaign picker, lose the connection, then get it back.
Expected: the campaign list and the active campaign are read again on reconnect, as their code comment says.
Actual: TanStack Query sets refetchOnReconnect off by default for queries with networkMode: 'always', which the campaign queries use, so they stay on the earlier data (with Shown as of) until something else reads them. WI-3.4’s exhibitor queries set refetchOnReconnect: true explicitly.
Fix: fixed on forge-et-pwa feature/F36-a-wave-3 in d25d00f; ships with the wave 3 capture-app PR. refetchOnReconnect: true on useCampaignList and useCampaignDetail, and the code comment that said it was on by default corrected. Two tests in campaignQueries.test.tsx check that a reconnect reads again; each failed before the fix.
Resolved: 2026-11-16 — fixed in altrium-io/forge-et-pwa#136 (F36-a wave 3, merged as 2037aa9e), as the Fix above describes. The PR’s review record (approve), Test record and full browser run (4,418 passed) were on its head dd7f75c.
Lesson: none — one-off: on forge-et-pwa main at 2037aa9e, every useQuery with networkMode: 'always' sets refetchOnReconnect (the campaign and exhibitor queries, each with a comment naming the default; useCustomerCheck turns it off on purpose), the other 'always' sites are mutations and one-off fetchQuery reads, where the option does not apply, and forge-et-fe has no networkMode: 'always' (git grep -n "networkMode" origin/main -- src in each repo).
BUG-428 — The OpenAPI document drops every schema that has a nullable record component with allowed values
Affects: F36-a (forge-et-be platform/web/NullableRecordComponentConverter.listNullInEnum; any response or request record with a @Nullable component that also carries @Schema(allowableValues = …)) · Found in: F36-a-WI-3.1’s build (forge-et-be#433), 2026-11-14 · Severity: low · Phase: 1
Repro: give a record a @Nullable String component with @Schema(allowableValues = {"a", "b"}), expose it on an endpoint, and fetch /v3/api-docs.
Expected: the schema is listed, its enum the allowed values plus null.
Actual: listNullInEnum adds null to the immutable list springdoc built from allowableValues and throws; springdoc then leaves out every schema that contains such a component. WI-3.1 worked around it: ExhibitorResponse.archiveReason and ExhibitorArchiveRequest.reason name their tokens in a description instead of an enum.
Fix: copy the list before adding null in listNullInEnum, with a converter spec row; then restore the two enums in WI-3.1’s records (and their CapturePlanOpenApiIT rows). Outside WI-3.1’s scope, so not fixed there.
Resolved: 2026-11-15 — fixed in altrium-io/forge-et-be#436. listNullInEnum copies the enum before adding null, and does the contains(null) check on the copy, because an immutable list throws on that check too; it does nothing when the property has no enum. Regression tests: NullableRecordComponentConverterSpec — a fixed-size, an immutable and a mutable enum each gain null, an enum that already lists null is left as it is, and a test-local record with a @Nullable component carrying allowableValues resolved through swagger-core’s own ModelConverters keeps its schema, with enum == ['a', 'b', null]. Before the fix, the fixed-size row threw UnsupportedOperationException, the immutable row NullPointerException, and the swagger-core row failed. Restoring ExhibitorResponse.archiveReason and ExhibitorArchiveRequest.reason as enums is left to F36-a wave 3 (forge-et-be#435), where those records are; it is a row in .forge/owed.md.
Lesson: none — one-off: the converter was the one place that added to an enum swagger-core had built from an annotation. The customizers that add enum values (OpenApiEnums, LeadOpenApiConfig, ActivityOpenApiConfig, TagOpenApiConfig) add to query parameters, or to nullable properties after this converter, whose enums are null or now a mutable copy, and a throw there fails /v3/api-docs as a whole, which OpenApiIT would catch.
BUG-427 — Flaky test: the Fast lane documentation-only suite runs git on the repository under Vitest’s 5 s default, and times out in a loaded full run (Windows)
Affects: F40 (forge-et-pwa src/test/scripts/localTestRecord.test.ts, “the Fast lane documentation-only decision › fails closed when git cannot list the paths”, line 294; the file is a byte copy of forge-et-fe’s, scripts/shared-with-fe.json) · Found in: F36-a-WI-2.4’s second /forge-test-verify (pnpm test:selected, full, on forge-et-pwa 313bae5, Windows 11, CPU at 99% from other agents, 2026-11-14) · Severity: low · Phase: 1
Repro: that run: Test timed out in 5000ms; 1,517 other tests passed (one more, BUG-350’s already-60 s suite, timed out at 60 s under the same load). Alone, with browserGateSelect.test.ts, both files pass (40 of 40, 63 s).
Expected: the test passes whatever else the machine is running.
Actual: the suite calls changedPaths / scope, which run synchronous git processes on the real repository, under Vitest’s 5 s default.
Fix: { timeout: 60_000 } on the suite’s describe (line 270), with the git-fixture suites’ comment; check the file’s other suites (the local test record, the test record on a loaded machine, the retry, the flaky and the slowest tests, and the machine lock) for git or process work under the default and give each that runs git the same. The fix lands in forge-et-fe and is copied byte for byte to forge-et-pwa. The same exception as BUG-350 and BUG-425: synchronous git with nothing racing.
Resolved: 2026-11-15 — fixed in altrium-io/forge-et-fe#537 (merged as 182b2217), with BUG-425, as the Fix above describes, and copied byte for byte to forge-et-pwa in altrium-io/forge-et-pwa#132 (merged 2026-11-15), its shared-file record updated.
Lesson: L-257 — another instance: git run on the repository itself counts as a git fixture for the timeout. L-257’s sources now list BUG-427.
BUG-426 — A multipart request whose body cannot be parsed answers 500, which the capture app resends without end
Affects: platform (forge-et-be platform/web/ApiExceptionHandler — every multipart route: F36-a’s POST /api/v1/capture/leads, the attachment upload) · Found in: the review of forge-et-be#432 (F36-a wave 2), 2026-11-14 · Severity: low · Phase: 1
Repro: send Content-Type: multipart/form-data; boundary=x with a body that does not use that boundary, or is cut short, to POST /api/v1/capture/leads as a signed-in capture.use holder.
Expected: a 400 in the RFC 9457 envelope (MALFORMED_REQUEST or the project’s equivalent), which the capture app keeps as Refused (F36-a wave 2 contract § What the capture app does with each E-1 answer).
Actual: Spring raises a MultipartException; ApiExceptionHandler has no mapping for it, so handleUnexpected answers 500, and the capture app resends a 5xx with backoff indefinitely. The browser’s own FormData never builds such a body, so the app cannot cause it; only a hand-made or truncated request does.
Fix: map MultipartException (other than the size-limit subclass, which already answers 413) to 400 with the project’s malformed-request code in ApiExceptionHandler, with an IT on one multipart route. Pre-existing platform behaviour, not changed by F36-a; left out of forge-et-be#432 as code the PR did not change.
Resolved: 2026-11-15 — fixed in altrium-io/forge-et-be#436, narrower than the Fix: line. Spring wraps every multipart parse failure in a MultipartException, the server’s own included: a full disk, an upload directory that is gone, a stored part that cannot be read. A 400 for those would be logged at DEBUG, where no alarm sees it, and the capture app would keep the capture as Refused and never send it again. So ApiExceptionHandler.handleUnreadableMultipart answers 400 MALFORMED_REQUEST only when the failure is the client’s. That covers:
- Spring’s own “not a multipart request” (no cause);
- Spring’s refusal of a part’s
Content-Disposition(anIllegalArgumentExceptionas the direct cause); - a
multipart/*content type with no boundary, or one that does not parse. Tomcat reports a missing boundary with the sameFileUploadExceptiona full disk gets, so the request’s own header decides; - a cause chain holding Tomcat’s
MalformedStreamException,InvalidContentTypeExceptionorInvalidFileNameException(a NUL in a part’s file name), anEOFException, aClientAbortExceptionor aSocketTimeoutException.
Every other MultipartException stays 500 INTERNAL_ERROR with an ERROR log. MaxUploadSizeExceededException keeps the base class’s 413. Found while fixing: a body whose boundary differs from the announced one is not a parse failure; it parses as no parts and already answered 400 VALIDATION_FAILED (capture is required). Regression tests:
CapturedLeadBodyUnreadIT, on a real Tomcat: a body cut short in a part’s value, one cut short in its headers, and amultipart/form-datatype with no boundary each answer400 MALFORMED_REQUEST; a wrong boundary answers400 VALIDATION_FAILED. Before the fix, the cut-short and no-boundary bodies answeredHTTP/1.1 500.ApiExceptionHandlerSpec: each client cause answers 400 at DEBUG. A full disk, a part not written to disk, a missing upload directory, a missing multipart configuration and an unreadable stored part answer 500 with one ERROR. The server rows failed against the review’s first version of the fix, which mapped everyMultipartExceptionto 400. The no-boundary, NUL-file-name andContent-Dispositionrows failed against its second version. Lesson: none — one-off:ResponseEntityExceptionHandlermaps the framework’s request exceptions, andMultipartExceptionis the one it leaves out that a client can cause. The other unmapped exceptions that answered 500 for client input came from the database (L-087, L-280), not from the advice.
BUG-425 — Flaky test: the browser gate’s selected-run suite times out building its git fixture in a full Vitest run (Windows)
Affects: F40 (forge-et-pwa src/test/scripts/browserGateSelect.test.ts, “the browser gate’s selected run (test-run-cost.md S-2) › writes a test list relative to e2e/ with the setup files, and the size matrix’s selection”, line 174; the file is a byte copy of forge-et-fe’s, scripts/shared-with-fe.json) · Found in: F36-a-WI-2.4’s /forge-test-verify (pnpm test:selected, full, on forge-et-pwa feature/F36-a-WI-2.4-verify-capture f597183, Windows 11, 2026-11-14) · Severity: low · Phase: 1
Repro: two full Vitest runs on that head: the first (CPU at 94% after it) timed it out with two SettingsRoute.test.tsx load timeouts; the second, started after three load samples under 65%, timed out this test alone (Test timed out in 5000ms, 14.1 s with its retry), 1,518 others passing. Alone, the file passes.
Expected: the test passes whatever else the machine is running.
Actual: the suite’s tests build a git repository fixture synchronously under Vitest’s default 5 s timeout. BUG-350 gave the file’s other git-fixture suite, “selectGate against a repository” (line 283), { timeout: 60_000 }; this suite, at line 80, kept the default.
Fix: { timeout: 60_000 } on the describe at line 80, as BUG-350’s suite has, with the same comment; the fix lands in forge-et-fe and is copied byte for byte to forge-et-pwa with the manifest updated (check-shared-drift.mjs --update). Not the lengthened timeout .forge/test-strategy.md § Flaky Tests bans, on the reasoning Nanndish accepted for BUG-350: the work under the limit is synchronous git and file copies with nothing racing.
Resolved: 2026-11-15 — fixed in altrium-io/forge-et-fe#537 (merged as 182b2217), with BUG-427, as the Fix above describes, and copied byte for byte to forge-et-pwa in altrium-io/forge-et-pwa#132 (merged 2026-11-15), its shared-file record updated.
Lesson: L-257 — another instance: BUG-350 fixed the suite that failed, not every suite in the file that builds git fixtures. L-257’s sources now list BUG-425.
BUG-424 — Flaky test: record-attachments.spec.ts AC-21 reloads while the delete’s list re-read is in flight, so its body cannot be read (macOS)
Affects: F33-a (forge-et-fe e2e/browser/attachments/record-attachments.spec.ts AC-21, through recordListAnswers in e2e/browser/attachments/attachments-support.ts) · Found in: the nightly full run on forge-et-fe main 0f99a28f, 2026-11-14 (macOS; forge-et-fe#535), and once before on 2026-11-11 · Severity: low — a test race, no product defect · Phase: 1
Repro: ET_BACKEND_PORT=<port> pnpm e2e e2e/browser/attachments/record-attachments.spec.ts --project=chromium --retries=0 --repeat-each=10 --workers=4 on forge-et-fe 0f99a28f against a local backend at forge-et-be 11bc8e41, nothing else running: 35 passed, 5 failed (2 on an account, 3 on a contact). On 2026-11-14 the nightly gate failed the contact case on both attempts and the account case once (passed on its retry); on 2026-11-11 the contact case failed once and passed on its retry.
Expected: the test passes whenever the card does what F33-a AC-21 asks.
Actual: every list answer the page received narrows as an AttachmentList fails with a list answer could not be read: Error: response.json: Protocol error (Network.getResponseBody): No resource with given identifier found … Response body is not available for a response that was navigated away from. The card takes a deleted file out of its list on the delete’s 204 and then re-reads the list in the background (recordAttachmentsQueries.ts). The test reloaded as soon as the row left, so the re-read’s answer could arrive while the page was navigating, and the listener could not read its body. The failed attempt’s trace shows the DELETE answered 204, the list GET answered 200 and the reload 18 ms after it.
Fix: in the test, start waiting for the list’s next answer (nextListAnswer, a new helper beside recordListAnswers) before confirming the delete, and wait for that answer and for every list body read (lists.settled()) before the reload. No product change.
Regression test: AC-21 itself, repeated 10 times per record type with retries off.
Resolved: 2026-11-14 — fixed in forge-et-fe#536. With the fix the same repro passed 40 of 40 (30 of 30 AC-21 attempts: 10 account, 10 contact, 10 lead); without it 5 of 30 failed.
Lesson: L-303 (new) — a test that reads answer bodies from a response listener lets every read still in flight land before it navigates; frontend-reviewer dimension 14 points at it.
BUG-423 — The Pricing page warns every week that a sale’s round is unpublished the day before it starts, which is noise for a Tea Room that prices after the auction
Affects: F26-g (FR-14 round_not_published, FR-15’s text; forge-et-fe PricingWarnings.tsx) · Found in: a review of the weekly pricing features against Empire’s practice, 2026-11-13 (.forge/engagement-log.md, 2026-11-13) · Severity: low — a notice only; nothing is blocked · Phase: 1
Repro: enter a sale that starts tomorrow and publish nothing for it; the Pricing page shows the round_not_published warning (“…will have no current price from tomorrow unless it is published”, with Review the Round). Empire never confirmed whether the Tea Room prices before a sale or after it, from its results (the same entry). If they price after it, the warning shows every week.
Expected: a notice that states the state and its consequence without reading as a fault, so it fits both ways of working (the lead, 2026-11-13: simple, minimal and flexible, no questions to Empire before UAT).
Actual: it reads as a warning.
Fix: decided by the lead on 2026-11-13: show it as information, not a warning, and keep its consequence, which holds whichever way the Tea Room works, for example “Prices for no_next_sale warning stays a warning, because without a next sale prices cannot end. An F26-g spec revision (FR-15’s text) in the fixing PR.
Regression test: the Pricing page test for round_not_published asserts the new text, its count and date, and the information styling.
Resolved: 2026-11-14 — fixed in altrium-io/forge-et-fe#533, with BUG-422 (F26-g spec Rev 5, FR-15). round_not_published reads “Prices for <date> the sale’s start; “1 standard has” in the singular) for every timing, with Review the Round, in its own block in the Alert’s neutral tone below the warnings (pricing-information, the schedule glyph); no_next_sale and sweep_overdue stay warnings. The server’s warning is unchanged. Regression tests: PricingWarnings.test.tsx (the sentence, its count and date for each timing and in the singular, the untinted information block holding the round line, the warning block holding the other two), and the browser warnings specs (expectInformationBlock); before the fix the sentence and the warning tint assertions fail (PricingWarnings.test.tsx against main’s PricingWarnings.tsx: 6 of 19 failed).
Lesson: none — one-off: the wording assumed the Tea Room prices before the auction, a practice Empire never confirmed; no other screen states that assumption.
BUG-422 — A mistyped auction sale cannot be corrected or removed, and one typo can empty a week’s price emails or end every current price that night
Affects: F26-c (FR-1, creating a sale), F26-g (FR-12, editing a sale’s dates; FR-2, the daily sweep), F26-d (the price email’s current sale) · Found in: a review of the weekly pricing features, 2026-11-13 · Severity: medium — a typo in a sale’s number or start date, recoverable only by publishing a round for the mistaken sale · Phase: 1
Repro: on forge-et-be main, AuctionSaleController has POST (create) and PATCH /{saleId} (dates only, and only while the sale has not started and has no published or withdrawn price); there is no way to change a sale’s year or number, or to delete a sale. Two typos then do real harm:
- Sale 433 entered instead of 43, with the same dates. The current sale is the latest start on or before today, ties going to the higher number (F26-g FR-2), so 433 is current; the real sale’s published prices are not anchored to it, so every customer’s price email lists them as left out (F26-d decision 2).
- Next week’s sale entered with today’s date. It has started, so its dates are locked, and that night the sweep ends every published price from earlier sales that no publish replaced (F26-g FR-2), leaving the standards unpriced until a round is published for the mistaken sale.
Expected: the Tea Room can put a mistake right before it does harm.
Actual: a mistaken sale stays forever, and its dates lock once it starts.
Fix: decided by the lead on 2026-11-13: (a) a sale with no
publishedand nowithdrawnprice can have its year, number and dates corrected (a started one included, so a sale entered with today’s date can be moved), or be deleted, its drafts with it; each with a reason and an audit event (AD #36); a sale with any published or withdrawn price stays as it is, except its end date (Plan), and its prices are put right through the existing withdraw and correct actions; (b) the Add a Sale dialog says in plain words, before it saves, when the sale starts today or earlier: from that day it is the current sale, and prices from earlier sales that its round does not replace end that night. Spec revisions to F26-c FR-1 and F26-g FR-12 in the fixing PR. Not reopened: prices the sweep already ended stay ended (AD #14); the dialog’s notice is what prevents it. Regression test: a backend IT per case — correct the number of a sale with only drafts; move the start of a started sale with only drafts; delete such a sale and its drafts; refuse a delete when aprice_emailrow references the sale; change only the end date of a sale with published prices, and refuse any other change to it; refuse every change to aheldsale; an edit of the current sale running alongside a sweep (Plan, the sweep); the audit events — and frontend tests of the plain-words notice for a start of today or earlier and of which actions each sale offers. Plan (Tharinda, 2026-11-14; fixed with BUG-423 in onefix/BUG-422-…branch per repo, both bugs claimed by harness#1390):
- Added to the fix (lead’s rule of 2026-11-13, flexible and simple): the end date can change on any sale that is not
held, published prices or not, becauseends_ondecides only when the sweep marks the saleheld(F26-g FR-3) and what the sale list shows; it never bounds a price’s validity, which runs to the next sale’s start. The new end must be on or after the start and on or after today. The start date, year and number keep the stricter rule below. No Start now or End now button (the lead, 2026-11-14): a sale starts by its first day and ends by its last, and editing those dates is the one way to change either. - Backend (
forge-et-be, packagecom.empireteas.pricing). TodayPATCH /api/v1/auction-sales/{saleId}(AuctionSaleController.changeAuctionSaleDates,AuctionSaleDatesRequest) runsAuctionSaleDatesService.changeDates: round lock (AuctionSaleRepository.lockForRoundWrite), thenRoundPriceWriteRepository.lockPublishedPrices, version check,409 SALE_STARTEDwhen held or started,startsOnafter today,409 SALE_PRICES_PUBLISHEDwhenAuctionSale.isPricesEverPublished(),409 LATER_PRICE_PUBLISHED(earliestPublishedStartOnOrAfter), thenAuctionSale.changeDates,DraftPriceWriteRepository.moveDraftsand auditauction_sale.dates_changed. Change it to:- a sale with no published and no withdrawn price (
!isPricesEverPublished()), started or not, but notheld, may changesaleYear,saleNo,startsOnandendsOn. KeepLATER_PRICE_PUBLISHEDfor a new start; drop the “start after today” rule for such a sale, since nothing has been closed at its start yet unless the sweep ran (prices the sweep already ended stay ended, AD #14, not reopened). A duplicate year and number is refused as create refuses it (409onauction_sale_unique_no). Auditauction_sale.dates_changedas now, and a newauction_sale.number_changed(old and new year and number) when they change, each with a requiredreason(1–500 characters) in the request — add the event codes toforge-et-be/docs/audit/README.md; - a sale with a published or withdrawn price: only
endsOnmay change (not held; on or afterstartsOnand today); anything else is409 SALE_PRICES_PUBLISHED. - Delete: new
DELETE /api/v1/auction-sales/{saleId}underbuying_standard.manage, body or queryversionandreason, same locks. Allowed only when!isPricesEverPublished()and noprice_emailrow references the sale (V48__price_email.sql,auction_sale_id NOT NULL). Deletes the sale’s drafts (buying_standard_pricerows withanchor_sale_id= the sale; that column isNOT NULL,V29__price_and_cost.sql:107) and clearsbuying_standard_cost.anchor_sale_id(nullable,V29:197) or refuses when one is set — check which is right in F26-b before choosing. Auditauction_sale.deletedwith the sale’s year, number, dates and the drafts removed.409 SALE_PRICES_PUBLISHEDotherwise. - The sweep.
PriceExpirySweepServicereads the current sale without a lock and locks it after, and its Javadoc relies on “C has started, so its dates cannot change meanwhile (F26-g FR-12 refuses a started sale)”. Once a started sale can be edited, an edit could land between that read and the lock. Re-read the current sale under its lock in the sweep (or take the locks in an order an edit cannot cross), add an IT for an edit of the current sale running alongside a sweep, update both services’ Javadoc, and say in the F26-g FR-12 revision why AD #14 still holds (no as-of answer for a day before today changes: the sweep’s close date is never before today). - What the sale list tells the screen.
SaleListing.canEditDates(“upcoming and never published”) andearliestStartsOn(“tomorrow”) decide today whether Edit Dates shows (AuctionSalesTable.tsx) and the Starts minimum (EditSaleDatesDialog.tsx). Replace them with what the new rules need — for examplecanEditSale,canEditEndOnly,canDelete, and the earliest allowed start and end — inAuctionSaleListItemandauctionSalesWire.ts; without them a started sale or an end-date edit is never offered. ASK FIRSTapplies: no schema change is expected; if one is needed, ask before writing a migration.
- a sale with no published and no withdrawn price (
- Frontend (
forge-et-fe,src/features/pricing).EditSaleDatesDialog.tsxbecomes Edit Sale (year, number, start, end and a reason; only End enabled, with a one-line reason why, when the sale has published prices); a Delete Sale action beside it with a confirmation naming the sale and its draft count, offered only when the sale can be deleted (the list item’s new flags, Backend above, with the matching wire and validation changes). Pricing labels are Title Case: add the new ones topricingTitleCase.test.ts’s lists (design-system §9, Rev 90).NewSaleDialog.tsx, and the edit dialog when the start moves to today or earlier: before saving, the plain-words notice “This sale starts on. From that day it is the current sale, and prices from earlier sales that its round does not replace end that night.” A past or today start stays allowed on Add a Sale, and becomes allowed on Edit Sale for a sale with no published price, so the two follow one rule (the lead, 2026-11-14): a sale that already started must be enterable on go-live day or when entered late. The end date stays on or after the start, checked on screen and on the server, with “Enter an end date on or after the start date”. Update crm-currentin.forge/design/prototypes/for the changed dialogs (.claude/rules/plans.md § Required Sections) and every test id and string the specs ande2e/assert (fewer-nightly-reds.mdR-3); addPAGE_STATESentries for new dialog states. - Docs in the same change: F26-c spec revision (FR-1: the start-today notice) and F26-g spec revision (FR-12: what each sale may change, delete, the end-date rule; FR-15 for BUG-423), each with
Approved by: Tharinda (lead)and the date in its### Revheading (the decisions are of 2026-11-13 and 2026-11-14);.forge/client/user-guide-weekly-prices.mdstep 1, § Auction sales (the table and the “What can be changed” bullet) and § Things to know (correct or delete a mistaken sale, the end date), and step 4 for BUG-423’s new notice; this section’sResolved:andLesson:lines andtracker.yamlstatus: fixed,fix_pr. - Process:
.claude/rules/tracker.md § Bug Tracking(no/forge-deliver): draft PRs,/forge-review-pron each (security pass applies — a new endpoint),python3 scripts/local-build-evidence.pyinforge-et-be,pnpm test:recordthenpnpm e2e:gate --selectedinforge-et-fe; merge the backend first, then the frontend within minutes (L-200: checke2e/for tests asserting the old409 SALE_STARTEDon a started sale’s edit). Resolved: 2026-11-14 — fixed in altrium-io/forge-et-be#430 and altrium-io/forge-et-fe#533, as the plan above sets out (F26-g spec Rev 5, F26-c spec Rev 6, the user guide’s step 1, step 4, § Auction sales and § Things to know, thecrm-currentprototype). The sweep re-reads the current sale’s id and start under the published-price lock and, when an edit, a delete or a create changed them since its first read, closes nothing and fails so the redelivery works the close out again (PriceExpirySweepService,CurrentSaleChangedException). The sale list carriescanEditSale,canEditEndOnly,canDeleteandearliestEndsOnin place ofcanEditDatesandearliestStartsOn; the screen offers Edit Sale and Delete Sale from them and checks a changed end against the server’s today. Decided while fixing: a cost row that names the sale refuses the delete (409 SALE_IN_USE, as a sent price email does) rather than having its anchor cleared, because no writer sets that column (F26-b Q-3) and a cost row is not updated in place (AD #14); a price email or cost row committed during a delete is answered as the same409from its foreign key; the list has no earliest start, since a start has no lower bound; and a price email being written when its sale is deleted is accepted for now (.forge/owed.md). Regression tests:PricingCalendarIT(a number corrected withnumber_changed; the year, number and dates in one request,number_changedbeforedates_changed; a sale started today and one started three days ago moved with their drafts; a duplicate number; an end moved on a sale with a published or only a withdrawn price; a new start, end or number on a held sale; an end of exactly today accepted, one before today refused; deletes and each refused delete; a price email committed during a delete; an edit of the current sale committed while the sweep waits),AuctionSaleControllerITandSaleListingSpec(the list flags). Before the fix asaleNoin the body was400 UNKNOWN_FIELD, a started sale409 SALE_STARTED, and theDELETEhad no handler; with the two fixes reverted on the branch, the delete race answered500(result.status.value() == statusfailed) and the sweep race closed the earlier price (failure != nullfailed). Inforge-et-fe:EditSaleDialog.test.tsx(the locked fields disabled with their reason; the notice in a live region),editSaleForm.test.ts(the notice at Colombo’s midnight and on a malformed date),DeleteSaleDialog.test.tsx(each refusal with Reload; an unknown outcome never says nothing was deleted),NewSaleDialog.test.tsx(the plain-words notice for a start of today or earlier),AuctionSalesTab.test.tsx,AuctionSalesTable.test.tsx(which actions each sale offers),auctionSalesWire.test.tsandpricingTitleCase.test.ts(Edit Sale, Delete Sale). Lesson: L-302 (new) — a record people type in needs a way to put a typo right, stated with the rules that lock it.
BUG-421 — At DEBUG, the S3 HTTP client logs object keys, including attachment keys
Affects: FS-11 (the file pipeline’s ObjectStore, forge-et-be application.yml logging levels) · Found in: F36-a WI-2.1’s log check (feature/F36-a-WI-2.1-captured-leads, 2026-11-13).
Repro: in forge-et-be, set the AWS SDK’s HTTP client logger to DEBUG and store or read an attachment through ObjectStore.
Expected: no object key appears in a log line. Keys are opaque and never on the wire or in a log (FS-11 FR-M3; F36-a FR-C7’s conditions).
Actual: the SDK’s HTTP client logs each request line, which carries the object key. WI-2.1’s log-check test skips that line so that its own assertion can pass.
Fix: pin the SDK’s HTTP wire logger (and any logger that prints the request line) at INFO or above in application.yml for every profile, then remove WI-2.1’s skip so the log check covers it.
Regression test: the capture log check (CapturedLeadLogHygieneIT or WI-2.1’s equivalent) passes with no skipped line, with the root logger at DEBUG.
Resolved: 2026-11-15 — fixed in altrium-io/forge-et-be#436. The request line comes from the client’s exec chain (ex-… executing PUT /<bucket>/<key>), not from its wire or headers loggers, so application.yml now caps the whole org.apache.hc.client5 at INFO in its base document. In httpclient5 5.6.5 every line at INFO or above names the host only, so an incident loses nothing. AiLoggerCapsSpec pins the cap, and CaptureLogHygieneIT drops its isS3RequestLine skip, so its key check covers every application event. Before the fix, without the skip, CaptureLogHygieneIT failed on [DEBUG] ex-0000000004 executing PUT /et-test-attachments/capture/2/<uuid>.
Lesson: none — one-off: no other log-hygiene test skips a third-party line. The other suites leave out org.apache.hc.client5 only on the test thread, for the test’s own HTTP client, which is a different exclusion.
BUG-420 — Flaky tests: the custom-fields size-matrix states at 768px in WebKit time out in their arrange step in a heavily loaded browser gate (macOS)
Affects: F34-c and F1-d (forge-et-fe e2e/browser/size-matrix/size-matrix.spec.ts, the states account-detail-custom-fields-* and contacts-page-custom-fields-* at size-768x1024-webkit; the failing step is makeSecondPrimary in e2e/browser/size-matrix/arrange.ts, through chooseContactAction in e2e/browser/accounts/detail-support.ts:1773) · Found in: forge-et-fe#485’s selected browser run (the type-scale change), head 1f925866, 2026-11-13 · Severity: low — a test failure under load, no product defect seen · Phase: 1
Repro: pnpm e2e:gate --selected on forge-et-fe 1f925866, a backend at forge-et-be 3a486043, on a macOS laptop with other desktop apps open. The gate’s load line: busy share 0.07 at the start, peak 0.99, average 0.85; load per core peak 11.29 (8 cores); the busiest processes at the peak were three WebKit content processes, the Claude desktop app’s renderer and a virtual machine.
Expected: the states pass in a gate run whatever the machine’s load, or the load check stops the run first.
Actual: account-detail-custom-fields-refreshing failed both attempts, and six more custom-fields states at 768px WebKit passed only on retry: locator.click on account-contacts-row-actions-make-primary timed out at 30 s with “element is not stable” before the menu item settled. The same gate on the same branch one commit earlier (977e39e, whose only difference is the price import’s CSS) passed all of them. Run alone with retries off on 1f925866 (pnpm exec playwright test e2e/browser/size-matrix/size-matrix.spec.ts --project=size-768x1024-webkit -g custom-fields --retries=0) all 55 pass. Two Chromium tests were flaky in the same run and passed alone: compose-schedule.spec.ts:385 (F20-g AC-29) and self-admin-next-request.spec.ts:117 (F5-d AC-24).
Fix: triage by .forge/test-strategy.md § Flaky Tests, steps 2–4: reproduce under forced load; if nothing reproduces it, close it wontfix as an overloaded machine with the load above. If it reproduces, make the arrange step wait for the row menu to finish opening before it clicks Make Primary.
Regression test: the states themselves, passing under the forced-load reproduction.
Resolved: 2026-11-16 — closed as wontfix (not reproduced), Tharinda decided. On a Mac whose load average was 18 to 43 on 8 cores (other sessions’ backend builds and browser runs), pnpm exec playwright test e2e/browser/size-matrix/size-matrix.spec.ts --project=size-768x1024-webkit -g custom-fields --retries=0, against a backend on its own database at forge-et-be 88309db5, passed every account-detail-custom-fields-* and contacts-page-custom-fields-* state with main’s helpers (forge-et-fe 70fc7709). The one failure in that run was custom-fields-edit, the Settings page’s state, whose page fixture’s setup timed out, which is not this bug. A wait for the row menu’s opening animation was tried and taken back out of forge-et-fe#553, since nothing showed it was needed. A recurrence is filed as a new bug that links here, with the run’s load.
Lesson: L-162 (a failure that appears only under load is reproduced before it is called flaky) — an instance where it was not reproduced, so nothing was changed.
BUG-419 — ViewsRow.test.tsx’s tooltip test fails its keyboard half whenever the whole file runs (Windows)
Affects: F32-d (forge-et-fe src/ui/filters/ViewsRow.test.tsx) · Found in: Nanndish’s Windows laptop while finishing forge-et-fe#493 (UI polish round 2), on main too; reported in the #493 handover, 2026-11-13 · Severity: low — one Vitest case on Windows only; the tooltip itself works
Repro: on Windows, in forge-et-fe: corepack pnpm exec vitest run src/ui/filters/ViewsRow.test.tsx.
Expected: every case passes, as on macOS (50 of 50 on 2026-11-13 at forge-et-fe 495fdbb6).
Actual: “shows a tab’s full name and mark in its tooltip, on hover and on keyboard focus” fails at line 234, the keyboard-focus half, every time the whole file runs; it passes alone or with only its own describe block. A Windows developer’s pnpm test:record fails whenever its selection includes this file.
Fix: find what an earlier case in the file leaves behind that the keyboard half depends on (an open tooltip, a pending timer, focus or the tooltip provider’s delay state) and reset it per test, or make the keyboard half wait for the tooltip’s own open state rather than a fixed timing; prove it on Windows.
Regression test: the whole file passes on Windows, three runs in a row.
Resolved: 2026-11-13 — closed as wontfix (not reproducible), Dinil decided. On a second Windows machine the whole file passed 50 of 50 in four runs, and in a run with every core loaded (32 busy processes on 16 cores, the file taking 70 seconds against 6) it passed too. The tooltip itself works and only this test case was affected, so no change was made. A recurrence is filed as a new bug that links here, with the Vitest failure message and DOM snapshot, the machine’s load, and the Windows and Node versions.
Lesson: L-162 (a failure that appears only under load is reproduced before it is called flaky) — an instance where it was not reproduced, so nothing was changed.
BUG-418 — Scheduled send: 3 of 10 emails due in the same minute got no answer from Microsoft (2 failed, 1 uncertain)
Affects: F20-g (forge-et-be ScheduledEmailSweep, EmailSendService; F20-g-WI-1.4 Decision 15) · Found in: the F20-g-WI-3.1 run on dev, 2026-11-12 (dev backend at image build-6edbbdd5, F20-g wave 1) · Severity: high — AC-31 requires each scheduled email to arrive within 2 minutes of its send_at; three of ten never arrived, with deferred=0 · Phase: 1
Repro: on dev, signed in as CRM01 (Microsoft 365 mailbox, send rate-limited like any connection), schedule ten untracked NEW emails to one recipient for the same whole minute (POST /api/v1/email/scheduled, ten requests, all 201; the run’s console helper burst). At the minute the sweep sends them. The run sent ten for 2026-11-12T10:17:00Z.
Expected: all ten arrive within 120 seconds of send_at, or any that the sweep cannot send in that minute are counted deferred and sent by the next sweep (FR-L5; WI-1.4 Decision 15 states the sweep meets NFR-3 for about 16–32 emails due in one minute).
Actual: the sweep at 10:17:04Z logged [due=10, claimed=10, sent=7, failed=2, uncertain=1, deferred=0, settled=0, unfinished=0, dismissed=0]. Seven arrived at the recipient, 10 to 17 seconds after send_at (sends 42, 43, 45, 46, 47, 48 and 51, all answered SENT between 10:17:10 and 10:17:12). The other three got no answer from Graph (httpStatus=0, errorCode=NO_ANSWER) at 10:17:19.6, 10:17:21.8 and 10:17:22.2: sends 44 and 49 failed at step prepare (scheduled emails 5 and 10, FAILED), and send 50 at step send (UNCERTAIN). None of the three reached the recipient (checked in the recipient’s inbox and Spam). The backend log for the same seconds shows the connection’s mailbox reads being throttled and refused (Mailbox Read Throttled, Mailbox Read Refused … token bucket empty, Mailbox Ingest Deferred … rate limit) as the seven sent messages were ingested into the CRM, and Mailbox Read Unavailable … concurrent reads late timeouts afterwards. The cause is not proven: the log shows these happening in the same seconds, not that one caused the other. The dead-letter queue stayed at 0.
Fix: first read the cause from the code and the log (.forge/plans/F20-g/F20-g-WI-3.1-record.md R-9 has the sweep and backend lines). Candidates: the sweep’s up-to-4 concurrent sends and the connection’s mailbox read budget share one rate guard, so the ingest of the first sends’ copies starves the later sends; or Graph’s per-connection concurrency limit. Whichever it is, a row that failed at step prepare never reached Graph’s send and is safe to retry on the next sweep, so it should be rescheduled, not left FAILED; an UNCERTAIN row (step send) is not safe to resend and keeps F20-c’s handling, but the sweep should send fewer rows at once when Graph stops answering. Settle the design in a short plan before code (WI-1.4’s Decision 15 is the one to revise, and WI-2.1’s queue timeout if it must rise).
Regression test: a sweep test with ten due rows and a provider that stops answering for the third, ninth and tenth calls: no row ends FAILED for a prepare-step timeout, all ten are sent by the following sweep, and the UNCERTAIN one is not resent; and the F20-g-WI-3.1 burst (R-9) repeated on dev, all ten within 120 seconds.
Resolved: 2026-11-16 — closed wontfix, decided by Dinil: the cause was never found and no code was changed. The same burst repeated on dev on 2026-11-16 (F20-g-WI-3.1 record R-9, attempt 2) sent all ten emails in one sweep of about 10 seconds, with no failure and nothing in the dead-letter queue. If a burst fails again, file a new bug with the new evidence.
Lesson: none — one-off: not reproduced on a repeat with no code change, and the cause is unknown, so no pattern can be named.
BUG-416 — The dev smoke account’s role holds email.manage, so pnpm e2e:smoke stops at its key check
Affects: F38-g (the smoke run and its account on dev) · Found in: the smoke run after the dev deploy of forge-et-fe fd888f07 (forge-et-fe#523, BUG-406’s fix), 2026-11-13 · Severity: medium — the smoke run after each dev deploy cannot run its checks until this is fixed, so dev deploys go unchecked; the product is not affected · Phase: 1
Repro: pnpm e2e:smoke from forge-et-fe main at fd888f07, signed in as CRM05@empire.lk. Check 1 (sign-in) passes; check 2 fails with “Key check failed, so no other check runs: the account holds keys that do not end in .view or .all: email.manage”, and checks 3–9 do not run. The run lists the role as smoke_test_view_only (Smoke test (view only)) with six keys: account.all, account.view, buying_standard.view, email.manage, lead.all, lead.view.
Expected: the role holds exactly the five keys F38-g set up on 2026-10-21 (environments-and-identity.md §5, Smoke account (F38-g); .forge/plans/F38-g/F38-g-WI-1.1-plan.md): account.all, account.view, buying_standard.view, lead.all, lead.view. The key check exists so the smoke account can write nothing on dev.
Actual: the role also holds email.manage. No migration grants it to an existing role (git grep email.manage over forge-et-be’s migrations on origin/main: V28 renamed mailbox.read to it, V46 and V57 only mention it), and no harness record asks for it, so it was most likely added through Settings › Roles on dev after 2026-10-21. CRM05 has a working Outlook mailbox (environments-and-identity.md §5) and has been used for email tests, which may be why. The audit trail (AD #36) records the role change.
Fix: remove email.manage from Smoke test (view only) on dev (Settings › Roles, a holder of settings.roles.manage), then run pnpm e2e:smoke again. Removing the key releases CRM05’s mailbox connection if it has one (F20-i FR-20, reason PERMISSION_REMOVED), so first ask whoever added it whether they still need CRM05 for email; an email test on dev uses another account (CRM01–CRM04), since CRM05 is reserved for the smoke run.
Regression test: the smoke run’s own key check, which caught it; pnpm e2e:smoke green after the change.
Resolved: 2026-11-13 — fixed on dev by a data change, so there is no PR: Tharinda removed email.manage from Smoke test (view only) in Settings › Roles. The next pnpm e2e:smoke listed the role with its five keys (account.all, account.view, buying_standard.view, lead.all, lead.view) and passed check 2, the key check, and every other check but check 8, which is BUG-417.
Lesson: none — one-off: a hand edit to the role of an account reserved for the smoke run; the run’s key check caught it as it was built to (F38-g).
BUG-417 — Smoke check 8 expects the SMOKE standard to have exactly two prices, but a carried-forward round adds more
Affects: F38-g (forge-et-fe e2e/smoke/checks.smoke.ts, check 8 Prices — SMOKE standard as of each SMOKE week) · Found in: the smoke run on dev after BUG-416’s fix, 2026-11-13 (dev at forge-et-fe fd888f07) · Severity: medium — the smoke run after each dev deploy fails for as long as dev holds a later SMOKE price, so a real failure in check 8 would not be seen; the product is not affected · Phase: 1
Repro: pnpm e2e:smoke against dev, signed in as CRM05@empire.lk. Check 8 reads GET /api/v1/buying-standards/{id}/prices?sort=validFrom,asc for the SMOKE standard and fails: “the SMOKE standard has exactly its two published prices, oldest first”, expected [4.11, 4.22], received [4.11, 4.22, 4.22]. Checks 1–7 and 9 pass.
Expected: check 8 proves what the smoke run promises (F38-g): the SMOKE standard’s two published prices and the price as of each SMOKE week, answers that never change once published (AD #14).
Actual: the check also requires that no later price exists. Carry forward (F26-g, forge-et-be CarryForwardService) drafts the current price of every active standard a round does not price, the SMOKE standard included, and publishing that round (F26-f) makes the draft a published price from the sale’s week. So each round published on dev after a carry forward gives the SMOKE standard one more 4.22 price, and the check fails from then on. The history endpoint returns only published rows (BuyingStandardPriceRepository), and no pricing change since 2026-11-04 touches publishing, so the third row is such a price. Its validFrom was not read, since the smoke account cannot read dev’s audit trail; the cause is inferred from the code.
Fix: check 8 asserts only what AD #14 keeps fixed: the first two prices, oldest first, are 4.11 and 4.22, the first ends where the second starts, and the price as of the first’s start, the day before the second’s start, and the second’s start. Any later price is allowed. A later price ends the second, so the check stops requiring the second to be open-ended, if it does. Excluding the SMOKE standard from carry forward is not the fix: the smoke run should see the product as it works.
Regression test: the smoke run on dev, green with dev’s three SMOKE prices; and, if the smoke helpers have unit tests, a case with three prices.
Resolved: 2026-11-13 — fixed in forge-et-fe#525, as the Fix above describes. The history check moved to e2e/smoke/smokePrices.ts (readSmokePrices): the first two prices, oldest first, are 4.11 then 4.22 and the first ends where the second starts; later prices are allowed; check 8’s three as-of reads are unchanged. src/test/scripts/smokePrices.test.ts covers the two prices, later prices (a carried-forward 4.22 and a 4.30), an empty or one-price history, either price wrong, and a gap. pnpm e2e:smoke against dev: from main (fd888f07) check 8 failed on [4.11, 4.22, 4.22]; from the fix (e1843be0) all nine checks passed. The F38-g WI-1.1 plan’s contract C-3 and its Risks entry carry a dated note, and e2e/README.md says later SMOKE prices are allowed.
Lesson: L-292 (new) — a check on shared, live data asserts only what nothing in the product can change.
BUG-415 — Flaky test: account-detail.spec.ts AC-8 times out settling the account page in a heavily loaded browser gate (macOS)
Affects: F1-d (forge-et-fe e2e/browser/accounts/account-detail.spec.ts, “AC-8 (smoke): an https website is a new-tab link…”, through settleAccountPage in e2e/browser/accounts/detail-support.ts) · Found in: forge-et-fe#523’s selected browser run (BUG-406’s fix), head a8658f22, 2026-11-13 · Severity: low — a test failure, no product defect seen · Phase: 1
Repro: ET_BACKEND_PORT=18080 pnpm e2e:gate --selected on forge-et-fe a8658f22, a backend at forge-et-be 9581cb58, started 00:00 Colombo time with other sessions’ records and gates queued on the machine. The run’s .browser-gate/result.json load: busy share 0.36 at the start, peak 1, average 0.91; load per core 0.66 at the start, peak 7.37, average 2.98 (8 cores); the busiest processes at the peak were the gate’s own WebKit content processes, WindowServer and a gitleaks scan. The test failed on both attempts, each at a different 5-second wait: first follow-ups-panel-loading still present (detail-support.ts:2085, settleAccountPage), then on the retry the page’s h1 not found. The rest of the run: 5,636 passed, 1 flaky (the size-matrix state account-detail-contact-gone at 768×1024 WebKit, passed on its retry).
Expected: the test passes in a gate run whatever the machine’s load, or the load check stops the run first.
Actual: it fails under that load. Run alone with retries off it passes, on the branch (a8658f22) and with forge-et-fe#523’s four changed app files put back to main’s (account-detail.spec.ts:351 and the size-matrix state: 2 passed each time), so BUG-406’s change does not cause it.
Fix: triage by .forge/test-strategy.md § Flaky Tests, steps 2–4: reproduce under forced load (one busy loop per core); if nothing reproduces it, close it wontfix as an overloaded machine with the load above. If it reproduces, find which read settleAccountPage waits on lands late (the follow-ups panel’s first read).
Regression test: the test itself, passing under the forced-load reproduction.
Resolved: 2026-11-13 — closed as wontfix, an overload outcome under .forge/test-strategy.md § Flaky Tests step 4, the lead’s decision (Tharinda): forced load did not reproduce it, and the failing run had other work beside it (a gitleaks scan among the busiest processes at the peak, other sessions’ records and gates queued). AC-8 alone on forge-et-fe main 50849579, Chromium, retries off, one worker, a backend at forge-et-be df279a0e: 10 of 10 passed with one busy loop per core (1-minute load up to 11.7 on 8 cores); 10 of 10 with four per core and a 6× CDP CPU throttle (load up to 36.3); 10 of 10 with eight per core and the 6× throttle (load up to 107, about 13 per core against the failure’s 7.37); 5 of 5 with two per core, a 4× throttle and the follow-ups panel’s read (/api/v1/activities?accountId=…) held 4 seconds. The retry failed at a different wait (the page’s h1), which points to the whole page being slow rather than one read landing late. No change was made. A recurrence on a machine that passed the load check is filed as a new bug that links here.
Lesson: L-162 (a failure that appears only under load is reproduced before it is called flaky) — an instance where forced load did not reproduce it, so nothing was changed; added to its source list, with the order the reproduction must keep with the heavy-run lock (test-strategy.md § Flaky Tests step 2, Rev 38).
BUG-414 — pnpm e2e:gate --selected runs the whole browser suite when the change selects nothing
Affects: F40 (the browser gate’s --selected mode, forge-et-fe and forge-et-pwa scripts/browser-gate/select.mjs) · Found in: forge-et-fe#511’s selected run, 2026-11-12 · Severity: low · Phase: 1
Repro: on a frontend branch whose change touches no gate input (forge-et-fe#511 changes only .github/workflows/pr-validation.yml), run pnpm e2e:gate --selected. selectGate returns { mode: 'full', why: 'no gate input changed since <merge base>' } (select.mjs:389-390 on 2026-11-12), and the gate starts all 5,635 tests.
Expected: R-1 (.forge/design/harness/fewer-nightly-reds.md, PRD Rev 93) asks for the run only “when its change selects at least one spec or size-matrix state”, and gives Selected browser run: none selected for the rest. The gate should print that line and stop, without needing a backend.
Actual: an hour or more of browser tests on a change that no browser test reads, holding the machine’s heavy-run lock, unless the developer stops it. Before R-1 the gate ran only for a PR into release/* (always --full), so the fallback had no cost.
Fix: in select.mjs, return a none selection when no gate input changed; have run.mjs print Selected browser run: none selected and exit 0 before the backend probe and the lock. forge-et-pwa’s select.mjs has the same fallback (line 390 on 2026-11-12) and gets the same change.
Regression test: a select.mjs unit test where the change touches only a workflow file, expecting the none selection, and a run.mjs test that such a run takes no lock and needs no backend.
Resolved: 2026-11-12 — fixed in forge-et-fe#520, copied byte for byte to forge-et-pwa#123. selectGate answers { mode: 'none' } when the change touches no gate input by touches-gate-input’s own rule (the dependency-only exception included), before the no-feature-map full run, so the capture app gets it too; runGate chooses before the machine’s lock and, on none, prints Selected browser run: none selected and exits 0 without the lock, the backend probe or a result file, choosing again if HEAD moved during the wait. The review of forge-et-fe#520 widened it to the same symptom by two other paths: a branch that needs a full run (a wave branch) now also answers none when nothing is selected, and forge-et-fe’s midnight start check (gate-start-check.mjs, BUG-268) no longer refuses such a run. New tests in browserGateSelect.test.ts (a workflow-only, a Vitest-only and a cleared dependency-only change, with and without a map), browserGateRun.test.ts (no lock, no backend, no Playwright, the earlier result kept; HEAD moving during the wait), browserGateCli.test.ts (a wave branch) and gateStartCheck.test.ts (23:50) failed before the fix. One side effect: without a map, --selected now reads the merge base first, so an unfetched origin/main is refused instead of running in full.
Lesson: none — one-off: R-1 changed when the gate is run and named the line a run with nothing selected prints, but no code change went with it; the fallback had been written when only --full runs on release/* PRs used the gate, where it cost nothing.
BUG-412 — While Disconnect runs, the Disconnect question’s Cancel stays available and a press is silently ignored
Affects: F20-a (forge-et-fe src/features/email/ConnectedPanel.tsx around line 240, the disconnect question) · Found in: the delta review of forge-et-fe#499 (F20-g wave 1), 2026-11-11, in code that PR did not change · Severity: low — keyboard and screen-reader users are told nothing; the disconnect itself is unaffected · Phase: 1
Repro (from the code): connect a mailbox, press Disconnect, confirm, and press the question’s Cancel while the request is held. Esc is refused while busy, but Cancel (AlertDialogCancel) has no busy state, so it looks available and a press does nothing.
Expected: Cancel is aria-disabled while the request runs, and a press keeps the question open (lesson L-277).
Actual: Cancel reads as available; its press is swallowed with no feedback.
Fix: pass disabled={busy} to AlertDialogCancel, whose disabled prop forge-et-fe#499 added (rendered as aria-disabled, press refused). Fixed in its own commit on forge-et-fe#499, under the overnight rule for a small, test-backed defect in old code.
Regression test: a ConnectedPanel test that holds the disconnect request and asserts Cancel is aria-disabled and a press keeps the question open.
Resolved: 2026-11-11 — fixed in forge-et-fe#499 (merged 0b08344d), commit ae859868: AlertDialogCancel disabled={busy}; the new MailboxRoute.test.tsx case failed before the fix with aria-disabled="false".
Lesson: L-277 — a second instance of its cause (a control that refuses its action while busy must say so with aria-disabled); no new lesson.
Filed as BUG-404 on the F20-g spec branch on 2026-11-11 and renumbered on 2026-11-12, first to BUG-410 and then to BUG-412, because main took BUG-404, then BUG-410 and BUG-411, while the spec PR was open (.claude/rules/id-numbering.md rule 5). The forge-et-fe#499 review record names it BUG-404.
BUG-411 — Show Suggestions opens a picker’s list but leaves focus in another field, so the keyboard and typing do not reach the list
Affects: F5-c (the Create User and user editor’s Manager picker), F1-e (the tag pickers) — every picker built on forge-et-fe src/ui/tags/AddTagPicker.tsx. Found by frontend-reviewer’s delta review of forge-et-fe#508 on 2026-11-12, in code that PR did not change.
Repro: in Create User, type in Email (or open the dialog, which since forge-et-fe#508 starts on Email), then press the Manager picker’s Show Suggestions with the mouse. The trigger cancels its mousedown to keep focus in the input, as its doc says “focus never leaves the input”, but focus was never in the input: the list opens while focus stays in Email. Arrow keys and typing go to Email, and typing does not close the list.
Expected: pressing Show Suggestions puts focus in the picker’s input and opens the list, so the keyboard drives the list.
Actual: the list opens with focus left in the other field.
Fix: in the trigger’s press, focus the input before toggling the list, as react-aria’s useComboBox does on press start; keep the mousedown cancel so a press with focus already in the input does not blur it.
Regression test: a unit test that focuses another field, presses the trigger and asserts the input has focus and the list is open. e2e/browser/users/see-and-add.spec.ts moves focus into the picker itself before filling Email (forge-et-fe#508); it can drop that step once this is fixed.
Resolved: 2026-11-12 — fixed in forge-et-fe#518. The trigger’s press focuses the input, then toggles the list; the mousedown cancel stays, so a press with focus already in the input keeps it there. The same trigger, copied, was in AttachAccountPicker (F1-f, Show Accounts) and RecordPicker (F3-a, Show accounts and the rest), and both get the same change. A unit row each in SearchSelect.test.tsx, AttachAccountPicker.test.tsx and RecordPicker.test.tsx focuses another field, presses the trigger and asserts the input has focus with the list open; each failed before the fix. see-and-add.spec.ts now asserts the Manager picker has focus after Show Suggestions, in place of moving focus there itself.
Lesson: none — one-off: three copies of one trigger shared the assumption that focus was already in the input, and all three are fixed together; nothing else in forge-et-fe cancels a mousedown to hold focus (git grep 'onMouseDown={(event) => event.preventDefault()}' src, which now finds only these triggers and the options’ rows).
BUG-410 — Tab pressed while the compose dialog’s AI Draft button loads moves focus to the dialog, not the next field
Affects: F11-h (the AI Draft button’s place), F20-c (the compose dialog). Found by forge-et-fe#508’s selected browser run on 2026-11-12, and confirmed on main e888816f in the same code path.
Repro: open Compose Email from an account (a draft context, so the AI Draft place renders) and press Tab within about half a second, while ComposeAiDraftPlaceholder still stands in for the button. When To loses focus, React re-renders and swaps the stand-in for the button; that removal lands while no element has focus, and Radix FocusScope’s observer of removed nodes then focuses the dialog element. The browser’s move to Cc is lost. Traced on forge-et-fe#508’s branch: focusout To (related target Cc), then focus() on the dialog from FocusScope’s handleMutations, and no focusin on Cc. On main the dialog’s own effect moves focus to To a render later, so the swap has usually happened before a Tab; forge-et-fe#508 puts focus on To during the dialog’s opening, so the browser tests, which press Tab at once, hit it every time.
Expected: Tab from To moves to Cc whenever it is pressed.
Actual: focus goes to the dialog element; the next Tab goes to the ✕.
Fix: keep the stand-in’s node in the page when the button arrives (render the button into the same element rather than replacing it), or swap it outside the blur, so no node is removed while focus moves. The browser tests wait for the swap in settleComposeDialog (forge-et-fe#508) and can stop waiting once this is fixed.
Regression test: a browser test that presses Tab from To before the AI Draft chunk has arrived (hold the chunk’s request with a route) and asserts focus is on Cc.
Resolved: 2026-11-12 — closed wontfix as a duplicate of BUG-402, decided by Tharinda. BUG-402 is the same defect, traced the same way (Radix FocusScope’s removed-node handling during the AI Draft stand-in’s swap), and forge-et-fe#505 had already fixed it on main by loading the button’s chunk with the page. This bug was filed from forge-et-fe#508’s branch, cut before #505 merged. settleComposeDialog’s comment names BUG-402 (forge-et-fe#518), and its wait for the stand-in stays, for a dialog opened before the chunk’s load has answered.
Lesson: none — one-off: a duplicate filed from a branch cut before the fix merged; the bug search before filing read bugs.md on that branch’s base rather than origin/main.
BUG-409 — The WebKit notes keyboard test reaches the delete question’s Delete backwards, though Delete now comes after Keep
Affects: F1-e (the account page’s Notes section; forge-et-fe e2e/browser/accounts/notes-tags-keyboard-webkit.spec.ts:96, “AC-29 (WebKit): deleting the middle note moves focus to the next note’s ⋯ menu, the last note to the Add a note line”) · Found in: the nightly full run on 2026-11-11 and 2026-11-12 (forge-et-fe#484) · Severity: low — test only; the product does what design-system §5 says · Phase: 1
Repro: in forge-et-fe at fe35c3d, with a backend at forge-et-be bb9fe28 on 18080: E2E_GATE=1 ET_BACKEND_PORT=18080 pnpm exec playwright test e2e/browser/accounts/notes-tags-keyboard-webkit.spec.ts --retries=0 --workers=1. The test fails every run, alone, with the CPU about 40% busy. The night of 2026-11-10 passed it; it first failed the night of 2026-11-11, after forge-et-fe#478 merged.
Expected: the test reaches Delete in the note’s delete question from Keep, where the question opens.
Actual: forge-et-fe#478 put the safe answer first (Keep, then Delete, design-system §5). The test still pressed the focus key backwards (Shift+Tab) from Keep, so focus left the question and expect(noteDeleteConfirm(item)).toBeFocused() failed. F1-e FR-P2 names both buttons and sets no order.
Fix: test only. Press the focus key forward from Keep. No other browser test moves backwards to a delete question’s Delete (grep over e2e/).
Regression test: the test itself. At fe35c3d it failed; after the fix it passed alone with retries off (2.4s).
Resolved: 2026-11-12 — fixed in forge-et-fe#517, test only, as the Fix above describes.
Lesson: none — one-off: the pattern (a product change the browser tests were not updated for) is already covered by frontend-reviewer dimension 32 (.forge/design/harness/fewer-nightly-reds.md R-3), which was added after forge-et-fe#478 merged.
BUG-408 — The record dialog’s ✕ sits 2px above its title’s centre, so the BUG-156 browser test fails
Affects: FS-3, F5-c (forge-et-fe src/ui/overlays/RecordEditDialog.tsx and .module.css, the header of every dialog built on it, and the three dialogs that copy that header: src/features/settings/SettingsDialog, src/features/notesAndTags/NotesDialog, src/features/priceEmail/PriceEmailDialog; e2e/browser/users/see-and-add.spec.ts:409, “the title’s line centres on the ✕”) · Found in: the nightly full run on 2026-11-11 and 2026-11-12 (forge-et-fe#484) · Severity: low — a 2px vertical offset in a dialog header · Phase: 1
Repro: in forge-et-fe at fe35c3d, with a backend at forge-et-be bb9fe28 on 18080: E2E_GATE=1 ET_BACKEND_PORT=18080 pnpm exec playwright test e2e/browser/users/see-and-add.spec.ts:409 --retries=0 --workers=1 on macOS. It fails every run with the same figures, both nights and alone: the title text’s centre 101.5625px, the ✕'s 99.609375px. The test came in with dd662c4a (forge-et-fe#476) and has not passed in a nightly run.
Expected: the title’s text line and the 36px ✕ share a centre to within 1px (design-system §5 Modal, BUG-156).
Actual: the text’s box, measured with a Range over the title’s text, sits 1.95px lower than the ✕'s centre.
Cause (measured on macOS, Chromium, at fe35c3d): the title row (.title) is 36px tall, the value of --et-layout-control-h-compact, and its text (20px box) is centred in it. The ✕ renders 32px tall, and the header aligns it to the top (align-items: flex-start), so its centre sits 2px above the title row’s. The text is not low; the ✕ is short.
Fix: the ✕ keeps its 32px size and sits in a row of the title row’s height (.closeRow, min-block-size: var(--et-layout-control-h-compact)), centred in it, so it shares the title’s centre with or without a description under the title. Every dialog built on RecordEditDialog gets the change, and the three dialogs that copy its header (Settings › Tags, notes and tags, price email) get the same .closeRow; the current prototype already draws the ✕ centred on the title. Moving the four copies of the header into one src/ui part is a follow-up in .forge/owed.md.
Regression test: the BUG-156 browser test. It failed before the fix (title 101.56px, ✕ 99.61px) and passed after (see-and-add.spec.ts, Chromium, retries off: 20 passed).
Resolved: 2026-11-12 — fixed in forge-et-fe#517, as the Fix above describes.
Lesson: none — one-off: the ✕ became the 32px icon-sm size in 8c7151bd (F5-d) while the title row kept the 36px compact control height, and BUG-156’s test, added later, was the first check of their alignment.
BUG-407 — The F20-g scheduling keyboard tests type the date month first, so they fail where the region orders dates year first (macOS)
Affects: F20-g (forge-et-fe e2e/browser/email/compose-schedule.spec.ts, “AC-29 (keyboard): Tab to the box…”, and e2e/browser/email/tracking-scheduling-smoke.spec.ts, “AC-29: keyboard only — schedule from the account header…”) · Found in: the nightly full run on 2026-11-11 and 2026-11-12 (forge-et-fe#484) · Severity: low — test only; the Send Later section works · Phase: 1
Repro: macOS with the language English and the region Sri Lanka (defaults read -g AppleLocale prints en_US@rg=lkzzzz), forge-et-fe at fe35c3d, a backend at forge-et-be bb9fe28 on 18080: E2E_GATE=1 ET_BACKEND_PORT=18080 pnpm exec playwright test e2e/browser/email/compose-schedule.spec.ts:427 e2e/browser/email/tracking-scheduling-smoke.spec.ts:1199 --retries=0 --workers=1. Both fail every run.
Expected: the keyboard walk enters tomorrow’s date (and, in the smoke test, a date three hours ahead) into the Send Later date input on any developer’s machine.
Actual: each test’s typeDate typed month, day, year. Chromium on macOS lays the input’s parts out from the system region, so the year came first and the input held 1113-02-26 in place of 2026-11-13.
Fix: test only. Each test reads the order once with dateOrder and types the date with typeDate from e2e/browser/custom-fields/values-support.ts, as BUG-122 did for the import walk; the second dialog’s walk now also asserts the date it typed.
Regression test: the two tests. On this Mac they failed before the fix and passed after (-g 'AC-29' over both specs: 17 passed).
Resolved: 2026-11-12 — fixed in forge-et-fe#517, test only, as the Fix above describes.
Lesson: L-143, a third instance of its date-order point (after BUG-122 and F34-c finding F-004). The lesson already said to read the order; nothing checked for it, so frontend-reviewer and test-strategy.md § Browsers now name it.
BUG-406 — In WebKit a dialog opened by a click gives focus to <body> on close unless its caller names returnFocusTo (macOS)
Affects: F5-d (which built RecordEditDialog), and the callers’ features, among them F5-c, F11-b and F26-c (forge-et-fe src/ui/overlays/RecordEditDialog.tsx and src/features/pricing/PricingDialog.tsx, used by every edit and confirmation dialog) · Found in: BUG-397’s fix (forge-et-fe#515), 2026-11-12 · Severity: low — a mouse user in Safari loses their place after closing a dialog; keyboard users are not affected, because a key press focuses the button · Phase: 1
Repro: in WebKit (Safari, or Playwright’s webkit), click a button that opens a RecordEditDialog or PricingDialog whose caller passes no returnFocusTo, then close the dialog: document.activeElement is <body>. In Chromium the same click focuses the button and focus comes back to it. BUG-397 was one such dialog. The callers without returnFocusTo in the dialog’s own file, as of forge-et-fe fe35c3df (git grep -L returnFocusTo over the files that render either dialog), include EditAccountDialog, EditContactDialog, ComposeEmailDialog, InquiryExtractionDialog, LeadFormDialog, EditSaleDatesDialog, NewSaleDialog, CreateRoleModal, SharingPanel, ChangeManagerDialog and CreateUserModal; some may receive it from their own caller, so each needs checking.
Expected: focus goes back to the control that opened the dialog, in every engine (design-system §5 Modal; WCAG 2.1 SC 2.4.3).
Actual: RecordEditDialog records document.activeElement when the dialog opens as the control to return to; in WebKit a click leaves focus on <body>, so there is no such control and, without returnFocusTo, focus stays on <body>.
Fix: decide between (a) RecordEditDialog recording the opener from the pointer event as well (for example the last pointerdown target that is a button or link, when document.activeElement is <body>), which fixes every caller at once, and (b) every caller naming its trigger in returnFocusTo. (a) is preferred if frontend-reviewer agrees it is safe with the menus whose items leave the page.
Regression test: a RecordEditDialog test that opens the dialog with focus on <body> after a pointer press on the trigger, closes it, and expects focus on the trigger; and the WebKit keyboard specs stay green.
Resolved: 2026-11-12 — fixed in forge-et-fe#523, by option (a), which fixes every caller at once. A new src/ui/overlays/openerControl.ts tracks pointer presses from start-up (installOpenerTracking() in src/main.tsx). WebKit gives a click’s focus to <body>, or to the nearest focusable ancestor of the button (the bulk bar’s tabindex section, a dialog’s scroll body); in either case openerControl() names the pressed control (a <label>'s control for a press on the label’s text), and otherwise the focused element. The focus that keeps the press is only the press’s own: the first focus between the press and its compatibility mouseup, which for a mouse click comes before pointerup and for a touch tap after it. Any other focus forgets the press, as does a cancelled press (a touch that became a scroll). RecordEditDialog (so PricingDialog), ConfirmDialog and ComposeEmailProvider read their opener through it. A recorded opener outside <main> and any open dialog (a menu item) or gone from the page still falls back to returnFocusTo, and so does one that takes no focus by the time the dialog closes. frontend-reviewer found the ancestor case, the late install and the compose opener on the first round, and on the second that a touch tap in WebKit moves focus after pointerup. Regression tests: openerControl.test.tsx (mouse and touch event orders, labels, a cancelled press), a RecordEditDialogFrame.test.tsx block (body, tabindex section, focus moved, press outside any control, menu item, removed and hidden openers) and a ConfirmDialog.test.tsx row, each fix row failing on the code before it; and e2e/browser/accounts/dialog-focus-return-webkit.spec.ts, which in WebKit opens Edit Account by mouse (closed by Cancel, ✕ and Esc) and the bulk bar’s Change Owner by mouse and by touch tap, and expects focus back on the button. It fails on main’s code.
Lesson: L-287 — this is the wider gap it named; its second point now records the decision (the dialogs read the pressed control), and that WebKit can focus a focusable ancestor of the button rather than <body>.
BUG-405 — A Pipedrive value holding U+0000 fails the migration’s write instead of being cleaned or reported
Affects: F4-b (the Pipedrive migration of leads, accounts and contacts). Found by backend-reviewer on forge-et-be#409 (BUG-403’s fix) on 2026-11-11.
Repro: a Pipedrive export whose organisation, person or lead text holds U+0000. BUG-403 added the control-character rule to the REST request records only. LeadFieldRules.check (shared by the REST create, F36-a’s submit and MigrationLeadWriter), AccountMigrationService and ContactMigrationService check lengths but not control characters, so the value reaches PostgreSQL, which refuses U+0000 (L-280). Not yet run against a real export: whoever takes this confirms it with a fixture first.
Expected: the run finishes, and each such record is either written with the NUL removed or listed in the migration report as left out, with the reason.
Actual: the record’s write fails at the database.
Fix: decide with the F4-b owner whether the migration strips U+0000 (and other control characters, for one-line fields) or reports the record; apply it in the migration writers, not in LeadFieldRules, whose other callers already refuse through the request records. F36-a’s submit path gets the same decision when F36-a is built.
Regression test: a migration fixture row holding U+0000 in a name, asserting the chosen outcome.
Resolved: 2026-11-12 — fixed in forge-et-be#413. The repro as filed cannot happen: XML 1.0 cannot hold U+0000 (raw or as �) or any other C0 control but tab, LF and CR, so an export holding one is refused as NOT_A_WORKBOOK at the dry run before any row is read (a 400, pinned by an ExportReaderSpec row). What a cell can bring is a tab, a line break, a C1 control or U+2028, and each was written as it was into a one-line column the CRM’s own requests refuse (BUG-403). The lead’s decisions (Tharinda, 2026-11-12): clean and import, never leave the record out; and read a C1 control as its Windows-1252 character (U+0092 is ’), the usual cause in real data. ExportReader.readRow reads every declared column through PlainText.repairWindows1252, then PlainText.toOneLine (a run of spaces holding a break or a removed control becomes one space, any other control character is removed, the ends are stripped), so the dry run and the write see the same text. Applied in ExportReader, upstream of the writers the Fix: line named, so the proposal digest covers the cleaned values; LeadFieldRules is unchanged. Regression tests: PlainTextSpec (the two methods), ExportReaderSpec (cleaning in shared and inline strings, and the U+0000 refusal) and PipedriveRunControlCharactersIT (a dry run and a run writing Synthetic Rcc O’Brien Co); each cleaning test failed on main’s reader. Left open, with owners in .forge/owed.md: F36-a’s submit path (Nanndish), and whether a Pipedrive export writes _xHHHH_ escapes, which the reader does not decode (Dinil).
Lesson: L-289 (new) — reproduce a defect found by reading code on an import path with a file in the real format first.
BUG-404 — Flaky test: OwnedRecordsList.test.tsx’s toast-focus case fails under a heavily loaded selected run (macOS)
Affects: F5-h, F32-c (forge-et-fe src/features/users/OwnedRecordsList.test.tsx, “after a pass that moved everything (a toast), moves focus to the list heading…”) · Found in: F32-d-WI-2.3’s finalize (pnpm test:selected on forge-et-fe 17ab89bb, 2026-11-10)
Repro: that run chose selected mode, ran 8588 tests and failed this one case, with the machine’s load average about 20 from other sessions’ browser runs. Run alone, the file passes 37 of 37. Neither F32-d wave 2 nor WI-2.3 changes src/features/users (git diff 5b2b2780 17ab89bb -- src/features/users is empty).
Expected: the case passes whatever the run’s load.
Actual: under heavy load it fails once; the failure’s message was not kept.
Fix: as BUG-382: hold the pass’s answer and release it inside act, so the renders it causes are flushed before findBy’s second starts. The case’s twin, “…when the confirming press focused nothing”, has the same shape and gets the same change.
Regression test: the file passes three runs in a row under one busy loop per core. The old file passes that too when run alone (10 of 10 runs of the two cases under 16 busy loops on 8 cores), so the timings in Resolved: are the evidence.
Resolved: 2026-11-12 — fixed in forge-et-fe#507 (1fb1e3be). No product defect. It failed again on forge-et-fe#504’s test record on 2026-11-12 (passed on its retry, the file taking 78 s with the machine at 13.7 load per core). Timed on macOS (8 cores) with the case alone, from the confirming press to the toast: 214–248 ms idle and 597–984 ms with 16 busy loops, against findByText’s 1 s default, so a full-suite run with other files beside it went over. With the answer held and released in act, the findBy left takes 12–101 ms under the same load.
Lesson: L-270 (another instance).
BUG-403 — A text field holding U+0000 (NUL) is accepted by the API and then refused by PostgreSQL, so the write answers 500
Affects: F1-a (account and contact writes) and F11-a (lead intake and edits): every text request field whose only check is a length. Found by backend-reviewer on forge-et-be#405 (F34-c wave 2) on 2026-11-11, which fixes the same gap for custom field text values (wave 2 contract Rev 2).
Repro: send a create or PATCH whose text member holds \u0000. Only the database half is proved so far: on Postgres 18, select length(chr(0)) gives null character not permitted and a UTF-8 0x00 byte gives invalid byte sequence for encoding "UTF8". The custom field text value showed the whole path (a 500, forge-et-be#405’s review); the other fields are not yet sent to the API, so whoever takes this bug confirms each one first.
Expected: 400 VALIDATION_FAILED, with an errors[] entry for the field.
Actual: the request passes validation and the flush error reaches ApiExceptionHandler’s catch-all, which answers 500. On forge-et-be main only InquiryExtractionRules refuses control characters (git grep -n isISOControl origin/main -- src/main).
Fix: one shared rule, used by every text request field, that refuses U+0000, and likely every control character, as F34-c’s custom text values do since contract Rev 2.
Regression test: a NUL row in each writer’s validation IT, failing before the fix.
Resolved: 2026-11-11 — fixed in forge-et-be#409. platform.text.PlainText holds one rule, the one F34-c’s custom text values use: a one-line field holds no control character (Unicode Cc) and no U+2028/U+2029; a multi-line field (a lead’s inquiryText, note bodies) may also hold a tab, a line feed and a carriage return. @OneLine and @MultiLine apply it to every text member of the account, contact and lead create and PATCH records, list elements included (preferredPackaging[0]), and the account, contact and lead note checks apply it to note bodies; each refusal is 400 VALIDATION_FAILED with an INVALID entry naming the field. Format characters (Cf) stay allowed, because a name in a right-to-left script may need them. The Pipedrive migration and the intake path write through services and are not covered by the request rule: they check lengths but not control characters, filed as BUG-405. Regression rows in AccountControllerIT (create and PATCH), ContactControllerIT, LeadControllerIT (create and PATCH), AccountNoteControllerIT (add and edit) and LeadNoteControllerIT send U+0000 or another control character and expect the 400; before the fix each reached the database (the bug’s repro). Text fields in other features (tags, email templates, sends) are not covered by this bug’s affects and keep their own checks.
Lesson: L-280 — this bug is that lesson’s second instance, already in its source list.
BUG-402 — Two compose keyboard tests in email-a11y.spec.ts fail on main: Tab never reaches the body, or Subject, within their step limit
Affects: F20-c (forge-et-fe e2e/browser/email/email-a11y.spec.ts:174, “AC-45 (keyboard), BUG-202: Escape with focus in the body closes an unchanged dialog…”, and :201, “AC-45 (keyboard): Send with To empty shows the error on To and moves focus there”) · Found in: the nightly full run of 2026-11-11 (forge-et-fe#484, “2 failing — e2e/browser/email/email-a11y.spec.ts”), and reproduced on 2026-11-11 by F20-g’s wave-1 frontend merge, on main 9d218cee and on feature/F20-g-wave-1 alike.
Repro: in forge-et-fe at origin/main 9d218cee, against a backend on forge-et-be main with a fresh database: ET_BACKEND_PORT=<port> E2E_GATE=1 pnpm exec playwright test --project=chromium --workers=1 --retries=0 e2e/browser/email/email-a11y.spec.ts. Two fail every time; ten pass.
Expected: Tab from the compose dialog’s fields reaches the body (the first test) and reaches Subject from To within 5 steps (the second).
Actual: “Tab reaches the control within N steps” fails in both: the tab order no longer reaches those controls within the limit.
Fix: load the AI Draft button’s chunk with the compose composition, and render the loaded button directly, choosing once per mount, so nothing in the dialog is swapped while a person tabs. The Rev 90 frame (e58e1a52) was not the cause.
Regression test: both tests pass on Chromium, and F20-h’s attachments-a11y.spec.ts:113 (AC-32 keyboard), which failed the same way.
Resolved: 2026-11-12 — fixed in forge-et-fe#505 (242f76c). Traced in Chromium: the first Tab out of To moved focus towards Cc, but the blur rendered (the field turned touched), and that render swapped the AI Draft button’s lazy-load stand-in for the button, removing a node while focus was on <body>; Radix FocusScope’s MutationObserver then focused the dialog container, so the next Tabs went Close, the body region, To. After the fix both email keyboard spec files pass (24 tests with setup), and the new composeEmailComposition.aiDraftPreloaded.test.tsx fails on the old code.
Lesson: L-284 (new, on main from forge-et-harness#1297).
BUG-401 — A subfolder’s item in the email folder picker fails axe’s label-content-name-mismatch: its visible text is not one run of its spoken name
Affects: F20-i (forge-et-fe src/features/email/FolderPicker.tsx, WI-2.4’s hidden “, in email-scheduled-picker and email-scheduled-picker-current, the first to open the picker, 2026-11-11 (WI-1.6 plan ## Notes item 7). Predates F20-g.
Repro: below bp/laptop, open the Email page’s folder picker with a folder tree that has a subfolder (the standard stub seed’s Ceylon under Buyers) and run axe on it.
Expected: each item’s visible text is one run inside its accessible name (WCAG 2.1 SC 2.5.3, Label in Name).
Actual: the item shows “Ceylon 2” and is named “Ceylon, in Buyers 2 unread”: the hidden “, in Buyers” sits between the name and the count.
Fix: put the hidden “, in expectPageClean at 768 × 1024 and 320px.
Resolved: 2026-11-12 — fixed in forge-et-fe#518. A subfolder’s item now reads “Ceylon 2 unread, in Buyers”: the hidden “, in FolderPicker.test.tsx asserts the new order. email-scheduled-picker and email-scheduled-picker-current use the standard tree again, with Ceylon under Buyers, so the size matrix’s axe check covers the subfolder item at every size; FLAT_FOLDER_TREE is removed.
Lesson: L-211 — another accessible name that breaks its visible label’s run, here by hidden text inserted between the visible parts rather than an aria-label replacing them; added to its sources.
BUG-400 — A custom field’s archived audit event keeps archivedAt to the 100 nanoseconds, but the stored value keeps microseconds, so CustomFieldAuditIT fails on Windows
Affects: F34-c (forge-et-be src/main/java/com/empireteas/customfield/service/CustomFieldDefinitionService.java around line 281; src/test/groovy/com/empireteas/customfield/CustomFieldAuditIT.groovy:119, “archived and restored: one row each…”) · Found in: F20-g WI-1.1’s full backend run (forge-et-be#399, head 4c89295b), 2026-11-10; reproduced on origin/main 6e39b716 · Severity: medium — a red test on main on Windows (and Linux, per BUG-288), which fails every Windows local build record until fixed · Phase: 1
Repro: on Windows, at origin/main 6e39b716: ./gradlew test --tests '*CustomFieldAuditIT*' — EXIT=1, 15 tests, 1 failed.
Expected: the archived event’s archivedAt equals the stored archived_at.
Actual: the event holds …890965100Z and the stored value …890965Z. CustomFieldDefinitionService sets archivedAt from clock.now() without cutting it to microseconds; Java’s clock on Windows has 100-nanosecond resolution and Postgres timestamptz keeps microseconds — BUG-288’s cause, in a value set by hand rather than by JPA auditing (MicrosecondDateTimeProvider covers only @CreatedDate and @LastModifiedDate).
Fix: cut the instant to microseconds where archivedAt (and any other hand-set timestamp in the customfield service) is set, as the time source platform/time or MicrosecondDateTimeProvider does; check the service’s other clock.now() uses.
Regression test: CustomFieldAuditIT passes on Windows. Fixed on forge-et-be#406 (F20-g wave 1, commit 914ec3c5, overnight 2026-11-11): archive() cuts clock.now() to microseconds, the only hand-set timestamp in the customfield package; the new CustomFieldArchiveTimeSpec (a fixed clock with sub-microsecond digits) failed 2 of 3 rows on the old code on any OS, and CustomFieldAuditIT failed before the fix and passes after on Windows.
Resolved: 2026-11-11 — fixed in forge-et-be#406 (merged 90b5031e), commit 914ec3c5; the wave’s full build record on Windows ran every test with none failing (14,677 tests).
Lesson: L-128 — another instance of BUG-288’s cause (a hand-set Instant written to timestamptz without cutting to microseconds); no new lesson.
BUG-399 — The CRM’s muted-text check and comments take WCAG’s large-text limits in points as pixels, so 18.66px text counts as large
Affects: F1-b and every CRM screen that sets text colour by size (forge-et-fe src/test/mutedTextBand.test.ts, which fails a CSS rule that puts --et-text-muted on normal-size text; the comments beside --et-text-muted in src/ui/styles/theme-empire.css, tokens.css and theme-neutral.css; design-system §2’s text/muted row). Found by frontend-reviewer on forge-et-fe#485 (the type-scale change, design-system Rev 109), 2026-11-10, by reading the code; that PR did not change these lines, so the fix is its own.
Repro (from the code): mutedTextBand.test.ts treats text as large at 18.66px, or at 14px bold, and the comments say “under 18.66px, or under 14px bold” is normal size.
Expected: WCAG 2.1’s large text (SC 1.4.3) is at least 18 point (24px), or at least 14 point (18.66px) bold. Below that, text needs 4.5:1, which Empire’s text/muted (4.09:1 on surface/card) does not reach.
Actual: points are read as pixels, so text from 18.66px to under 24px at normal weight, and from 14px to under 18.66px bold, counts as large. Under Rev 109’s scale that would let type/stat (20px) and type/subhead (18px) take text/muted. Nothing fails today: no rule puts muted text on those levels.
Fix: mutedTextBand.test.ts’s limits are 24px, and 18.66px at weight 700 or more (14pt is 18.666…px, taken as 18.66 as axe-core’s rounding does), in one isLargeText function; the three comments and design-system §2’s row (Rev 132) say the same.
Regression test: isLargeText checked at each limit (24px and 20px at 400; 18.66px, 18px and 14px at 700), and a synthetic rule putting --et-text-muted on the 20px stat level at its own weight 600, which the check must report with its exact message; 22 tests passed.
Resolved: 2026-11-17 — fixed in forge-et-fe#553 (merge 029affc1; Test record and selected browser run on d4f3352). forge-et-pwa#154 re-copies the changed tokens.css into the capture app: forge-et-fe#553 had no companion re-copy, so the shared-file drift check stopped the capture app’s nightly browser gate on its 2026-11-17 20:25 hand run (forge-et-pwa#140).
Lesson: an instance of L-190 (a file shared by copy makes its fix a two-repo change): the fix changed only a comment in tokens.css, which forge-et-pwa copies byte for byte, and no capture-app PR went with it. The unit confusion itself (WCAG’s points read as pixels) is one-off.
BUG-398 — Since F32-d the accounts Country list has no All option, so the BUG-386 keyboard test and the accounts-country-open page states fail
Affects: F32-d (list views on accounts) and F1-b (the account list). Found by the nightly full run on 2026-11-10 (forge-et-fe#484): the BUG-386 keyboard test in e2e/browser/accounts/accounts-a11y.spec.ts:229, and the size-matrix state accounts-country-open at seven sizes in Chromium and WebKit, 9 tests in all.
Repro: run them alone in gate mode against the night’s commits (see BUG-397’s repro). They fail every run. They pass at forge-et-fe#472’s merge (BUG-386, which added them) and fail from #470 (feature/F32-d-wave-1), the next merge.
Expected: the tests wait for getByRole('listbox').getByRole('option', { name: 'All', exact: true }) to take focus when the list opens.
Actual: the element is not found. Since #470 the Country filter is a checkbox list starting at “Afghanistan”, with no All option. The list itself opens and scrolls correctly (checked at 1440 × 900 on the local CRM).
Fix: decide whether F32-d meant to remove All. If it did, update the two tests to the checkbox list. Keep BUG-386’s point, that the list is reachable by keyboard and shows every option, and update the accounts-country-open page state. If it didn’t, restore the All option.
Regression test: the two existing tests, updated or kept.
Resolved: 2026-11-12 — fixed in forge-et-fe#515. F32-d removed All on purpose: FR-F2 makes Country a multi-value filter, the shared FilterMultiSelect (src/ui/filters), whose list is checkboxes where none ticked means All, so it has no All row. The tests follow the product. The BUG-386 keyboard test opens the Country filter with Enter, checks that the first box (Afghanistan) has focus and is in view, types Zim to reach the last box (Zimbabwe), Shift+Tab to Zambia and Sri to Sri Lanka, each focused and in view, then Escape gives focus back to the trigger; page.clock holds the typeahead’s time between keys. The accounts-country-open page state opens the same list with no box ticked and the first box focused. Before the fix the tests failed on the missing All option at forge-et-fe fe35c3df (the nightly run of 2026-11-12); after it, accounts-a11y.spec.ts passes in Chromium and the page state passes at every size in both engines. The Country list was also the proof behind the accepted axe finding on a shadcn Select’s scrolling list (SELECT_VIEWPORT, e2e/support/axeExemptions.ts), which other Selects (Role, Manager, mail folders, custom-field options) still need with real data; the proof moved to a new page state, user-editor-role-long-open: the Role select over forty roles, where End and Home focus the last and first option, each scrolled into view, and the state expects the exemption.
Lesson: none — the tests were left behind by F32-d’s merge. The rule that now catches this is fewer-nightly-reds.md R-1 and R-3: a frontend PR runs the browser tests its change selects, and frontend-reviewer checks every changed string and test id against the tests.
BUG-397 — Since the UI revamp (forge-et-fe#442) the weekly-prices import’s create button does not take focus in WebKit
Affects: F26-c (the weekly workbook import on the Pricing page). Found by the nightly full run on 2026-11-10 (forge-et-fe#484), page state pricing-import-create-unknown at size-768x1024-webkit and size-1024x768-webkit. The same night’s other failures from #442 are already filed as BUG-390 to BUG-393. This is a separate defect from BUG-013, the Windows-only click that times out under load.
Repro: in forge-et-fe at 5b2b278, with a backend at forge-et-be 6e39b71 on 18080: E2E_GATE=1 ET_BACKEND_PORT=18080 pnpm exec playwright test e2e/browser/size-matrix/size-matrix.spec.ts --project=size-768x1024-webkit -g pricing-import-create-unknown --retries=0. On macOS it fails every run, in about 7 s, and it passes in Chromium at every size. A bisect over forge-et-fe’s first-parent merges finds it passing at #452 and failing from #442 (feature/ui-rev100).
Expected: focus is on getByTestId('pricing-import-create') in that state.
Actual: the button is not focused in WebKit.
Fix: find what #442 changed in the import panel’s focus handling, or in WebKit-specific focus styles or tabindex. Restore the focus move, or update the page state if the revamp deliberately changed where focus goes.
Regression test: the existing size-matrix state; it fails today in WebKit.
Resolved: 2026-11-12 — fixed in forge-et-fe#515. #442 moved the Replace and Create confirmation from a native <dialog> to the Radix PricingDialog, which gives focus back on close to the control that had focus when it opened (RecordEditDialog), and falls back to returnFocusTo when there is none. In WebKit a click does not focus a button, so the page state’s click on Create left focus on <body>, the dialog recorded no opener, and with no returnFocusTo focus stayed on <body>. CreateStep now passes returnFocusTo={[createRef]}. A CreateStep test opens the confirmation with no focus move, as WebKit does, presses Replace and Create and expects focus on Create; it fails without the change. The size-matrix state pricing-import-create-unknown passes at every size in both engines. Other dialogs whose callers name no returnFocusTo have the same gap in WebKit: BUG-406.
Lesson: L-287.
BUG-396 — Focus is lost when Retry on New User’s or Edit User’s failed Role or Manager list is pressed and the read had never loaded
Affects: F5-c (forge-et-fe src/features/users/CreateUserModal.tsx and UserEditorForm.tsx, the Role and Manager reads’ failure boxes and their Retry: users-create-roles-retry, users-create-managers-retry, users-editor-roles-retry, users-editor-managers-retry) · Found in: the second and fourth delta reviews of forge-et-fe#476, 2026-11-09, from reading the code on main; not yet run · Severity: low — keyboard focus only (WCAG 2.4.3), when a read failed on first load · Phase: 1
Repro (from the code): open New User or Edit User, make the roles read or the managers read (useManagerCandidates) fail on first load so the failure box and Retry show, and press Retry from the keyboard. Each box is rendered from the query’s isError alone, and under TanStack Query v5 a key that never held data goes back to pending and clears its error at the press, so the box and the pressed Retry unmount; Retry’s busy can never show. In New User the Role select then remounts in its loading state with the id Retry had.
Expected: focus stays on Retry while the retry runs and if it fails again, and moves to a named target once a successful retry removes the button (lesson L-183).
Actual (from the code): focus falls to <body>.
Fix: hold each failure box through the retry with useRetryHold, as UserRoleAccessCard and DraftPriceDialog do (forge-et-fe#476 moves all four onto src/ui/fields/FieldReadFailure), and give Retry a focus repair to its field when the read succeeds.
Regression test: CreateUserModal and UserEditorRoute component tests with the request held: Retry stays the same element, aria-busy and focused; a retry that fails again keeps focus; one that succeeds moves focus to the field.
Resolved: 2026-11-10 — fixed in forge-et-fe#486. A new hook, src/lib/query/useHeldFieldRead.ts (over useHeldRead), holds each field’s failure from the Retry press until the read answers, and after a successful retry moves focus to the field’s control, in a layout effect that runs before Radix’s focus trap would move lost focus onto the dialog. Focus the user moved elsewhere is left alone, and a re-read nobody pressed Retry for moves none. New User and Edit User use it for both the Role and the Manager reads; while a failure is held, the Manager list’s empty line says it failed and Edit User’s Role field drops its Loading roles helper. The review moved isEnclosingDialog (BUG-321’s rule) to one shared src/lib/browser/isEnclosingDialog.ts, used by useFocusAfterWrite too. The four CreateUserModal and UserEditorRoute rows failed before the fix at their first aria-busy="true" check, the pressed Retry having left the page; useHeldFieldRead.test.tsx covers a press that focuses nothing, focus left on the enclosing dialog, focus the user moved and re-reads nobody pressed Retry for, and each of its guards was removed in turn with a row failing. Checked in the browser on the branch (Chromium, the local stub session) with GET /api/v1/roles failed and then held from the page: in New User the same Retry stayed busy and focused through a second failure, and a successful retry moved focus straight to the Role select.
Lesson: L-183 — another instance: the users forms’ Role and Manager failure boxes were rendered from isError alone. L-183’s sources now list BUG-396; a field that also needs focus moved after its retry uses useHeldFieldRead.
BUG-395 — renderWithQueryClient(..., { strict: true }) does not run a component’s effects twice, so the tests that use it do not test StrictMode’s mount, unmount and mount again
Affects: F20-i (forge-et-fe src/test/queryClient.tsx, added in 8925c301), and every test that passes { strict: true }: among them RoleEditorRoute.test.tsx (“raises exactly one ‘Role created’ toast under StrictMode, whose effects run twice”), and the RolesRoute, AccountDetailRoute, AccountListScreen, LeadsScreen.handout, FollowUpEditBody and recordAttachmentsQueries tests. Found by F34-c wave 1’s review fixes on forge-et-fe#474, 2026-11-09.
Repro: a probe component whose effect logs on run and cleanup. renderWithQueryClient(<Probe/>, undefined, { strict: true }) logs one effect run; render(<StrictMode><Probe/></StrictMode>) logs effect, cleanup, effect.
Expected: the option mounts the provider and everything under it, unmounts it and mounts it again, as its JSDoc says and as main.tsx does (lesson L-182).
Actual: the helper puts <StrictMode> inside RTL’s wrapper component. React rehearses a first mount only for a tree whose root is inside StrictMode, and the wrapper sits above it, so effects run once.
Fix: pass RTL’s own option, render(ui, { wrapper: Wrapper, reactStrictMode: strict }), and drop the inner <StrictMode>. Run every test that passes { strict: true }; a test that now fails has found a real double-effect defect, which gets its own bug.
Regression test: a unit test of the helper with a probe component asserting effect, cleanup, effect under { strict: true }, and one run without it.
Resolved: 2026-11-09 — fixed in forge-et-fe#479 with the fix above. The helper sends reactStrictMode only when strict is given, so a test that set RTL’s configure({ reactStrictMode: true }) keeps it. The review found the same shape in seven more test files, fixed in the same PR:
- the
recordAttachmentsQueries,leadRankingQueries,leadTriageQueries,leadResearchQueriesanduseNewMailNoticehook tests nested<StrictMode>insiderenderHook’s wrapper; they passreactStrictMode; - the
EmailTemplateEditorRouteandLeadApproveSendActiontests passed<StrictMode>inside theuigiven to the helper; they pass{ strict: true }; CustomFieldsSettingsRoute.test.tsx, which had worked around this bug withconfigure, passes{ strict }to the helper.
Tests: src/test/queryClient.test.tsx checks effect, cleanup, effect under strict; the query provider mounted twice and unmounted once; one run without strict; RTL’s configured StrictMode kept without strict; and a rerender that keeps the provider. Its first row fails on main’s helper. Every test file that uses strict: true or reactStrictMode, and every changed file, passes under the real double mount (24 files, 1033 tests), so no hidden double-effect defect was found.
Lesson: L-182 — a new update: StrictMode goes at the render root, never inside a wrapper or inside the ui handed to a helper that wraps it.
BUG-394 — A read that failed before shows its old error and Retry while a screen reopened soon after re-reads it, though nobody pressed Retry
Affects: F1-a, F11-b, F32-c (forge-et-fe src/lib/query/readFailed.ts and its callers: the account list’s Owner and Tag filters, AccountsFilters.tsx and tagFilter.ts; the lead list’s Owner and Tag filters, LeadFilters.tsx; the lead page’s Tags section, LeadTagsSection.tsx; the Pipedrive Owner selects, src/lib/bulkReassign/pipedriveOwners.ts) · Found in: frontend-reviewer’s review of BUG-388’s fix in forge-et-fe#475, 2026-11-09, from reading the code; the lead chose to fix it (the owed.md row it replaces) · Severity: low — a misleading state for a few minutes after a failure; nothing is lost or blocked · Phase: 1
Repro (from the code): make one of these reads fail, leave the screen and come back within the query cache time (5 minutes). TanStack Query re-reads the failed key on mount. readFailed is true for any re-read of a key that failed and never held data (isFetching && data === undefined && errorUpdatedAt > 0), so the old error and its Retry show while the read runs. On the lead page’s Tags section, Retry is also busy.
Expected: a read the user did not retry shows its loading state; the error is held through a re-read only from a Retry press, as useRetryHold holds it everywhere else (BUG-266, lesson L-183).
Actual (from the code): the old error, with a Retry the user never pressed.
Fix: start the hold from the press. Add a small shared hook over useRetryHold that reports whether the read failed (an error now, or one held from a Retry press) and gives a retry() that holds before it re-reads. Move every readFailed caller to it, and remove readFailed.
Regression test: for each caller, a component test where the read fails, the screen is unmounted and mounted again within the cache time, and the re-read shows the loading state, not the error; the existing held-then-fails-again and focus rows still pass.
Resolved: 2026-11-09 — fixed in forge-et-fe#477 with the fix above. src/lib/query/useHeldRead.ts, over useRetryHold, returns failed, the shown error and a retry() that holds before it re-reads. These callers moved to it:
- the account list’s Owner and Pipedrive Owner selects (
AccountsFilters.tsx); - the account and contact Tag filters (
useTagFilter); - the lead list’s Tag, Owner and Pipedrive Owner selects (
LeadFilters.tsx); - the lead Tags section, which dropped its own kept error.
pipedriveOwnerSelectState takes failed from the caller, and readFailed is removed.
Tests:
useHeldRead.test.tsx: a press held through a second failure; a press that succeeds; a read that holds data; a mount re-read shows loading.- Reopened-screen rows: the account list’s Owner and Tag (
AccountListScreen.test.tsx), the lead list’s Tag and Owner (LeadsScreen.test.tsx), and the lead Tags section. - The Pipedrive Owner selects re-read without a press (
AccountsFilters.test.tsx,LeadFilters.test.tsx).
All seven new component rows fail on main’s code. The review ran two checks, swapping in a stand-in useHeldRead:
- with
readFailed’s old behaviour, exactly the new rows fail; - with no hold at all, the BUG-387 and BUG-388 rows fail too, so the pressed-Retry hold and its focus repairs stay pinned.
The contacts Tag filter has no screen row; it shares useTagFilter with the account Tag filter.
Lesson: L-183 — a new update: a hold worked out from the query’s state broke its 2026-10-27 rule again; a field’s Retry uses useHeldRead, and only a press starts a hold.
BUG-393 — The Standards book’s Add Standard has no accessible name after a refusal says the client is archived
Affects: F26-b (forge-et-fe src/features/pricing/StandardsToolbar.tsx, pricing-add-standard, its ARCHIVED_CLIENT_REASON) · Found in: a browser gate run on forge-et-fe main at 2391ea95, 2026-11-09 (Windows, Chromium; reproduced alone with no retry) · Severity: medium — WCAG 2.1 SC 4.1.2 if the button itself has no name; F26-b AC-25 · Phase: 1
Repro: pnpm e2e e2e/browser/pricing/standards-book-writes.spec.ts:298 --project=chromium --retries=0 with the local backend running. The test adds a standard, the server answers 422 ACCOUNT_NOT_ACTIVE, and the toolbar’s Add Standard should then say the client is archived.
Expected: pricing-add-standard’s accessible name matches An archived client cannot get new standards (F26-b AC-25).
Actual: its accessible name is the empty string.
Fix: not yet known — check whether pricing-add-standard is still the button (the Rev 100 icon change or the Tooltip may have moved the test id onto a wrapper with no name), or whether the reason no longer reaches the button’s name; then fix the code or the test.
Regression test: the failing browser test, and a StandardsToolbar component test that the unavailable button’s name carries the reason.
Resolved: 2026-11-12 — fixed in forge-et-fe#515. No product defect: the button keeps its name. The test read Add Standard’s name while the Add Standard dialog was still open after the 422, and an open Radix modal sets aria-hidden on everything outside it, so Playwright computed the hidden button’s name as "" (shown on a bare page: a button under an aria-hidden ancestor gives ""). The test now presses Cancel, waits for the dialog to close, then checks aria-disabled and the name. The component test the entry asked for already exists: PricingRoute.test.tsx “disables Add standard with the archived-client reason for an archived client” checks the name carries the reason. standards-book-writes.spec.ts passes in Chromium.
Lesson: L-279 (another instance: a name read behind an open modal, not a row).
BUG-392 — Three Standards book browser tests still expect Cost saved on the Set Cost status line, which now says it only when the toast cannot be heard
Affects: F26-b (forge-et-fe e2e/browser/pricing/standards-book-writes.spec.ts:549 and :611, standards-book-a11y.spec.ts:100) · Found in: a browser gate run on forge-et-fe main at 2391ea95, 2026-11-09 (Windows, Chromium; reproduced alone with no retry) · Severity: low — the tests are out of date; the product does what design-system Rev 103 says · Phase: 1
Repro: pnpm e2e e2e/browser/pricing/standards-book-writes.spec.ts:549 e2e/browser/pricing/standards-book-a11y.spec.ts:100 --project=chromium --retries=0 with the local backend running.
Expected: the browser tests check the success the way the code reports it since cf61a77c (2026-11-07, “repeat a pricing dialog’s success only when the toast cannot be heard”): the pricing-set-cost-toast toast, as PricingRoute.test.tsx already does.
Actual: each test waits for pricing-set-cost-status to read Cost saved; the line is empty because the toast region can be heard.
Fix: in the three tests, assert the toast (pricing-set-cost-toast, Cost saved) in place of the status line, keeping the checks that the dialog stays open and shows the new current cost.
Regression test: the three browser tests themselves, passing on main.
Resolved: 2026-11-12 — fixed in forge-et-fe#515. The three tests expect exactly one Cost saved toast (pricing-set-cost-toast) and the status line empty; the test that saves twice waits for the first toast to leave before the second save, so the count cannot pass on the first toast. standards-book-writes.spec.ts and standards-book-a11y.spec.ts pass in Chromium.
Lesson: none — left behind by cf61a77c; the rule that now catches this is fewer-nightly-reds.md R-1 and R-3.
BUG-391 — The price email layout browser test still expects layout/dialog-max, which F26-d Rev 4 replaced with the lg size
Affects: F26-d (forge-et-fe e2e/browser/price-email/price-email-layout.spec.ts:65) · Found in: a browser gate run on forge-et-fe main at 2391ea95, 2026-11-09 (Windows, Chromium; reproduced alone with no retry) · Severity: low — the test is out of date; the dialog is the size F26-d Rev 4 sets · Phase: 1
Repro: pnpm e2e e2e/browser/price-email/price-email-layout.spec.ts:65 --project=chromium --retries=0 with the local backend running.
Expected: the test checks F26-d AC-26 as Rev 4 states it: the dialog is design-system Rev 90’s lg size (800px, layout/dialog-lg), narrowing with the viewport with space/8 at each side.
Actual: at 1280 × 720 the test requires a width ≤ min(layout/dialog-max, viewport − 2 × space/8) = 664.7px and the dialog is 800px.
Fix: read --et-layout-dialog-lg in place of --et-layout-dialog-max in the test’s width bound.
Regression test: the browser test itself, passing at 1280 × 720 and 768 × 1024.
Resolved: 2026-11-12 — fixed in forge-et-fe#515. The helper is expectDialogWithinLg, reading --et-layout-dialog-lg, and the test’s title and messages name layout/dialog-lg. price-email-layout.spec.ts passes in Chromium.
Lesson: none — left behind by F26-d Rev 4; the rule that now catches this is fewer-nightly-reds.md R-1 and R-3.
BUG-390 — At 1024 × 768 the lead page’s Notes and Email tab overflows <main> by 58px
Affects: F33-a, F11-f (forge-et-fe the lead page’s fill layout from bp/laptop, src/features/leads/LeadView.module.css; e2e/browser/attachments/record-attachments-layout.spec.ts:238) · Found in: a browser gate run on forge-et-fe main at 2391ea95, 2026-11-09 (Windows, Chromium; reproduced alone with no retry) · Severity: low — the page scrolls by 58px where the tab should fill the window and scroll inside its cards · Phase: 1
Repro: pnpm e2e e2e/browser/attachments/record-attachments-layout.spec.ts:238 --project=chromium --retries=0 with the local backend running: the lead page at 1024 × 768, Notes and Email tab, with the attachments card filled on Overview.
Expected: from bp/laptop the tab fills <main> and <main> does not scroll (overflow ≤ 1px); the notes and the attachments scroll inside their cards (F33-a AC-10, WI-1.5 clause).
Actual: <main> overflows by 58px. At 1280 × 800 on a lead with no attachments the same tab did not overflow (checked by hand, 2026-11-09).
Fix: not yet known — measure which block of the tab is taller than its share at 1024 × 768 (likely a fixed-height block above the notes, or the attachments card, not shrinking).
Regression test: the failing browser test.
Resolved: 2026-11-12 — fixed in forge-et-fe#515. Measured in Chromium at 1024 × 768: the lead header wraps to 411px (269px at 1366 × 768), which leaves the tab 238px, and the email column (the Follow-up draft and Email cards) needs 321px and cannot shrink, so <main>'s content was 826px against 768. Under the fill layout the email column now scrolls inside itself (LeadView.module.css, on the notes-and-email grid’s second column, which RecordColumns always renders), with position: relative and overscroll-behavior: contain as the contact page’s rail has, so the hidden text placed absolutely inside it cannot escape and lengthen the page again (L-176; notesScrollerPosition.test.ts now checks both columns, and fails without position: relative). Nothing changes where the column fits (1366 × 768) or below bp/laptop, where the fill layout is off. The browser test failed with “overflow 58px” before the change at forge-et-fe fe35c3df; it now also checks that the email column scrolls at 1024 × 768 and that Compose Email is reachable inside it, and passes at both laptop sizes.
Lesson: none — one-off: the lead’s email column is the one fill-layout column made of cards that cannot shrink.
BUG-389 — After a contact is archived from the Contacts list, its row leaves the table while the list re-reads
Affects: F1-f (forge-et-fe the Contacts list, src/features/accounts/ContactsScreen.tsx, ContactsTable.tsx; e2e/browser/accounts/contacts-list.spec.ts:575) · Found in: a browser gate run on forge-et-fe main at 2391ea95, 2026-11-09 (Windows, Chromium; reproduced alone with no retry) · Severity: medium — F1-f AC-12’s refreshing state: the row disappears before the re-read answers · Phase: 1
Repro: pnpm e2e e2e/browser/accounts/contacts-list.spec.ts:575 --project=chromium --retries=0 with the local backend running. The test pins the list, archives a contact from its row’s ⋯ menu, and holds the list’s re-read.
Expected: while the re-read is held, the rows stay on screen (the archived contact’s included), with no skeleton and the status line saying the list is updating (F1-f AC-12).
Actual: the archived contact’s row is gone (contacts-row for it not found) before the held re-read is released.
Fix: not yet known — check whether the archive’s success now removes the row from the cached page (or changes the list’s query key) before the re-read, which F1-f AC-12 does not allow.
Regression test: the failing browser test, and a ContactsScreen component test that the row stays while the re-read after an archive is pending.
Resolved: 2026-11-10 — fixed in forge-et-fe#487, tests only: the row was never removed. A contact write settles once the lists are re-read (F1-d WI-1.2 plan, subtask 3: usePatchContact’s onSuccess returns the invalidation), so the Archive Contact? confirmation stays open and busy over the held re-read, and an open modal hides the rest of the page from the accessibility tree (Radix’s aria-hidden and inert now, the native showModal() before). The test looked the row up by role (getByRole('link', …)), which skips hidden content, so it found nothing; Playwright’s snapshot at the failure showed the dialog open with both buttons disabled. The browser test now checks the archive’s submit is busy over the held re-read, finds the kept row with includeHidden (contactsRowNamed(page, name, { includeHidden: true })) and checks the confirmation closes after the release; it failed on main d48c664f alone (Chromium, retries off) and passes on the branch. A new ContactsScreen.actions.test.tsx row holds the re-read after an archive and finds the row still in the table, hidden from assistive tech, with no skeleton and Updating…; with onSuccess changed to not await the re-read, it failed. The review found three absence checks with the same cause, which passed at once behind the open confirmation whether or not the row left (contacts-writes.spec.ts, both archives; contacts-access.spec.ts, the sharee’s archive): each now waits for the confirmation to close first. The title and severity record what was reported; whether the confirmation should keep waiting for the re-read is F1-d’s design and is unchanged.
Lesson: L-279 — a new lesson: a test of the page during a write’s re-read runs while the write’s dialog is still open, so it finds page content with includeHidden / hidden: true or waits for the dialog to close.
BUG-388 — Focus is lost when Retry on the lead page’s Tags section is pressed and the tag options read had never loaded
Affects: F11-b (forge-et-fe src/features/leads/LeadTagsSection.tsx, the Add tag read error and its Retry, leads-tags-retry) · Found in: frontend-reviewer’s review of forge-et-fe#471 (BUG-387’s fix), 2026-11-09, from reading the code; not yet run · Severity: low — keyboard focus only (WCAG 2.4.3), when the tag options read failed on first load · Phase: 1
Repro (from the code): on a lead’s page, as a user who can manage it, make the tag options read (leadsKeys.tags()) fail on first load, so the section shows its read error and Retry; press Retry from the keyboard. readFailed is tagOptions.isError && !loading, and under TanStack Query v5 a key that never held data goes back to pending and clears its error at the press, so the error and the pressed Retry unmount.
Expected: focus stays on Retry while the retry runs and if it fails again, and moves to a named target once a successful retry removes the button (lesson L-133).
Actual (from the code): focus falls to <body>.
Fix: as BUG-387 did on the lists: show the error while readFailed(tagOptions) from src/lib/query/readFailed.ts holds, and give the Retry a focus repair (useFocusAfterWrite) with a target that stays on the page.
Regression test: a LeadTagsSection or lead page component test where the read fails, Retry is pressed and fails again (focus stays on Retry), and one where it succeeds (focus on the target).
Resolved: 2026-11-09 — fixed in forge-et-fe#475 with the fix above. The section shows the Add tag read error and Retry while readFailed(tagOptions) holds. Retry notes the press and arms a focus repair (useFocusAfterWrite, timed to the tag read, scoped to the whole section because the error block mounts only on failure). The repair moves focus to Add tag once a successful retry removes the button. The review found two older defects on the same button, fixed in the same PR:
- a mouse press on Retry while the Add tag search was open could fold the search and move the button from under the pointer; it now has BUG-378’s
keepSearchFocusguard; - the failed read showed no trace id; the read’s error is now kept through a retry and shown.
Retry is also busy while it re-reads. Tests are in LeadTagsSection.test.tsx:
- held, then fails again (anchored on the query settling as an error);
- success → Add tag;
- focus moved elsewhere stays;
- a press that focused nothing, then success → Add tag;
- a press that focused nothing, then fails again;
- the read’s trace id and busy state;
- the BUG-378 press with the search open and folded.
On main’s code, three of the first four rows fail; the “moved elsewhere” row passes either way.
Lesson: L-183 — another instance: the lead page’s Tags section showed its read error from isError alone. L-183’s sources name BUG-388 as fixed.
BUG-387 — Focus is lost when Retry on the account or lead list’s Owner filter is pressed and its read had never loaded
Affects: F1-a, F11-b (forge-et-fe src/features/accounts/AccountsFilters.tsx and src/features/leads/LeadFilters.tsx, the Owner select’s read error, people.isError) · Found in: frontend-reviewer’s review of forge-et-fe#464 (F32-c wave 2), 2026-11-08, in code that PR did not change · Severity: low · Phase: 1
Repro (from the code): make the Owner select’s directory read fail on first load, so its read error shows Retry; press Retry from the keyboard. Under TanStack Query v5 a query that never held data goes back to pending and clears error on the press, so people.isError is false while the retry runs and the read error and its Retry unmount at the press.
Expected: focus stays on Retry while the retry runs, stays on the new Retry if it fails again, and moves to the screen’s fallback (the h1) if it succeeds (WCAG 2.4.3; the Tag select’s readFailed hold, L-133, L-183).
Actual: on the lead list focus falls to <body>; on the account list ownerFocus finds the old button gone and moves focus to the h1 even when the retry fails again.
Fix: hold the failure on screen while the retry runs, as F32-c wave 2 does for the Pipedrive Owner select (readFailed-style isError || (isFetching && data === undefined && errorUpdatedAt > 0), or useRetryHold), and add the focus repair on the lead list (useFocusAfterWrite, as LeadsScreen already imports).
Regression test: on both lists, a component test where the directory read fails, Retry is pressed and fails again (focus stays on Retry), and one where it succeeds (focus on the h1).
Resolved: 2026-11-09 — fixed in forge-et-fe#471 with the fix above. Both Owner selects show their read error while readFailed(people) holds, which stays true while a retry re-reads; the rule moved to one shared src/lib/query/readFailed.ts, replacing the private copies in accountsQueries.ts and pipedriveOwners.ts. The lead list gained what the account list had: its Retry notes the press and arms a focus repair (useFocusAfterWrite, timed to the directory read) that moves focus to the h1 once a successful retry removes the button. The review found the lead list’s Tag select had the same defect; it was fixed in the same PR the same way (tagFocusAfterRetry). The lead page’s Tags section has the same pattern on another screen: BUG-388. Tests: AccountListScreen.test.tsx (the account Owner select held, then failed again, anchored on the query settling as an error) and LeadsScreen.test.tsx (for the lead Tag and Owner selects: held then failed again; success → h1; focus moved elsewhere stays; a press that focused nothing → h1); on main’s code seven of these rows fail, and the two “moved elsewhere” rows pass either way.
Lesson: L-183 — another instance: the Owner and lead Tag selects showed their error from isError alone. L-183’s sources now list BUG-387 and BUG-388.
BUG-386 — An open shadcn Select list taller than the window fails axe’s scrollable-region-focusable
Affects: every CRM screen with a shadcn Select whose option list can be longer than the window (forge-et-fe src/ui/shadcn/select.tsx; for example the account list’s Country filter at 320 × 640) · Found in: F32-c WI-2.1’s review-fix round on forge-et-fe#464, 2026-11-08: a page state that opened the Pipedrive Owner list with one long name grew taller than 320 × 640 and failed; the state was changed to avoid it · Severity: medium — WCAG 2.1 SC 2.1.1, axe scrollable-region-focusable (serious) · Phase: 1
Repro: at 320 × 640, open a shadcn Select whose options are taller than the window and run axe with the page check (expectPageClean).
Expected: no axe violation; the open list is reachable and scrollable from the keyboard (arrow keys move through Radix’s options).
Actual: axe reports scrollable-region-focusable on the scrolling div[data-radix-select-viewport], which holds no element in the Tab order (Radix moves real focus to each option, which has tabindex="-1"; axe counts only tabbable descendants). No page state opens a list that long today, so the nightly size matrix does not see it.
Fix: not decided. Confirm the keyboard can reach every option of a scrolled list (Radix’s own behaviour); if it can, the finding is axe not seeing Radix’s focus model and the fix is a reasoned entry in e2e/support/axeExemptions.ts scoped to [data-radix-select-viewport]; if it cannot, make the viewport focusable. Then add a page state that opens a long list at 320 × 640. BUG-123 is the same rule on the tag suggestions list.
Regression test: a size-matrix state with a shadcn Select list longer than the window, open, passes the page check at 320 × 640.
Resolved: 2026-11-09 — fixed in forge-et-fe#472 by the first branch of the fix above. Reproduced on the account list’s Country select (250 options) at 320 × 640 and 768 × 1024 with the shared page check. The F32-c Pipedrive Owner list with the long name passed: three options are not tall enough to scroll. The keyboard reaches every option: accounts-a11y.spec.ts “BUG-386 (keyboard)” at 320 × 640 checks that End, ArrowUp, Home and typing each focus an option and bring it into view, and that Escape returns focus to the trigger. So the finding is accepted as a reasoned entry in e2e/support/axeExemptions.ts, scoped to [data-radix-select-viewport] (SELECT_VIEWPORT). Radix also holds Tab inside the open list, so a Tab stop on the viewport would add nothing. The size-matrix state accounts-country-open opens the list at every size and expects the exemption to apply; without the entry it fails at 320 × 640 with the viewport finding. The Pipedrive Owner -open states carry the long name again and now check it is listed.
Lesson: none — one-off: an axe rule that counts only tabbable descendants, against a list whose options take focus out of the Tab order, accepted like SELECT_OPEN_MAIN; no other Lesson: line names this rule on a Radix viewport.
BUG-384 — A refused mailbox connect start moves its Reconnect or Try Again button into an error panel, so keyboard focus falls to the page body
Affects: F20-a, F20-b (forge-et-fe src/features/email/connect/NextStepAction.tsx, the mailbox screen’s Reconnect / Try Again; src/features/email/SyncIndicator.tsx, the lapsed indicator’s Reconnect) · Found in: frontend-reviewer’s review of forge-et-fe#462 (BUG-351’s fix), 2026-11-08, from reading the code; not yet run · Severity: low — keyboard focus (WCAG 2.1 SC 2.4.3), only after the server refuses a start; the button itself still works · Phase: 1
Repro (by reading): on /email/mailbox with the mailbox needing a reconnect, Tab to Reconnect and press Enter while the start is refused (a 500 on POST of the connect start, as sync-indicator.spec.ts’s “a reconnect that cannot begin” row pins it).
Expected: focus stays on Reconnect (now inside the error panel), so a keyboard user can try again at once (lesson L-133).
Actual (by reading): before the failure each component renders its button on its own; after it, the same button is passed as <ErrorPanel action={button}>. That is a different parent, so React unmounts the focused button and mounts a new one, and focus falls to <body>. Nothing restores it: a refused start never reaches the focus-restoring re-read, and ErrorPanel only sets role="alert". This is the pattern of BUG-014 and BUG-302. NextStepAction.test.tsx asserts no focus, and the browser row presses with the mouse.
Fix: the fix (2026-11-09) puts focus on the new button, per L-133, rather than keeping one parent. Keeping the button under one parent would take it out of the error panel, which design-system §5 says carries the retry, a visible change. A hook, useRefusedStartFocus, moves focus to the button after each render that moves it (into the panel on a refusal, and out of it on a retry pressed from the panel), only when the press left focus on the button or on nothing and nothing holds focus once the move has rendered.
Regression test: per component, a case that presses the button with the keyboard, refuses the start, and asserts the button still has focus; it fails before the fix.
Resolved: 2026-11-09 — fixed in forge-et-fe#468 (25531363, with the review’s changes in f3eb7062). src/features/email/connect/useRefusedStartFocus.ts is used by NextStepAction and SyncLapsedBody. Tests, each failing with the component change reversed: per component, a keyboard press refused (both of NextStepAction’s labels), a press that focused nothing (fireEvent.click from <body>), and a retry pressed inside the panel, held, then refused (focus on the button during the wait and after; these failed on 25531363’s first version of the hook, which the review found); a moved-elsewhere row in each passes both ways, as it should. The browser row sync-indicator.spec.ts “BUG-384: … pressed with the keyboard …” passes in Chromium and fails with the change reversed. Checked by hand on /email/mailbox against a local backend: after Reconnect pressed with Enter and a refused start, focus was on Reconnect inside the error panel. The lapsed indicator’s copy was not checked by hand, because a lapsed subscription cannot be set up locally.
Lesson: L-133, another instance — a control remounted under a new parent by its own refusal; L-133’s sources now list BUG-384.
BUG-385 — At 320px on a user’s page, a third-level name in the Reporting line tree is about 8px wide, and a deactivated third-level person pushes <main> sideways
Affects: F5-c (forge-et-fe src/features/users/ReportingLineChain.module.css, the .tree indent per level; ReportingLineChain.tsx, FLAT_FROM_LEVEL) · Found in: BUG-381’s fix (forge-et-fe#454), 2026-11-08: its size-matrix state user-editor-reporting-deep once moved to the user’s page, and a held page on main · Severity: medium — 320px reflow (WCAG 2.1 SC 1.4.10) and the 24px target floor (design-system §9), for any tree three levels deep · Phase: 1
Repro: on a user’s page (/settings/users/<id>, view mode) at 320 × 640, give the user people below them three levels deep. The navigation rail keeps about 75px at that width, and each tree level indents about 41px (--rail-x about 23px plus --elbow about 18px), so after the initials a third-level name has a column about 8px wide. The size matrix reports each third-level name link as “8.13px × 312px, under the 24 × 24px floor” (axe target-size). With a deactivated person at the third level, the Inactive label, which does not wrap, passes <main>'s edge by about 24px.
Expected: a third-level name is readable, its link meets the 24px floor, and nothing scrolls sideways at 320px.
Actual: each third-level name shows one letter per line, and a deactivated third-level person makes <main> scroll sideways.
Why no state caught it: the size-matrix states with a deeper tree ran on the edit route, where the Edit user dialog covers the page, so the page check’s target-size scan never reached the tree; the view route’s own state (user-view) has two levels.
Fix: decided by Tharinda on 2026-11-09: below bp/tablet only, BUG-381’s flat levels start at the third level below the user instead of the fourth (FLAT_FROM_LEVEL 3, and the CSS steps back one indent, level two’s, instead of two), so a third-level person lines up with the user’s own reports and names their level and manager; at bp/tablet and wider every tree looks as it does now. This narrows BUG-381’s condition that a tree of three levels or fewer looks unchanged at every width: it now holds at bp/tablet and wider. The option not taken was a smaller indent per level below bp/tablet.
Regression test: remove the BUG-385 knownFailures mark from user-editor-reporting-deep in forge-et-fe e2e/pageStates.ts; the state then passes the page check at 320 × 640.
Resolved: 2026-11-09 — fixed in forge-et-fe#466, as Tharinda decided. FLAT_FROM_LEVEL is 3, and .flatLevel steps back one indent (level two’s), so below bp/tablet the third level and deeper line up with the user’s own reports and each person on them reads “Level N, reports to bp/tablet and wider every level still indents. The flat block still covers the rail of a level-two person with a later sibling. Tests: ReportingLineChain.test.tsx (levels three to five flat at 320px, a three-level tree flat at its third level, a two-level tree unchanged, nothing flat at 768px, a branching tree; three of its cases fail with the source change set aside); the size-matrix state user-editor-reporting-deep passes at every size with its BUG-385 mark removed, and its pixel check now reads ul[data-level="3"]; on the previous code it fails at 320 × 640.
Lesson: L-272 — the same cause: a fix measured on a page state that did not show the section on the page where it is seen. This bug is what L-272’s own BUG-381 found once the state moved; L-272’s sources list both. The fix’s first browser runs also used another worktree’s dev server on port 5173 and measured that worktree’s code; they were repeated with E2E_GATE=1 (L-274, which now lists BUG-385, and a new rule in .claude/rules/test-record.md).
BUG-383 — After a refused New user save, focus can stay on Save: the focus move runs in a frame callback that can come before the error summary renders
Affects: F5-c (forge-et-fe src/features/users/CreateUserModal.tsx, the save’s onError) · Found in: the frontend review of forge-et-fe#459 (the BUG-355 fix), 2026-11-08, in code, not in a run · Severity: low — focus only; the summary’s role="alert" still announces the refusal · Phase: 1
Repro: in onError, the dialog sets the field errors and the summary, then calls requestAnimationFrame(() => focusFirst(mapped.fields, mapped.summary)), whose comment says it runs “after the render that shows the messages”. On a loaded machine the frame can run before React renders them, as BUG-355 showed for the user page. A refusal that only fills the summary then finds no summary element (submit had just cleared it, and it renders only with lines), so focus stays on Save. A field error focuses the input before its aria-invalid and message are rendered, so a screen reader reads the field without its error. A component test that runs frame callbacks at once (BUG-355’s UserEditorRoute.test.tsx cases) shows the order.
Expected: after a refused save, focus moves to the first error, or to the summary, once they are on screen.
Actual: under load the move can run before they render.
Fix: as BUG-355: onError sets a focus request in state, and an effect after the render that shows the messages moves focus (the pattern UserActivationCard’s failureFocus uses).
Regression test: a CreateUserModal component test that mocks requestAnimationFrame to run at once and asserts the summary, then the first field, has focus after a refused save; it fails on the current code.
Resolved: 2026-11-09 — fixed in forge-et-fe#468 (6617f1cc, with a comment fix in f3eb7062), as the Fix line says: the save sets a focus request in state in the same batch as the messages, and an effect moves focus after that render. The client-side check before sending used the same frame-free path but focused at once, before aria-invalid rendered; it uses the same request now. Tests: two CreateUserModal.test.tsx cases run frame callbacks at once — a 500 focuses the summary, and a 409 focuses Email with aria-invalid="true" already set when focus arrives (recorded by a focus listener); both failed on the old code. Checked by hand against a local backend: a duplicate email moved focus to Email, already invalid with its message; a 500 (answered in the page) moved focus to the error summary.
Lesson: L-271, another instance — the same frame-callback focus move as BUG-355; L-271’s sources now say it is fixed.
BUG-382 — Flaky tests: the account and lead lists’ clean-pass focus cases miss the result toast within findBy’s 1 s default under load (Windows)
Affects: F32-c (forge-et-fe src/features/accounts/AccountListScreen.selection.test.tsx and src/features/leads/LeadsScreen.selection.test.tsx, “a confirm press that focused nothing still leaves focus on the
after a clean pass”) · Found in: forge-et-fe#442 (feature/ui-rev100) after its merge of main, Windows 11, 2026-11-08 · Severity: low — tests only; a failing run blocks a local pnpm test or pnpm test:record · Phase: 1
Repro: three runs on 2026-11-08 failed both cases, each with Unable to find an element by: [data-testid="account-list-result-toast"] (and leads-result-toast). The test record on af5b8dc6 failed them on the first try and the retry, with the machine at 0.93 average load: a size-matrix run was going beside it. A Vitest batch over src/features/{users,accounts,activities}, src/app and others failed the account case. A batch with the toast raised through notifySuccess failed both. Run alone straight after, both passed, 2 of 2 twice.
Expected: each case waits for what the clean pass’s answer causes, whatever the run’s load.
Actual: await screen.findByTestId('<list>-result-toast') gives the pass’s answer, the dialog’s close, the list’s re-read and the toast together only findBy’s default second.
Fix: hold the pass’s answer and release it inside act, so the React work it causes is flushed before findBy’s second starts; findBy then waits only for Sonner’s toast.
Regression test: both cases pass three runs in a row under one busy loop per core (pnpm exec vitest --run <both files> --retry=0 -t "focused nothing"), where the old code failed three of three. The whole files under that load also lose other cases to the 20 s test timeout and the “a search clears it” cases to their own wait (see Resolved:), so the whole-file runs are not this bug’s check.
Resolved: 2026-11-08 — fixed in forge-et-fe#456 (85def261, with the review’s changes in 6dac638a). No product defect. After fireEvent.click(confirm), one findByTestId('<list>-result-toast') covered four renders of the whole list: the confirm turning busy, the re-read’s busy rows, the dialog closing, and Sonner’s toast. Timed on Windows, the toast came 0.28–0.88 s after the click on an idle machine and 2.0–2.8 s with one busy loop per core (16), each step taking 300–700 ms. With that load both cases failed 3 of 3 runs with this bug’s error. Each case now holds the bulk-reassign route behind a gate that answers from the test’s store, and releases it inside await act(async () => …). That flushes the dialog’s close, the cleared selection, the focus move to the <h1> and the re-read, so findBy waits only for the toast. Both cases passed 3 of 3 under the same load, and in all six later loaded whole-file runs. The sibling Change Owner cases were measured and left unchanged: they confirm through user.click, which flushes the chain inside act, and their findBy took 7–316 ms under load. In the loaded whole-file runs, with Docker and other sessions also running (100% CPU), other cases failed: “keeps the selection across the page, the sort and the page size” (both files) and several others hit the 20 s test timeout, and “a search clears it” (both files) missed its own wait once per set. Those are not this bug; they are recorded in owed.md for the flaky report to watch.
Lesson: L-270 (new, with BUG-365).
BUG-381 — At 320px a deactivated person four or more levels down the Reporting line tree, or anyone five levels down, pushes <main> sideways
<main> sidewaysAffects: F5-c (forge-et-fe src/features/users/ReportingLineChain.module.css, the .tree indent; src/ui/components/StatusIndicator/StatusIndicator.module.css, .label { white-space: nowrap }) · Found in: frontend-reviewer’s review of forge-et-fe#450 (BUG-354’s fix), 2026-11-08, from computed widths; not yet seen in a browser · Severity: medium — 320px reflow (WCAG 2.1 SC 1.4.10), on every system; only a reporting tree deeper than any page state has · Phase: 1
Repro (expected from the measurements): on a user’s page at 320 × 640, give the user people below them four levels deep, with a deactivated person at the fourth level. Each tree level indents about 41px (--rail-x about 23px plus --elbow about 18px). BUG-354’s measurement put the third level’s text column at about 59px, so the fourth level’s is about 18px. The Inactive status label cannot wrap (white-space: nowrap) and is about 56px wide (the 8px dot, a 6px gap and the word at 11.5px), so it passes <main>'s inner edge by about 35–40px. At the fifth level the 36px initials alone are wider than the column, for any person.
Expected: no horizontal scroll on the page or in <main> at 320px, however deep the tree.
Actual (expected from the measurements): <main> scrolls sideways.
Fix: decided by Tharinda on 2026-11-08: below bp/tablet, stop indenting past a set depth and show a deeper person’s depth another way (a small level label), so any depth fits; the option not taken was a smaller indent per level. Tharinda expects the reporting line to stay within three levels in practice, and chose the fix as a safety net for a deeper tree, on the condition that a tree of three levels or fewer looks exactly as it does now at every width: the cap starts at the fourth level below the user, and applies only below bp/tablet. The regression tests check both sides (no change at three levels; no overflow at four and five). Confirm the overflow first in a browser (a size-matrix state or a held page with a four-level tree).
Regression test: a size-matrix state with a deactivated person four levels down passes the page check at 320 × 640.
Resolved: 2026-11-08 — fixed in forge-et-fe#454, as Tharinda decided. Below bp/tablet (useMediaQuery(mq.tablet)), each level from the fourth below the user (FLAT_FROM_LEVEL 4) is a .flatLevel list: the first steps back the two indents of levels two and three, so it lines up with the user’s own reports, the deeper ones add none, they draw no rail or elbow, and each person on them reads “Level N, reports to position: relative; z-index: 1; background: var(--et-card)); the rails resume below it. A tree of three levels or fewer is unchanged at every width: the only DOM change is a data-level attribute on each level’s list, which no style reads. Tests: ReportingLineChain.test.tsx (levels four and five flat at 320px, a branching tree, nothing flat at 768px, a three-level tree unchanged at 320px); the size-matrix state user-editor-reporting-deep, a five-level branching tree on the user’s page, passes at all seven sizes, and its pixel check of the gap above each flat person failed on the previous head at 320px (“a rail at 23px across the gap 3px down”). Moving that state from the edit route, where the dialog covers the tree, to the user’s page showed that the third level is already too narrow there at 320px: BUG-385, marked as a known target-size failure on the state at that size.
Lesson: L-272 — a size-matrix state checks what is visible on its page; a section covered by a dialog is not checked by it.
BUG-380 — A recreate of the All-mail folder can be handed back the gone id by the resolve it already joined, and answers 503
Affects: F20-i (forge-et-be mailbox/service/AllMailFolders.java, singleFlight) · Found in: the nightly full run on main fb140884, 2026-11-08 (macOS), as AllMailFoldersSpec “two recreates that join one resolve answering the gone id create one folder between them” (forge-et-be#385) · Severity: low — a request that recreates the folder while another request’s resolve is finishing answers 503 MAILBOX_UNAVAILABLE once; the next request recreates it · Phase: 1
Repro: the spec passed alone at fb140884 (22 of 22 in 1.1 s), so the window is narrow. A test that holds a flight which has already answered in the resolving map until the next request looks for a running flight fails on the old code with the nightly’s exception, MailboxUnavailableException: Microsoft could not be reached from AllMailFolders.recreate.
Expected: a recreate whose joined resolve answered the gone id finds or creates the folder again in its second attempt.
Actual: singleFlight completed the flight’s future before removing the flight from the map. The recreate woke on the answer, started its second attempt, found the finished flight still in the map, joined it, got the gone id again and threw UNREACHABLE. A busy machine widens the window between the two steps.
Fix: the flight leaves the map before it answers (on success and on failure), so a request that has its answer starts a new flight. FolderTreeCache.read has the same order and is left as it is: a request that joins its finished read gets a tree read moments before, which is correct. A rarer order still ends in the same 503 and is accepted, as recreate’s Javadoc already states (“a gone id answered again is 503”): a second resolve that read held as empty before the first resolve stored the gone id starts its own flight after the first leaves the map, reads the gone id from held, and the recreate’s second attempt joins it. It needs that resolve paused across the whole window, and the next request recreates the folder; a later nightly failure of this spec is checked against that order before the fix is called broken.
Regression test: AllMailFoldersSpec “a recreate’s second attempt never joins the resolve that already answered it the gone id (BUG-380)”, which fails before the fix and passes after; the eight All-mail specs pass (67 tests).
Resolved: 2026-11-08 — fixed in forge-et-be#386 (38a0b992, with the review’s failure-path change in 3d4f0eeb).
Lesson: none — one-off: it is the only single flight whose caller makes a second attempt that must not see the first answer; the other one, FolderTreeCache.read, is correct with either order.
BUG-379 — On the Activities page, focus can fall to the page body when Reopen in a Completed follow-up’s dialog is followed by Escape before the list’s re-read lands
Affects: F3-a (forge-et-fe src/features/activities/ActivitiesScreen.tsx, the edit FollowUpDialog, about l.304–331) · Found in: F3-a WI-4.1, the test-fix of R3’s WebKit keyboard walk, 2026-11-07 · Severity: low — keyboard focus only (NFR-1), in a narrow timing window; nothing is lost or changed · Phase: 1
Repro (by reading, not yet reproduced): on /activities → Completed, open a done follow-up, press Reopen (200), then press Escape before the Completed list’s re-read has answered.
Expected: focus goes to the page’s h1 once the re-read drops the reopened row, as it does after a save whose re-read drops the row (WI-2.2 plan Decision 13, the savedFrom check).
Actual (by reading): the screen passes no onReopened to the edit dialog, so savedFrom is armed only by onSaved. At the close the row is still listed, so focus returns to its title; the re-read then removes that row and focus is left on body.
Fix: not decided. The likely fix is onReopened={() => { savedFrom.current = { list: shownListRef.current }; }} on the edit dialog, so the same check covers a reopen; confirm the repro first (a component test that holds the re-read).
Regression test: an ActivitiesScreen component test: Reopen, close, then release the held re-read without the row; focus is on the h1.
Resolved: 2026-11-08 — fixed in forge-et-fe#454 with the likely fix above: the edit FollowUpDialog gets onReopened={() => { savedFrom.current = { list: shownListRef.current }; }}, so the check that moves focus to the h1 when a re-read drops the saved row also covers a reopen. The repro was confirmed by the regression test, ActivitiesScreen.test.tsx “BUG-379: Reopen, then Escape before the re-read lands — when the re-read drops the row, focus goes to the h1”: it holds the list’s re-read, closes the dialog with Escape (focus returns to the row’s title), then answers the re-read without the row; it failed without the fix, with focus on body.
Lesson: L-133 — another instance: a second path (Reopen) that removes the row was not armed with the focus target the save path already had. L-133’s sources now list BUG-379.
BUG-378 — A press on the tag options Retry while the Add tag search is open is lost: the search folds on mousedown and Retry moves from under the pointer
Affects: F1-e (forge-et-fe src/ui/tags/AddTagControl.tsx, which folds the search back to the chip when focus leaves it; src/features/notesAndTags/RecordTagsSection.tsx, whose Retry sits under the field) · Found in: the nightly full run on main 6e1b32c2, 2026-11-07 (macOS), as e2e/browser/accounts/account-tags.spec.ts AC-21 “Retry reads the options again”, titled “Try Again reads the options again” since forge-et-fe#554 (forge-et-fe#430) · Severity: low — the first click does nothing and a second click works; the keyboard path is not affected · Phase: 1
Repro: on an account’s Overview tab, with the tag options read failing (pinTagOptions(page, 'account', { kind: 'failed' })), open the Add tag search, press Escape once to close its list, and click Retry with the mouse. No options read is sent. A click listener on the document shows the click target is the account-tags-panel card, and Retry is 45px higher after the press than before it.
Expected: the click on Retry reads the options again.
Actual: mousedown moves focus from the search to Retry, the search folds back to the chip, the field gets shorter and Retry moves up 45px before mouseup, so the click lands on the card. The test passed before F3-a (forge-et-fe#431) only because of where the page was scrolled. forge-et-fe#443 changes the test to fold the search first; this bug is the lost press itself.
Fix: decided by Tharinda on 2026-11-08: keep focus in the search while Retry is pressed (preventDefault on its pointerdown), so the search stays open, nothing moves and the click reaches Retry. The option not taken: wait until the press ends before folding when focus moves inside the Tags panel.
Regression test: a browser case that clicks Retry with the search open and expects the options read; it fails before the fix.
Resolved: 2026-11-08 — fixed in forge-et-fe#446, as Tharinda decided. RecordTagsSection’s options Retry prevents its mousedown while the Add tag search (its combobox) has the focus, so the press leaves the focus in the search, the search does not fold and Retry stays under the pointer. The handler is on mousedown, not pointerdown, because focus moves as mousedown’s default action. With the search not focused, a click focuses Retry as before, which AC-23’s focus rules need; an always-on version failed RecordTagsSection.test.tsx’s “failing again leaves focus on it” case. The keyboard path is unchanged. Tests: account-tags.spec.ts AC-21 now presses Retry with the search open (Escape once) and checks the read, that the search stayed open, and that Retry did not move; RecordTagsSection.test.tsx asserts the press’s default is prevented while the search has the focus, and a second case presses Retry with the search open and checks the focus stays in the search when the read fails again and when it succeeds; both failed with the source change set aside. The browser test also asserts the focus stayed in the search.
Lesson: none — one-off: the first press lost because a control that folds on focusout moved its neighbour before mouseup; no other Lesson: line or lesson names the cause.
BUG-377 — Below 1024px the folder picker’s loading skeleton has no width, so Loading your folders is never seen
Affects: F20-i (forge-et-fe src/features/email/FolderPicker.tsx, the loading state; InboxToolbar.module.css .picker) · Found in: the nightly full run on main 6e1b32c2, 2026-11-07 (macOS), size-matrix state email-folders-loading at 768 × 1024 (Chromium and WebKit) and 320 × 640 (forge-et-fe#430) · Severity: medium — below bp/laptop the folder picker’s place in the data bar is empty while the tree loads, instead of the design-system §5 loading state · Phase: 1
Repro: pnpm exec playwright test e2e/browser/size-matrix --grep 'email-folders-loading' --project='size-*' --repeat-each=3 --retries=0: 9 of 9 fail at 768 and 320 with getByTestId('inbox-folders-loading') resolved but hidden.
Expected: the one-line skeleton in the picker’s place, named Loading your folders.
Actual: BUG-361 (forge-et-fe#435) made the data bar’s picker slot as wide as its content, and a skeleton bar has no width of its own, so the skeleton was zero pixels wide.
Fix: the skeleton sits in a box five space/8 steps wide (FolderColumn.module.css .pickerLoading), about the Folder: Inbox trigger’s width. A fixed length, not a percentage: the slot’s width comes from the box, so a percentage of it resolves to nothing.
Resolved: 2026-11-07 — fixed in forge-et-fe#443 (a2acffe4). The size-matrix state failed 9 of 9 at 768 and 320 before and passes 31 of 31 at every size after.
Lesson: none — one-off: the folder picker’s skeleton is the only loading placeholder in a data bar slot sized to its content; the other SkeletonLines uses (git grep -ln SkeletonLines -- src/features, as of 2026-11-07) sit in panels, columns and the message pane, which have a width of their own.
BUG-376 — Cancel closes the Publish, Withdraw, Remove Draft and Correct price dialogs while their request is still running, where ✕ and Esc refuse
Affects: F26-f (Publish the Round, Withdraw), F26-g (Correct), and Remove Draft (forge-et-fe src/features/pricing/PublishRoundDialog.tsx, WithdrawDialog.tsx, RemoveDraftDialog.tsx, CorrectPriceDialog.tsx) · Found in: the frontend review of forge-et-fe#442, 2026-11-07, by reading; the behaviour predates that PR · Severity: medium · Phase: 1
Repro: on the Pricing screen, open Publish the Round (or Withdraw, Remove Draft, Correct), press the main action, and press Cancel while the request is still running.
Expected: Cancel is unavailable while the request runs, as ✕ and Esc are (each dialog sets closeDisabled) and as New Sale and Edit Dates already do.
Actual: Cancel calls onClose and the dialog closes; the request still completes or fails on the server, and its result or refusal is not shown.
Fix: disable Cancel on each dialog’s own pending flag, as NewSaleDialog.tsx and EditSaleDatesDialog.tsx do, with a unit test per dialog pressing Cancel while the request is pending. The fix found the same defect in DraftPriceDialog.tsx (Add a Price and Edit), which is fixed the same way.
Resolved: 2026-11-09 — fixed in forge-et-fe#468 (135c958d, with the review’s changes in f3eb7062). Each of the five dialogs disables Cancel on the flag it passes as closeDisabled, and PricingDialog’s comment says so. A test per dialog (two for Withdraw: one price and the round) holds the request, waits for Cancel to be aria-disabled, presses it and asserts no onClose, then answers the request with a failure and asserts Cancel closes the dialog; all six failed with the source change reversed. Checked by hand against a local backend, with the removal held in the page: Remove Draft’s Cancel was greyed out and did nothing while the request ran; after a 500 the dialog showed its failure and Cancel worked. SetCostDialog, StandardFormDialog and ArchiveStandardDialog pass no closeDisabled and are not this defect; whether they may close while their write runs is in owed.md.
Lesson: L-277 (new) — a dialog that refuses its ✕ and Esc while a request runs refuses its own Cancel on the same flag; frontend-reviewer dimension 9 points at it.
BUG-375 — A lead edit with an extreme estimatedValue (a huge exponent or scale) gets a 500 instead of a 400
Affects: F11-b (forge-et-be lead/web/LeadController.estimatedValueBounds, LeadController.java:340, used by the lead edit) · Found in: F4-b, the third review round of forge-et-be#383, 2026-11-07 · Severity: low — a signed-in user with lead.manage sending a crafted number; no data is changed · Phase: 1
Repro: PATCH /api/v1/leads/{id} with "estimatedValue": 100E2147483647, or with 0E-16384 (zero with a scale above 16383). Not yet run through Jackson; found by reading the code.
Expected: 400 VALIDATION_FAILED on estimatedValue (out of range), as for any other out-of-range value.
Actual (by reading): stripTrailingZeros() runs first and throws ArithmeticException: Overflow for 100E2147483647, and the digit count is an int; a zero with a scale above 16383 passes the check and the database refuses it (invalid scale in external "numeric" value). Either way the request ends in a 500.
Also likely (not checked): the REST create (accept / submit, newLead) — a JSON 0E-16384 may pass @Digits(fraction = 2), which strips trailing zeros, and then fail the insert.
Fix: not decided. Use the rule F4-b’s wave PR puts in LeadFieldRules (estimatedValue: count integer digits as a long before stripping; storedEstimatedValue: store an accepted value at the column’s scale), and call it from the controller’s create and edit paths rather than keeping a second copy. Confirm each repro through Jackson first.
Regression test: a LeadController case with each value, expecting 400; it fails with 500 before the fix.
Resolved: 2026-11-08 — fixed in forge-et-be#384. Through Jackson, the PATCH with 100E2147483647 answered 500 and the POST already answered 400 (@Digits). The patch now checks the value with the shared LeadFieldRules.estimatedValue, which counts integer digits as a long before any stripTrailingZeros(), and the controller’s own copy (estimatedValueBounds) is removed. Every create (the intake’s newLead) and every patch now stores the value at the column’s scale through LeadFieldRules.storedEstimatedValue, as the migration path already did. A zero with a scale above 16383 is therefore stored as 0.00 on a patch, where this entry expected a 400: the shared rule accepts a zero at any scale. On a create, @Digits still refuses that zero with a 400. Neither answers 500. Tests: LeadControllerIT NFR-4 gains 100E2147483647 and 1E-2147483647 on POST and PATCH (the first failed on the PATCH with 500 before the fix), and a new case pins the zero (POST 400, PATCH stored 0.00), which failed before the fix.
Lesson: L-264 (new) — an untrusted BigDecimal’s size is bounded from its precision and scale, counted as a long, before any operation whose cost or result depends on the exponent. BUG-374 is the same cause.
BUG-374 — A workbook number cell with a huge exponent is turned into a plain string of up to about a gigabyte when it is read
Affects: F4-a, F26-b (forge-et-be platform/spreadsheet/SheetCell.of, SheetCell.java:121, read by the Pipedrive import and the price import) · Found in: F4-b, the second review round of forge-et-be#383, 2026-11-07 · Severity: low — only a user who may run an import can upload the file, and Pipedrive and Excel do not write such numbers; one crafted cell could exhaust the backend task’s memory · Phase: 1
Repro: an .xlsx whose number cell holds 1E999999999 as its cached value, uploaded to the Pipedrive dry run or the price import.
Expected: the cell is read as a number and refused by the field’s rule, with no large allocation.
Actual: SheetCell.of builds the cell’s text with number.stripTrailingZeros().toPlainString(), which writes out every digit (about 10⁹ characters) before any rule runs; F4-b’s range check on a lead’s value runs after it.
Fix: not decided. Keep the number as a BigDecimal and build its text with toString() (or refuse an exponent beyond a bound) in SheetCell.of, then check the price import’s and the Pipedrive reader’s specs for any test that reads the plain text. A regression case reads a workbook with such a cell and asserts the refusal without the allocation.
Regression test: a SheetCell case with 1E999999999, which fails with an out-of-memory error before the fix.
Resolved: 2026-11-08 — fixed in forge-et-be#384. SheetCell.of now reads a non-zero cached number outside a double’s range as an ERROR cell with Excel’s #NUM! text, no number and no date. Outside a double’s range means the power of ten of its first digit, counted as a long from precision and scale, is above 308 or below −324. No spreadsheet program writes such a number. Every number and date rule already refuses an error cell: the price import’s figures, and the Pipedrive dry run’s decimal and timestamp columns. A Pipedrive text column reads it as the text #NUM!, as it reads any error cell. This leaves no number for CellText’s setScale(4) to expand either. The largest and smallest doubles and a zero at any scale still read as numbers. Test: SpreadsheetReaderSpec “BUG-374: a number outside a double’s range reads as Excel’s #NUM! error…” (1E999999999, -1E999999999, 1E-999999999, a date-formatted 1E400, and the edges). Before the fix it failed with java.lang.OutOfMemoryError: Java heap space.
Lesson: L-264 (new) — the same cause as BUG-375.
BUG-373 — A paged list answers 500 for a page whose offset passes Integer.MAX_VALUE, where the collection contract says 200 with no items
Affects: FS-10 and every paged list built on it (forge-et-be platform/web/collection/ListQueryResolver bounds page only at Integer.MAX_VALUE; RecordNoteService, LeadNoteService, the account timeline, ActivityService and others call Spring Data with that Pageable) · Found in: F33-a (the backend review of forge-et-be#382, 2026-11-07) · Severity: low · Phase: 1
Repro: signed in, GET a paged list with ?page=21474837&size=100 (an offset above Integer.MAX_VALUE). Not run; reasoned from Spring Data JPA, which refuses a Pageable offset above Integer.MAX_VALUE with InvalidDataAccessApiUsageException.
Expected: docs/api/collections.md § The envelope: a page past the last is 200 with items: [] and the true totals (or the request is refused as 400 OUT_OF_RANGE).
Actual: the exception reaches the advice’s catch-all, so the answer is 500 with an ERROR log.
Fix: not decided. Once, in the platform: either ListQueryResolver refuses page × size above Integer.MAX_VALUE as 400 OUT_OF_RANGE, or ListQuery.toPageable() callers answer an empty page as ActivityListQueriesImpl does. F33-a’s record-attachment list carries its own guard in its wave.
Regression test: a CollectionBindingSpec/resolver case for the bound, and one list IT with that page.
Resolved: 2026-11-08 — fixed in forge-et-be#384, with the 200 answer. That is what docs/api/collections.md, F33-a AC-4 and ActivityListQueriesImpl already gave, so no contract or spec changes (chosen in the fix; Tharinda asked for the bug to be fixed and can reverse the choice in review). The repro’s ?page=21474837&size=100 starts at row 2,147,483,600, still inside the int range. The first page past it at size 100 is 21474838. platform.persistence.EtRepositoryImpl is now the base class of every repository (JpaConfig repositoryBaseClass). For a page that starts past the int row offset, its findAll(Pageable) and findAll(Specification, Pageable) answer no items and the count of the same specification, without running the select, so the reach predicate holds for the total as it does for the rows. That covers the lists built on findAll: the notes, accounts, contacts, tags, certifications, users and the rest. The two native paged reads, the cost list and the published price history, check EtRepositoryImpl.startsPastIntOffset and read new count queries. The timeline and the round reads already bind a long offset, and the activity and attachment lists keep their own guards. Tests, each failing before the fix: CertificationControllerIT (page=21474838 and page=2147483647 at size 100), BuyingStandardCostControllerIT AC-21 and BuyingStandardPriceControllerIT at page 21474838 of 100. No resolver case: the resolver’s bound did not change.
Lesson: none — one-off: the cause is Spring Data’s int first result, which F33-a and the activity list each guarded locally before this fix answered it once in the repository base class. No list built on findAll can reach it again, and docs/api/collections.md tells a hand-written paged query to use startsPastIntOffset.
BUG-372 — Flaky tests: two LeadsScreen.test.tsx cases time out at 20 s in a full Vitest run on a loaded machine (Windows)
Affects: F11-b, F11-a (forge-et-fe src/features/leads/LeadsScreen.test.tsx: “a failed option read stays inside its select (AC-14a) › a failed tag list shows its error beside the Tag select, and the list, filters, search and lead page work” and “the trace id beside each read failure (coding-standards §3.2) › leads-stale-trace: a list refresh shows the trace id when the problem body has one, and nothing when it has none”) · Found in: F3-a wave 3’s local test record (pnpm test:record, every test, as a moved file made the selection fall back to all) on forge-et-fe feature/F3-a-wave-3 4ccb40cc, Windows 11, 2026-11-06 · Severity: low — tests only; a failing run blocks a local pnpm test or pnpm test:record · Phase: 1
Repro: that run, 12349 passed, 1 skipped and 2 failed (897.6 s), both Error: Test timed out in 20000ms. on the first try and on the retry (40.9 s and 46.0 s in all), with the machine’s load at 0.89 average and 1.00 peak from other sessions’ work (load 0.24 at the start). The file run alone straight after passed 60 of 60 (117.8 s). The tag-list case also timed out once in F3-a WI-3.1’s subtask 6 run of src/app src/features/leads src/features/accounts (2026-11-03) and passed when re-run alone.
Expected: each case finishes well inside its timeout whatever the run’s load.
Actual: under a full run’s load each case runs past 20 s; alone the whole file takes about 2 minutes, so each case is slow even unloaded.
Fix: both cases did more work than their budget allowed (.forge/test-strategy.md § Flaky Tests cause 6), most of it one step: choosing Sweden from the Country filter, whose popover renders about 250 countries and whose findByRole('checkbox', { name }) works out the accessible name of every one. Timed on an idle machine, one pick took about 1.5 s (0.5–0.8 s to open, about 0.9 s to find and tick); leads-stale-trace picked Country twice, once per problem body, and the tag-list case once plus its whole walk. The refresh in leads-stale-trace is now the Approved status pill, and each trace-id row is two cases (with a trace id, without), so no case renders the screen twice. The AC-14a case is four, each with its own render: the Tag select’s error, its empty and usable list, and the Owner filter (three options) reaching the list read; the search reaching the list read; Retry re-reading only the tags; a lead opening from the list while the tag read is still failing. The first two end by checking the tag failure is still beside its select only, after the list re-read. The Country filter’s own failed refresh stays covered by “a failed refresh keeps the rows, marked stale, with Retry”. No timeout was changed.
Regression test: the file passes in three full pnpm test runs in a row on a loaded Windows machine.
Resolved: 2026-11-11 — fixed in forge-et-fe#490 (424b3ed7; review fixes 5d1344cb), test only. Windows 11, 12 cores. Under forced load (two busy worker_threads loops per core), main’s copy of the file took 11.1 s for the tag-list case and 16.9 s for leads-stale-trace, against the 20 s timeout; with the fix the split cases took 1.5–2.8 s. Three full pnpm test runs in a row on the fix: LeadsScreen.test.tsx passed 82 of 82 each time, the split cases 1.2–5.2 s, at 84%, 98% and 97% average CPU load (peak 100% each, sampled every 10 s). Each of those runs also failed one unrelated case, a different one each time, each once at about 100% load, so none is filed (test-strategy.md § Flaky Tests F-2): LeadsScreenActions.test.tsx “a create re-reads the list and opens the new lead’s page…”, MailboxConnectProvider.test.tsx “the page button shows Try again…”, OwnedRecordsList.test.tsx “after a pass that moved everything (a toast)…”. /forge-review-pr with frontend-reviewer found no Blocker or Important; its four Nits: three fixed in 5d1344cb (comments in the present tense, case names that say the filter and search reach the list read, the tag failure checked after the re-read), and the fourth, five more cases in the file that pick Sweden from Country without testing it, has a row in .forge/owed.md. The cases are now named “a failed option read stays inside its select (AC-14a) › a failed tag list shows its error beside the Tag select, which stays usable, and another filter still reaches the list read”, “… › with a failed tag list, the search still reaches the list read”, “… › Retry beside a failed tag list re-reads only the tags”, “… › with a failed tag list, a lead still opens from the list”, and “the trace id beside each read failure › leads-stale-trace: a list refresh › shows the trace id when the problem body has one” / “… › shows nothing when the problem body has none”.
Lesson: L-125 — another case that did several cases’ work in one test budget, its largest step a pick from the Country filter’s 250 options, the same as BUG-347 in AccountListScreen.test.tsx. Added to L-125’s sources.
BUG-371 — Flaky test: the size matrix’s leads-lead-ranking-done state at 1280 × 800 in Chromium finds no lead view within the 5 s ready wait (Windows)
Affects: F11-g (forge-et-fe e2e/browser/size-matrix/size-matrix.spec.ts:65, waitUntilReady on getByTestId('leads-view'); state leads-lead-ranking-done, project size-1280x800-chromium) · Found in: F3-a wave 3’s direct size-matrix run (ET_BACKEND_PORT=8091 ET_E2E_WORKERS=4 pnpm e2e --forbid-only e2e/browser/size-matrix/size-matrix.spec.ts --grep "account-detail|leads-lead", the app served by vite preview) on forge-et-fe feature/F3-a-wave-3 b2fc26b4, against forge-et-be main 5a14884b on a freshly reset database, Windows 11, an idle machine (CPU 10–19% before the run), 2026-11-06 · Severity: low — one test fails now and then; F3-a does not change the Priority panel or the lead view’s loading · Phase: 1
Repro: that run, 1569 passed and 2 failed (38.6 min): expect(locator).toBeVisible() failed, getByTestId('leads-view') not found within 5000 ms, 5.7 s into the test; the same state passed at the other four Chromium sizes in the run. Run alone straight after on the same head and backend (--project=size-1280x800-chromium --project=size-768x1024-webkit --grep "leads-lead-ranking-done$|leads-lead-notes-refreshing$" --repeat-each=5 --retries=0, ET_E2E_WORKERS=2), the state passed 10 of 10.
Expected: the lead view shows within the ready wait at every size.
Actual: now and then the lead view has not rendered when the 5 s ready wait ends, with no other failure in the test.
Fix: not decided. Triage per .forge/test-strategy.md § Flaky Tests: find what holds the lead view back (its pinned reads, the Priority panel’s lazy chunk, or the route’s first render). This is the ready-wait miss that BUG-218’s vite preview build ended for the dev server; a longer ready wait is not the fix.
Regression test: the state passes at every size in three full size-matrix runs in a row.
Investigated (2026-11-08, Nanndish, Windows; not reproduced, no code change): the leads-view frame renders even while the lead is loading, so the only steps before it are the page load, /me, the LeadPageRoute chunk (about 35 JS files and 5 stylesheets, requested once /me answers) and a render. The Priority panel’s lazy chunk and the pinned reads are not on that path, and the ready wait already targets that first render. 30 repeats at size-1280x800-chromium with 4 workers on the shared, loaded machine passed 30 of 30, with the lead view 0.8–2.1 s after goto started. 20 repeats under one busy loop per core passed 20 of 20: goto took 0.7–5.5 s, and the span the 5 s ready wait covers (from goto returning to the lead view) stayed at or under 2.6 s. A miss on an idle machine therefore needs one of those steps to stall about 4 s, and the original run kept no trace to name which. Two possible changes, neither made: start loading the matched route’s chunk alongside /me rather than after it (shortens the chain, but does not prevent a stall), or have the size matrix attach its pending-request list when a ready wait misses, so the next occurrence names the late step. Left in-progress for the lead to decide (owed.md).
Resolved: 2026-11-09 — closed as wontfix: not reproduced; Nanndish decided to check once more and close it if the miss did not recur. The original command run again (ET_E2E_WORKERS=4 pnpm e2e --forbid-only e2e/browser/size-matrix/size-matrix.spec.ts --grep "account-detail|leads-lead", the app served by vite preview) on forge-et-fe 25531363 (main with BUG-376/383/384’s fixes, none of which touches the lead view) against forge-et-be main 0cbda764, Windows 11, CPU about 12–28% at the start: 1662 passed, none failed or retried, in 30.1 min, leads-lead-ranking-done included at all seven sizes. With the 50 passing repeats of 2026-11-08, the one miss in 1571 stays unexplained, and neither code change was made. A recurrence is a new bug that links this one, and its run keeps a trace.
Lesson: none — one-off: a single ready-wait miss on an idle machine that 51 later runs of the state, including the same command, never repeated; no product or test cause was found.
BUG-370 — Flaky tests: four of leadPanels.approveSend.test.tsx’s cases miss the lazily loaded Approve & send within findBy’s 1 s default in a batched Vitest run (Windows)
Affects: F11-h, F3-a (forge-et-fe src/app/leadPanels.approveSend.test.tsx:121, :137, :146, the cases that wait for LeadApproveSendAction’s chunk: “the arriving button is drawn as its stand-in was, and takes its place”, the two BUG-269 “once the chunk has loaded …” cases and “once the chunk has loaded, shows the button with no stand-in”) · Found in: F3-a wave 3’s check of its origin/main merge, pnpm exec vitest run over 88 files (src/app/leadPanels*.test.*, src/features/leads/LeadView.test.tsx, src/lib, src/app/leadFollowUpsFlow.test.tsx, src/features/leadRanking) on forge-et-fe feature/F3-a-wave-3 7d9f6df4, Windows 11, 2026-11-06 · Severity: low — tests only; a failing run blocks a local pnpm test or pnpm test:record · Phase: 1
Repro: that run, 1033 passed and 4 failed (130.6 s): Unable to find role="button" and name "Approve & send" and Unable to find an element by: [data-testid="ai-drafts-toast-live"], each after about 1 s, with the group still holding ai-drafts-approve-send-placeholder (the chunk had not resolved). The file alone straight after passed 10 of 10, and the same 88-file batch run again passed 1037 of 1037 (81.1 s, CPU about 52% from other work).
Expected: each case waits for what the chunk’s arrival causes, whatever the run’s load.
Actual: under a batched run’s load the import() of LeadApproveSendAction takes longer than findBy’s default 1 s, so the case fails on the stand-in.
Fix: not decided. Triage per .forge/test-strategy.md § Flaky Tests: resolve the chunk before the assertion (await the loader the composition exports, loadLeadApproveSendAction, in the case’s setup, so the wait no longer measures module loading), rather than a longer timeout.
Regression test: the file passes in repeated loaded batch runs on Windows.
Resolved: 2026-11-08 — fixed in forge-et-fe#451. Each case that needs the loaded chunk now calls loadChunk() before it asserts. It calls loadLeadApproveSendAction(), releases the chunk, waits for a mark the mock’s factory sets in a finally when its import has finished, then awaits that loader call and checks it answers with the real component. Awaiting the loader alone, as the Fix: line proposed, was not enough: an import() made while the mock’s factory is running skips the mock and answered after 61 ms. The loader is called before the release so that each case also runs on its own; the first version waited for an import only an earlier case had started, and the review of forge-et-fe#451 found its loaded-chunk cases hung when run alone. composeEmailComposition.aiDraft.test.tsx held the AI draft button’s chunk the same way and had the same race (the review found it): its old version failed 2 of 3 with the import held 1.5 s, and it gets the same loadChunk. Fails before: with the chunk’s import held 1.5 s longer in the mock, the old file failed 2 of 10, the arriving-button case and the archived-lead BUG-269 case, each after about 1 s, as in the repro. Passes after: with the same hold, 10 of 10 in three runs, and each loaded-chunk case alone; 10 of 10 without the hold. The AI draft file: 3 of 3 held, three times, and each case alone (Windows 11).
Lesson: L-116 — a component test that renders a lazy() module loads it before it asserts; this PR adds BUG-370 to the lesson’s sources, with an update on waiting for a mocked module’s import.
BUG-369 — Flaky test: the size matrix’s account-detail-share-dialog state at 1024 × 768 in WebKit waits out the test timeout on a click on the share dialog’s search box (Windows)
Affects: F5-f (forge-et-fe e2e/browser/size-matrix/arrange.ts:2533, openPinnedShareDialog’s click on sharing-dialog-search; state account-detail-share-dialog, project size-1024x768-webkit) and F11-f (e2e/browser/leads/support.ts:1695, showLeadTab’s click on the Notes and email tab; state leads-lead-notes-refreshing, project size-768x1024-webkit) · Found in: F3-a wave 3’s direct size-matrix run (ET_E2E_WORKERS=4 pnpm e2e --forbid-only e2e/browser/size-matrix/size-matrix.spec.ts --grep "account-detail|leads-lead", the app served by vite preview) on forge-et-fe feature/F3-a-wave-3 27edbb39, against forge-et-be main c26cdba1 on a freshly reset database, Windows 11, CPU about 19% at the start, 2026-11-06 · Severity: low — one test fails now and then; the wave does not change the sharing dialog · Phase: 1
Repro: that run, 1444 passed and 1 failed (30.8 min): locator.click: Test timeout of 30000ms exceeded, the search box resolved and Playwright “waiting for element to be visible, enabled and stable” until the timeout. Run alone straight after on the same head and backend (--project=size-1024x768-webkit --project=size-768x1024-webkit --grep "account-detail-share-dialog(?!-)" --repeat-each=5), the state passed 10 of 10. In an earlier run on the same head under heavy load it failed once at 1280 × 800 in Chromium, among 214 load timeouts.
Also seen (F3-a wave 3’s final direct run on b2fc26b4, idle machine, 2026-11-06): the same stall on another control. leads-lead-notes-refreshing at size-768x1024-webkit: locator.click on leads-view-tab-notes (the Notes and email tab, e2e/browser/leads/support.ts:1695, showLeadTab from settleLeadNotes) resolved the tab and waited for “visible, enabled and stable” until the 30 s test timeout. Run alone straight after (--repeat-each=5 --retries=0), it passed 10 of 10.
Expected: the arrange step’s click lands once the search box is visible, enabled and stable, at every size.
Actual: now and then in WebKit the click waits the whole test timeout on a search box that never reads as stable.
Fix: not decided. Triage per .forge/test-strategy.md § Flaky Tests: find what keeps moving (the dialog’s open animation, a layout change as the candidate list loads) or whether this is the WebKit stall BUG-013 and BUG-183 recorded, which BUG-218’s vite preview build was meant to end. A longer timeout is not the fix.
Regression test: the state passes in WebKit at both tablet sizes in three full size-matrix runs in a row.
Investigated (2026-11-08, Nanndish, Windows; no code change): both original call logs (worktrees/F3-a/verify-runtime-w3/matrix-run5.log and matrix-run7.log on Nanndish’s machine) end at attempting click action → waiting for element to be visible, enabled and stable, then the test timeout, with no element is not stable line and no retry. So the element never moved: Playwright’s in-page stability check never finished, because it needs animation frames (on Windows WebKit, Playwright 1.63 asks for 5 frames at least 15 ms apart), and the page drew none for about 25 s. The same signature hit unrelated controls in the overloaded run on the same head (matrix-run.log: the compose Send button, leads-customer-change, leads-match-save), and matrix-run7 also lost a Chromium state to a load timeout. Neither control has anything that moves: the Access dialog fades opacity only, and the tabs’ transition-all changes colour only. Not reproduced: under 16 busy loops with 4 workers, both states at both WebKit sizes ×5 gave 20 failures, all ordinary load timeouts in other waits, none with this signature; a 90-test WebKit probe run and a 500-state WebKit run with a frame-and-timer probe showed no such stall, and every frame gap came with a timer gap (the page’s main thread was blocked, not a WebKit painting fault). While re-testing, Windows’ System process used about 490% CPU and Defender about 110%, beside other sessions. A test-side change would have to skip the actionability check (force, dispatchEvent) or lengthen a timeout, and both weaken the test. Proposed wontfix as an overloaded machine (.forge/test-strategy.md § Flaky Tests step 4), for the lead to decide (owed.md); the evidence is weaker than step 4 asks, because those were direct pnpm e2e runs with no result.json load record. A stability-wait call log with no element is not stable line is the sign to look for in triage.
Re-checked (2026-11-09, Nanndish, Windows; no code change): the original command run again on forge-et-fe 25531363 (main with BUG-376/383/384’s fixes, none of which touches the sharing dialog or the lead tabs) against forge-et-be main 0cbda764, CPU about 12–28% at the start: 1662 passed, none failed or retried, in 30.1 min, account-detail-share-dialog and leads-lead-notes-refreshing passing at both WebKit tablet sizes. One full run toward the regression test’s three; still a direct pnpm e2e run with no result.json load record.
Resolved: 2026-11-09 — closed as wontfix: an overloaded machine, not the test or the product (.forge/test-strategy.md § Flaky Tests, step 4); Nanndish decided, after the re-check above passed. Both failures end in a stability wait with no element is not stable line: the page drew no frames for about 25 s, so Playwright’s in-page check never finished, and the same signature hit unrelated controls in an overloaded run on the same head. Neither control moves, forced load did not reproduce the signature, and the original command passed in full on a near-idle machine. A test-side change would have to skip the actionability check or lengthen a timeout, both of which weaken the test, so nothing was changed. The evidence is weaker than step 4 asks, because the failing runs were direct pnpm e2e runs with no load check or result.json load record. A recurrence is a new bug that links this one; its run goes through the browser gate (or keeps its own load record), and its call log is checked for an element is not stable line first.
Lesson: L-267, another instance — the failing runs were direct pnpm e2e runs with no load check or load record, so overload could be inferred from the call logs but not shown; L-267’s sources now list BUG-369.
BUG-368 — Every email send from the CRM fails on dev: Graph answers the send of the draft with 411 Length Required
Affects: F20-c (forge-et-be GraphMailboxProvider.sendDraft) · Found in: F38-g WI-4.2’s walk on dev, 2026-11-06 · Severity: high — no email can be sent from a deployed CRM · Phase: 1
Repro: on dev, as a user holding email.manage with a connected mailbox, Compose email on an account, any recipient, Send. The answer is 502 EMAIL_SEND_FAILED “The email was not sent.” (trace 6aedc52dd25ea201ac3d09b42630c93b). /ecs/dev/backend logs Email Send Failed - [sendId=8, connectionId=10] [step=send, httpStatus=411, graphErrorCode=null], and the same for send 7 (connection 5) at 11:39 UTC. The draft is created; the next call, POST /me/messages/{draftId}/send, fails.
Cause: sendDraft posts with no body. The send path’s RestClient runs over the JDK HttpClient, which speaks HTTP/2 to Graph and sends no content-length on an empty request, and Graph’s front end refuses such a POST with 411 before it reads the token. Reproduced with the JDK client and no token, so nothing was sent: HTTP/2 with no body → 411, HTTP/2 with an empty byte body → 411, HTTP/2 with the JSON text {} and a stated length → 401 InvalidAuthenticationToken, HTTP/1.1 with no body → 401. A bodiless HTTP/2 DELETE answers 401, so the adapter’s deletes are not affected. The local stand-in mailbox and MockRestServiceServer never open a connection, so no test could see it.
Expected: the draft is sent and the send answers 201.
Actual: 502 EMAIL_SEND_FAILED, the draft deleted.
Fix: sendDraft sends the JSON text {} as a String with Content-Type: application/json; the string converter states its length, so the request goes with content-length: 2. (A Map.of() body, as the reply and forward calls send, goes through the Jackson converter with no stated length — the review of forge-et-be#378 caught that.) Regression test: GraphMailboxProviderSpec’s send case expects that body, content type and Content-Length: 2. Proved on dev after the merge’s deploy by a real send.
Resolved: 2026-11-06 — fixed in forge-et-be#378 (4b866e60, c57abefa). GraphMailboxProviderSpec failed 1 of 91 on the code before the fix (“Content type not set”) and on the first fix’s Map.of() body (“Expected header c57abefa: 12910 tests, 0 failed. Every other POST in forge-et-be src/main/java already carries a body (a grep of .post() on the branch). The real send on dev is §B F20-c item (a) in .forge/cloud-gated-validation.md, walked after the deploy.
Lesson: none — one-off: the adapter’s only bodiless POST; a fake that opens no connection (the stand-in mailbox, MockRestServiceServer) cannot see a protocol-level refusal, which is what §B F20-c item (a)'s real send on dev exists for.
BUG-359 — The Leads list’s priority filter and sort take 2–4 s at 20,000 leads, because the list rebuilds every lead’s ranking fingerprint per request
Affects: F11-g (forge-et-be LeadRankingRepository.CURRENT_PRIORITY_SQL, findListPriorities, priorityIn, the priority sort) · Found in: the F11-g wave 1 review, forge-et-be#372, re-timed on realistic data, 2026-11-06 · Severity: medium — the list misses its 300 ms target (F11-b NFR-1) on these requests; at today’s volume of a few hundred leads users do not feel it · Phase: 1
Repro: load src/test/resources/db/testfixture/lead-volume.sql and lead-ranking-volume.sql (20,000 leads, 90% with research done and findings), then time GET /api/v1/leads?sort=priority,asc, ?priority=P1 and ?priority=P1&sort=priority,asc as the administrator. The medians are 2.2 s, 3.3 s and 4.4 s; sort=receivedAt,desc is 65 ms.
Expected: each under 300 ms at 20,000 leads.
Actual: F11-g contract D-2 makes the list decide whether each ranking is current by recomputing the lead’s input fingerprint in SQL (LeadRankingSql.FINGERPRINT: two SHA-256s, the domain expressions and the findings aggregation). A priority filter or sort does that for every lead in reach, for the page and again for the count.
Fix: decided by Dinil, 2026-11-06: mark a ranking out of date the moment its lead’s ranked details change, so the list reads a stored mark instead of recomputing. The plan:
- Migration
V63:lead_ranking.stale boolean NOT NULL DEFAULT false, and a partial index on(lead_id)wherestalefor the sweep’s repair. - Recompute the mark after every lead
PATCHand every research write. In a transaction of its own after the commit, runUPDATE lead_ranking SET stale = (input_fingerprint <> :current) WHERE lead_id = :id, with:currentfrom the Java reader (LeadRankingInputsReader, which applies spec Rev 2’s kept inquiry-only ranking).- It runs whether or not a re-rank is queued. A
contact_namechange sends research back topendingand so changes the ranking basis, though F11-g FR-J1 queues nothing for it. - It recomputes the mark, never only sets it. A grade changed A→B→A before the re-rank must end unmarked, or the mark would stay true forever.
- It leaves the row’s
versionand the ranking fields unchanged. - What cannot make a ranking stale: a status change, an owner change or the research wait passing (spec FR-A3 Rev 2).
- It runs whether or not a re-rank is queued. A
- The writer sets the mark:
LeadRankingWriterwrites it the same way under the lead lock, and records the audit event below when it changes it. In practice that isfalse, because the writer already skips a message whose fingerprint is no longer current. - The sweep repairs missed marks, in both directions: the hourly sweep, which keeps the SQL fingerprint, first finds without writing the rows whose mark disagrees with it and the rows already marked stale at any lead status, up to its batch size. It then recomputes each one as after a
PATCH: the lead’s lock first, then the Java fingerprint, then the mark, so a job or aPATCHcommitting meanwhile cannot leave a wrong mark. A mark left on because the step after aPATCHfailed is cleared this way. This run’s candidates come first, so rows already marked cannot crowd them out of the batch. It logsLead Ranking Stale Mark Repairedat WARN once per run, with the count and the lead ids, so a missed recompute shows in the log;Lead Ranking Stale Repair Unchangedat WARN names rows whose recompute changed nothing (the SQL and Java fingerprints disagree on them, or a write since the select already set them); a failed recompute is logged at WARN, or at ERROR when every row failed or the select itself failed (coding-standards §8.3). - Audit: each change of the mark records
lead.ranking_stale_changed(staleold and new) in the same transaction (AD #36). The actor follows the audit service’s rule: the editing user after aPATCH;SYSTEMwith joblead-researchafter a research write;SYSTEMwith joblead-ranking-sweepin the sweep;SYSTEMwith joblead-rankingwhen a ranking run’s write changes it. - Its own transaction: the recompute runs in a
REQUIRES_NEWtransaction after the commit, which locks the lead row first, as the writer does. It skips the lock when the lead has no ranking row. A failed recompute is logged at WARN and does not stop the re-rank from being queued. - The list reads the mark: the current priority becomes
r.prioritywherer.state = 'done' AND NOT r.stale, joined bylead_id, for the row member, thepriorityfilter and the sort. The single ranking read (GET /api/v1/leads/{id}/ranking) keeps computing the fingerprint, which is one lead. - No backfill: dev has no ranking rows (its queue does not exist yet), and a local database’s rows are repaired by the first sweep.
- Wire shape unchanged: a stale ranking still reads
priority: null, matchesNONEand sorts last (contract D-3).
The contract’s D-2 and the spec’s list paragraph are revised in this PR. The tests:
- the stale-list cases (AC-19a) through the new mark, and every case of the new AC-19b:
- grade A→B marks and B→A clears;
contact_namemarks;- owner and estimated value do not mark;
- a research write marks;
- a
PATCHduring the AI call never leaves a current-looking row; - one audit event per change;
versionis unchanged;
- the sweep’s repair and its WARN;
LeadRankingFingerprintIT’s list part replaced by the mark;- a re-timing at 20,000 leads on the realistic fixture, each median under 300 ms, recorded in the PR.
Regression test: the re-timing above, recorded in the fixing PR; and an IT that a stale ranking marked by a
PATCHreadsnullin the list without the SQL fingerprint being evaluated. Resolved: 2026-11-06 — fixed in forge-et-be#376. - The fix: V63 adds
lead_ranking.staleand a partial index on the marked rows (079548e3), and the list reads the mark (4426d4c5). - The end-to-end tests:
2fc00b9f, plus the timing fixture’s out-of-date rows (77c9f573). - Review fixes:
- the sweep finds repairs without writing and recomputes each under the lead’s lock, in both directions (
5d780cb5;LeadRankingStaleRaceIT); - a run’s write that clears the mark is audited (
2a52a542); - the candidates come first and every repair outcome is logged at the coding-standards §8.3 levels (
be1cf329;LeadRankingSweepServiceSpec); - Javadoc and log text (
0ac20cad,b5486083,5961ff1b).
- the sweep finds repairs without writing and recomputes each under the lead’s lock, in both directions (
- Fails before (PR comments 6680959243 and 6682184550):
LeadRankingStaleMarkITat079548e3: 4 of 6 failed;LeadRankingStaleRaceITat77c9f573: 2 of 2 failed, with a current row leftstale = true;LeadRankingSweepIT’s clear case at77c9f573failed;LeadRankingSweepServiceSpecat6a5afc4a: 4 of 4 failed.
- Timing: medians at 20,000 leads on the realistic fixture were 67, 86, 108 and 102 ms, against 65 ms, 2.2 s, 3.3 s and 4.4 s before.
- Full build at
b5486083: 12,910 tests, 0 failed. Lesson: L-258 — this is the defect its fixture rule found; the re-timing used the realistic fixture. L-156 — another instance: the sweep’s one-statement repair decided a write from rows a concurrent writer changes without taking the writers’ lock.
BUG-358 — A lead whose only detail is a website makes a paid ranking call with no buyer text when its research finds nothing
Affects: F11-g (forge-et-be LeadRankingInput.nothingToRank(), userText()) · Found in: the F11-g wave 1 review, forge-et-be#372 (adversarial pass), 2026-11-06 · Severity: low — rare (a lead created by hand or migrated with only a website filled in) and cheap, but it ranks such a lead P3 against the prompt’s “never lower for no research” · Phase: 1
Repro: create a lead with only website set (no inquiry text, company name or tea requirement) while the search stub answers nothing_found (STUB-NOTHING) or research is unavailable; let research write and ranking run.
Expected: no AI call; the lead is written no_input and the panel says there is not enough to rank it, as for an empty lead (spec FR-A4).
Actual: nothingToRank() counts the website as something to rank, but userText() never sends the website (WI-1.1 plan Decision 13). So the model is called with an empty user text and only the country, buyer type and source, and is likely to answer P3 with low confidence.
Fix: decided by Dinil and Tharinda, 2026-11-06 (option (a)). The website only feeds research:
nothingToRank()is true when there is no inquiry text, company name or tea requirement and no company research findings.- A website-only lead still waits for its research (FR-J3). Research that finds the company ranks it from the findings. Research that finds nothing, cannot run, or has not finished within the 24-hour research wait writes
no_input, with no call. This is the one stated exception to decision 3’s “never left unranked” (spec Rev 3).
Spec FR-A4 is revised in this PR (Rev 3).
Regression test: a website-only lead with nothing_found research, and one with unavailable, are each written no_input with no ai_invocation row; a website-only lead whose research found the company is ranked with basis = 'inquiry_and_research'.
Resolved: 2026-11-06 — fixed in forge-et-be#376.
- The fix:
LeadRankingInput.nothingToRank()no longer counts the website, so a lead with only a website is ranked only from its research findings (6a0df32f). - Test: a lead with only a website and a country is also covered (
6a5afc4a). - Fails before: on
5a14884b, 5 of 37LeadRankingJobITtests failed (each website-onlyno_inputcase was rankeddone), and 3 of 30LeadRankingAnalyserSpectests failed (PR comment 6680959243). - After: both pass, and so does the full build. Lesson: none — one-off: the check that decides “nothing to rank” counted a field the AI call never sends; the spec now says the website only feeds research.
BUG-367 — Flaky test: the compose template test whose fill-in 404 re-reads the chosen search finds the settled search instead (Windows)
Affects: F20-d (forge-et-fe src/features/email/compose/ComposeEmailDialog.insertTemplate.test.tsx:949, “a fill-in 404 re-reads the search the user chose from, not one settled after the choice”) · Found in: BUG-360’s Vitest run, 2026-11-06 · Severity: low — its file passed when run alone · Phase: 1
Repro: Windows 11, pnpm exec vitest --run src/features/email src/ui src/app/shadcnOnly.test.ts (174 files) on fix/BUG-360-email-screen-layout, which changes no compose or template code, with the CPU at 100% (three other Claude sessions and Docker Desktop running). The re-read requests carried q=intro where the test expects none. The file passed 46 of 46 when run alone straight after, at the same load.
Expected: the re-read after the 404 uses the search the user chose from.
Actual: once, the search that settled after the choice.
Fix: the case typed intro with user-event and then pressed ArrowDown and Enter, assuming both finish inside the search’s 300 ms debounce (SEARCH_DEBOUNCE_MS). When the typing took longer, the search settled before the choice, so the choice came from the searched list and the re-read rightly carried q=intro. The product is right; the test’s timing is wrong. Set the text, press ArrowDown and click the focused item as synchronous events, so no timer can run between them.
Regression test: the same case: the old version fails when 350 ms pass between the typing and the choice, and the new one has no such gap.
Resolved: 2026-11-08 — fixed in forge-et-fe#451. The case now runs fireEvent.change, fireEvent.keyDown (ArrowDown) and fireEvent.click on the item ArrowDown focused, all in one synchronous step, and waits SEARCH_DEBOUNCE_MS + 100 after the choice for the search to settle. Vitest’s fake timers were tried first and dropped: Testing Library’s user-event wrapper waits on a setTimeout(0) that only Jest’s fake timers advance, so type hung. Fails before: the old case with a 350 ms pause before the choice failed with expected [ 'intro', 'intro' ] to deeply equal [ null, null ], the repro’s failure. Forced load alone (one busy loop per core) did not reproduce it with one file running: old and new each passed 5 of 5. The file passes 46 of 46 (Windows 11).
Lesson: L-125 — a test that assumes a fixed amount of work finishes inside a real-time window fails on a loaded machine; this PR adds BUG-367 to the lesson’s sources.
BUG-366 — Flaky test: the WebKit 1024x768 case of the Email panes test still shows Loading your mail when its 5 s wait for the list ends (Windows)
Affects: F20-i (forge-et-fe e2e/browser/email/inbox-message-view.spec.ts:261, “the two panes fill the page in WebKit at 1024x768 too”; e2e/browser/mailbox/support.ts:581 settleMailboxScreen) · Found in: BUG-362’s browser check, 2026-11-06 · Severity: low — passed when run alone · Phase: 1
Repro: Windows 11, pnpm exec playwright test e2e/browser/email/inbox-message-view.spec.ts --project=chromium -g "fill the page" --retries=0 on fix/BUG-360-email-screen-layout, with the CPU at 100% (three other Claude sessions and Docker Desktop running). The WebKit case’s settleMailboxScreen found no inbox-list within toBeVisible’s 5 s; the page snapshot showed Loading your mail. Run alone straight after, at the same load, it passed in 10.6 s.
Expected: the wait for the list allows for a first WebKit page load on a loaded machine, or the run starts only below the load limit (.claude/rules/test-record.md).
Actual: a 5 s wait that a slow first load can exceed.
Fix: wait for the first-page list read’s answer before the list’s wait, so the 5 s covers only the render.
Regression test: with the list read held back 4.5 s, the old test failed 4 of 4 (snapshot Loading your mail) and the new one passed 6 of 6.
Resolved: 2026-11-08 — fixed in forge-et-fe#459 (3cc46298, with the review’s named 20 s bound on the response wait in 21a0a442). No product defect. The test went to /email and gave the list 5 s, which covered the whole chain the list read ends: the page load, /me, the EmailRoute chunk, the connection read, a render, then the list read, which itself answers within about 100 ms. Normally the list showed 0.9–1.1 s after goto started. Under one busy loop per core it came 3.2–4.3 s after goto returned, close to the limit. A cold WebKit browser launched inside the test adds more. expectPanesFillPage now registers a wait for the first-page list read’s response (LIVE_PAGE_URL, no cursor parameter) before goto, then calls settleMailboxScreen. With the read held back 4.5 s the old test failed 4 of 4 and the new one passed 6 of 6. Without the hold, it passed 20 with --repeat-each=5, and 12 after the review’s change. Under forced load one WebKit case ran past its 30 s total with goto alone taking 3–7 s, which is overload, not this race.
Lesson: L-101 (another instance: a wait for an element that covered a chain of steps; it now waits for the step’s outcome first; BUG-366 added to its sources).
BUG-365 — Flaky test: the compose forward test that moves focus to the first original’s Remove after a Retry finds focus on the page body (Windows)
Affects: F20-h (forge-et-fe src/features/email/compose/ComposeEmailDialog.forwardAttachments.test.tsx, “moves focus to the first original’s Remove once a Retry lists them (F20-h WI-1.2 plan Decision 15)”) · Found in: BUG-360’s Vitest run, 2026-11-06 · Severity: low — passed when its file ran again · Phase: 1
Repro: Windows 11, pnpm exec vitest --run src/features/email src/ui/shadcn src/ui/icons src/app/shadcnOnly.test.ts (100 files) on fix/BUG-360-email-screen-layout, which changes no compose code, with the CPU at 100% (three other Claude sessions and Docker Desktop running). toHaveFocus on email-compose-original-attachment-remove found focus on <body>. The same file passed in the next run with three other files.
Expected: focus is on the first original’s Remove once Retry has listed the originals.
Actual: once, focus on <body> at the assertion.
Fix: release the held Retry read inside act, so the three renders before the focus move are flushed before waitFor’s second starts.
Regression test: the 101-file batch above, under one busy loop per core, failed this case and its Safari sibling on the old code and passed both twice after.
Resolved: 2026-11-08 — fixed in forge-et-fe#456 (38d150c3). No product defect. After the held Retry read is released there are three renders of the whole compose dialog before focus moves: the list, the settled Retry’s count (whose effect asks for focus), and the focus effect. They took 0.36–0.6 s idle and up to 2.2 s with one busy loop per core, against waitFor’s 1 s. The product’s order is correct, and guaranteed by the code: TanStack’s refetch resolves only after the observer’s result is updated, so the render the count causes always sees the list loaded (10 of 10 timed runs agreed). The case alone did not fail under that load (5 of 5 passed). The original 101-file batch did: this case and its sibling “…when the Retry press focused nothing (Safari)” were its only 2 failures of 2195, both with focus on <body>. Both cases now release the held read inside await act(async () => …) before the existing waitFor. Both passed the same loaded batch twice (one unrelated failure in run 1, EmailRoute.test.tsx “reads the page and the tree again and shows the list, not the prompt”; run 2 passed 2195 of 2195).
Lesson: L-270 (new, with BUG-382).
BUG-364 — An open message’s primary Reply button comes first among its actions, where every other screen puts the primary action last
Affects: F20-i (forge-et-fe src/features/email/MessageView.tsx) · Found in: Nanndish’s UI review of the Email screen, 2026-11-06 · Severity: low — the order differs from the rest of the CRM; nothing fails · Phase: 1
Repro: open Email and choose a message, in the reading pane or on its own page.
Expected: the primary action last, at the right end, as design-system §6 Screen header and Rev 72 set for every screen. Nanndish’s decision, 2026-11-06: Reply stays the primary button and moves to the right end.
Actual: Reply (primary), Reply all, Forward, then Mark as unread and Refresh.
Fix: the two action groups swap and each is reversed: Refresh, Mark as unread, then Forward, Reply all, Reply. The actions are right-aligned, so Reply stays at the right edge when they wrap under the subject in the reading pane. Tab order follows the visual order. The design is the Email Screen Redesign artifact (https://claude.ai/artifact/Mw9xaBAjGoxAbbSGMnM7Cy); design-system Rev 101 records it.
Regression test: MessageView.test.tsx › “renders the header with the actions in it …” (Forward, Reply all, Reply, with Reply last) and “Tab reaches the breadcrumb, Back to Email, Refresh, Mark as unread, Forward, Reply all, Reply and the download in order …”; both fail on main. In the browser after the fix: Refresh message, Mark as unread, Forward, Reply all, Reply, right-aligned in the reading pane.
Resolved: 2026-11-06 — fixed in forge-et-fe#435 (a067dba3, 039a1fcb), with BUG-360 to BUG-363.
Lesson: L-254 — another instance: the message view took its own order for a screen’s actions where Rev 72 sets the primary action last everywhere. L-254’s sources now list it.
BUG-363 — A received attachment is a full-width row with a Download button, not a tile that downloads when clicked as in Outlook
Affects: F20-i (forge-et-fe src/features/email/MessageAttachments.tsx) · Found in: Nanndish’s UI review of the Email screen, 2026-11-06 · Severity: low — the download works; the list takes a row per file and does not look like Outlook’s · Phase: 1
Repro: open a message that carries a file attachment.
Expected: Nanndish’s decision, 2026-11-06: Outlook-style tiles in a wrapping row, each a file-type icon, the name and the size; clicking anywhere on a tile downloads the file, and a download icon at its right edge, always shown, says so. Outlook opens a preview on click and keeps Download in a menu; the CRM has no preview, so the click downloads.
Actual: one AttachmentRow per file across the pane, with an outline Download button (design-system §5 Attachment row, the received-file variant, Rev 87).
Fix: each file is one shadcn outline Button drawn as a tile, named Download downloadAttachment, savedFileName, the object URL’s lifetime) and the focus rules are unchanged, with focus returning to the tile. A download icon concept joins the icon registry. The design is the Email Screen Redesign artifact; design-system Rev 101 records it.
Regression test: MessageAttachments.test.tsx › “draws each file as one tile button holding its icon, name, size and download icon, with the size as its description”; it fails on main, where each row holds a separate Download button. In the browser after the fix: four 224 × 55px tiles in a wrapping row on Stub: attachments (not clicked, because a click saves a file).
Resolved: 2026-11-06 — fixed in forge-et-fe#435 (a067dba3, 039a1fcb), with BUG-360, BUG-361, BUG-362 and BUG-364.
Lesson: none — one-off: a design choice changed (Nanndish’s, to follow Outlook), not a defect that recurs.
BUG-362 — The Email message list stops after a few rows and scrolls inside that short box instead of filling the window
Affects: F20-i (forge-et-fe src/features/email/EmailRoute.tsx, EmailRoute.module.css) · Found in: Nanndish’s UI review of the Email screen, 2026-11-06 · Severity: medium — at desktop width only a few messages show at once, with unused height under the list · Phase: 1
Repro: open Email at bp/laptop or wider with a mailbox holding more messages than fit.
Expected: the list card runs from under the data bar to the bottom of the window, level with the folder pane and the reading pane, as design-system §6 Two-pane list + detail draws it; only its rows scroll.
Actual: the list card shows a few rows and scrolls inside that height, with the pager under it and unused space below.
Cause: measured in the browser on main (4a91ed9c) at 1440 × 900, signed in as a local test user with a stub mailbox: the list pane was 727px tall, but the selected tab’s panel only 357px. Radix keeps the tab not chosen (Other) in the page as an empty TabsContent with the hidden attribute, and EmailRoute.module.css .tabPanel set display: flex (with flex: 1 1 0 from bp/laptop), which beats the attribute’s display: none; the empty panel took the other half of the pane (357 + 12 + 357). The email templates screen and the record page had each fixed the same cause on their own (L-262).
Fix: .tabPanel[hidden] { display: none; }. The Focused and Other tabs move to the top of the list card and the pager to its foot (BUG-361), so the card holds the tab strip, the rows and the pager; every box from the list pane down passes the pane’s height on, and only the rows scroll. The pager’s footer shows only when there is a page to go to. The design is the Email Screen Redesign artifact; design-system Rev 101 records it.
Regression test: mailboxTokens.test.ts › “the tab panel sets display, so it keeps [hidden] hidden (BUG-362)” and “from bp/laptop every box from the list pane to the rows passes the height on (BUG-362)”; EmailRoute.test.tsx › the Focused/Other test asserts the two panels, one hidden, inside the list card; e2e/browser/email/inbox-message-view.spec.ts › “the two panes fill the page …” now measures the list card’s own bottom edge against the reading pane’s (the old check compared the pager’s, which sat under both panels, so it passed on the bug). In the browser after the fix: the list card, the folder rail and the reading pane all end at 876px, on the Focused and the Other tab.
Resolved: 2026-11-06 — fixed in forge-et-fe#435 (a067dba3, 039a1fcb), with BUG-360, BUG-361, BUG-363 and BUG-364.
Lesson: L-262 — new: the third time a class’s display beat the hidden attribute on a tab panel (after the email templates screen and the record page); frontend-reviewer check 30 now looks for it.
BUG-361 — The Email data bar’s Category filter is stretched to a fixed slot and the Focused/Other tabs sit beside it as pill tabs, unlike the other list screens’ filters
Affects: F20-i (forge-et-fe src/features/email/InboxToolbar.tsx, CategoryFilter.tsx, InboxTabs.tsx, EmailRoute.tsx) · Found in: Nanndish’s UI review of the Email screen, 2026-11-06 · Severity: low — the filters work; the row mixes three control styles · Phase: 1
Repro: open Email at bp/laptop or wider, on the Inbox.
Expected: the data bar’s filters drawn as the Leads and Accounts filters are (design-system §6 List screen): each select as wide as its text, 40px tall, space/4 apart. Nanndish’s decision, 2026-11-06: the Focused and Other tabs move to the top of the list card, because they switch the list rather than filter it.
Actual: the Category select fills a 288px slot (up to layout/form-width) around one word, and the Focused/Other pill tabs, 36px tall, follow it in the data bar (F20-i WI-3.2 plan D-12, D-15).
Fix: CategoryFilter drops fullWidth and its slot takes the Leads .filter rule; the tab row (InboxTabList) renders as an underlined tab strip at the top of the list card, inside the card, so the list card and the reading pane still start level. The tabs’ rules stay: Inbox only, gone during a search, manual activation. This reverses D-15’s placement; design-system Rev 101 records it. The design is the Email Screen Redesign artifact. The folder picker below bp/laptop is sized to its text the same way.
Regression test: mailboxTokens.test.ts › “the picker and the category filter are as wide as their text, at most layout/form-width (BUG-361)”; InboxToolbar.test.tsx › “the data bar holds no tabs (BUG-361)”; EmailRoute.test.tsx › the Focused/Other test asserts the tab strip is the list card’s first part and not in the data bar. In the browser after the fix: Category: All 134px wide, 40px tall; the tabs at the top of the list card, the selected one underlined 3px in primary.
Resolved: 2026-11-06 — fixed in forge-et-fe#435 (a067dba3, 039a1fcb), with BUG-360, BUG-362, BUG-363 and BUG-364.
Lesson: L-254 — another instance: the Email data bar’s filter took its own width where the Leads filters set the look. L-254’s sources now list it.
BUG-360 — The Email data bar’s search box is 552px wide, where every other list screen’s is 288px
Affects: F20-i (forge-et-fe src/features/email/InboxToolbar.module.css) · Found in: Nanndish’s UI review of the Email screen, 2026-11-06 · Severity: low — the box is longer than its content needs and than every other screen’s · Phase: 1
Repro: open Email at bp/laptop or wider.
Expected: Nanndish’s decision, 2026-11-06: 288px (nine space/8 steps), the Leads, Accounts and Users width, not the 420px layout/search-max.
Actual: 1.5 × layout/list-pane-width plus space/5 (552px), so it lines up with the folder and list columns (design-system Rev 84), and the collapsed folder rail’s variant of the same.
Fix: .filter takes the Leads .search rule, flex: 0 1 calc(var(--et-space-8) * 9), in both folder-pane states; the collapsed-rail variant and the data-folders and data-with-picker attributes it read go. The design is the Email Screen Redesign artifact; design-system Rev 101 records it.
Regression test: mailboxTokens.test.ts › “.filter is the other list screens’ 288px (nine space/8 steps), not the width of the columns below it (BUG-360)”, which also reads the Leads rule it matches; it fails on main. In the browser after the fix: the search box 288px at 1440 × 900.
Resolved: 2026-11-06 — fixed in forge-et-fe#435 (a067dba3, 039a1fcb), with BUG-361 to BUG-364.
Lesson: L-254 — another instance: the Email search box took its own width where every other list screen’s is 288px. L-254’s sources now list it.
BUG-357 — Flaky tests: two seam files’ clean-up requests find the browser gate’s app server not answering, at the same moment (Windows)
Affects: F1-c, F11-g (forge-et-fe e2e/seam/account-timeline.seam.spec.ts:1801 and e2e/seam/lead-ranking.seam.spec.ts:1704, both failing in their afterAll clean-up) · Found in: the F11-g wave 1 browser gate (full profile, 6735 passed, 43.5 min) on forge-et-fe feature/F11-g-wave-1 2bbcdcee, Windows 11, 16 cores, load 0.66 average and 0.77 peak, 2026-11-06 · Severity: low — both passed on the gate’s retry; costs a retry, hides nothing · Phase: 1
Repro: in that run both clean-ups failed with apiRequestContext.get: connect ETIMEDOUT ::1:15173 followed by connect ECONNREFUSED 127.0.0.1:15173, on the gate’s app server (port 15173), within the same few seconds. The tests’ own assertions had passed.
Expected: the gate’s app server answers every request for the whole run.
Actual: for a short window the app server neither accepted on ::1 nor on 127.0.0.1.
Fix: not decided. Triage per .forge/test-strategy.md § Flaky Tests: read the gate’s app-server output for that window (a restart, a crash or a blocked event loop), and whether localhost resolving to ::1 first adds a connect timeout before the IPv4 fallback. A longer timeout is not the fix.
Investigated (2026-11-08, Nanndish, Windows): the server was not down; the client gave up. The clean-ups send backend API calls to the app origin, which vite preview proxies. Given localhost, vite preview listens on ::1 only on Windows. Playwright 1.63’s request client connects with Node’s autoSelectFamily, which gives every address but the last a 5-second timer. When a worker’s event loop is held up for 5 s while a connect is pending, the timer fires before the finished connection is read: ETIMEDOUT ::1, then ECONNREFUSED 127.0.0.1, where nothing listens. That is exactly this bug’s signature. Scratch servers showed that a stalled, overloaded or stopped server gives ECONNREFUSED ::1 instead. Against the real built app with the client held 6 s, 5–8 of 20 requests failed with this signature in each of 3 runs; with the server on 127.0.0.1 and the client calling http://127.0.0.1:<port>, 20 of 20 succeeded each time. A candidate fix, forge-et-fe#455, serves the gate’s app on 127.0.0.1 and makes it the gate’s base URL. It overlaps BUG-196 (Dinil, in progress), which records the same signature and that serving on 127.0.0.1 is not the fix (the draft forge-et-fe#369). #369 changed only the server’s host, so its clients’ ::1 attempt was refused and the last attempt hit the system’s own connect timeout (ECONNREFUSED ::1, then ETIMEDOUT 127.0.0.1). #455 also moves the client, which removes this bug’s signature but not a connect that gets no answer at all. BUG-196 also records that one client blocked for 6 s after starting a connect still connected, where this run’s 20 concurrent connections did not. Whether #455 resolves this bug, and whether this bug is a recurrence of BUG-196, is in owed.md. What held two workers up for 5 s at once was not found; the run’s output no longer exists.
Regression test: the two files pass their clean-up in repeated loaded gate runs (the nightly full run on main once #455 has merged).
Resolved: 2026-11-09 — fixed in forge-et-fe#455 (4b5c5134, 9ecc76e0; merged 2026-11-08 as f22ebc1e): the browser gate serves its app on 127.0.0.1 and calls it there, so no client tries ::1 first and Node’s per-address 5-second timer cannot end a connect that has already succeeded. Nanndish decided on 2026-11-09 that this bug is closed on its own, separate from BUG-196: #455 removes this bug’s signature (ETIMEDOUT ::1 then ECONNREFUSED 127.0.0.1), which the investigation reproduced and #455’s host change ended in 20 of 20 requests three times. What #455 does not cover (a connect that gets no answer at all, and runs outside the gate) stays with BUG-196 (owed.md). The nightly full run on main is the regression check from here; a recurrence of this signature in a gate is filed as a new bug that names this one.
Lesson: L-158, another instance — on Windows localhost is ::1 first; a server bound there and a client that calls localhost add a connect attempt that a stalled event loop can time out, so the lesson gains a bullet.
BUG-356 — Flaky test: the size matrix’s contacts-edit-from-account state at 1024 × 768 in WebKit finds a colour-contrast failure on the dialog’s description (Windows)
Affects: F1-f (forge-et-fe e2e/browser/size-matrix/size-matrix.spec.ts:263, state contacts-edit-from-account, project size-1024x768-webkit) · Found in: F11-g wave 1’s WI-1.4 run of playwright test e2e/browser/size-matrix (4643 passed, 2 failed, 32.9 min) on forge-et-fe feature/F11-g-WI-1.4-e2e-ai-lead d6397815, Windows 11, while another agent ran a Vitest suite, 2026-11-06 · Severity: low — passes on a re-run; the wave does not change contacts or the dialog · Phase: 1
Repro: that run: axe color-contrast (serious) on <p data-slot="dialog-description"> of the Edit contact dialog. --repeat-each=3 on the same stack passed 3 of 3.
Expected: the page check passes whatever else the machine is running.
Actual: under load axe sometimes reads the description’s contrast while the dialog is still fading in.
Fix: at its 5-second cap the shared page check finishes each finite animation that ends within the cap, and fails on one that ends after it, instead of scanning the page part-way.
Regression test: the four BUG-356 self-tests in e2e/browser/support/page-check.spec.ts: a paused 150 ms fade, slowed to playback rate 0.001 and then played, is finished before the scan; a 60 s animation fails the check by name; a transition started by a finished animation’s animationend is finished too; an animation its end handler restarts after every finish fails the check after three rounds. Each fails on the code before the fix and passes on it, in Chromium and in WebKit. The original failure was not reproduced, so repeated loaded runs of the state prove nothing either way.
Resolved: 2026-11-08 — fixed in forge-et-fe#457 (3bf83984, with the two review rounds’ changes in 46666f37 and 7634174c) and its byte copy forge-et-pwa#114; not reproduced. settleAnimations (e2e/support/pageCheck.ts) waited for every finite animation, but at its 5-second cap axe scanned the page as it stood. Likely cause, not proved: the browser advances an animation only when it draws a frame, while the cap counts wall time. So if the page drew no frame for about 4.85 s, the Edit contact dialog’s 150 ms fade was still part-way at the cap, and axe read the description as low contrast. BUG-369’s call logs show WebKit pages on a loaded Windows machine going about 25 s without frames. Nothing in the state restarts the fade (its key and content element do not change), so the cap was the only path. At the cap, each finite animation still running is now judged by its own end time (getComputedTiming().endTime). One that ends within the cap is finished and annotated animation-finished-at-cap (visible in the HTML report). One that ends after it fails the check by name, so a product animation that never settles is not passed silently. After a finish, the page is listed again after two rendered frames, up to three rounds, by the same rule, and anything still being started after the third fails the check. No finite animation in src/ runs over the cap (150–250 ms). Reproduction attempts did not show the original failure: under 16 busy loops, 12 of 12 runs of the state passed and none reached the cap, and a 500-state WebKit probe run had no color-contrast failure. On the final code the page-check spec passed 73 in Chromium (--repeat-each=3), its BUG-356 and BUG-138 self-tests 25 in a one-off WebKit run, the WebKit dialog states 58 with none failing on the new errors, and forge-et-pwa’s copy of the spec 21.
Lesson: L-101 (another instance: the shared check’s own wait could end on a page still changing, this time at its time cap; BUG-356 added to its sources).
BUG-355 — Flaky test: F5-c AC-8’s reactivation browser test finds the user page heading not focused (Windows)
Affects: F5-c (forge-et-fe e2e/browser/users/activation.spec.ts:286, “F5-c AC-8: reactivating the user shows them active, and they can sign in again”) · Found in: F11-g wave 1’s WI-1.4 run of playwright test e2e/browser/leads e2e/browser/sharing e2e/browser/users --project=chromium (362 passed, 1 failed) on d6397815, Windows 11, while another agent ran a Vitest suite, 2026-11-06 · Severity: low — passes on a re-run; the wave does not change the users screens · Phase: 1
Repro: that run: expect(locator).toBeFocused() on the user’s <h1 tabindex="-1"> timed out at 5 s (“inactive”, resolved 14 times). --repeat-each=3 of the file on the same stack: 25 passed.
Expected: the test passes whatever else the machine is running.
Actual: under load the focus move to the heading after reactivation is not seen within 5 s.
Fix: the activation’s answer sets a focus request in the same batch as the dialog’s close, and an effect after the render that removed the dialog moves focus, in place of a requestAnimationFrame.
Regression test: two UserEditorRoute.test.tsx cases, reactivation and deactivation, run frame callbacks at once and assert the h1 has focus; both fail on the old route code (“Received element with focus”) and pass on the fix. The browser test passed 6 of 6 under a 6x CPU throttle, where the old code failed 5 of 6.
Resolved: 2026-11-08 — fixed in forge-et-fe#459 (17f47ca1, with the review’s changes in 21a0a442). A product defect, not a slow test. After an activation succeeded, UserEditorRoute.tsx moved focus to the page’s h1 in a requestAnimationFrame, assuming the closed dialog would have rendered by then. Under load the frame ran first: focus was still on the dialog’s confirm, so mayMoveFocus() refused, and the dialog’s close returned focus to the trigger (by then the Deactivate button). A console probe under a 6x CDP CPU throttle showed exactly that in 5 of 6 runs, and once in 5 with no throttle on the loaded machine. The answer now sets page state in the same batch as the card’s setDialog(null). An effect, which must stay a passive effect, moves focus after the render that removed the dialog: it runs after Radix’s focus trap and aria-hidden are undone and before Radix’s setTimeout return of focus, which the dialogs’ onCloseAutoFocus then leaves alone because focus is not lost. UserActivationCard’s onChanged doc says to move focus from an effect. The browser test, activation.spec.ts --grep "AC-8" under the 6x throttle, failed 5 of 6 on the old code and passed 6 of 6 on the fix. The deactivation path’s focus step passed 4 of 4. The same frame-before-render defect in CreateUserModal.tsx’s refused-save focus is BUG-383.
Lesson: L-271 (new, with BUG-383).
BUG-354 — At 320px the user page under the Edit user dialog’s leave question overflows <main> by 2px again (BUG-313 recurred) (Windows)
Affects: F5-c (forge-et-fe the user page at 320; size-matrix state user-editor-leave-dialog) · Found in: F11-g wave 1’s WI-1.4 size-matrix run on d6397815, then on forge-et-fe main 032c7517 against the same backend, Windows 11, 2026-11-06 · Severity: medium — 320px reflow (WCAG 2.1 SC 1.4.10) · Phase: 1
Repro: pnpm exec playwright test e2e/browser/size-matrix --grep user-editor-leave-dialog --project=size-320x640-chromium --repeat-each=3: 3 of 3 fail on the F11-g wave branch and 3 of 3 on main 032c7517.
Expected: no horizontal scroll on the page or in <main>.
Actual: “<main> overflows by 2px”, the same detail as BUG-313, which was resolved on 2026-11-03 with no traced fix and its mark removed in forge-et-fe#385.
macOS: the page is clean on the lead’s Mac. The 2026-11-07 nightly run on main 6e1b32c2 failed the state because the mark expected an overflow that did not happen, and the same command as the repro passed 3 of 3 there on 2026-11-07. The overflow is on Windows only, which also explains BUG-313 resolving with no traced fix on a Mac. forge-et-fe#443 attaches the mark only when the browser runs on Windows (e2e/pageStates.ts USER_LEAVE_MARKS).
Fix: on Windows, find the 2px-wide element on the page under the dialog (the page check’s overflow detail) and let it wrap. A font that measures wider on Windows is the likely cause.
Regression test: the size-matrix state, unmarked. Until the fix, the state carries a BUG-354 known-failure mark on Windows (F11-g wave 1; Windows only since forge-et-fe#443).
Resolved: 2026-11-08 — fixed in forge-et-fe#450. The cause was the scrollbar, not a font. At 320 × 640 <main> is 244px wide, and on Windows its classic vertical scrollbar takes 15px, leaving 229px inside (macOS overlay scrollbars take none). The one element outside its own scroll box and past that inner edge was a deactivated person’s note in the Reporting line card, three levels down the tree: “Shares to them count for nobody until reactivated”. It is a flex item in ReportingLineChain.module.css’s .inactive, so it could not shrink below the word “reactivated” (61px), and it ended at 307px against an inner edge at 305px. The note’s own span (.inactiveNote) now has min-inline-size: 0 and overflow-wrap: anywhere, as the card’s person links (.link) and the effective-permissions section’s .meta already do, and the Windows-only mark is removed from e2e/pageStates.ts. Fails before: with the old CSS and no mark, the state failed with “<main> overflows by 2px”. Passes after: 3 of 3 (--repeat-each=3). Every size-matrix state matching “user” passed at the four Chromium sizes (136), and e2e/browser/effective-permissions passed 38 (Windows 11, Chromium). At 320 the third tree level is narrow enough that names already break every few letters, and the note now wraps the same way; a layout with more room for deep levels is a design question, recorded in owed.md. The review of forge-et-fe#450 found that one level deeper the overflow returns on every system, because the Inactive label cannot wrap: filed as BUG-381.
Lesson: L-224 — a layout proved where scrollbars overlay the content can be wrong where they take room; this PR adds BUG-354 (and BUG-313 before it) to the lesson’s sources.
BUG-353 — MessageView.test.tsx’s StrictMode rows put StrictMode inside the query provider
Affects: F20-i (forge-et-fe/src/features/email/MessageView.test.tsx:193, from dcb5f804) · Found in: the forge-et-fe#424 round-2 delta review (2026-11-05) · Severity: low · Phase: 1
Repro: read the test’s render helper: it wraps the app in <StrictMode> and passes that to renderWithQueryClient, which puts QueryClientProvider outside it.
Expected: StrictMode outside the provider, as main.tsx mounts it (L-182), so the provider’s mount-unmount-mount runs in the test.
Actual: the provider mounts once; a double request caused by the provider’s StrictMode remount would not be caught by these rows.
Fix: use renderWithQueryClient(app, undefined, { strict: true }), the option F20-i wave 4 added to src/test/queryClient.tsx (forge-et-fe#424).
Regression test: the existing StrictMode rows, rendered through the strict option.
Resolved: 2026-11-06 — fixed in forge-et-fe#429: the helper calls renderWithQueryClient(app, undefined, { strict }), so StrictMode wraps the query provider. The bug was a test gap, so there is no failing run before the fix; MessageView.test.tsx passes, 71 tests.
Lesson: L-182 — another instance; wave 4’s own StrictMode rows were fixed the same way, and L-182’s sources now list BUG-353.
BUG-352 — Seventeen size-matrix page states failed at 1280 × 800 on main and passed on the next run (Chromium)
Affects: F40 (forge-et-fe/e2e/browser/size-matrix/size-matrix.spec.ts:263, “every page state passes the shared page check at this size”; the offboarding, role and accounts-tag-retired page states) · Found in: F20-i-WI-4.3 merge-base baseline (2026-11-05) · Severity: low · Phase: 1
Repro: on the lead’s Mac, the whole size-1280x800-chromium project of forge-et-fe at a740830c (main that day) against forge-et-be 0b52c57f on a fresh isolated database, with another session’s verify stack up: 17 failed, 615 passed (45.0 minutes).
Expected: the page states pass at 1280 × 800, as they do in the nightly full run.
Actual: 17 rows failed on the baseline; the same project passed in full (640) on the F20-i wave-4 stack 90 minutes later in 6.5 minutes, with no wave-4 change to those pages. The long baseline time points at load or a wait, not at the pages. The run is recorded in .forge/plans/F20-i/F20-i-WI-4.3-test-verify.md (baselines).
Fix: not yet investigated: check whether the nightly run on main has seen these rows fail, and run the project again under load (L-107).
Triage (2026-11-08, test-strategy.md § Flaky Tests steps 1–2): the nightly full run on main ran every size-1280x800-chromium state on 2026-11-05, -06, -07 and -08 with no failed attempt (~/.et-nightly/logs/<date>/fe-gate.log; the only failures at that size were the hand runs of 2026-11-02, with hundreds of rows failing on an environment fault). Under forced load — one busy loop per core (8) started once the run held the heavy-run lock, load average 16 to 18 — the offboarding, role and accounts-tag-retired states at 1280 × 800, three times each (--repeat-each=3 --retries=0, forge-et-fe main a068aee0, an isolated backend on main fb140884), passed 144 of 144. The CDP CPU throttle was not applied (the Playwright CLI has no switch for it). Not reproduced; the failed baseline ran beside another session’s verify stack. Under step 4 the outcome is wontfix for an overloaded machine, which is Tharinda’s decision: brought to him, not closed.
Regression test: the project green across repeated runs on a loaded machine.
Resolved: 2026-11-08 — closed as wontfix: the machine was overloaded, not the test (.forge/test-strategy.md § Flaky Tests, step 4); Tharinda decided. The failing run had another session’s verify stack up and took 45 minutes against 6.5 minutes for the same project 90 minutes later. The triage above found no failed attempt of these states in the four nightly full runs since, and 144 runs of them under forced load all passed. A recurrence on a machine within the load limit is a new bug that links this one.
Lesson: none — one-off: a failure from other work loading the machine, which the browser gate’s load check now refuses to start beside; no test or product cause was found.
BUG-351 — Flaky tests: two mailbox connect browser tests miss their wait when the machine is loaded (macOS)
Affects: F20-a, F20-b (forge-et-fe/e2e/browser/mailbox/connect-a11y.spec.ts:160 “AC-28: reconnect required is reachable and activatable from the keyboard”; e2e/browser/mailbox/sync-indicator.spec.ts:504 “D10: a reconnect that lands straight back in FAILED re-renders”) · Found in: F20-i-WI-4.3 orchestrator browser run (2026-11-05) · Severity: low · Phase: 1
Repro: on the lead’s Mac, run the whole chromium project of forge-et-fe while another session’s verify stack and browser runs are up (on 2026-11-05, feature/F20-i-WI-4.3-verify-threads-and at 6865571b against forge-et-be feature/F20-i-wave-4 13bea834; the project took 17.2 minutes, against 10.5 minutes for the same project at the merge-base the same day).
Expected: the result does not depend on how busy the machine is.
Actual: connect-a11y.spec.ts:160 did not see mailbox-connected within 5 s after the reconnect, and sync-indicator.spec.ts:504’s page.waitForResponse(isStubConnect) timed out at 30 s. Both passed on the merge-base baseline and both passed, unchanged, when re-run at once on the same stack (c61aee73, 15 passed in 9.0 s). Wave 4 changes no file in the mailbox connect flow.
Fix: not yet investigated. Force the failure under load (--repeat-each while another suite runs, L-107) before changing a wait.
Triage (2026-11-08, test-strategy.md § Flaky Tests steps 1–3), forge-et-fe main a068aee0, an isolated backend on main fb140884, one busy loop per core (8) started once the run held the heavy-run lock, no CDP throttle (the Playwright CLI has no switch for it):
- The two tests alone, ten times each with one worker: 20 of 20 passed, about 1 s each, at a load average of 33.
- The two tests ten times each with the default workers: three failures, but each was the repeats colliding on the one signed-in user’s single mailbox connection (a
404forcing a state, a500on a second connect) — an artefact of running one test in parallel with itself, not this bug. - The whole
chromiumproject under the same load: 1,386 passed, 2 failed, both inconnect-a11y.spec.ts:160(the file’s four state tests run in one worker): connected — Enter on Disconnect mailbox opened no dialog within 5 s, and the page snapshot at the failure shows the connected panel intact, no dialog, and focus no longer on the button the test had just asserted was focused; connect failed — the activation did not reachmailbox-connectedwithin 5 s, BUG-351’s own symptom. The nightly full run has passed both tests on each night since (2026-11-05 to -08). - It reproduces, so it is a timing defect to fix (step 3), not an overloaded machine. The lead: something re-renders the mailbox screen after the test has focused the control — a background read landing late — and the press or the panel’s
confirmingstate is lost with it; a re-render that drops focus is also an accessibility defect. Next: hold back each read the screen makes after it settles (the connection read, the sync status, the automatic-connect check; L-162) until the failure is deterministic, then fix the cause. Not fixed in the overnight batch; brought to Tharinda. Regression test: the two tests green across repeatedchromiumruns on a loaded machine. Resolved: 2026-11-08 — fixed in forge-et-fe#462 (tests only). Holding back the reads, as the Triage planned, found the cause. The mailbox screen mounts with the connection already cached (the app-wide connect check read it), shows it, and re-reads it, because TanStack re-reads a stale query on mount. While that re-read runs, every connection action is busy and refuses a press, which is the designed behaviour (BUG-026, design-system §5 Busy row). The tests pressed once the controls were visible; on a loaded machine the re-read was still running then, so the press was refused. With every connection read after the first held 3 s, the failure happened every time: all fourconnect-a11yAC-28 keyboard rows (no confirm dialog; nomailbox-connected) and threesync-indicatorreconnect rows (the stub connect never sent).settleMailboxScreenandconnect-a11y’sarrangenow wait until no button on the page is busy (expectNoBusyAction). The committed regression row,connect-a11y.spec.ts“BUG-351: a key press waits for the screen’s own re-read of the connection, however late it lands”, holds the re-read 2 s; it fails without the wait and passes with it. The mailbox and email specs pass in Chromium (308), as do the size matrix’semail-*andmailbox-*states at 1280 × 800 and 320 × 640 (202). The Triage’s “focus no longer on the button” was not a second defect: a probe logged focus on Disconnect before the refused press, after it and after the re-read landed, and the failure’s page snapshot marks no focused element even when focus is on the button. The review found a separate, older focus defect in the same screens: BUG-384. Lesson: L-101 — another instance: “settled” included the screen’s own background re-read, which keeps its actions busy after they are visible. L-101’s sources now list BUG-351.
BUG-350 — Flaky test: the browser gate’s selection test that runs in full without a map times out in a loaded run (Windows)
Affects: F40 (forge-et-pwa src/test/scripts/browserGateSelect.test.ts, “selectGate against a repository (test-run-cost.md S-2) › runs in full without a map (forge-et-pwa), when only Vitest tests changed, and refuses an unknown base”, line 310; the file is a copy of forge-et-fe’s, scripts/shared-with-fe.json) · Found in: forge-et-pwa#112’s local test record (pnpm test:record on 758cfd72), Windows 11, 16 cores, load 0.12 at the start and 0.94 at its peak from other sessions’ work, 2026-11-05 · Severity: low — one test fails now and then under load; the gate’s selection is unaffected · Phase: 1
Repro: that run: Test timed out in 5000ms, after its one retry (13.2 s in all). The other 748 tests in the run passed. A second record on the same commit (load 0.33 at the start, 1.00 at its peak, 0.76 on average) timed it out again, together with the same describe’s “selects a feature’s change and records the merge base” (line 299). Run alone, the file passed 12 of 12.
Expected: the test passes whatever else the machine is running.
Actual: the test builds a git repository fixture (repoWithMap(false)) and runs the selection over it, within Vitest’s default 5 s timeout.
Fix: give the suite “selectGate against a repository” the git-fixture timeout, { timeout: 60_000 } on its describe, as browserGateCli.test.ts and browserGateRules.test.ts set theirs. The file is shared, so the fix lands in forge-et-fe and is copied here.
Regression test: the test passes in repeated loaded runs on Windows.
Resolved: 2026-11-05 — fixed in forge-et-fe#425 (bb9510dd) and copied by forge-et-pwa#112 (513b20c, the manifest recorded with --from=bb9510dd…). Nanndish decided on 2026-11-05 that this is a fix, not the lengthened timeout .forge/test-strategy.md § Flaky Tests bans, and not step 4’s wontfix. The suite’s two tests are synchronous (createGitFixture and the selection run git through execFileSync and spawnSync), wait on nothing and measure no time, so there is no race for a longer limit to slow down, and a hung git blocks the worker whatever the limit. The cost is fixed work, and the 5 s default was the odd one out among the repo’s git-fixture suites (60–120 s). Step 4 does not fit either: the tests failed in four records of four, so closing it wontfix would leave forge-et-pwa#112 unable to record. Triage step 2’s forced load (14 busy loops) did not reproduce it without the fix (12 of 12 passed): it needs the record’s own parallel git spawning. The regression check is the record itself: forge-et-pwa#112’s pnpm test:record on 513b20c, with the fix, passed 749 tests (4 skipped), at a load of 0.39 at the start, 0.99 at its peak and 0.70 on average, where the same record had failed these two tests four times out of four at a similar load.
Lesson: L-257 — a suite that builds git fixtures takes the git-fixture timeout on its describe; BUG-349 is a second instance of the same cause.
BUG-349 — Flaky test: the shared-file list test that refuses an absolute or repeated path times out building its template workspace in a loaded run (Windows)
Affects: F40 (forge-et-pwa src/test/scripts/sharedDrift.test.ts, “the shared-file list (F40 WI-1.1 plan Decision 14) › refuses a list with an absolute, parent-relative or repeated path”, line 120) · Found in: two Vitest runs of src/ui/styles and src/test/scripts/sharedDrift.test.ts on branch chore/shared-tokens-dialog-sizes (forge-et-pwa#112, d5017efe), Windows 11, with other sessions’ backend, browser runs and builds holding the CPU at 94–100%, 2026-11-05 · Severity: low — one test fails now and then under load; the drift check it tests is unaffected · Phase: 1
Repro: both runs: Test timed out in 5000ms (the second took 13.4 s). Alone, with -t "refuses a list", it passed on the same head.
Expected: the test passes whatever else the machine is running.
Actual: registerWorkspaces() builds the template workspace, two git repositories of 76 files, the first time a test calls workspace(). The first caller is this test, in the first describe, which keeps Vitest’s default 5 s timeout, while every later describe sets { timeout: 60_000 }; so the build is charged to a 5 s test. The local test record selects this file for every PR that changes a shared file (node scripts/test-selected.mjs --base origin/main --list).
Fix: on branch fix/BUG-349-shared-drift-template-timeout, three changes in forge-et-pwa src/test/scripts/:
sharedDriftFixture.ts:registerWorkspaces()builds the template in abeforeAllwith the git-fixture timeout (60 s), so the build belongs to no test. It covers all three files that use the fixture. ItsafterEachandafterAlltake the same timeout: a Vitest hook gets its own 10 shookTimeoutwhatever itsdescribesets, and theafterEachdeleting a case’s copy timed out once at 10 s insharedDriftFeSide.test.ts.sharedDrift.test.ts, “refuses a list”: writes its list into an empty temporary folder instead of copying a workspace it never used as a repository.sharedDrift.test.ts, thereadBlobssuite: each case copies the workspace and commits in it, so the suite takes{ timeout: 60_000 }as the file’s other git-fixture suites do.
None of these lengthens a timeout .forge/test-strategy.md § Flaky Tests bans. The work is synchronous git and file copies with no race, the cause BUG-350 settled (§ Flaky Tests, cause 6).
Regression test: loaded Vitest runs of src/ui/styles and src/test/scripts/sharedDrift (206 tests) on Windows 11, 16 cores, under other sessions’ load, 2026-11-06. Each run’s CPU busy share (average) was sampled every second. The per-run log is the scratch script’s output, not kept.
- Before,
origin/main’s fixture, 5 runs at loads 0.99–1.00: “refuses a list” timed out in all 5 (15.2–31.2 s), andreadBlobstimed out in all 5. - The
beforeAllalone, 8 runs: at loads 0.95–1.00, “refuses a list” still timed out (9.9–10.5 s); at 0.64–0.83 it passed in 1.9–3.0 s.readBlobstimed out in 4 of 8. - With the temporary folder and the
readBlobstimeout, 6 runs at loads 0.58–0.84: “refuses a list” passed in 17–97 ms in all 6, andreadBlobspassed in all 6. One run (0.84) failed the fixture’safterEachat its 10 s hook timeout, which led to the hook timeouts. - With the hook timeouts, 1 run at load 0.88: 206 of 206 passed. A second run was stopped from outside after 8 minutes, with no result.
forge-et-pwa#113’s pnpm test:record on its head a1147fe (the commits after these runs fix only the review’s Nits) selected 21 test files, the three sharedDrift* files included. It passed 749 of 753 tests with 4 skipped, 0 failed and no flaky test, in 7 min. The load was 0.23 at the start, 0.97 at the peak and 0.59 on average. The record before it, on the same commit, passed every test that ran (682, 4 skipped), but three files unrelated to this fix never started: Vitest’s forks pool timed out waiting for their workers at a peak load of 0.99. One occurrence under load stays in the weekly flaky report (.forge/test-strategy.md § Flaky Tests).
Resolved: 2026-11-06 — fixed in forge-et-pwa#113. This is not the lengthened timeout .forge/test-strategy.md § Flaky Tests bans, on the reasoning Nanndish accepted for BUG-350: the work under the 60 s limit is synchronous git and file copies with nothing racing, and the one case that did not need it now does no git work at all.
Lesson: L-257 — the same cause as BUG-350. This fix adds an update to L-257: build a lazily built shared fixture in a beforeAll; give the suites that copy it, and its hooks, the git-fixture timeout too; give a case that needs only a folder an empty one.
BUG-348 — With a row’s menu opened, a table’s pinned first-column cell keeps its card fill and hides the row’s open tint
Affects: F11-b and every table that pins its first column (forge-et-fe src/ui/dataTable.ts DATA_TABLE_CELL_FILL, DATA_TABLE_MENU_CELL; src/ui/pinnedFirstColumn.ts) · Found in: the review of forge-et-fe#415 (BUG-310), 2026-11-05; the same on main · Severity: low — the open row’s tint stops at its first cell; nothing is hidden or unreachable · Phase: 1
Repro: at bp/tablet or wider, on a table that pins its first column (the leads list), open a row’s ⋯ menu by keyboard and compare the first cell’s painted colour with the row’s other cells.
Expected: the whole row, pinned cells included, shows the open-menu tint.
Actual: the pinned first-column cell paints its opaque card fill over the row’s tint. Since forge-et-fe#415 the pinned ⋯ cell paints the tint as a layer over its fill; the first cell does not.
Fix: move the tint layer from DATA_TABLE_MENU_CELL into DATA_TABLE_CELL_FILL, so it covers both pinned cells (the review’s suggestion).
Regression test: extend forge-et-fe#415’s tint test to the row’s first cell; it fails before the fix.
Resolved: 2026-11-06 — fixed in forge-et-fe#429: the tint layer moved from DATA_TABLE_MENU_CELL into DATA_TABLE_CELL_FILL. leads-layout.spec.ts’s BUG-310 tint test now checks the first cell too. With main’s dataTable.ts it fails on the first cell (matches: false, row tinted: true); with the fix it passes in Chromium. pinnedFirstColumn.test.tsx checks that the fill carries the layer.
Lesson: none — one-off: BUG-310’s fix put the tint layer on the one pinned cell it changed; the layer now sits in DATA_TABLE_CELL_FILL, which every pinned body cell uses, so a pinned cell added later gets it too. Not L-219, whose cause is a control scrolled under a pinned column.
BUG-347 — Flaky test: AccountListScreen.test.tsx’s “a filter writes the URL, returns to page 1 and narrows the request” times out at 20 s on a loaded machine (Windows)
Affects: F1-b (forge-et-fe src/features/accounts/AccountListScreen.test.tsx, “a filter writes the URL, returns to page 1 and narrows the request”) · Found in: the review of forge-et-fe#411 (BUG-314), a run of the PR applied on main of the tests that use useListSearchParams or read route state, Windows 11, on a machine shared with other sessions, 2026-11-05 · Severity: low — one test fails now and then under load; the list is right · Phase: 1
Repro: that run: the case hit the 20-second timeout. On plain main on the same machine it took 17.9 s, near the limit whatever the change. A second AC-4 case in the same describe, “the Owner filter lists directory names and sends ownerId”, failed with Timed out in waitFor in forge-et-fe#408’s test record on aa56fbdf (machine load 0.94 average, 1.00 peak, 2026-11-05), and passed alone on that head (5 of 5 matching cases); fix the two together.
Expected: the test passes whatever else the machine is running.
Actual: under load the case runs past the 20-second budget.
Fix: both cases did more work than their budget allowed (.forge/test-strategy.md § Flaky Tests cause 6). The filter case chose three data-bar selects, Customer type, Region and Country, in one case. Country’s Radix list renders all 250 countries, so choosing one took about 5 s on a loaded machine: about 1.8 s to open, 0.7–1.1 s for findByRole('option', { name }) to compute every option’s accessible name, and 2.3–2.9 s for the click. It is now one case per select under a describe that keeps the case’s name. Each later case opens on the URL the earlier choices wrote (?customerType=Importer, then ?customerType=Importer®ion=Middle+East), so the final combined-URL assertion is kept. The Owner case opened and closed the select inside a waitFor until it offered three options. That waitFor had the one-second default, and one opening takes longer than that under load. It is now two cases, one counting the options and one choosing a person. Each waits once for the directory, by the owner name the rows show from the same query, then opens the select once. The file’s choose helper now finds the option by its text inside the open listbox and clicks its role="option" element, which saves the name computation (30–60 ms against 0.7–1.1 s for Country). No timeout was raised. Test only.
Regression test: the split cases, run on Windows 11 with the machine at 100% CPU (Get-Counter) from the other sessions’ work. Before, the AC-4 describe alone (-t), three runs: the filter case took 16.2, 12.6 and 14.1 s, and the Owner case 4.8, 4.6 and 5.4 s. After, the whole file three times, 103 of 103 each time: Customer type 3.8, 5.4 and 3.2 s; Region 4.7, 4.9 and 3.2 s; Country 6.5, 6.9 and 5.3 s; the Owner cases 4.8, 2.3 and 1.9 s (options) and 4.7, 4.5 and 2.9 s (choice). None of them is among the file’s three slowest cases in any run.
Resolved: 2026-11-06 — fixed in forge-et-fe#434 (a86df4a5; review fixes f7c305f4), test only. /forge-review-pr with frontend-reviewer found no Blocker or Important. It checked choose()‘s text lookup against every caller in the file and the installed Radix Select source. Three Nits were fixed: ownersNamed() now waits inside Nile Leaf’s row, the AC-4 comment says that the build-up of writes in one session is covered by useListSearchParams.test.tsx, and a typo. One Nit was left as is: the one-second waits after a choice, because the request is recorded during the click. After the review fixes, the AC-4 describe passed twice more at 100% CPU, Country at 6.2 and 5.8 s. The PR’s test record on f7c305f4: 2,798 tests in 66 files, 2,797 passed, 0 failed, 1 skipped, in 5 min 01 s (machine load 0.27 at start, 1.00 peak, 0.58 average). One test, knownFailureMarks.test.ts’s “the e2e tree has no stray test-fail call and no mark without a bug id”, passed only on its retry. It is not this change’s, and the weekly flaky report takes it.
Lesson: L-125 — another case whose arrangement did several cases’ work in one test budget (.forge/test-strategy.md § Flaky Tests cause 6), after BUG-331 in the same file. Added to L-125’s sources.
BUG-346 — Inside a dialog, a Tab pressed before the shared Select’s focus return has run lands on the dialog’s first control, not after the Select
Affects: F5-c and every dialog with a Select: the shared forge-et-fe src/ui/shadcn/select.tsx inside a Radix dialog (every RecordEditDialog Select: Edit user’s Role, New follow-up’s Type, …) · Found in: the review of forge-et-fe#410 (BUG-293), 2026-11-05; older than that PR · Severity: low — only a key press faster than a person’s (an automated one) reaches it · Phase: 1
Repro: in a component test, open a Select inside the shared Dialog, close its list by Escape, and press Tab at once, before Radix’s return of focus to the trigger has run.
Expected: the Tab starts from the Select’s trigger and moves to the next control.
Actual: takeTabFromUntilFocused stops on the focus trap’s focusin, so the Tab starts from the dialog’s content and lands on the dialog’s first control (✕).
Fix: none — closed as won’t fix (below).
Regression test: none, as no fix is made.
Resolved: 2026-11-08 — closed as wontfix; Tharinda decided. Only a key press faster than a person’s lands in the few milliseconds before Radix returns focus to the Select’s trigger, so no user is affected. Changing the shared Select, which every dialog uses, would carry more risk than the case is worth. A test that presses Tab after closing a Select waits for the trigger to have focus first, as takeTabFromUntilFocused already does.
Lesson: none — one-off: a timing window only an automated key press reaches, closed without a fix; no other Lesson: line names the same cause.
BUG-345 — At 768px and wider, a control partly scrolled under a table’s pinned first column or pinned ⋯ column fails axe target-size
Affects: F11-b (forge-et-fe the leads table, measured), and any table that pins its first column (usePinnedFirstColumn) and is wider than its region at bp/tablet · Found in: BUG-310’s fix, a scroll sweep of the leads table at 768 × 1024 on forge-et-fe fix/BUG-310-leads-sort-target (eba3984b) and on main 4421bd61, Chromium and WebKit, 2026-11-04 · Severity: low — axe target-size (serious) at some horizontal scroll positions; no size-matrix state stops at one · Phase: 1
Repro: at 768 × 1024 on /leads with the size matrix’s pinned inbox (pinLeadsInbox), set leads-table-scroll’s scrollLeft in 20px steps from 0 to its end and run axe’s target-size rule at each. The probe used for BUG-310 is a Playwright spec that answers every API read in the browser; it is not committed.
Expected: axe is clean at every scroll position.
Actual: on main: leads-sort-country at scrollLeft 180 (12px clear) and leads-sort-buyer-type at 300 (11.5px), each partly under the pinned Name header’s sort button, and leads-sort-est-value at the end (22.7px; BUG-310’s position, which a user who scrolls there by hand still reaches). With BUG-310’s pinned ⋯ column, the same three, plus a row’s +N tag button (+1: show all 3 tags, +3: show all 5 tags) at 380, 400 and 560 (7.4 to 17.6px clear), partly under the row’s pinned ⋯ button. axe counts only the overlap of one control with another: a header or tag passing under a pinned cell’s background is not reported, only under its button.
Fix: decided by Tharinda on 2026-11-08: snap the scroll to column edges (the first direction below), and measure every pinned table with the sweep before and after. The directions that were weighed: snap the region’s scroll to column edges (scroll-snap-type: x with a scroll-snap-align on each column and scroll-padding-inline the width of the pinned columns), so the table never rests with a column part-way under a pinned one; or accept the in-between positions as a reasoned page-check exemption. Measure every pinned table with the sweep before and after.
Measured (2026-11-08, Chromium, 768 × 1024, forge-et-fe fix/BUG-345-pinned-table-snap 72c0df02, the sweep e2e/browser/leads/pinned-scroll-sweep.spec.ts): on main the leads table fails at five resting positions — leads-sort-priority at 180, the +1 and +3 tag buttons at 600 to 740, and the scroll end at 780 — and the accounts table passes. With the snap rules injected for the measurement (scroll-snap-type: x mandatory, scroll-padding-inline the pinned Name and ⋯ columns’ widths, measured): scroll-snap-align: start on the header cells leaves 576 (a +1 tag button) and the end; start end leaves only the end, 780, where the first column after Name is still cut. Snapping cannot move the scroll end, so the end needs a further choice before the fix lands: a trailing spacer that makes the end a column edge, or a reasoned page-check exemption for the end position alone. Brought to Tharinda; the fix is not built.
Decided (2026-11-08, Tharinda, on the measurement above): the scroll end, which snapping cannot move, gets a trailing spacer after the last column, so the scroll can also stop on a column edge at the end; no page-check exemption.
Regression test: the sweep, as a browser test over the leads table and one other pinned table.
Resolved: 2026-11-08 — won’t fix, Tharinda’s decision, made on the review of forge-et-fe#453 (snapping with the end spacer, closed unmerged). The review showed that snapping keeps only one edge clean at a time: depending on column widths, which change with the data and the window, a column can still rest part-way under the pinned Name column; and the end spacer moved the pinned ⋯ column at every scroll position. The in-between scroll positions are accepted as a reasoned exception: a user who scrolls a little further sees the whole column, and no table behaviour changes. No size-matrix state rests at such a position today; a page state that does takes an axe exemption (e2e/support/axeExemptions.ts) that names BUG-345. The sweep (e2e/browser/leads/pinned-scroll-sweep.spec.ts) and the review’s probe are in forge-et-fe#453’s last commit, d1a9aff4; the branch fix/BUG-345-pinned-table-snap was deleted on 2026-11-13, and GitHub keeps that commit with the closed PR.
Lesson: none — one-off: a layout trade-off accepted, not a defect class; no other Lesson: line names a pinned column’s scroll positions.
BUG-344 — A Retry on the Users and email templates lists’ first-load error panels drops focus to the page
Affects: F5-c (forge-et-fe src/features/users/UsersScreen.tsx, users-retry; src/features/users/usersQueries.ts useUsersList); F20-d (src/features/email/templates/EmailTemplatesScreen.tsx, email-templates-retry), added by the review of forge-et-fe#406 · Found in: reading useUsersList while fixing BUG-317, 2026-11-03; not yet run · Severity: low — a keyboard user loses their place when retrying a failed first load · Phase: 1
Repro: open Settings › Users while the users read fails, so the error panel shows with no rows; Tab to Retry and press it.
Expected: Retry stays, busy, until the read answers; on a second failure focus is still on Retry (lesson L-183), and a good answer moves focus to the h1, as on the Contacts list’s first-load error.
Actual (from the code): useUsersList maps a key with no data and no snapshot to loading while it re-reads, and TanStack v5 returns such a key to pending at the press, so the panel and Retry unmount and focus falls to <body>. BUG-317 added useRetryHold to useUsersList for the stale notice only; the error state does not use the hold. On the email templates list the failure is held through the re-read (L-183), but its Retry is not busy, and a successful read unmounts the panel with no focus repair, so focus falls to <body>.
Fix: use Contacts’ pattern on both lists: one retry handler for both panels (notePress, then arm once the read settles), retrying on the error view, and the error Retry busy; on Users also map a held failure with no snapshot to error.
Regression test: per list, after a first-load failure, Retry stays the same element, busy, and focused through a second failure, and a successful retry moves focus to the h1.
Resolved: 2026-11-06 — fixed in forge-et-fe#429, the fix above on both lists. Regression tests: the BUG-344 blocks in UsersScreen.test.tsx and EmailTemplatesScreen.test.tsx, and a useUsersList row in usersQueries.test.tsx. All failed on main’s code; the hook row read loading where error+retrying is expected. A Chromium probe against a running backend showed the same on both lists: Retry busy and focused during the read, still focused after a second failure, and focus on the h1 after a good read.
Lesson: L-183 — another instance: the error state did not use the hold that the stale notice already used. L-183’s sources now list BUG-344.
BUG-343 — The Fast lane’s Set up Trivy step fails when get.trivy.dev is unreachable, blocking every ready app-repo PR (Linux CI)
Affects: the CI of forge-et-be, forge-et-fe and forge-et-pwa (.github/workflows/pr-validation.yml, step Set up Trivy: aquasecurity/setup-trivy@v0.3.1 with version: v0.74.0) · Found in: forge-et-be#361’s ready run on 2026-11-04 (run 41763237262, both attempts); the latest forge-et-fe and forge-et-pwa runs the same hour; feature/F20-i-wave-2 the night before · Severity: high — while the host is down no app-repo PR can merge, because the Fast lane is a required check · Phase: 1
Repro: the action runs trivy/contrib/install.sh from Trivy’s main. It downloads the binary from https://get.trivy.dev/trivy?... (GET_DOWNLOAD in the script), not from GitHub’s releases. On 2026-11-04 around 03:00 UTC, get.trivy.dev resolved to no address and a request from the lead’s Mac timed out. The step logs “found version: 0.74.0 for v0.74.0/Linux/64bit” and then “Process completed with exit code 6” (curl: could not resolve host).
Expected: Trivy installs from a source that does not depend on one third-party host, so a PR’s Fast lane fails only on its own code.
Actual: the step fails, and so does the Fast lane, in all three repos.
Fix: install the pinned version from GitHub’s release assets, which the lead chose on 2026-11-04 over waiting for the host.
- In all three workflows, Set up Trivy is a
runstep. It downloadstrivy_0.74.0_Linux-64bit.tar.gzfromgithub.com/aquasecurity/trivy/releases/download/v0.74.0/with boundedcurl. - It checks the tarball against a SHA-256 pinned in the workflow beside the version (
VandSHA, bumped together by hand). The pin equals the release’s published digest, and the release is immutable. - It unpacks only the
trivymember into$RUNNER_TEMP/trivyand puts it onPATH. aquasecurity/setup-trivyis no longer used.
Regression test: each fix PR’s own full Fast lane. In each one, Set up Trivy and the Trivy scan passed with the new install while get.trivy.dev still did not resolve.
Resolved: 2026-11-04 — fixed in forge-et-be#363, forge-et-fe#402 and forge-et-pwa#109. The blocked PRs (F20-i wave 2’s forge-et-fe#395 and forge-et-be#356, F43’s forge-et-be#361) merged on a fresh run after them. .forge/security/README.md states the new install.
Lesson: none — one-off: an upstream host outage behind a third-party action’s install script. The fix removes the dependency, so the Fast lane’s downloads now come only from github.com releases, checked against pinned or published checksums, as the gitleaks step’s already did.
BUG-342 — A model reply refused by Haiku 5.5’s safety classifiers is retried as a malformed answer, spending a second call
Affects: F12 (forge-et-be BedrockAiProvider, AiCallRunner; every AI feature goes through them) · Found in: the Haiku 5.5 readiness review of 26 October (model-choice-evaluation-plan.md P-4; .forge/owed.md, “Haiku 5.5 readiness in the backend”, item 2), not seen live: no feature runs on Haiku 5.5 yet · Severity: low — no user sees a wrong result; a refused call costs one extra Bedrock call before the manual path · Phase: 1
Repro: on a fake Bedrock client, answer a Converse call with a reply that has no tool call and the stop reason refusal, which the SDK’s StopReason does not name, so it arrives as a raw string.
Expected: the call ends INVALID_OUTPUT after one attempt, and the log says it was refused.
Actual: BedrockAiProvider classifies the reply INVALID_OUTPUT (NO_TOOL_USE stopReason=refusal), so the runner retries it once. The same request usually refuses again, so the call takes two attempts before the manual path.
Fix: a reply with no tool call and a stop reason of refusal, content_filtered or guardrail_intervened (compared in lower case) is a new ModelCallException.Kind.REFUSED, detail REFUSED stopReason=<value>, and AiCallRunner ends the call INVALID_OUTPUT at once, unretried. The ai_invocation.outcome value and every feature are unchanged, and the WARN Bedrock Attempt Failed line names kind=REFUSED. The lead chose this over a new REFUSED outcome (2026-11-04), which would need a Flyway migration of the outcome check and a change in every place that branches on AiFailure.
Regression test: BedrockAiProviderSpec, “a reply stopped for #stop is REFUSED …” (three stop reasons), “a refusal ends INVALID_OUTPUT after one client call, not retried (BUG-342)”, and the WARN-line case; AiCallRunnerSpec, “a REFUSED attempt ends INVALID_OUTPUT at once, unretried …”. On main’s code six cases failed: the runner case with TooFewInvocationsError for record(INVALID_OUTPUT, attempts == 1), because the runner made a second call.
Resolved: 2026-11-04 — fixed in forge-et-be#361. Which stop reason Haiku 5.5 sends for a refusal on Bedrock is proved once it answers on dev (register §B F12 item 11).
Lesson: none — one-off: a difference of a newer model from Haiku 4.5, found before any feature moves to it. Like BUG-248, which was a request shape a later model rejected, it was caught by the step ai-provider-strategy.md §4d already requires before a switch (the readiness review and the feature’s evaluation run). The two causes differ, so they are not one pattern, and no lesson in .forge/lessons.md names this cause.
BUG-341 — The mailbox settings have two ways in: the Email screen’s settings button and a Mailbox item in the sign-out menu
Affects: F20-i (forge-et-fe src/app/CurrentUserMenu.tsx; F20-i FR-19, AC-40; design-system Rev 80) · Found in: a UI review of main by Nanndish and Tharinda, 2026-11-03 · Severity: low — a duplicate way to the same page, in a menu where nobody looks for email settings · Phase: 1
Repro: sign in as a user whose role holds email.manage; open the current-user menu in the sidebar footer.
Expected: one way to the mailbox settings, on the Email screen.
Actual: the menu lists Mailbox (to /email/mailbox) before Sign out, beside the Email screen’s settings button. F20-i added it as the way in with no connection, where the Email screen hides its button; in that state the Email screen already shows the same Try again or Reconnect as /email/mailbox, so the item led to a page that offered nothing more.
Fix: remove the item; the menu holds Sign out only. F20-i Rev 4 changes FR-19 and AC-40, and design-system Rev 88 the §6 Current-user menu. The Email screen’s button, and its rule for when it shows, are unchanged.
Regression test: src/app/CurrentUserMenu.test.tsx, “lists Sign out only for an email.manage holder” and “… for a user without email.manage”; e2e/browser/shell/current-user-menu.spec.ts asserts one item.
Resolved: 2026-11-03 — fixed in forge-et-fe#396 (spec and design in forge-et-harness#1105). The holder row failed on main’s menu (“expected [ ‘Mailbox’, ‘Sign out’ ] to deeply equal [ ‘Sign out’ ]”) and passes on the fix. Checked in a running browser on 2026-11-04 against main abce02e1 and the local backend, signed in through the stub as the all.permissions test user (who holds email.manage): the open menu’s items read ["Sign out"], and no app-user-menu-mailbox element is in the page. The changed browser specs first run in the nightly full run on main.
Lesson: none — one-off: a spec choice reversed at a UI review, not a code defect. The rule it broke is already design-system §5’s Section settings affordance: no way in to a sub-view that shows nothing the current screen does not (BUG-015, BUG-020). No other Lesson: line or lesson in .forge/lessons.md names this cause.
BUG-340 — The price round’s AC-25 keyboard case finds no tooltip on focus once the verify stack’s database holds an earlier run’s corrections
Affects: F26-f (forge-et-fe e2e/browser/pricing/round-correct.spec.ts:257, “AC-25 rendered: the new row shows Published · Corrected … and its reason, who and when on hover and on keyboard focus”) · Found in: the second whole chromium run of F20-i WI-2.6’s Phase 2 on the lead’s Mac (forge-et-fe WI-2.6 branch 12103701), on a database that the first run and the seam runs had already used, 2026-11-03 · Severity: low — a browser test fails on a reused database; the product is not affected as far as is known · Phase: 1
Repro: on the F20-i verify stack, run the whole chromium project twice on the same database, or run round-correct.spec.ts alone after one whole run. The hover half passes; after keyTo focuses the Published — corrected status button, getByRole('tooltip') with the correction’s text is not found (spec line 282). The row’s accessible text holds the expected “Corrected by … on Tue 3 Nov, 22:53: E2E correction w3 …”. The case passed on the merge-base baseline and on the first whole run on the wave, at the same pricing code: F20-i changes nothing in src/features/pricing, src/ui, e2e/browser/pricing or e2e/support. Run alone with the file three times in a row, it failed each time.
Expected: the tooltip opens on keyboard focus whatever earlier runs left in the database.
Actual: no tooltip on focus.
Fix: not decided. First decide which state the case depends on (the round’s rows or the corrections an earlier run left), then whether the cause is the test’s arrangement or the Tooltip’s focus behaviour.
Regression test: the file run twice on one database, passing both times.
Not reproduced, 2026-11-04 (Tharinda’s session): on the lead’s Mac, against the primary forge-et-be compose database, which earlier runs that day had used: the whole e2e/browser/pricing folder with the two pricing seam specs (268 passed), then round-correct.spec.ts alone three times in a row (17 of 17 each), all passing; each run adds a correction of 90302 to sale 2026 / 40, so the round held several earlier corrections. What differs from the failing stack is not yet known. A lead for the next attempt: the shared src/ui/components/Tooltip/Tooltip.tsx is controlled and has its own handling of a close right after a focus open (BUG-181) and of focus arriving during the close delay (BUG-182); the case hovers, moves the pointer away, then tabs from the search box to the status button, past any earlier rows’ tooltips, so a test of two tooltips focused one after the other in quick succession may show it.
Tooltip test, 2026-11-04: a component test of that sequence in Tooltip.test.tsx (an input, three triggers with delayMs={0}; hover the last, move the pointer away, focus the input, Tab three times to the last), once with the Tab inside the close delay (fake timers) and once after the close delay: the last trigger’s reason was shown and the other two were closed both times; the file passed 24 of 24 on each of three runs. The F20-i verify stack’s database, the only state that showed the failure, is gone (no container or volume is left). The test was not committed: it found nothing, and the existing BUG-182 test covers the closest path.
Resolved: 2026-11-04 — closed as wontfix (not reproducible), the lead’s decision (Tharinda, 2026-11-04: write the Tooltip test, then close if it passes). The evidence is the two attempts above: the pricing browser specs on a database holding several earlier corrections, and the component test of the sequence. No change was made. A recurrence is filed as a new bug that links here, with the database kept (a pg_dump of the stack) so the failing state can be read.
Lesson: none — one-off: a browser failure seen only on one stack’s database, which was removed before anyone triaged it. When a browser failure depends on data, keep the database (or a dump) until the bug is triaged.
BUG-339 — Flaky test: Leads inbox AC-14a, the directory case: the Country select’s list is still open after Sweden is chosen (macOS)
Affects: F11-b (forge-et-fe e2e/browser/leads/leads-inbox.spec.ts:407, “AC-14a (smoke): with the directory read failing, its select shows the error and Retry; the list, Country and the lead page still work; Retry re-reads only that list”) · Found in: the whole chromium project in F20-i WI-2.6’s Phase 2 browser run (forge-et-fe WI-2.6 branch d05cda88 on feature/F20-i-wave-2, which changes nothing in leads), the lead’s Mac, 2026-11-03 · Severity: low — one browser test fails now and then; the product is not affected · Phase: 1
Repro: the whole chromium project (1,294 tests, 3 workers) with a load average of about 5.6 on 8 cores from desktop apps outside the run. The case failed in chooseFilter (leads-inbox.spec.ts:1473): after the Sweden option was clicked, getByRole('listbox') still counted 1 at the 5-second timeout. It passed on the merge-base baseline the same evening and on a re-run of the case alone straight after. BUG-182 is the same test’s tags case on Windows, with a different symptom (one request too many in Retry’s counting window).
Expected: choosing an option closes the select’s list, and the case passes in a full run.
Actual: the list was still open when the case checked it.
Fix: not decided. Triage per .forge/test-strategy.md § Flaky Tests: first decide whether this is the overload outcome (the run had no load check of its own) or a click that lands before the option is ready (compare the option’s wait with the Radix Select arrow-key timing F20-i WI-2.4 found in inbox-a11y.spec.ts).
Regression test: the case ten times in a full chromium run within the load limit, ten in ten.
Triage, 2026-11-04 (step 2 of .forge/test-strategy.md § Flaky Tests): on the lead’s Mac, forge-et-fe main 4421bd61 against a backend on port 8081, playwright test e2e/browser/leads/leads-inbox.spec.ts --project=chromium -g 'with the directory read failing' --repeat-each=10 with one busy loop per core (yes > /dev/null × 8) beside it and other sessions’ work going, a starting load average of 18 on 8 cores: 10 of 10 passed (1.9 min). The CDP CPU throttle was not tried. Forced load did not reproduce it.
Full step 2, 2026-11-04 07:07–07:12, after the nightly run, on a quiet machine (load average 1.8 at the start): the same case, --repeat-each=10 each time, at forge-et-fe main 4421bd61 against forge-et-be main d3df4fc2 on port 8081: one busy loop per core alone, 10 of 10 passed; with the Chromium CPU throttle added (Emulation.setCPUThrottlingRate, a temporary uncommitted edit) at rate 4, 10 of 10, and at rate 6, 10 of 10. A run at rate 10 failed all 10 earlier, at settleLeadsList (the list never loaded within 5 s), with 16 busy loops by mistake: a browser too slow for any test, not this bug’s symptom. No step can be held back: the Country options are a fixed list in the code (COUNTRY_CODES), so nothing is in flight when Sweden is clicked. Step 2 did not reproduce it.
Resolved: 2026-11-04 — closed as wontfix (machine overload), the lead’s decision (Tharinda, 2026-11-03: reproduce, then close if it does not). The evidence: the failing whole-project run had no load check of its own and ran at a load average of about 5.6 from desktop apps outside it, it passed on the merge-base baseline and alone straight after, and the full step 2 (forced load, the CPU throttle at rates 4 and 6, nothing to hold back) did not reproduce it. The rule’s process-list test (another gate, build, backend or container at the peak) is not met by desktop apps, so this closure rests on the lead’s decision. No change was made. A recurrence in a run within the load limit is filed as a new bug that links here.
Lesson: L-162 (a failure that appears only under load is reproduced before it is called flaky). An instance where forced load did not reproduce it, so nothing was changed, as for BUG-195 and BUG-259.
BUG-338 — Flaky test: the router test that resolves /pricing does not find the Pricing heading before its timeout in a loaded run (Windows)
Affects: F5-b (forge-et-fe src/app/router.test.tsx, “router — Phase-1 route skeleton (AC-3) › resolves /pricing to its screen, not a 404”) · Found in: a Vitest run of src/features/{settings,users}, three pricing test files, src/features/email/InboxToolbar.test.tsx and src/app (88 files) on branch fix/BUG-332-ui-consistency (1989c312’s tree, before commit), Windows 11, 16 cores, with other sessions’ backend and dev server running, 2026-11-03 · Severity: low — one test fails now and then under load; the route resolves · Phase: 1
Repro: that run: Unable to find role="heading" and name "Pricing" after 10.3 s. Alone it passed two runs in two on the branch and one in one without the branch’s changes.
Expected: the test passes whatever else the machine is running.
Actual: under load the lazily loaded Pricing screen has not rendered its heading when the query gives up.
Cause (2026-11-06, Windows 11, other sessions holding the CPU at 100%): the test paid the Pricing module’s first import inside its own heading wait. Vitest transforms and evaluates a module the first time a file imports it. router.test.tsx imports router.tsx statically, but each screen is a lazy() import that runs only when a case first renders its route. A diagnostic file timed import('../features/pricing/PricingRoute') from a fresh worker at 15.0 s cold, and at 1.5 s even after router.tsx and its shared modules had loaded. The second import of the same module took 0.5 ms. The case’s wait is 10 s, so a loaded run ran out before the screen rendered. Every case in the file whose route is the first to reach its lazy() module carries the same risk; Pricing failed because its tree, which brings React Aria’s combobox, is the heaviest.
Fix: src/app/router.test.tsx now imports every module router.tsx loads through lazy() while the file loads, before any test, the same as its static imports. It reads the import('…') specifiers from router.tsx, with comments removed, as e2e/fixtures/routes.setup.ts does (BUG-079). The file sits beside router.tsx, so each relative specifier names the same module the router imports. The file fails to load if the count of import('…') specifiers differs from the count of lazy() calls, or from the count of = lazy( declarations in the source before comments are removed (the review’s Nit 2). The warm-up is a top-level await, not a beforeAll: loading a file has no time limit, while a hook has Vitest’s 10 s default. Each case’s wait now covers only the route resolving and its screen rendering. The wait is unchanged at 10 s; the constant is renamed from LAZY_ROUTE_TIMEOUT_MS to ROUTE_RENDER_TIMEOUT_MS, and its comment now says what it covers. Only the test changes; no app code and no shared test helper.
Regression test: proof by a forced failure (L-107). A temporary Vite plugin in vitest.config.ts (reverted after) added 600 ms to the transform of each src/features/pricing/ module. With it, the pre-fix file run with -t "resolves /pricing to its screen" failed with this bug’s symptom, Unable to find role="heading" and name "Pricing", at 10,336 ms. The fixed file under the same plugin passed, with the cost paid while the file loaded (transform 80% of the run). Then the whole fixed file ran five times in a row on Windows 11 with other sessions’ work holding the CPU at 100% at each start: 80 of 80 passed each time (54–79 s).
Resolved: 2026-11-06 — fixed in forge-et-fe#433 (b14ab623, with the review’s Nit fixes in 04109937). /forge-review-pr recorded approve on 04109937 with frontend-reviewer: no Blocker or Important; Nits 2–4 fixed, Nits 1 and 5 left as is with reasons in the record.
Lesson: L-116 — another instance, in a Vitest file rather than a browser spec: a test that renders a lazy() screen pays its module’s first load inside its own wait. Added to L-116’s sources with an update.
BUG-337 — Flaky test: three held URL-state tests (Weekly prices, Activities, Standards book) find the read idle where they wait for it to load (Windows)
Affects: F26-f (forge-et-fe src/features/pricing/pricingUrlState.held.test.tsx:52, “a sale chosen, then a search before the sale has rendered, keeps both”); F1-c (src/features/activities/activitiesUrlState.held.test.tsx:41) and F26-b (src/features/pricing/standardsBookUrlState.held.test.tsx:50), added 2026-11-05 from the review of forge-et-fe#411, which saw both fail with the same expected 'idle' to be 'loading' under load and pass alone · Found in: a Vitest run of src/features/{users,email,pricing,settings}, src/ui and src/app/shadcnOnly.test.ts together (248 files, 279 s) on branch fix/BUG-332-ui-consistency from main d0408c73, Windows 11, 16 cores, with another session’s backend and dev server running, 2026-11-03 · Severity: low — one test fails now and then under load; the hook is right · Phase: 1
Repro: that run: AssertionError: expected 'idle' to be 'loading' at line 58. Alone it passed three runs in three on the branch and two in two on main; the branch does not change the hook or the test.
Expected: the test passes whatever else the machine is running.
Actual: under load the read the test waits for is already idle when it checks for loading.
Also seen (F3-a-WI-3.1’s local runs, 2026-11-03): two full pnpm test runs on forge-et-fe e94ce82d (forge-et-fe#393, Windows, other sessions’ work running), each 11248 passed, 1 skipped, 1 failed: the first lost activitiesUrlState.held.test.tsx, the second pricingUrlState.held.test.tsx › “a sale chosen, then a search before the sale has rendered, keeps both”, both expected 'idle' to be 'loading'. Each file passed alone 3 of 3; that branch changes neither file nor the hooks they test.
Cause: the review of forge-et-fe#411 was right. React Router never sets navigation.state to loading during the router’s first load (7.18.5, handleLoaders: shouldUpdateNavigationState = !isUninterruptedRevalidation && !initialHydration), so each test’s first waitFor(… navigation.state … 'idle') passed at once, before router.state.initialized was true and before the route’s probe had rendered. The probes rendered null and captured the hook’s update function into a module variable, so under load the test called the module default (a no-op) or the previous test’s function, which reached no router, and no navigation ever started. Not the loading state being short-lived: the held loader keeps it until release(). Reproduced every time by delaying the first loader call in src/test/heldLoader.ts by 100 ms, for the run only: all five tests in the three files failed with expected 'idle' to be 'loading', with initialized: false and the probe rendered 0 times at the failing line. No other heldLoader user waits for idle before its first render: the rest wait for a rendered element first (useListSearchParams.test.tsx), or wait for idle only after a held navigation is released.
Fix: test-only, in the three files. Each probe renders <span data-testid="probe" /> and the setup awaits screen.findByTestId('probe') before acting. The captured functions reset in afterEach to one that throws (the probe has not rendered yet). The waitFor on loading became a synchronous check of what the change caused: the pending navigation’s URL holds the first value and the rendered URL does not. The final checks wait for the whole end state (both values in the URL) in one waitFor, not for idle and then a read. These hook tests render no rows, so the URL is the end state. No wait was lengthened; src/test/heldLoader.ts is unchanged.
Regression test: the three files themselves, on forge-et-fe branch fix/BUG-337-held-url-state-tests (e8d54317, ffc48bc3). Fails before: with the first loader call delayed by 100 ms, the five tests from main (4a91ed9c) failed five in five with expected 'idle' to be 'loading'; after: the fixed tests passed five in five under the same delay. Loaded runs on Windows 11, 16 cores, with other sessions’ Vitest runs keeping the CPU at 100% (sampled before each run): 20 runs of the three files, 20 passed, 5 tests each. pnpm typecheck and biome check on the three files exit 0.
Resolved: 2026-11-06 — fixed in forge-et-fe#432 (e8d54317, and ffc48bc3 for the review’s four Nits: the Auction sales and Activities state read through a function that throws before render, exact on-screen values in the not-yet-rendered checks, a location-key check in the Standards book’s on-and-off test, docblock wording). Reviewed by /forge-review-pr with frontend-reviewer: no Blocker or Important, record approve on ffc48bc3.
Lesson: L-261 — a test starts once what it drives has rendered, not on a state that already holds before the first load.
BUG-336 — The Settings tag pages have no pager, unlike every other list
Affects: F1-e (forge-et-fe src/features/settings/SettingsTagsRoute.tsx, TagSettingsTable.tsx) · Found in: Nanndish’s UI review, 2026-11-03 · Severity: low — a long tag list is one long table; nothing is lost · Phase: 1
Repro: open Settings › Account tags with more than 25 account tags (the local data has 119).
Expected: the shared ListPager under the table, as on Users, Contacts and Leads (design-system §6 Pagination): 1–25 of 119, the page numbers and Rows per page.
Actual: every tag in one table, with no pager (F1-e FR-T2 as written before Rev 11).
Fix: page in the browser: both lists are already read whole and sorted there (F1-e D-9), and the tag API sorts by name only, so server paging would lose the Colour and Status sorts. tagSettingsPage.ts reads page and size from the URL (25, 50, 100; the defaults never written), pageOf takes the page and shows the last page for a page past the end, and the route renders ListPager under the table. A page change adds a history entry; a size change and a new order return to page 1. The panel shares one ListSearchParamsProvider, so a header press that writes the sort and the page keeps both. A retire or a restore that moves the tag to another page turns to that page, replacing the history entry, and focuses its ⋯ menu, so F1-e NFR-7 holds with the pager. F1-e Rev 11 and design-system Rev 85 record the change.
Regression test: SettingsTagsRoute.test.tsx › “the pager (BUG-336)”: 30 tags show 25 rows, 1–25 of 30, Next goes to ?page=2 with five rows, a new order returns to page 1; a page past the end shows the last page; size=50 shows all 30; a retire on page 1 turns to page 2 with focus on the tag’s ⋯ menu, and its restore back to page 1; a sort header pressed during the first load keeps focus. The size matrix gains settings-tags-paged. The browser specs AC-27 and the retire order check compare the page shown, not the whole list.
Resolved: 2026-11-03 — fixed in forge-et-fe#392 (f5fb1a2c, 1989c312, 92515f99, a3bd735d). The page turn gives way to a user who moved while the write ran (a pager press, Back), so it never overrides their page or focus (L-180); when the control that started the write has left the page and focus is nowhere, focus goes to the tag’s ⋯ menu on the user’s page, else to the h1. Fails before: the pager tests fail with the fix removed, the two focus tests on f5fb1a2c (expected '' to be '?page=2', expected false to be true), the moved-user test on 1023ba78 (expected '' to be '?page=2'), and the Back-during-retire test on 92515f99 (focus on <body>); after: SettingsTagsRoute.test.tsx 55 passed, three runs in three. The changed browser specs were not run: another session’s dev server held port 5173, which pnpm e2e would have reused (L-251) and the gate needs free; the nightly full run on main runs them. In the browser, 119 local account tags showed 1–25 of 119, Next went to ?page=2 (26–50 of 119), and a sort returned to 1–25 of 119 with ?sort=name%2Casc.
Lesson: none — one-off: the tag pages were the one table outside the shared pager, by an earlier decision (F1-e D-9) that this change keeps.
BUG-335 — The Settings tag pages’ first load is skeleton lines, not the table’s skeleton rows
Affects: F1-e (forge-et-fe src/features/settings/SettingsTagsRoute.tsx, TagSettingsTable.tsx) · Found in: Nanndish’s UI review, 2026-11-03 · Severity: low — the page jumps from grey lines to a table when the tags arrive · Phase: 1
Repro: open Settings › Account tags at bp/tablet or wider and watch the first read.
Expected: the table’s header over skeleton rows of the populated row’s height, as every list’s first load (design-system §6 Data table › Empty and loading), the way Users and Contacts load.
Actual: four SettingsSkeleton lines with no header, then the table.
Fix: TagSettingsTable takes rows="loading": from bp/tablet up it draws its header (sortable as ever) over TAG_SKELETON_ROWS (5) placeholder rows of four cells, aria-busy, with the container’s test id settings-tags-loading; below bp/tablet, where the rows are blocks, it keeps the skeleton lines.
Regression test: SettingsTagsRoute.test.tsx › “at bp/tablet and wider, the first load is the table’s header over five skeleton rows (BUG-335)”; it fails without the fix.
Resolved: 2026-11-03 — fixed in forge-et-fe#392 (f5fb1a2c). Fails before: the test fails with the fix removed; after: SettingsTagsRoute.test.tsx 51 passed.
Lesson: L-253 — the same cause as BUG-334: a list-wide loading rule that did not reach every list.
BUG-334 — The Users list keeps its old rows while a sort, page or filter is read, where every other list draws skeleton rows
Affects: F5-c (forge-et-fe src/features/users/UsersScreen.tsx, UsersTable.tsx) · Found in: Nanndish’s UI review, 2026-11-03 · Severity: low — the list shows the previous order until the new one arrives · Phase: 1
Repro: on Settings › Users, press the Email header.
Expected: design-system Rev 83 (§6 Data table › Empty and loading): as many skeleton rows as it showed, under its header, with the pager busy and the status line saying the list is updating — as Contacts, Accounts, Leads and the Pricing tabs do.
Actual: the old rows stay until the new page arrives, and the status line is empty. Users was built before Rev 83, and Rev 83’s carry-forward did not name F5-c.
Fix: as ContactsScreen.tsx: the rows branch passes rows={view.kept ? 'loading' : view.list.items} and skeletonRows={view.list.items.length}; UsersTable draws skeletonRowKeys(skeletonRows) (five on the first load); the status line says Updating… while kept. A same-request re-read still keeps the rows. F5-c FR-2 already asks for skeleton rows, so its text is unchanged.
Regression test: UsersScreen.test.tsx › “draws as many skeleton rows as it showed while a sort is read, under the same header (BUG-334)” and “says Updating… while a page change is in flight”; both fail without the fix.
Resolved: 2026-11-03 — fixed in forge-et-fe#392 (f5fb1a2c). Fails before: both tests fail with the fix removed; after: UsersScreen.test.tsx 49 passed. In the browser, with the users read delayed, pressing Email showed 25 skeleton rows under the same header and Updating…, then 25 rows, with focus kept on the header button.
Lesson: L-253 (new) — Rev 83’s scope named the lists from memory, and the oldest one was left out.
BUG-333 — The Weekly prices status filters fill the one in force near-black, unlike the Leads status pills
Affects: F26-f (forge-et-fe src/features/pricing/RoundSummaryBar.tsx, RoundSummaryBar.module.css) · Found in: Nanndish’s UI review, 2026-11-03 · Severity: low — a different look for the same control · Phase: 1
Repro: open Pricing › Weekly prices; compare the selected All filter with the selected New pill on Leads.
Expected: the same pill as Leads (design-system §6 Filter pills row): a shadcn Button, the primary fill for the one in force and the outline for the others.
Actual: hand-styled links whose selected fill is --et-text-primary, the near-black text colour.
Fix: each filter is a shadcn Button asChild (size sm, rounded-full) around its Link, default when current and outline otherwise; it stays a link with aria-current, since it changes the URL. Can’t be published keeps its danger tone while not current. The not-in-round button is an outline pill with a dashed border. The CSS module keeps only what the variants do not set.
Regression test: RoundSummaryBar.test.tsx › “draws the filters as the Leads status pills … (BUG-333)”; it fails without the fix.
Resolved: 2026-11-03 — fixed in forge-et-fe#392 (f5fb1a2c). Fails before: the test fails with the fix removed; after: RoundSummaryBar.test.tsx passed. In the browser, the selected All computes to the Leads pill’s fill (rgb(138, 90, 18), white text, 32px pill), the others to its outline.
Lesson: L-254 — another instance: the Weekly prices filters were a screen’s own variant of the filter pill, built apart from the Leads pills; added to L-254’s sources. src/app/shadcnOnly.test.ts does not see a styled <Link>, so frontend-reviewer dimension 28 is where the next one is caught.
BUG-332 — The Email header’s Mailbox settings button is the only icon-only action beside labelled ones
Affects: F20-b (forge-et-fe src/features/email/InboxToolbar.tsx) · Found in: Nanndish’s UI review, 2026-11-03 · Severity: low — a gear with no visible word or tooltip beside Refresh and Compose · Phase: 1
Repro: open Email with a mailbox connected.
Expected: Nanndish chose, 2026-11-03, the icon with the words Mailbox settings, in Refresh’s outline style (design-system Rev 85).
Actual: a ghost icon button (design-system §5 Section settings affordance, Rev 25 and Rev 84).
Fix: an outline Button with the settings icon and the words Mailbox settings, with no aria-label: the words are the accessible name (L-211), so F20-b AC-25b and the browser specs that find it by name are unchanged.
Regression test: InboxToolbar.test.tsx › “renders a button with the accessible name first among the actions …” now also asserts the words, no aria-label, the icon and the outline variant; it fails without the fix.
Resolved: 2026-11-03 — fixed in forge-et-fe#392 (f5fb1a2c, 1023ba78). Fails before: the test fails with the fix removed; after: the 1,368 tests of src/features/email passed. Not seen in the browser: the local test user’s mailbox is disconnected, so the button is absent by design (AC-25c), and connecting it would have changed a local backend another session was using.
Lesson: none — one-off: a design choice changed, not a defect that recurs.
BUG-331 — Flaky test: AccountListScreen.test.tsx’s stale-view pager row times out at 20 s in a test record on an overloaded machine (Windows)
Affects: F1-b (forge-et-fe src/features/accounts/AccountListScreen.test.tsx, “the data states (AC-7) › a stale view shows the pager only when its rows answer the same request”) · Found in: pnpm test:record on forge-et-fe#390’s head f4e61f2d (6,660 tests; the PR does not touch the account list), Windows 11, 2026-11-03 · Severity: low — one unit test times out under load; the behaviour works · Phase: 1
Repro: the record’s Machine load: line read “busy 0.45 at start, 1.00 peak, 0.95 average (16 cores, Windows)”; the machine was busy with Windows’ own processes (system, svchost, msmpeng), and other worktrees’ Postgres and LocalStack containers were running. The row failed with “Test timed out in 20000ms”; the file alone (pnpm exec vitest run src/features/accounts/AccountListScreen.test.tsx) passed 98 of 98 three times straight after.
Expected: the row passes in a full run.
Actual: the row ran past Vitest’s 20-second timeout.
Fix: a slow row (.forge/test-strategy.md § Flaky Tests cause 6). The row checks three requests — other filters, another page, the same request after an archive — and renders the account list once for each, with a Radix select, a page change and a row menu between them, all inside one 20-second budget. Split it into one case per request under a describe that keeps the row’s name; each case keeps its own render and assertions.
Regression test: the file ten times in a full run within the load limit, ten in ten.
Resolved: 2026-11-03 — fixed in forge-et-fe#413 (d6bd3bc1; review fixes d4996df1, be5c7cf9), test only, with BUG-329. Checked on Windows 11 while about twelve other sessions’ work (Claude, Docker, Chrome) held the machine at 100% CPU (Get-Counter), over the record’s 0.65 limit, so these runs stand in for triage step 2’s forced load. Before: the file alone failed the row at 22.2 s (“Test timed out in 20000ms”; five other rows in the file also timed out under that load, 1,083 s for the file), and the row alone (-t) took 21.0 s and 21.5 s, both failing, and 18.7 s. The single-request rows beside it took 9–11 s, so three requests in one case did not fit the 20 s budget. After: the three cases alone, three runs, passed 9 of 9 at 7.5–19.6 s each (the slowest is the first case of a -t run, which carries the file’s first render); the whole file twice, 100 of 100 both times, the three cases at 3.4–7.1 s. The ten-in-ten full run within the load limit was not run here, because the sessions sharing this machine were told not to run full suites; the PR’s test record and the nightly run take it.
Lesson: L-125 — another case whose arrangement did several cases’ work in one test budget (.forge/test-strategy.md § Flaky Tests cause 6), after BUG-128, BUG-213 and BUG-220. Added to L-125’s sources.
BUG-330 — A Playwright run started outside the gate holds port 5173 while it waits for the heavy-run lock, so the lock holder’s browser gate cannot start
Affects: FS-4 (forge-et-fe e2e/support/heavyRunLock.ts, playwright.config.ts globalSetup and webServer; harness design test-run-cost.md M-5) · Found in: the nightly full run started by hand on the lead’s Mac at 09:51 on 2026-11-03, while a pnpm exec playwright test of one spec (forge-et-fe#375’s BUG-283 test) waited on the lock · Severity: medium — one waiting single-spec run stopped the nightly run’s CRM browser gate from starting; the product is not affected · Phase: 1
Repro: while another heavy run holds the lock (a test record, a backend build record or the nightly run), start pnpm exec playwright test <one spec> --project=chromium in a frontend worktree. It prints “Waiting for … to finish” and keeps waiting; lsof -nP -iTCP:5173 -sTCP:LISTEN shows its vite preview listening. When the lock holder’s own browser gate starts, it refuses on port 5173.
Expected: a run that cannot start yet holds nothing another run needs: one heavy run at a time on a machine, and the gate’s port free for the run that holds the lock (test-run-cost.md M-5).
Actual: Playwright starts its webServer before globalSetup, and the lock wait is in globalSetup (heavyRunLock.ts). The waiting run therefore holds port 5173 for as long as it waits. heavyRunLock.ts’s header states this (“a run that is still waiting already holds port 5173, so a gate that takes the lock while this run waits refuses on the port”), so the limit is known but not prevented. On 2026-11-03 the waiting run held the port for 36 minutes, until the nightly session stopped it.
Fix: take the lock before Playwright starts its server: pnpm e2e (and a documented wrapper for one spec) takes the lock in a script that then runs playwright test, as pnpm e2e:gate does. Or, when a run outside the gate finds the lock held, refuse at once with the holder’s name instead of waiting, so its server stops.
Regression test: a heavyRunLock test that, with the lock held by another process, shows a run outside the gate either refusing without waiting or waiting without anything listening on 5173.
Update 2026-11-04: forge-et-fe#399 and forge-et-pwa#108, both merged on 2026-11-04, moved the gates to ports of their own, 15173 and 15174, by the lead’s decision. A waiting run still holds 5173 but no longer stops a gate: the case in Found in cannot happen.
Resolved: 2026-11-04 — fixed by forge-et-fe#399 (8cae2485) and forge-et-pwa#108 (498c09cb), closed by the lead’s decision (Tharinda, 2026-11-04), with no work started on the fix above. The gates no longer serve on 5173, so the run that holds the lock never needs the port a waiting run holds. What is left is a waiting run holding 5173 against the developer’s own pnpm dev, which then starts on 5174, as Vite does whenever 5173 is taken. That is not worth the lock-before-server change. The regression proof: forge-et-pwa’s full gate passed (134 tests) while another server held 5173, and the shared browserGateRun.test.ts asserts that the gate’s port is not 5173.
Lesson: none — one-off: the cause was two kinds of run sharing one port, and the fix gave the gate a port of its own, so it cannot recur in this form.
BUG-329 — Flaky test: leadResearchComposition.test.tsx’s 404-removal row still finds the research panel when it checks, in a full selected run (Windows)
Affects: F11-c (forge-et-fe src/app/leadResearchComposition.test.tsx, “a removal answered 404 (F11-c Wave 1 contract client-side fact 7) › hands the lead to the view’s out-of-reach handling: the selection clears and the notice says so”) · Found in: pnpm test:selected on forge-et-fe fix/BUG-301-small-a11y-batch (6,650 tests; the branch does not touch the file), Windows 11, with another session’s dev server running, 2026-11-03 · Severity: low — one unit test fails now and then; the behaviour works · Phase: 1
Repro: a full or selected Vitest run on a busy Windows machine. The run’s recorded load was not printed (test:selected prints none). The test took 2,085 ms and failed on expect(element).not.toBeInTheDocument(); run alone (pnpm exec vitest run src/app/leadResearchComposition.test.tsx) it passed 4 of 4 straight after.
Expected: the row passes in a full run.
Actual: the element the row expects gone after the 404 was still in the document when it checked.
Fix: the assertion names the research panel, which the row waits to leave with waitFor at Testing Library’s default one second. React Router 7 renders the navigation to /leads in a startTransition, so the lead view, panel included, stays on screen until React has rendered the list screen; under load that takes more than a second. Nothing races: the outcome is the same every time, only later. The 404 row now waits for the list’s notice with the five-second page budget LeadsScreenActions.test.tsx already gives the same wait (PAGE_TIMEOUT, a named constant that says why), then checks that the panel has gone; the notice can only appear once the lead view has unmounted.
Regression test: the file ten times in a full run, ten in ten.
Resolved: 2026-11-03 — fixed in forge-et-fe#413 (d6bd3bc1; review fixes d4996df1, be5c7cf9), test only, with BUG-331. Checked on Windows 11 while the machine was at 100% CPU from about twelve other sessions’ work (Claude, Docker, Chrome; Get-Counter 100% average), far over the record’s 0.65 limit, so these runs stand in for the forced load of triage step 2. Before: the file alone, five runs, failed the row once with the reported error (not.toBeInTheDocument() at line 208, 3,089 ms). Holding the slow step (L-107): a probe copy of the file whose /leads route suspends 1.5 s on a timed promise, which the router’s transition keeps off screen, failed the original row 3 of 3 at the same assertion and passed the fixed row 3 of 3; the probe was not committed. After: the file alone passed 5 of 5, the row in 3.6–4.9 s. The ten-in-ten full run was not run here, because the sessions sharing this machine were told not to run full suites; the PR’s test record and the nightly run take it. The “while the panel is pending” row in the same file also failed twice in the five before-runs under this load (a 20 s timeout and its 10 s lazy-panel wait) and passed after; it is not this bug, and a recurrence goes through the weekly flaky report.
Lesson: L-256 (new) — a wait on what a navigation renders needs the page budget; the default second measures the machine.
BUG-328 — The New template test asserts the create’s arrival state after the template page has cleared it
Affects: F20-d (forge-et-fe src/features/email/templates/EmailTemplateEditorRoute.test.tsx, “asks a settings.email_templates.manage holder Shared or Personal …”) · Found in: the test record of forge-et-fe#385 after main d5866286 was merged in, 2026-11-03; it also fails on forge-et-fe#381’s head 0d484040 and on main · Severity: medium — the nightly CRM Vitest suite fails; the product is right · Phase: 1
Repro: pnpm exec vitest run src/features/email/templates/EmailTemplateEditorRoute.test.tsx -t "asks a settings.email_templates.manage holder" on main d5866286: expected null to deeply equal { createdTemplate: true }, every run.
Expected: the test checks what a create’s arrival does: the new template’s page takes focus on its h1.
Actual: TemplateView reads the arrival state and clears it (navigate(…, { replace: true, state: null }), F20-d), so a reload does not announce the create again; the test read the state after that.
Fix: wait for the page’s h1 to have focus, then assert the state is cleared.
Regression test: that test, five passes in five.
Resolved: 2026-11-03 — fixed in forge-et-fe#385 (bc01c84d). forge-et-fe#381’s test record passed on 0d484040 (6,992 selected tests, 0 failed); on this macOS machine the same test fails every run on that commit, so whether the page clears the state before the test reads it depends on the machine. Fails before: every run on d5866286 and 0d484040; after: 46 passed five times.
Lesson: none — one-off: a test that reads a value its component clears in an effect races that effect; assert what the value causes instead.
BUG-327 — The F43 busy spec for the lead’s Compose email clicks a button on a tab it never opens
Affects: F43 (forge-et-fe e2e/browser/leads/follow-up-busy.spec.ts, “in the lead’s Compose email … 429 AI_BUSY”) · Found in: a run of the lead browser specs on main 65705a8d with forge-et-be c052a30b, 2026-11-03 (two in two fail) · Severity: medium — the nightly CRM gate fails; the product is right · Phase: 1
Repro: pnpm exec playwright test e2e/browser/leads/follow-up-busy.spec.ts --grep "Compose email" --project=chromium → locator.click: … email-lead-compose … element is not visible.
Expected: the spec opens the lead’s Notes and email tab, where the email panel and its Compose email are since the UI revamp (F11-f Rev 5), as compose-entry-points.spec.ts does.
Actual: it clicks Compose email on the lead’s first tab, where the panel is hidden.
Fix: showLeadTab(page, 'notes') before the click.
Regression test: that test, passing against c052a30b.
Resolved: 2026-11-03 — fixed in forge-et-fe#385: the spec opens the panel with showLeadCompose, which selects the tab and waits for the compose place, as its sibling specs do. Fails before: two in two on main 65705a8d; after: 12 passed for the spec file.
Lesson: none — one-off; BUG-323 is the same shape for other panels the revamp moved.
BUG-326 — account-briefing-busy.spec.ts imports a helper another PR removed, so pnpm typecheck fails on main
Affects: F43 (forge-et-fe e2e/browser/accounts/account-briefing-busy.spec.ts) · Found in: pnpm typecheck on main 65705a8d, 2026-11-03, while running the dialog specs for BUG-309 · Severity: high — main fails the typecheck, so the nightly CRM static checks fail and the spec file does not load in the browser gate · Phase: 1
Repro: pnpm typecheck on main → TS2305: Module './briefing-support' has no exported member 'editAccountButton'.
Expected: main typechecks.
Actual: 9a0c39eb (the clean-up after forge-et-fe#371’s review) removed editAccountButton from briefing-support.ts, which nothing used then; forge-et-fe#384 (F43 wave 2) merged afterwards with a spec that imports it. Each PR passed on its own; together they break — the case git-conventions.md § Branching leaves to the nightly run.
Fix: read the header’s Edit account by its test id (account-detail-edit) in the spec.
Regression test: pnpm typecheck, exit 0, and the spec file loads in the accounts folder’s browser run.
Resolved: 2026-11-03 — fixed on main first: forge-et-fe#381 and forge-et-fe#353 each restored editAccountButton in briefing-support.ts (95e7613f, 579c5d57), and together defined it twice, which failed the typecheck again (TS2393); forge-et-fe#385 removes the second copy and keeps the spec as main has it. Fails before: TS2305 on 65705a8d, then TS2393 on d5866286; after: pnpm typecheck exit 0 on ced9c024. forge-et-fe#389, titled fix(BUG-328) for a duplicate of this bug that was dropped before merge (the number BUG-328 went to the New template test), made the same deletion and merged 90 seconds after #385 (0c1118aa), so it changed nothing on main.
Lesson: none — one-off: the nightly static checks are the guard for two PRs that break each other.
BUG-325 — Flaky (macOS): the Edit user dialog’s keyboard walkthrough starts wherever Radix’s focus trap last saw focus
Affects: F5-c (forge-et-fe src/features/users/UserEditorRoute.test.tsx, “tabs through the Edit user dialog in reading order”) · Found in: the nightly full run’s CRM Vitest on main 4f9f1252, 2026-11-03 02:37 (macOS); 1 of 8 runs alone on that commit · Severity: low — a test defect; the dialog’s order is right · Phase: 1
Repro: for i in $(seq 8); do pnpm exec vitest run src/features/users/UserEditorRoute.test.tsx -t "tabs through"; done on 4f9f1252: one fails.
Expected: the walk reads the ✕, the fields, Cancel and Save every time.
Actual: sometimes it starts at Display name and ends on the ✕: the test blurred the focused element and pressed Tab, so the walk began wherever Radix’s focus trap had last recorded focus, which depends on when the dialog’s first focus landed.
Fix: start the walk on the ✕ by focusing it, and check that Tab from Save wraps back to it.
Regression test: that test, ten passes in ten.
Resolved: 2026-11-03 — fixed in forge-et-fe#385. Fails before: 1 of 8 runs on 4f9f1252; after: ten in ten.
Lesson: none — the same shape as the New account walkthrough, which already starts on its ✕.
BUG-324 — Four CRM browser specs reject AD #41’s Settings keys: a key with a dotted resource, and a role’s keys returned in ascending order
Affects: F5-a, F5-d (forge-et-fe e2e/seam/auth.seam.spec.ts, e2e/browser/roles/support.ts) · Found in: the CRM browser gate of the nightly full run on main 612d3739 with forge-et-be bb1a2a8b, 2026-11-03 (forge-et-fe#370) · Severity: medium — the nightly CRM gate fails; the product is right · Phase: 1
Repro: pnpm exec playwright test e2e/seam/auth.seam.spec.ts e2e/browser/roles/self-admin.spec.ts e2e/browser/roles/self-admin-next-request.spec.ts --project=seam --project=chromium against a backend at bb1a2a8b: four fail.
Expected: a /me key matches the catalog’s rule, where the resource may be dotted (forge-et-be PermissionCatalog KEY_PATTERN), and a role read’s keys are compared ascending (RoleDetailResponse).
Actual: forge-et-be#342 (AD #41) gave each Settings page a key of three parts (settings.account_tags.manage), and the seam test’s pattern allowed exactly one dot; the self-admin roles, now seeded with six keys in catalog order, were compared in that order with a read that returns them ascending.
Fix: the seam test uses the backend’s pattern; the roles’ reset compares the seeded keys sorted.
Regression test: those four tests, passing against bb1a2a8b.
Resolved: 2026-11-03 — fixed in forge-et-fe#385. Fails before: the four fail against bb1a2a8b; after: 16 passed for the three spec files.
Lesson: L-200’s check before merge covers only a backend PR that turns a refusal into an acceptance; a backend PR that changes what a response holds (a key’s form, a list’s order) breaks the frontend’s specs the same way. Whether to widen the check is in owed.md (“Check a backend response change against the frontend’s specs before merge”).
BUG-318 — The email templates list’s refresh-failed notice gives no as-of time for the rows it keeps
Affects: F20-d (forge-et-fe src/features/email/templates/EmailTemplatesScreen.tsx, REFRESH_FAILED and the stale render) · Found in: frontend-reviewer’s review of forge-et-fe#375 (BUG-281), 2026-11-02, from reading the code · Severity: low — the kept rows are shown undated · Phase: 1
Repro: open Settings › Email templates, make its re-read fail (switch tab or page while the backend refuses), and read the notice above the rows.
Expected: coding-standards §3.2 and §3.3: rows served from a cache carry their as-of time, as the Users, Leads, Accounts and Contacts lists say (“Showing users as of 14:32”).
Actual: “We couldn’t refresh the templates. Showing the last list read. Nothing was lost. Try again” — no time.
Fix: build the message from the snapshot’s dataUpdatedAt, with the same formatter the Users list uses (UsersScreen.tsx formatAsOf), and update the tests that read the text.
Regression test: an EmailTemplatesScreen test whose stale notice names the read’s time.
Resolved: 2026-11-03 — fixed with BUG-317 on forge-et-fe fix/BUG-317-refresh-notices (51fd53e9): the notice reads “We couldn’t refresh the templates. Showing the last list read, as of 14:32. Nothing was lost. Try again”, from the kept snapshot’s dataUpdatedAt (templatesQueries.ts stale.asOf). The time comes from src/lib/format/businessDate.ts formatColomboTime, which Leads, Accounts, Contacts and Roles use, rather than UsersScreen.tsx formatAsOf, so the templates feature does not import from the Users feature; “the last list read” stays because the kept rows can answer another tab, search or page. Fails before: the two BUG-318 tests fail on main 45585e24; after: both pass.
Lesson: none — one-off: the notice was written before the other lists’ as-of wording, and no other refresh-failed notice in src/features lacks a time (git grep -n "Showing" -- 'src/features/*.tsx' | grep -v "as of" lists only page-range labels, on 51fd53e9).
BUG-317 — On the Users and email templates lists, a successful Retry on the refresh-failed notice removes the button and drops focus to the page
Affects: F5-c (forge-et-fe src/features/users/UsersScreen.tsx, users-stale-retry), F20-d (src/features/email/templates/EmailTemplatesScreen.tsx, email-templates-stale-retry) · Found in: frontend-reviewer’s review of forge-et-fe#375 (BUG-281), 2026-11-02, from reading the code · Severity: low — a keyboard user loses their place after a successful retry · Phase: 1
Repro: on either list, make a re-read fail so the refresh-failed notice shows over the kept rows, let the backend answer again, Tab to Retry and press it.
Expected: focus moves to a stable place once the notice and its Retry go — the page’s h1 — as on the Leads, Accounts and Contacts lists (useFocusAfterWrite, BUG-262), and as forge-et-fe#375 does for Roles.
Actual: the notice and Retry unmount when the re-read succeeds and nothing moves focus, so it falls to <body>.
Fix: useFocusAfterWrite(regionRef, titleRef, renderedAt) on each list, as ContactsScreen.tsx does: notePress in Retry’s onClick, arm(...) once the re-read settles; mark Retry busy while it re-reads.
Regression test: per list, toHaveFocus() on the h1 after a successful retry, focus left alone when moved elsewhere during the wait, and focus kept on Retry when the retry fails again.
Resolved: 2026-11-03 — fixed in forge-et-fe#406 (51fd53e9; the review’s fixes are in 0ccca3e4: the useUsersList comment and the anchor of the L-183 templates test). useUsersList and useTemplatesList hold a stale failure through its Retry (useRetryHold; Users had none, so its notice left at the press and came back with a new Retry on a second failure) and return renderedAt and a refetch that resolves with the read’s update time; each screen calls useFocusAfterWrite(screenRef, titleRef, renderedAt), notePress in the stale Retry’s onClick and arm(...) once the re-read settles, and marks that Retry busy. Tests: UsersScreen.test.tsx and EmailTemplatesScreen.test.tsx “focus after the refresh-failed Retry (BUG-317)” (keyboard press, a click that focuses nothing, focus moved during the read; Users also a second failure, which the templates list’s existing L-183 test covers), and usersQueries.test.tsx’s two retry-from-stale tests. Fails before: 7 of the BUG-317 tests fail on main 45585e24 (the two “focus moved during the read” guards pass before and after); after: 4 passed (4) files, 124 passed (124) tests for the four files.
Lesson: L-133 — a control removed by its own press’s result with no focus target named, the case L-133 records; BUG-317 is added to its sources.
BUG-316 — A Retry on the Roles list’s first-load error panel removes the button at the press and drops focus to the page
Affects: F5-d (forge-et-fe src/features/roles/RolesRoute.tsx, roles-list-retry) · Found in: frontend-reviewer’s review of forge-et-fe#375 (BUG-281), 2026-11-02, from reading the code; not yet run · Severity: low — a keyboard user loses their place when retrying a failed first load · Phase: 1
Repro: open Settings › Roles while the roles read fails, so the error panel shows with no rows; Tab to Retry and press it.
Expected: Retry stays, busy, until the read answers; on a second failure focus is still on Retry (lesson L-183).
Actual (from the code): the panel is rendered from status === 'error' alone, and with no cached rows TanStack v5 returns the query to pending at the press, so the panel and Retry unmount and focus falls to <body>.
Fix: hold the failure through the retry with src/lib/hooks/useRetryHold.ts, as the other lists’ first-load errors do, and mark Retry busy.
Regression test: a RolesRoute test where a held retry keeps Retry mounted, busy and focused, and a retry that fails again leaves focus on it.
Resolved: 2026-11-03 — fixed in forge-et-fe#405 (9d94ae98, 3024241e): RolesRoute holds the first-load failure with useRetryHold from the press until the read answers, marks Retry busy, and gives both Retries the stale Retry’s useFocusAfterWrite repair, so a good answer moves focus to the h1 (L-183’s third rule). RolesRoute.test.tsx’s two BUG-316 tests failed on 45585e24 (2 failed: Retry never busy, as it unmounts) and the file passes with the fix (34 passed). In a running browser (Chromium, the CRM against a local backend, the roles read failed and the re-read held by a patched fetch): after Enter on Retry the same element stayed, aria-busy="true" and focused, with no skeleton; a failed re-read left it focused and not busy; a good one showed the rows and focus on the h1. The review of forge-et-fe#405 added a test for a press that focused nothing followed by a failed re-read (35 tests in the file).
Lesson: L-183 — another first-load error panel rendered from status === 'error' alone; added to L-183’s sources.
BUG-315 — Enter in the activity dialog’s record search with no option highlighted submits the form, saving the activity
Affects: F3-a (forge-et-fe src/features/activities/RecordPicker.tsx ~131–136, the RecordField inside FollowUpDialog.tsx’s <form>) · Found in: the delta review of forge-et-fe#374 (F1-f Rev 7, New contact), by reading the code, 2026-11-02 · Severity: medium — an activity can be saved by an Enter meant for the search · Phase: 1
Repro: not reproduced. Open the activity dialog, type in the record search so no option is highlighted, press Enter. RecordPicker’s Enter case calls preventDefault() only when an option is active, so the keydown reaches the form and submits it.
Expected: Enter in the search box picks the highlighted option, or does nothing; it never submits the dialog (lesson L-198).
Actual: with no option highlighted, Enter submits the activity form.
Fix: cancel Enter always, before any early return (a blocked or read-only state included), and pick only when an option is active — as forge-et-fe#374 did for AttachAccountPicker (commits d3c8fe63, 6d5dd5c5).
Regression test: a component test that fireEvent.keyDown(<record search>, { key: 'Enter' }) with no active option returns false, and a browser row that no activity is saved.
Resolved: 2026-11-03 — fixed in forge-et-fe#390: RecordPicker cancels Enter before its blocked return and picks only when an option is active, as AttachAccountPicker does. RecordPicker.test.tsx’s two BUG-315 tests (Enter with nothing highlighted inside a <form>, and Enter while blocked) failed without the fix and pass with it. No browser row was added: the cancelled keydown is what stops the submission (L-198’s test rule).
Lesson: L-198 — another text box inside a form that cancelled Enter only when it picked something; added to L-198’s sources.
BUG-323 — Browser specs that open a lead or an account wait for panels the UI revamp hides (the notes list, the account’s Buying standards panel), so they fail on main
Affects: forge-et-fe e2e/browser/leads/ helpers that settle a lead’s page on its notes list (settleLeadDetail, through openLead) and e2e/browser/accounts/detail-support.ts settleStandards (:1708, through settleAccountPage); every spec on them: F11-h’s follow-up specs, F13’s account-briefing specs, F20-c’s compose specs, F26-d’s price-email specs, and the size-matrix states leads-lead-draft* and account-detail-briefing* · Found in: F43 wave 2’s baseline run (O0), 2026-11-02, on forge-et-fe f0c21837 with forge-et-be b3c5c374; Dinil saw the price-email half the same day (BUG-224’s “Blocked 2026-11-02” line) · Severity: medium — the browser suites of four features cannot pass on main, so neither a change to them nor a new spec on the same helpers can be proved in the browser · Phase: 1
Repro: on an isolated stack, pnpm exec playwright test --project=chromium e2e/browser/leads/follow-up- e2e/browser/accounts/account-briefing e2e/browser/email/compose- and the size-matrix states above at 1280×800.
Expected: each spec opens its lead or account and reaches the panel it tests.
Actual: 80 failures in F43’s O0, every one a toBeVisible() that times out on an element that is present but hidden: getByTestId('account-standards') (52) or leads-notes-list / -empty / -error (42). The UI revamp (forge-et-fe#364) moved both panels out of the first view, and the helpers still wait for them.
Fix: test-only, in progress: the open draft forge-et-fe#371 (Nanndish) updates the browser specs for the revamp, these helpers included. F43-WI-2.3’s B3–B5 and its leads-lead-draft-ai-busy and account-detail-briefing-ai-busy page states use the same helpers, so #371’s change covers them too.
Regression test: after #371 merges, the repro above and F43-WI-2.3’s three busy specs and four -ai-busy page states are green on main (the nightly full run, or a run by hand).
Resolved: 2026-11-04 — fixed in forge-et-fe#371 (merged 2026-11-03, 23014f39), which updates the helpers for the revamp. Proved by the nightly full run of 2026-11-04 on forge-et-fe 4421bd61 with forge-et-be d3df4fc2 (the harness issue labelled nightly-log): the CRM browser gate passed with no failure outside its 14 known-failure marks, none of them this bug’s. The follow-up, account-briefing, compose and price-email specs passed, among them follow-up-busy.spec.ts and account-briefing-busy.spec.ts, as did the leads-lead-draft* and account-detail-briefing* size-matrix states at every size.
Lesson: none — one-off: the revamp (forge-et-fe#364) moved the two panels and its spec updates were made in their own PR, #371, by decision; since PRD Rev 81 the nightly full run is the guard for a UI change that leaves browser specs behind.
BUG-321 — The Access dialog keeps no list with an out-of-date mark and Retry when a read fails
Affects: F5-g (forge-et-fe src/features/sharing/AddColleagues.tsx, SharingPanel.tsx) · Found in: reading the product against F5-g FR-D2 and FR-P5 for the review of forge-et-fe#371, main ad7c6e63, 2026-11-02 · Severity: medium — a failed read shows old rows as current · Phase: 1
Repro: open a record’s Access dialog as a holder of <record>.manage, search Add colleagues, then make the next candidate read fail (or a re-read of the shares).
Expected: FR-D2’s error with the list still shown: the last candidate rows stay, marked out of date, with Retry. FR-P5 asks the same of People with access inside the dialog.
Actual: AddColleagues.tsx:116-117 sets readFailure only when there is no data or a Retry is held, so a failed re-read keeps the previous rows (sharingQueries.ts:66-69) with no mark and no Retry. In SharingPanel.tsx:694-695 the dialog draws a stale shares read as its plain people list; the stale band (failurePanel(current, true), :663) shows only on the card behind the dialog.
Fix: draw the out-of-date mark and Retry over the kept rows in both lists when a read with data fails.
Regression test: a size-matrix state for each list’s stale state (leads-lead-share-dialog-stale was the earlier one), plus a component test in AddColleagues.test.tsx.
Resolved: 2026-11-03 — fixed in forge-et-fe#408 (c795f1d4; review fixes aa032f30, 073c7aa6). Add colleagues keeps the rows of a failed re-read under a band that says they are out of date and from what time (The list of colleagues could not be refreshed. Showing it as of HH:MM. Nothing was lost.), with Retry and the trace line (sharing-dialog-candidates-stale); a pick and Share still work over them. People with access draws its own stale band over the kept rows (sharing-dialog-people-stale), with test ids apart from the card’s, and its first-read error moves to sharing-dialog-people-* for the same reason. A Retry that answers puts focus on the search or the dialog’s title (useFocusAfterWrite); the hook now counts focus on a dialog that holds its region as nowhere, because Radix’s focus trap puts focus there when the focused control leaves, so the repair never ran inside a dialog. New tests: AddColleagues.test.tsx (three), four SharingPanel.test.tsx cases (each band’s out-of-date state, its Retry focus, and a Retry that fails again), and src/lib/hooks/useFocusAfterWrite.test.tsx (the hook inside a dialog, beside one, and after a press that focused nothing). The four written first failed before the fix and pass after it; in the review round, three hook tests failed on the earlier hook and pass on the fix. The new size-matrix states account-detail-sharing-dialog-stale and leads-lead-share-dialog-stale were run once in Chromium at 1280 and 768 px, with every API call answered in the browser, in a probe that was not committed and that left out the size matrix’s page check: both bands sit inside the dialog with no horizontal overflow, the rows stay, and after Retry focus is on the title or the search. The same probe over main’s code shows the three rows with no band, and both arrangements fail on the band. The nightly full run is their first run with the full page check.
Lesson: L-133 — another instance, for the focus half: a Retry that answered left with its band, and inside a Radix dialog the lost focus lands on the dialog’s content, not <body>, so useFocusAfterWrite did not move it (L-133’s Update of 2026-11-03). The missing band itself is one-off: the UI revamp moved the list into the dialog without its stale state, and the colleagues list’s stale state was never built.
BUG-320 — In the Edit role dialog no access control is described by its permission’s catalog description
Affects: F5-d (forge-et-fe src/features/roles/RoleAccessEditor.tsx) · Found in: e2e/browser/roles/editor.spec.ts “AC-18: in the Edit role dialog the Leads choice is described…”, Chromium, on main ad7c6e63 with the specs updated, 2026-11-02 · Severity: medium — the descriptions are visible text with no programmatic tie to their control (WCAG 2.1 SC 1.3.1) · Phase: 1
Repro: pnpm exec playwright test e2e/browser/roles/editor.spec.ts -g "Leads choice is described" --project=chromium with the mark removed.
Expected: each radio of an entity’s access group, and each Allowed tick, has its catalog description as its accessible description, as the editor before F5-d Rev 8 did (AC-18: “each other key is a checkbox with its label and description”).
Actual: the Leads None radio’s description is empty. LevelChoice sets no aria-describedby; AllowedTick sets it only to the blocked reason (:178); AllRecordsBox has only aria-label (:145). The description is row text only.
Fix: point each control’s aria-describedby at its row’s description (and, for a blocked tick, at the reason too).
Regression test: that browser test, unmarked. The unmarked AC-18 test in the same file pins a blocked tick’s description to the reason exactly; once the fix adds the catalog description, that check becomes the pattern the marked test uses (the reason as a part of the description).
Resolved: 2026-11-03 — fixed in forge-et-fe#390: AccessMatrix gives each row’s description an id and passes it to renderAccess; the role editor points each level radio’s aria-describedby at it, and each single-key Allowed tick’s at it too, with a blocked tick’s reason after it. Two boxes are left as they were: Allowed: Settings, whose row text is a computed summary of the pages held, not a catalog description, and All records, which the row’s level description does not describe. RoleEditorRoute.test.tsx’s three description assertions failed without the fix and pass with it. editor.spec.ts’s AC-18 description test is unmarked, and the other AC-18 test now expects the reason at the end of the description. Those browser tests were not run before the PR (another session’s dev server held port 5173); the nightly full run is their first run.
Lesson: none — one-off: the access-table editor (F5-d Rev 8) drew the descriptions as row text and did not carry over the earlier editor’s aria-describedby; the unit tests now assert the description on each control.
BUG-319 — The user page shows the shares line for a role that grants nothing
Affects: F5-e (forge-et-fe src/features/users/EffectivePermissionsSection.tsx, the record layout) · Found in: e2e/browser/effective-permissions/reach.spec.ts “AC-19: E — no shares line for a role that grants nothing”, Chromium, on main ad7c6e63 with the specs updated, 2026-11-02 · Severity: low — one line of text the role makes meaningless · Phase: 1
Repro: pnpm exec playwright test e2e/browser/effective-permissions/reach.spec.ts -g "no shares line" --project=chromium with the mark removed.
Expected: no shares line (F5-e FR-15 lists what the section shows for a role that holds nothing; F5-e WI-1.2 plan Decision 11: with no view key a share gives nothing to see). The section layout does this, and its component test asserts it.
Actual: the record layout, used on the user page since the UI revamp, renders SharesLine in its side column whenever the read is ready, whatever the role holds.
Fix: render the side column’s SharesLine only when the answer holds a key, as the section layout does.
Regression test: that browser test, unmarked, and a record-layout case in EffectivePermissionsSection.test.tsx.
Resolved: 2026-11-03 — fixed in forge-et-fe#390: the record layout renders SharesLine only when the role holds a key, read from the same ready value the stacked layout uses. EffectivePermissionsSection.test.tsx’s record-layout AC-19 test failed without the fix and passes with it; reach.spec.ts’s AC-19 test is unmarked (first run: the nightly full run).
Lesson: none — one-off: the UI revamp’s second layout drew the shares line outside the check the first layout used; both now read one value.
BUG-313 — At 320px the user page under the Edit user dialog’s leave question overflows <main> by 2px
Affects: F5-c (forge-et-fe the user page at 320) · Found in: the size matrix’s user-editor-leave-dialog state at 320 × 640 on main ad7c6e63 with the specs updated, 2026-11-01 · Severity: medium — 320px reflow (WCAG 2.1 SC 1.4.10) · Phase: 1
Repro: pnpm exec playwright test e2e/browser/size-matrix --grep user-editor-leave-dialog --repeat-each=3 --project=size-320x640-chromium with the mark removed: 3 of 3 fail.
Expected: no horizontal scroll on the page or in <main>.
Actual: “<main> overflows by 2px” (with BUG-309’s axe finding on the dialog body).
Fix: find the 2px-wide element on the page under the dialog (the page check’s overflow detail) and let it wrap.
Regression test: the size-matrix state, unmarked.
Resolved: 2026-11-03 — no longer reproduces on main 612d3739 (the nightly full run of that day reported the mark as passing, “BUG-309, BUG-313 no longer fails these here”), and the state, unmarked, passed three in three at 320, 768, 1024 and 1280 on 65705a8d; no commit has been traced as the fix. The mark is removed in forge-et-fe#385.
BUG-312 — At 320px the Edit role dialog’s body scrolls sideways by 34px
Affects: F5-d (forge-et-fe the Edit role dialog — RecordEditDialog’s body and roles-editor-name) · Found in: the size matrix’s role-editor, role-editor-refused, role-editor-stale and role-editor-leave-dialog states at 320 × 640 on main ad7c6e63 with the specs updated, 2026-11-01 · Severity: medium — 320px reflow (WCAG 2.1 SC 1.4.10) · Phase: 1
Repro: pnpm exec playwright test e2e/browser/size-matrix --grep role-editor-refused --repeat-each=3 --project=size-320x640-chromium with the mark removed: 3 of 3 fail.
Expected: no horizontal scroll at 320 outside the access table’s own scroll region.
Actual: “div scrolls sideways by 34px”, and roles-editor-name and roles-editor-description span x 51px to 321px. The sideways scroll shows in every run; the page check reports the two fields clipped only in some (in role-editor-stale, two runs of three; why is not traced), so the mark (ROLE_DIALOG_OVERFLOW in e2e/pageStates.ts) lists clipped-control as optional.
Fix: let the dialog’s form shrink to the viewport (a min-inline-size: 0 on the field column, or the Name field at 100% of its cell).
Regression test: those size-matrix states, unmarked.
Note (2026-11-03): since forge-et-fe main’s LOCK_SCROLL_BEHIND_QUESTION (merged into forge-et-fe#385 at ced9c024), the leave question locks the dialog body’s scrolling, so role-editor-leave-dialog no longer shows the overflow and is unmarked; the body is still 34px too wide under it. The regression test is the other three states (role-editor, role-editor-refused, role-editor-stale).
Resolved: 2026-11-03 — fixed on forge-et-fe branch fix/BUG-277-dialog-short-window (with BUG-277’s RecordEditDialog part). The cause: RoleEditorForm.module.css’s .fields grid gave each auto-fit column a fixed minimum of 0.75 × layout/list-pane-width (270px), wider than the dialog body’s 208px content box at 320px; the minimum is now capped at the body’s width (min(100%, …)), as every other auto-fit grid in the CRM already caps it. Checked in a running browser (headless Chromium, the live role editor) at 320 × 640: the body scrolls sideways by 0px (before: 34px), and Name and Description end inside the dialog; at 1280 × 800 the dialog is unchanged. The size-matrix mark ROLE_DIALOG_OVERFLOW is removed from role-editor, role-editor-refused and role-editor-stale, and the four role-editor dialog states pass the size matrix at 320 × 640 unmarked; and e2e/browser/shell/record-edit-dialog-short-viewport.spec.ts asserts the body does not scroll sideways at 320 × 640.
Lesson: none — one-off: every other repeat(auto-fit, minmax(…)) grid in forge-et-fe src already caps its column minimum with min(100%, …); this was the one that did not.
BUG-311 — At 320px the price import’s Choose file button reaches past the viewport
Affects: F26-c (forge-et-fe src/features/pricing/WorkbookStep.tsx, pricing-import-file-choose) · Found in: the size matrix’s pricing-import-check-failed and pricing-import-file-failed states at 320 × 640 on main ad7c6e63 with the specs updated, 2026-11-01 · Severity: medium — 320px reflow (WCAG 2.1 SC 1.4.10) · Phase: 1
Repro: pnpm exec playwright test e2e/browser/size-matrix --grep "pricing-import-(check|file)-failed" --project=size-320x640-chromium with the marks removed.
Expected: the page check is clean at 320.
Actual: “<main> overflows by 2px” and pricing-import-file-choose spans x 173px to 322.13px, outside the 320px viewport.
Fix: let the import step’s row wrap at 320 (the button onto its own line).
Regression test: those size-matrix states, unmarked.
Resolved: 2026-11-03 — fixed in forge-et-fe#390. The buttons already wrapped onto their own lines; at 320 × 640 the Workbook step leaves the failure box 93px for a 149px Choose another file whose label could not wrap. FileUpload’s failure actions (.actions > button) now wrap their labels, with block padding, and keep the Button’s own minimum height. Checked in a browser at 320 × 640 in the failed state: <main> and the page overflow by 0px and no control reaches past the viewport (before: 2px, and x 173 to 322.13px); at 1024px both buttons stay 40px tall on one row, as before. The size-matrix marks are removed (first run: the nightly full run).
Lesson: none — one-off: the only file zone placed in a column this narrow; the shared component now wraps any action label that does not fit.
BUG-310 — At 768px, with a lead row’s menu open, the leads list’s Est. value sort button fails axe target-size
Affects: F11-b (forge-et-fe the leads list header, leads-sort-est-value) · Found in: the size matrix’s leads-row-menu state at 768 × 1024, Chromium and WebKit, on main ad7c6e63 with the specs updated, 2026-11-01 · Severity: low — axe target-size (serious) on one header control in one state · Phase: 1
Repro: pnpm exec playwright test e2e/browser/size-matrix --grep leads-row-menu --project=size-768x1024-chromium with the mark removed.
Expected: axe is clean.
Actual: target-size on leads-sort-est-value. Not traced why only with the row menu open; the menu’s popover may sit close enough to the button to fail axe’s spacing rule.
Cause (found 2026-11-03): not the menu. axe’s finding is partiallyObscured (22.7 × 24px) plus target-offset 20.4px, and the related node is the pinned Name header’s sort button (leads-sort-name), not the menu. At 768 × 1024 the table’s scroll box (leads-table-scroll) is 654px wide and its content 1,242px, so the ⋯ column is off screen. The state clicks the pinned lead’s ⋯, Playwright scrolls it into view, and the box ends at scrollLeft 588, its end. There the Est. value header button (x 92.5–205.2) sits under the pinned Name header (button x 104–182.5, cell to 191.5), and only 22.7px of it is clear of the Name button. Measured on forge-et-fe 45585e24 with every API read pinned in the browser (Chromium and WebKit, the same numbers). With the menu closed, the table scrolled to 588 fails the same way, and at 540, 500, 450 and 400 it passes. So any user who scrolls this table to its end at 768px gets the same layout. It is the BUG-232 pattern (a control scrolled under the pinned first column) on a header at 768px, where design-system §9 pins the first column. Which header lands under the pinned column at the end depends on the column widths, so other pinned tables can do the same.
Fix: a layout choice. The options were: (a) pin the ⋯ column at the inline end from bp/tablet up as well, so reaching a row’s menu never scrolls the table, across every table with a ⋯ column; (b) record that a header partly scrolled under the pinned column is accepted at the end of the scroll, as a reasoned, permanent page-check exemption (design-system §9); © add space at the end of the scroll so the last position clears the pinned column, which depends on this table’s column widths. Changing the test so the table does not scroll would hide the finding.
Decision: Nanndish’s, 2026-11-04: option (a), pin the ⋯ column (design-system Rev 94). Shown the scroll sweep below, he kept it on 2026-11-05 and asked for the remaining scroll positions to be filed as their own bug (BUG-345).
Regression test: the size-matrix state, unmarked; and a scan in src/ui/pinnedFirstColumn.test.tsx that fails any table pinning its first column whose row ⋯ column is not pinned.
Resolved: 2026-11-05 — fixed in forge-et-fe#415 (eba3984b; review fixes da262da3, 6376378e, 52cd0226, 3eb61225). src/ui/pinnedFirstColumn.ts adds PINNED_END_CELL and PINNED_END_CORNER, and src/ui/dataTable.ts adds DATA_TABLE_MENU_CELL and DATA_TABLE_HEAD_MENU, which pin the ⋯ column at the inline end while the table pins its first column (bp/tablet up). The nine tables with a row ⋯ menu use them: Accounts, Contacts, Leads, Users, Roles, Email templates, Tag settings, the price round table and the Standards table. The auction-sales and not-in-round tables, whose last column holds plain buttons, are unchanged. The size-matrix mark on leads-row-menu is removed. Fails before: with the mark removed and the source on main, leads-row-menu at size-768x1024-chromium and -webkit failed 2 of 2 (axe target-size on leads-sort-est-value); the new scan failed for all nine tables. After: the state passes in both engines (12 passed, setup included), against a backend from forge-et-be main on an isolated stack; the scan passes (10 of 10); the nine tables’ component tests pass (35 files, 767 tests); tsc -b is clean. In a browser at 768 × 1024 with every API read pinned, opening the pinned lead’s menu leaves the table at scrollLeft 6 (588 before) and axe clean, the ⋯ cells and header are sticky, and at 320 × 640 nothing is pinned, as before. A sweep of the table’s scroll in 20px steps still finds target-size at other positions, three of them where a tag’s +N button passes under the pinned ⋯ button: BUG-345. The review of forge-et-fe#415 added: the ⋯ cell paints the row’s open-menu tint over its fill (a browser test, failing before); a rendered check of the Standards table’s pinned header; and a Tab test showing a focused control lands clear of the pinned ⋯ cell in Chromium and WebKit. The review also asserts the pinned ⋯ cell keeps its opaque fill, and exports PINNED_END_CELL and PINNED_END_CORNER from the ui barrel. The pinned first column’s own missing tint is BUG-348.
Lesson: L-219 — another instance: a control partly scrolled under a pinned column, here at bp/tablet on a header; L-219 gains an update.
BUG-309 — At 320px the scrolling body of RecordEditDialog is not keyboard-focusable
Affects: F1-f, F5-c, F5-d (forge-et-fe RecordEditDialog’s body, shared by the Edit contact, Edit user and Edit role dialogs) · Found in: the size matrix’s contacts-edit-leave, contacts-edit-leave-saving, user-editor-leave-dialog, user-editor-role-open and role-editor-leave-dialog states at 320 × 640 on main ad7c6e63 with the specs updated, 2026-11-01 · Severity: medium — WCAG 2.1 SC 2.1.1, axe scrollable-region-focusable (serious) · Phase: 1
Repro: pnpm exec playwright test e2e/browser/size-matrix --grep contacts-edit-leave --project=size-320x640-chromium with the marks removed.
Expected: axe is clean: a region that scrolls can be reached by the keyboard.
Actual: axe reports scrollable-region-focusable on the dialog’s <div class="_body_…">, which scrolls at 320 (BUG-123 is the same rule on another list).
Fix: give the scrolling body tabIndex={0} with an accessible name, as the other scroll regions have.
Regression test: those size-matrix states, unmarked.
Resolved: 2026-11-03 — fixed in forge-et-fe#385: the body is a <section> named “
tabIndex={0}, one Tab stop after the ✕ in every RecordEditDialog (Edit contact, Edit account, New account, Edit user, Edit role, the sharing dialog); the keyboard walkthroughs that pass through it name it. By the nightly run of 2026-11-03 the defect also showed at 1024 and 1280 (Chromium) and 1024 (WebKit) in role-editor-leave-dialog, where AD #41’s Settings rows made the Edit role dialog tall enough to scroll; the marks covered 320 only. Fails before: those states fail axe scrollable-region-focusable with the marks removed; after: the five states pass three in three at 320, 768, 1024 and 1280 (Chromium) and 1024 (WebKit).
BUG-308 — At 320px a contact name with a long unbroken word does not wrap, so the contact page under the Edit contact dialog overflows <main> by 1–13px
Affects: F1-f, F20-l (forge-et-fe the contact page’s Other emails) · Found in: e2e/browser/accounts/contact-email-addresses-a11y.spec.ts AC-20 at 320 × 640 on main ad7c6e63 with the specs updated, 2026-11-01 · Severity: low — an intermittent sub-pixel reflow (WCAG 2.1 SC 1.4.10) · Phase: 1
Repro: pnpm exec playwright test e2e/browser/accounts/contact-email-addresses-a11y.spec.ts --grep "320" --repeat-each=5 --project=chromium. On one Windows machine, 2 of 6 runs failed: “<main> overflows by 8px”, then “by 1px”; 4 passed. Seen again on 2026-11-03 in forge-et-fe#371’s browser gate on e4ac76e7 (Windows, “by 1px”), passing on its retry.
Expected: no horizontal scroll at 320 on the contact page with three other addresses, one very long, while the Edit contact dialog is open over it.
Actual: sometimes <main> is 1–8px wider than the viewport. The dialog renders outside <main>, so the overflow is in the contact page under it.
Fix: on a failing run, read which element is wider than <main> (the page check’s overflow detail), then let it wrap (overflow-wrap: anywhere on the long address is the likely one).
Regression test: the AC-20 320 test with --repeat-each=10, ten in ten.
Resolved: 2026-11-03 — fixed in forge-et-fe#385. Not a flake and not Windows-only: on macOS it failed 2 of 3 and 5 of 6 runs. The element is the contact’s name (ContactProfileCard .name), a flex item of its name line whose test value ends in a run token with no space; overflow-wrap: break-word does not lower a flex item’s minimum width, so the name never wrapped and its width, which depends on the token’s letters, decided each run (1–13px over). The name now has overflow-wrap: anywhere and min-inline-size: 0. Fails before: 5 of 6 runs; after: ten in ten (Chromium, macOS).
BUG-307 — The account timeline’s pressed All filter chip shows its count below the contrast floor
Affects: F1-c (forge-et-fe src/features/timeline/AccountTimelinePanel.module.css, .filterCount) · Found in: e2e/browser/accounts/account-timeline-a11y.spec.ts AC-17 on main ad7c6e63 with the specs updated, 2026-11-01 · Severity: medium — WCAG 2.1 SC 1.4.3, axe serious; F1-c NFR-2 and AC-17 · Phase: 1
Repro: open an account with timeline entries on Notes and timeline; run axe.
Expected: axe is clean with the panel filled.
Actual: color-contrast on the pressed All chip’s count (<span class="_filterCount_…">20</span>): .filterCount sets opacity: 0.7, and on the pressed chip’s primary fill that falls below 4.5:1.
Fix: drop the opacity on the pressed chip, or give the count a token that meets 4.5:1 on both fills.
Regression test: the AC-17 axe scans, unmarked.
Resolved: 2026-11-03 — fixed in forge-et-fe#390: the pressed chip draws its count at full opacity (.filter[aria-pressed='true'] .filterCount), told apart by weight. Measured in a browser: 5.91:1 on the primary fill (3.8:1 at 0.7). The AC-17 axe scan and the three timeline size-matrix states are unmarked (first run: the nightly full run).
Lesson: L-252 (new) — normal-size text drawn with a palette step or an opacity is measured against whatever theme it lands in; muted text takes --et-text-secondary. BUG-301 and BUG-298 are the same cause.
BUG-306 — Account notes show Imported for a note with no author, and head their section Notes
Affects: F1-e (forge-et-fe src/features/notesAndTags/NoteItem.tsx importedLabel, src/ui/notes/NoteEntry.tsx, RecordNotesSection.tsx) · Found in: e2e/browser/accounts/account-notes.spec.ts AC-22 on main ad7c6e63 with the specs updated, 2026-11-01 · Severity: low — shipped wording contradicts an approved spec · Phase: 1
Repro: open the seeded account with an authorless note, on Notes and timeline.
Expected: F1-e FR-P2 and AC-22: a note with no author shows —, and the section is headed Notes ({author ?? importedLabel}), and the heading renders Notes and the total in a separate span, reading Notes 0.
Fix: decided by Tharinda on 2026-11-08: the code follows F1-e. A note with no author shows — (no import writes account or contact notes, so Imported misleads), and the section is headed Notes (.forge/owed.md “Settle the small points left by the UI revamp preview”, item 1.
Regression test: the two AC-22 assertions, unmarked.
Resolved: 2026-11-08 — fixed in forge-et-fe#446, as Tharinda decided (the code follows F1-e; the spec is unchanged). A note with no author on an account or contact shows — as its author, its delete menu names “the note with no author”, and RecordNotesSection heads the section Notes (NoteEntry takes noAuthor: 'imported' | 'none', which picks the label and the avatar; the lead page keeps Imported and the glyph). No account or contact note lacks an author in normal use — every note records the signed-in person who wrote it — so this covers seeded data and any future note written by a job or an import. The crm-current prototype’s account and contact Notes headings now read Notes (n). Tharinda then decided that every card heading puts its count in parentheses; changing the Contacts and Follow-ups cards is an owed.md row. The two AC-22 marks in account-notes.spec.ts are removed; the component tests of the heading (RecordNotesSection, AccountNotesAndTagsPanels, ContactNotesAndTagsAction) and of the authorless note now expect the new wording and failed with the source change set aside. This settles .forge/owed.md “Settle the small points left by the UI revamp preview”, item 1.
Lesson: none — one-off: the same one-time restyle by the UI revamp preview as BUG-305, whose line records the judgement.
BUG-305 — The Mailbox Connection card’s name and connected date are not at type/meta in text/secondary
Affects: F20-a (forge-et-fe src/features/email/ConnectedPanel.module.css, .name and .fact dd) · Found in: e2e/browser/mailbox/connect-screen.spec.ts AC-19/AC-22a on main ad7c6e63 with the specs updated, 2026-11-01 · Severity: low — visual hierarchy against an approved criterion · Phase: 1
Repro: connect the mailbox (local stub) and open the Mailbox settings page.
Expected: F20-a AC-19 (kept by Rev 10): the display name and the connected date at type/meta in text/secondary.
Actual: the name uses --et-type-body-size (13px) and --et-text-primary; the date uses --et-type-secondary-size and --et-text-primary.
Fix: use the meta tokens, or revise AC-19 if the revamp meant the larger name.
Regression test: the AC-19 type-scale assertions, unmarked.
Resolved: 2026-11-03 — fixed in forge-et-fe#412 (6954c28; review fixes 22c37b46, cadcf37b). The code now matches AC-19; the spec is unchanged. The display name and the connected date use the stylesheet’s existing .meta rule (type/meta in --et-text-secondary), and .fact dd keeps the primary colour only for a value that is not a .meta line (.fact dd:not(.meta)), because on revamp routes the Status value’s StatusIndicator label inherits its type and colour from its <dd>, and AC-19 does not set them. The Connected on label and the address are unchanged. In the running page (revamp theme), --et-text-secondary (rgb(110, 95, 78)) measures 6.16:1 on the white card, above the 4.5:1 that normal-size text needs (L-252); the name and the date compute 11.5px, and the address stays 13px in --et-text-primary. connect-screen.spec.ts no longer marks the AC-19 test as a known failure: built from main’s styles it fails (the name computes 13px, not 11.5px), and on the first fix commit the whole spec passed in Chromium. mailboxTokens.test.ts binds the panel’s .meta rule, and MailboxRoute.test.tsx checks that the name and the date carry it.
Lesson: none — one-off: the UI revamp preview (forge-et-fe#364) restyled the Connection card without the AC-19 type clause, and the update of its browser specs caught it. BUG-300, BUG-304 and BUG-306 came from the same one-time restyle, and BUG-323’s line records the same judgement.
BUG-304 — The Mailbox page is capped at 900px, where design-system Rev 77 §6 keeps it at 720px
Affects: F20-f (forge-et-fe src/features/email/MailboxRoute.module.css, .screen) · Found in: e2e/browser/mailbox/inbox.spec.ts on main ad7c6e63 with the specs updated, 2026-11-01 · Severity: low — layout against the design system · Phase: 1
Repro: open the connected Mailbox page at 1440px.
Expected: design-system Rev 77 §6: the one-column rule stays for the Mailbox page, capped at twice layout/list-pane-width (720px) and centred.
Actual: max-inline-size: calc(2.5 * var(--et-layout-list-pane-width)), 900px.
Fix: use twice the width, or record the wider cap in design-system §6.
Regression test: the inbox.spec.ts column test, unmarked.
Resolved: 2026-11-04 — fixed in forge-et-fe#397 and this PR. The lead decided on 2026-11-03 to keep 900px: each Settings sections row holds its question beside its card, and twice leaves both too narrow. Design-system Rev 89 records the 2.5 × cap; the CSS is unchanged (comment only). expectCentredColumn takes the cap factor, and the Mailbox column test checks 2.5 ×, unmarked; it also checks that the column reaches its cap when the canvas is wider, so a cap of twice fails it (shown by setting the CSS to 2 × for one run).
Lesson: none — one-off: the code chose 2.5 the day before design-system Rev 77 restated twice for the page; the review found the test’s one-sided cap check would not have caught a return to twice, and the check now has both sides.
BUG-303 — The Weekly prices summary bar adds up its All count in the browser
Affects: F26-f (forge-et-fe src/features/pricing/RoundSummaryBar.tsx, chips()) · Found in: the update of main’s browser specs after forge-et-fe#364, 2026-11-01 · Severity: low — against forge-et-fe CLAUDE.md’s rule that the browser never recomputes a total; the figure is right while the three counts are complete · Phase: 1
Repro: read chips(): All is drafts + published + withdrawn.
Expected: the total comes from the server, as every other total does.
Actual: the browser sums it.
Fix: take All from the counts read, adding it to the backend’s answer if it is missing.
Regression test: a component test that All shows the server’s total when it differs from the sum.
Resolved: 2026-11-04 — fixed in forge-et-be#358 (d3df4fc2), forge-et-fe#398 and this PR (F26-f Rev 3). The round summary’s counts gain total, count(*) over the round’s rows; the bar shows it. Tests: RoundReadServiceSpec (the service passes the repository’s 11, not a sum of 10), AuctionSaleRoundReadIT AC-4 (the member list, and total equal to the unfiltered rows list’s totalItems), RoundSummaryBar.test.tsx (All 9 where the statuses sum to 5), roundWire.test.ts (an omitted or string total refused), and the browser AC-18, whose helper had also summed All.
Lesson: none — one-off: the rule that the browser never recomputes a total already covers it. The review found the browser helper computing the expected All the same way the bar did, so the browser suite could not have caught the bug; it now reads the server’s value.
BUG-302 — After the first tag is attached to a record with none, focus does not return to the Add tag chip
Affects: F1-e (forge-et-fe src/features/notesAndTags/RecordTagsSection.tsx, AddTagControl) · Found in: e2e/browser/accounts/account-tags.spec.ts AC-21 on main ad7c6e63 with the specs updated, 2026-11-01 · Severity: medium — keyboard focus lost after an action (WCAG 2.1 SC 2.4.3) · Phase: 1
Repro: on an account with no tags, open Add tag and pick a tag.
Expected: F1-e FR-P1 and AC-21 (Rev 9): the pick folds the search back to the chip with focus on it.
Actual: focus is lost. The empty state draws the chip inside account-tags-none’s list and the tags state inside TagList’s, so the attach answer remounts AddTagControl, whose focusChip starts unset. With tags already on the record the chip keeps focus.
Fix: keep one list element across the empty and tags states, or carry the focus request across the remount.
Regression test: the AC-21 first-pick focus test, unmarked.
Resolved: 2026-11-03 — fixed in forge-et-fe#409 (c2618ed, 1c5bbd0a): the empty state renders the same TagList as the tags state, with no tags, so the first tag keeps the list and the chip; the list’s test id is <prefix>-none with no tags and <prefix>-list after. The review found the same remount on a change to the stale state (a pick whose outcome is unknown, then a failed re-read), so the empty, tags and stale cases now return one shape and keep the list in all three. Component tests of the first pick and of the stale case each failed on the code before their fix (toHaveFocus on the chip) and pass. Checked in a browser on a new account: after the first pick the chip has focus, the tag list is the same <ul> node as the empty state’s, and the tag is announced. The AC-21 first-pick focus test is unmarked (first run: the nightly full run).
Lesson: L-133, another instance — a control that stays across two states but sits under a different parent in each is remounted and loses focus; BUG-014 had the same cause, so the lesson gains a bullet and frontend-reviewer dimension 18 a sentence.
BUG-301 — No phone number and No email address on the contact page fail axe colour contrast
Affects: F1-f (forge-et-fe src/features/accounts/ContactProfileCard.module.css .missing, color: var(--et-neutral-500)) · Found in: e2e/browser/accounts/contact-email-addresses-a11y.spec.ts AC-20 on main ad7c6e63 with the specs updated, 2026-11-01 · Severity: medium — WCAG 2.1 SC 1.4.3, axe serious; F1-f AC-13 and NFR-6 · Phase: 1
Repro: open a contact with no phone number; run axe.
Expected: axe is clean.
Actual: color-contrast on contact-page-phone-missing (“No phone number”). No email address uses the same .missing style; the test’s contact has an email, so only the phone text is scanned.
Fix: use text/secondary (or the muted token that meets 4.5:1) for .missing.
Regression test: the AC-20 axe test, unmarked.
Resolved: 2026-11-03 — fixed in forge-et-fe#390: .missing uses --et-text-secondary. Measured in a browser on a contact with no phone number: 6.16:1 on the card (4.09:1 with --et-neutral-500). The AC-20 axe test is unmarked (first run: the nightly full run).
Lesson: L-252 (new) — muted normal-size text took a palette step, which the revamped theme sets under 4.5:1; BUG-298 and BUG-307 are the same cause.
BUG-300 — The user page’s access table contradicts F5-e AC-14, AC-14a, FR-15 and AC-19
Affects: F5-e (forge-et-fe src/lib/permissionCatalog/accessMatrix.ts, src/features/users/AccessMatrix.tsx, EffectivePermissionsSection.tsx) · Found in: the update of main’s browser specs after forge-et-fe#364, on main ad7c6e63, 2026-11-01 · Severity: medium — the shipped screen does not meet four criteria of an approved spec · Phase: 1
Repro: open a user’s page. The leads row shows the group’s name, not the lead.manage label, and no “includes View”; a role holding only buying_standard.manage shows the group’s name too; a role that grants nothing still shows every group, greyed.
Expected: F5-e AC-14 and AC-14a: the held key’s label and “includes View”; FR-15 and AC-19: no group for a role that grants nothing.
Actual: accessMatrix.ts titles a row with the key’s label only when its group has more than one area (title: entities > 1 ? item.manage.label : null), AccessMatrix.tsx never renders “includes View”, and EffectivePermissionsSection.tsx renders the table and the Reporting line card whenever the read is ready. F5-e Rev 2’s FR-12 (“every area appears”) also conflicts with FR-15 inside the spec.
Fix: decided by Tharinda on 2026-11-08, in two parts. (1) A row keeps the area’s name with None, View and Manage and the held level marked, as the role editor shows it; the spec changes: F5-e AC-14 and AC-14a are reworded to the built rows (no key label, no “includes View” — Manage already covers View). (2) A role that grants nothing shows only FR-15’s sentence and no access table; the code changes (EffectivePermissionsSection.tsx renders no AccessMatrix for an empty role), and F5-e FR-12’s “every area appears” is limited to a role that grants something, so FR-12 and FR-15 agree.
Regression test: section.spec.ts AC-14, reach.spec.ts AC-14a and AC-19, unmarked.
Resolved: 2026-11-08 — fixed in forge-et-fe#446 and this PR (F5-e Rev 3), as Tharinda decided. The spec follows the built rows: F5-e AC-14 and AC-14a name each row as the role editor names it (its catalog group’s name, or its manage key’s label in a group with more than one entity) at its held level, with no key label and no “includes View”, FR-12’s table appears only for a role that holds at least one key, and AC-19 is unticked until the merged code re-proves it, so it agrees with FR-15. The code follows FR-15 and AC-19: EffectivePermissionsSection’s record layout renders no access table for a role that grants nothing, only the sentence. The reach.spec.ts AC-19 mark is removed; the two marked label tests (section.spec.ts AC-14, reach.spec.ts AC-14a) are deleted, because the unmarked AC-14 and “AC-16, AC-14a” tests already check the reworded criteria (one row per area, the held level, the reach, no row for an implied key). EffectivePermissionsSection.test.tsx gains a record-layout case with no table for an empty role, which failed with the source change set aside, and one with the table for a role that holds a key.
Lesson: none — one-off: the same one-time restyle by the UI revamp preview as BUG-305, whose line records the judgement.
BUG-299 — The role editor hides All records while None is chosen; F5-d AC-19 wants it shown aria-disabled with its reason
Affects: F5-d (forge-et-fe src/features/roles/RoleAccessEditor.tsx renderReach, src/lib/permissionCatalog/accessMatrix.ts) · Found in: the update of main’s browser specs after forge-et-fe#364, 2026-11-01 · Severity: low — an approved criterion not met; a decision is owed · Phase: 1
Repro: open a role’s Edit role dialog and set an entity to None.
Expected: F5-d AC-19: All records shown aria-disabled with the reason “Choose View or Manage first”.
Actual: the box is not rendered at None (row.reach is null at level === 'none').
Fix: the decision in .forge/owed.md (“Agree the UI revamp preview’s reversals of earlier decisions”, item 3): keep the hidden form and revise AC-19, or restore the shown-and-blocked form.
Regression test: editor.spec.ts’s two AC-19 tests and roles-a11y.spec.ts’s AC-19 tooltip tests, unmarked (or rewritten with AC-19 if the spec changes). The tooltip tests hold only the steps up to the tooltip while marked; the fix also puts back their AC-25 page check over the tooltip at each size, as the comment above the test describes (at 320, expectOnlyKnownPageFailures with ROLE_DIALOG_OVERFLOW for BUG-312).
Resolved: 2026-11-03 — closed as wontfix: overtaken by F5-d Rev 9 (forge-et-harness#1061), which keeps the hidden All records box and rewords AC-19 to it; the browser tests marked under this id are rewritten to the hidden form in forge-et-fe#371; Tharinda decided.
Lesson: none — one-off: the requirement was changed to the built form.
BUG-298 — On the lead page’s Customer card, the names of an account and contact the user cannot open fail colour contrast
Affects: F11-f (forge-et-fe the lead page’s Customer card, leads-customer-account, leads-customer-contact) · Found in: the size matrix’s leads-lead-unreachable-match state on main ad7c6e63, 2026-11-01 · Severity: medium — WCAG 2.1 SC 1.4.3, axe color-contrast serious · Phase: 1
Repro: pnpm exec playwright test e2e/browser/size-matrix --grep=leads-lead-unreachable-match --project=size-1280x800-chromium.
Expected: the page check is clean.
Actual: axe reports color-contrast on the spans “An account you cannot open” and “A contact you cannot open”.
Fix: give the muted text a token that meets 4.5:1 on the card’s background.
Regression test: the size-matrix state, unmarked.
Resolved: 2026-11-03 — fixed in forge-et-fe#390: the Customer card’s .muted uses --et-text-secondary, the change BUG-301 makes (6.16:1 on a white card, 4.09:1 before). The size-matrix state is unmarked (first run: the nightly full run).
Lesson: L-252 (new) — the same cause as BUG-301.
BUG-297 — The Auction sales table’s sale-round links are 15px tall, under the 24px target-size floor
Affects: F26-g (forge-et-fe src/features/pricing/AuctionSalesTable.tsx, pricing-sales-round-<id>) · Found in: the size matrix at 1280 on main ad7c6e63, 2026-11-01 · Severity: medium — the browser gate’s target-size floor (design-system §9) · Phase: 1
Repro: pnpm exec playwright test e2e/browser/size-matrix --grep=pricing-auction-sales --project=size-1280x800-chromium.
Expected: every link at least 24 × 24px (or the spacing exception).
Actual: each sale-round link is about 69 × 15px; it fails pricing-auction-sales, -stale, -sorted and the three pricing-edit-dates states.
Fix: give the link the table’s link-cell height, as the other tables’ row links have.
Regression test: those size-matrix states, unmarked.
Resolved: 2026-11-03 — fixed in forge-et-fe#390: .roundLink is an inline-flex box with a space/7 minimum height, as AccountStandardsPanel.module.css’s .manage is. Measured in a browser: 69.47 × 24px (69.47 × 15px before). The size-matrix marks are removed (first run: the nightly full run).
Lesson: L-148 — a link that is not inline in running text needs its own 24px target, the lesson’s second bullet; BUG-292 is the same. Added to L-148’s sources.
BUG-296 — The role page’s holders card says a role “can be deleted from its edit page”
Affects: F5-d (forge-et-fe src/features/roles/RoleHoldersCard.tsx) · Found in: the update of main’s browser specs after forge-et-fe#364, 2026-11-01 · Severity: low — wrong guidance text · Phase: 1
Repro: open a role that nobody holds.
Expected: F5-d Rev 8 FR-18: Delete role is on the role’s page, so the text points there.
Actual: the card says it “can be deleted from its edit page”.
Fix: reword the line.
Regression test: a component test of the card’s line.
Resolved: 2026-11-03 — fixed in forge-et-fe#390: the line reads “Nobody holds this role, so it can be deleted with Delete role on this page.” RoleEditorRoute.test.tsx’s BUG-296 test failed without the fix and passes with it.
Lesson: none — one-off: wording left from before F5-d Rev 8 moved Delete role to the role’s page.
BUG-295 — The Edit user dialog’s manager list and the Access dialog’s Add colleagues list are announced as “Tags”
Affects: F5-c, F5-g (forge-et-fe src/ui/tags/AddTagPicker.tsx, the SearchSelect that reuses it for Manager, and the sharing Access dialog’s Add colleagues search) · Found in: the update of main’s browser specs after forge-et-fe#364, 2026-11-01 · Severity: low — a wrong accessible name · Phase: 1
Repro: open Edit user, open Manager; or open a record’s Access dialog and search Add colleagues. A screen reader announces either list as “Tags”.
Expected: each list is named for its field.
Actual: AddTagPicker.tsx hard-codes aria-label="Tags" on its listbox, and both reuse it.
Fix: take the listbox’s name from the caller.
Regression test: a component test of SearchSelect’s listbox name.
Resolved: 2026-11-03 — fixed in forge-et-fe#390: AddTagPicker takes the listbox’s name from a listLabel prop, Tags when not given; SearchSelect requires one (the Edit user dialog passes Manager), and Add colleagues passes Colleagues. The BUG-295 tests in SearchSelect.test.tsx and SharingPanel.test.tsx failed without the fix and pass with it.
Lesson: none — one-off: a shared component kept the name its first caller needed; SearchSelect can no longer be used without naming its list.
BUG-294 — At 320 × 640 the account page’s Buying standards tab collapses to 0px tall
Affects: F1-d, F26-b (forge-et-fe src/features/accounts/AccountDetailPage.module.css, the Buying standards tab’s fill layout) · Found in: e2e/browser/accounts/account-detail-layout.spec.ts AC-21 on main ad7c6e63 with the specs updated, 2026-11-01 · Severity: medium — 320px reflow (WCAG 2.1 SC 1.4.10) · Phase: 1
Repro: at 320 × 640, open an account page and select Buying standards.
Expected: the standards Card keeps its content’s height, and <main> scrolls once the table region reaches its minimum height (the stylesheet’s own comment).
Actual: with that tab selected the page takes <main>'s height, and every level from the tabs down to the Card may shrink to nothing (flex: 0 1 auto; min-block-size: 0). The header (breadcrumb, h1, wrapped actions, four-cell summary strip) and the tab list use all the height, so the tab panel and the Card are 0px tall while the table spills out below them. 768 × 1024 is fine.
Fix: a minimum block size on the panel and the Card does not work: a level with min-block-size: 0 takes no floor from the region inside it, and removing the 0 stops the table scrolling in its card at 768. So the Buying standards tab bounds the page’s height only from bp/tablet (useMediaQuery(mq.tablet) in AccountDetailPage.tsx), as the Notes and timeline tab does from bp/laptop; below 768 the panel and Card keep their content’s height and <main> scrolls (forge-et-fe e760d58d, in forge-et-fe#371).
Lesson: none — one-off. The Notes and timeline tab’s fill layout was already limited to bp/laptop. The Buying standards tab’s fill had no such limit, and no test ran the 320px layout spec against the revamp until its browser specs were updated.
Regression test: AccountDetailRoute.test.tsx “the Buying standards tab bounds the page from bp/tablet only (BUG-294)” (failed without the fix); account-detail-layout.spec.ts AC-21 at 320, unmarked (failed before, passes after); the size matrix’s account-detail-tab-standards at every size.
Left open: a window shorter than the table region’s floor at 768 or wider can still collapse the same way; no supported size is that short (.forge/owed.md).
Resolved: 2026-11-03 — fixed in forge-et-fe#371 (e760d58d; merged 23014f39), a product defect, carried in the browser-spec PR whose gate it blocked.
BUG-293 — In the Edit user dialog, closing the Role select’s list, by Escape or by choosing a role, does not return focus to the Role trigger
Affects: F5-c (forge-et-fe the Edit user dialog’s Role select — a shadcn Select inside the Radix dialog) · Found in: e2e/browser/users/users-a11y.spec.ts NFR-2 on main ad7c6e63 with the specs updated, 2026-11-01 · Severity: medium — keyboard focus lost inside a modal (WCAG 2.1 SC 2.4.3, F5-c NFR-2) · Phase: 1
Repro: open Edit user, focus Role, open its list with the keyboard, then either press Escape or choose a role.
Expected: the list closes and focus is back on the Role trigger, so Tab goes on to the next control. It did this when the editor was a page.
Actual: the list closes (aria-expanded="false") and the dialog stays open, but focus is not on the Role trigger, after Escape and after a choice alike.
Fix: not traced; likely the Select’s close-and-restore-focus inside the dialog’s focus trap. Restore focus to the trigger on close.
Regression test: users-a11y.spec.ts NFR-2, the Escape test and the choose-a-role test, both unmarked.
Resolved: 2026-11-03 — fixed in forge-et-fe#410 (9454f5b8; review fixes 6fe705b3, cf86e8be), in the shared src/ui/shadcn/select.tsx. Cause: when the list closes, its focused option leaves the page, and the dialog’s Radix focus trap moves focus to the dialog’s content (its MutationObserver fallback for focus on <body>). The Select’s close handler from BUG-251 then saw focus on an element other than <body>, took it for a Tab the user had pressed, and cancelled Radix’s return to the trigger. The handler now also lets the return run when focus is on an element that holds the trigger and has tabindex="-1", as the trap’s container does, so a focusable region that holds the trigger and that the user Shift+Tabbed to keeps focus (6fe705b3). Every Select inside a dialog had the defect; in a running browser (the user’s own dev server against a local backend) the Edit user dialog’s Role select, after Escape and after a choice, and the New follow-up dialog’s Type select, after Escape, log focus as option → dialog content → trigger, and Tab from Role reaches See the … role; with the new condition disabled, focus stays on the dialog content, as reported. Two new tests in select.test.tsx (a Select in the shared Dialog, Escape and a choice) failed before the fix (2 failed | 10 passed) and pass after (12 passed). A row with the Select inside a tabIndex={0} region failed on the first fix (1 failed | 13 passed) and passes after the review fix (14 passed). A Tab pressed inside a dialog before the return runs is BUG-346. The two NFR-2 browser tests are unmarked and the third Role-select test no longer puts focus back by hand (first browser run: the nightly full run).
Lesson: L-133 — “nothing holds focus” is not only <body>: inside a focus trap the trap’s own fallback holds it; added to L-133’s sources.
BUG-292 — Links on the user page are 19.5px tall, under the 24px target-size floor
Affects: F5-c, F5-e (forge-et-fe the user page’s role link and the Reporting line card’s person links) · Found in: effective-permissions-a11y.spec.ts and the size matrix’s user-view state on main ad7c6e63, 2026-11-01 · Severity: medium — the browser gate’s target-size floor (design-system §9) · Phase: 1
Repro: pnpm exec playwright test e2e/browser/size-matrix --grep=user-view --project=size-1280x800-chromium.
Expected: every link at least 24 × 24px (or the spacing exception).
Actual: the role link (“Pinned Administrator”) and each person link are about 19.5px tall.
Fix: give inline record links in the side cards the 24px minimum block size.
Regression test: the size-matrix user-view state and effective-permissions-a11y.spec.ts’s A section checks, unmarked.
Resolved: 2026-11-03 — fixed in forge-et-fe#390: the Reporting line card’s person and manager links (ReportingLineChain.module.css .link) are inline-flex boxes with a space/7 minimum height. The summary strip’s role link already measured 24px (min-h-6), and the role link in the section’s sentence is inline in running text, which the floor exempts. Checked in a browser with the page check’s own target-size rule on two user pages at 1024 and 768px: clean; with the old style the two person links measure 19.5px, as reported. The user-view and Access card marks and effective-permissions-a11y.spec.ts’s expected-failure check are removed (first run: the nightly full run).
Lesson: L-148 — the same cause as BUG-297; added to L-148’s sources.
BUG-322 — Four Stop hooks run past their 5-second timeout on almost every turn, so each turn waits for them and the local mirrors they rebuild are left stale
Affects: the harness’s Stop hooks in .claude/settings.json — regen-reference-graph.sh, track-token-usage.sh, regen-usage-dashboard.sh, regen-docs-html.sh · Found in: a Claude Code /doctor check on the lead’s Mac, 2026-11-03 · Severity: low — no product effect; every turn in the harness ends about 5 seconds late, and the reference graph, the doc site and the usage dashboard are rarely rebuilt · Phase: 1
Repro: count the Stop hook records in the session transcripts (~/.claude/projects/*/*.jsonl, "type":"attachment" lines whose attachment.hookEvent is Stop). In the 50 most recent sessions (2026-10-28 to 11-02), 1,001 of 1,005 Stop records were hook_cancelled with timeoutMs: 5000: regen-reference-graph.sh 653, track-token-usage.sh 176, regen-usage-dashboard.sh 106, regen-docs-html.sh 65. Run alone in a clean worktree of main (a9375f80), python3 .claude/lib/harness-graph.py --write takes 5.6 s, longer than its timeout on its own.
Expected: the hooks that only rebuild gitignored local files finish, and the turn does not wait for them.
Actual: they run in the foreground with a 5-second timeout. Claude Code starts the hooks of one event together, not in the order they are registered, so each turn waits for the slowest (about 5 s) and then stops it. The graph generator has outgrown the limit; the usage hooks read a session transcript and a global usage file that grow with use (118 MB on this machine), and they wait on each other’s lock. .claude/hooks/README.md § Stop hook execution order says the hooks run in registration order, which is not how Claude Code runs them.
Fix: mark the four hooks "async": true (Claude Code runs an async hook in the background without blocking the turn) with a 60-second timeout, and keep the hooks whose output the turn needs — tracker-freshness.sh and forge-recap.sh (their systemMessage), verify-before-stop.sh (it can block), regen-tracker-dashboard.sh (fast) — in the foreground. The generators already tolerate an overlapping run: the graph writes through os.replace, and the usage hooks take a lock and resume from saved state. Correct the README’s order section. A case in .claude/hooks/tests/governance-guards.test.sh checks the settings: the four hooks are async with a timeout of at least 60 s, and the three that report or block are not async.
Resolved: 2026-11-03 — fixed in forge-et-harness#1080 (21e6694d). The four hooks run with "async": true and a 60-second timeout; tracker-freshness.sh, forge-recap.sh, verify-before-stop.sh and regen-tracker-dashboard.sh stay in the foreground. async is in Claude Code 2.1.316’s hook schema (“If true, hook runs in background without blocking”). The new governance-guards.test.sh case failed before the fix (93 passed, 4 failed, one per hook) and passes after (97 passed, 0 failed). The hooks README now describes each hook as foreground or background, not as an ordered list.
Lesson: L-250 (new) — a hook whose work grows with the repo outgrows its fixed timeout without anyone seeing it.
BUG-314 — useListSearchParams matches a landed write by its query string, so a query written twice leaves a stale pending entry and a later link to it is ignored
Affects: the shared forge-et-fe src/lib/hooks/useListSearchParams.ts (BUG-221, BUG-228, BUG-290; lesson L-201) and every screen on it · Found in: the review of forge-et-fe#372 (BUG-290, BUG-291), 2026-11-02; the matching line is unchanged from main · Severity: low — needs the same query written twice among writes the router has not rendered, then a link or path change to one of the stale queries · Phase: 1
Repro: a hook test with a held route loader (src/test/heldLoader.ts) under <StrictMode>, on forge-et-fe cdbb2ca6: hold; write tab=mine, then '', then tab=mine (each replace); release — the URL and the screen show tab=mine. Then router.navigate('/list'): the URL is '' but the screen still shows tab=mine (expected 'tab=mine' to be '').
Expected: a navigation the screen did not write is adopted, so the screen shows the URL.
Actual: the router renders only the last of the three writes. The hook finds its query with pending.indexOf(urlKey), which matches the first tab=mine, so '' and the second tab=mine stay pending. The later navigation to '' matches the stale '' and is taken for the screen’s own write landing, so the screen keeps tab=mine.
Fix: not decided. Match a landed write by its identity, not its query: give each write an id in the navigation state (NavigateOptions.state) and read it back from location.state, keeping any state a caller passes; or clear the pending entries when the router settles with nothing in flight.
Regression test: the repro above as a hook test, and one where the router renders the first write before the next is made (the in-order case indexOf gets right today).
Update (2026-11-02): claimed and released the same day (Tharinda’s session) after checking the fix options against main (122e1fe8); nothing was changed. None of the simple fixes is safe:
- Matching the last pending entry (
lastIndexOf) instead of the first breaks the in-order case: withtab=mine,'',tab=minepending, a commit of the firsttab=minewould clear all three, and the commit of''that follows would be adopted as a navigation from outside. - An id in the navigation state changes
location.statefromnullto an object on every write.AccountListScreen.tsx’s return-from-form effect (around line 296) focuses the page title whenever the state is an object, so each filter change on Accounts would move focus out of the search box; every reader oflocation.stateon a screen that uses the hook (useRecordTab, the leads, roles, users, email templates and Pricing screens) would need checking for the same. - Clearing the pending writes when the router is idle also clears writes made after the render being committed that the router has not shown yet, which are the quick second changes the hook exists to keep (BUG-228, BUG-290). Reading the router’s navigation state also needs a data router, and some component tests render under a plain
<MemoryRouter>. A fix needs its own design, written with BUG-261 (the same hook, in progress), whose stale-write symptom may share the cause. Resolved: 2026-11-03 — fixed in forge-et-fe#411 (5123331b,64ddbf7a,05be2712), a product defect, by the id option. Each write carries an id in its route state (listSearchParamsWrite, beside any state the caller passes), and a committed navigation is the screen’s own write landing only when it carries a pending id; any other is adopted. The other options stay rejected for the reasons above. The focus risk was real and is handled: the hook exportscallerRouteState, which strips the id, and the Accounts list’s return-from-form effect reads it, so its own writes no longer look like arrivals. Without that changeAccountListScreen.test.tsx’s “Clear filters empties the URL and moves focus to the search box” fails (focus leaves the search box); with it the file passes. Every otherlocation.statereader on a screen that uses the hook checks a key of its own (noticeFrom,listReturnToFrom,asksToEdit,focusLeadFrom, …), so the extra key changes nothing there. New tests inuseListSearchParams.test.tsx, under<StrictMode>: the repro (three writes held, only the last rendered, then a link to'') fails onmain’s matching (expected 'tab=mine' to be '') and passes now; the in-order case (each write rendered before the next, then links to''and back) passes before and after; and a caller’s route state is kept beside the id. The hook file passed 14/14 three runs in a row, and every test file of the screens on the hook (Accounts, Contacts, Leads, Lead view, Users, Roles, Email templates, Activities, Pricing and its URL-state hooks,useListSort,useRecordTab, BUG-290’s interrupted test) passes when run on its own; in one 26-file run on a machine shared with 12 other sessions, six cases timed out or read the page before its first render, and each passed on a rerun of its file. In a running browser (Accounts list, stub sign-in): Shared with me clicked on, off and on in one task sent three writes and only the third (listSearchParamsWrite: 3) landed, with the switch on; the sidebar’s Accounts link to/accountsthen showed the switch off; Back showed it on again and Forward off, each matching the URL; and typing a search kept focus in the search box once the write landed. BUG-261 (Biseni, in progress) was not touched. Lesson: L-201 — a write compared with its own landing by the value it wrote, not by its identity; listed among the lesson’s sources, with an update on matching writes by an id.
BUG-277 — At a short window, such as 320×256 (400% zoom), the dialogs whose footer band stays in view leave little or no room for the form
Affects: F3-a’s New follow-up dialog (forge-et-fe src/features/activities/FollowUpDialog.tsx, .footer in FollowUpDialog.module.css), and the dialogs built on RecordEditDialog (EditAccountDialog, EditContactDialog, NewAccountDialog, the role editor and role view) · Found in: the delta review of forge-et-fe#353’s merge of main (F3-a wave 2), 2026-11-01 · Severity: medium — a reflow loss (WCAG 2.1 SC 1.4.10, PRD § Accessibility › Supported sizes) at 400% zoom, which the browser gate’s 320px size (at a 640px height) does not reach · Phase: 1
Repro: not reproduced; read from the code. Open the New follow-up dialog in a 320×256 window (or 1280×1024 at 400% zoom). The dialog is at most 100vh − 32px tall; its header band and its sticky footer band (about 118px, the buttons in one column) stay in view, leaving about 46px for the form, which scrolls behind them. RecordEditDialog’s footer is flex: none beside a scrolling body, the same shape.
Expected: at 320×256 the form can be read and filled: the bands take a small share of the dialog’s height (F3-a’s edit dialog test in forge-et-fe#353, e2e/browser/activities/activities-edit-footer-short-viewport.spec.ts, asserts at most 60%), or the footer scrolls with the form at that height.
Actual: the header and footer bands take most of the dialog’s height.
Fix: measure each dialog at 320×256 first. Then either wrap a narrow footer’s buttons in rows, as forge-et-fe#353 does for F3-a’s edit dialog (Cancel and the main action on one row), or let the footer scroll with the form below a height threshold. Design-system Rev 74’s “kept in view while the body scrolls” stands at ordinary heights.
Regression test: a browser test per dialog at 320×256 that the footer band is at most 60% of the dialog’s visible height and the first field can be scrolled fully into view above it.
Note: the number BUG-277 was first used on F3-a’s wave 2 harness branch for a duplicate of BUG-276, dropped before that branch merged.
Decision: Nanndish’s, 2026-11-05: the fix follows design-system Rev 90 (only the body scrolls, so the title and the buttons stay in view) at every height. The whole-dialog scroll below a height threshold, in forge-et-fe#407 as first opened, is dropped, and #407 finds the room another way within Rev 90.
Note (2026-11-05) — the RecordEditDialog part is fixed in forge-et-fe#407 (branch fix/BUG-277-dialog-short-window, af6b8b5e), reworked to design-system Rev 90: only the body scrolls, at every height, so the title and the buttons stay in view. Measured in a running browser (headless Chromium, the live dialogs) at 320 × 256: the role editor, New account, Edit contact and the user editor were each 190px tall, with a 73px title row and a 73px footer (its buttons already on one row), which left the body 44px, less than its own padding; no field could be seen. Below a window height of 320px (ORDINARY_HEIGHT, decided in script with useMediaQuery, as the CRM’s other breakpoints are) the frame sets data-height="short" and takes space/3 above and below the title row and the footer and space/4 above and below the body; the sides are unchanged. After: at 320 × 256 each of the four dialogs has a 49px title row, a 49px footer and a 92px body (74px inside its padding), the first labelled field (56 to 64px) fits in it, the title and every button are in view, and the dialog itself does not scroll. At 320 × 640 and 1280 × 800 the frame is unchanged (73px bands, the same body heights). Regression tests: e2e/browser/shell/record-edit-dialog-short-viewport.spec.ts, one test per dialog at 320 × 256 (the dialog does not scroll, the title and each footer button are in view, the footer at most 60% of the dialog, the first field fits in the body with its label), the role and user editors at 320 × 320 (the ordinary spacing still fits a field at the lowest height it serves) and the role editor at 320 × 640; 8 passed, and without the short-window rules the four 320 × 256 tests fail on the field check. src/ui/overlays/RecordEditDialogFrame.test.tsx pins data-height at 320px and 256px and the footer’s test id; on main’s frame two of its three tests fail. Not fitted — two other dialogs on the same frame, not among this bug’s dialogs, where within Rev 90 the frame’s spacing alone cannot make room, so each needs a decision (owed.md): the template editor (F20-d), whose three buttons (Delete template, Cancel, Save template) wrap to two rows at 320px, a 94px footer even with the short spacing, leaving a 47px body that a 56px field does not fit even with no body padding (a decision on its buttons); and the sharing dialog (F5-g), whose title (Access to and the record’s name) wraps to four lines for the size matrix’s long pinned account name, a 123px title row at 320 × 256 (147px at 320 × 320) that leaves an 18px body (a decision on the title, such as limiting it to one or two lines); with a short record name its title is one line and the frame fits. The New follow-up dialog’s part lands with F3-a wave 3’s PR (WI-3.1, first reviewed as forge-et-fe#393, which closed as superseded by the wave branch feature/F3-a-wave-3 on 2026-11-06), which reworks that dialog, so the bug stays in progress until it does.
Note (2026-11-08) — measured again, on main 8f780295 and on forge-et-fe#442 c7d82619 (Chromium, the size-matrix states). On main, after forge-et-fe#419’s Rev 90 frame, the template editor and the sharing dialog with the long name both fit at 320 × 256: the template editor’s two button rows leave a 102px body, and the sharing dialog’s three-line title leaves 92px, each holding its first field. Neither needs the decision owed.md asked for, and both rows are removed. F3-a wave 3 (forge-et-fe#431) did not move the follow-up dialogs onto the frame: on main the whole dialog still scrolls, with a 118px sticky footer, and the title scrolls away. forge-et-fe#442 moves them onto the frame, but there, at 320 × 320, still on the ordinary spacing, the follow-up dialogs’ title row and description left a 39px body, and the template editor’s title-cased buttons wrapped to three rows (120px), leaving 40px. In the Edit follow-up dialog the narrow footer’s two columns (minmax(min-content, 1fr)) could not shrink below Save Follow-Up, so Delete and Cancel spilled past the dialog’s left edge, 13px out of view at 320 × 400.
Decision: Nanndish’s, 2026-11-08: fix it in forge-et-fe#442; the frame takes its short spacing below 400px tall instead of 320px; below bp/tablet the template editor’s buttons show Delete, Save and Create, and their accessible names keep “Template” (WCAG 2.5.3). The follow-up dialogs’ save button follows the same rule (Save, named Save Follow-Up), because the Edit follow-up footer did not fit otherwise.
Resolved: 2026-11-08 — fixed in forge-et-fe#442 (cbf56e7e, then 80ac1077, 1e0f130d and 8b56362a from its review), after forge-et-fe#407 and #419. ORDINARY_HEIGHT is (min-height: 400px); TemplateEditorForm.tsx (TEMPLATE_BUTTONS) and the follow-up dialogs (followUpLabels.ts) shorten the labels below bp/tablet. After: at 320 × 256, 320 × 320 and 320 × 400 every dialog on the frame keeps a body of at least 92px with its first field in it, the dialog itself never scrolls up, down or sideways, and every button is in view and inside the dialog. e2e/browser/shell/record-edit-dialog-short-viewport.spec.ts now covers eight dialogs (the four before, the New and Edit follow-up dialogs, the template editor and the sharing dialog with the long name), each at 320 × 256, 320 × 320 and 320 × 400 (the new lowest ordinary height), and checks that each footer button lies inside the dialog’s box. With activities-edit-footer-short-viewport.spec.ts, 37 passed; with #442’s c7d82619 source 10 fail (every dialog at 320 × 320, and the Edit follow-up dialog at all three heights), and the containment check alone catches the spill (Delete at 19.25px to 264px against the dialog’s 32px to 288px). RecordEditDialogFrame.test.tsx pins 400px as ordinary and 320px as short. Unit rows at 320px and 1280px (src/test/viewport.ts’s stubViewportWidth) check each shortened button’s shown text and that it starts the full name; with the narrow branch removed, all four narrow rows fail. Whether the other framed dialogs shorten their labels too is a design question in owed.md. The size-matrix states for every framed dialog at the four Chromium sizes: 393 passed, 1 failed, email-template-refresh-gone at 320 × 640 (the editor body scrolls sideways by 8px), which fails the same way on c7d82619 and passes on main: a separate #442 regression, its Back to Email Templates button (201px, title-cased with an icon, unable to wrap) indented beside the panel’s icon. By the same rule (Nanndish’s decision, 2026-11-08) it shows Back below bp/tablet and keeps its full name (8b56362a); the state then passes 3 of 3 at 320 × 640, and every email-template size-matrix state passes at the four Chromium sizes (98).
Lesson: L-265 — a size check for a shared dialog frame covers every dialog that uses it, so a dialog moved onto the frame later, or a label that widens, is measured too.
BUG-291 — Four URL-state hooks write with the functional setSearchParams, which starts from the URL of the render that made it, so a second quick change can put back what the first removed
Affects: F26 and F1-c (forge-et-fe src/features/activities/activitiesUrlState.ts, src/features/pricing/standardsBookUrlState.ts, src/features/pricing/weeklyPricesUrlState.ts, src/features/pricing/auctionSalesUrlState.ts) · Found in: the review of forge-et-fe#367 (BUG-223), by reading the code, 2026-10-31 · Severity: low — not reproduced; the symptom would be a filter or search change undone by a second change made before the first rendered · Phase: 1
Repro: not reproduced. The same class as BUG-221 and BUG-228 (lesson L-201): React Router’s functional setSearchParams builds the next URL from the render that called it, so a write fired before an earlier one has rendered starts from the URL before the earlier one.
Expected: each change starts from the change before it, as on the screens moved onto useListSearchParams.
Actual: these four hooks still call the functional setSearchParams from plain useSearchParams.
Fix: not decided. For each, a held-loader test (src/test/heldLoader.ts, as BUG-221’s) of two quick changes; then move each onto a hook that reads the URL it last asked for, useListSearchParams. The Standards book is done (forge-et-fe#368, for BUG-211, which turned out to be this defect): moved before BUG-290’s fix because the old code drops a change whenever the timing allows, while BUG-290 only falls back, inside a discarded render, to the committed URL the old code always started from, and on the shared hook BUG-290’s fix covers it too. Activities, Weekly prices and Auction sales remain.
Regression test: each hook’s held-loader test of two quick changes.
Resolved: 2026-11-02 — fixed in forge-et-fe#372 (b8567df2, cdbb2ca6; merged 122e1fe8), a product defect found by reading the code. Activities, Weekly prices and Auction sales now write through useListSearchParams, so each change starts from the query string the screen last asked for. New held-loader tests of two quick changes (activitiesUrlState.held.test.tsx, pricingUrlState.held.test.tsx) fail on main — the first change is gone (expected 25 to be 50, the round back on tab: 'standards', sale: null, scope missing) — and pass now. The PR’s review found a fifth place: the Weekly prices status filters (RoundSummaryBar) built their links from the rendered URL, so a filter pressed right after a sort dropped it. They now read the page’s shared copy, and a PricingRoute.test.tsx held-loader case fails before and passes after. The Standards book was moved earlier (forge-et-fe#368, BUG-211).
Lesson: L-201 — a write built from, or compared with, the URL of the render that made it; listed among the lesson’s sources.
BUG-290 — useListSearchParams takes a render that shows an older location for Back, so a quick second change to a tab, select or switch can undo itself
Affects: the shared forge-et-fe src/lib/hooks/useListSearchParams.ts (BUG-221, BUG-228; lesson L-201) and its callers: Accounts, Contacts, Leads, Users, useRecordTab, useListSort, and the Standards book (src/features/pricing/standardsBookUrlState.ts, moved onto the hook by forge-et-fe#368 for BUG-211) · Found in: the review of forge-et-fe#367 (BUG-223), 2026-10-31, when the email templates list was built on the hook (commit 4cec0232) · Severity: medium — the screens that use the hook can show and write a different tab, filter or sort than the user chose, after quick changes · Phase: 1
Repro: on forge-et-fe 4cec0232, EmailTemplatesScreen.test.tsx’s “ArrowRight, ArrowLeft, ArrowRight then a search ends on Mine with the search” as 16 parallel copies: 15 of 16 failed, and 8 of 10 in a second run with only that case running (so not machine load). The whole file as 16 copies failed it 9, 10 and 13 times in three rounds, always at the third ArrowRight’s Mine aria-selected="true".
Expected: three quick tab changes end on the third one, and the URL follows.
Actual: the reviewer’s trace, the same in every failing copy: ArrowRight writes tab=mine; React starts the router’s render of that navigation, in which the hook sets seen to its location key and clears pending, but that render is discarded and never commits. The hook’s state change is kept. Later committed renders still show the router’s first entry (location.key default, navigation type POP, as a memory router’s first entry is), so on the third ArrowRight the hook sees a key other than seen with type POP, takes it for Back, adopts '', and the tab goes back to Shared. By the code, any older location whose query is not pending would be adopted the same way.
Fix: not decided. Count a location as a navigation only the first time its key appears (keep the keys seen, ignore an older one), or adopt only in a layout effect keyed on location.key, so only committed renders count. Then move the email templates list onto the hook (it keeps its own mirror for now, BUG-223’s fix in forge-et-fe#367).
Regression test: a hook or screen test of three quick writes from key events, also under a root <StrictMode> as main.tsx renders the app; and the templates keyboard case as 16 parallel copies with no failure once the list is on the hook.
Resolved: 2026-11-02 — fixed in forge-et-fe#372 (575c8199, cdbb2ca6, c1f1fc7a, dc2eea21; merged 122e1fe8), a product defect. The hook now checks each committed navigation once, in a layout effect keyed on location.key, so a render React starts and throws away changes nothing; the email templates list moved onto the hook. A deterministic test, useListSearchParams.interrupted.test.tsx, renders outside act with slow children so the next key press interrupts the router’s transition render: on main’s hook it failed 3 of 3 runs (the tab back on shared after the third key), and it passes 5 of 5 now. The templates test file as 16 parallel copies failed the keyboard case 3, 5 and 9 times in three rounds on main’s hook and 0 times on the fix (496 passed each). The PR’s review found that the Pricing page’s tabs and panels each held their own copy of the hook, so a panel mounted by a tab switch the router had not rendered started from the old URL and its first write put the page back on the old tab: ListSearchParamsProvider now gives a page one shared copy, PricingRoute uses it, and a held-loader test fails before and passes after. Deferred from the review: BUG-314 (a query written twice leaves a stale pending entry), and an owed.md check of whether /leads/:id’s two copies lose a change.
Lesson: L-201 — the shared hook, the fix for that lesson’s class, itself read state from a render React had not committed; listed among the lesson’s sources.
BUG-289 — Flaky test: ContactAddressLookupIT’s NFR-1(a) plan check fails when Postgres serves the other-address branch from a different index (Windows)
Affects: F20-l (forge-et-be src/test/groovy/com/empireteas/account/repository/ContactAddressLookupIT.groovy:139, “NFR-1(a) — each branch of the lookup can be served by its address-key index”) · Found in: a full ./gradlew build --no-daemon --rerun-tasks on forge-et-be main (ef022576), Windows 11, 16 cores, run beside three full pnpm test runs of forge-et-fe while checking BUG-215, 2026-10-30 · Severity: low — one IT fails now and then, which fails the build record; the lookup itself works · Phase: 1
Repro: a full ./gradlew build on Windows while the machine is loaded. Not yet tried alone.
Expected: the plan of the lookup’s statement names both contact_address_key_idx and contact_email_address_key_idx.
Actual: plan.contains('contact_email_address_key_idx') was false: Postgres served the contact_email branch with an index scan on contact_email_contact_id_lower_email_key and a filter on the address, under a merge join on contact.id. The test asserts which index the planner picks on a small shared table, which depends on its statistics and on the rows other tests left (.forge/test-strategy.md § Flaky Tests, cause 1).
Fix: not decided. Make the plan independent of the shared table’s statistics (for example ANALYZE the two tables after seeding a fixed set of rows, or check the plan in a container of its own), or assert the property NFR-1(a) needs (each branch is an index scan with the address in its index condition) rather than an index name. A longer wait is not the fix.
Regression test: the IT passes in repeated full builds on Windows.
Resolved: 2026-11-03 — fixed in forge-et-be#365 (1785ce82; review fixes f58a0b3, 94c6065d), a test defect; the lookup was always served by its indexes. The cause is not the machine’s load: the lookup’s UNION can remove duplicates with a Merge Append of inputs already in id order, so under enable_seqscan = off alone an index leading with the id (contact_pkey, contact_email_contact_id_lower_email_key), scanned whole with the address as a filter, competes with each address-key index, and the shared tables’ statistics decide which wins. Reproduced on a scratch Postgres 18 with skewed statistics, and on every table size and statistics state tried once enable_sort and enable_hashagg were also off. The check now also switches off enable_indexscan and enable_indexonlyscan: a bitmap scan is built only from an index condition and Postgres 18 counts disabled plan nodes before costs, so each branch shows its address-key index exactly when its expression matches it. It asserts the Bitmap Index Scan on <index> lines, and still names the indexes NFR-1(a) names. A check that accepted “some index, no sequential scan” was rejected: a new negative control (the branch on lower(e.email)) is served by contact_email_contact_id_lower_email_key, which that check would pass. A new case with sorts and hash aggregation also off failed on the old check (plan.contains('contact_address_key_idx') false, both branches on id-ordered full scans with filters) and passes now; the spec passed three runs in a row on Windows 11 (11 tests, 0 failures each).
Lesson: L-255 — new: to ask whether an index can serve a predicate, leave the planner only bitmap scans; backend-reviewer § Test isolation checks it.
BUG-288 — A contact’s create response carries modifiedAt to the 100 nanoseconds, but the stored value keeps microseconds, so the next read returns a different modifiedAt
Affects: F20-l (forge-et-be src/test/groovy/com/empireteas/api/ContactOtherEmailsControllerIT.groovy:220, “a PATCH sending the same set in the same order leaves version and modifiedAt unchanged (AC-4)”), and possibly every create response that returns a timestamp set in memory · Found in: two full ./gradlew build --no-daemon --rerun-tasks runs on forge-et-be main (ef022576), Windows 11, 2026-10-30 · Severity: low — the IT fails on every Windows build; a client comparing a created record’s modifiedAt with a later read sees a change that did not happen · Phase: 1
Repro: on Windows, ./gradlew test --tests 'com.empireteas.api.ContactOtherEmailsControllerIT' — failed in both full builds that ran it.
Expected: a no-change PATCH answers the same modifiedAt the create answered (AC-4).
Actual: the create answered 2026-10-30T16:12:54.249790200Z and the PATCH 2026-10-30T16:12:54.249790Z. Java’s clock on Windows has 100-nanosecond resolution and the create response carries the in-memory value; Postgres timestamptz keeps microseconds, so every later read carries the truncated value. macOS’s clock reports microseconds, so the two match there.
Fix: give Spring Data JPA auditing a DateTimeProvider (@EnableJpaAuditing(dateTimeProviderRef = …) in forge-et-be platform/config/JpaConfig.java) that returns the system clock’s instant truncated to microseconds, so every @CreatedDate and @LastModifiedDate, and the contact change stamp that reuses the auditing handler, carries only what timestamptz keeps. Then look for other timestamps set in memory and returned before a read, and record what is found.
Regression test: the IT passes on Windows, and a unit test of the timestamp source shows a value with sub-microsecond digits truncated.
Resolved: forge-et-be#344. JPA auditing takes its time from MicrosecondDateTimeProvider (platform/config), the system clock cut to microseconds, so every @CreatedDate and @LastModifiedDate and the contact change stamp carry only what timestamptz keeps. ContactOtherEmailsControllerIT passes on Windows 11 (31 tests, AC-4 included), MicrosecondDateTimeProviderSpec proves the cut with a fixed clock on any operating system, and a new AuditingSpec feature fails when the provider is not wired. Not Windows only: Java’s clock carries nanoseconds on Linux too, so CI and every deployed environment had the mismatch, and only the developers’ Macs (microseconds) hid it; the (Windows) suffix was dropped from the title. Other timestamps set in memory: the mailbox services already cut theirs (MailboxConnectionService.now(), MailboxIngestService); AccountMigrationService, ContactMigrationService and RecordTagService bind a bare Instant.now() into bulk SQL but return only ids or a count, so no in-memory value reaches a client.
Lesson: L-128 — the same cause as F26-b’s buying-standards create; this fix makes the backend change that lesson named as a candidate.
BUG-287 — The account.repository package comment says only account.service reaches it, but the lead package does too
First filed on forge-et-harness#1038 as BUG-286; renumbered because an open PR opened earlier holds BUG-286 (.claude/rules/id-numbering.md rule 5).
Affects: F1-a (forge-et-be src/main/java/com/empireteas/account/repository/package-info.java, line 4) · Found in: F20-l WI-2.1’s pre-PR review (forge-et-be#339), 2026-10-30 · Severity: low · Phase: 1
Repro: on main, git -C forge-et-be grep -n 'account.repository' -- src/main/java/com/empireteas/lead lists LeadIntakeService and LeadMatchReach importing ContactRepository. After F20-l WI-2.1, LeadTriageRepository’s matcher reaches ContactAddressLookup too.
Expected: the package comment names every package that reaches it, or the code keeps to what the comment says.
Actual: the comment says “Reached only from account.service”, which was already false before F20-l. It is a comment only; the ArchUnit layering rules pass.
Fix: not decided. Either reword the comment to name the lead package’s readers, or route those reads through account.service. Whichever is chosen, the layering rule that allows the lead package’s access should be stated.
Resolved: 2026-11-02 — fixed in forge-et-be#345 (e6e56374) by rewording the comment, the first of the two options. account/repository/package-info.java names every reader outside account — lead.service’s LeadIntakeService, LeadMatchReach, LeadService and LeadTriageMatcher — and the two rules that allow them: LayeringSpec (any ..service.. package may reach any ..repository.. package) and FeatureSlicesSpec rule (a) (lead may depend on account). It also gives the git grep that re-checks the list. Comment only; no test applies.
Lesson: none — one-off: a package comment that went stale when another feature added a reader.
BUG-286 — Initials circles show a bracket, such as “R(”, for a name whose last word starts with a bracket
Affects: F5-g, F5-c (forge-et-fe src/ui/components/Avatar/Avatar.tsx initialsOf, used by every initials circle: the Shared with card’s faces, the people in the access dialog, the users list) · Found in: the UI revamp preview’s Shared with card on a lead (forge-et-fe#364), local stack, 2026-10-30 · Severity: low — display only; each circle’s accessible name is the full name.
Repro: share a lead with a user whose display name ends in a bracketed word, such as the local seed’s “Lead (test)” or “Rep (test)”, and look at the Shared with card’s faces.
Expected: the circle shows the name’s initials, such as “L” or “LT”, never a bracket.
Actual: the circle shows “L(” or “R(”: initialsOf takes the first character of the first and the last word, and the last word is “(test)”.
Fix: take each initial from the first letter of a word, skipping words that start with no letter (a bracket, a digit, punctuation); a name with no such second word yields one letter. Add the bracketed name to Avatar.test.tsx’s initialsOf cases.
Regression test: initialsOf('Rep (test)') gives R, and initialsOf('Ama (x) Silva') gives AS.
Resolved: 2026-11-02 — fixed in forge-et-fe#375 (5b369e9e). initialsOf (src/ui/components/Avatar/Avatar.tsx), which every initials circle uses (PersonAvatar, RoleHoldersCard, ReportingLineChain, ContactProfileCard, ConnectedPanel), takes initials only from words that start with a letter (\p{L}), so a bracket, digit or quote is never an initial and a name with no such word yields nothing. Regression tests in Avatar.test.tsx: “Rep (test)” and “Lead (test)” give one letter, words starting with a digit or punctuation are skipped at either end, non-Latin letters count, and “(test) 42” gives nothing — three failed before the fix; “Ama (x) Silva” already gave “AS”.
Lesson: none — one-off: the only place that derives display text from the first character of each word of a name.
BUG-285 — Flaky test: MailboxIngestFailurePathsIT’s transient-503 row ran out its dead-letter wait in a full build under heavy load (macOS)
First filed on forge-et-harness#1025 as BUG-282; renumbered because main had filed its own BUG-282 first (.claude/rules/id-numbering.md rule 5).
Affects: F20-a (forge-et-be src/test/groovy/com/empireteas/mailbox/…/MailboxIngestFailurePathsIT.groovy, “a transient 503 failure is retried to the queue’s bound and leaves NO claim”) · Found in: F20-l WI-1.1’s full ./gradlew check on feature/F20-l-WI-1.1-other-addresses-table (622bbf0d), 2026-10-28 · Severity: low · Phase: 1
Repro: not reproduced. The full build ran 58m 54s at a load average above 10; the row waited 238 s for the message to reach the dead-letter queue and gave up. The branch changes nothing under mailbox.
Expected: the 503 is retried to the queue’s bound and the message reaches the dead-letter queue within the row’s wait.
Actual: the wait ran out once. The spec on its own passed 10 of 10 in 71 s, and the next full build passed (BUILD SUCCESSFUL, 26m 16s, 10,796 tests, 0 failed).
Fix: not decided. Triage per .forge/test-strategy.md § Flaky Tests: the same class as BUG-195 and BUG-259 (an SQS-backed IT’s dead-letter wait under load). If it recurs within the load limit, raise its wait or bound the retries’ delay in the test profile.
Triage, 2026-11-04 (step 2 of .forge/test-strategy.md § Flaky Tests): on the lead’s Mac, forge-et-be main d3df4fc2, ./gradlew test --tests 'com.empireteas.mailbox.MailboxIngestFailurePathsIT' --rerun three times with one busy loop per core (yes > /dev/null × 8) beside it and other sessions’ Gradle runs going, a load average of 41–43 on 8 cores: 10 of 10 each time (131.8 s, 80.1 s, 82.2 s). Forced load did not reproduce it.
Resolved: 2026-11-04 — closed as wontfix (machine overload), the lead’s decision (Tharinda, 2026-11-03: reproduce, then close if it does not). The evidence: the failing full build ran at a load average above 10, forced load did not reproduce it, the spec alone passed 10 of 10, and the next full build passed. No change was made. The build record’s load check (limit 0.65) refuses a build on a machine that busy. A recurrence on a machine within the limit is filed as a new bug that links here, with the record’s Machine load: line.
Lesson: L-162 (a failure that appears only under load is reproduced before it is called flaky). An instance where forced load did not reproduce it, so nothing was changed, as for BUG-195 and BUG-259.
BUG-284 — In the role editor, a blocked “All records” tick shows its reason only on mouse hover, not when reached with Tab
First filed on forge-et-harness#1015 as BUG-277; renumbered because forge-et-harness#1005, opened earlier, files its own BUG-277 (.claude/rules/id-numbering.md rule 5).
Affects: F5-d (forge-et-fe src/features/roles/AllRecordsTick.tsx, GatedTick, and the Tooltip it uses; likely RequirementTick too, the same component for Publish weekly prices) · Found in: F38-g’s retest on dev, WI-4.1 check F5-d.4, 2026-10-28, against CRM deploy-dev run 40635347050 (a3ed965a) and backend run 40571164275 (fdca82c1); record .forge/plans/F38-g/F38-g-WI-4.1-record.md R-7 on feature/F38-g-wave-4 · Severity: medium — a keyboard user gets no visible reason why the tick is blocked; WCAG 2.1 SC 1.4.13 asks content shown on hover to be shown on focus too · Phase: 1
Repro: on dev, as CRM01, open Settings › Roles, a role, Edit. With Leads on None, press Tab until All records under Leads has the focus ring. Chrome, Windows 11.
Expected: F5-d AC-19: All records is aria-disabled="true", stays focusable, and carries “Choose View or Manage first” as a tooltip on focus and in its accessible name. The component’s own comment says “a tooltip on focus and hover”.
Actual: focused with Tab, the box shows its focus ring and no tooltip. Hovering the box with the mouse shows “Choose View or Manage first”. The accessible name was not read in this walk.
Fix: not decided. Find why the tooltip does not open on keyboard focus — whether Tooltip opens only on pointer hover, or the focus lands on an element the tooltip does not watch (the checkbox input inside the label rather than the trigger) — and open it on focus as well, for AllRecordsTick and RequirementTick alike. Check it in a browser with the keyboard before and after.
Regression test: a browser test that tabs to a blocked All records (and Publish weekly prices) and asserts the tooltip is visible; the existing unit test in RoleEditorRoute.test.tsx checks the tooltip after an event the browser may not send on Tab.
Resolved: 2026-11-02 — fixed by forge-et-fe#364 (the UI revamp preview, merged be8df6d3), which replaced the role editor. In its Edit role dialog (RoleAccessEditor.tsx) an area’s All records box is not shown while None is chosen and appears once View or Manage is chosen, so the blocked box this bug was about no longer exists; the lead kept that form on 2026-11-02 (F5-d Rev 9, AC-19). The one tick that is still blocked, Publish weekly prices until Pricing is chosen, shows its reason as text under it and is described by it, not by a tooltip. Regression tests: RoleEditorRoute.test.tsx, “RoleEditorRoute — All records (AC-19)” (no box while None is chosen; cleared by choosing None) and the Publish weekly prices cases beside it. AllRecordsTick.tsx’s tooltip form is no longer rendered by the editor. No code was changed for this bug.
Lesson: none — one-off: the component that had the defect was replaced before it was fixed.
BUG-283 — A saved account’s website, country, region or Est. LTV cannot be removed: the account patch reads an empty value as unchanged
Affects: F1-a (forge-et-be src/main/java/com/empireteas/account/web/AccountPatchRequest.java), F1-b (forge-et-fe account edit form, FR-C2) · Found in: BUG-272’s fix in forge-et-fe#356, 2026-10-28; the lead decided on 2026-10-29 that these values must be removable · Severity: low — a wrong value can be replaced but not removed; nothing is lost · Phase: 1
Repro: edit an account that has a website, empty the Website field and save.
Expected: the website is removed, and the account page no longer shows it.
Actual: the form refuses the empty value before any request (F1-b FR-C2, Rev 10), because AccountPatchRequest reads a null field as “leave unchanged” and cannot tell it from “remove” (F1-a WI-1.2 plan, finding N-4). Est. LTV has the same rule, and country and region the same limit.
Fix: revise F1-a so the patch can say “remove” for its optional fields (for example, an empty string clears a text field, or a list of fields to clear), change AccountPatchRequest and the service to match, with an audit event per AD #36 as for any write; then have the edit form send the removal instead of refusing it, and revise F1-b FR-C2. A backend API change: the frontend tests that assert the refusal are updated in the same window (L-200).
Regression test: a backend spec that patches a saved website to the “remove” form and reads it back absent; a form test that empties Website and sees the removal sent; both fail before the fix.
Resolved: 2026-11-02 — fixed in forge-et-be#345 (b4ac89c0, 7740de38) and forge-et-fe#375 (073294bc), by F1-a Rev 12 and F1-b Rev 12. The account PATCH takes an optional clear list naming fields to set to null (country, region, customerType, website, estimatedLtv); an absent or null member still leaves its field unchanged, so no existing caller changes. customerType was added by the delivery team as the one other nullable single-value field, and the lead agreed on 2026-11-02. Removing the estimate keeps ltvCurrency. Any other name, a null element, or a field both set and cleared is 400 VALIDATION_FAILED naming clear; a clear that is not an array of names is 400 MALFORMED_REQUEST. Each removal is in the patch’s account.updated event as old to new: null (AD #36). The edit dialog sends an emptied field in clear instead of refusing it, and its selects always offer the empty option. Regression tests: AccountControllerIT (each clearable field and all five, the refusals, the OpenAPI enum) and AccountAuditHttpIT failed before the fix (./gradlew test on those two specs: 13 of 50 failed); the review added cases for a name listed twice, an already-empty field and a malformed clear (129 tests across the affected specs passed after). accountForm.test.ts and EditAccountDialog.test.tsx failed 4 of 50 before the fix.
Lesson: L-249 — a PATCH whose null means “leave unchanged” needs a stated way to remove each optional value.
BUG-282 — Flaky test: the account form’s Country and Region are measured while the New account dialog zooms in, so AC-15’s top edges differ by more than 1px (macOS)
Affects: F1-b (forge-et-fe e2e/browser/accounts/accounts-layout.spec.ts:280, “AC-15 (supporting): on New account and Edit, Country and Region sit side by side at 768, 1024 and 1280px and stack at 320px”; the New account dialog, src/ui/shadcn/dialog.tsx’s DialogContent) · Found in: the nightly full run on main of 2026-10-29, forge-et-fe 470c3e40, on the lead’s Mac (forge-et-fe#354) · Severity: medium — a test-timing defect: the form lays out as F1-b specifies, but the failure turned the nightly run red and can fail any frontend browser gate · Phase: 1
Repro: macOS, the nightly full run’s CRM browser gate (Chromium, 8 workers) on main 470c3e40 with forge-et-be 944c6e73; the gate’s full log is ~/.et-nightly/logs/2026-10-29/fe-gate.log on the lead’s Mac. The test failed its run and its retry; it had passed in forge-et-fe#350’s full gate on f69662f8, and the commits between that gate and 470c3e40 change only the Biome configuration and the lockfile.
Expected: at 768px and wider, the top edges of the Country and Region fields are within 1px of each other.
Actual: /accounts at 768px: Country and Region share a top edge — received 1.46px, then 1.09px on the retry. The New account dialog zooms in over 200ms as it opens (data-[state=open]:zoom-in-95, duration-200), and openNewAccount waits only for the form to be visible and its certifications loaded. The test reads the two boxes one after the other while the zoom runs, so each is scaled by a different amount. The Edit form, a page with no entry animation, did not fail.
Fix: call settleAnimations(page) (e2e/support/pageCheck.ts) before the two boxes are read, so the test measures the dialog once it has opened. Test only; the product is unchanged.
Regression test: the AC-15 test itself, repeated.
Resolved: forge-et-fe#359 (01378dad, 27ed0264), 2026-10-29. The test calls settleAnimations(page) before it reads the two boxes, and openNewAccount (e2e/browser/accounts/support.ts) now ends with the same wait, so every caller gets the dialog at its final size (frontend-reviewer’s Nit). Fails before: the nightly run’s failure and its retry above. With the fix, on macOS: pnpm exec playwright test e2e/browser/accounts/accounts-layout.spec.ts -g "AC-15 \(supporting\)" --repeat-each=20 --retries=0 --workers=4 --project=chromium passed 30 of 30 (the 20 repeats and the setup tests) while other sessions’ backend builds raised the machine’s load from 7 to 33. An earlier run with 8 workers at a load of 60 on 8 cores timed out in 16 of 30 tests, none of them on this test’s top-edge check, so it proved nothing either way.
Lesson: L-101 — a test checked the page while it was still changing, here two boundingBox() reads taken during a dialog’s entry zoom. Added to L-101’s sources.
BUG-281 — A failed refresh of the Roles list replaces its rows with the error panel, back from a role’s page or on a refetch
Affects: F5-d (forge-et-fe src/features/roles/RolesRoute.tsx ~150–175, RolesState; settings/roles/:id is a sibling route of the list, src/app/router.tsx ~384) · Found in: frontend-reviewer’s second delta review of forge-et-fe#356, 2026-10-28, from reading the code; not yet run · Severity: low — rows the client already holds are hidden behind an error panel until Retry; nothing is changed or lost · Phase: 1
Repro: open Settings → Roles (its rows load), make its re-reads fail, then open a role and go back — or let a refetch run while the list is open.
Expected: the list’s last good rows stay on screen, marked stale, with Retry (coding-standards §3.2: a refresh failure never blanks good data).
Actual (from the code): the Roles list keeps no last-good snapshot at all, and RolesState checks query.status === 'error' before it looks at the data, so any failed refresh shows roles-list-error in place of the rows. It is BUG-278’s and BUG-280’s defect in a stronger form, since there is no snapshot to fall back on.
Fix: move the Roles list onto src/lib/hooks/useLastGoodSnapshot.ts (forge-et-fe#356), as the leads, accounts, contacts, users and email templates lists use it, and map a failure with a snapshot to stale with Retry. Check first whether the roles read is paged; the hook expects PagedParams (a single-page read passes { page: 1 }).
Regression test: a RolesRoute component test mirroring UsersScreen.test.tsx’s “back on the list from a user’s page (BUG-278)”, with createAppGcTestQueryClient and a route for settings/roles/:id, that fails before the fix.
Resolved: 2026-11-02 — fixed in forge-et-fe#375 (74c6fcaa, a4a34095). RolesRoute.tsx: a failed re-read whose cache still holds rows (status: 'error' with data) keeps them under the refresh-failed notice with Retry and the trace id (roles-list-stale, roles-list-stale-retry, roles-list-stale-trace); a failed first read, or one with no rows to keep, is still the error panel. Not on useLastGoodSnapshot: the Roles read is one unpaged key returning a plain array, so the cached data is the last good read itself and the hook’s paged shape (ListPage) does not fit; its comment now says so. Regression tests: RolesRoute.test.tsx, “a failed refresh keeps the rows (BUG-281)” — a refetch that fails while the list is open, a list whose refresh failed shown again from a role’s page, a re-read that fails on return, and Retry recovering; all four failed before the fix (Unable to find … roles-list-stale). The size matrix’s new roles-list-stale state opens a role and goes back while the list’s re-read fails: it passed at every size in both engines, twice, and failed with main’s RolesRoute.tsx at getByTestId('roles-list-stale').
Lesson: L-238 — the same defect on the one list the hook did not cover, with its rows in the query cache.
BUG-280 — Back on the email templates list from a template’s page, a failed re-read shows the error panel with no rows although the cache holds that list
Affects: F20-d (forge-et-fe src/features/email/templates/templatesQueries.ts useTemplatesList; settings/email-templates/:id is a sibling route of the list, src/app/router.tsx) · Found in: frontend-reviewer’s delta review of forge-et-fe#356, 2026-10-28, from reading the code; confirmed by a failing component test · Severity: low — rows the client already holds are hidden behind an error panel until Retry; nothing is changed or lost · Phase: 1
Repro: open Settings → Email templates (its rows load), make its re-reads fail, open a template, go back.
Expected: the list’s cached rows stay on screen, marked stale, with Retry (coding-standards §3.2).
Actual: BUG-278’s cause on the email templates list: useTemplatesList kept its own last-good snapshot in a ref, one per mount, with no fallback to the key’s cached data, so the new mount showed email-templates-read-failed and Retry with no rows.
Fix: useTemplatesList moves onto src/lib/hooks/useLastGoodSnapshot.ts ({ page: params.page ?? 1 }) and its isPastTheLastPage; its returned view is unchanged. The hook’s comments now name the lists that share it (leads, accounts, contacts, users, email templates) instead of “every list”; Activities keeps its own copy, having no sibling record page.
Regression test: EmailTemplatesScreen.test.tsx, “back on the list from a template’s page (BUG-280)”: the refresh-failed case failed before the fix (email-templates-stale not found) and passes after; the whole file passes (30 tests).
Resolved: forge-et-fe#356, 2026-10-28 (Tharinda).
Lesson: L-238 — the fifth instance, after BUG-267, BUG-270, BUG-273 and BUG-278. Added to L-238’s sources.
BUG-279 — The email templates table’s header row reads a CSS Module class that was never defined
Affects: F20-d (forge-et-fe src/features/email/templates/EmailTemplatesTable.tsx:92, styles.headerRow; EmailTemplatesScreen.module.css) · Found in: frontend-reviewer’s review of forge-et-fe#356, 2026-10-28, from reading the code · Severity: low — nothing visible changes: the build maps the missing name to undefined, which TableRow’s cn() drops · Phase: 1
Repro: read the code: EmailTemplatesTable.tsx passes styles.headerRow to the header row, and git log -S headerRow -- src/features/email/templates shows .headerRow was never defined in EmailTemplatesScreen.module.css since F20-d (a7cb17ef).
Expected: every CSS Module class a component reads is defined in its module.
Actual: the header row gets no class from it. It was probably meant to cancel the row’s border and hover tint, as the tag table’s .headRow does; the header cells’ opaque card fill already hides both. Under Vitest a CSS Module maps every name to itself, so no component test could see it.
Fix: remove the reference (<TableRow>), and add src/app/cssModuleClasses.test.ts, which fails when a component reads styles.x or styles['x'] that its imported module does not define. A one-off scan found no other such reference in src/.
Regression test: cssModuleClasses.test.ts failed before the fix (1 failed, 600 passed, naming styles.headerRow) and passes after.
Resolved: forge-et-fe#356, 2026-10-28 (Tharinda). Filed during forge-et-fe#356’s review as BUG-278; renumbered to BUG-279 when BUG-277 was taken first (forge-et-harness#1015) and this bug’s first number went to the Users-list bug.
Lesson: none — one-off: a CSS Module reference with no class behind it; the new scan test stops the next one, and no other bug has this cause.
BUG-278 — Back on the Users list from a user’s page, a failed re-read shows the error panel with no rows although the cache holds that list
Affects: F5-c (forge-et-fe src/features/users/usersQueries.ts useUsersList; settings/users/:id is a sibling route of the list, src/app/router.tsx) · Found in: frontend-reviewer’s review of forge-et-fe#356, 2026-10-28, from reading the code; confirmed by a failing component test · Severity: low — rows the client already holds are hidden behind an error panel until Retry; nothing is changed or lost · Phase: 1
Repro: open the Users list (its rows load), make its re-reads fail, open a user’s page, go back.
Expected: the list’s cached rows stay on screen, marked stale, with Retry (coding-standards §3.2).
Actual: BUG-267’s, BUG-270’s and BUG-273’s cause on the Users list: useUsersList kept its last-good snapshot in a ref, one per mount, with no fallback to the key’s cached data, so the new mount showed the error panel with no rows.
Fix: one shared hook, src/lib/hooks/useLastGoodSnapshot.ts (with isPastTheLastPage), now used by the leads, accounts, contacts and users lists; the users list gains the cached-rows step the other three had. The four copies differed only in that step and in their list and params types. This also settles owed.md’s row on extracting the snapshot into one hook.
Regression test: UsersScreen.test.tsx, “back on the list from a user’s page (BUG-278)”: the refresh-failed case failed before the fix (users-stale not found; the page showed “We couldn’t load the user list”) and passes after; useLastGoodSnapshot.test.tsx covers the hook’s rules (9 tests).
Resolved: forge-et-fe#356, 2026-10-28 (Tharinda). Filed during the review as BUG-277; renumbered to BUG-278 because forge-et-harness#1015 filed BUG-277 first.
Lesson: L-238 — the fourth instance, after BUG-267, BUG-270 and BUG-273; the shared hook is that lesson’s fix for all four lists. Added to L-238’s sources.
BUG-276 — Flaky test: effective-permissions AC-21a’s Retry is gone before the test clicks it, because the view’s own read succeeds after the editor’s failed one (Windows)
Affects: F5-e (forge-et-fe src/features/users/usersQueries.ts useEffectivePermissions, src/features/users/EffectivePermissionsSection.tsx; the test e2e/browser/effective-permissions/refresh.spec.ts:528) · Found in: the full pnpm e2e run with the Windows PR profile (ET_GATE_SCOPE=pr-chromium) in the ui-revamp worktree, forge-et-fe feature/ui-revamp, 2026-10-27/28 · Severity: low — a test-timing defect; the product shows the error panel and Retry as F5-e specifies, then its own later read replaces them with the rows · Phase: 1
Repro: on Windows, with other Playwright and Vitest runs loading the machine (four sessions running tests at once; no load figure was recorded, since the runs were outside the gate), run pnpm exec playwright test e2e/browser/effective-permissions/refresh.spec.ts --project=chromium --grep "AC-21a" --repeat-each=3. All three runs failed. A later re-run on a quieter machine (--last-failed, 2 workers) passed.
Expected: after the save, the section’s re-read fails (the test pins a 500 on the next GET), the error panel shows with Retry and none of the old rows, and the test presses Retry.
Actual: from the trace: after the PATCH, the editor’s section reads at the new version and gets the pinned 500. The page then returns to view mode and the view’s section mounts; useEffectivePermissions has gcTime: 0, so the failed entry was already dropped, and the view reads again and gets 200. The error panel and its Retry are replaced by the rows before the test can click. On a quiet machine the test clicks Retry before the view’s read lands.
Fix: not decided. Either the test pins the failure on every read until it presses Retry (the behaviour it asserts is the editor’s section, not the view’s), or the save’s return to view mode keeps the failed read so the view shows the same error instead of reading again. The user editor’s logic was not changed by the ui-revamp work (only its heading and placeholders), so this is a race in the existing code and test.
Regression test: AC-21a passing 3 of 3 with --repeat-each=3 while another test run loads the machine.
Resolved: forge-et-fe#355 (93155b85, fd423ded), 2026-10-28, by the first fix above. The test now pins the failure on every read of the section’s address (failEveryGet in e2e/browser/effective-permissions/support.ts, beside failNextGet), waits for view mode, and removes the pin only after it has checked the error panel and its trace line, before it presses Retry; the product code is unchanged. Fails before: in the browser gate on 69b40e0d (8 workers) and on 127ca6e5 (4 workers) AC-21a failed its run and its retry, once with the panel shown but its trace line gone and once with no panel, the snapshot each time showing the rows the view’s later read had loaded. With the fix, refresh.spec.ts passed 30 of 30 (--repeat-each=2, 4 workers, Windows).
Lesson: none — one-off: the only test in which a save moves the page to a second component that reads the same address again after the pinned failure; failNextGet’s other uses fail the first read of a page that reads it once.
BUG-275 — The email templates list’s row menu items fail axe colour contrast, so the templates accessibility browser test fails on main
Affects: F20-d (forge-et-fe the templates list’s row ⋯ menu, its dropdown-menu-item elements with data-testid="email-templates-row-…"; e2e/browser/email/templates-a11y.spec.ts:581, “AC-37 (rendered — axe)”) · Found in: forge-et-fe#349’s browser gate on bb48944d, 2026-10-28, after main was merged in; reproduced alone on a clean origin/main checkout (a5224f39, forge-et-fe#337’s merge) · Severity: medium — a WCAG AA contrast failure, and it fails every frontend browser gate on a branch that has main’s templates page · Phase: 1
Repro: on main, with the backend running: pnpm exec playwright test e2e/browser/email/templates-a11y.spec.ts --project=chromium --retries=0 -g "AC-37 \(rendered — axe\)" failed 1 of 1 on main and 3 of 3 on forge-et-fe#349’s branch (macOS, a quiet machine); the row-menu scan moved alone into a test of its own passed once, so the failure may depend on the steps before it in the same test.
Expected: axe finds no violation on the templates list with a row’s ⋯ menu open (F20-d AC-37).
Actual: axe reports color-contrast (serious) on two role="menuitem" items of the row menu. Run alone on #349’s branch, the same file’s keyboard tests at lines 409 and 520 also failed; not yet checked on main.
Fix: a test fix, not a colour fix. Read in the browser at the moment the test scanned, the open menu was at opacity 0, still in its fade-in (animate-in fade-in-0), while Playwright already reported it visible; 1.5 seconds later the items were rgb(44,33,23) on white, well over 4.5:1, and the scan passed. expectNoAxeViolations (e2e/support/pageCheck.ts) ran axe without waiting for animations, unlike collectPageCheckFailures; it now awaits settleAnimations first. No token changes. The keyboard tests at lines 409 and 520 passed on main with the fix (the whole file 21 of 21, Chromium).
Regression test: the existing AC-37 axe test passes: 1 of 1 failed on main, 3 of 3 passed with the fix (--repeat-each=3).
Resolved: forge-et-fe#356, 2026-10-28 (Tharinda).
Lesson: L-101 — a test checked the page while it was still changing, BUG-033’s and BUG-127’s cause again, here in the axe-only check that skipped the shared wait. Added to L-101’s sources.
BUG-274 — A busy shadcn button with a leading icon draws its spinner beside the icon, partly over it, instead of in its place
Affects: every screen moved to shadcn/ui with a busy button that carries a leading icon — the Refresh and Retry buttons of the Leads, Contacts, Activities and Email screens and the message view (F1-a, F1-f, F3-a, F20-b, F20-c), among others; the cause is forge-et-fe src/ui/shadcn/button.tsx · Found in: frontend-reviewer’s delta review of forge-et-fe#348 (bb53615f), 2026-10-27, from reading the code; not yet seen in a browser · Severity: low — the button still refuses the press and sets aria-busy; only where the spinner is drawn is wrong · Phase: 1
Repro: on /email at 1280px, press Refresh (an outline button with the retry icon) and look at it while the list re-reads; or any moved screen’s busy Retry.
Expected: design-system §5 Busy: the spinner takes the icon’s place and the label and the button’s width do not move, as the older src/ui Button’s LeadingSlot did.
Actual: the spinner sits in a strip w-4 (16px) wide at the button’s inline start, sized for the default px-4 padding. A button whose first child is an <svg> takes has-[>svg]:px-3 (12px), so the 12px ring, centred in the 16px strip, covers 2px to 14px while the icon starts at 12px: the ring overlaps the icon’s edge and shows beside it, and the icon stays.
Fix: in button.tsx, while busy, hide a leading svg (aria-busy:[&>svg:first-child]:invisible or the same in the spinner’s markup) and draw the spinner over the icon’s box, or size the strip from the padding the button actually has. Measure it in a browser before and after.
Regression test: a button.test.tsx row for a busy button with a leading icon (the icon hidden, one spinner), and a browser check that the spinner’s box does not overlap the label.
Fixed as: the first option. The spinner’s strip carries data-slot="button-spinner"; while busy, the button hides an icon that directly follows it ([&>[data-slot=button-spinner]+svg]:invisible, which keeps the icon’s box), and the strip moves onto that box through has-[+svg]: classes per size — the size’s has-[>svg] padding as its start, the icon’s width as its width. Measured on /contacts’ busy Retry at 1280 and 768px: before, the spinner box (x 1145–1161) overlapped the visible icon (1157–1173); after, it sits on the hidden icon’s box (1157, 16px wide), and the button’s width is unchanged. button.test.tsx’s new row failed before the fix and passes after.
Resolved: forge-et-fe#356, 2026-10-28 (Tharinda).
Lesson: L-212 — another prop path of the adapted shadcn Button (busy with a leading icon, at the has-[>svg] padding) that broke unseen because its test covered only the default case. Added to L-212’s sources.
BUG-273 — Back on the contact list from a contact’s page, a failed re-read shows the error panel with no rows although the cache holds that list
Affects: F1-f (forge-et-fe src/features/accounts/contactsQueries.ts, useContactsList’s snapshotRef, as forge-et-fe#337 leaves it) · Found in: fixing BUG-270 in forge-et-fe#349, 2026-10-27, from reading forge-et-fe#337 (draft, branch feature/ui-standardisation, head 9a0addd1) · Severity: low — rows the client already holds are hidden behind an error panel until Retry; nothing is changed or lost · Phase: 1
Repro: not yet run; it needs forge-et-fe#337’s contact page. Open /contacts (its rows load), open a contact with every list read failing, go back, then open a contact and go back again.
Expected: the list’s cached rows stay on screen, marked stale, with Retry (coding-standards §3.2).
Actual (from the code): BUG-270’s cause on the contact list. main has no contact page, so the defect is not on main; forge-et-fe#337 adds contacts/:id as a sibling route of contacts (src/app/router.tsx), so opening a contact unmounts the list, and useContactsList keeps its last-good snapshot in a ref, one per mount. A key whose refresh failed keeps status: 'error' with its data, and on return the new mount has no snapshot and no fresh data, so the view is error with no rows. Raised on forge-et-fe#337.
Fix: BUG-270’s: in useContactsList, when the snapshot is empty and there is no fresh data, take the key’s own cached data (query.data that is not placeholder data, with rows and not past the last page) as the snapshot.
Regression test: BUG-270’s component tests (AccountListScreen.test.tsx, “back on the list from an account’s page (BUG-270)”), written for the contact list against a route for contacts/:id.
Correction (2026-10-28): the defect is on main now: forge-et-fe#337 merged (a5224f39) with the contact page as a sibling route, and useContactsList had no fallback to the cached data.
Fixed as written. ContactsScreen.test.tsx’s “back on the list from a contact’s page (BUG-273)”: the cached-rows case failed without the fix (contacts-stale not found) and passes with it; its neighbour (the re-read fails on return while the key still has good data) passes either way, as BUG-270’s does.
Resolved: forge-et-fe#356, 2026-10-28 (Tharinda).
Lesson: L-238 — the third instance, after BUG-267 and BUG-270, as that lesson foresaw. Added to L-238’s sources.
BUG-272 — No screen can set an account’s website: the create and edit forms leave it out, yet the account page shows it and email templates fill it in
Affects: F1-b (the account create and edit forms, forge-et-fe src/features/accounts/AccountFormPage.tsx, accountWriteApi.ts), F1-d (the profile grid’s Website row), F20-d (the Company website merge field) · Found in: F38-g WI-4.2’s setup on dev, 2026-10-27 · Severity: medium — a user cannot record or correct an account’s website, so the account page shows — and the Company website merge field stays empty for every account created in the app · Phase: 1
Repro: on dev, as a user whose role holds account.manage, open Accounts → New account; then open any account → Edit account.
Expected: both forms have a Website field, since the account page shows the website (F1-d FR-V4) and email templates fill it in (company.website, F20-d).
Actual: neither form has one. The F1-b spec’s create form (FR-C1) lists the fields without the website, and the edit form (FR-C2) covers the same fields, so the code does what the spec says. The backend accepts a website on create (AccountCreateRequest) and on update (AccountPatchRequest), but nothing calls it with one — accountWriteApi.ts leaves it out, and the Pipedrive import (AccountMigrationService) does not set it — so only a direct API call can.
Fix: add an optional Website field to the account create and edit forms, sent as website on POST and PATCH /api/v1/accounts, with an F1-b spec revision adding it to FR-C1. Whether the Pipedrive import should also carry the website is a separate question for F4.
Regression test: a form test that fills Website on create and on edit and asserts the request carries website, and a browser test that the account page then shows it as a link.
Fixed as: an optional Website field (type="url", account-form-website) after Est. LTV and currency on both forms, sent trimmed as website on create and on edit only when it changed (F1-b Rev 10). F1-a’s patch reads a null website as “leave unchanged” and has no way to clear one, so emptying a saved website is refused before any request — “Enter a website. The website can be changed but not removed” — as the estimate already is. The backend has no length or format rule on the account website (website TEXT, V15), so the form adds none; the account page links it only for an http(s) value (websiteHref). Tests: accountForm.test.ts, NewAccountDialog.test.tsx, a new AccountFormRoute.edit.test.tsx case, and AC-9’s browser test, which now fills Website and checks the account page’s link.
Resolved: forge-et-fe#356, 2026-10-28 (Tharinda).
Lesson: none — one-off: a spec’s form left out a field the API, the record page and a merge field already used; no other bug has this cause.
BUG-271 — At 768px, with the account list scrolled to its end, the pinned Name column covers part of the Country sort button
Affects: F1-b (forge-et-fe src/features/accounts/AccountsTable.tsx, src/ui/pinnedFirstColumn.ts) · Found in: the browser gate on forge-et-fe#337 (69e331a8), 2026-10-27 · Severity: low — the button is only partly covered while the table is scrolled sideways at tablet width; sorting still works from its visible part or after scrolling back · Phase: 1
Repro: ET_E2E_WORKERS=2 pnpm exec playwright test e2e/browser/size-matrix --grep=accounts-pinned --project=size-768x1024-chromium with the BUG-271 mark removed from e2e/pageStates.ts → axe target-size on account-list-sort-country. Also accounts-archiving at 768 Chromium, and accounts-pinned at 768 WebKit.
Expected: every visible control keeps the 24px target floor (design-system §9), the table scrolled or not.
Actual: the size matrix scrolls the account table to its end before the page check. The first column is pinned from bp/tablet up (design-system Rev 68), so the Country header’s sort button ends partly under the sticky Name column, with less than 24px of it uncovered. With real data it scrolls fully under the column, which axe does not count; with the pinned sample names, whose widths differ, only part is covered. forge-et-fe#337’s narrower ⋯ actions column changed the table’s width, which is why main does not fail.
Fix: not decided. A pinned first column covers part of some control at many scroll positions on any table; the other tables pass only because of where their columns end, so changing a width here would make the test pass without fixing anything. The real choices are design ones: accept it as a trade-off of the pinned column (an entry in e2e/support/axeExemptions.ts, with the lead’s agreement, since the pin at tablet width is his decision of 2026-10-21, design-system Rev 68), or pin the first column only from bp/laptop up. Nanndish decided on 2026-10-27 to ship forge-et-fe#337 with the mark and decide this with the UI/UX engineer. Remove the COUNTRY_SORT_UNDER_PIN mark in e2e/pageStates.ts in the PR that settles it.
Regression test: the size matrix’s accounts-pinned and accounts-archiving states at 768, unmarked.
Update, 2026-10-28: on forge-et-fe feature/ui-revamp every table takes its cell classes from the shared src/ui/dataTable.ts, which changed the column widths, and the mark went stale in Chromium at 768: both states passed there with the mark applied, which the size matrix counts as a failure. The mark is removed from accounts-archiving and narrowed to WebKit on accounts-pinned (e2e/browser/size-matrix/arrange.ts, e2e/pageStates.ts). The Windows PR profile runs no WebKit, so WebKit at 768 is still to be checked on a macOS run; the cause and the design choice above are unchanged.
Update, 2026-10-28 (later): WebKit runs on Windows too, so it was run there: ET_BACKEND_PORT=8081 pnpm exec playwright test e2e/browser/size-matrix --project=size-768x1024-webkit --project=size-1024x768-webkit --grep "accounts-pinned$|accounts-archiving$", with the mark removed, passed both states at both sizes. The last mark and its COUNTRY_SORT_UNDER_PIN helper are removed on feature/ui-revamp (forge-et-fe#355, d0fecebb), so the size matrix’s regression rows above run unmarked. The bug stays open: the shared column widths moved the Country button out from under the pinned column at these sizes, but a pinned first column can still cover part of a control at some scroll position on any table, and the design choice above is still to make.
BUG-270 — Back on the account list from an account’s page, a failed re-read shows the error panel with no rows although the cache holds that list
Affects: F1-b (forge-et-fe src/features/accounts/accountsQueries.ts, useAccountsList’s snapshotRef) · Found in: fixing BUG-267 in forge-et-fe#345, 2026-10-27, from reading the code · Severity: low — rows the client already holds are hidden behind an error panel until Retry; nothing is changed or lost · Phase: 1
Repro: not yet run. Open /accounts (its rows load) and let a refresh of that list fail, so it shows its rows marked stale; open an account, then go back to the list with the next list read failing too.
Expected: the list’s cached rows stay on screen, marked stale, with Retry (coding-standards §3.2).
Actual (from the code): BUG-267’s cause on the account list. The last-good snapshot is a ref, one per mount, and the account’s page is a sibling route of the list (src/app/router.tsx, accounts/:id), so opening an account unmounts the list. A key whose refresh failed keeps status: 'error' with its data, and on return the new mount has no snapshot and no fresh data, so the view is error with no rows. The contacts list has no page of its own on main; forge-et-fe#337 adds contacts/:id, which gives useContactsList the same gap.
Fix: BUG-267’s: in useAccountsList, when the snapshot is empty and there is no fresh data, take the key’s own cached data (query.data that is not placeholder data, with rows and not past the last page) as the snapshot; the same in useContactsList once a contact page is a sibling route.
Regression test: BUG-267’s component test (LeadsScreen.test.tsx, “back by ← Leads to a list whose refresh failed …”), written for the account list against a route for accounts/:id.
Resolved: forge-et-fe#349, 2026-10-27. When the new mount has no snapshot and no fresh data, useAccountsList takes the key’s own cached data (not placeholder data, with rows, not a page past the last) as the snapshot, so the view is stale over the cached rows. Fails before: in AccountListScreen.test.tsx, “back on the list from an account’s page (BUG-270)” › “a list whose refresh failed shows its cached rows, marked stale, with Retry” failed against main’s accountsQueries.ts; its sibling, a re-read that fails only on return, already passed on main and stays as a guard. With the fix the file passed 96 of 96, and accountsQueries.test.tsx 41 of 41. Checked in Chromium against the local backend (a throwaway spec, not committed): on /accounts, an account opened twice with every list read failing showed, on the second return, the error panel with no rows on main’s code, and the 25 cached rows under the stale notice with the fix. The contact list’s copy arrives with forge-et-fe#337’s contact page: BUG-273.
Lesson: L-238 — the second instance, as that lesson foresaw.
BUG-269 — On an archived or converted lead, the loaded Approve & send leaves its toast’s nodes in the heading row’s actions group, which then adds a gap
Affects: F11-h, F11-f (forge-et-fe src/features/aiDrafts/LeadApproveSendAction.tsx, the Toast it hosts beside the button; src/features/leads/LeadActionsPlace.module.css, .inline:empty) · Found in: frontend-reviewer’s review of forge-et-fe#345 (BUG-265’s fix), finding 1, 2026-10-27, from reading the code; main behaves the same · Severity: low — Edit and Restore move left by one gap when the action’s chunk arrives; nothing is changed or lost · Phase: 1
Repro: not yet run. As a role holding email.manage and lead.manage, open an archived (or converted) lead and watch the heading row as Approve & send’s chunk loads.
Expected: a role whose action shows no button leaves the actions group :empty, so the group takes no place in the row and adds no gap (BUG-012), before and after the chunk loads.
Actual (from the code): LeadApproveSendAction hosts its outcome Toast for as long as the role held both keys (hostsToast), whether or not the button shows, and Toast renders two nodes — the live region and the [popover] — inside the action’s slot. On an archived or converted lead the stand-in from BUG-265 renders nothing, so the group is :empty until the chunk arrives; then the toast’s nodes make it a zero-width item in the heading row’s .actions, which adds one gap, and the controls before it move.
Fix: decided by the lead (Tharinda, 2026-10-27): render the action’s Toast outside the group, through a portal to document.body (the popover is in the top layer and the live region is out of flow, so nothing on screen moves). The other candidate, hosting the outcome toast somewhere the lead view owns, is not taken. The toast’s caller rule in Toast.tsx (mounted across its open flip) still holds.
Regression test: in src/app/leadPanels.approveSend.test.tsx, after the chunk loads, the group of an allowed role on an archived lead is still empty.
Resolved: forge-et-fe#349, 2026-10-27. LeadApproveSendAction renders its Toast through createPortal(…, document.body); the toast keeps its lifetime (mounted from the first allowed render until the action unmounts), so Toast.tsx’s caller rule still holds. Fails before: the two new cases in leadPanels.approveSend.test.tsx (an archived and a converted lead, after the chunk loads) failed against main’s action; with the fix that file, LeadApproveSendAction.test.tsx and aiDraftFlow.test.tsx passed 46 of 46. Checked in Chromium against the local backend (a throwaway spec, not committed; the lead’s read answered as archived, the action’s module held until the row was measured): on main’s code the group held 2 nodes after the chunk loaded and Restore moved from x 1061.45 to x 1052.45; with the fix the group held none, Restore stayed at x 1061.45, and the toast’s live region was a child of <body>.
Lesson: L-239 (new) — a place that collapses with :empty is kept open by any node a filler renders, an invisible one included; BUG-012 was the first instance.
BUG-268 — Flaky test: the F26-d price-email browser tests find no prices when a browser gate run crosses midnight (Windows)
Affects: F26-d (forge-et-fe e2e/fixtures/priceEmail.setup.ts, which seeds sales E and C and their prices from the server’s Colombo “today” once per Playwright run; the tests in e2e/browser/price-email/) · Found in: F3-a wave 1’s second overnight pnpm e2e:gate run on forge-et-fe#296 (6db65d72), 2026-10-26 to 27; the setup ran at about 23:50 and the price-email tests at about 00:05 (the machine’s clock is UTC+05:30) · Severity: low — a test-data defect; the product is unaffected, and a gate run that does not cross midnight passes · Phase: 1
Repro: start pnpm e2e:gate in forge-et-fe late enough that the price-email tests run after midnight Colombo time (on Windows with 3 workers, about 15 minutes after the start).
Expected: the price-email tests find the prices the setup published, whatever the time of day.
Actual: every price-email test that ran after midnight failed — eleven of them, across price-email-a11y, price-email-dialog, price-email-layout and price-email-send — with the dialog in its no-prices state and Continue disabled. The setup had dated sale C, its publish and its self-check to the day before. The same tests passed in the next full run, begun at 00:25 on the same machine. The exact rule that hides the prices on the next day was not traced.
Fix: decided by the lead (Tharinda, 2026-10-27): pnpm e2e:gate refuses to start a run that would cross midnight Colombo time (Asia/Colombo, whatever the machine’s own time zone), and the refusal says when a run can start. The other candidates, a server clock the test profile can pin and setup prices that also cover the next day, are not taken: each could be proved only by a run that crosses midnight. Until the fix merges, start a full gate run early enough to finish before midnight Colombo time.
Regression test: unit tests of the gate’s start check: a start whose run would end after midnight Colombo time is refused, and one that would end before it is allowed, with the clock and time zone faked.
Resolved: forge-et-fe#349, 2026-10-27. pnpm e2e:gate now runs scripts/gate-start-check.mjs, which takes the run’s scope the way the gate does, estimates the run at 90 minutes for the full profile and 45 for the PR profile (or the last run of that profile in the checkout, when longer), refuses with exit 2 when the run would end after the next midnight in Asia/Colombo, saying when a run can start, and otherwise runs scripts/browser-gate.mjs with the same arguments. The check is a CRM-only entry script rather than a change under scripts/browser-gate/, whose files forge-et-pwa carries as byte copies. The nightly run calls scripts/browser-gate.mjs directly at 02:00, so it is not checked. Regression test: src/test/scripts/gateStartCheck.test.ts, 26 of 26, with Vitest’s fake clock and the machine’s time zone set to Colombo, UTC, New York and Kiritimati, including a start whose run ends inside one UTC day and still crosses midnight in Colombo; the file does not exist on main, where nothing refuses such a start.
Lesson: none — one-off: the first fixture found to depend on a run staying within one day, and the gate now refuses the run rather than the fixture being changed; a second day-bound fixture would make it a lesson.
BUG-267 — Back on the Leads list from a lead’s page, a failed re-read shows the error panel with no rows although the cache holds that list
Affects: F11-b (forge-et-fe src/features/leads/leadsQueries.ts, useLeadsList’s snapshotRef) · Found in: frontend-reviewer’s delta review of forge-et-fe#344 (round 2, finding 2), 2026-10-26, from reading the code; main behaves the same · Severity: low — rows the client already holds are hidden behind an error panel until Retry; nothing is changed or lost · Phase: 1
Repro: not yet run. Open /leads (its rows load), open a lead, then go back by ← Leads with the next list read failing (or after a refresh of that list failed).
Expected: the list’s cached rows stay on screen, marked stale, with Retry (coding-standards §3.2: a failed refresh does not blank good data).
Actual (from the code): the last-good snapshot that stale shows is a ref, one per mount, and the lead’s page is a sibling route of the list (src/app/router.tsx, leads/:id), so opening a lead unmounts the list. On return the key’s failed re-read has no fresh data and the new mount has no snapshot, so the view is error with no rows.
Fix: the candidate the review named: when the snapshot is empty and the key is in error, use the key’s own cached data (query.data that is not placeholder data) as the snapshot.
Regression test: a component test that returns to the list from a lead with the list read failing, and expects the cached rows under the refresh-failed notice.
Resolved: forge-et-fe#345, 2026-10-27. When the new mount has no snapshot and no fresh data, useLeadsList takes the key’s own cached data (not placeholder data, with rows, not a page past the last) as the snapshot, so the view is stale over the cached rows; while the return’s re-read runs it shows them refreshing. Fails before: LeadsScreen.test.tsx “back by ← Leads to a list whose refresh failed shows its cached rows, marked stale, with Retry (BUG-267)” failed against main’s leadsQueries.ts (no leads-stale); its sibling, a re-read that fails only on return, already passed on main (that key still held fresh data at the new mount) and stays as a guard. With the fix the leads tests passed 695 of 695. Checked in Chromium against the backend: on /leads?status=all, a lead opened twice with every list read failing showed, on the second return, the cached rows under the stale notice; on main’s code the stale notice was missing. The account list has the same gap: BUG-270.
Lesson: L-238 (new) — a list’s last-good snapshot held for one mount is lost when opening a record unmounts the list; BUG-270 is the second instance.
BUG-266 — On the account and contact lists, a filter change back to a list that failed earlier shows the old failure, busy, while it reads
Affects: F1-b, F1-f (forge-et-fe src/features/accounts/accountsQueries.ts, lastErrorRef and retryingFailedKey in useAccountsList; src/features/accounts/contactsQueries.ts, the same rule in useContactsList) · Found in: frontend-reviewer’s delta review of forge-et-fe#344 (round 1, finding 1), 2026-10-26 · Severity: low — a failure message that no longer applies shows, with a busy Retry, for the length of one read; nothing is changed or lost · Phase: 1
Repro: not run on these lists. The same rule, copied into the Leads list in forge-et-fe#344’s first fix, failed this component test: the first read fails; choose a Country (that read succeeds); then Clear filters with the read held, using a query client with the app’s five-minute gcTime (the test client’s gcTime: 0 drops the failed key and hides it). While the read is held, the failure panel showed over the kept rows. The account and contact lists’ rule is the same code.
Expected: a failure is kept on screen only through its own Retry (lesson L-183); any other re-read of a key that failed earlier shows the kept rows refreshing, or the skeleton.
Actual: the rule keeps each key’s last failure and shows it as retrying on any re-read of that key: TanStack v5 puts a key with no data back to pending with its errorUpdatedAt kept, which the rule reads as a Retry. A key that has data and whose refresh failed keeps status: 'error' while it re-reads, and the error branch shows it as retrying: query.isFetching on any re-read too (accountsQueries.ts and contactsQueries.ts, the query.isError branch).
Fix: use src/lib/hooks/useRetryHold.ts with hasData set to “fresh data for the current key”, held from the Retry press, and set retrying only from the hold, as useLeadsList does since forge-et-fe#344; the rejected-request state of both lists stays as it is.
Regression test: the Leads tests (LeadsScreen.test.tsx, “a list that failed before and is read again by a filter change …” and “a list whose refresh failed, read again after a filter change and back …”), written for each list.
Resolved: forge-et-fe#345, 2026-10-27. useAccountsList and useContactsList hold a failure through useRetryHold, only from a refetch (refetch calls hold(): Retry and Reset list on both lists, and a row action’s re-read on the contact list), never on a filter change, with hasData the current key’s fresh data, and set retrying only from the hold; lastErrorRef and retryingFailedKey are gone. A 400 is never held (isError is passed as false for it), so the rejected-request state is unchanged. In contactsQueries.ts only useContactsList and its imports changed, because forge-et-fe#337 edits the same file. Affects: corrected from F1-a to F1-b, which owns the account list. Fails before: the two new tests in each of AccountListScreen.test.tsx and ContactsScreen.test.tsx (“… shows the kept rows while it reads, not the old failure (BUG-266)”) failed against main with the failure panel on screen; with the fix the src/features/accounts/ and src/lib/hooks/ tests passed 1003 of 1003, the BUG-208 Retry tests included. Checked in Chromium against the backend: on /accounts, after the list’s refresh failed and a Customer type filter was set and cleared with the read held, the kept rows showed with Updating…; on main’s code the stale panel showed with a busy Retry.
Lesson: L-183 — the failure was held for any re-read of a failed key rather than from the Retry press; L-183 gains an update saying so, and BUG-266 joins its sources.
BUG-265 — On a lead’s page, Approve & send loads after the heading row and moves Edit and Archive / nurture to the next row
Affects: F11-h, F11-f (forge-et-fe src/features/leads/LeadActionsPlace.tsx:99, the actions place’s <Suspense fallback={null}>) · Found in: BUG-260’s trace under forced load, and frontend-reviewer’s review of forge-et-fe#344 (finding 5), 2026-10-26 · Severity: low — a click made as the action arrives can land where Archive / nurture was; seen by a browser test on a loaded machine, not reported by a user · Phase: 1
Repro: macOS, 8 busy loops beside the run: pnpm exec playwright test e2e/browser/leads/lead-detail-status.spec.ts --project=chromium -g AC-20 --repeat-each=6 --retries=0 --trace=on on main (b0bebb28) failed once in the 403 row. Its trace’s screencast shows the heading row with Edit and Archive / nurture beside the heading, then, 128 ms later, Approve & send added and Edit and Archive / nurture moved to a row below the heading.
Expected: the heading row keeps its shape while its lazily loaded action arrives, so a control does not move under the pointer (L-208).
Actual: the action’s place renders nothing until its chunk loads (fallback={null}, kept so a role without the action leaves the group :empty, BUG-012), and the row re-wraps when it arrives.
Fix: decided by the lead (Tharinda, 2026-10-27): a fallback the size of the action, hidden and inert, rendered only where the action will show (a role holding email.manage and lead.manage, a new or approved lead), so the row takes its final shape at the first paint and an empty group still collapses.
Regression test: a component test that the heading row’s group holds a placeholder of the action’s size while its chunk is held, and a browser check that Archive / nurture does not move between the view’s first paint and the action’s arrival.
Resolved: forge-et-fe#345, 2026-10-27. src/app/leadPanels.ts puts LeadApproveSendPlace in the actions list: the lazy action inside its own Suspense whose fallback is LeadApproveSendPlaceholder (src/features/aiDrafts), the same shadcn Button with the same icon and label, hidden and inert, rendered only where the action will show, as ComposeAiDraftPlaceholder does for AI draft. The placeholder’s module also owns the rule and the label (useApproveSendAllowed, approveSendLabel), which LeadApproveSendAction now uses. Tests: src/app/leadPanels.approveSend.test.tsx (the group empty for a role without either key and for an archived or converted lead; the stand-in for a new and an approved lead; the arriving button drawn with the stand-in’s classes) and lead-detail-status.spec.ts “BUG-265: …”, which holds the action’s module and compares Archive / nurture’s box before and after it arrives. Fails before: against main’s composition 4 of the 8 component tests failed and the browser test found no stand-in; a copy of it without that assertion measured Archive / nurture moving from x 1010, y 91.6 to x 462.6, y 131.1 in Chromium. With the fix both pass.
Lesson: L-208 — a lazily loaded control rendered into a settled row with a null fallback; BUG-265 joins its sources.
BUG-264 — After a capture-app deploy, the new service worker’s cache holds no shell until a second online page load
Affects: F38-d (forge-et-pwa public/sw.js, the install and activate handlers) · Found in: frontend-reviewer’s review of forge-et-pwa#95 (BUG-258’s fix), finding 2, 2026-10-26 · Severity: low — opening the app with no connection is not a V1 requirement (PRD Rev 64); an offline launch between the two loads fails, and nothing is lost · Phase: 1
Repro: not run. Open the capture app online, so the worker is installed and the shell is cached. Deploy a build that changes the app. Open the app online once (the old worker serves this load and the browser installs the new worker, which takes over and deletes the old cache). Then go offline and open the app.
Expected: the app opens from the new cache.
Actual (from the code): the first online load after a deploy runs under the old worker, which stores the new index.html and its entry files in the old cache. The new worker skips waiting, and its activate deletes that cache, so the new cache starts empty and gains the shell only at the next online page load. An old tab left open across the update also loses its offline copies of chunks it has not loaded yet. This already happened on every production release, which changes the version in the cache name; since BUG-258’s fix (forge-et-pwa#95) the name also carries a build hash, so it happens on every dev deploy that changes the app too.
Fix: install stores the new shell in the new cache before the worker takes over: / and every file the build put in dist/assets/ (source maps excluded). The page’s entry files alone would not open the app offline, because the capture screen is a lazily loaded chunk (CaptureRoute-*.js) that only the entry script names. The list is written into sw.js by the build step that already substitutes the version and hash (vite.config.ts), so the worker keeps no hand-written precache list. Every file is fetched with cache: 'no-cache' and must pass shouldStore (a 200, and no HTML under a file’s key) before anything is stored; one that does not fails the install, so the browser keeps the old worker and its cache and tries the update again on a later load, rather than taking over with an incomplete cache. scripts/validate-pwa.mjs checks that the built list names exactly the files in dist/assets/.
Regression test: a worker test, in src/lib/sw/swParity.test.ts, that the shipped worker’s install leaves / and the build’s files in the new cache, and stores nothing when one file answers with the HTML fallback.
Resolved: forge-et-pwa#96, 2026-10-27, as the Fix: line says, with one check added for the deploy order: the deploy uploads sw.js before index.html, so for a moment the distribution can answer / with the previous build’s page, and install stores the page only when every /assets/ file it names is on the worker’s list. The first version compared a hash of the page’s bytes with the build hash; frontend-reviewer found that a wrong hash at build time, or an antivirus or proxy that adds to the HTML, would then fail every install and keep the old worker for good, so it checks the files the page names instead. scripts/validate-pwa.mjs also fails a build whose page would fail that check, whose worker no longer carries the same page pattern, or which writes a file of a type the worker does not serve. Fails before: the five install tests in swParity.test.ts failed against main’s worker, the old-page test fails with checkPage removed, and the no-built-file test fails without its rule; with the fix pnpm exec vitest --run src/lib/sw passed 50 of 50. Checked in Chromium with two builds served as the distribution serves them (a missing file answered with index.html, status 200): build A loaded online, replaced by build B, one online load, then offline. With main’s worker no cache was left after that one load and the offline load failed (net::ERR_FAILED); with the fix the new cache held / and all five of B’s files and the offline load showed B’s heading. With B’s worker and files but A’s index.html, B’s worker did not install and A’s cache stayed.
Lesson: none — one-off: the only cache this project replaces on a new version is the capture app’s worker cache, and its install now fills the new cache before activate deletes the old one. L-099 (every cache names the event that clears it) is about when a cache is cleared, not about filling its replacement first, and no other Lesson: line names this cause.
BUG-263 — check-done-acs.py passes an unticked criterion whose own text contains “*Not met”
Affects: F38-f (.claude/lib/check-done-acs.py, unticked_without_note; F38-f spec AC-6) · Found in: F38-f-WI-1.1’s check-done-acs.py --feature F38-f run, 2026-10-26 · Severity: low — at close, one criterion can stay unticked without the check failing; nothing in the product is affected · Phase: 1
Repro: on feature/F38-f-WI-1.1-read-checks (6c6f4eac), python3 .claude/lib/check-done-acs.py --feature F38-f exits 1 and lists AC-1 – AC-5, AC-7, AC-8, AC-10 and AC-11, but not AC-6, which is unticked. AC-6’s own wording ends “an open Sonnet 5.5 row is cited by a Not met note on the parent’s AC-29 first two bullets.”
Expected: AC-6 is listed until it is ticked or carries its own *Not met: …* note.
Actual: the check treats any *Not met anywhere in the bullet as the note (NOTE not in " ".join(bullet)), so a criterion that mentions a Not met note in its requirement passes unticked.
Fix: match only a note that starts a sentence or a continuation line in the form .claude/rules/specs.md gives (*Not met: <why>, <where recorded>.*, for example \*Not met: rather than *Not met), and add a self-test case with a criterion whose text mentions a Not met note.
Resolved: forge-et-harness#966, 2026-10-26. Matching only \*Not met: would have failed two notes already in specs, F1-a AC-29’s “Not met by design…” and F11-c’s “Not met (in part):…”, so a note is instead *Not met followed by anything but another *, outside inline code: *Not met* is how a criterion’s text mentions a note, and no note is written that way. The self-test gains a case with a mention, a note quoted in code and the three note forms in use. Fails before: unticked_without_note on origin/main returned [] for “- [ ] AC-6 cited by a Not met note”; with the fix --self-test passes, a full run passes on all 47 done features (EXIT=0), and --feature F38-f lists AC-6 with the other unticked criteria.
Lesson: none — one-off: a substring check that took a mention of the note for the note itself. The nearest earlier case, L-234 (pgrep -f matching its own command line), is a different mechanism with its own fix; no other Lesson: line names this cause.
BUG-262 — On the Leads list, Show all leads, Go to page N and Retry drop focus to the page body when they remove themselves on their own press
Affects: F11-b (forge-et-fe src/features/leads/LeadsScreen.tsx: Show all leads in the empty-status state, Go to page N in the past-the-end state, Retry in the error state) · Found in: the review of forge-et-fe#338, 2026-10-23 — carried over unchanged from the old LeadListPane · Severity: low — a keyboard or screen-reader user loses their place; nothing is changed or lost · Phase: 1
Repro: on /leads with a status that has no leads, Tab to Show all leads and press Enter; the list shows every lead and document.activeElement is <body>. The same with Go to page N on /leads?page=99 and with Retry on a failed list read.
Expected: focus moves to a control that is still on screen — the list’s heading, or the table region — as the other list screens do when the control that held it goes away (WCAG 2.1 SC 2.4.3, F11-b NFR-7).
Actual: each control unmounts with the state it sits in, and nothing moves focus, so it falls to <body>.
Fix: after each of the three actions, once the new state has rendered, focus the list’s heading (LeadsScreen’s focusHeading) when focus was on the removed control. A component test per control asserts where focus lands.
Resolved: forge-et-fe#344, 2026-10-26. LeadsScreen uses useFocusAfterWrite scoped to the data region, as ContactsScreen does: Retry arms it once its re-read settles, and Show all leads and Go to page N arm it once the list for the URL they asked for has answered; focus moves to the <h1> only when it was lost with the control. The refresh-failed Retry (leads-stale-retry), which leaves the page the same way, has the same wiring. While either Retry reads, useLeadsList keeps its failure on screen through useRetryHold, held from the press for any failure with no fresh data (retrying, the Retry busy), so a failed retry leaves the same Retry in place and focused (L-183); any other re-read of a key whose last read failed — a filter change back to it — shows the kept rows refreshing, as on main, not the old failure (BUG-266 records the account and contact lists’ rule, which does not). useLeadsList also returns renderedAt, and its refetch resolves with the read’s update time. The arming for a press that moves the URL, which the account and contact lists each had a copy of, moved to src/lib/hooks/useArmOnSettle.ts, used by all three lists, with its own tests. Fails before: the first five focus tests failed against main’s LeadsScreen and leadsQueries, the two failed-retry tests against the first fix, which had no hold, the filter round-trip test against the second, which copied the contact list’s rule, and the round trip after a key’s own refresh failed against the third; with the fix the three leads test files passed 110 of 110, and useArmOnSettle.test.ts with the contacts and accounts list test files 176 of 176. Checked in Chromium against the backend: after Enter on Go to page N (/leads?status=all&page=999) and on Show all leads (an empty status, its list read routed empty), document.activeElement was H1#leads-heading.
Lesson: L-133 — a state change that removed the focused control with no focus target named, the case L-133 records for Retry and Go to page N on the contacts list; BUG-262 is added to its sources.
BUG-261 — Flaky test: BUG-228’s WebKit keyboard walk loses the second AI tag and the search, about 1 run in 4 on forge-et-fe#338’s branch (Windows)
Affects: F11-d (forge-et-fe e2e/browser/leads/lead-filters-keyboard-webkit.spec.ts:18, the walk in e2e/browser/leads/leadFiltersWalk.ts:63; the cause is in the shared src/lib/hooks/useListSearchParams.ts) · Found in: repeated runs of the spec while preparing forge-et-fe#338, 2026-10-23, and its full gate runs · Severity: low — a filter change made in quick succession after another can be lost; the gate fails about 1 run in 4 · Phase: 1
Repro: Windows 11, 16 cores, idle apart from the run. node node_modules/@playwright/test/cli.js test --project=chromium e2e/browser/leads/lead-filters-keyboard-webkit.spec.ts --repeat-each=20 (8 workers): on feature/shadcn-leads-table, 11 of 20 failed at first, then 3 and 5 of 20 after forge-et-fe#338’s own search-box fixes (53f6cdc1, 8b303074); on main (c342dc28), 1 of 20. In the gate it failed in 1 of the 4 runs on 99fcd7cb.
Measured again on Windows, 2026-10-26 (forge-et-fe#343, which does not change the Leads filters or useListSearchParams): the same command, one after the other on the same machine and backend: on main (326c481a) 7 of 20 failed, and on forge-et-fe#343’s 9c07224a 8 of 20 failed, all but one reading SPAM true null (one on main read SPAM true <token>, one on #343 null true null). It also failed in that PR’s full gate and again when re-run on its own. The rate on main is no longer the 1 of 20 above.
Measured again on Windows, 2026-10-27 (forge-et-fe#296, F3-a wave 1, which does not change the Leads screen or useListSearchParams): it was the only failure of the full gate on 6db65d72 (5,771 passed) and one of three in an earlier full run on the same commit. The same command, 20 repeats a run, against the same backend: on 6db65d72 11, 6 and 7 failed (24 of 60); on main as merged into it (b0bebb28, a separate worktree) 6, 5 and 3 failed (14 of 60), the last four runs alternating branch and main. Failures read SPAM true null, SPAM true <token> and once null true null.
Seen in Chromium too, on Windows, 2026-10-27 (forge-et-fe#348, the Email screen, which does not change the Leads screen or useListSearchParams): the Chromium run of the same walk, e2e/browser/leads/lead-triage-inbox.spec.ts:167 (BUG-228 (keyboard), walkAiTagFiltersByKeyboard), failed once and passed on its retry in the PR-profile gate on eca5f0fa (4,094 passed, 1 flaky). Machine load: 0.85 average, 0.91 peak, on 16 cores.
Expected: after Spam, a search, Duplicate and Existing customer? in quick succession, the URL reads aiTag=DUPLICATE&customerMatchProposed=true&q=<token>.
Actual: every failure since the fixes reads SPAM true null or SPAM true <part of the token>: the switch’s write started from a query string older than the search write and the Duplicate choice. Instrumented runs showed the Duplicate choice never reached the Select’s onValueChange.
Fix: not decided. The search box’s part was fixed in forge-et-fe#338 (it reset itself when its own debounced write landed, dropping keys typed meanwhile). What remains is a write that starts from a stale snapshot in useListSearchParams, which BUG-228 and L-201 cover; the slower table render gives it more room. An instrumented run logging requested.pending, latestKey.current and each adopted navigation at the switch’s write shows which path is stale.
Fix (decided 2026-11-06): the stale write was fixed by BUG-290’s forge-et-fe#372 (2026-11-02) and BUG-314’s forge-et-fe#411 (2026-11-05), which changed useListSearchParams after the last measurement above. What was left was the slow render: the Leads screen held its search box’s text, so every key press re-rendered the screen, its filter bar and its table. Move the text, the debounce and the rule that the box follows the URL into a LeadsSearchBox that owns them; Clear filters empties it through a counter prop.
Regression test: forge-et-fe src/features/leads/LeadsSearchRenders.test.tsx counts the table’s and the filter bar’s renders while a word is typed (5 and 5 on the earlier code, 0 and 0 after), and a LeadsScreen.test.tsx case covers Clear filters pressed while a search waits on its debounce (it fails without the counter). The browser walk itself, lead-filters-keyboard-webkit.spec.ts, stays the end-to-end check.
Resolved: 2026-11-07 — fixed in forge-et-fe#437 (2c5a05ce; review fixes 7d1b342c, 7c9d4085), a product defect. Measured on Windows 11 (16 cores), backend on main 5a14884b, 3 workers, no retries:
- On
main4a91ed9c, after #372 and #411: the walk failed 1 of 60 at first (2026-11-06), every failure a test timeout or the old URL symptom, and many more in rounds run beside another session’s backend build. A trace showed typing the 17-character token took 11.8 s of the 30 s. Typing the token into the Leads search box took a mean of 6.9 s (5.9–8.0 s, 8 runs) against 0.38 s for a bare input appended to the same page, so the cost was React work on each key. - With the fix: typing took a mean of 1.1 s (0.26–3.4 s). Alternating
mainand the fix in one quiet window on 2026-11-07 (each round started only with no other test run or doc build on the machine and the CPU under 40%):mainfailed 9 of 60 (all 30 s test timeouts at the typing and key steps, in its first round), the fix 0 of 60; with two earlier quiet rounds, the fix passed 100 of 100. The Chromium walk (lead-triage-inbox.spec.tsBUG-228 (keyboard)) passed 20 of 20 on the fix. - Beside another session’s backend build (a Gradle test JVM or
javadocrunning, 2026-11-06), both versions failed many runs at every step, including page load (10–14 s for the list response); that is the overload outcome of.forge/test-strategy.md § Flaky Tests, not this test. Lesson: L-201 for the stale write (fixed by #372 and #411) — another instance of a write starting from a query string the screen has moved past. L-273 (new) for the slow render — a text box’s typed text lives in the box, not in the screen around it;frontend-revieweritem 10 points at it, and.forge/owed.mdholds the check of the other list screens that keep their search text the same way.
BUG-260 — Flaky test: lead-detail-status.spec.ts AC-20’s Archive click never lands under load, because the lead detail closes first (macOS)
Affects: F11-f (forge-et-fe e2e/browser/leads/lead-detail-status.spec.ts:269, “AC-20 (smoke): a real 404 after the lead left reach says it is no longer available, …”, project chromium) · Found in: pnpm e2e:gate on forge-et-fe#341 (e74f100e, a change to the Vite config’s imports only — no app code), macOS, 2026-10-26 · Severity: low — the gate fails; the product is not known to be affected · Phase: 1
Repro: under load. pnpm exec playwright test e2e/browser/leads/lead-detail-status.spec.ts --project=chromium --repeat-each=5 --retries=0 with 8 busy loops (one per core) beside it, at a load average of about 23 (2.9 per core): 39 of 40 passed and AC-20 failed once (2026-10-26). Without the busy loops: 40 of 40 passed, AC-20 in about 1.5 s. The gate’s own run: Machine load: busy 0.15 at start, 0.99 peak, 0.88 average; load average 4.50 per core peak (8 cores, macOS), 5628 passed, 1 failed (this test).
Expected: after the test reassigns the lead to an executive (patchLead, so it leaves the user’s reach), the click on Archive / nurture sends the archive request, the backend answers 404, and the app shows the “no longer available” notice and returns to /leads?status=all.
Actual: locator.click: Test timeout of 30000ms exceeded at line 283, the call log repeating “element was detached from the DOM, retrying”, and page.waitForResponse for the archive POST never resolved. In the failure’s page snapshot the app is already on the leads list with the “Leads” heading focused and the lead still listed, so the lead detail, with its Archive button, closed before the click landed and no archive request was sent.
Fix: not decided. First suspect, not confirmed: between patchLead and the click, a read of the lead (the detail’s own query, or a re-read another component starts) gets the 404 the reassignment causes, and the app’s “lead gone” handling closes the detail; on a loaded machine that read lands before the click. Find which request closes it (a trace of a failing run), then make the test wait for nothing to be in flight before reassigning, or hold the detail’s reads until the click; not by lengthening the timeout (.forge/test-strategy.md § Flaky Tests).
Regression test: the spec’s AC-20 row passes 40 of 40 with the forced-load repro above.
Trace (2026-10-26, macOS, 8 busy loops beside the run): -g AC-20 --repeat-each=6 --trace=on failed once, in the 403 row (lead-detail-status.spec.ts:247), with the same cause. Playwright found Archive / nurture visible and stable beside the heading, then waited 128 ms before it clicked; in that gap Approve & send (F11-h’s action, lazily loaded into the heading row with no loading marker) mounted and moved Edit and Archive / nurture to a row below the heading, so the click landed on empty space and no request was sent. The lazy panels’ first reads (/shares, /triage, /research) went out at the same moment: settleLeadDetail returns before the lazy panels have loaded and read. For the 404 row, holding the sharing panel’s /shares read until just after patchLead (a temporary copy of the test) failed 3 of 3 exactly as in the gate: SharingPanel hands a 404 on its read to onRecordGone, LeadPage’s out-of-reach handling returns to /leads, and the click finds its button detached.
Resolved: forge-et-fe#344, 2026-10-26 (test only). The AC-20 rows answer the lead’s AI summary and Research reads with the contract’s final not_analysed and not_researched answers (pinAiAnswers, each body checked by its wire reader), so nothing re-reads and no AI read reaches the backend after the reassignment; and they call a new settleLeadForAction before they change the lead or click: settleLeadDetail, settleLeadPanels (the sharing panel past its first read, every panel place’s chunk loaded), both AI panels showing the pinned answers, and Approve & send in the heading row. Under the forced load (load average up to 63, 8 cores), lead-detail-status.spec.ts --repeat-each=8 passed 48 of 48, and the 404 and 403 rows --repeat-each=40 passed 80 of 80. The layout shift itself is filed as BUG-265.
Lesson: L-101 — the test clicked straight after a settle that did not wait for the lazily loaded parts of the view (BUG-150, BUG-151 and BUG-230’s cause again), here a heading-row action with no loading marker and a panel read that changes the view; BUG-260 is added to its sources.
BUG-259 — Flaky test: ScanResultQueueIT’s AC-3 dead-letter row ran out its wait again in a full build, on a machine within the load limit (macOS)
Affects: F38-c (forge-et-be src/test/groovy/com/empireteas/attachment/service/ScanResultQueueIT.groovy:696, “AC-3: a body that is not JSON is not acknowledged, reaches the dead-letter queue, and claims nothing”) · Found in: python3 scripts/local-build-evidence.py on forge-et-be#311 (8ed7cd69, a Spotless reformat of one test line in BuyingStandardPriceControllerIT), macOS, 2026-10-25 · Severity: low — one IT failed, which failed the build record; the product is unaffected · Phase: 1
A recurrence of BUG-195, closed wontfix as machine overload, whose resolution asks a recurrence on a machine within the limit to be filed as a new bug with the record’s load line. This build started within the limit: Machine load: busy 0.23 at start, 0.71 peak, 0.33 average; load average 1.56 per core peak (8 cores, macOS), the busiest processes at the peak being the build’s own java (222%), the Docker VM (95%), com.docker.backend (23%) and another session’s node (21%).
Repro: a full ./gradlew build on macOS; not reproduced on demand.
Expected: the poison message reaches the dead-letter queue within the 60-second PollingConditions wait.
Actual: SpockTimeoutError: Condition not satisfied after 90.42 seconds and 11 attempts (dlqCarries(marker) false). Eleven polls in 90 seconds is about 8 seconds a poll, against dlqCarries’s 1-second SQS wait, so LocalStack answered slowly rather than not at all (BUG-195’s run took about 90 seconds a poll). In the same run the AC-4 COMPLETED/UNSUPPORTED row passed but took 413.6 seconds, against 30-second waits, so the slowness was in LocalStack calls outside the polls as well. The build ran 58 minutes; earlier full builds of the same branch that day (93978b7, 0e602f9) passed the spec. Alone right after, ./gradlew test --tests 'com.empireteas.attachment.service.ScanResultQueueIT' --rerun passed 14 of 14 (EXIT=0), the AC-3 row in 27.6 seconds.
Fix: decided 2026-11-13 (Tharinda), after the reproduction below: the row sets the queue’s retry limit to 2 for its own message and puts 8 back afterwards. The row proves the message is not acknowledged and reaches the dead-letter queue, not how many receives that takes; at 8, every receive is a LocalStack round trip plus the 3-second visibility timeout, so a slow LocalStack’s delay is multiplied by eight. The queue keeps 8 for the other rows, because AC-2 needs that many deliveries to commit its row in time.
Regression test: the AC-3 row run alone while the test run’s LocalStack container is frozen 7 seconds of every 8 — failing on main, passing with the fix.
Triage (2026-11-13, .forge/test-strategy.md § Flaky Tests step 2):
- Recurrence. None since the report: in every nightly full run from 2026-10-27 to 2026-11-13 (the
nightly-logissue) the row took 27.5–30.1 seconds, except the night of 2 November, when the lid was closed and the Mac slept (937.6 s, forge-et-be#343). - Host load does not slow LocalStack. BUG-195’s three runs with a busy loop per core passed; limiting the LocalStack container to 0.25 CPU (
docker update --cpus) left the row at 36 seconds, and 0.1 CPU stopped LocalStack starting at all. - A slow LocalStack reproduces it. A helper froze the test run’s LocalStack container (
docker pause, thendocker unpause) for 7 seconds of every 8, starting once the spec’s queues existed, and the row was run alone (./gradlew test --tests 'com.empireteas.attachment.service.ScanResultQueueIT.AC-3: a body that is not JSON*' --rerun, forge-et-bemaindf279a0e): 3 of 5 runs failed withCondition not satisfied after 64.7–65.4 seconds and 8 attempts— about 8 seconds a poll, this bug’s signature — and 2 passed at 73 and 81 seconds. Freezing it 5 seconds of every 7 did not reproduce it. What slowed LocalStack in the 2026-10-25 build is not known; the build took 58 minutes against the 25–30 the nightly runs took then. Resolved: forge-et-be#422, 2026-11-13 (test only).ScanResultQueueITgainsMALFORMED_MAX_RECEIVE_COUNT = 2: the not-JSON row sets the scan-result queue’s redrive policy to it before it sends (SetQueueAttributes), and itscleanup:block waits for the queue to empty at that limit, then putsMAX_RECEIVE_COUNT(8) back (review fix13c846c7), so feature (a) and every other row still run on 8;setup()andcleanup()already wait for the queue to be empty. With the fix, the same stall passed 4 of 4 (the row 19.8–32.7 s), and the whole spec unstalled passed 14 of 14 (EXIT=0) with the row at 6.5 seconds. Lesson: L-301 — a test that waits on an emulator is reproduced by slowing the emulator, not the host, and takes no more round trips than its assertion needs. BUG-195 and BUG-285 (both closedwontfixwhen host load did not reproduce them) are added to its sources.
BUG-258 — The capture app’s service worker keeps one cache across dev deploys and can store the HTML fallback under a script’s key
Affects: F38-d (forge-et-pwa public/sw.js, the CACHE name at line 16 and the fetch handler at lines 74–96, with .github/workflows/build.yaml sync::prune; src/app/router.tsx:10, the lazily loaded NotFoundRoute) · Found in: frontend-reviewer’s review of forge-et-pwa#91 (F38-d wave 5), findings 1 and 2, 2026-10-25 · Severity: low — a tab left open across a deploy sees an error page on an unknown path, and the stale entries are for URLs no new build asks for · Phase: 1
Repro: not yet run. The capture app has deployed to dev since 2026-10-25 (forge-et-pwa build.yaml), so it can now be run there: open https://capture.dev.crm.empire.lk and leave the tab open; let a second deploy run; in the old tab, open a path that loads NotFoundRoute for the first time.
Expected: the route opens, or the app reloads onto the new build, and the worker stores no HTML under a script’s URL. Each deploy’s build starts from an empty cache.
Actual (from the code):
sync::prunedeletes the previous build’s hashed chunks. The distribution answers a missing object with/index.htmland status 200, and the worker caches any status-200 response for a shell extension whatever itsContent-Type. So the old tab’s chunk request fails to load, and the worker stores the HTML under that.jskey.RouteErrorElementrecovers withnavigate('/'), not a reload. This is BUG-139’s pattern, fixed forforge-et-feonly.- The cache name carries only
package.json’s version (et-capture-shell-__APP_VERSION__), which a dev deploy frommaindoes not change. The comment above it says “a deploy orphans the old cache”, which holds only for a release, so on dev every deploy reuses one cache and old hashed assets pile up on test phones. Production is not affected in the same way, because release tags bump the manifest version (.claude/rules/release-and-branch-protection.md). Fix: not started. In a laterforge-et-pwaPR: - cache a non-navigation response only when its
Content-Typeis nottext/html, and updatesrc/lib/sw/shouldCache.tswith it (scripts/validate-pwa.mjschecks that the two agree); - offer a reload on a failed chunk load, as BUG-139’s fix does in
forge-et-fe; - either put the build’s commit in the cache name, or reword the comment to say “a release”.
Regression test: a worker test that a status-200
text/htmlanswer to a.jsrequest is not cached, and a component test that a failed chunk load offers a reload, asforge-et-fe’ssrc/app/routerErrorElements.test.tsxdoes. Resolved: 2026-10-26 — fixed in forge-et-pwa#95, all three parts: public/sw.jsandsrc/lib/sw/shouldCache.tsgain ashouldStorerule: only a 200 is stored, and HTML only for a page load, under the one shell key.swParity.test.tsruns the shipped worker’s rule against the tested one and its fetch handler on a 200text/htmlanswer to a.jsrequest (nothing stored);scripts/validate-pwa.mjschecksHTML_TYPEand the new guards.RouteErrorElementoffers Reload the page when the route error is a failed chunk load.src/app/chunkLoadError.tsis a byte-for-byte copy offorge-et-fe’s, listed inscripts/shared-with-fe.jsonsopnpm check:sharedkeeps them the same.src/app/routerChunkLoad.test.tsxrenders the shipped routes withNotFoundRoutefailing as a pruned chunk does, ande2e/browser/shell/a11y.spec.ts’s route-error test now answers the capture chunk withindex.htmland status 200, as the distribution does, rather than aborting it.- The cache name is
et-capture-shell-<version>-<12 hex of sha256(dist/index.html)>, set byvite.config.ts. The entry chunk names every lazily loaded chunk and stylesheet, so any change to a bundled file gives a new cache, which the worker’sactivatedrops the old one for; a change topublic/alone keeps the cache.validate-pwa.mjsrefuses a build without the hash.
Checked in Chromium on the built app served with the distribution’s fallback: with NotFoundRoute’s chunk removed, /nowhere offered Reload the page (44px, one <main>), and with the chunk restored the click reloaded onto the route. The browser pane could not register a service worker, so the cache rule is proved by the worker tests only. The review’s finding that a new cache holds no shell until a second online load is BUG-264.
Lesson: L-194 — the worker and the browser test were both written against a refused request, not the distribution’s real answer to a pruned file (index.html, status 200). BUG-258 is added to its sources.
BUG-257 — The backend’s build record dirties its own working tree when it checks an earlier record, then refuses to record
Affects: forge-et-be scripts/local-build-evidence.py (the Test record check) · Found in: /forge-deliver F38-d --wave 4, the second build record on forge-et-be#306, 2026-10-23 · Severity: medium — a full ./gradlew build (31 minutes on 2026-10-23) runs and passes, then nothing is recorded, on every PR that already carries a record; the workaround costs a second full run · Phase: 1
Repro: on a PR whose earlier head already has a Test record comment, push a commit that changes a file the build reads, and run python3 scripts/local-build-evidence.py from a clean tree with no scripts/__pycache__/.
Expected: the build runs and the record is written on the new head.
Actual: the build passes (BUILD SUCCESSFUL in 31m 37s), then the script prints “The working tree changed while the build ran, so the build may not have been of the commit the record would name. Commit or stash the changes and run it again. Nothing was recorded.” and exits 2. git status shows only ?? scripts/__pycache__/fast-lane-scope.cpython-314.pyc: the script loads scripts/fast-lane-scope.py with importlib.util.spec_from_file_location (around line 249) to decide whether the earlier record can be carried, Python writes that file’s bytecode cache into the repository, and the script’s own clean-tree check then sees it. forge-et-be .gitignore has no __pycache__ rule. A PR’s first record never imports the module, which is why it does not happen there.
Fix: not started. Either set sys.dont_write_bytecode = True in the script before the import, or add __pycache__/ to forge-et-be .gitignore (check that the clean-tree test honours ignore rules). Workaround until then: run the record as PYTHONDONTWRITEBYTECODE=1 python3 scripts/local-build-evidence.py, after removing any scripts/__pycache__/.
Resolved: forge-et-be#316 (468d03a6; the review’s fix to the test, 53b10a1f), 2026-10-26, by the first option: the script sets sys.dont_write_bytecode = True before any import, which fixes it on every machine whatever its .gitignore. Loading the scope module moved into a fast_lane_scope(directory) helper. The self-test loads a copy of the module from a temporary folder, in a child Python with PYTHONDONTWRITEBYTECODE and PYTHONPYCACHEPREFIX removed, and checks no __pycache__ appears beside it. The review found the first version of the test passed without the fix wherever the workaround variable was set, as it was in the shell that first ran it. Fails before, with that variable still set: on a copy without the line, SELF-TEST FAIL [the scope module loads from the folder it is given, and leaves no __pycache__ beside it]: expected (0, 'True False'), got (0, 'True True'); with the fix self-test passed — 92 cases.
Lesson: none — one-off: the only module the record scripts load from their own folder; the frontends’ records are Node scripts, which write no cache into the checkout.
BUG-256 — A Pipedrive label longer than 39 characters makes the write run create a tag over F1-e’s 50-character name limit
Affects: F4-d (forge-et-be tag/service/TagService.java findOrCreate, called from migration/service/MigrationTagResolver.resolve), F1-e (the 1–50 character rule, enforced only in TagController, MAX_NAME_LENGTH) · Found in: the adversarial review of forge-et-be#310 (F4-d wave 2), finding I-1, 2026-10-24 · Severity: low — no real label is that long: the 2026-09-28 exports carry 3 distinct organisation labels and 6 distinct person labels (.forge/discovery/source-schemas/pipedrive-*-export-schema.md); a longer one would make one tag a person could not have created, nothing else · Phase: 1
Repro: not run. A synthetic export with an organisation labelled with 40 or more characters, then a dry run and a write run: the run creates Pipedrive: <label> (51+ characters) and reports nothing; the dry run does not flag it either.
Expected: the tag-name rule holds for every tag, whoever creates it; a label that cannot become a valid tag name is visible in the dry run, so a reviewed dry run does not turn into a surprise at write time.
Actual: the service does not check the length and the database has no CHECK on tag.name (V22), so the run stores it.
Fix: decided by the lead (Tharinda, 2026-10-27), after the deferral from #310 (Biseni’s instruction to finish the feature): the dry run and the write run leave such a label out, with an entry under values left out, so a reviewed dry run shows it before the write; and TagService.findOrCreate checks the length as a backstop. Refusing or warning in the dry run is not taken. Refusing only at write time would turn a reviewed dry run into a 500 with a full rollback.
Regression test: a dry run and a write run over a label of 40 characters, asserting the label is listed under values left out and that no tag name over 50 characters is stored; and a TagService.findOrCreate case refusing a name over 50 characters.
Resolved: forge-et-be#321, 2026-10-27. OrganizationMapper.labelTags, which the organisation, person and lead mappers all use, leaves out a label whose Pipedrive: <label> name would be over TagService.MAX_NAME_LENGTH (50) and returns it as a Values left out entry naming the file, the Pipedrive ID, the labels column and the label; DryRunEngine files the organisation’s and the lead’s entries. The write run runs the same engine on the same files, so it skips the same labels and its report lists them too. TagService.findOrCreate refuses a blank name or one over 50 characters after trimming (IllegalArgumentException), and TagController.MAX_NAME_LENGTH now points at the service’s constant. Lead labels are covered too, beyond the entry’s organisation and person labels, because they go through the same helper. Spec revisions: F4-a Rev 11 (FR-P1 item 10), F4-d Rev 2 (FR-T1). Fails before: with the production code set aside, MigrationTagResolutionIT’s new dry-run-and-write-run case and TagServiceFindOrCreateIT’s 51-character and blank refusals failed, 3 of 12 (EXIT=1); with the fix, the ten touched suites passed, 129 tests, 0 failures (EXIT=0).
Lesson: L-073 — a bound checked only on the REST route did not cover an in-process caller of the same service (“Check the other entry points”); L-073 gains an update saying it applies to any bound a route checks, and BUG-256 joins its sources.
BUG-255 — On an account or contact, the Add tag list sets its option count from the window’s width alone, so ten options can run off a short window
Affects: F1-e (forge-et-fe src/features/notesAndTags/RecordTagsSection.tsx:232, addTagOptions.ts ADD_TAG_OPTIONS_MAX and ADD_TAG_OPTIONS_MAX_REFLOW) · Found in: frontend-reviewer’s review of forge-et-fe#335 (BUG-206’s fix), finding 6, 2026-10-23 · Severity: low — arithmetic only so far; a short or zoomed window shows it, and no supported size is affected · Phase: 1
Repro: not yet seen in a browser. Open an account or a contact whose tag read holds more than ten tags, in a window at least 768px wide and under about 765px tall (for example 1280 × 720 or 1280 × 600), scroll the Add tag field to the window’s middle and open Add tag.
Expected: the options and the “Type to find RecordTagsSection lists ADD_TAG_OPTIONS_MAX (ten) whenever the window is bp/tablet wide or wider, and ADD_TAG_OPTIONS_MAX_REFLOW (seven) below that, with no height step. By BUG-206’s measurements in Chromium (a row 30.1px, the popover the rows plus 40.5px, 6.6px from the 44px field, the library’s 12px edge padding), ten options need a window about 765px tall with the field mid-window and seven about 584px, so the list can run off a shorter window.
Fix: mirror the lead’s rule (src/features/leads/addTagWindow.ts): ten from bp/tablet wide and a height that fits ten, seven from about 600px tall, and three below. The two features may not import each other (addTagOptions.ts’s header), so the window rule moves to src/lib or src/ui, or is copied with a pointer, in the fixing PR.
Regression test: browser checks on the account and contact views as lead-detail-a11y.spec.ts’s “the tag list fits a … window, with the field mid-window” cases do: the field mid-window at a height just under each step.
Measured with the old rule (2026-10-26, Chromium, 15 pinned tags): on an account with the field mid-window, the tenth option ran 28.3px past the bottom of a 1280 × 599 window; at 1280 × 720 the ten fitted. In the contact Modal, where the field sits where the dialog places it, the ten fitted at both sizes.
Resolved: forge-et-fe#344, 2026-10-26. The lead’s window rule moved from src/features/leads/addTagWindow.ts to src/ui/addTagWindow.ts, with a useAddTagOptionsMax hook (src/ui/useAddTagOptionsMax.ts) that LeadTagsSection and RecordTagsSection both call: ten options from bp/tablet wide and ADD_TAG_ROOMY_HEIGHT tall, seven from ADD_TAG_SHORT_HEIGHT tall, three below. ADD_TAG_OPTIONS_MAX_REFLOW and notesAndTags’ own copy of ADD_TAG_OPTIONS_MAX are gone. New e2e/browser/accounts/add-tag-window.spec.ts checks the account page, with the field mid-window, and the contact Modal, where the Modal places the field (it cannot scroll it to the middle), at 1280 × 860, 1280 × 600 and 1280 × 599; its first version, at 1280 × 860, 1280 × 720 and 1280 × 599, failed 4 of 6 with the old rule (ten options listed at the two shorter sizes), and all 6 pass with the fix; src/ui/useAddTagOptionsMax.test.ts checks the hook at each step.
Lesson: L-170 — the account’s Add tag rule was a copy of the lead’s, and BUG-203 and BUG-206 changed only the original; the shared rule now lives in src/ui. BUG-255 is added to its sources.
BUG-254 — A CRM dev deploy run stayed in its first S3 sync until the deploy job’s 10-minute timeout stopped it
Affects: F38-d (forge-et-fe .github/workflows/build.yaml, job deploy, step sync::assets) · Found in: F38-d-WI-4.2’s read of the build runs since F38-d’s wave 2, 2026-10-22 · Severity: low — sync::html never ran, so the bucket kept the previous index.html and the site served the earlier build; the next run deployed main; the run billed 16 minutes against the 2 estimated · Phase: 1
Repro: not reproduced. Run 39510630707, a push of 5b4a0138 on 2026-10-20 at 22:47 UTC: build succeeded; deploy ran artifact::download, configure::aws and identity::aws, then started sync::assets at 22:48:12 UTC and never finished it. Every other build run of forge-et-fe and forge-et-be since 19 Oct succeeded (environments-and-identity.md §4 “F38-d — the dev deploy pipelines, checked”, AC-12).
Expected: the staged sync of the bundle finishes in seconds, as in the other runs (each deploy job 13–26 s).
Actual: the run’s annotation reads “The job has exceeded the maximum execution time of 10m0s”; the job ran 22:48:04–23:03:04 UTC and ended cancelled. The job’s log is not kept: gh run view --job 118930653473 --repo altrium-io/forge-et-fe --log answers log not found: 118930653473, so where the sync stopped is unknown.
Fix: to find. If it happens again, keep the log before it expires (download it from the run page) and check whether one upload hung; a candidate change, not confirmed, is a shorter --cli-read-timeout and --cli-connect-timeout on the aws s3 sync and aws s3 cp steps, so a stalled request fails fast and is retried instead of holding the job for the whole timeout.
Not seen again (2026-11-08): the 60 most recent forge-et-fe build.yaml runs, back to 2026-10-29, all succeeded (gh run list -R altrium-io/forge-et-fe --workflow build.yaml -L 60).
Resolved (mitigated, cause not found): forge-et-fe#528 and forge-et-pwa#125, 2026-11-13. The capture app’s deploy has the same S3 steps, so both frontends get the change. In each deploy job, every S3 step and cache::invalidate gives up on a request after 10 s without a connection or 20 s without a reply, where the CLI defaults are 60 s for each. The CLI retries such a request up to five times (AWS_MAX_ATTEMPTS: "5", where the default is three; AWS_RETRY_MODE: standard, already the CLI v2 default, pinned). Each S3 step fails after timeout-minutes: 3, and the invalidation after 2. A hang now fails the run in minutes, with the step named, where it used to be cancelled after the job’s 10 minutes; a failure before sync::html still leaves the previous index.html serving. The first commit set --cli-read-timeout 60, the CLI’s own default, which changed nothing; review caught it (frontend-reviewer), and the fix is 20 s, so five attempts and their backoff (about 115 s) fit inside the step. No automated regression test is possible: no test in either repo reads build.yaml, and the workflow’s paths-ignore: .github/workflows/** keeps the PRs’ own merges from deploying, so the first ordinary merge to main after them runs the new steps. If a run hangs again, download its log from the run page before it expires.
Lesson: L-294 — before setting a client’s timeout or retry option as a fix, read its default, and bound each network step of a workflow with its own timeout-minutes, below the job’s.
BUG-253 — The lead research panel’s AI source label writes its date with a weekday
Affects: F11-c (forge-et-be LeadResearchService, the Research panel’s source label) · Found in: Tharinda, 2026-10-22, while fixing BUG-240 · Severity: low — cosmetic; once BUG-240 is fixed, the other three AI labels on the same lead view read <source>, d MMM and this one alone keeps the weekday · Phase: 1
Repro: on a lead whose research has finished, the Research panel’s provenance line reads from: public web search, Wed 21 Oct.
Expected: from: public web search, 21 Oct — design-system §5, AI provenance — all four elements, Source: what the output was made from, a comma, then the date as d MMM.
Actual: LeadResearchService.LABEL_DATE is EEE d MMM, as the F11-c Wave 1 contract’s AiProvenance table set it.
Fix: not started. Left out of BUG-240’s fix because forge-et-fe#329 (F11-c wave 1, open on 2026-10-22) asserts the weekday form in its seam spec (e2e/seam/lead-research.seam.spec.ts, the real backend’s string), its size-matrix pin and its fixtures; changing the backend first would fail that PR. Once #329 has merged: build the label with AiProvenance.Source.dated(LABEL_SOURCE, <date>) (added by BUG-240), update LeadResearchControllerIT’s formatter, the public web search example in forge-et-be docs/api/payloads.md and that file’s sentence naming F11-c’s label as the one exception (§ AI provenance), the does not yet (BUG-253) clause in AiProvenanceSpec’s Javadoc, and the forge-et-fe tests, fixtures and pin #329 added, gating the forge-et-fe PR against the backend branch, and add a dated revision note to the F11-c Wave 1 contract.
Regression test: LeadResearchControllerIT reads the label as public web search, <d MMM>, which fails before the change.
Resolved: forge-et-be#303 (706112c7) and forge-et-fe#334 (3ced7ddd), 2026-10-22, to merge backend first. LeadResearchService builds the label with AiProvenance.Source.dated and its own EEE d MMM formatter is gone; payloads.md and AiProvenanceSpec no longer name it as the exception. The frontend’s seam spec, browser support (now expectedLabelDate), size-matrix pin and fixtures expect public web search, 21 Oct. The F11-c Wave 1 contract’s Rev 3 records the change on forge-et-harness#846, where the contract still is. Fails before: LeadResearchControllerIT with the new expectation on the unchanged service, 1 of 18 failed on "public web search, Thu 22 Oct" (EXIT=1); with the fix the research specs passed, 25 suites with no failure (EXIT=0).
Lesson: L-170 — the same cause as BUG-240: a feature formatted a shared display value in its own code, after its own contract, instead of using the shared builder; BUG-253 is added to its sources.
BUG-252 — Flaky test: ScheduledFreshTraceIT AC-4 found the second scheduled message’s log line with no traceId once (Linux CI)
Seen on: Linux, the GitHub Actions ubuntu-latest runner (forge-et-be’s Slow lane); not seen on macOS, where it passed in the local build record of forge-et-be#299 the same day.
Affects: FS-9 (forge-et-be src/test/groovy/com/empireteas/platform/messaging/ScheduledFreshTraceIT.groovy, “AC-4 (listener path) — two attribute-less scheduled messages get distinct, non-empty traceIds on the worker lines”) · Found in: forge-et-be#293’s Slow lane, PR validation run 39769292262, head 2996be3e, 2026-10-22 11:47 UTC · Severity: low — one failure in 10,158 tests; the same test passed in the local build record on cc63dbf and in the dispatched run 39728456188 that day; the product is not known to be affected · Phase: 1
Repro: not reproduced. The run was the Slow lane of a Dependabot PR that changes only @commitlint/cli and @commitlint/config-conventional; nothing it changes is read by the backend tests. Machine load on a CI runner is not recorded.
Expected: both scheduled messages’ handler log lines carry a non-empty, distinct traceId in their MDC within 25 s.
Actual: Condition not satisfied after 25.00 seconds and 249 attempts at ScheduledFreshTraceIT.groovy:153; the last attempt failed on tb != null && tb.length() > 0 with tb null (line 161). So the second message’s log event was captured, but its MDC map held no traceId.
Fix: to find. First suspect, not confirmed: Logback’s LoggingEvent copies the MDC lazily, the first time getMDCPropertyMap() is called, from whichever thread calls it; if no appender read it on the worker thread, the test thread’s call copies the test thread’s empty MDC. Check whether the ListAppender the spec attaches calls prepareForDeferredProcessing(), and if not, call it in the appender before storing the event. Per .forge/test-strategy.md § Flaky Tests, do not fix it by lengthening the 25 s wait.
Cause (confirmed 2026-10-22, logback-classic 1.5.38 source): the suspect holds. LoggingEvent’s constructor does not copy the MDC; getMDCPropertyMap() copies the calling thread’s on first call and keeps it. Logger.callAppenders runs the logger’s own appenders before its parents’, so the spec’s ListAppender on ScheduledDemoJobHandler holds the event before the root console appender reads the MDC on the worker thread. A test-thread poll that finds the event in that moment copies the test thread’s empty MDC into it, and every later poll reads the same empty map, which is why the 25 s wait ran out. Not Linux-specific: a slower runner widens the moment. QueueTraceIT (FS-7) reads its captured events’ MDC the same way.
Fix (approach): a shared test appender, com.empireteas.support.MdcCapturingAppender, a ListAppender whose append calls event.prepareForDeferredProcessing() on the logging thread before keeping the event, and whose list is safe to read while the worker appends to it. ScheduledFreshTraceIT and QueueTraceIT use it. Regression test: MdcCapturingAppenderSpec holds the logging thread inside the moment (a parent appender that waits) and reads the event from the test thread; a plain ListAppender loses the traceId every time, the new appender keeps it.
Resolved: forge-et-be#304 (bbad0beb; review fixes cc81203a, 79507a0b; build record 10,164 tests, 0 failed), 2026-10-23, as the approach above, reproduced and fixed on macOS. The review found the same race in MailboxBackfillIT on the event’s thread name, which Logback fills in lazily the same way; it uses the appender too. Fails before: MdcCapturingAppenderSpec with the appender not yet freezing the MDC, 1 of 2 failed (EXIT=1), the plain-ListAppender case reading a null traceId; with the fix MdcCapturingAppenderSpec 2 of 2, ScheduledFreshTraceIT 4 of 4 and QueueTraceIT 2 of 2 passed (EXIT=0). The 25 s wait is unchanged.
Lesson: L-228 — a test that reads a Logback event’s MDC (or thread name, or formatted message) on another thread than the one that logged it reads its own unless the event was frozen on the logging thread; new in this PR, with a row in backend-reviewer § Test isolation.
BUG-251 — A Tab pressed right after choosing an option in a shadcn Select is lost, and focus goes back to the Select’s trigger
Affects: the shared shadcn Select (forge-et-fe src/ui/shadcn/select.tsx, AD #39) and so every screen with one; seen on F11-d’s AI tag filter on the Leads inbox · Found in: frontend-reviewer’s delta review of forge-et-fe#330 (BUG-228’s fix), finding D3, 2026-10-22; confirmed by a browser probe the same day · Severity: low — a keyboard user who presses Tab within the list’s close animation has to press it again; nothing is changed or lost · Phase: 1
Repro: on /leads, focus AI tag, press Enter to open it, ArrowDown to Spam, Enter to choose it, then Tab at once and type abc. A throwaway Playwright probe (Chromium, against forge-et-fe#330’s branch) did exactly this 3 of 3 times: focus ended on the AI tag trigger (leads-filter-ai-tag) and the search box stayed empty. The same probe with a click on the search box instead of Tab kept focus there and typed abc every time.
Expected: Tab after choosing moves focus to the next control, as it does once the list has closed.
Actual: Radix Select keeps its list mounted through the shadcn exit animation (about 150 ms). A Tab pressed then is handled inside the closing list, whose focus scope keeps it there, and when the list unmounts Radix focuses the trigger. Keys typed meanwhile reach the trigger, where the Select’s type-ahead can act on them. A click elsewhere is not affected, because it moves focus out before the unmount.
Fix: not decided. Options: an onCloseAutoFocus on the shadcn SelectContent that does not move focus back to the trigger when focus has already left the list for another element; or no exit animation on the Select’s content; or both, checked against Radix’s own focus return on Escape.
Regression test: a browser test on one Select screen: choose by keyboard, Tab at once, type, and check that focus and the typed text are in the next control, repeated (--repeat-each 10) in Chromium and WebKit.
Cause (found 2026-10-23, by a browser probe logging focus, keys and the list’s state): Radix keeps the list mounted, data-state="closed", through the closing animation; while it is, its SelectContentImpl prevents Tab and re-runs focusSelectedItem, which focuses the chosen option again. Radix returns focus to the trigger only from FocusScope’s unmount timer, about 36 ms after the list unmounts. The trigger loses aria-controls once closed, and InertWhileOpen kept the page inert until unmount. Neither option above is enough on its own: without the animation the Tab, about 10 ms after Enter, lands on the page body before Radix’s timer; and moving focus to the trigger inside the Enter keydown makes Enter’s default action click the trigger and reopen the list.
Fix (settled 2026-10-23 by the probe, before the final code): both options and a third part. (1) No closing animation on SelectContent, so the list unmounts at once. (2) The Select holds its own open state, so Radix reports a close as it happens, and on close, until focus lands anywhere or 500 ms pass, a Tab, Shift+Tab or Option+Tab that arrives while focus is still on the page body starts from the trigger: focus moves to the trigger before the browser acts on the key, and the browser’s own Tab runs from there (takeTabFromUntilFocused, src/lib/browser/takeTabFrom.ts). So each engine keeps its own tab order, WebKit’s skipping of buttons where macOS keyboard navigation is off included. A first version computed the next control itself and diverged from WebKit’s order; the second delta review replaced it. (3) SelectContent’s onCloseAutoFocus leaves focus where it is once it is somewhere other than the body. SelectTrigger gives its element to the Select through a context.
Resolved: forge-et-fe#335 (018ae688; review fixes 5ade6b0c, 63ad0caa, 9ee53cb2), 2026-10-23. Fails before: the new select-tab-after-choose.spec.ts (Chromium) and select-tab-after-choose-webkit.spec.ts (WebKit), Tab and Shift+Tab, each target measured with the list closed and read again after Radix’s own focus return has run, on main’s Select: 4 of 4 failed (focus ended on leads-filter-ai-tag); without SelectContent’s onCloseAutoFocus guard, 1 of 4 failed in the browser and the unit row every time. With the fix, the Leads and shell specs that touch the Select passed 270 of 270 with --repeat-each=5. Unit tests: select.test.tsx (10) and takeTabFrom.test.ts (5).
Lesson: L-133 — closing the list removes the focused option, and where focus goes had to be said and had to arrive before the next key; added to its sources, with a bullet on an overlay’s closing animation.
BUG-250 — scripts/check-incremental-compile.py fails on F4-d’s migration write path: the migration methods take Map<String, @Nullable Object> audit contexts that are never null
Affects: F4-d (forge-et-be account/service/AccountMigrationService.java and ContactMigrationService.java, wave 1, on main since forge-et-be#260; and, on F4-d-WI-2.1’s branch feature/F4-d-WI-2.1-group-writes, forge-et-be#301: TagService.findOrCreate, the RecordTagService.detach overload and their callers migration/service/MigrationGroupWriter.java and migration/service/MigrationTagResolver.java — at forge-et-be 49b90f11 the error sites are MigrationGroupWriter.java:211, :229, :269, :292, :340 and MigrationTagResolver.java:84), F1-e (attachIncludingRetired, same parameter shape) · Found in: F4-d-WI-2.1 subtask 8, python3 scripts/check-incremental-compile.py in its worktree, Windows, 2026-10-22 · Severity: low — developer loop only: an incremental compileJava of the writer fails NullAway; ./gradlew check (exit 0 at 684d4dc3) and CI’s clean compile pass · Phase: 1
Repro: reported by the F4-d-WI-2.1 implementer, not yet re-run by anyone else: in a worktree on feature/F4-d-WI-2.1-group-writes with a compiled build/, python3 scripts/check-incremental-compile.py exits 1 on MigrationGroupWriter.java. The writer is the first Java caller of these methods, so the problem showed only now.
Expected: the check passes; forge-et-be/CLAUDE.md (Keep @Nullable out of type arguments that other files read) allows Map<K, @Nullable V> in a non-private signature only when a value can really be null.
Actual: recompiled alone, the writer reads the methods’ class files, where javac has dropped the @Nullable in the type argument (the BUG-237 pattern), and NullAway refuses the Map<String, @Nullable Object> it passes. Workaround: ./gradlew compileJava --rerun.
Fix: declare the audit-context parameters as Map<String, Object> (no value in them is ever null) in AccountMigrationService, ContactMigrationService, TagService.findOrCreate, the RecordTagService.detach overload and, if it is the same shape, F1-e’s attachIncludingRetired; then re-run python3 scripts/check-incremental-compile.py. Candidate for the F4-d wave 2 PR’s review, since that PR adds the first caller; ContactMigrationService is outside WI-2.1’s ## Files to Modify, so the implementer left it.
Resolved: forge-et-be#310 (1af3d68c; merged 2026-10-26 as 30fd601c), F4-d’s wave 2 PR, as a review fix: the audit-context parameters are Map<String, Object> (attachIncludingRetired had been removed by WI-2.1 revision 2, so it needed no change). python3 scripts/check-incremental-compile.py passes on the fix, where it exited 1 on the writer’s and resolver’s calls before; the review record (forge-et-be#310, comment 6387675412) lists it as fixed.
BUG-249 — Flaky test: the dev smoke run’s check 6 (Accounts) found no SMOKE Account row within 5 s once, while check 9 found the account through the API in the same run (Windows)
Seen on: Windows 11 (Git Bash), against dev; not seen on macOS, where the smoke run has not yet been run.
Affects: F38-g (forge-et-fe e2e/smoke/checks.smoke.ts, check 6 “Accounts — SMOKE account through account.all”, the toHaveCount(1) on account-list-row) · Found in: F38-g-WI-2.1 R-4 attempt 2, pnpm e2e:smoke against dev at 2026-10-22 05:23–05:27 UTC, after the forge-et-be deploy run 39727038992 (e841198b) · Severity: low — a flaky check; one failure in five complete smoke runs on dev, and the next run passed; the product is not known to be affected · Phase: 1
Repro: not reproduced. Signed in as crm05@empire.lk (role Smoke test (view only)), the run opened /accounts?q=SMOKE; the check took 11.6 s in all. The four other smoke runs that day (R-3, R-4 attempts 1 and 3, R-5 attempt 1) and R-6’s deliberate failure passed check 6.
Expected: the accounts list shows one row, SMOKE Account (Japan), reached through account.all.
Actual: expect(getByTestId('account-list-row')).toHaveCount(1) received 0 after 5,000 ms (e2e/smoke/checks.smoke.ts:140). In the same run, check 9 found SMOKE Account through /api/v1/accounts?q=SMOKE, so the record and the reach were in place. The run’s cleanup deleted the error context, so the page’s state at the failure is not known (.forge/plans/F38-g/F38-g-WI-2.1-record.md, R-4 attempt 2).
Fix: to find. First suspects: the list’s first read after a fresh sign-in landing late, or an empty or loading state that has no rows while the search query is in flight. Keep the error context of a failed smoke run so the next failure shows the page. Per .forge/test-strategy.md § Flaky Tests, do not fix it by lengthening the 5 s wait.
Investigated 2026-11-02 (Dinil), cause still unknown:
- The backend was not slow. The dev load balancer’s one-minute metrics (
et-dev-backend, read-only CloudWatch reads) over attempt 2 (05:23–05:27 UTC) show at most 1.56 s for the slowest answer in any minute, no 5xx answers, and the same 4xx count (10) as attempt 3, which passed. - The task change was not the cause. The backend logs show the new task started at 05:19:34, and the old one shut down cleanly at 05:25:51–05:26:01, before the checks phase.
So the empty list came from the browser side. Playwright’s error context would hold dev data and outlive the run (plan F38-g-WI-1.1 D-7), so forge-et-fe#376 keeps a data-free summary in check 6’s failure message instead. It gives:
- the page’s path;
- which of the Accounts list’s state test ids it showed, and the row count;
- each
/api/v1/accountsread, with when it was sent, its status, when its headers and body came, or why it failed.
The report prints only the first line, so those lines are in Playwright’s output above it (e2e/README.md § The smoke suite).
Resolved: 2026-11-05 — closed as wontfix, the lead’s decision (Tharinda), proposed by Dinil: one failure in five smoke runs on 2026-10-22 and none recorded since; the backend was not slow; the product is not known to be affected. No fix was made. forge-et-fe#376 stays, so a recurrence’s failure message shows the page’s state and each list read’s timing; a recurrence is filed as a new bug that links here, with those lines.
Lesson: none — one-off: a single failure whose page state was not kept. forge-et-fe#376 now keeps a data-free summary in check 6’s failure message, and F38-g-WI-1.1 D-7 already records why the smoke run keeps no trace.
BUG-248 — Every AI call to Sonnet 5.5 is expected to fail: BedrockAiProvider forces a named tool, which Sonnet 5.5 rejects
Affects: F12’s AIProvider seam (forge-et-be src/main/java/com/empireteas/ai/service/BedrockAiProvider.java: the request builder sets toolChoice(ToolChoice.fromTool(SpecificToolChoice…)) at line 138, and the error mapping turns a ValidationException into MODEL_UNAVAILABLE at line 168), and through it the two features configured on Sonnet 5.5: the email draft (F11-h, et.ai.features.email-draft.model) and the account briefing (F13, et.ai.features.account-briefing.model) · Found in: reading BedrockAiProvider against Anthropic’s Sonnet 5.5 migration guide (the claude-api reference, cached 2026-10-06) while preparing .forge/design/ai/model-choice-evaluation-plan.md, 2026-10-22 · Severity: high — on their configured model both features would fall back to the manual path on every call; hidden on dev, whose task definition overrides both to Haiku 4.5 while the account’s Sonnet 5.5 quota is 0 · Phase: 1
Repro: not yet run: no call to Sonnet 5.5 has been made from any environment (cloud-gated-validation.md §C-5). Once dev’s Sonnet 5.5 quota is raised, and with the lead’s approval of the Bedrock calls: AWS_PROFILE=et-dev ./gradlew liveEvaluation -Pfeature=email-draft -PconfirmBedrockCalls=true -Pmodel=global.anthropic.claude-sonnet-5-5 in forge-et-be.
Expected: each case returns a draft through the prompt’s tool, as it does on Haiku 4.5.
Actual (predicted from the guide): Anthropic’s guide lists, as a breaking change on Sonnet 5.5, that tool_choice any or a named tool returns a 400 invalid_request_error (“tool_choice: type “tool” and “any” are not supported for this model”), and that the same holds for its Bedrock forced-tool path. Bedrock answers such a request with a ValidationException, which BedrockAiProvider maps to MODEL_UNAVAILABLE: the feature answers as unavailable, the person writes the email or reads the account themselves, and the alarm names the model as if it were missing. Haiku 4.5 and Haiku 5.5 accept a forced tool, so the fault shows only on Sonnet 5.5 (and on Opus 5.5 and Fable 5.1, which no feature uses).
Fix: build every request with toolChoice auto for every model (strategy §4c rule 2, one request shape), the prompt stating that the answer is given only through the named tool, and the existing rule kept that an answer without that tool call is INVALID_OUTPUT and retried once (et.ai.call.max-invalid-output-retries). Prove it with one live call to Sonnet 5.5 on dev once its quota allows, and re-run the Haiku 4.5 live evaluations of inquiry-extract, lead-triage, email-draft and account-briefing to show the change costs nothing on Haiku. A side effect to measure in the same runs: Haiku 5.5 skips thinking entirely when a tool is forced, so with auto it may think first, which changes its token use (model-choice-evaluation-plan.md P-1).
Regression test: a BedrockAiProvider spec asserting the built ConverseRequest’s tool choice is auto for each configured model — it fails on today’s code — plus the live call above, recorded in the cloud-gated register.
Status (2026-10-26): open, no owner. The lead’s decision: fixed when the team takes it up, not now. It blocks nothing in Phase 1, because every AI feature runs on Haiku 4.5 until AWS lifts the account’s new-account limits (ai-provider-strategy.md §5 row 10), and Haiku 4.5 accepts a forced tool. It must be fixed before any feature is moved to Sonnet 5.5, and it is one of the three changes that make a model switch a configuration change (ai-provider-strategy.md §4d; owed.md, “The AI layer under Bedrock’s new-account limits”). Since it was filed, a third feature defaults to Sonnet 5.5: lead research (F11-c, et.ai.features.lead-research.model, merged in forge-et-be#297), so the fix covers three features. The code is as described above on forge-et-be origin/main at 25e61c91 (26 Oct): the forced tool at BedrockAiProvider.java line 138, the ValidationException mapping at line 168. Where to pick up: the code change and the regression test need no AWS and can be done and merged on their own; the live proof waits until a Sonnet 5.5 call answers on dev.
Resolved: forge-et-be#316 (9f808935; review fixes 53b10a1f, 31e527d5; a Fast-lane fix to a comment, d0871bb8), 2026-10-26: every request uses toolChoice auto. Every prompt version already says “Answer only by calling the tool you are given”, so no prompt version changed, and a reply without the tool call stays INVALID_OUTPUT, retried once. BedrockAiProviderSpec asserts auto for every model a feature names or a price row lists, that every version of each feature’s prompt carries that sentence, and (owed.md, “Haiku 5.5 readiness in the backend”, item 3) that a reply with reasoning blocks before the tool call is read from the call; docs/ai/bedrock-cloud-contract.md states the new request. Fails before: the spec against origin/main’s BedrockAiProvider, EXIT=1, with the AC-2b request test, the F11-h AC-6 draft test and every per-model case failing; with the fix 41 tests, 0 failures (EXIT=0), and 38 after the review made the per-model case one per distinct model, 0 failures (EXIT=0). docs/ai/README.md now names the prompt sentence the request relies on. The live proofs, one Sonnet 5.5 call on dev and the Haiku 4.5 re-runs of the four live evaluations, are row 10 of cloud-gated-validation.md §B’s F12 block, needing the lead’s approval of the Bedrock calls.
Lesson: none — one-off: a request shape that a later model refused. ai-provider-strategy.md §4d already requires each feature’s evaluation on the new model before it is switched, which would have shown it before any user did.
BUG-247 — The backend test JVM runs out of its 2g heap part-way through the full suite
Affects: FS-14 (forge-et-be build.gradle, tasks.named('test'), maxHeapSize = '2g', L622 on main) · Found in: the local build record (python3 scripts/local-build-evidence.py) on F11-c’s backend wave, forge-et-be feature/F11-c-wave-1 at 26d5830a, Windows 11, 2026-10-22 · Severity: medium — the full ./gradlew build dies with OutOfMemoryError after about 44 minutes, so no build record can be written; the product is not affected · Phase: 1
Repro: ./gradlew build on F11-c’s backend wave (every spec, F11-c’s lead-research integration tests included).
Expected: the whole suite runs in the test JVM’s heap.
Actual: after about 9,738 tests the executor ran out of heap; SearchSeamGuardIT died building its Spring context. No test failed on its own, and the machine was not loaded (load average 0.19). The suite’s cached Spring contexts each hold a datasource, an SQS container and a Hikari pool, and F11-c’s integration tests add more of them.
Fix: raise the test task’s maxHeapSize to '3g', with the comment beside it naming this bug, in its own commit in F11-c’s backend wave (c9b8a638).
Regression test: the wave branch’s full ./gradlew build (the local build record) runs the whole suite to the end.
Resolved: 2026-10-22 — fixed in forge-et-be#297 (F11-c wave 1, merged as 3ef0064f) by commit c9b8a638.
Lesson: L-090 — another instance (the size of the whole tree, not one test: raise the heap in build.gradle); no new lesson.
Amended 2026-11-17 — 3g is outgrown: the local build record on F36-c’s backend wave (forge-et-be feature/F36-c-wave-1 at b8482957, with F37-a’s wave 1 and F36-b’s wave 2 merged in from main) ran 17,762 tests on Windows 11; 17 failed to load their Spring context with OutOfMemoryError: Java heap space near the end of the suite (PriceBroadcastSweepBudgetIT, PriceExpirySweepIT, PriceImportAuditIT), at a peak load of 0.77. The test task’s maxHeapSize is raised to '4g', its comment naming this bug, in its own commit on that branch (12f6b7e2, forge-et-be#456). Regression test: that branch’s next full build record runs the whole suite to the end. main has the same tests and the 3g heap, and was not checked the night of 2026-11-16 (forge-et-be issue #453): until #456 merges, a full run on main can fail the same way.
BUG-245 — The size matrix’s contacts-gone state fails: BUG-232’s mark expects a failure that Windows does not have (Windows)
Windows only: the per-OS twin of BUG-232 (macOS).
Affects: F1-f (forge-et-fe e2e/pageStates.ts, contacts('contacts-gone', '/contacts') and its knownFailures entry for BUG-232 on size-320x640-chromium; checked by e2e/browser/size-matrix/size-matrix.spec.ts:217) · Found in: F11-c WI-1.6’s size-matrix runs on Windows 11, forge-et-fe feature/F11-c-WI-1.6-e2e-lead-research and main 21288b2c, 2026-10-22 · Severity: medium — every full size-matrix run, and so every pnpm e2e:gate, fails this state on Windows; the product is not affected · Phase: 1
Repro: Windows 11, an isolated stack (et-f11-c-wave1): pnpm exec playwright test e2e/browser/size-matrix --project=size-320x640-chromium --grep contacts-gone — 1 failed, 10 passed, on F11-c’s branch and on main’s own code (worktrees/F11-c/verify-main/forge-et-fe at 21288b2c), 3 of 3 runs.
Expected: the run passes on both systems: the mark applies where the failure happens.
Actual: BUG-232 no longer fails these here; remove them from the mark in e2e/pageStates.ts — the axe target-size failure BUG-232 records on macOS does not happen on Windows, so the mark’s “a named kind that did not fail” check fails the state.
Fix: limit the mark to macOS (the platform BUG-232 was found on) — the entry’s projects is ['size-320x640-chromium'] only when process.platform === 'darwin', with a comment naming this bug — in its own commit in F11-c’s frontend wave; the mark’s own tests (src/app/knownFailureMarks.test.ts) keep passing. forge-et-fe#320 (BUG-232’s fix, draft) deletes the whole mark, so whichever PR merges second resolves one small conflict in that entry, keeping the deletion.
Regression test: the same size-matrix command passes on Windows; on macOS the mark still applies until BUG-232’s fix lands.
Resolved: 2026-10-22 — forge-et-fe#320 (BUG-232’s fix) merged and deleted the mark, so the state carries none on either system. F11-c’s interim fix (586883c4, the mark applied on macOS only), which made F11-c’s size-matrix runs pass on Windows that night (E-5: 28 passed at 320 × 640), was dropped in favour of main’s deletion when main was merged into F11-c’s frontend wave (7715492d).
Lesson: none — one-off: a known-failure mark recorded on one operating system and checked on another; BUG-232’s fix removed it, and tracker.md § Bug Tracking already files a test failure per operating system.
BUG-243 — compileTestGroovy runs out of heap at the forked compiler’s 1536m
Affects: FS-14 (forge-et-be build.gradle, groovyOptions.forkOptions.memoryMaximumSize = '1536m', L541 on main) · Found in: F11-c WI-1.2’s full ./gradlew build runs on Windows 11, forge-et-be feature/F11-c-WI-1.2-exa-search-provider (on WI-1.1’s ba30d74e), 2026-10-21 · Severity: medium — a full build fails at random with OutOfMemoryError in compileTestGroovy, which wastes a 45-minute local build record or a CI run; the product is not affected · Phase: 1
Repro: ./gradlew build on F11-c’s backend wave (WI-1.1 + WI-1.2 specs; WI-1.3 adds more): compileTestGroovy ran out of heap in 5 of 7 full compiles at 1536m.
Expected: the test compile fits its heap every time.
Actual: it fits only sometimes; removing any one new spec makes it fit, so the limit is the size of the whole test source set, not one construct. Every compile passed at 2g (an init script outside the worktree, never committed).
Fix: raise groovyOptions.forkOptions.memoryMaximumSize to '2560m', with the comment beside it naming this bug, in its own commit in F11-c’s backend wave (f9f1e456, comment reworded in 8e9aa07b). 2g was enough for WI-1.1 and WI-1.2’s specs; WI-1.3’s full build, with every spec, needed 2560m.
Regression test: the wave branch’s full ./gradlew build (the local build record) compiles the whole test source set, WI-1.3’s specs included, with no init script.
Resolved: 2026-10-22 — fixed in forge-et-be#297 (F11-c wave 1, merged as 3ef0064f) by commits f9f1e456 and 8e9aa07b.
Lesson: L-090 — another instance (“the size of the whole tree, not one test: raise the compiler heap in build.gradle”); no new lesson.
Amended 2026-11-16 — 2560m is outgrown: with F36-c WI-1.2’s specs added (forge-et-be feature/F36-c-WI-1.2-plan-research-job, stacked on WI-1.1), compileTestGroovy failed with Java heap space on every compile at 2560m; removing about a hundred lines from one spec let it pass and adding one trivial method back failed it again — the same source-set-size cause. At 3584m (an init script, not committed) it compiled in about a minute and the whole ./gradlew check passed on 85940229. The line is raised to '3584m', its comment naming this bug, in its own commit on that branch (F36-c WI-1.2 plan Rev 2, approved by Dinil). F37-a’s wave branch (origin/feature/F37-a-wave-1) separately raises it to '3072m', untested against F36-c’s specs: whichever of the two merges to main second keeps '3584m'.
BUG-242 — smokeRun.test.ts’s output-folder check fails (Windows)
Affects: F38-g (forge-et-fe src/test/scripts/smokeRun.test.ts:320, “passes, runs both phases with the dev target and the run flag, and deletes both afterwards”; added in 9813a9ff) · Found in: F11-c WI-1.4’s full pnpm test on Windows 11, forge-et-fe feature/F11-c-WI-1.4-research-panel at c98dcb23, 2026-10-21 — the test and everything it imports are unchanged from main (git diff --quiet origin/main HEAD -- src/test/scripts/smokeRun.test.ts scripts/smoke e2e/smoke e2e/support/smokeSession.ts src/lib/config exits 0) · Severity: medium — pnpm test, and so pnpm test:record, fails on every Windows machine, which blocks every frontend PR’s local record there; the product is not affected · Phase: 1
Repro: on Windows, pnpm exec vitest run src/test/scripts/smokeRun.test.ts.
Expected: the test passes on Windows and macOS: the Playwright --output= folder lies inside the run’s report folder.
Actual: it fails on Windows: the check is output.startsWith(\${reportDir}/`), and the code builds the folder with path.join, which writes ` on Windows.
Fix: a duplicate of BUG-244, which F13 found the same day: forge-et-fe#327 (merged 2026-10-22) fixed it by asserting the exact path.join result, the separator being the platform’s. F11-c’s own fix (7a73692c, a path.relative check that the --output= folder lies inside the report folder) was replaced by main’s when main was merged into F11-c’s frontend wave (56104ae5). Test-only; the smoke-run code is correct on both systems.
Regression test: the same test, passing on Windows (this machine) and on the Linux CI Slow lane.
Resolved: 2026-10-22 — a duplicate of BUG-244, closed by BUG-244’s fix in forge-et-fe#327 (merged as 1592be74).
Lesson: L-103 — another instance (a test compared a path.join result with a forward-slash literal; CI on Linux cannot catch it); no new lesson.
BUG-246 — In Chromium the pinned-first-column tables are off by 10px: the Name cell moves 10px when pinned at 768px, and lags the scroll by 10px when unpinned at 320px, on main (Windows)
Affects: BUG-232’s fix (forge-et-fe#320, merged 2026-10-21 23:31Z: src/ui/pinnedFirstColumn.ts and the tables that use it — the accounts list and the contacts list), and its tests: e2e/browser/accounts/pinned-first-column.spec.ts:37 (the 768px pinned row), e2e/browser/accounts/accounts-layout.spec.ts:188 (AC-15 at 320 × 640), e2e/browser/accounts/contacts-layout.spec.ts:419 (AC-22 at 320 × 640) and the size matrix’s accounts-pinned state at size-320x640-chromium · Found in: F13 wave 1’s Stage 12 pnpm e2e:gate on forge-et-fe feature/F13-wave-1 (07eac450, after main was merged in), 2026-10-22 · Severity: high — every full browser gate on a Windows machine fails on these four tests, main included, so no forge-et-fe PR can be marked ready from Windows; and it may be a real defect for Windows Chrome users (Empire’s) · Phase: 1
Repro: Windows 11, 16 cores, Chromium (Playwright chromium-1243), against a local forge-et-be on 8880. On main’s own code (forge-et-fe origin/main 10696c8d, a detached worktree), run alone on a quiet machine: ET_BACKEND_PORT=8880 pnpm exec playwright test e2e/browser/accounts/accounts-layout.spec.ts e2e/browser/accounts/contacts-layout.spec.ts e2e/browser/accounts/pinned-first-column.spec.ts --project=chromium → 3 failed, 26 passed; … e2e/browser/size-matrix --grep "accounts-pinned|account-detail-compose-sent" --project=size-320x640-chromium --project=size-1024x768-webkit → 1 failed (accounts-pinned), 13 passed. The F13 wave head fails the same four the same way. forge-et-fe#320’s own gate passed on another machine (Browser gate: passed · 478002329939 · 5362 passed).
Expected: at 768px the first cell stays put while the table scrolls (Math.abs(measured.moved) < 1); at 320 × 640 it scrolls away with the table (moved ≈ scrolled, within 1px).
Actual: pinned-first-column.spec.ts:62 — “contacts-table-scroll’s first cell stays put”: expected < 1, received 10. accounts-layout.spec.ts:221 and contacts-layout.spec.ts:452 — “the Name cell scrolls away with the table”: expected ≤ 1, received 10. size-matrix.spec.ts:235 — “the Name cell scrolls away with the table below bp/tablet”: expected ≤ 1, received 10. The same 10px in all four, pinned and unpinned, which suggests a Windows-only difference in the scroll region (a classic scrollbar’s width, or a border or padding the measurement or the sticky offset does not account for) rather than load: each failed identically when run alone.
Cause (found 2026-10-22): every pinned table’s scroll region sets scrollbar-width: thin with scrollbar-gutter: stable both-edges. Where the scrollbar takes room (Windows, a 10px thin scrollbar) both-edges reserves a 10px strip on the left too, so the first cell starts 10px in from the region’s edge; Chromium pins a sticky start-0 cell at the region’s edge, inside that strip. So at 768px the pinned Name cell jumps 10px left as soon as the table scrolls — a real shift Windows Chrome users see — and at 320px the unpinned cell scrolls correctly but starts 10px in, where the two layout tests and the size matrix assume the region’s edge. On macOS the scrollbar overlays the content, the strip is 0px, and every test passes. A plain page in Playwright’s Chromium on Windows shows it alone: a sticky cell in a stable both-edges region starts 10px in and moves 10px; in a stable region it starts at 0 and moves 0.
Fix: reserve the scrollbar strip on the right only (scrollbar-gutter: stable) on every table that pins its first column — the accounts, contacts, auction sales, round, standards, roles and users tables — so the first cell starts at the region’s edge and stays there when pinned. Tables that do not pin (email templates) keep both-edges. Trade-off: on Windows these tables no longer have the 10px strip on their left, so they sit against the card’s left side instead of evenly between its sides; macOS is unchanged. A source scan in src/ui/pinnedFirstColumn.test.tsx fails a file that pins a first column and reserves the strip on both edges.
Regression test: the scan in src/ui/pinnedFirstColumn.test.tsx — with the seven tables at origin/main it fails (EXIT=1, 1 failed, 7 passed, listing them), and passes on the fix (8 passed); it runs on every OS, where the browser specs catch the shift only on Windows. e2e/browser/pricing/weekly-prices.spec.ts’s pinned-column test, which had accepted a move of up to the left strip, now expects no left strip and the pinned cell to stay put. On Windows 11 Chromium against an isolated backend, the four specs that failed on main and the weekly-prices spec pass (39 passed), and the size matrix’s accounts-pinned row passes at every size project (17 passed).
Resolved: forge-et-fe#331, 2026-10-22.
Lesson: L-224 (new): anything that depends on a scrollbar’s width is proved on a machine where scrollbars take room; L-220 gains an update (find the cause before marking a platform-only failure).
BUG-244 — smokeRun.test.ts’s full-run test fails: the smoke script builds the Playwright output folder with \ (Windows)
Affects: FS-1’s smoke runner (forge-et-fe scripts/smoke/run.mjs runPhase, and its test src/test/scripts/smokeRun.test.ts:320, “runSmoke — a full run › passes, runs both phases with the dev target and the run flag, and deletes both afterwards”) · Found in: F13-WI-1.2’s full pnpm test, 2026-10-21 · Severity: medium — pnpm test, and so pnpm test:record, fails on every Windows machine, main included; the product is not affected · Phase: 1
Repro: on Windows, pnpm exec vitest run src/test/scripts/smokeRun.test.ts in forge-et-fe. On main 21288b2c (the files unchanged in F13-WI-1.2’s worktree): EXIT=1, 1 failed, 36 passed; the expected --output argument starts with <reportDir>/ and the received one with <reportDir>\.
Expected: the test passes on every operating system the team builds on.
Actual: runPhase passes join(reportDir, \output-${name}`)toplaywrightArgs; node:path's joinuses` on Windows, while the test compares with a /-joined string. It passes on Linux and macOS, where join uses /. The same shape as BUG-038.
Fix: to be made in F13’s frontend wave, in its own commit, under Dinil’s standing answer for small test-backed defects in old code (.forge/plans/F13/F13-overnight-log.md standing answer 4).
Resolved: 2026-10-22 — fixed in forge-et-fe#327 (F13 wave 1, merged as 1592be74) by commit a9b733ca, which builds the test’s expected --output path with path.join, so the separator is the platform’s.
Lesson: L-103 — another instance (a test compared a path.join result with a forward-slash literal); no new lesson.
BUG-241 — F1-c’s AC-17 WebKit timeline tests fail when keyboard navigation is off: loadEverythingByKeyboard presses plain Tab (macOS)
Affects: F1-c (forge-et-fe e2e/browser/accounts/account-timeline-keyboard-webkit.spec.ts, both AC-17 rows; e2e/browser/accounts/timeline-support.ts loadEverythingByKeyboard) · Found in: the full pnpm e2e:gate on forge-et-fe#317 (BUG-236’s fix), 2026-10-21 · Severity: medium — every full browser gate on such a Mac fails on both rows, main included; the product is not affected · Phase: 1
Repro: macOS with keyboard navigation off (the default). Run pnpm exec playwright test e2e/browser/accounts/account-timeline-keyboard-webkit.spec.ts --project=chromium --workers=1 against a local backend. On main b2dc3ee7: EXIT=1, 2 failed, 10 passed, both with Tab reaches the control within 15 steps at timeline-support.ts:535. The same result on forge-et-fe#317’s head before this fix, alone on an idle machine, so load is not the cause.
Expected: both rows reach Show older entries by keyboard and pass, whatever the machine’s keyboard-navigation setting.
Actual: WebKit’s Tab reaches text fields only on such a machine, the cause BUG-231 found in the notes-and-tags WebKit tests. loadEverythingByKeyboard called tabTo with its default key, Tab, so focus never reached the button. The spec’s header comment said “a button, which WebKit’s Tab reaches”, which is true only with keyboard navigation on.
Fix: loadEverythingByKeyboard takes a key option, Tab by default. account-timeline-keyboard-webkit.spec.ts passes the key e2e/support/focusKey.ts webkitFocusKey finds on the machine (Tab or Alt+Tab). account-timeline-a11y.spec.ts, which runs in Chromium, keeps Tab. The header comment now states the rule. git grep -n "keyboard.press('Tab')" e2e found no other WebKit spec that presses plain Tab between buttons. Every hit is in a Chromium spec or helper. The other WebKit keyboard specs choose their key already: notes-tags-keyboard-webkit and follow-up-keyboard-webkit through webkitFocusKey, offboarding-keyboard-webkit with its own probe, and lead-triage-keyboard-webkit, which tries both keys.
Regression test: the two AC-17 rows themselves. They failed on main b2dc3ee7 (EXIT=1, 2 failed) and pass with the change: with account-timeline-a11y.spec.ts too, EXIT=0, 15 passed.
Resolved: forge-et-fe#317 (86291164, review wording fix 56c908cb), 2026-10-21, folded into BUG-236’s fix PR so that its gate could pass. The delta review found one more state that presses plain Tab in WebKit: the size matrix’s leads-lead-share-unowned. It passes today only because of the older Tooltip’s explicit tabIndex, and it has a row in owed.md.
Lesson: L-168, another instance: its sources gain BUG-241. The rule it states, choose the key from the machine’s behaviour (webkitFocusKey), already covers it. F1-c’s WebKit spec was written before that rule reached the accounts helpers.
BUG-240 — The AI provenance “from:” date is written in three different ways
Affects: F11-d (forge-et-be LeadTriageService, the AI summary panel’s source label), F11-h (forge-et-be emaildraft/service/DraftSourceLabel.java, the Follow-up draft panel’s), F12 (forge-et-be InquiryExtractor, the pasted-inquiry draft’s) · Found in: Dinil, on dev, 2026-10-21, looking at one lead’s AI summary and Follow-up draft panels side by side · Severity: low — cosmetic; the meaning is clear, but one lead view shows two formats · Phase: 1
Repro: on dev, create a lead with inquiry text, let its AI summary arrive, and press Draft follow-up. The two panels’ provenance lines read from: inquiry, Tue 20 Oct and from: inquiry 20 Oct.
Expected: one format everywhere, the design system’s (.forge/design/ui/design-system.md §5, AI provenance — all four elements, Source: from: inquiry email, 14 Aug): what the output was made from, a comma, then the date as d MMM.
Actual: each feature formats its label on the server in its own way, as its own contract set it:
- F11-d lead triage:
inquiry,+EEE d MMM(inquiry, Tue 20 Oct;lead fields, Tue 20 Octwith no inquiry text) — F11-d Wave 1 contract,source.label. - F12 inquiry extraction:
pasted inquiry,+EEE d MMM. - F11-h follow-up draft:
inquiry+d MMM, no comma (inquiry 20 Oct,inquiry 20 Oct · 2 sent emails,email from <name>, 20 Oct · inquiry 14 Aug) — F11-h Wave 1 contract, Source labels. The F11-c lead research label (public web search,+EEE d MMM) has the same defect; it is BUG-253, left out of this fix because forge-et-fe#329, open on 2026-10-22, asserts its weekday form. Fix: align the three to<source>, d MMM, through one shared builder so they cannot drift again. Inforge-et-be:AiProvenance.Sourcegainsdated(what, date)(<what>, d MMM) andlabelDate(date);LeadTriageService,InquiryExtractorandDraftSourceLabelbuild their labels with it (inquiry, 20 Oct,lead fields, 20 Oct,pasted inquiry, 20 Oct,inquiry, 20 Oct · 2 sent emails,email from <name>, 20 Oct · inquiry, 14 Aug), and F13’sBriefingSourceLabel, alreadyd MMM, takes its date fromlabelDate. Every label is built by the server and shown as given, so no product code changes inforge-et-fe; its component and API tests, wire fixtures, size-matrix pins, browser support files and the doc comments that copy the old strings change. Five of its tests read the real backend’s string —e2e/seam/lead-triage.seam.spec.ts,e2e/seam/inquiry-extraction.seam.spec.ts,e2e/seam/ai-drafted-follow-up.seam.spec.ts,e2e/browser/leads/lead-triage-summary.spec.tsande2e/browser/leads/lead-inquiry-support.ts(found while fixing, the last two by the backend review; the report named only the first) — so theforge-et-fePR is gated against the BUG-240 backend branch and merges straight after the backend PR. A dated revision note goes on each frozen contract that set the old form (F11-d D-3, F11-h Source labels, F12 Wave 2). Regression test:AiProvenanceSpecpinsdatedandlabelDate(<what>, d MMM, no weekday);DraftSourceLabelSpec,InquiryExtractorSpec,LeadTriageControllerITandAiDraftControllerITassert each builder’s label in that form, and fail before the change. Resolved: forge-et-be#302 (f37f298d, review fixes09dd9da9,05e52143,dd23ce02) and forge-et-fe#332 (25e79b15,ef6a9761), 2026-10-22, merged backend first.AiProvenance.Source.datedandlabelDatehold the form; the lead triage, inquiry extraction and follow-up draft labels usedated, F13’s briefing useslabelDate. Fails before:DraftSourceLabelSpecandInquiryExtractorSpecwithDraftSourceLabel.javaandInquiryExtractor.javaput back toorigin/main, 6 of 19 failed (EXIT=1); with the fix the label specs passed 30 of 30. The research label is BUG-253. The backend review found two more frontend tests that read the real label (ai-drafted-follow-up.seam.spec.ts,lead-inquiry-support.ts), fixed inef6a9761. Lesson: L-170 — each AI feature formatted the same label in its own code instead of one shared builder; BUG-240 is added to its sources, with a bullet on shared display formats.
BUG-239 — Flaky test: F26-f AC-24’s Weekly prices test hovers Published and no “cannot be withdrawn” tooltip appears within 5 s, once in a full gate (macOS)
macOS only so far: seen on macOS (8 cores); no Windows run is recorded.
Affects: F26-f (forge-et-fe e2e/browser/pricing/weekly-prices-states.spec.ts:414, “AC-24: a published row with canWithdraw false offers no Withdraw, and its status says why”; the hover at :422 and the wait at :425) · Found in: the second full pnpm e2e:gate on forge-et-fe#314 (BUG-200/201/212, adf37715), 2026-10-20 · Severity: low — one test failed once in a full gate run and passed 5 of 5 alone; the product is not known to be affected · Phase: 1
Repro: macOS, 8 cores. ET_BACKEND_PORT=8084 pnpm e2e:gate (5,207 passed, 1 failed, 5 skipped, 4 known failures; 50.5 min; 4 workers) against forge-et-be main 5d1ecd0c on an isolated stack (et-bug-200-e2e). The run’s load: busy 0.28 at start, 0.96 peak, 0.86 average, load average 3.33 per core at the peak; the busiest processes at the peak were WindowServer (89.5 %), the run’s own chrome-headless-shell processes (76.6 %, 52.2 %, 48.5 %) and the Claude app (45.0 %); no stalls were recorded. Run alone straight after on the same head and backend: pnpm exec playwright test e2e/browser/pricing/weekly-prices-states.spec.ts --project chromium --grep "AC-24: a published row with canWithdraw false" --repeat-each=5 --workers=1, 5 of 5 passed (EXIT=0). The first full run on the same head passed this test.
Expected: after the row menu closes, hovering the row’s Published status shows its tooltip, “In effect since expect(getByRole('tooltip').filter({ hasText: /In effect since Tue 3 Nov( \d{4})? — cannot be withdrawn/ })).toBeVisible() failed after 5 s with no such element: the tooltip never opened.
Fix: test-only, decided 2026-10-22 before the code; the two suspects below are ruled out (2026-10-22, by reading the code), so the cause comes from the trace. Ruled out: (a) the row menu is a non-modal shadcn DropdownMenu (src/ui/shadcn/dropdown-menu.tsx, RoundTable.tsx passes no modal), so the page never gets pointer-events: none; (b) react-aria’s tooltip hover opens only under pointer modality, which Escape changes — but src/ui’s Tooltip opens on pointerenter through its own 400 ms timer (Tooltip.tsx handlePointerEnter), whatever the modality; a test change built on (b) was reverted in forge-et-fe#332 after its frontend review. Candidates left: the pointer never entered, the timer never fired, or the tooltip opened and closed again (a late scroll closing a hover open through useCloseOnScroll, whose BUG-181 reopen covers focus opens only, or a pointerleave). Original first suspect: the hover at :422 lands while the row menu Escape closed is still closing. A Radix menu sets pointer-events: none on the page while it is open, so a hover sent before that is removed reaches nothing, and the tooltip waits for a pointer-enter that never comes. Reproduce it under forced load first, with a Playwright trace (.forge/test-strategy.md § Flaky Tests, triage step 2). If the trace confirms it, wait for the menu to be hidden and the row-actions trigger to hold focus before hovering, through one shared e2e helper that BUG-219’s marker hover uses too (the same pattern: a hover straight after a Radix layer closes). If the trace shows another cause (a scroll that closes the tooltip, the trigger re-rendered), fix that instead. No longer timeout and no retry.
Regression test: the spec, passing in full gate runs on macOS with the default worker count.
Resolved: forge-et-fe#332 (3894ba45; 436c1535 answers the same read empty in weekly-prices.spec.ts and round-correct.spec.ts, whose status hovers run against the real backend too), 2026-10-22. The cause was neither suspect: the Pricing page’s warnings read (GET /api/v1/pricing/warnings, F26-g), which sits between the header and the tabs, came from the shared backend, and a warning card that landed after the hover pushed the round table down, so the pointer resting where Published had been never entered it. The state tests now answer that read empty, as the size matrix does (BUG-233). Not reproduced by load alone: 10 runs under one busy loop per core and 10 more with Chrome throttled 6× all passed. Reproduced by holding the step that lands late (.forge/lessons.md L-162): the warnings answer held until the pointer was over Published, then a warning card — the old spec failed 5 of 5 with the bug’s own error (toBeVisible on the tooltip, element not found); with the fix, the same hold then the spec’s empty answer passed 5 of 5. A first change, built on react-aria’s pointer modality, was reverted after the frontend review showed src/ui’s Tooltip opens on its own pointerenter timer.
Lesson: L-101 — a read that lands after the settle moves the control under the pointer, BUG-233’s cause again on the next tab of the same page; added to its sources.
BUG-238 — The account list’s Archive and Restore say “Nothing was changed” when a write’s outcome is unknown
Affects: F1-b (forge-et-fe src/features/accounts/AccountListScreen.tsx archiveFailureMessage, ~146 on main 41852183; src/features/accounts/accountsQueries.ts useArchiveAccount, ~358) · Found in: frontend-reviewer on forge-et-fe#316 (BUG-207’s fix), outside its diff, 2026-10-20 · Severity: low — misleading text; no wrong data is written · Phase: 1
Repro: on /accounts, archive or restore an account from its row while the PATCH gets no answer, a 5xx with no problem body, or a 2xx body the client cannot parse.
Expected: the page says it could not confirm the archive or restore and re-reads the list (L-191), as BUG-207’s fix does for contacts.
Actual: archiveFailureMessage has no isWriteOutcomeUnknown branch, so it shows “The account was not archived. Nothing was changed. Try again”, although the server may have committed; useArchiveAccount re-reads the lists only on a 409 or a 404, so a committed archive stays on screen.
Fix: branch on isWriteOutcomeUnknown before the generic message, with “couldn’t confirm” text in the attach dialog’s wording, and re-read the lists on an unknown outcome in useArchiveAccount’s onError, as contactsQueries.ts refreshAfterFailure does.
Regression test: component tests for the three unknown cases, for archive and for restore: the couldn’t-confirm text, no “Nothing was changed”, and the list re-read.
Resolved: 2026-10-22 — fixed in forge-et-fe#330 (d7da8f1c). archiveFailureMessage checks isWriteOutcomeUnknown first and says “We couldn’t confirm whether the account was archived (restored). Check the refreshed list before trying again”, the wording of useContactRowActions.ts; the reference is still appended. useArchiveAccount re-reads the lists on an unknown outcome as on a 409 or 404. Regression test: AccountListScreen.test.tsx’s six cases (a network failure, a 503 with no problem body, a 200 that does not narrow; for Archive and for Restore) failed before the fix with “The account was not archived. Nothing was changed. Try again” and pass after; a seventh case keeps a 500 with a problem body as a refusal that says nothing changed and does not re-read.
Lesson: L-191 — another instance; added to its source list in BUG-207’s harness PR.
BUG-237 — An incremental compileJava in forge-et-be fails NullAway on EmailSendController’s List<@Nullable JsonNode> reads after an unrelated class changes; a clean compile passes
Affects: FS-4 (the Error Prone + NullAway gate, forge-et-be build.gradle) and F20-h (forge-et-be src/main/java/com/empireteas/mailbox/web/EmailSendController.java:157 and :170, reading SendEmailRequest.attachmentIds() and removedSourceAttachmentIds(), declared @Nullable List<@Nullable JsonNode>) · Found in: BUG-199’s fix (forge-et-be#285), in its worktree, macOS, 2026-10-20 · Severity: low — developer loop only: a targeted test run fails to compile until a clean compile; CI and clean builds pass · Phase: 1
Repro: in a forge-et-be worktree with a compiled build/, change attachment/service/AttachmentService.java (seen after adding a constructor parameter, and again after editing a method body and adding a private constant and an anonymous class), then run ./gradlew test --tests '<any spec>'. Seen twice; a third attempt, with the inputs already in Gradle’s build cache, did not reproduce.
Expected: the incremental compileJava passes, as the clean one does.
Actual: compileJava fails with error: [NullAway] Cannot assign from type List<JsonNode> to type List<@Nullable JsonNode> due to mismatched nullability of type parameters at EmailSendController.java:157 and :170. ./gradlew compileJava --rerun-tasks then passes, and the same tests run green.
Cause (confirmed 2026-10-21): on JDK 21, javac drops a type-use annotation inside a type argument when it reads a class from its class file (JDK-8225377, fixed in JDK 22). javap -v shows @Nullable at METHOD_RETURN, location=[TYPE_ARGUMENT(0)] on SendEmailRequest.attachmentIds(), so the class file is right; javac does not attach it to the type. When Gradle recompiles EmailSendController and reads SendEmailRequest from build/classes, NullAway sees List<JsonNode>. A deterministic reproduction: compile, then change only EmailSendController.java and run ./gradlew compileJava --no-build-cache; Gradle recompiles one class and both errors appear. AttachmentService played no part. The same failure came from LeadService (:494, :499, reading LeadIntakeService.auditedFields), TemplateFillService (:74, :75) and WeeklyPricesTemplateService (:73), both reading MergeValues.valuesFor and passing it to TemplateFiller.fill. -XDaddTypeAnnotationsToSymbol=true on Oracle JDK 21.0.10 does not help. Compiling with a JDK 25 toolchain and options.release = 21 does, with the pinned Error Prone 2.38.0 and NullAway 0.12.7.
Fix: option A, code only, chosen by Tharinda on 2026-10-21 (no build.gradle change). SendEmailRequest’s lists are @Nullable List<JsonNode>, because Jackson 3.1.7 reads a JSON null entry as a NullNode, never a Java null. MergeValues.valuesFor returns Map<String, String> and leaves out a field with no value, which TemplateFiller already treated like null. LeadService reads the audit images into var. The rule is in coding-standards.md §7.2 (with a var exception in §2), the backend CLAUDE.md and backend-reviewer. Option B, the JDK 25 compile, is an owed.md row for the next analyzer upgrade.
Regression test: python3 scripts/check-incremental-compile.py (forge-et-be). It changes each production file that names a @Nullable type argument, on its own, runs an incremental compileJava, restores the file and recompiles before the next one, and stops if Gradle recompiled everything. Before the fix (the finished script run against main’s code, f2a2f443): EXIT=1, with 9 of 26 candidates failing and the 7 distinct errors above. After the fix: EXIT=0, 19 of 19 candidates compile on their own. Also EmailSendForwardAttachmentsIT sends a null and a number in removedSourceAttachmentIds (INVALID), and MergeValuesSpec pins that a field with no value has no entry.
Resolved: forge-et-be#288, 2026-10-21. The cause itself was removed on 2026-11-19 by forge-et-be#487: compileJava runs on Temurin 21 with -XDaddTypeAnnotationsToSymbol=true, so javac reads the annotation from class files and an incremental compile sees what a clean one does. The var workaround and check-incremental-compile.py were retired with it (.forge/engagement-log.md, 2026-11-19).
Lesson: L-216 — new: a check that runs only on clean builds cannot see a failure that needs an incremental one; reproduce it by changing only the failing file, and make a check force the incremental case and confirm it happened.
BUG-236 — The shadcn AlertDialog lets Option+Tab leave it after its last control: Radix’s trap loops a plain Tab only
Affects: F20-d, and every alert dialog built on forge-et-fe src/ui/shadcn/alert-dialog.tsx (AD #39’s shadcn layer) · Found in: the frontend-reviewer pass on forge-et-fe#315 (BUG-231’s fix), finding 3, 2026-10-20 · Severity: low — Safari keyboard users with the default setting move by Option+Tab, and focus leaves the alert dialog for the browser’s own controls; nothing is written · Phase: 1
Repro: read, not yet run: Radix react-focus-scope@1.3.0’s handleKeyDown loops only when event.key === 'Tab' && !event.altKey, and AlertDialog uses the same FocusScope as Dialog. BUG-231 showed the Dialog case in WebKit on macOS: Option+Tab from the last control landed on <body>. An alert dialog opened inside a Dialog is not covered by the outer dialog’s loopOptionTab, because focus is outside that dialog’s DOM.
Expected: Option+Tab past the last control moves to the first, and Option+Shift+Tab before the first to the last, as in the shadcn Dialog since forge-et-fe#315.
Actual: focus leaves the alert dialog.
Fix: move loopOptionTab and its tabbable rule out of src/ui/shadcn/dialog.tsx into a shared src/ui/shadcn module and call it from AlertDialogContent’s onKeyDown the same way, with a unit row like dialog.test.tsx’s. Left out of forge-et-fe#315 because another session was changing alert-dialog.tsx (BUG-201) at the same time.
Resolved: forge-et-fe#317 (380b5816, review fix 67577031; the PR also fixes BUG-241), 2026-10-21. loopOptionTab, isTabStop and isShown moved unchanged from dialog.tsx into src/ui/shadcn/loop-option-tab.ts. DialogContent and AlertDialogContent both call it from onKeyDown, after the caller’s handler. Two new rows in alert-dialog.test.tsx failed without the call and pass with it. Their disabled and hidden controls sit at each end, so removing either rule fails the row. The rows in dialog.test.tsx were moved the same way. In the browser pane (Chromium) on the template delete question, a real Alt+Tab from Delete reached Cancel, and Alt+Shift+Tab from Cancel reached Delete. The keydown carried altKey: true and was prevented. alert-dialog.tsx’s InertWhileOpen probably shares BUG-235’s plain-Tab gap; that is left to BUG-235.
Lesson: L-168, another instance: its sources gain BUG-236, and a bullet that every focus-trapping shadcn content calls loopOptionTab, because Radix loops a plain Tab only. frontend-reviewer dimension 23 now checks it on a new shadcn content.
BUG-235 — Plain Tab leaves a shadcn Dialog in WebKit when Tab reaches text fields only: InertWhileOpen makes Radix’s focus guards inert
Affects: F1-e and F1-f (every dialog on forge-et-fe src/ui/shadcn/dialog.tsx; select.tsx carries a copy of the same helper) · Found in: the frontend-reviewer pass on forge-et-fe#315 (BUG-231’s fix), finding 1, with a WebKit probe on macOS, 2026-10-20 · Severity: low — only where Tab skips buttons (Safari’s and macOS’s default); there, Tab from the dialog’s last text field, or from any control with no text field after it, moves focus to the browser’s own controls · Phase: 1
Repro: WebKit on macOS with keyboard navigation off. Open the note-delete dialog (focus on ✕) and press Tab: focus lands on <body>. The reviewer’s static-page probe: Tab from the last control reaches Radix’s end focus guard (<span data-radix-focus-guard tabindex=0>) when the guard is not inert, and FocusScope’s focusin handler then returns focus inside; with the guard inert it reaches <body>.
Expected: focus stays inside the open dialog (F1-e AC-29).
Actual: focus leaves it. Radix’s hideOthers marks the guards data-aria-hidden="true", and InertWhileOpen (dialog.tsx) makes every such element inert, the guards included, so the fallback never runs.
Fix: a design call, not yet made. Keeping the guards focusable while they are aria-hidden trips axe aria-hidden-focus, which is why InertWhileOpen exists. One option: skip [data-radix-focus-guard] in InertWhileOpen and take aria-hidden off the guards while the dialog is open, restoring it on close; then prove it with a plain-Tab WebKit row (Tab from ✕ stays inside) and an axe scan of an open dialog. BUG-231’s tests use Alt+Tab on such machines, so nothing covers plain Tab there today.
Decision: fix it, not accept it — the lead (Tharinda), 2026-10-21.
Cause: Radix’s two focus guards are shared by every open layer, sit at the start and end of <body>, and are how its trap catches a Tab that leaves the content: Radix’s own Tab handling wraps only at the last control it counts, buttons included, so where WebKit’s Tab skips buttons, Tab from the ✕ or Cancel goes past the content to a guard. InertWhileOpen made the guards inert with the rest of what hideOthers hid, so the Tab left the page. alert-dialog.tsx carried the same copy of the helper and had the same gap. select.tsx’s copy does not matter: Radix Select calls preventDefault on Tab inside its list.
Regression test: forge-et-fe e2e/browser/accounts/notes-tags-keyboard-webkit.spec.ts, “a plain Tab or Shift+Tab from each control stays inside the note-delete dialog”, which also runs the page check on the open dialog and checks, on any machine, that focus() on the end guard moves focus to it (the focused control’s focusout names it) and the trap returns it inside: on macOS with keyboard navigation off it failed before the fix (“Tab from delete-dialog-close stays inside the dialog”) and passes after (the whole spec 31 of 31 with --repeat-each=3). src/ui/shadcn/dialog.test.tsx’s two guard rows (open and after close; an AlertDialog opened on top) failed before and pass after; the nested row also fails with only dialog.tsx fixed; src/ui/shadcn/alert-dialog.test.tsx’s guard row covers a standalone AlertDialog.
Resolved: forge-et-fe#321, 2026-10-21: src/ui/shadcn/focus-guards.ts — the Dialog’s and the AlertDialog’s InertWhileOpen leave the guards out of what they make inert and keep their aria-hidden off while open (a MutationObserver takes it off again when a layer on top marks them). It is not put back on close, which the option above suggested: hideOthers owns the attribute, counting the layers that set it and removing it when the last of them closes, so there is nothing to put back.
Lesson: L-212 — an adaptation of a shadcn component broke what Radix keeps from upstream, here the focus trap’s guard fallback, and the helper’s tests covered only its own marks; the lesson now covers behaviour the adaptation can break, not only props.
BUG-234 — Flaky test: F11-f AC-31’s keyboard tag test sends no attach request after ArrowDown and Enter on an existing tag, once in a full gate (macOS)
macOS only so far: seen on macOS (8 cores); no Windows run is recorded.
Affects: F11-f (forge-et-fe e2e/browser/leads/lead-detail-a11y.spec.ts:138, “AC-31 (keyboard): choose an existing tag, create a tag and remove it, then add a note”; the wait at :158) · Found in: F38-g wave 1’s third full Stage 12 pnpm e2e:gate run on forge-et-fe feature/F38-g-wave-1 (53eef407), 2026-10-20 · Severity: low — one test failed once in a full gate run and passed 15 of 15 alone; the product is not known to be affected · Phase: 1
Repro: macOS, 8 cores. pnpm e2e:gate (5,198 passed, 1 failed, 5 skipped, 4 known failures; 48.2 min; default workers) against forge-et-be main bb7a2e6f on a freshly reset isolated stack (et-f38-g-wave1). The run’s load: busy 0.24 at start, 0.96 peak, 0.82 average, load average 2.79 per core at the peak; the busiest processes at the peak were the run’s own WebKit content processes (91.9 %, 38.9 %, 36.3 %), the Docker VM holding the run’s own database (64.4 %) and WindowServer (39.8 %). Run alone straight after on the same head and backend: pnpm exec playwright test e2e/browser/leads/lead-detail-a11y.spec.ts:138 --project=chromium --repeat-each=5 --workers=1, 15 of 15 passed (EXIT=0). The previous full run on the same head passed this test and failed BUG-233 instead.
Expected: typing the tag’s name, ArrowDown and Enter attach the existing tag: PUT on the lead’s tag path answers 200.
Actual: page.waitForResponse for that PUT timed out after 30 s (:65 via :158): no attach request was sent.
Fix: test-only, decided 2026-10-22 before the code. The Add tag input is src/ui’s react-aria Combobox, which sets aria-busy while its options load; the test types the name, waits only for the option to be visible, then presses ArrowDown and Enter at once — .forge/test-strategy.md § Flaky Tests cause 5 (.forge/lessons.md L-155). Reproduce it under forced load first, in Chromium and WebKit. Then, in both the choose an existing tag and the create a tag steps, wait for the input’s aria-busy="false" before ArrowDown, and for its aria-activedescendant to name the intended option after ArrowDown before pressing Enter.
Regression test: the test under forced load, repeated in Chromium and WebKit, fails before the change (or, if forced load does not reproduce it, the held options read of L-162 does) and passes after. (The browser specs run in Chromium only; WebKit runs only the size matrix.)
Resolved: forge-et-fe#332 (e1ccd824), 2026-10-22 — fixed without a reproduction, by the lead’s decision (Tharinda, 2026-10-22, in session). The change is the one .forge/test-strategy.md § Flaky Tests cause 5 prescribes: arrowDownAndChoose waits for aria-busy="false" before ArrowDown and for aria-activedescendant to name the option before Enter, in both tag steps and the match picker of the same file. Not reproduced on macOS against origin/main’s test: 10 runs under one busy loop per core and 10 more with Chrome throttled 6× all passed; a hold-back probe (a tag-list re-read held for 1 s between ArrowDown and Enter) did not start the re-read, so it proved nothing. With the change, 10 runs under the busy loops passed. The failed run’s record shows no other work loading the machine (its busiest processes were its own), so it is not an overload wontfix. A recurrence is a new bug that links this one.
Lesson: L-155 — the keyboard choice was made without the waits it prescribes; BUG-234 is added to its sources.
BUG-233 — Flaky test: the size matrix’s pricing-archive-refused state at 320 × 640 cannot click Archive in the row-actions menu, which Playwright finds outside the viewport, once in a full gate (macOS)
macOS only so far: seen on macOS (8 cores); no Windows run is recorded.
Affects: F26-b and F40, through the Pricing row-actions menu moved to shadcn/ui by forge-et-fe#262 (forge-et-fe e2e/browser/size-matrix/arrange.ts:1766 openRowDialog, used by the pricing-archive-refused state’s after step, arrange.ts:4148; state pricing('pricing-archive-refused', '/pricing', …), e2e/pageStates.ts:2396) · Found in: F38-g wave 1’s Stage 12 pnpm e2e:gate on forge-et-fe feature/F38-g-wave-1 (53eef407, main d9450ece merged in), 2026-10-20 · Severity: low — one test failed once in a full gate run and passed 15 of 15 alone; the product may be affected (a menu item drawn outside the viewport at 320px), which is not yet known · Phase: 1
Repro: macOS, 8 cores. pnpm e2e:gate (5,198 passed, 1 failed, 5 skipped, 4 known failures; 47.8 min; default workers) against forge-et-be main bb7a2e6f on a freshly reset isolated stack (et-f38-g-wave1). The run’s load: busy 0.46 at start, 0.94 peak, 0.82 average, load average 3.03 per core at the peak; the busiest processes at the peak were the run’s own chrome-headless-shell processes (102.6 %, 92.8 %, 40.1 %, 37.6 %) and node (49.8 %), with no other gate, build or backend besides the run’s own; no stalls recorded. Run alone straight after, on the same head and backend: pnpm exec playwright test e2e/browser/size-matrix/size-matrix.spec.ts --project=size-320x640-chromium --grep pricing-archive-refused --repeat-each=5 --workers=1, 15 of 15 passed (EXIT=0).
Expected: after the row’s actions button is pressed, Archive in its menu is clicked and the archive dialog opens.
Actual: locator.click on getByTestId('pricing-row-actions-990101-menu').getByRole('menuitem', { name: 'Archive', exact: true }) timed out after 30 s: the item resolved and was visible, enabled and stable, but after each scroll Playwright reported “element is outside of the viewport” and retried until the test timed out.
Fix: a test defect, not a product one. The size matrix’s Standards book (F26-b) and Weekly prices (F26-f) states pinned every read of their own but not the page’s warnings read (GET /api/v1/pricing/warnings, rendered by PricingWarnings above the tabs), which F26-g added later: 52 of the 72 size-matrix states on /pricing read it from the backend. A fresh stack answers no warnings; in a full gate the backend holds the sales other tests created and answers warning cards. When the cards landed after a state had opened its row menu, they pushed the table down; at 320px the trigger sits inside the table’s own scroll box, the menu (position: fixed, non-modal, src/ui/shadcn/dropdown-menu.tsx) followed its trigger below the 640px viewport, and Playwright cannot scroll a fixed element into view, so every retry failed. The same gate had failed pricing-set-cost the same way in another run. The product is correct: the menu stays attached to its trigger and the page still scrolls, so a user can reach it; cards that load after first paint cannot reserve their height, since their number is not known. size-matrix.spec.ts now answers the warnings read empty (pinWarnings(page, NO_WARNINGS)) for every state before its arrangement runs; pricing-warnings still pins its three cards, as the newer route, and an empty answer adds no box, so it moves nothing whenever it lands (forge-et-fe fe326488, comments e5b1a486). Triage (.forge/test-strategy.md § Flaky Tests): forced load (one busy loop per core, 4 workers) did not reproduce it, 60 of 60, nor did load with a ×6 Chromium CPU throttle, 60 of 60; holding the warnings answer until the menu was open did: with three pinned cards the menu moved to y = 971, and with the real backend’s one card 9 of 10 runs failed with “element is outside of the viewport”.
Regression test: the backend answering a warning card held until the row menu is open, pricing-archive-refused and pricing-set-cost at 320 × 640 Chromium, --repeat-each=5: 9 of 10 failed before the fix, 10 of 10 passed after; with the fix under forced load and the ×6 throttle, 30 of 30; every size-matrix state on /pricing at 320 × 640 gets a pinned warnings answer (72 of 72, none from the backend) and 112 of 112 passed; and the full gate on forge-et-fe#313.
Resolved: forge-et-fe#313, fe326488, 2026-10-20.
Lesson: L-101 — a click after the settle, with a read that lands later moving the controls under it, is BUG-151’s and BUG-230’s cause again; here the late read was one a later feature added to the page and the state never pinned, and the lesson now covers that.
BUG-232 — The size matrix’s contacts-gone state at 320 × 640 fails axe’s target-size rule on the contact rows’ Edit buttons, on main
Affects: F1-f (forge-et-fe src/features/accounts/ContactsTable.tsx, the contacts-row-edit buttons; the page state contacts('contacts-gone', '/contacts'), e2e/pageStates.ts:1933, checked by e2e/browser/size-matrix/size-matrix.spec.ts:216) · Found in: F38-g wave 1’s Stage 12 pnpm e2e:gate on forge-et-fe feature/F38-g-wave-1 (fd9a317f), 2026-10-20 · Severity: medium — a WCAG 2.1 AA target-size failure on a CRM screen at the 320px reflow size, and every full browser gate fails on it · Phase: 1
Repro: macOS, 8 cores. The gate (4,983 passed, 3 failed, 43.1 min; load busy 0.27 at start, 0.90 peak, 0.77 average) against forge-et-be 0e0a766a on a fresh isolated stack (et-f38-g-wave1). It fails the same way run alone, and on main’s own code, not only on the F38-g branch, which changes no screen: pnpm exec playwright test e2e/browser/size-matrix/size-matrix.spec.ts --project=size-320x640-chromium --grep contacts-gone, 1 failed and 10 passed, on forge-et-fe main 8c0bc93a and again on d9450ece (after forge-et-fe#262 and #308 merged), against forge-et-be main bb7a2e6f on a freshly reset isolated stack.
Expected: the shared page check (e2e/support/pageCheck.ts expectPageClean) finds nothing at 320 × 640.
Actual: one axe finding, target-size (impact serious), on every contacts-row-edit button (data-variant="outline", data-size="sm", aria-label="Edit Pinned …").
Fix: not yet known. The likely cause is the small outline button’s size at 320px, or another control placed too close to it; F1-f’s f2d8f020 gave the table’s links the 24px floor, but not these buttons. Find the change on main that made it fail, then give the buttons the 24px target (.forge/design/ui/design-system.md §9) without a page-check exemption.
Cause (found 2026-10-20, released for a layout decision): not the button’s size, and no later merge broke it: the state fails the same way on F1-f’s last gated head (3f020350), on its branch tip (a4262861) and on its merge (d6199fe4), so it has failed since F1-f added it; F1-f’s forge-et-fe#287 merged with the browser gate skipped, and its evidence did not list this state among the failures. axe’s finding is partiallyObscured, related node the row’s pinned contacts-row-name cell, plus target-offset 17.8px to Archive. Measured at 320 × 640 (forge-et-fe 41852183): the app shell keeps a 76px sidebar rail, so the table’s scroll box (contacts-table-scroll) is 206px wide; the pinned Name column is 101px, which leaves 105px beside it, and the actions cell is 154px (Attach to account is 136px). The state presses Attach to account, and Playwright scrolls it into view, so the box ends at scrollLeft 965 of 1,171 (its end), where Edit (x 156.5, 49.6px wide) sits under the pinned column (x 95–196) and 18.7px of it shows. Any table scrolled to its end at 320px does this, so it is not a test defect. Every fix is a layout choice, which is the lead’s: (a) pin the Name column only from bp/tablet (§9’s breakpoint table pins it for 768–1023; at 320px it lets wide tables scroll inside their region), which would change the Accounts and Users tables too, since they pin it “at every width”; (b) icon-only, or a row-actions menu for, the row actions below bp/tablet; © a narrower rail, or none, at 320px. scroll-padding-inline-start on the box would keep scrolled-to controls beside the pinned column but not the end of the scroll, so it is not enough on its own. forge-et-fe#311’s known-failure mark on contacts-gone stays until the fix.
Decision: the lead’s (Tharinda, 2026-10-21): keep the 320px reflow and take option (a) — pin a table’s first column from bp/tablet up only, across every table (design-system §9, Rev 68).
Fixed by: a hook, usePinnedFirstColumn (forge-et-fe src/ui/pinnedFirstColumn.ts), sets data-pin-first on a table from bp/tablet up. The shadcn tables (contacts, accounts, users, buying standards, the account page’s Buying standards panel) key their first-column cells on it through a named Tailwind group (PINNED_CELL, PINNED_CORNER); the CSS-module tables (the weekly prices round table, auction sales, the not-in-round dialog, roles, and RowSelection’s column in the two offboarding lists) select on [data-pin-first]. Two unused pinned rules in UsersScreen.module.css are deleted. src/ui/styles/shadcn.css names pinnedFirstColumn.ts as a Tailwind @source, because the classes are defined there and reach the tables only by name; without it the build generated neither rule, and the first review found that. src/ui/pinnedFirstColumn.test.tsx fails any stylesheet that pins at the inline start outside [data-pin-first], any class string with a bare sticky plus a bare start offset (left-0, start-0), and a shadcn.css without the @source line. forge-et-fe#311’s known-failure mark is removed. The regression test, contacts-gone at size-320x640-chromium, failed before the fix (1 failed, axe target-size) and passes after (11 passed); ContactsTable.test.tsx’s BUG-232 cases failed before (3) and pass after. In Chromium the contacts, accounts and roles tables pin nothing at 320px (first cell static; the first Edit, with the table scrolled to its end, is the element at its own centre) and pin at 1024px (sticky). A browser test, e2e/browser/accounts/pinned-first-column.spec.ts, scrolls the contacts and accounts tables to their end: at 768 × 1024 the first cell is sticky and moves less than 1px, and at 320 × 640 it is static and moves by the scroll distance (2 passed, Chromium). It also settles owed.md’s 2026-10-15 row on a focused row action scrolling under the sticky first column at 320px, which this change removes.
Resolved: forge-et-fe#320 (a4a18f8f, review fixes 00d98e99, c006f3a5; the three older tests that asserted the pin at 320 × 640 — F1-b AC-15, F1-f AC-22 and the size matrix’s accounts-pinned — brought into line by 960977e3 and 47800232), fixed 2026-10-21; merged as 10696c8d on 2026-10-22.
Lesson: L-219 — new: a table column pinned at every width can cover the last column’s controls at the narrowest width.
BUG-231 — The notes-and-tags keyboard tests “the dialog traps focus” and “keeps Tab inside the Modal” fail on main: Tab leaves the dialog
Affects: F1-e (forge-et-fe e2e/browser/accounts/notes-tags-keyboard-webkit.spec.ts:82 and :113, both AC-29; the check expectFocusInside, e2e/browser/accounts/notes-tags-support.ts:1005) · Found in: F38-g wave 1’s Stage 12 pnpm e2e:gate on forge-et-fe feature/F38-g-wave-1 (fd9a317f), 2026-10-20 · Severity: medium — either the note-delete dialog and the contact’s notes-and-tags Modal no longer keep keyboard focus inside them (an accessibility defect, F1-e AC-29), or the tests no longer match the dialogs; every full browser gate fails on it · Phase: 1
Repro: macOS, 8 cores, the same gate run as BUG-232. Both tests also fail run alone, and on main’s own code: pnpm exec playwright test e2e/browser/accounts/notes-tags-keyboard-webkit.spec.ts --project=chromium, 2 failed and 12 passed, on forge-et-fe main 8c0bc93a and again on d9450ece, against forge-et-be main bb7a2e6f on a freshly reset isolated stack.
Expected: after Tab, focus stays inside the open dialog (:82, deleting a note) and inside the contact Modal (:113).
Actual: Error: Tab stays inside the dialog and Error: Tab stays inside the Modal: expect.poll timed out after 5 s with Expected: true, Received: false, from deleteByKey (:100) and from the test body (:133).
Fix: not yet known. First find the merge on main that made them fail. The shadcn moves of dialogs and the app frame (for example forge-et-fe#262) are the first suspects. Then fix the focus trap, or the test if the dialog is correct and the test reads the wrong container. When it is fixed: F38-g’s wave branch split each AC-29 test so only the Tab-inside step carries the mark (forge-et-fe 5def28bf), and the unmarked note-delete test reaches Confirm with focus() instead of tabTo. The fix removes both markKnownFailure('BUG-231', …) calls and puts tabTo back in that test (e2e/browser/accounts/notes-tags-keyboard-webkit.spec.ts).
Cause: no merge on main made the tests fail; they were written on a machine whose WebKit Tab reaches buttons. On macOS with keyboard navigation off (the default, defaults read -g AppleKeyboardUIMode unset), Playwright’s WebKit Tab reaches text fields only, as Safari’s does, so a Tab from the dialog’s ✕ left the page (logged document.activeElement: BODY). Option+Tab is Safari’s key for every control, but Radix’s FocusScope loops only a plain Tab (isTabKey excludes altKey), so Option+Tab walked ✕ → Cancel → Delete → <body>. The plain-Tab case has a second cause, filed as BUG-235.
Regression test: the WebKit trap tests (Tab or Alt+Tab, chosen by e2e/support/focusKey.ts webkitFocusKey; the note-delete one now also checks the wrap at both ends, and the contact-Modal one walks one press more than the Modal has controls) and src/ui/shadcn/dialog.test.tsx’s Option+Tab row. With main’s dialog.tsx: 2 failed (“Alt+Tab past the last control wraps to the first”, “Alt+Tab stays inside the Modal”); with the fix, the file’s 16 passed on macOS. The unit row fails without loopOptionTab and passes with it.
Resolved: forge-et-fe#315 (19310fde, review fixes fa0b1537, 539d6000, 4f90bd24), 2026-10-20: loopOptionTab in the shadcn Dialog wraps Option+Tab at the first and last control as Radix wraps Tab, and the WebKit tests choose Tab or Alt+Tab by whether Tab moves between two buttons on the machine. Plain-Tab containment where Tab reaches text fields only stays open as BUG-235, and the same Option+Tab gap in AlertDialog as BUG-236.
Lesson: L-168 — a keyboard test that depends on how the machine moves focus (here macOS’s keyboard-navigation setting, not the OS) passes where it was written and fails elsewhere; the lesson now names the setting and the probe.
BUG-230 — Flaky test: the size matrix’s leads-lead-draft-compose state at 1024 × 768 in WebKit finds no AI draft subject within 5 s, once in a full gate run (Windows)
Renumbered: filed as BUG-227 on F11-d’s plan branch; an entry on main took 227 first, so this one renumbered (.claude/rules/id-numbering.md rule 5). The forge-et-fe commits 9f1ebb75 and 590e49f8 and forge-et-fe#308’s review record still say BUG-227; the code comments were corrected to BUG-230 by forge-et-fe#309.
Windows only: seen only on Windows 11; no macOS run is recorded as reproducing it.
Affects: F11-h (forge-et-fe e2e/browser/size-matrix/arrange.ts:2258, showAiDraft, used by the leads-lead-draft-compose page state’s after step, arrange.ts:7778) · Found in: F11-d’s Stage 12 pnpm e2e:gate on the forge-et-fe wave branch feature/F11-d-wave-1 at 6b26261a, 2026-10-20 · Severity: low — a flaky test; it failed one full gate run and passed 10 of 10 alone; the product is not known to be affected · Phase: 1
Repro: Windows 11, 16 cores. ET_BACKEND_PORT=8980 pnpm e2e:gate (5,158 tests, 40.9 min, default workers) against the F11-d wave backend 6fac7522 on a freshly reset isolated compose stack (et-f11-d-wave1). The run’s load: busy 0.04 at start, 0.91 peak, 0.82 average; the busiest processes at the peak were system 154.1 %, node 92.4 %, 57.0 % and 46.2 %, and chrome-headless-shell 38.5 % — the gate’s own work, no other gate, build or backend. The state failed once, in size-1024x768-webkit only; every other size passed it. Run alone on the same head with --project=size-1024x768-webkit -g leads-lead-draft-compose --repeat-each=5 --workers=1: 5 of 5 passed; the same under forced load (one busy loop per core): 5 of 5 passed. .browser-gate/result.json recorded no stalls.
Expected: after AI draft is pressed on the pinned lead, the draft’s subject (ai-drafts-subject) appears and the page check runs.
Actual: expect(page.getByTestId('ai-drafts-subject')).toBeVisible() timed out after 5 s with “element(s) not found”. The failed run’s error-context.md is under the wave worktree’s test-results/browser-size-matrix-size-m-c95c0-al-leads-lead-draft-compose-size-1024x768-webkit/.
Fix: the size-matrix states clicked a control in a lead-view panel right after settlePinnedLead, which returns before the lazy panels have mounted and before F11-d’s AI summary panel — above the draft panel — has read its triage and grown. The click landed on a shifted layout and pressed nothing. Proven on F11-d’s own state of the same shape (leads-lead-triage-tag-changed): an instrumented run showed no request after the click and the button 65 px lower than where it was clicked; waiting for the panels took it from 6 failures in 80 to 0. The second full gate run on forge-et-fe#308 failed leads-lead-draft-unavailable at the same size the same way. settlePinnedLeadPanels (settle, every leads-panel-*-loading gone, the triage summary shown) is now used by every pinned-lead state that clicks after settling, and settleLeadViewPanels by the F11-h follow-up specs (forge-et-fe 9f1ebb75). Those specs read triage from the real backend, so the panel could still grow from Preparing the summary… to the summary mid-test; their openLead now routes the lead’s triage read to a pending answer, holding the panel at one height (590e49f8).
Regression test: leads-lead-draft-unavailable and leads-lead-draft-compose at 1024 × 768 WebKit, 25 of 25 each; the F11-h follow-up specs 42 of 42 at 590e49f8; and the full gate on forge-et-fe#308.
Resolved: forge-et-fe#308 (F11-d wave 1), 9f1ebb75 and 590e49f8, 2026-10-20.
Lesson: L-101 — a click straight after a settle that does not wait for the lazily loaded panels above it is BUG-150/BUG-151’s cause again.
BUG-229 — The AI summary panel keeps offering Link to this customer while a new lead version’s first read is in flight, after a refusal whose re-read removed the proposal
Affects: F11-d (forge-et-fe src/features/leadTriage/leadTriageQueries.ts, useLeadTriage’s read and placeholderData: keepPreviousData) · Found in: the adversarial delta pass, round 3, on forge-et-fe#308 at 992601d4, 2026-10-20 (the frontend-reviewer rated the same item a Nit) · Severity: low — the window lasts one read; a second press sends the used-up ids and the server refuses it (409 STALE_VERSION or 400); the panel corrects itself when the read lands · Phase: 1
Repro: a component test (reproduced in a scratch copy, not committed): the panel on version 3 with a proposal; Link to this customer refused 409 STALE_VERSION; the view moves to version 4; the refusal’s re-read of version 3 answers with no proposal while version 4’s first read is still held.
Expected: once the refusal has judged the proposal gone (matchOpen: false), the button leaves and the refusal’s message takes focus.
Actual: the panel keeps version 3’s earlier body, with an enabled Link to this customer, until version 4 answers: TanStack 5.104.2 freezes a placeholder once shown (queryObserver reuses prevResult.data), and the panel’s useLatestCachedTriage takeover (992601d4) runs only when the key has no data, not while it has a placeholder. Focus is not lost: when version 4 answers, the button goes and the message takes focus. If version 4 brings a proposal back, the refusal’s focus request outlives the refusal.
Fix: while query.isPlaceholderData, show the lead’s latest cached answer (cached?.data ?? query.data, ranked as latestTriage ranks it), with a test row for the held-read window. In the same change: a hook row for the takeover’s isNewer guard (no test fails without it today), !Number.isFinite(version) in the ranking filter, the takeover comment (“no fresh answer: pending, placeholder, or a failed re-read”), and a comment that the ranking assumes the panel’s key is the highest cached version. Left out of forge-et-fe#308 because the review had run its three delta rounds (overnight decision 18, .forge/plans/F11-d/F11-d-overnight-log.md).
Regression test: the component row above: no lead-triage-link and the message focused before version 4 answers.
Resolved: forge-et-fe#315 (af1536a7), 2026-10-20. While the key shows a placeholder, the panel takes kept ?? cached?.data ?? query.data: the hook’s kept last good answer (which takes over every later cached answer by isNewer, and outlasts its cache entry once nothing observes it), then the cache’s latest, then the placeholder. The Fix: line’s cached?.data ?? query.data alone was not enough: an entry nothing observes is removed when its read settles (gcTime: 0 in the test client), and cached then fell back to the frozen placeholder. The regression row (TriageMatchBlock.test.tsx) failed before and passed after; the isNewer row (leadTriageQueries.test.tsx) fails with the guard removed. Seen in the browser pane against a local backend, with the reads held by a fetch wrapper: before, Link to this customer stayed enabled and focused beside the refusal’s message while version 1 was pending; after, it was gone and the message had focus.
Lesson: L-214 (new) — a TanStack placeholder is the earlier key’s answer as it was when the key changed, and does not follow later changes to that entry.
BUG-228 — The Leads inbox reads its filters back from the URL, which lags a render, and F11-d’s AI tag select and Existing customer? switch compare against that value (L-201)
Affects: F11-b (forge-et-fe src/features/leads/useLeadsUrlState.ts:84-98, the shared list-state hook), and through it F11-d’s src/features/leads/LeadAiFilters.tsx:45-47 (Radix Select, value={params.aiTag ?? ANY}) and :63-68 (Radix Switch, checked={params.customerMatchProposed}) · Found in: frontend-reviewer on forge-et-fe#308 (F11-d wave 1), 2026-10-20 · Severity: low — not reproduced; the same pattern caused F20-d’s tab-and-search race, seen only in a real-browser keyboard walk · Phase: 1
Repro: not reproduced. useLeadsUrlState parses the list parameters from useSearchParams() on every render, and React Router 7 updates them a render late. Radix Select skips onValueChange when the chosen item equals what it holds as the current value, and the Switch computes the new value from checked; so a second change made before the URL catches up (an arrow key straight after a selection, a fast second press, or a search typed straight after a filter change) can be lost or write back the old value. F11-d’s filters are the first Radix controls on this screen that compare against the URL value.
Expected: every filter change applies, however fast the next one follows.
Actual: unproven; the risk is the one L-201 records.
Fix: not decided. As F20-d did (forge-et-fe#286): hold the list parameters as React state in useLeadsUrlState, let the URL follow them, and read the URL back only on a Back or Forward navigation. It is the shared hook of F11-b’s inbox and changes every Leads filter, so it is fixed on its own rather than inside F11-d’s wave (overnight decision 16, .forge/plans/F11-d/F11-d-overnight-log.md). Decided (2026-10-22, with BUG-221): useLeadsUrlState reads and writes the list parameters through the shared src/lib/hooks/useListSearchParams.ts (BUG-221’s Fix), so params change in the render that changed them and the URL follows. The selection stays in the path.
Regression test: a component test with a held route loader: a second filter change made while the first has not rendered keeps both, and AI tag and Existing customer? show the new value at once. Then a real-browser keyboard walk that changes AI tag by arrow key and then types a search at once, 10 of 10 runs in Chromium and WebKit.
Resolved: 2026-10-22 — fixed in forge-et-fe#330 (089e746e; review fixes a9e86330, test 344caf21, 7973ec5f, 57b7eedf). useLeadsUrlState reads and writes through src/lib/hooks/useListSearchParams.ts (BUG-221’s Resolved line), so params change in the render that changed them; useLeadForm takes the screen’s own url instead of calling the hook a second time, so one screen holds one copy of the state. The selection stays in the path. First reproduced here: two held-loader tests over the real Radix controls in useLeadsUrlState.test.tsx — the select still showed the old tag while the first change was pending, and a second press of the switch was lost — failed before the fix and pass after. The real-browser walk (e2e/browser/leads/leadFiltersWalk.ts, run by lead-triage-inbox.spec.ts and the WebKit twin lead-filters-keyboard-webkit.spec.ts) passed 10 of 10 in Chromium and 10 of 10 in WebKit, and also 10 of 10 on the code before the fix: a real browser seldom leaves the render of delay the race needs, so the walk checks the controls by keyboard and the held-loader tests are the regression test (said in the walk’s header and the PR’s Deviations). Full pnpm e2e:gate on 57b7eedf: 5461 passed, 5 skipped, 0 known failures.
Lesson: L-201 — another instance; its sources gain BUG-221 and BUG-228, and it now names the shared hook and the held-loader test as the way to reproduce the lag in jsdom; frontend-reviewer dimension 4 says the same.
BUG-226 — Flaky test: the size matrix’s offboard-records-stale page state ran past its 30-second timeout once in WebKit at 768×1024 (Windows)
Windows only: seen only on Windows 11.
Affects: F5-h (forge-et-fe e2e/browser/size-matrix/size-matrix.spec.ts:216, page state offboard-records-stale, project size-768x1024-webkit) · Found in: F1-c WI-1.4’s Step 5e size-matrix run on feature/F1-c-WI-1.4-e2e-timeline-suite at c879d593 (based on feature/F1-c-wave-1 1f9d170f), 2026-10-19 · Severity: low — one page state, passes on re-run; a timeout fails the browser gate · Phase: 1
Repro: Windows 11, pn exec playwright test -c alt-port.local/playwright.config.ts e2e/browser/size-matrix (3,358 tests, 2 workers) against an isolated backend. The test timed out at 30 s inside collectPageCheckFailures at pageCheck.ts:355 (page.evaluate(measureControls, …)), about 5 hours into the run; the run’s load was not recorded. Re-run alone with its neighbours (--project=size-768x1024-webkit --grep offboard-records-stale) it passed in 8.6 s. F1-c changes no file the page renders (the offboarding page, not the account page). The same run had two other timeouts at the moment the machine went to sleep and woke (pricing-import-file, pricing-import-file-dragover, 768×1024 Chromium, 17:46 and 21:22 local), which are not filed: both passed on re-run in under 7 s.
Expected: the page state passes its page check within the timeout under a full run’s load, or the failure is shown to be overload only.
Actual: the page check’s control measurement in WebKit did not return within 30 s once.
Fix: not decided. Triage by .forge/test-strategy.md § Flaky Tests (read the load, reproduce under forced load in WebKit); a timing defect is fixed, an overload-only failure is closed wontfix with its evidence (the lead decides). Related: BUG-184 (the same class in forge-et-pwa’s size matrix).
Resolved: 2026-10-30 — fixed by BUG-218’s fix, forge-et-fe#365: the browser suite runs against a development build served by vite preview instead of the dev server, so each test’s fresh browser context loads some 70 files instead of 270 source modules, and Windows starts at most 6 browsers. In BUG-218’s three full gates in a row (--full on ece00c5a, a full ./gradlew build alongside), offboard-records-stale and every other WebKit 768×1024 state passed in all three, and no size-matrix test was flaky. BUG-184 (forge-et-pwa) is not changed by it.
Lesson: L-248.
BUG-225 — Flaky tests: two LeadsScreenActions edit-404 cases time out waiting for Select a lead in a leads-folder Vitest run under load (Windows)
Windows only: seen only on Windows 11; no macOS run is recorded as reproducing it.
Affects: F11-b (forge-et-fe src/features/leads/LeadsScreenActions.test.tsx, “out of reach and gone (AC-26a, AC-26b)”: the it.each “an edit that drops the lead (%s 404) sends no further read of it (BUG-144)” at line 477, the patch-404 case, and “an edit answered 404 NOT_FOUND says the lead is no longer available and drops it” at line 532) · Found in: the shadcn Leads list backfill (feature/shadcn-leads-list, uncommitted on #262’s ac79f4bb), 2026-10-19 · Severity: low — two unit tests; both pass alone; pnpm test:record records nothing when one fails · Phase: 1
Repro: Windows 11, pnpm exec vitest --run src/features/leads src/app src/test (83 files, 2,079 tests), with the Vite dev server on 5173 serving the same worktree, the browser pane open on it, and the backend on 8080 running beside it. No load figure was recorded: these were direct Vitest runs, not pnpm test:record. In two such runs the patch-404 case failed (once on its own, once together with the NOT_FOUND case), each at about 6 to 7 seconds, with Unable to find an element by: [data-testid="leads-no-selection"]; LeadsScreenActions.test.tsx took 118 and 130 seconds in those runs. Run alone (pnpm exec vitest --run src/features/leads/LeadsScreenActions.test.tsx), the file passed 23 of 23 both times. The branch changes only the list pane’s components and styles, which jsdom does not apply, and neither test reads a placeholder; it has not been run on main.
Expected: after the edit’s 404, the selection clears and Select a lead to see its details (leads-no-selection) shows within the wait, however loaded the machine is.
Actual: under the load of the leads-folder run the element is not found before the default findBy wait ends.
Fix: not decided. Follow .forge/test-strategy.md § Flaky Tests: reproduce under forced load (the file as parallel copies, as BUG-128 did), time the steps between the 404 and the cleared selection, and cut the slow one; or give the wait a budget that states why. Check main first, to confirm the shadcn branch plays no part. Same class as BUG-213, BUG-214 and BUG-220.
Regression test: the file run as 16 parallel copies for three rounds with no failure.
Not reproduced on macOS (2026-10-23, 8 cores, forge-et-fe main 3935f926): the file as 16 parallel copies in one Vitest run, 368 of 368 passed in 26 s; the same with eight busy loops beside it, 368 of 368 in 50 s, no timeout. Reproduce on Windows.
Not reproduced on Windows (2026-10-30, 16 cores, forge-et-fe main efd36474): the file as 16 parallel copies, three rounds, 624 of 624 passed each, no failure (slowest case 5.1, 8.3 and 7.1 s, none of them the two edit-404 cases); and three full pnpm test runs with a full Gradle build beside them, the two cases at most 2.3 s. The shadcn branch it was found on played no part. No change to the test was made.
Resolved: 2026-11-02 — closed as wontfix, an overload outcome under .forge/test-strategy.md § Flaky Tests step 4, the lead’s decision (Tharinda), proposed by Dinil: the bug’s own regression test passes on Windows and on macOS, and the report’s runs had a dev server, the browser pane and the backend beside them. No change was made. A recurrence on a machine that passed the load check is filed as a new bug that links here.
Lesson: L-162 (a failure that appears only under load is reproduced before it is called flaky) — an instance where forced load did not reproduce it, so nothing was changed; added to its source list.
BUG-224 — Flaky test: the third keyboard attach in F20-h’s AC-32 test opens no file chooser, beside F11-h’s follow-up specs (Windows)
Affects: F20-h (forge-et-fe e2e/browser/email/attachments-a11y.spec.ts:98, “AC-32 (keyboard): files attached from the keyboard; each Remove is named with its file…”, failing at line 154 in attachByKey, e2e/browser/email/attachments-support.ts ~812) · Found in: F11-h Wave 1, Stage 7 Step 5e, the orchestrator’s browser runs on forge-et-fe feature/F11-h-WI-1.4-e2e-ai-drafted-follow-up (ee7244b5) against forge-et-be 9b1131b6 on a fresh isolated stack, Windows, 2026-10-19 · Severity: low — one test fails now and then in a large parallel run and fails the gate when it does; the product is not known to be affected · Phase: 1
Repro: against a backend on a fresh local stack, run pnpm exec playwright test e2e/browser/leads/follow-up- e2e/browser/email/attachments-a11y.spec.ts --project chromium --repeat-each 2 (three workers): it failed in 2 of 4 such runs (verify-runtime/triage-wave-attach-fu.log, wi14-2-3.log). It passed 6 of 6 alone and 2 of 2 beside e2e/browser/leads/lead-detail-.
Expected: after the only attached row is removed and focus returns to Attach files, Space on the focused file input opens the file chooser, as it did twice earlier in the test.
Actual: page.waitForEvent('filechooser') times out (30 s test budget) although both focus assertions passed just before Space. e2e-test-implementer found no race in the test (the <input type="file"> is never remounted) and no product focus defect, and nothing F11-h’s specs share with it but machine load. The same symptom with Enter predates F11-h and was handled by switching to Space (forge-et-fe d0fd2097).
Fix: not decided. Reproduce under forced load with trace: 'on' (.forge/test-strategy.md § Flaky Tests, step 2) and read, for the Space press: the console’s “File chooser dialog can only be shown with a user activation”, document.activeElement at keyup, and with DEBUG=pw:protocol whether Page.setInterceptFileChooserDialog precedes the keyUp and Page.fileChooserOpened follows. A longer wait is not the fix.
Regression test: the same test passes in repeated runs beside e2e/browser/leads/follow-up- on a fresh stack.
Not reproduced on macOS (2026-10-23, 8 cores, forge-et-fe main 3935f926, against a main backend): attachments-a11y.spec.ts:98 beside follow-up-a11y, follow-up-compose and follow-up-draft specs, --repeat-each=4, four workers, eight busy loops beside the run: the AC-32 test passed 4 of 4 (three other failures were the follow-up specs’ own data under repeated runs, “No mailbox is connected”). Reproduce on Windows.
Cause found, not yet fixed (2026-10-31, Windows, forge-et-fe 12aff918; Dinil’s overnight run): Playwright turns file-chooser interception on when a page gets its first filechooser listener and off when the last one is removed, and sends the switch-on without waiting for it (playwright-core 1.63 coreBundle.js: the client’s _updateSubscription calls updateSubscription with .catch(() => {}); the server’s setFileChooserInterceptedBy sends Page.setInterceptFileChooserDialog only when the state changes). page.waitForEvent('filechooser') adds a listener and removes it when the event comes, so every attachByKey switches interception off after its chooser and on again before the next key press — a race the key press can win under load, and then Chromium opens no chooser that Playwright reports. Every attach after the first runs that race; the reports caught the third. Reproduced: the BUG’s own run (follow-up specs beside attachments-a11y.spec.ts, --repeat-each=2, three workers) with 8 busy loops beside it, four times: one waitForEvent('filechooser') timeout in eight passes of the test.
Proposed fix: keep one filechooser listener on the page for the whole test (added once by attachByKey/attachByPicker, or in the test’s set-up, followed by one awaited round trip), so interception never switches off between attaches. Not made tonight: the test also opens the account page the UI revamp changed, and fails there on main before it reaches the attach (forge-et-fe#364; its browser-test update is owed). Its regression test then needs many runs before and after, since the race hit once in eight passes.
Investigated 2026-11-03 (Dinil), not reproduced on main after forge-et-fe#371: Windows 11, 16 cores, forge-et-fe 612d3739 against a main backend on a fresh local stack, every run with busy loops beside it and --retries=0. The test passed 23 times and failed none:
- the bug’s own run (the follow-up specs beside
attachments-a11y.spec.ts,--repeat-each 2, three workers, 8 busy loops), eight times: 16 passes; - the price-email, email, follow-up and accounts folders, 8 workers beside 8 busy loops, four times: 4 passes;
- the whole
chromiumbrowser project, 8 workers beside 8 busy loops, three times: 3 passes.
The cause above is not confirmed. A script on a bare page with one file input pressed Space on it 1,800 times, with waitForEvent('filechooser') before each press as attachByKey does, under 16 busy loops, with tracing on and off. Every chooser was reported, with the listener removed between presses or kept. The proposed fix was written and not committed, because nothing shows the race it removes.
Resolved: 2026-11-03 — closed as wontfix, the lead’s decision (Tharinda), proposed by Dinil: forced load did not reproduce it on main since the UI revamp’s browser-test update, and the interception race recorded on 2026-10-31 did not show in direct tries. No change was made. A recurrence is filed as a new bug that links here, with the run’s machine load, its trace and the test’s step.
Lesson: L-162 (a failure that appears only under load is reproduced before it is called flaky) — an instance where forced load did not reproduce it, so nothing was changed; added to its source list.
BUG-223 — Flaky tests: two ComposeEmailDialog forward-attachment cases and one EmailTemplatesScreen case time out in a full Vitest run beside a backend build (Windows)
Windows only: seen only on Windows 11.
Affects: F20-h (forge-et-fe src/features/email/compose/ComposeEmailDialog.forwardAttachments.test.tsx, two cases) and F20-d (src/features/email/templates/EmailTemplatesScreen.test.tsx, one case) · Found in: F1-c WI-1.2’s full pnpm test on its branch (forge-et-fe#300, head 270a3d58), 2026-10-19 · Severity: low — unit tests that pass alone; pnpm test:record records nothing when one fails · Phase: 1
Repro: Windows 11, a full pnpm test (433 files, about 8,085 tests) with default workers while another session’s ./gradlew build with Testcontainers ran on the same machine. The three cases timed out in that one run. The three files passed alone (pnpm exec vitest run of the three, 64 passed) and in the next two full runs, which failed only on BUG-220. The run’s recorded load was not kept, so step 1 of .forge/test-strategy.md § Flaky Tests (read the run’s load) cannot be done for it.
Expected: each case passes in a full run however loaded the machine is, or the failure is shown to be overload only.
Actual: each timed out once under a concurrent backend build.
Fix: not decided. Triage by .forge/test-strategy.md § Flaky Tests: reproduce under forced load (the order D-4 gives); a timing defect is fixed, an overload-only failure is closed wontfix with its evidence (the lead decides). Related: BUG-220 (same class, F26-d).
Not reproduced on macOS (2026-10-23, 8 cores, forge-et-fe main 3935f926): each file as 16 parallel copies in one Vitest run — ComposeEmailDialog.test.tsx 432 of 432 (17 s), EmailTemplatesScreen.test.tsx 400 of 400 (14 s) — and the same with eight busy loops beside them (34 s, 24 s), no timeout. Reproduce on Windows.
Resolved: 2026-10-31 — fixed in forge-et-fe#367 (5be41157, 18fe6705, 39dc1cd4), a product defect found by the triage. EmailTemplatesScreen.test.tsx as 16 parallel copies on Windows 11 (main efd36474) failed the Clear-search case (“writes ?q= after the debounce, and Clear search removes it and focuses the search box”) 7 times in 3 rounds with expected '?q=zzz' to be '' — not a timeout. The list mirrors its state into the URL from an effect and skipped the write when the new URL matched the last rendered one; React Router renders a navigation a render or more after the write, so Clear search right after a search was written compared '' with the rendered '' and never wrote it, leaving ?q= in the URL while the list showed no search. The hook now compares with the URL it last asked for until a render shows it, and Back or Forward clears a pending request. A new case holds the router with heldLoader: it fails every run on the old code and passes now. Three rounds of 16 copies after the fix: 0 failures (496 passed each). Building the list on the shared useListSearchParams instead failed its keyboard tab case — BUG-290, filed. The two ComposeEmailDialog forward-attachment cases passed every round with no change (slowest 1.2 s).
Lesson: L-201 — a write compared with, or started from, the URL of the render that made it misses a write the router has not rendered yet; this is a fourth instance, on a list that keeps its own state.
BUG-220 — Flaky test: PriceEmailDialog’s 51-contacts Also copy case times out at 20 seconds in a full Vitest run under load (Windows)
Windows only: seen only on Windows 11; no macOS run is recorded as reproducing it.
Affects: F26-d (forge-et-fe src/features/priceEmail/PriceEmailDialog.test.tsx:232, “PriceEmailDialog — Also copy (AC-27) › with 51 contacts and 49 ticked, the unticked boxes are aria-disabled and the note shows”) · Found in: forge-et-fe#284’s pnpm test:record on 087f112b, 2026-10-18 · Severity: low — one unit test; pnpm test:record records nothing when it fails · Phase: 1
Repro: Windows 11, 12 cores, pnpm test:record (427 files, 8,017 tests) with the backend running and nothing else heavy; the script’s load line read “busy 0.54 at start, 1.00 peak, 0.99 average”, the top processes being the Vitest workers and Windows Defender. The case timed out at 20,214 ms, and again in a second full run (load 0.99 average). Run alone (pnpm exec vitest run src/features/priceEmail/PriceEmailDialog.test.tsx --reporter verbose), the file passed 18 of 18 and the case took 5,810 and 6,663 ms on the F1-e branch, and 6,840 and 6,271 ms on main 640e1451, so F1-e’s changes do not slow it; the full suite’s 12 workers make it about four times slower.
Expected: the case passes however loaded the machine is.
Actual: it renders 51 contacts and ticks 49 of them, which under full-suite load runs past the 20 s budget.
Fix: not decided. As BUG-128 did: run the file as parallel copies, time each step, and cut the slow one (for example, tick the boxes through fewer user events); or give the case a budget that states why.
Regression test: the file run as 16 parallel copies for three rounds with no failure.
Not reproduced on macOS (2026-10-23, 8 cores, forge-et-fe main 3935f926): the file as 16 parallel copies in one Vitest run, 288 of 288 passed in 13 s; the same with eight busy loops beside it, 288 of 288 in 26 s, no timeout. Reproduce on Windows.
Resolved: 2026-10-31 — fixed in forge-et-fe#367 (359a1176), test only. The case’s work was the cause, as in BUG-128: it clicked 49 boxes with userEvent, each a pointer sequence and a re-render of the dialog. The first 48 ticks, which only arrange the case, are now one fireEvent.click each; the 49th (the one that reaches the cap) and every later click stay userEvent; the walk to the cap is toggleAlsoCopy’s unit test. On Windows 11, the file as 16 parallel copies: before, the case took 12.0 s median and 12.7 s at most of its 20 s; after, three rounds, 6.1 s median and 6.4 s at most, no failure. Alone, about 6 s before and 2.2 s after. The timeout was not raised.
Lesson: L-125 — a case whose arrangement walks many items one at a time runs close to its timeout under load; a fourth instance after BUG-128 (added to its sources).
BUG-219 — Flaky test: F26-d’s send-and-marker smoke fails when another test runs beside it (Windows)
Affects: F26-d (forge-et-fe e2e/browser/price-email/price-email-send.spec.ts:122, “AC-15, AC-16, AC-17, AC-18, AC-19 (smoke): a sender without account.manage sees Send this week’s prices…”) · Found in: F1-e’s targeted browser run before merging forge-et-fe#284, 2026-10-17 · Severity: low — one test; passes alone · Phase: 1
Repro: Windows 11, backend forge-et-be main 5c457091 on a fresh database, E2E_GATE=1.
- F1-e branch,
e2e/browser/{accounts,price-email,access,shell,settings,sharing,auth}with 3 workers: failed once at:224, the marker’s tooltip not found within 5 s afterhover()(Expected: "Prices sent Sat 17 Oct to … by F26-d send (test)", element not found). - F1-e branch, the test alone with
--repeat-each 3 --workers 2: 2 of 3 failed. forge-et-femain640e1451(no F1-e),--repeat-each 4 --workers 2: 3 of 4 failed, withno price email on this account for sale Cand atoBemismatch.- Run alone with 1 worker, three times on each of
mainand the F1-e branch: 6 of 6 passed. Expected: the test passes whatever runs beside it. Actual: with a second worker it fails in most runs. Themainerrors point at shared data: two copies (or another price-email test) reach the same pool account or sale, so one copy’s price email is not the one the other reads. The tooltip failure may be the same cause or a hover that lands before the marker settles; not confirmed. Fix: test-only, decided 2026-10-22 before the code. The--repeat-eachfailures are not this flake: the setup project makes the pool accounts once per Playwright run and eachprice-email-browser-*key belongs to exactly one test (e2e/browser/price-email/support.ts, THE DATA RULES), so a repeated copy always findsprice-email-browser-1already sent to —no price email on this account for sale Cis that, by the data design. The genuine failure is the first one, on Windows: the marker’s tooltip did not open afterhover()at:224in an ordinary 3-worker run. Reproduce it on Windows under forced load with a Playwright trace (--trace=retain-on-failure), fix the cause the trace shows, and keep each test on its own pool account. No longer timeout and no retry. Tried on macOS, 2026-10-22 (Tharinda), not reproduced: againstforge-et-befix/BUG-240-source-label-formaton a fresh isolated stack,price-email-send.spec.tswithprice-email-dialog.spec.ts,--workers=3,--project=chromium: 3 runs with oneyesbusy loop per core (each run 23 of 23 passed, setup included), 3 runs with the busy loops and Chrome’s CPU throttled 6× through CDP (each 23 of 23), and 3 runs with the marker’s box and the element under the pointer logged at 0, 0.3, 1 and 3 s after the hover (all passed; the marker never moved, the pointer stayed on it, and the tooltip was open by 1 s). Two leads ruled out while reading the code: the marker stays mounted across the price-email re-read (AccountPriceEmailAction.tsxkeeps it whileread.datais defined), and the toasts sit at the bottom centre, away from the marker. The test’s hover is not after a Radix layer on the page (the compose dialog is a native<dialog>). The release (forge-et-harness#869) frees it for a Windows run, the operating system it failed on (.claude/rules/tracker.md § Bug Tracking, one bug per operating system). Seen again on Windows, 2026-10-23 and 2026-10-25 (forge-et-fe#338, which does not touch price email): Windows 11, 16 cores, 8 workers, in fullpnpm e2e:gateruns. On8b303074(busy 0.38 at start, 1.00 peak, 0.97 average)price-email-send.spec.ts:122failed at:224, the marker’s tooltip not found within 5 s after the hover. On99fcd7cb(busy 0.59 at start, 1.00 peak, 0.95 average)price-email-dialog.spec.ts:320(“AC-16, AC-19 (smoke): a contact named with markup…”) failed the same way at:358,getByRole('tooltip')not found for Prices sent Sun 25 Oct to F26-d Bold …. Both passed when re-run on their own. Seen again on Windows, 2026-10-26 (forge-et-fe#343, which does not touch price email): in the fullpnpm e2e:gateon9c07224a(busy 0.22 at start, 1.00 peak, 0.96 average, fresh database),price-email-dialog.spec.ts:320failed at:358the same way, the tooltip not found for Prices sent Mon 26 Oct to F26-d Bold …. It passed when its file was re-run on its own. Regression test: on Windows,e2e/browser/price-emailbeside other browser folders with several workers under forced load, green three times in a row (not--repeat-each, which the one-test-per-account rule rules out). Replaces the earlier the file with--repeat-each 4 --workers 2, green three times in a row. Blocked 2026-11-02 (Dinil): cannot be reproduced onmainuntil the UI revamp’s browser-test update merges (forge-et-fe#371, a draft). Onmain122e1fe8, the price-email, accounts and leads folders under forced load ran 8 workers beside 8 busy loops, with a fresh local backend. Every price-email test failed (20 of 20) before reaching the marker:settleStandards(e2e/browser/accounts/detail-support.ts:1708) waits foraccount-standards, which the revamped account page hides. forge-et-fe#371 changes that helper. Run the regression test above once it has merged. BUG-224 waits on the same PR. Investigated 2026-11-03 (Dinil), not reproduced onmainafter forge-et-fe#371: Windows 11, 16 cores,forge-et-fe612d3739against amainbackend on a fresh local stack, every run with--retries=0and--trace=retain-on-failure. Both tooltip tests (price-email-send.spec.ts:122andprice-email-dialog.spec.ts:320) passed 7 times each and failed none: - the price-email, email, follow-up and accounts folders, 8 workers beside 8 busy loops, four times;
- the whole
chromiumbrowser project (1,223 tests), 8 workers beside 8 busy loops, three times.
The regression test above passed: the price-email folder beside other browser folders, under forced load, green more than three times in a row. A possible cause found by reading the tooltip, not confirmed: Radix opens it 400 ms after the pointer moves onto the marker and never if the pointer leaves first, and once open it closes on any scroll of an element that holds the marker and does not open again until the pointer leaves and returns. A layout shift or a scroll just after the hover would therefore leave no tooltip.
Resolved: 2026-11-03 — closed as wontfix, the lead’s decision (Tharinda), proposed by Dinil: forced load did not reproduce it on main since the UI revamp’s browser-test update, and the bug’s own regression test passes. No change was made. A recurrence is filed as a new bug that links here, with the run’s machine load and its trace, which would show whether the marker moved or the page scrolled after the hover.
Lesson: L-162 (a failure that appears only under load is reproduced before it is called flaky) — an instance where forced load did not reproduce it, so nothing was changed; added to its source list.
BUG-218 — Flaky tests: the full browser gate fails on load timeouts in screens the PR does not touch, so a whole-tree run rarely passes on one machine (Windows)
Affects: the browser gate (forge-et-fe pnpm e2e:gate), across F5-c (users), F5-h (offboarding), F11 (leads), F20-c (email), F26 (pricing) and F1-d (account page) specs · Found in: F1-e wave 1, 2026-10-16 · Severity: medium — the gate cannot certify a large PR, so the developer has to decide on a skip · Phase: 1
Repro: Windows 11, ET_E2E_WORKERS=3 pnpm e2e:gate over the whole tree (about 4,100 tests, about 3 h), with other sessions’ work on the same machine.
- F1-e gate 2 (
forge-et-fe25f60867, fresh database): 4177 passed, 25 failed, 6 skipped. The run’s own load line read “busy 0.57 at start, 1.00 peak, 0.95 average”.playwright test --last-failed(2 workers): 33 passed, 1 failed. The one was a real F1-e change, fixed ina641c14a.- The failures were page headings not visible within 5 s, 30 s test timeouts in
keyboard.press,page.evaluateandframe.evaluate, and one horizontal-scroll reading. They cluster in WebKit 1024×768 and 768×1024 size states (offboard-*,account-detail-*),users/journey.spec.ts,users/users-a11y.spec.ts,pricing/weekly-prices.spec.tsand thepricing-*states.
- F1-e gate 3 (
a641c14a, stopped part-way): 14 failures in about 1,900 tests. They were inemail/reply-forward.spec.ts,leads/leads-a11y.spec.ts,pricing/weekly-prices-permissions.spec.ts,pricing/weekly-prices-states.spec.ts,users/journey.spec.ts, and the size statespricing-archived,account-detail-contacts-moreandaccount-detail-sharing-stale. users/journey.spec.tsfailed in all three runs.- forge-et-fe#338’s gate runs, 2026-10-23 to 2026-10-25 (Windows 11, 16 cores, the default 8 workers, the local backend on its compose stack, the database reset twice for BUG-173’s limit; a Leads change, so every failure below is in a screen it does not touch). Seven full runs; each of the last five failed between 1 and 6 tests, a different set each time, all at
1.00peak load with the run’s own browsers the busiest processes, and every failed test passed when re-run on its own. The developer recordedBrowser gate: skipped — <reason>on the PR.8b303074(busy 0.38 at start, 1.00 peak, 0.97 average): 5,637 passed, 6 failed —select-tab-after-choose-webkit.spec.ts(a Tab after choosing landed onapp-sidebar-toggle; 20 of 20 alone),price-email-send.spec.ts(BUG-219),standards-book-a11y.spec.tsAC-31 (BUG-211),weekly-prices.spec.tsAC-17/19 (page.reloadpast the 30 s timeout), theaccount-detail-contact-notes-tags-stalesize state in WebKit 1024×768 (the dialog’s close click past 30 s), and one Leads test the PR then fixed.3d761f8d(busy 0.14 at start, 1.00 peak, 0.96 average, fresh database): 5,641 passed, 2 failed —account-tags.spec.tsAC-21 (account-contacts-loadingpast 5 s) and one Leads test the PR then fixed.99fcd7cb, three runs: 5,642 passed and 1 failed (theaccountssize state in Chromium 1024×768,account-list-tablenot visible in 5 s, with Windows Defender and Windows Update among the busiest processes); 5,641 passed and 2 failed (BUG-261’s walk andselect-tab-after-choose-webkit.spec.tsagain); 5,638 passed and 5 failed (account-briefing.seam.spec.tsAC-15 andai-drafted-follow-up.seam.spec.tsAC-17 together, bothconnect ECONNREFUSED 127.0.0.1:5173— the dev server stopped answering for a moment —price-email-dialog.spec.ts(BUG-219),standards-book-a11y.spec.ts(BUG-211) and theaccount-detail-price-email-fill-refusedsize state in WebKit 1024×768, its briefing place still loading after 5 s).
- forge-et-fe#343’s gate, 2026-10-26 (the same machine and setup, fresh database):
9c07224a(busy 0.22 at start, 1.00 peak, 0.96 average): 5,641 passed, 2 failed — BUG-261’s walk andprice-email-dialog.spec.ts(BUG-219), neither in a screen the PR changes.
Expected: a whole-tree gate on an unchanged screen passes regardless of the machine’s other work, or the gate refuses to start or report while the machine is overloaded. Actual: at about 0.95 average CPU, some 0.5–0.7 % of the tests time out somewhere in the tree. Every one of those tested passes when re-run, so a full gate on a large PR rarely passes in one run. Fix: not decided. Options:
- the gate refuses to run, or marks its result as load-affected, above a load limit (as
scripts/local-build-evidence.pydoes at 0.65); - the gate retries a failed test once and reports it as flaky rather than failed;
- the specs named above get budgets that hold under load.
Related: BUG-128, BUG-214, BUG-215 and BUG-216 (unit tests), BUG-173 (a grown database). The design
.forge/design/harness/test-run-cost.md(draft, 2026-10-25) proposes the second option and a cheaper gate (its F-1, P-1, P-2 and S-2); this bug is resolved by its Goal 3. Regression test: a whole-tree gate at 3 workers passes three times in a row on the team’s Windows machine while another session runs a Gradle build. Resolved: 2026-10-30 — fixed in forge-et-fe#365, with its companion forge-et-pwa#103 (the PWA’sdevServerCommand.test.tsstops being a copy of the CRM’s, which #365 renames). - Cause: every browser test opens a fresh browser context, and the dev server sent the app as some 270 separate source modules, which each test fetched and compiled again. Under a full gate’s own load a page sometimes had not appeared within the size matrix’s 5-second ready wait, a WebKit page check ran out its 30 seconds, or the browser failed to fetch a lazily loaded route module, in bunches a few seconds apart. A watcher probing the dev server, its proxy and the backend every second through one full gate (2026-10-29) saw no answer slower than 0.65 s, so the server was not stalling; the cost was in the browsers.
- Fix: Playwright’s webServer runs
pnpm e2e:app, a development build (NODE_ENV=development, so the stub sign-in and React’s development build are in it) served byvite previewthrough the same/apiproxy. A cold page load on a quiet machine fell from 1.3 s to 0.8 s (/accounts, 273 responses to 73). A test that names a lazily loaded module goes throughappModulePath(e2e/support/appModules.ts), andsettleLazyFilleralso waits for the stylesheets Vite’s preload helper waits for. On Windows the gate starts at most 6 browsers (src/lib/config/e2eWorkers.ts,ET_E2E_WORKERSoverriding). - Measured before the fix (full gates on the dev server, Windows 11, 16 cores, 2026-10-29): 8 browsers, 4 flaky at 0.84 average CPU; 4 browsers, 5 flaky at 0.50 and 17 minutes longer; 6 browsers, three runs with 1, 7 and 20 flaky (the last two failing the gate), mostly size-matrix pages not ready within 5 s.
- Regression test, run at the new default of 6 browsers instead of 3: three full gates (
pnpm e2e:gate --full, 5,886 tests, both engines, every size) in a row onece00c5a, each on a fresh database with a full./gradlew build --rerun-tasksof forge-et-be running alongside for 59 minutes, all passed: 3, 5 and 4 flaky, CPU 0.78, 0.77 and 0.76 on average. No size-matrix test was flaky in any of them. Of the 12 flaky tests, 9 wereconnect ETIMEDOUT ::1:5173from Playwright’s request context (BUG-196, still open), 2 were BUG-282 (fixed onmainafter the branch was cut) and 1 the attachments keyboard test already filed. Lesson: L-248 — a browser suite that opens a fresh context per test pays the dev server’s whole unbundled module graph in every test; run it against a build.
BUG-217 — The shadcn RadioGroup misses the selection when an arrow key is released before Radix’s roving-focus timer
Affects: AD #39’s shadcn layer (forge-et-fe src/ui/shadcn/radio-group.tsx, which passes Radix’s RadioGroup through unchanged); today F20-d’s template editor Kind group (src/features/email/templates/TemplateEditorForm.tsx) and F26-d’s price-email To group · Found in: F26-d-WI-1.4’s browser suite, run by the orchestrator, 2026-10-17 · Severity: medium — a keyboard user can move focus to a radio without choosing it (WCAG 2.1 SC 2.1.1); intermittent, more often on a loaded machine · Phase: 1
Repro: in a real browser, focus the checked radio of a shadcn RadioGroup and press ArrowDown once (page.keyboard.press('ArrowDown')); on the first press on a page, focus moves to the next radio and it stays unchecked. In jsdom, a plain user.keyboard('{ArrowDown}') (not a held {ArrowDown>}) shows the same.
Expected: the selection follows focus, as Radix’s RadioGroup intends.
Actual: @radix-ui/react-radio-group 1.5.0 checks an item on focus only while an arrow key is still down (a keydown/keyup flag on document), but @radix-ui/react-roving-focus 1.2.0 moves focus inside a setTimeout; when the key is released before that timer runs, focus moves and nothing selects the item. Measured in Chromium and WebKit: the first press on a page logs keydown, keyup, timer.
Fix: in RadioGroup (radio-group.tsx), an onKeyDown after {...props} that calls props.onKeyDown and, on an arrow key, schedules a setTimeout (queued after Radix’s) that .click()s document.activeElement when it is a role="radio" in the group that is not aria-checked="true" and not aria-disabled="true" — so Radix’s own onCheck, onValueChange and the disabled rule all still apply. Then remove F26-d WI-1.2’s local workaround on the To group. The shared file needs Dinil’s go-ahead: the overnight session’s edit to it was refused by the permission system as a change to a shared resource, so F26-d ships the local workaround only.
Regression test: a radio-group.test.tsx case — focus the checked radio, user.keyboard('{ArrowDown}') (a quick press), and expect the next radio aria-checked="true"; the same case in TemplateEditorForm’s tests for the Kind group.
Resolved: 2026-10-17. Fixed in the shared src/ui/shadcn/radio-group.tsx by F1-e’s forge-et-fe#284, which had made the same fix for its own radio groups before this bug was filed: the group records that the last key inside it was an arrow key and checks the item that takes focus next through the item’s own click, so aria-disabled still refuses it. Quick-press cases are in radio-group.test.tsx. Removing F26-d’s local workaround in PriceEmailDialog.tsx and adding the Kind group case are in .forge/owed.md.
BUG-216 — Flaky test: the browser gate’s stopped-backend case measures its refusal against a 5-second wall clock, and took 15 seconds in a full Vitest run under load (Windows)
Windows only: seen only on Windows 11; no macOS run is recorded as reproducing it.
Affects: F40 (forge-et-fe src/test/scripts/browserGateRun.test.ts:323, “the backend check (AC-6a, forge-et-fe) › refuses within the probe bound when the backend is stopped, printing the start-up steps”, the bound at :340; forge-et-pwa carries a byte copy of the file, where the case does not run because the capture app has no backend probe) · Found in: BUG-128’s full Vitest run, 2026-10-12 · Severity: low — a flaky test; pnpm test:record records nothing when it fails · Phase: 1
Repro: Windows 11, 16 cores. pnpm exec vitest run (304 files, 5,324 tests, 799 s, 8 failed) on forge-et-fe fix/BUG-128-row-selection-timeout, which was main at 68c0be7 plus BUG-128’s test change. The session’s commit hooks (lint-staged Biome) and a git push ran during the run; the machine’s load was not measured (the load check came later). The case failed on expect(Date.now() - started).toBeLessThan(5000): expected 15283 to be less than 5000. Run alone on 2026-10-16 (pnpm exec vitest run src/test/scripts/browserGateRun.test.ts, at 21566ac) it passed in 1,340 ms, and the file passed 47 of 48 with 1 skipped.
Expected: the case passes however loaded the machine is.
Actual: the case starts its clock before runGate, which runs git in a fixture repository as well as the probe, so the measured time is the whole gate start-up, not the probe’s own wait. A loaded machine spends more than 5 s on that.
Fix: not decided. L-125 applies: keep what AC-6a protects — the refusal and its start-up steps, and a probe that gives up — without timing the whole run against the wall clock. Options: assert the probe’s configured timeoutMs is what was used (through the harness), or time only the probe call. Re-copy the file into forge-et-pwa with its drift manifest.
Fix (decided 2026-10-23, before the code): the first option. The case wraps probeHttp and records each call; it asserts the gate refused (exit 2) with its start-up steps, and that the probe was called once, with the configured URL and its 5,000 ms timeoutMs. No wall-clock assertion remains in it. That the probe gives up within its bound stays the job of the case “treats a backend that accepts but never answers as down once the bound passes”, which times only the probe. (The test’s own comment calls it “the next case”; it is the one after next.) A probe that never gave up would still fail the case, at its 60 s test limit. Regression test, changed from the line above (the lead accepted a deterministic proof, 2026-10-23): the case fails on main’s code when the gate’s start-up is held back past 5 s (a slow git wrapper first on PATH, since the start-up’s git calls cannot be injected), and passes with the change under the same wrapper; no Windows run is needed, because nothing the case asserts depends on the machine’s speed.
Seen again 2026-10-15 in F1-e WI-1.4’s full pnpm test on Windows 11 (feature/F1-e-WI-1.4-notes-tags-panels at 57230656, about 5,560 tests, another session’s Gradle build running): “refuses within the probe bound…” failed on its timing bound and passed alone.
Regression test: the full pnpm test on Windows, green across repeated runs, with the case still failing when the probe never gives up.
Resolved: forge-et-fe#335 (7fa2fceb) and forge-et-pwa#88 (the byte copy, its drift manifest re-recorded from #335’s merge commit), 2026-10-23, test only, as the decided fix above. Fails before: main’s case with a git wrapper that sleeps 6 s on status, expected 6076 to be less than 5000 (EXIT=1); with the change, the backend-check block passed 3 of 3 under the same wrapper (EXIT=0).
Lesson: L-125 — a time limit that assumes an idle machine; here a wall-clock bound around work the case did not mean to measure. Added to its sources.
BUG-215 — Flaky tests: three browserGateCli git-fixture cases run past their 60-second limit in a full Vitest run under load (Windows)
Windows only: seen only on Windows 11; no macOS run is recorded as reproducing it.
Affects: F40 and F41 (forge-et-fe src/test/scripts/browserGateCli.test.ts: “covers --head (C-2, C-8) › covers a later commit that changes only README.md, and not one that changes src/” :132, “› a branch that narrows the list does not make the result cover (C-9)” :172, both under the { timeout: 60_000 } describe at :84; and “dependency-only changes (F41 AC-9, AC-11a) › covers --head B: a result on A covers a #91-pair B byte for byte, and not a vite B” :326, under the describe at :283; forge-et-pwa carries a byte copy) · Found in: BUG-128’s full Vitest run, 2026-10-12 · Severity: low — flaky tests; pnpm test:record records nothing when one fails · Phase: 1
Repro: the same run as BUG-216 (Windows 11, 16 cores, pnpm exec vitest run, 304 files, 5,324 tests, 799 s, at main 68c0be7 plus BUG-128’s test change, with commit hooks and a push running; load not measured). The three cases timed out at 60 s, after 75,474 ms, 74,455 ms and 92,191 ms. Each builds throwaway git repositories and runs many separate git processes, which start slowly on Windows and more slowly while Vitest’s other workers use every core (BUG-172’s cause). Run alone on 2026-10-16 at 21566ac, the file passed 30 of 30, these cases in 4,309 ms, 4,032 ms and 7,442 ms.
Expected: the cases pass however loaded the machine is.
Actual: about ten times slower under full-suite load than alone, past a limit set at roughly eight times the idle time.
Fix: not decided. L-125 and BUG-172’s fix (a longer per-test limit for a git-heavy test, never a higher global testTimeout) is one option; fewer git calls per case (one fixture repository built once per describe and reused) is the other, and lowers the cost rather than allowing more time. Re-copy the file into forge-et-pwa with its drift manifest.
Seen again 2026-10-15 in F1-e WI-1.4’s full pnpm test on Windows 11 (57230656, another session’s Gradle build running): “covers --head B…” timed out and passed alone.
Regression test: the full pnpm test on Windows, green across repeated runs.
Not reproduced on macOS (2026-10-23, 8 cores, forge-et-fe main 3935f926): browserGateCli.test.ts as 16 parallel copies in one Vitest run, 480 of 480 passed in 39 s; the same with eight busy loops beside it, 480 of 480 in 52 s, no timeout (git runs much faster here than on Windows). Reproduce on Windows.
Not reproduced on Windows (2026-10-30, 16 cores, forge-et-fe main efd36474): three full pnpm exec vitest --run runs in a row (540 files, 10,629 passed, 1 skipped each, 394 s, 369 s and 354 s, no retries), with a full ./gradlew build --no-daemon --rerun-tasks of forge-et-be running beside them the whole time. The slowest browserGateCli.test.ts cases took 5.5 to 9.2 s against their 60 s limit (alone: 1.8 to 3.6 s); when this bug was filed they took 75 to 92 s. These runs repeat the second report’s condition (a Gradle build beside the full run) without a failure. Since the reports, the full run has gone from 799 s to about 6 minutes, and the test record refuses to start while other work keeps the CPU more than 0.65 busy, and runs one at a time with the gate and the other records (harness design test-run-cost.md); work started during a record is still not stopped. No change to the test was made.
Resolved: 2026-11-02 — closed as wontfix, an overload outcome under .forge/test-strategy.md § Flaky Tests step 4, the lead’s decision (Tharinda), proposed by Dinil: not reproducible on a loaded machine within the limit, and both reports ran beside other heavy work. No change was made. A recurrence on a machine that passed the load check is filed as a new bug that links here.
Lesson: L-162 (a failure that appears only under load is reproduced before it is called flaky) — an instance where forced load did not reproduce it, so nothing was changed; added to its source list.
BUG-214 — Flaky tests: AccountContactsPanel’s Add contact case times out at 20 seconds in a full Vitest run under load, and the next case then misses its heading (Windows)
Windows only: seen only on Windows 11; no macOS run is recorded as reproducing it.
Affects: F1-d (forge-et-fe src/features/accounts/AccountContactsPanel.writes.test.tsx: “Add contact (AC-13; AC-25 clause) › posts the four fields with the account id, re-reads, shows and announces the new contact” :236, and the next case, “› moves the Primary badge to a contact added as primary” :271) · Found in: BUG-128’s full Vitest run, 2026-10-12 · Severity: low — flaky tests; pnpm test:record records nothing when one fails · Phase: 1
Repro: the same run as BUG-216 (Windows 11, 16 cores, pnpm exec vitest run, 304 files, 5,324 tests, 799 s, at main 68c0be7 plus BUG-128’s test change, with commit hooks and a push running; load not measured). The first case timed out at 20,039 ms. The next case failed 4,656 ms later with Unable to find role="heading" and name "Contacts (3)". Run alone on 2026-10-16 at 21566ac, the file passed 33 of 33: the first case in 1,869 ms, the second in 573 ms.
Expected: each case passes however loaded the machine is, and one case’s failure does not change the next case’s result.
Actual: the first case ran past its budget under load. The second case’s failure suggests the first left work running after its timeout (a pending request or a re-read that later changed the shared mock data, so the list held a different number of contacts), but that is not confirmed.
Fix: not decided. First find the slow step in the first case (as BUG-128 did: run the file as parallel copies, time each step) and cut it. Then check whether a timed-out case’s unfinished work reaches the next case: each case should reset its mocks and wait for its own requests to finish.
Seen again 2026-10-15 and 2026-10-16 in F1-e’s full pnpm test runs on Windows 11: “posts the four fields…” failed twice on 5b45a136 (a 20 s timeout while typing about 90 characters, another session’s Gradle build running) and passed alone and on the next full run. In the forge-et-fe#284 review the same day, src/app/router.test.tsx’s four “resolves /accounts|/leads|/email|/pricing … not a 404” cases failed when run with 62 other files and passed alone (48/48) — timing-dependent findByRole checks on lazy routes, the same class.
Regression test: the file run as 16 parallel copies for three rounds with no failure, as BUG-128’s was.
Not reproduced on macOS (2026-10-23, 8 cores, forge-et-fe main 3935f926): the file as 16 parallel copies in one Vitest run, 368 of 368 passed in 10 s; the same with eight busy loops beside it, 368 of 368 in 17 s, no timeout. Reproduce on Windows.
Not reproduced on Windows (2026-10-30, 16 cores, forge-et-fe main efd36474): the file as 16 parallel copies, three rounds, 672 of 672 passed each, no failure (the slowest case 4.1 s of 20 s); and three full pnpm test runs with a full Gradle build beside them, the two cases at most 3.8 s and 1.1 s. No change to the test was made.
Resolved: 2026-11-02 — closed as wontfix, an overload outcome under .forge/test-strategy.md § Flaky Tests step 4, the lead’s decision (Tharinda), proposed by Dinil: the bug’s own regression test passes on Windows and on macOS, and the reports’ runs had commit hooks, a push or another session’s work beside them. No change was made. A recurrence on a machine that passed the load check is filed as a new bug that links here.
Lesson: L-162 (a failure that appears only under load is reproduced before it is called flaky) — an instance where forced load did not reproduce it, so nothing was changed; added to its source list.
BUG-213 — Flaky test: LeadFormDialog’s Tea requirements create case times out at 20 seconds in a full Vitest run under load (Windows)
Windows only: seen only on Windows 11; no macOS run is recorded as reproducing it.
Affects: F12 (forge-et-fe src/features/leads/LeadFormDialog.test.tsx:1052, “the Tea requirements group (F12 AC-21) › sits after the inquiry in the add form; a create sends each set tea member and no empty one”) · Found in: BUG-128’s full Vitest runs, 2026-10-12 · Severity: low — a flaky test; pnpm test:record records nothing when it fails · Phase: 1
Repro: the same run as BUG-216 (Windows 11, 16 cores, pnpm exec vitest run, 304 files, 5,324 tests, 799 s, at main 68c0be7 plus BUG-128’s test change, with commit hooks and a push running; load not measured): timed out at 20,193 ms. In an earlier full run the same day, cut short when the branch moved under it, the case took 9,245 ms and passed. Run alone on 2026-10-16 at 21566ac, the file passed 62 of 62 and this case took 2,394 ms, the slowest in the file. The same file’s AC-26 case timed out beside BUG-128 on 2026-10-09 (BUG-128’s Repro:).
Expected: the case passes however loaded the machine is.
Actual: the case is the file’s slowest, and a full-suite load takes it past its 20 s budget.
Fix: not decided. BUG-128’s cause is the first suspect: count the case’s interactions and lookups (each userEvent step re-renders the dialog), reproduce with the file run as 16 parallel copies, time each step, and cut the work rather than raise the timeout (L-125; test-strategy.md § Flaky Tests).
Regression test: the file run as 16 parallel copies for three rounds with no failure.
Not reproduced on macOS (2026-10-23, 8 cores, forge-et-fe main 3935f926): the file as 16 parallel copies in one Vitest run, 992 of 992 passed in 50 s; the same with eight busy loops beside it, 992 of 992 in 88 s, no timeout. Reproduce on Windows.
Resolved: 2026-10-31 — fixed in forge-et-fe#367 (51e6aef9; the fill helper moved to src/test/fill.ts in 4cec0232), test only. The case typed Grade and Volume key by key, 52 keystrokes each re-rendering the form; they are now pasted (one input event), and the request body, Grade trimmed, is still what the case checks (.forge/test-strategy.md § Flaky Tests cause 6). On Windows 11, the file as 16 parallel copies: before, the case took 7.1 s median and 7.4 s at most of its 20 s; after, three rounds, 2.5 s median and 2.7 s at most, no failure. Alone, 2.4 s before and 1.5 s after. The timeout was not raised.
Lesson: L-125 — typing text a case does not test is most of its time under load; added to its sources.
BUG-212 — The sign-in page’s “service unavailable” notice shows no reference id (BUG-087’s rule)
Affects: F5-a (forge-et-fe src/features/auth/LoginRoute.tsx, the unavailable sign-in marker; src/lib/session/sessionContext.ts) · Found in: the second project audit (.forge/audits/2026-10-05-project-audit-2.md), kept as an owed.md row until 2026-10-16 · Severity: low — the notice is shown; only the support reference is missing · Phase: 1
Repro: stop the backend, or make the session bootstrap fail, then open the sign-in page.
Expected: the notice ends with the failed request’s Reference <traceId> line (TraceLine), as every other failed request does since BUG-087.
Actual: no reference line. The session context keeps only a status ('loading' | 'authenticated' | 'unauthenticated') and no error object, so the sign-in page has no failed request to read a trace id from.
Fix: keep the bootstrap’s failed request (or its problem) in the session context and pass it to the notice, which renders TraceLine. It changes the session context’s interface. (The other screen the audit named, the role editor’s failed Reload, was fixed by BUG-091 in forge-et-fe#168.)
Regression test: LoginRoute.test.tsx — with the bootstrap failing with a problem that carries a trace id, the unavailable notice shows Reference <traceId>.
Resolved: 2026-10-20 — fixed in forge-et-fe#314 (24583bff, review fixes c158117a). The session context’s unauthenticated arm keeps the bootstrap’s failure as failedRequest beside reason: 'unavailable' (SessionProvider’s bootstrap catch). LoginRoute’s notice renders TraceLine (login-service-trace) outside its status region, for the last failed Try again, or else for the bootstrap; a failed refresh still leaves the store as it was. Regression tests: LoginRoute.test.tsx (the reference after the bootstrap, after a failed retry, and held while the next check runs; no line without a trace id) and SessionProvider.test.tsx (the store keeps the failure; a 401 keeps none). The three that need the fix failed before it (no login-service-trace; the probe read none) and pass after. The size matrix’s login-service-unavailable state now stubs a trace id and checks the line at every size.
Lesson: L-144 — another case of BUG-087’s “every error state” rule that its fixes did not list; BUG-212 added to its sources.
BUG-211 — Flaky test: the Standards book keyboard walk’s Add standard stays aria-disabled past the 5 s wait BUG-189’s fix added, once in a full gate run (Windows)
Affects: F26-b (forge-et-fe e2e/browser/pricing/standards-book-a11y.spec.ts:93, “AC-31: every Standards book action works from the keyboard alone, with focus into and out of each dialog”, the wait at line 168) · Found in: BUG-118’s pnpm e2e:gate on forge-et-fe#200 (331a034), Windows 11, 2026-10-16 · Severity: low — a flaky test; it failed one full gate run, and the product is not known to be affected · Phase: 1
Repro: Windows 11, 16 cores. ET_E2E_WORKERS=6 pnpm e2e:gate (4,065 tests, 54.5 min) against forge-et-be main at 6801ac6e on port 8080, on a freshly reset isolated compose stack (et-bug-118-gate). The run’s load: busy 0.12 at start, 0.99 peak, 0.87 average. The busiest processes at the peak were system 217.9 %, vmmemwsl 104.4 % (the Docker VM: the run’s own Postgres and LocalStack, and any other compose stack left up on the machine), chrome-headless-shell 84.8 %, svchost 77.2 % and node 54.5 %. None is another gate, build or backend by name. Whether another stack in the Docker VM was busy was not recorded. The test failed once in that run. Run alone against the same backend with --repeat-each=3, it passed 3 of 3, and the next full gate on the same commit with ET_E2E_WORKERS=4 passed (busy 0.70 average, 0.77 peak).
Expected: once the client lookup by id has answered, Add standard becomes allowed, and the walk tabs to it and presses Enter.
Actual: expect((await checked).ok()) passed, so the lookup had answered with a success. Then expect(add).not.toHaveAttribute('aria-disabled', 'true') timed out after 5 s. The button was resolved 14 times, every time with aria-busy="false" and aria-disabled="true". So the button was not waiting on the lookup any more, but it was still blocked for some reason. The failed run’s error-context.md and result.json were overwritten by the next run, so the reason the button showed was not kept.
Fix: not decided. First find out which block the button showed after the lookup answered. Possible causes: the lookup’s answer took longer than 5 s to render under load; or another worker changed the shared seeded client (Al Rifai Trading) or its standards while this test ran, so the answer was a block other than “Checking the client in this link”. To trace it, record the button’s reason text (its aria-describedby target) when the wait fails, and repeat the test beside the size matrix or under forced load (.forge/test-strategy.md § Flaky Tests, triage step 2). If another worker’s data is the cause, give the walk a client of its own. A longer wait is not the fix. If only an overloaded machine reproduces it, close it wontfix as triage step 4 sets out.
Regression test: the spec, passing in full gate runs on Windows with the default worker count.
Linked: BUG-189 (fixed in forge-et-fe#250), where this same walk failed under load at the step just after this one. Its fix added the wait that failed here.
Not reproduced on macOS (2026-10-23, 8 cores, forge-et-fe main 3935f926, against a main backend): standards-book-a11y.spec.ts:100 (the AC-31 keyboard walk), --repeat-each=10, four workers, eight busy loops beside the run: 10 of 10 passed. Reproduce on Windows.
Seen again on Windows, 2026-10-23 to 2026-10-25 (forge-et-fe#338, which does not touch pricing): Windows 11, 16 cores, 8 workers. It failed in two full pnpm e2e:gate runs, on 8b303074 (busy 0.38 at start, 1.00 peak, 0.97 average) and on 99fcd7cb (busy 0.59 at start, 1.00 peak, 0.95 average), and once when run on its own with five other specs right after a backend restart. Each time it was page.waitForResponse: Test timeout of 120000ms exceeded at line 145, the wait for the client lookup by id, which is earlier than the step above, so the lookup’s answer never came. It passed in the next runs on its own (70 of 70 with the five other specs, then 18 of 18).
Resolved: 2026-10-31 — fixed in forge-et-fe#368 (5660b6c5, c98bcad8, 74632890; merged ad7c6e63), a product defect. useStandardsBookUrlState wrote with React Router’s functional setSearchParams, which starts from the URL of the render that made it, so a change made before the previous one had rendered dropped it: Show archived pressed right after a client was chosen wrote the URL without accountId, and Add standard stayed blocked with no client — the block this walk saw after the lookup had answered (BUG-291’s class; BUG-221 and BUG-228 before it). The hook now writes through useListSearchParams, which starts each change from the URL the screen last asked for. Two held-loader tests fail on main and pass now: the hook (expected null to be '7') and the screen (PricingRoute.test.tsx: the real Show archived switch does not show on while the client is pending; after the fix the URL keeps accountId=12, the book reads with the client, and Add standard is allowed). On Windows the spec ran --repeat-each=5 beside 8 busy loops: 45 of 45. The 2026-10-23 to 25 symptom (no lookup by id at all) fits the same cause if a later write built on the stale URL replaced the client before it rendered; that was not reproduced. The book moved onto the hook before BUG-290’s fix, with the reason in BUG-291 and the PR.
Lesson: L-201 — a write started from the URL of the render that made it loses a change the router has not rendered yet; added to its sources.
BUG-210 — Contact email links do not encode ?, & or #, so a stored address can pre-fill extra mail headers
Affects: F1-d, F1-f (forge-et-fe src/features/accounts/contactDisplay.tsx ~34, and F1-d’s ContactRow) · Found in: frontend-reviewer on the F1-f wave diff (98106758), 2026-10-16 · Severity: low — it cannot run script, because the scheme always comes first; a crafted stored email could pre-fill a BCC or a subject in the user’s mail client · Phase: 1
Repro: save a contact whose email is a@b.example?bcc=x@evil.example, then open the email link from the contacts list, the contact view or the account page.
Expected: the link opens a message to that one address, with nothing pre-filled.
Actual: the link is mailto:a@b.example?bcc=x@evil.example, so the mail client adds a BCC.
Fix: percent-encode ?, & and # in the address when building the mailto: href (one helper used by both places), or confirm that the backend’s email validation rejects those characters and say so in the helper.
Regression test: a component test that a stored address with ?bcc= gives an href with no query.
Fixed by: contactLinkHref (src/features/accounts/contactLinkHref.ts) percent-encodes %, ?, & and #; F1-d’s ContactRow and F1-f’s contactDisplay both build their mailto: and tel: links with it. A , stays as it is: it adds a recipient the user sees in To. The regression tests (contactDisplay.test.tsx, AccountContactsPanel.test.tsx, “percent-encodes ?, & and #”) failed before the fix, 2 of 2, and pass after; a tel: # test was added in review. In Chromium against main’s backend, the contacts row’s link is mailto:a@b.example%3Fbcc=x@evil.example, with no query.
Resolved: forge-et-fe#316, 81dad2d0, 2026-10-20.
Lesson: none — one-off: of the links the CRM builds from a stored value, the account’s website already goes through its own check (websiteHref, AccountProfileGrid.tsx) and message-body links through the reviewed SafeMessageBody allow-list; the contact email and phone links were the ones without one, and both builders now share contactLinkHref. No other Lesson: line or lesson names this cause.
BUG-209 — The shared account picker’s Retry unmounts when pressed, because loadFailed clears during the refetch
Affects: F1-d, and every consumer of src/lib/accountPicker/useAccountPicker.ts (~93) · Found in: frontend-reviewer on the F1-f wave diff, 2026-10-16 · Severity: low — a keyboard user loses their place; nothing is written · Phase: 1
Repro: in an account picker whose search failed, press Retry with the keyboard.
Expected: Retry stays (busy) until the next result, and focus stays on it if the retry fails again.
Actual: loadFailed is TanStack Query’s isError, which v5 clears when the refetch starts, so Retry unmounts at the press and focus falls to the page.
Fix: hold the failed state until the next result in the hook (as F1-f’s AttachAccountPicker does locally with src/lib/hooks/useRetryHold.ts), and drop F1-f’s local hold once the hook does it.
Regression test: a hook or component test that Retry stays rendered and focused through a retry that fails again.
Fixed by: usePickerData (src/lib/accountPicker/useAccountPicker.ts, was pickerData), used by useAccountPicker and useContactPicker, holds the failure from the Retry press until the next result (useRetryHold, own data only, never the previous search’s placeholder); RecordSearchPicker’s Retry is busy meanwhile, and, from the review, its input takes focus when Retry leaves after a retry that answers. AttachAccountPicker’s local hold is removed; LeadMatchModal and AttachContactDialog get the hold through the hook. The regression tests (useAccountPicker.test.tsx, one per picker, “…through a retry that fails again”) failed before the fix, 2 of 2 (Retry gone after the press), and pass after. In Chromium against main’s backend, focus stays on the attach dialog’s Retry through a keyboard retry that fails again.
Resolved: forge-et-fe#316, 20c8644d and 4565759e, 2026-10-20.
Lesson: L-183 (a control removed by its own action loses focus) — another instance; the lesson now says to hold the failure in the shared hook and to give the Retry that leaves after a success a focus target.
BUG-208 — The account list’s first-read Retry unmounts when pressed, so focus falls to the page
Affects: F1-b (forge-et-fe src/features/accounts/accountsQueries.ts ~215, useAccountsList) · Found in: frontend-reviewer on the F1-f wave diff, 2026-10-16 · Severity: low — a keyboard user loses their place; nothing is written · Phase: 1
Repro: on /accounts, make the first list read fail, then press Retry with the keyboard.
Expected: the error view stays with Retry busy, and a retry that fails again leaves focus on Retry.
Actual: with no snapshot the view returns to loading at the press, Retry unmounts, and focus falls to the page.
Fix: keep the error view during the retry ({ state: 'error', error, retrying }), as the account record read already does (accountsQueries.ts ~355-363) and as F1-f’s contacts list now does.
Regression test: a component test with a retry that fails again and focus kept on Retry.
Fixed by: useAccountsList keeps { state: 'error', error, retrying } through a first read’s Retry when no rows are kept, as the account record read does; AccountListScreen shows Retry busy and treats a retrying error as not settled. The regression test (AccountListScreen.test.tsx, “…one that fails again keeps focus on Retry (BUG-208)”) replaced the test that asserted the old behaviour; it failed before the fix and passes after. In Chromium against main’s backend, focus stays on the same Retry through a keyboard retry that fails again.
Resolved: forge-et-fe#316, d965d288, 2026-10-20.
Lesson: L-183 — another instance.
BUG-207 — F1-d’s contact edit and archive Modals say “Nothing was changed” when a write’s outcome is unknown
Affects: F1-d (forge-et-fe ContactFormModal.tsx ~114, ArchiveContactModal.tsx ~42, AccountContactsPanel.tsx), and F1-f’s contacts page, which reuses both Modals unchanged (useContactRowActions.ts saveEdit ~336, confirmArchive ~377) · Found in: frontend-reviewer on the F1-f wave diff, 2026-10-16; widened by its delta round 2 the same day · Severity: low — misleading text; no wrong data is written · Phase: 1
Since F1-f’s review fix N-d, usePatchContact re-reads the lists before rejecting on an unknown outcome, so on the contacts page an archive that did commit now removes the row behind a Modal that says nothing changed, and pressing Archive again answers 409. The fix below covers the Modals, the contacts page’s saveEdit / confirmArchive, and F1-d’s account panel, with an unconfirmed outcome as F1-f’s attach dialog has.
Repro: in either Modal, make the save’s outcome unknown: a network failure after the request is sent, a 5xx with no problem envelope, or a 2xx body the client cannot parse.
Expected: the Modal says it could not confirm the change and offers a re-read (L-191, coding-standards.md §3.2), as F1-f’s attach dialog and restore now do.
Actual: each shows CONTACT_NOT_SAVED_MESSAGE (“Nothing was changed”), although the server may have committed.
Fix: branch on isWriteOutcomeUnknown before the generic failure, with “couldn’t confirm” text and a re-read of the contact.
Regression test: per Modal, component tests for the three unknown cases.
Fixed by: every contact write in AccountContactsPanel (add, edit, archive, Make primary, Restore) and useContactRowActions (saveEdit, confirmArchive) checks isWriteOutcomeUnknown before the generic failure, in F1-f’s attach-dialog wording. Edit: the form says “We couldn’t confirm whether the contact was saved. Reload to check before saving again”; Reload re-reads the contact, and still at the version the save was sent with it was not saved, so what was typed stays under “The contact was not saved. Nothing was changed”, while a newer version refills the form; a failed Reload keeps the notice and asks for Reload again. Archive: the confirmation’s notice with Reload, which closes it and re-reads the lists. The panel’s row actions: the failure panel’s couldn’t-confirm text with Reload in place of Try again. Add: the lead’s decision (Tharinda, 2026-10-20) — a create has no version, so a second Save would add the contact twice; the form closes, the panel re-reads its lists and shows “We couldn’t confirm whether the contact was added. Check the list before adding again”, as the contacts page does for an unconfirmed Restore. The new states have size-matrix entries (account-detail-contact-unconfirmed, -contact-add-unconfirmed, -archive-unconfirmed, -contact-row-unconfirmed, contacts-edit-unconfirmed, contacts-archive-unconfirmed). The regression tests (AccountContactsPanel.writes.test.tsx, ContactsScreen.actions.test.tsx, “BUG-207”) failed before the fix, 19 of 19, and pass after; the review added Restore, the committed add and the committed edit. In Chromium against main’s backend, the edit and archive Modals show their notices on a 503 with no body.
Resolved: forge-et-fe#316, a7531281, 4565759e and 6b729c14, 2026-10-20.
Lesson: L-191 — another instance of BUG-132’s class; the lesson now covers an unknown create, which has no version to make a second Save safe, and a re-read that proves nothing committed.
BUG-227 — Flaky test: the price import seam’s VALIDATION_FAILED case times out walking the auction-sales list in a full gate (Windows)
Affects: F26-c (forge-et-fe e2e/seam/price-import.seam.spec.ts:1784, “VALIDATION_FAILED: row 1 refuses endsOn before startsOn, a missing saleYear, out-of-range numbers and an undeclared status, writing nothing”; the auction-sales list walk at :1022) · Found in: forge-et-fe#262’s pnpm e2e:gate runs on Windows: on 2026-10-19 with 8 workers (stopped run, ac79f4bb) and on 2026-10-20 with 4 workers (480cd60e, busy 0.58 at start, 0.79 peak, 0.72 average), against forge-et-be main 0e0a766a · Severity: low — one seam test fails the gate in some full runs; the product is not known to be affected · Phase: 1
Repro: not reproduced alone. Run the full pnpm e2e:gate on a local database a few gate runs old (the 2026-10-20 failure had 219 auction sales and about 800 leads).
Expected: the test’s result does not depend on how much data earlier runs left or how busy the machine is.
Actual: the test hit its 30 s timeout while reading GET /api/v1/auction-sales?page=2&size=100, walking every page of a list that each gate run adds to; the request context was then disposed. Run alone, pnpm exec playwright test e2e/seam/price-import.seam.spec.ts:1784 --project=seam --repeat-each=3 --workers=1 passed 3 of 3. The same gate run passed every other test (4889 passed).
Fix: test-only, decided 2026-10-22 before the code. GET /api/v1/auction-sales filters by status only, sorts by startsOn only (default startsOn,desc) and has no read by id, so the test cannot ask for one sale. Each run adds sales on 2026-10-20, and the VALIDATION_FAILED case walks the whole list once per refused request (five times). Add a bounded walk that stops at the first item older than a given startsOn, give the refused requests a date range no other test writes so a wrongly saved one would sort to the top, and bound the file’s other walks the same way where the test knows the date (saleListItemById, the duplicate case). Also check the list query for a per-row query (N+1) that makes a 100-row page take seconds under load; a product defect found there is filed as its own bug. A longer timeout is not the fix (.forge/test-strategy.md § Flaky Tests).
Regression test: on a database with several hundred extra auction sales, the case under forced load times out before the change and passes after; then the case passes in repeated full gates on a database several runs old.
Resolved: forge-et-fe#332 (c9234f36, review fixes e19fb942, 93a1fa81, 436c1535, 7a76697b), 2026-10-22. The list query was not the cost (one statement; draftCount is a @Formula subquery on the indexed anchor_sale_id); the number of pages walked was. Refused row 1 requests now use 1901 dates and are checked oldest first up to them (one page); AC-1’s sale and duplicate check use 1902 dates the same way; findSale stops at the sale; the held walk stops after the seeded held sale; saleListItemById reads backwards from the last page, since within one start day the list’s id tie-break is ascending and a just-created sale follows every earlier sale of its day; it stops at a page holding a sale that lists before it and reads the next page once more, for a sale a concurrent create moved on. On an isolated database with 2,000 extra sales on 2026-10-20, under one busy loop per core: the VALIDATION_FAILED case on origin/main’s spec failed 3 of 3 (Test timeout of 30000ms exceeded at the walk, the bug’s own failure); with the fix it passed 3 of 3, twice; the whole file then failed AC-18 at saleListItemById, which 93a1fa81 fixed, and passed 35 of 35. Remaining growth: the sales sharing 2026-10-20 (the contract’s test-sale date) are still read by seededScheduledSale, which stops at sale 42, the first of that day, and the browser suite’s findSeededRound (e2e/browser/pricing/round-support.ts readSalesViaApi) still reads every sale — with the 2,000 extra sales it timed out weekly-prices.spec.ts’s hook, and on the database without them that file passed. The gate’s local-data limit resets the database after about four runs, so a real gate database stays a few pages; a test that must survive more is a new bug under L-147.
Lesson: L-147 — run-unique fixtures grow a shared list, here the auction sales, which the gate’s limit does not count; BUG-227 is added to its sources.
BUG-222 — Flaky gate setup: the route warm-up itself fails to load the login route on a dev server’s first start (Windows)
Affects: F5-a (forge-et-fe e2e/fixtures/routes.setup.ts:89, “load every lazy route, panel and compose module once, so no spec pays the first transform”, BUG-079’s fix; every project depends on setup) · Found in: forge-et-fe#262’s first pnpm e2e:gate --push on ac79f4bb, the first dev-server start after a merge of main (d44d2877), Windows, busy 0.22 at start and 0.31 peak, 2026-10-19 · Severity: low — the gate stops in setup, skips every other test and pushes nothing; a rerun passed · Phase: 1
Repro: not reproduced on demand. Start pnpm e2e:gate on a checkout whose Vite cache predates a large merge, so the dev server’s first requests transform much of the app at once.
Not reproduced on macOS (2026-10-23, 8 cores): pnpm exec playwright test --project=setup against a main backend, 8 attempts, all 10 of 10 passed: 3 with node_modules/.vite deleted first (cold cache, 4.5–11.9 s); 3 with the cache warmed on main at 3f697cb8 (2026-10-14), then the checkout moved to main at 3935f926 (743 files and the lockfile changed) and pnpm install run, so the dev server’s first start followed a large merge as in the report (5.4–16.4 s); and 2 of those with eight busy loops beside them (7.9–11.7 s). So the cause is not a plain cold or stale Vite cache on macOS; reproduce on Windows, where it was seen.
Expected: the warm-up loads every route module, however long the first transforms take.
Actual: the warm-up’s import of src/features/auth/LoginRoute.tsx failed with “Failed to fetch dynamically imported module” after 2.3s, and the gate reported 9 passed, 1 failed and 4887 skipped. pnpm exec playwright test --project=setup then passed 10 of 10, and the gate’s rerun passed setup in 3.1s.
Fix: not decided. Find why the first import failed (a dev-server restart or dependency re-optimisation during the first requests, or a request that timed out), then make the warm-up survive it — for example by retrying a module whose import failed once, after the dev server has answered. A longer timeout alone is not the fix (.forge/test-strategy.md § Flaky Tests).
Regression test: the gate’s setup passes on the first start after a merge that changes many modules.
Resolved: 2026-10-30 — fixed by BUG-218’s fix, forge-et-fe#365 (merged efd36474): the browser suite no longer runs against the dev server, so nothing transforms or re-optimises dependencies on its first requests. Playwright starts pnpm e2e:app, which builds the app from scratch (--emptyOutDir) and serves the files with vite preview; every gate is a first start, and routes.setup.ts imports the built chunks. Checked on Windows 11 at efd36474 against a fresh local backend: pnpm e2e --project=setup three times, each with node_modules/.vite and .browser-gate/app deleted first and 16 busy loops on the 16 cores — 10 of 10 passed each time (1.2 to 1.3 minutes). Every full gate since the change has passed its setup too (BUG-218’s three, and forge-et-fe#365’s final one).
Lesson: L-248.
BUG-221 — Clear filters can be undone by the search box’s pending debounce, which puts the cleared filters back
Affects: F1-f (forge-et-fe src/features/accounts/ContactsScreen.tsx, clearFilters and the search debounce), F5-c (src/features/users/UsersScreen.tsx, the same) · Found in: forge-et-fe#262 after its merge of main (d44d2877), where e2e/browser/accounts/account-tag-filter.spec.ts:78 failed one run in three: after Clear filters the list was read with no filters, then about 580ms later with tagId again; 2026-10-19 · Severity: medium — Clear filters appears to work, then the list narrows again without the user doing anything · Phase: 1
Repro: on a list screen with a search applied and a filter set (on the account list before 5f21b53f: /accounts?q=<text>&tagId=<id>), press Clear filters while the page renders slowly (a loaded machine; the account list’s own browser test hit it one run in three).
Expected: the URL and the list stay cleared.
Actual: the cleared filters come back. Clear filters empties the search box at once and changes the URL in a navigation that renders later. In between, the box (empty) and the URL (still holding the search) disagree, so the search debounce sets a 300ms timer. Its write goes through React Router’s functional setSearchParams, which starts from the URL of the render that made it, the one still holding the filters, so if the clear has not rendered within 300ms the timer writes the old filters back with an empty search.
Fix: on the account list, forge-et-fe#262 (5f21b53f) builds every write from the URL the screen last asked for, reset from the rendered URL after a navigation from elsewhere. The contacts and Users screens share the pattern (setQueryText('') beside a functional setSearchParams) and are not fixed or reproduced; the same change fits each, or one shared hook for the three. The email templates screen’s Clear search clears only the search, so a late write there changes nothing. Decided (2026-10-22, with BUG-228): one shared hook, src/lib/hooks/useListSearchParams.ts, a drop-in for React Router’s useSearchParams: it holds the URL the screen last asked for as React state, renders from it, starts every functional write from it, and adopts the rendered URL only when it changed from outside (a URL this screen did not write: Back, Forward, a link). The account list moves onto it from its local ref, and the contacts, Users and Leads screens take it.
Regression test: for each screen, a component test with a held route loader (the account list’s a write while an earlier navigation is pending, forge-et-fe#262), which holds the Clear filters navigation pending while the search debounce fires; plus the hook’s own tests. The browser test this line first named: for each screen, a browser test that applies a search and a filter, presses Clear filters, and checks the URL is still clear a moment later, repeated (--repeat-each 10), as account-tag-filter.spec.ts:78 does for the account list.
Resolved: 2026-10-22 — fixed in forge-et-fe#330 (1bb3f16f; review fixes a9e86330). A new hook, src/lib/hooks/useListSearchParams.ts, is a drop-in for React Router’s useSearchParams: it holds the query string the screen last asked for as React state, renders from it, starts every functional write from it, checks each rendered navigation (location.key) once, adopts every Back and Forward, and treats any other navigation as its own write only when its query string is one of its pending writes. The account list moves onto it from forge-et-fe#262’s ref; the contacts and Users screens take it, and Contacts’ Clear filters starts from the params last asked for, as the account list’s does. src/test/heldLoader.ts holds the account list test’s held-loader helper for the four test files that use it. Regression tests: the contacts and Users screens’ held-loader tests failed before the fix (?hasAccount=true and ?active=true came back after Clear filters) and pass after; the hook’s eight tests, of which the three race tests fail on plain useSearchParams and two of the three navigation tests added in review fail on the hook’s first version. The browser test this entry first named was not added: the held-loader tests reproduce the race every run, where the browser test caught it about one run in three (the PR’s Deviations). Full pnpm e2e:gate on 57b7eedf: 5461 passed, 5 skipped, 0 known failures.
Lesson: L-201 — another instance (with BUG-228); its sources gain both, and its lesson now covers a timer’s write from a stale URL and the held-loader way to reproduce the lag.
BUG-206 — On a lead, the Add tag list can still run off a window shorter than about 640px, a zoomed page included
Affects: F11-f (forge-et-fe src/features/leads/LeadTagsSection.tsx, ADD_TAG_OPTIONS_MAX_SMALL; src/features/leads/addTagWindow.ts) · Found in: the round-3 delta review of forge-et-fe#262, which reviewed BUG-203’s fix (22e30406), 2026-10-16 · Severity: low — no supported size is that short, but a short browser window or a zoomed-in page is · Phase: 1
Repro: not yet reproduced; the arithmetic says it fails. On forge-et-fe feature/shadcn-ui-review (cda68e13 or later), open a lead whose tag read holds more than seven tags, in a window under about 640px tall — for example 1280 × 600, or a 1280 × 720 window zoomed to 200%, which is 640 × 360 CSS px — scroll the field to the window’s middle and open Add tag.
Expected: the options and the “Type to find ADD_TAG_ROOMY_HEIGHT). Seven rows and the line need about 290px on one side of the 44px field, so they fit only from about 640px tall with the field mid-window (ADD_TAG_OPTIONS_MAX_SMALL’s comment: seven fit at 320 × 640). Below that the list can run off the window as BUG-203’s did, and the size matrix’s shortest size, 320 × 640, does not check it.
Fix: not decided. Options: fewer options again below about 640px tall (a second min-height step); or let the list scroll in a window too short for any bounded count, which reopens BUG-123’s axe finding there.
Regression test: a browser check in a window under 640px tall with the field scrolled to the window’s middle, as lead-detail-a11y.spec.ts “the tag list fits the shortest window that lists ten, with the field mid-window” does at ADD_TAG_ROOMY_HEIGHT.
Fix (decided 2026-10-23, from measurements before the code): the first option, a second step. Measured in Chromium: an option row is 30.1px, and the popover is the rows plus 40.5px, opening 6.6px from the 44px field; with the field mid-window and the library’s 12px edge padding, seven options fit from about 584px tall, and three from about 345px. So below ADD_TAG_SHORT_HEIGHT (600px) Add tag lists ADD_TAG_OPTIONS_MAX_SHORT (three), at any width; src/features/leads/addTagWindow.ts holds both, and the browser helper reads them. The arithmetic above (seven fit only from about 640px) was pessimistic: 1280 × 600 already fit.
Resolved: forge-et-fe#335 (8f05743b), 2026-10-23. Fails before: the new lead-detail-a11y.spec.ts checks at 1280 × 600 and 640 × 360 (a 1280 × 720 window zoomed to 200%), with the field mid-window: 640 × 360 failed, the last part ending at 227.9px against the popover’s edge at 147.5px; 1280 × 600 passed. With the fix all three sizes (with 1280 × 860) passed 3 of 3 each.
Lesson: L-195 — a list kept short so it never scrolls is sized by the height it needs; BUG-206 is the height case BUG-203 left below 640px. Added to its sources.
BUG-205 — Flaky test: in F20-h AC-32’s keyboard test, the attachment list’s Retry after an armed failure on a forward still shows the error in most runs (Windows)
Affects: F20-h (forge-et-fe e2e/browser/email/attachments-a11y.spec.ts:164, “AC-32 (keyboard): on a forward, the list’s Retry after a failed read puts focus on the first original’s Remove…”; armAttachmentsFailure and settleForwardList in e2e/browser/email/attachments-support.ts; the stub’s armAttachmentsFailure in forge-et-be MailboxStubService) · Found in: F20-d Wave 1, Stage 12, a targeted run of e2e/browser/email on forge-et-fe feature/F20-d-wave-1 (1d4b3d6f) against forge-et-be d1982f18 on a fresh isolated stack, Windows, 2026-10-16 · Severity: low — one test fails in most runs, which fails the gate when it does; the product is not known to be affected · Phase: 1
Repro: against a backend on a fresh local stack, run pnpm exec playwright test e2e/browser/email/attachments-a11y.spec.ts:164 --repeat-each 3 --workers 1 --project chromium.
Expected: the test arms one failed read of the original’s attachments (stub-attachments-failure FAILED, which the stub says fails the caller’s next read once), sees the error, presses Retry, and the originals are listed.
Actual: Retry lists the originals times out with the list still error (attachments-support.ts:1067). On forge-et-fe main (0e07dd7a, no F20-d code) it failed 2 of 3 serial runs; on F20-d’s wave head before and after its review fixes, 2 of 3 and 4 of 5. The same test passed in F20-d’s full gate on 926d0a4f against forge-et-be 3082b86c, and F20-d changes nothing in the forward’s attachment list or the stub. So Retry’s read sometimes meets a failure too. Possibly the armed failure is consumed by a read other than the one the test expects (an extra read on open), or a later read is still served the failure; not traced.
Fix: not decided. Log the stub’s reads of the message’s attachments during the test to see which read takes the armed failure, then make the arming or the test match the reads the dialog makes. A longer wait is not the fix (.forge/test-strategy.md § Flaky Tests).
Regression test: the same test passes in repeated serial runs on a fresh stack.
Also seen in F1-f’s browser gate, 2026-10-16: it failed in attempts 2 and 3 on forge-et-fe feature/F1-f-wave-1 with ET_E2E_WORKERS=2 (“the forward list is loaded, failed with Retry idle, or absent” after 14 s), passed alone 13 of 13, and passed in a three-worker run. The test budgets exactly five of the default principal’s ConnectionRateLimiter tokens (F20-h WI-1.4 plan Decision 5), so another email spec running beside it under the same principal can also spend them — a second cause to rule out while tracing.
Resolved: 2026-10-19 — fixed in forge-et-fe#301 (F11-h wave 1, commit 3b90b5a8, fix(BUG-205)). The cause was in the test helper, not the stub or the product: forwardListState counted the error block and then called the auto-waiting getAttribute('aria-busy') on Retry, so when Retry’s read succeeded between the two calls it waited for a Retry that had gone until the 14 s poll gave up; a failure’s trace showed the read answering 200 in 59 ms with the list loaded on screen. The helper now reads the list’s state in one evaluateAll. The test itself is the regression test: it failed in every full run of F11-h’s browser suite before the fix and in 2 of 4 runs beside F11-h’s follow-up specs, and after it passed in F11-h’s command 2 (run 21, 329 passed) and in the wave’s browser gate (0084aef6, 4985 passed) (.forge/plans/F11-h/F11-h-WI-1.4-test-verify.md).
Lesson: L-206 (new, in this PR): a polled state reader reads the page once and makes no auto-waiting call. The token budget named in the Also seen line was ruled out: the limiter is per connection.
BUG-204 — Two tag seam tests on forge-et-fe main expect account tags to be refused, which F1-e’s backend on main now accepts
Affects: F11-f and F11-b (forge-et-fe e2e/seam/lead-detail.seam.spec.ts:1477, “F11-f Wave 1 contract row 7: TagCreateRequest refusals…”, and e2e/seam/leads-inbox.seam.spec.ts:1516, “AC-9 (smoke): row 2 lists the active lead tags…”); caused by F1-e merging its backend before its frontend · Found in: F20-d Wave 1, Stage 12’s pnpm e2e:gate on forge-et-fe feature/F20-d-wave-1 at 926d0a4f against forge-et-be feature/F20-d-wave-1 at 3082b86c (which contains main through 934cac3e), 2026-10-16 · Severity: medium — every frontend browser gate run against the backend on main fails these two tests until the fix lands; the product is not affected · Phase: 1
Repro: run either test against a backend built from forge-et-be main at or after 934cac3e (forge-et-be#265, F1-e Wave 1).
Expected: each seam test states the contract the backend on main serves.
Actual: F1-e’s Wave 1 contract makes account a valid entityType for POST /api/v1/tags and adds an active filter to GET /api/v1/tags. The backend half (forge-et-be#265) is on main; the frontend half that updates these two tests (forge-et-fe#284, a draft) is not. So POST /api/v1/tags with entityType: account answers 201 where the F11-f test expects 400, and GET /api/v1/tags?entityType=lead&active=true answers 200 where the F11-b test expects 400.
Fix: forge-et-fe#284 (F1-e Wave 1) already changes both tests to F1-e’s contract; when it merges, it also removes the markKnownFailure('BUG-204', …) lines that F20-d’s frontend wave PR adds to both tests, if F20-d merges first.
Regression test: the two tests themselves, unmarked once forge-et-fe#284 merges.
Resolved: 2026-10-17. Fixed in forge-et-fe#284 (F1-e Wave 1): both tests state F1-e’s contract and carry no markKnownFailure('BUG-204', …); the seam project passed on the PR’s branch against the backend on main.
BUG-203 — On a lead, the Add tag list runs off the bottom of a 1280 × 720 window now that the frame has no top bar
Affects: F11-f (forge-et-fe src/features/leads/LeadTagsSection.tsx, ADD_TAG_OPTIONS_MAX and its shortList Combobox) · Found in: the browser run of forge-et-fe#262 (shadcn/ui backfill, which removes the app frame’s top bar), Windows 11, Chromium, 2026-10-16 · Severity: low — the list still works; its last row and the line under it sit about 10px below the window · Phase: 1
Repro: on forge-et-fe feature/shadcn-ui-review (551cebd7), in a 1280 × 720 window, open a lead whose tag read holds more than ten tags and open Add tag. Or run e2e/browser/leads/lead-detail-a11y.spec.ts › “axe: a live lead with the tag combobox open” (it pins MANY_PINNED_TAGS).
Expected: the ten options and the “Type to find main still has the top bar, so it does not show there.
Fix: decided by Nanndish, 2026-10-16: fewer options when the window is short. Add tag lists ten only in a window at least bp/tablet wide and 860px tall, the height at which ten rows and the line under them fit on either side of the field when the field sits mid-window; any other window lists seven, the count the 320px reflow already used. Rejected: sizing the count from the room measured when the list opens (layout code in the section for an exact fit), and tighter option rows (they would change every short-list Combobox).
Regression test: the existing lead-detail-a11y.spec.ts check above, which fails on forge-et-fe#262 before the fix; plus a 1280 × 860 check with the field at the window’s middle.
Resolved: 2026-10-16 — fixed in forge-et-fe#262 (22e30406). LeadTagsSection.tsx gains ADD_TAG_ROOMY_WINDOW (mq.tablet and min-height: 860px) and lists ADD_TAG_OPTIONS_MAX only when it matches, otherwise ADD_TAG_OPTIONS_MAX_SMALL (seven, renamed from ADD_TAG_OPTIONS_MAX_REFLOW because it now covers short windows too). expectShortTagList in lead-detail-support.ts expects seven below 768px wide or 860px tall, so the size matrix’s 1024 × 768 and 1280 × 800 now check seven. lead-detail-a11y.spec.ts gains “the tag list fits the shortest window that lists ten, with the field mid-window” at 1280 × 860, which first asserts the field’s centre is within 2px of the window’s middle. The planned 1280 × 720 size-matrix case was not added: the chromium project already runs the axe check above at 1280 × 720, and a new size project would run every page state at one more size. Proof, in the fix worktree: pnpm typecheck exit 0; pnpm check exit 0; Vitest on LeadTagsSection.test.tsx and shadcnOnly.test.ts, 57 passed; lead-detail-a11y.spec.ts --grep tag in chromium, 12 passed (both tag checks included); --grep leads-lead-tag-picker-many, 16 passed (all seven sizes, Chromium and WebKit). Not yet run on the new head: the whole browser gate and the test record, which run when #262 is marked ready. The fix was written before the claim PR (forge-et-harness#730) merged, which tracker.md § Bug Tracking step 3 forbids; the claim merged before the fix reached a record.
Lesson: L-195 (new — a list kept short so it never scrolls is sized by the height it needs, not by the window’s width); BUG-123 is its first instance.
BUG-199 — An attachment upload writes its object to S3 inside the database transaction, holding a pooled connection for the whole upload
Affects: FS-11 (forge-et-be src/main/java/com/empireteas/attachment/service/AttachmentService.java:247, objectStore.put inside the @Transactional upload and uploadOwnRoute); F20-h’s email uploads are its busiest caller · Found in: the review of forge-et-be#262 (F20-h Wave 1), backend-reviewer, 2026-10-15 · Severity: low — no wrong result; a slow or large upload holds a database connection for its duration · Phase: 1
Repro: read AttachmentService.upload: the object store put of the whole file (up to 25 MB for an email upload) runs inside the method’s transaction, after the row is written.
Expected: no network call of unbounded length inside a database transaction (.forge/design/coding-standards.md §2 Transactions).
Actual: the transaction, and the pooled connection behind it, stay open for the whole upload. Several large uploads at once can exhaust the pool for every other request.
Fix: not decided. Put the object first, outside any transaction, then insert the row in a short transaction, keeping the existing compensating deleteAllVersions for a row that fails to insert. F20-h did not change the placement, so its review deferred it here.
Regression test: a spec that asserts no transaction is active when ObjectStore.put is called.
Resolved: 2026-10-20 — fixed in forge-et-be#285 (345b7546, c46a05ee, 9e048ccb, a579299a).
- Neither upload route is
@Transactionalnow.AttachmentServicestill checks visibility, type and size first, then puts the object with no transaction open. - One
TransactionTemplatetransaction follows the put. The constructor takes aPlatformTransactionManager, asPriceExpirySweepServicedoes. The transaction inserts the row, recordsattachment.uploaded(andattachment.scan_skippedwhile scanning is off) in that same transaction (AD #36), and requests the scan. The counts still wait for the commit. - The compensating
deleteAllVersionsruns only on a known roll-back. The callback’s first statement registers anafterCompletionsynchronization that records the outcome. The object is deleted when the transaction never began or rolled back, including abeforeCommitrefusal such asUnauditedWriteCheck’s. It is kept, with one ERROR line naming the owner and the exception’s class (never the key), when the commit’s outcome is unknown or an after-commit action failed after the row committed. The first commit deleted it on any exception, and the PR review caught that. - An upload called inside a caller’s transaction scope is refused before anything is read. “Scope” means
TransactionSynchronizationManager.isSynchronizationActive(), so aSUPPORTSscope counts too. No caller does this:AttachmentControllerandEmailUploadService.uploadare not transactional. - Scan verdicts are unaffected.
ScanResultHandlerand the GuardDuty consumer see the same order as before (put, then the row’s commit), andLocalScanDispatcherstill defers to after commit.
Regression test:
UploadTransactionSpec(T1) wraps the real service in Spring’sTransactionInterceptorreading@Transactional, with a new resourceless testRecordingTransactionManager. It records whether a transaction was active at each collaborator call. Its rows cover:- both routes, scanning on and off;
- a failed row insert and a failed audit insert;
- a commit that rolls back at the database, and one refused in
beforeCommit; - an unknown outcome, and an after-commit failure;
- a caller’s transaction, and a synchronization-only scope.
- Against the pre-fix
AttachmentServiceit failed 8 of 8 on its assertions. For example, it recorded[put in a transaction, row in a transaction, delete in a transaction]and0 * visibility.canManage(*_) (1 invocation). With the fix, all 13 pass. - The other specs and ITs pass:
UploadValidationSpec28/28 andScanOffUploadSpec6/6;- on the real stack:
AttachmentAuditIT10 (AC-7b: a failed audit insert leaves no row and no object),AttachmentPipelineIT21,AttachmentVisibilityIT9,EmailUploadControllerIT20,EmailUploadTypesIT22,ScanOffUploadIT3,EmailUploadScanOffIT1,DefaultScanPathIT1,EmailAttachmentAuditIT4; - the ArchUnit suite.
The full build record is on the PR. Developing it hit BUG-237, a bogus NullAway error on an incremental compile.
Lesson: L-213 (prove where a transaction starts through Spring’s own interceptor; undo an external effect only after a known roll-back). backend-reviewer’s transactions checklist points at it.
BUG-198 — Flaky tests: a size-matrix state’s route has not rendered when the 5-second ready wait ends, once per full browser gate (Windows)
Affects: F26-f (forge-et-fe e2e/pageStates.ts:1814, weeklyPrices('pricing-weekly-sales-empty'); its arrangement e2e/browser/size-matrix/arrange.ts:4176) and F5-f (leads-lead-share-dialog) · Found in: F20-h Wave 1’s Stage 12 pnpm e2e:gate on forge-et-fe feature/F20-h-wave-1 (3af44ebf) against the forge-et-be wave backend (7a18eac6) on an isolated stack, Windows, 2026-10-15 · Severity: low — one test fails now and then, which fails the gate; the product is unaffected · Phase: 1
Repro: run pnpm e2e:gate in forge-et-fe on Windows.
Expected: the Weekly prices tab with no sales renders its Pricing heading within the size matrix’s 5-second ready wait at every size.
Actual: at size-320x640-chromium, waitUntilReady (size-matrix.spec.ts:60) found no level-1 Pricing heading within 5 seconds, 45 minutes into the run, and no page snapshot was captured (the route had not rendered). The other 3,783 tests passed (6 expected skips). Run alone straight after, on the same head and backend, the state passed 3 of 3 (--project=size-320x640-chromium --grep pricing-weekly-sales-empty --repeat-each=3). F20-h changes nothing under pricing. A second gate on the same head and a freshly reset backend failed one other state the same way: leads-lead-share-dialog at size-1024x768-chromium found no leads-view within 5 seconds (size-matrix.spec.ts:62), 3,783 others passing; alone straight after it passed 3 of 3. So it is not one state: late in a full gate, a route now and then has not rendered by the end of the ready wait.
Fix: not decided. Find what holds the route back late in a long gate (a lazy chunk’s first transform, the pinned reads’ order, or the dev server under load — compare BUG-194 and BUG-196), then address it. A longer ready wait is not the fix (.forge/test-strategy.md § Flaky Tests).
Regression test: the state passes inside repeated full gates on a loaded Windows machine.
Resolved: 2026-10-30 — fixed by BUG-218’s fix, forge-et-fe#365. What held the route back was the dev server’s module graph: every test’s fresh browser context fetched and compiled some 270 source modules, so under a full gate’s load the route sometimes had not rendered within 5 seconds. The suite now runs against a development build served by vite preview (a cold page load 1.3 s → 0.8 s on a quiet machine), and Windows starts at most 6 browsers. In BUG-218’s three full gates in a row (--full on ece00c5a, a full ./gradlew build alongside, CPU 0.76–0.78 on average), no size-matrix state missed its ready wait. The ready wait stays 5 seconds.
Lesson: L-248.
BUG-139 — A tab left open across a dev deploy fails to open a route it has not loaded yet, and the error page’s way back does not reload
Affects: F38-d (forge-et-fe .github/workflows/build.yaml sync::prune; src/app/router.tsx:71, the / route’s errorElement; src/app/RouteErrorElement.tsx) · Found in: frontend-reviewer’s review of forge-et-fe#259 (BUG-136), 2026-10-14 · Severity: low — the route error page shows, and a full reload recovers · Phase: 1
Repro: open https://dev.crm.empire.lk and leave the tab open. Merge a frontend change to main, so the dev deploy runs. In the old tab, open a screen whose route chunk the tab has not loaded yet.
Expected: the screen opens, or the app reloads itself onto the new build.
Actual: each deploy’s last step deletes the previous build’s hashed chunks, so the old tab gets index.html back for the chunk (the distribution answers a missing object with it) and the dynamic import fails. RouteErrorElement catches it, and its own comment names this case, but the / route’s error element uses the default recovery="home" (“Back to the dashboard”), which navigates inside the app without reloading; opening the failed route again fails again until the person reloads the page. Since BUG-136, a reload always fetches the new index.html, so a reload is enough.
Fix: not decided. Either reload on a failed dynamic import (a vite:preloadError listener, or RouteErrorElement choosing reload when the error is a failed chunk load), or keep the previous build’s assets for one deploy (leave assets/ out of the prune and delete old assets one deploy later). The first needs no change to the AWS side.
Regression test: a component test that a failed chunk load on a route offers a reload, or a browser test that fails a route chunk’s request and sees the app reload.
Resolved: 2026-10-15 — fixed in forge-et-fe#280 by the first option above. RouteErrorElement offers Reload the page, in the same in-app layout, when the error is a failed chunk load; every other error keeps its recovery. src/app/chunkLoadError.ts recognises each engine’s message:
- Chromium: “Failed to fetch dynamically imported module”.
- WebKit: “Importing a module script failed”, and, for a chunk answered with
index.htmlas the distribution does, “‘text/html’ is not a valid JavaScript MIME type for module script”. - Firefox: “error loading dynamically imported module”.
- Vite: “Unable to preload CSS”, for a stylesheet.
Nothing reloads by itself, so a missing chunk cannot cause a reload loop.
Evidence:
- Component tests in
src/app/routerErrorElements.test.tsx: one per engine’s message, the/leadssection on the shipped route table, and aReact.lazyscreen. All fail on the pre-fix component. - A throwaway Playwright check against a local backend: it answered the Leads chunk with
index.html(200,text/html) and passed in Chromium and WebKit. The first WebKit run of that check is what found the MIME message.
The missing <main> on this error page is older than this bug and is BUG-200.
Lesson: L-194 (a chunk-load failure is tested the way the server really answers a missing file, in every engine).
BUG-138 — Flaky test: the size matrix’s user-editor-stale state fails axe color-contrast on Save in WebKit at 1024×768, now and then (macOS)
Affects: F5-c, through F40’s shared page check (forge-et-fe e2e/browser/size-matrix/size-matrix.spec.ts:216, state user-editor-stale in e2e/browser/size-matrix/arrange.ts; e2e/support/pageCheck.ts settleAnimations) · Found in: pnpm e2e:gate on forge-et-fe#257 (95ba635), macOS, ET_E2E_WORKERS=4, an isolated backend stack, 2026-10-14 · Severity: low — one size-matrix state fails now and then, which fails the whole gate; the product is unaffected · Phase: 1
Repro: on macOS, with nothing else running but the backend: ET_BACKEND_PORT=<port> pnpm exec playwright test e2e/browser/size-matrix/size-matrix.spec.ts --project size-1024x768-webkit -g user-editor --repeat-each=10.
- On the BUG-154 branch: it failed once in the gate’s 3,700 tests and once in 160 runs of that command; 340 more runs and 30 runs of the state alone passed.
- On forge-et-fe
main(0ea8055): 320 runs of that command and 30 of the state alone passed.
Two failures in about 520 runs cannot be told apart from none in 350, so the failure is not attributed to #257, whose change is a column’s margins.
Expected: the state passes the shared page check every time.
Actual: axe reports color-contrast (serious) on Save (users-editor-save, enabled). Every other check is clean, and the same state passes at every Chromium size.
Likely cause, not proved:
- The shadcn
Buttonhastransition-all. - The arrange step saves, gets the pinned
409 STALE_VERSION, and waits for Save’saria-busyto readfalse. settleAnimationsthen waits only for animations already running. If WebKit has not yet recalculated the button’s style, its colour transition has not started, so it is not ingetAnimations(), and axe reads a colour part-way.- BUG-033 had the same kind of cause. Fix: not decided. Two options:
- in
pageCheck.ts(shared with forge-et-pwa, L-190), flush style and wait one animation frame beforegetAnimations(), so a transition that was just triggered is listed; - or have the arrange step wait for Save’s transitions to finish.
Trace a failing run first.
Regression test: user-editor-stale passes 100 of 100 in WebKit at 1024×768 under the repro command.
Cause (found 2026-10-14): a probe that logged every animation the page started (20 runs, WebKit 1024×768) showed Save starting a new background-color transition in every run, about 100 ms after the previous one and after settleAnimations had made its one listing of the running animations — most likely the stale notice moves Save from under the pointer and WebKit updates the hover state late. When axe scanned before that transition ended, it read Save part-way. So the transition was listed late, not created late: neither option above alone would have caught it.
Resolved: forge-et-fe#260, with its byte copy forge-et-pwa#76 (2026-10-14). settleAnimations in the shared e2e/support/pageCheck.ts now lists the page again two frames after each wait, until a listing finds nothing running, and stops after 5 seconds whatever still runs; it is exported, and e2e/browser/users/support.ts’s copy of the old single-listing wait now re-exports it. A new self-test in e2e/browser/support/page-check.spec.ts starts a second transition while the check waits for a first: it failed 5 of 5 on the old code and passes on the new. The regression test passed 100 of 100, and the probe on the fix showed the wait returning 20–35 ms after Save’s late transition ended in 20 of 20 runs. The arrange step was left as it is, so the state still exercises the wait.
Lesson: L-101 — a test checked the page while it was still changing; this time the shared check’s own wait missed a transition that started during it. Added to L-101’s sources.
BUG-137 — The offboarding sub-view, the account page and the message view are not on the detail-page rule: full width and left-aligned
First filed on forge-et-harness#673 as BUG-136; renumbered because main took BUG-136 first (.claude/rules/id-numbering.md).
Affects: F5-h (forge-et-fe src/features/users/Offboard.module.css:10), F1-d (src/features/accounts/AccountDetailPage.module.css:7), F20-c (src/features/email/MessageView.module.css:13) · Found in: frontend-reviewer’s review of forge-et-fe#257 (BUG-154), 2026-10-14 · Severity: low · Phase: 1
Repro: at 1440px, open a user’s editor, then Offboard (/users/:id/offboard). Then open an account (/accounts/:id) and a message (/email/messages/:id).
Expected: every drill-in page follows one placement rule.
Actual: since BUG-154, the user editor is a centred, capped column (design-system §6 Detail page, Rev 56). Its own offboarding sub-view is full width and left-aligned, so the page jumps sideways between the two. The account page and the message view are also full width.
Fix: decide for each page.
- Offboarding. It holds three-column record tables, which each scroll in their own card, so capping it changes how those tables lay out.
- The account page and the message view. These are multi-panel detail pages, not one column of forms.
Then widen Rev 56’s rule to the pages that take it, and list the rest as full width with a reason.
Regression test: for each page that takes the rule, a row in e2e/browser/shell/detail-page-column.spec.ts through centredColumn.ts.
Update (2026-10-28): decided and built. All three pages take the centred column, capped at two list-pane widths (Nanndish’s decision, design-system Rev 74), in forge-et-fe#355 (AccountDetailPage.module.css .page, MessageView.module.css .screen, Offboard.module.css .screen), and their centring was measured in the browser at 1280px. The bug stays open until the regression rows above exist: detail-page-column.spec.ts checks the user editor, the account edit page and the role editor, not these three.
Resolved: 2026-11-02 — closed as wontfix, the lead’s decision (Tharinda). The centred column these regression rows would have checked was reversed for record pages by design-system Rev 77 (forge-et-fe#364, 2026-10-30), which the lead agreed on 2026-11-02: the account page is full width on main. The message view and the offboarding page are still centred, but the draft forge-et-fe#371 removes e2e/browser/shell/detail-page-column.spec.ts and centredColumn.ts with the rest of the centred-column checks, so no rows are added for them. Rows for all three pages were written on a local branch on 2026-10-29 and are dropped unpushed.
Lesson: none — one-off: the layout rule the check was for was reversed before the check was added.
BUG-136 — The CRM frontend’s dev deploy uploads files with no cache headers, so a browser can show a blank page after a deploy
Affects: F38-d (forge-et-fe .github/workflows/build.yaml, the sync::build step, added by the DevOps team on 2026-10-13 in 9c78916…bf4adc8) · Found in: the after-the-fact check of the DevOps team’s pushes, 2026-10-14 (.forge/design/environments-and-identity.md §4 row 15 (e)) · Severity: medium — dev only today, but every deploy can leave a person on a blank page until a hard refresh
Repro: curl -sI https://dev.crm.empire.lk/ and curl -sI on any file under /assets/ return no Cache-Control header (checked 2026-10-14). A request for a hashed script that no longer exists, /assets/<old-hash>.js, returns 200 with content-type: text/html, because the distribution answers a missing object with /index.html.
Expected: the staged sync in architecture.md § CI/CD Pipelines: hashed assets uploaded first with Cache-Control: max-age=31536000, immutable; other files (favicons, manifest.webmanifest and the like) with max-age=0, must-revalidate; index.html last with max-age=0, must-revalidate; --delete only on the last step.
Actual: one aws s3 sync dist/ s3://et-dev-spa --delete with no headers. A browser may reuse an index.html it fetched before a deploy (with no Cache-Control, browsers apply their own freshness rule). That page asks for the old hashed scripts, which --delete removed, and gets HTML back, so the app does not start.
Fix: replace the one sync with the staged sync above, in a PR to forge-et-fe with the lead’s approval as a CI change (the DevOps team may push it directly instead — .claude/rules/git-conventions.md § Branch Protection). The same fix applies to forge-et-pwa’s deploy when it is written.
Regression test: after a deploy, curl -sI shows immutable on a hashed asset and max-age=0, must-revalidate on / and on /favicon.ico.
Resolved: forge-et-fe#259 (2026-10-14). build.yaml now syncs dist/assets/ with public, max-age=31536000, immutable, then the fixed-name files from public/ and index.html with max-age=0, must-revalidate, then a last --delete --size-only step that only deletes (a filtered sync --delete deletes only what the filter includes, so the prune is a step of its own; --size-only re-uploads nothing, so no object loses its header). A workflow-level concurrency group, never cancelled part-way, keeps two deploys from pruning each other’s assets. Dry runs of each stage against the bucket before merge selected exactly the 155 hashed assets, the three icons and index.html. Regression test, after the merge’s deploy (run 38514857599, every step green), 2026-10-14: curl -sI on https://dev.crm.empire.lk/ and /favicon.ico shows cache-control: max-age=0, must-revalidate, and on /assets/index-f2ia76fV.js cache-control: public, max-age=31536000, immutable. architecture.md § CI/CD Pipelines corrected in the same change: it said every stage runs with --delete, which would remove the old scripts before the new index.html landed. Found in review and filed: BUG-139 (a tab left open across a deploy). A browser that cached the old header-less index.html may still show one blank page after the first deploy with this fix.
Lesson: none — one-off: a deploy file pushed outside the PR flow left out a step the design records; the after-the-fact check of those pushes (.claude/rules/git-conventions.md § Branch Protection) is the control, and it caught this one.
BUG-132 — Eight more pricing write paths say “Nothing was changed” when a write’s outcome is unknown
Filed as BUG-196 and renumbered to the lowest free number before its record merged (.claude/rules/id-numbering.md rule 2). forge-et-fe commits and code comments merged with forge-et-fe#250 say BUG-196.
Affects: F26-b, F26-c, F26-e and F26-f screens in forge-et-fe src/features/pricing/:
ArchiveStandardDialogPublishRoundDialogRemoveDraftDialogWithdrawDialogSetCostDialogStandardFormDialogNewSaleDialog’s generic failure- the restore action in
StandardsBook
Found in: BUG-190’s fix, which its Lesson field asked to widen to every write dialog, 2026-10-13 · Severity: low — misleading text; no wrong data is written · Phase: 1
Repro: in any of those paths, make the write’s outcome unknown: a network failure after the request is sent, a 5xx with no problem envelope, or a 2xx body the client cannot parse.
Expected: the dialog says it could not confirm the change and offers a re-read (L-191, coding-standards.md §3.2), as CorrectPriceDialog, NotInRoundDialog and, since BUG-190, DraftPriceDialog do.
Actual: each shows its generic “nothing was changed / saved” failure, although the server may have committed.
Fix: branch on isWriteOutcomeUnknown in each, before the generic failure, with “couldn’t confirm” text and a re-read of what the write changes. Known refusals stay unchanged.
Regression test: per dialog, component tests for the three unknown cases, and one test that a 5xx with an envelope keeps the old text.
Resolved: 2026-10-13 — fixed in forge-et-fe#250 (3db8533…7a6d9e0; review fixes 9a7cf8b, 14ee0e0, ba73699, 41cef7a, 1a5a158, 57c6bf5), merged as 6908626a.
- The fix: all eight paths branch on
isWriteOutcomeUnknownand re-read what a success re-reads. Where a repeat is not refused (withdraw whole round, add standard, set cost), the text says to check first. - Tests: 40 new cases failed before the fix and pass with it.
- Browser gate: passed on
b39c599, 3,446 tests.
Lesson: L-191 (a write has three outcomes; never say nothing changed when the outcome is unknown). This is its widening: the rule now also stands in coding-standards.md §3.2 and frontend-reviewer (harness #653), so the next write dialog is checked against it.
BUG-014 — A permission-guarded button that switches between blocked and allowed while focused drops focus to the page
Filed as BUG-195 and renumbered to the lowest free number before its record merged (.claude/rules/id-numbering.md rule 2). forge-et-fe commits and code comments merged with forge-et-fe#250 say BUG-195.
Affects: FS-6 / F26-b (forge-et-fe src/ui/permissions/PermissionGuard.tsx). Seen on the Standards book’s Add standard (StandardsToolbar.tsx, addBlockedReason) and possible on any guarded control whose blocked reason can change while it holds focus. · Found in: BUG-189’s investigation, 2026-10-13 · Severity: low — a keyboard user loses their place; nothing is written wrongly · Phase: 1
Repro:
- On
/pricing, pick a client in the Standards book’s client filter. That starts a lookup of the client by id. - While the lookup runs, Add standard is blocked (“Checking the client in this link”), and
PermissionGuarddraws it as<span><Tooltip>…button…</Tooltip></span>. - Tab to it. When the lookup answers, the guard draws the bare button. React replaces the element, and focus falls to
<body>.
Expected: focus stays on Add standard as it changes from blocked to allowed, and the other way round. Actual: focus is lost to the page. A following Enter does nothing. Fix: draw the same element tree in both states, so React keeps the focused button and only its blocked attributes and tooltip change. Add a component test where a focused guarded button switches state and keeps focus, with a press that focused nothing as one row (L-133). Regression test: that component test, and BUG-189’s walk. Side effects of the fix (forge-et-fe#250):
- Import page, failed Create: after a failed Create that re-checks, focus now stays on Create. The F26-c WI-1.2 plan (Rev 1) moved it to the report heading, but only because Create used to be replaced. Now it survives, so under L-133’s rule focus does not move. The repair to the Check heading still applies when Check again leaves the page.
- Sidebar nav rows: they now pass their ref on, so a blocked nav entry’s
aria-disabledand reason reach the link, which they did not before. - Notes composer: the lead view’s notes composer moves to a single guard and tooltip, because the guard now wraps allowed children too.
- Shared helper: the four copies of the contextual-block resolver became one
ContextualBlockinsrc/ui/permissions.
Resolved: 2026-10-13 — fixed in forge-et-fe#250, merged as 6908626a.
- Commits:
f42e7f2; follow-upsf7c7ce5(notes composer),a81d989(matrix),877da9f(ContextualBlock),b39c599(the import e2e tests). - Tests: five cases failed before the fix and pass with it, including the Standards book’s Add standard keeping focus when the client lookup answers.
- Browser gate: passed on
b39c599.
Lesson: none — one-off. A wrapper drawn only in one state replaces the element in the other. The general rule is L-133, focus survives only on an element that stays; this adds it to a shared component.
BUG-202 — Escape typed in the compose dialog’s body does not close the dialog or ask to discard
Affects: F20-c (forge-et-fe src/features/email/compose/RichTextBody.tsx / ComposeEmailDialog.tsx, the TipTap body inside the native <dialog>) · Found in: F20-d-WI-1.4’s browser suite, 2026-10-16; reproduced on main 0e07dd7a · Severity: medium — a keyboard user writing the body cannot leave the dialog with Escape (they must Tab out first); F20-c FR-C6 says Escape closes an unchanged dialog and asks on a changed one · Phase: 1
Repro: open Compose email by keyboard, Tab into the body, press Escape. A throwaway probe (orchestrator, F20-d overnight run) logged keydown Escape on email-compose-body prevented=true, and the dialog stayed open with no cancel event; with focus in To the same Escape fired cancel and closed the dialog.
Expected: Escape in the body behaves as anywhere else in the dialog: an unchanged email closes, a changed one shows the discard question.
Actual: the body’s keydown is default-prevented before it reaches the dialog, so the native cancel never fires and nothing happens. The preventing handler is not in RichTextBody’s own handleKeyDown (it handles only Ctrl/Cmd+K); it is somewhere in the TipTap / ProseMirror stack, not yet identified.
Fix: not decided. Find the handler (break on preventDefault for Escape in the body), then let Escape through to the dialog, for example by returning it from handleKeyDown before the editor’s own keymaps or re-dispatching a close request. F20-c AC-45’s keyboard test covers Escape only with focus in To; add the body case.
Regression test: e2e/browser/email/templates-a11y.spec.ts’s BUG-202 test (marked known failure by F20-d) and an F20-c AC-45 case with focus in the body, both unmarked once fixed.
Resolved: 2026-10-22 — fixed in forge-et-fe#330 (77a04a11; review fix a9e86330). The handler was ProseMirror’s own: prosemirror-view’s captureKeyDown returns true for Escape (“Enter, Esc”) and the view then calls preventDefault(), so the native <dialog> never fired cancel. RichTextBody.tsx adds a handleDOMEvents.keydown that returns true for Escape, which skips ProseMirror’s own handling and leaves the browser’s default; IME composition is unaffected (ProseMirror returns early during it) and the link row’s own Escape is in a separate input. Regression tests: RichTextBody.test.tsx’s Escape case (not default-prevented, text unchanged) failed before the fix and passes after; templates-a11y.spec.ts’s BUG-202 test lost its known-failure mark, and email-a11y.spec.ts gained the AC-45 body case (an unchanged email closes; a changed one asks to discard, the typed body kept). Both browser tests passed in a targeted Chromium run and in the full pnpm e2e:gate on 57b7eedf (5461 passed, 5 skipped, 0 known failures).
Lesson: none — one-off: an editor library that default-prevents Escape inside a native <dialog>. No other bug or lesson has that cause; the nearest, BUG-251 (filed from this PR’s review), is Radix taking focus back, not a prevented key.
BUG-201 — The shadcn Button with asChild gives Radix’s Slot two children, which Slot refuses
Affects: AD #39’s shadcn layer (forge-et-fe src/ui/shadcn/button.tsx, the render at about l.83) · Found in: F20-d-WI-1.2, adding the shadcn alert-dialog, 2026-10-15 · Severity: low — nothing on main renders <Button asChild>, so no screen is affected today · Phase: 1
Repro: render <Button asChild><a href="/x">Go</a></Button>, or a shadcn component that wraps its action in Button asChild, such as the generated alert-dialog.
Expected: the Slot merges the button’s props onto the one child element.
Actual: the component passes Slot.Root two children — the busy spinner’s slot (null when asChild is set) and children — and Slot requires exactly one, so it throws or drops the merge.
Fix: when asChild is set, pass children alone (no spinner slot), and add a test that renders Button asChild around a link. F20-d-WI-1.2 works around it in its own src/ui/shadcn/alert-dialog.tsx; remove that workaround in the same fix.
Regression test: a button.test.tsx case rendering asChild with a link child: one <a> with the button’s class and data-slot, no error.
Resolved: 2026-10-20 — fixed in forge-et-fe#314 (1a474c2f; review fixes c158117a, adf37715). With asChild the Button passes children alone to Slot, with no spinner. The alert dialog’s action and cancel are back in shadcn’s form, Button asChild around Radix’s own buttons, so F20-d’s workaround is gone; disabled is left out of their props. PriceEmailDialog.tsx’s comment that asChild could not wrap a link was corrected. Regression test: the new src/ui/shadcn/button.test.tsx (one <a> with the button’s class, data-slot and data-variant, no error; no spinner beside the child while busy) failed before the fix (Slot failed to slot onto its children) and passes after; the new src/ui/shadcn/alert-dialog.test.tsx covers the dialog’s answers.
Lesson: L-212 — the second path of the adapted shadcn Button that broke unseen until a feature used it (after BUG-192’s busy); written in the fixing harness PR, with a row in frontend-reviewer’s shadcn dimension.
BUG-200 — The route error page has no main landmark and no app shell when the / route’s error element replaces AppLayout
Affects: FS-3 (forge-et-fe src/app/RouteErrorElement.tsx, the default branch; src/app/router.tsx:71, the / route’s errorElement) · Found in: frontend-reviewer’s review of forge-et-fe#280 (BUG-139), 2026-10-15 · Severity: low — a screen-reader user lands on a page with no main region and no navigation; the page still works · Phase: 1
Repro: make any screen under / fail. For example, block its chunk in a browser, or throw in its component. Then read the page’s landmarks.
Expected: the error page has one <main>, as every page does (design-system §9). The reload branch already has one.
Actual: the / route’s error element renders in place of AppLayout, and its default branch draws only the title and the failure card. So the page has no <main>, no sidebar and no navigation. Since BUG-139, that includes the failed-chunk page a stale tab sees after every deploy. The browser gate does not catch it: axe’s landmark-one-main and region rules are best-practice rules, and e2e/support/pageCheck.ts does not gate them.
Fix: not decided. Either give the default branch a <main> too, as the reload branch has (it also replaces the whole page), or render the error inside AppLayout so the shell stays. The second needs the layout to catch errors below it.
Regression test: a component test, on the shipped route table, that a failing section’s error page has exactly one main.
Resolved: 2026-10-20 — fixed in forge-et-fe#314 (a431b850; review fixes c158117a). The lead chose the first option: every layout of RouteErrorElement carries the page’s one <main>, the default and the 404 ones as the reload one already did, because each use replaces a whole page. The error page still has no sidebar or navigation; rendering it inside AppLayout was not taken. router.tsx is unchanged. Regression test: routerErrorElements.test.tsx “on the shipped routes, a failing section’s error page has exactly one main landmark” failed before the fix (Unable to find an accessible element with the role "main"), as did the four engine cases once they asserted one main; all pass after. The size matrix gained the app-section-load-error state (the Accounts chunk aborted), which checks the page and its one <main> at every size in both engines.
Lesson: L-106 — a screen only a failure renders had no page state, so the gate never checked it; BUG-200 added to its sources, and forge-et-fe#314 added the state.
BUG-197 — Flaky test: the size matrix’s pricing-set-cost-read-error state times out opening its row menu in WebKit 1024×768 in a full gate (macOS)
Affects: F40 and F26-g (forge-et-fe e2e/browser/size-matrix/arrange.ts:1575, the pricing-set-cost-read-error state, project size-1024x768-webkit) · Found in: pnpm e2e:gate on forge-et-fe#272 (4f8653b), macOS, 8 cores, default workers, 2026-10-15, the first gate run with the load recorded · Severity: low — the gate fails; the product is unaffected · Phase: 1
Repro: a full pnpm e2e:gate on macOS; not yet reproduced on demand.
Expected: the state’s Set cost menu item opens within the 30-second test timeout.
Actual: after the row-actions trigger was clicked, the gate waited the whole 30 seconds for the Set cost menu item, then timed out. The result was 3,700 passed, 1 failed and 5 skipped, in 36.4 minutes.
- Recorded load: a busy share of 0.33 at the start (within the limit), 1.00 at the peak and 0.92 on average; a load average of 2.22 per core at the start and 5.56 at the peak.
- Busiest processes at the peak:
nodeandchrome-headless-shell, which cannot be told apart from the run’s own, andmediaanalysisd. - Not a stall:
stallswas empty. The test ran its 30 seconds and no longer. - Re-run: the same commit’s next gate passed (3,701 passed).
This is the macOS twin of BUG-129’s pricing sightings on Windows (its F26-g notes): the same wait on a WebKit row action, in a different state. Recorded first under BUG-194 as an “Also seen” note, and moved here because the symptom is BUG-129’s, not BUG-194’s stall.
Fix: waits on BUG-129’s cause (.claude/rules/tracker.md § Bug Tracking, a twin on the other operating system). Then prove the fix, or the overload outcome of .forge/test-strategy.md § Flaky Tests, on macOS.
Regression test: the WebKit 1024 pricing states inside repeated full gates on macOS.
Triage (2026-10-15, .forge/test-strategy.md § Flaky Tests):
- The record: the run started within the limit (busy 0.33), and its peak processes were the run’s own. There was no stall: the test ran its 30 seconds and no longer.
- Forced load: the
size-1024x768-webkitpricing-*states were run twice with 8 busy loops beside them, at load averages up to 16. Both runs passed all 104 states, the slowest in 5 seconds. Then this one state ran with--repeat-each=40, also under 8 busy loops: 40 of 40 passed. - The step itself: the click on Actions for BS-990101 toggles the shared
DropdownMenu(src/ui/components/DropdownMenu/DropdownMenu.tsx), which opens with the popover API. The failing run waited 30 seconds for the menu’s Set cost item, so the menu either never opened or closed before the item could be found. No step that lands late was found to hold back. - Outcome: not reproduced, and the record shows no other work, so this is not the overload case. It stays open as a defect to find. It waits on BUG-129’s cause, which covers the same symptom in pricing states on Windows. A recurrence on macOS is recorded here with the gate’s
loadandstallsand a trace. BUG-129’s cause (recorded 2026-10-22): the full run’s own browsers saturate the CPU, and under that load WebKit stops drawing for seconds, so Playwright’s stability wait and the axe scan run out of the 30 seconds (BUG-129’s “Cause found”). This run’s record is consistent with it: busy 1.00 at the peak and 0.92 on average, the busiest processes the run’s ownnodeandchrome-headless-shell. Step 4 above ruled out overload because no other work was running; the cause is the run’s own work, so it is not ruled out. Not yet proved on macOS: the forced-load runs in step 2 ran only the pricing states, not a full gate. The fix follows BUG-129’s (the number of browsers at once, or one of the options it puts to the lead), then is proved on macOS with repeated full gates. Resolved: closedwontfix2026-10-23 as an overload outcome; Tharinda (the lead) decided. Nothing on macOS can show a fix: the forced-load attempts in step 2 (8 busy loops beside all 104size-1024x768-webkitpricing states, twice, and this state 40 times) never reproduced it, and the one failure came in a full gate whose own browsers held the CPU at busy 1.00 at the peak and 0.92 on average (load average 5.56 per core at the peak; the busiest processes the run’s ownnodeandchrome-headless-shell, and macOS’smediaanalysisd)..forge/test-strategy.md § Flaky Testsstep 4 names overload by other work; this closure applies it to overload by the run’s own work, which is BUG-129’s cause, as the lead’s decision for this bug. The prevention, the number of browsers a full gate starts on each operating system, is BUG-129’s change (Nanndish). A recurrence on macOS after that change is a new bug, recorded with the gate’sload,stallsand a trace. Lesson: L-175 — the full browser gate’s own load fails a different test now and then, here on macOS at the default 4 workers on 8 cores; BUG-197 is added to its sources.
BUG-196 — Flaky test: a seam test’s request to the Vite dev server is refused once in a full gate on a loaded machine (Windows)
Affects: F5-c (forge-et-fe e2e/seam/users-edit.seam.spec.ts:215, “Tea Room PATCH of a test user with a valid body → 403 FORBIDDEN, and the user is unchanged”) · Found in: BUG-186’s pnpm e2e:gate run on forge-et-fe main (a18d03c) against forge-et-be#251, Windows, 2026-10-14, with another session’s backend test suite running on the same machine · Severity: low — one test fails now and then, which fails the gate; the product is unaffected · Phase: 1
Repro: run pnpm e2e:gate in forge-et-fe on Windows while the machine is loaded.
Expected: every request a seam test makes through the dev server’s proxy is answered.
Actual: apiRequestContext.get: connect ETIMEDOUT ::1:5173 / connect ECONNREFUSED 127.0.0.1:5173 on GET /api/v1/users/128 inside expectUnchanged (e2e/seam/support.ts:387), 41 minutes into the run. The dev server logged “Importing a module script failed” in the same minutes. The other 3,699 tests passed. Run alone right after, users-edit.seam.spec.ts passed 20 of 20.
Fix: not decided. Find whether the dev server stops accepting connections under load (its event loop blocked by a transform) or Node resolves localhost to ::1 first while Vite listens on IPv4 only, then address the cause. A longer timeout is not the fix (.forge/test-strategy.md § Flaky Tests).
Regression test: the seam suite passes inside repeated full gates on a loaded Windows machine.
Also seen (2026-10-16, F20-d overnight run, Windows): the same refusal on the browser side — apiRequestContext.post: connect ETIMEDOUT ::1:5473 / ECONNREFUSED 127.0.0.1:5473 at signInAs, three minutes into WI-1.4’s e2e/browser/email + e2e/browser/mailbox run (verify-runtime/wi14-2-2.log) on a dev server started on port 5473 by an override config; the same run on the same commit before and after passed 204 of 204.
Also seen (2026-10-17, F26-d overnight runs, Windows): twice. (1) F26-d-WI-1.3’s baseline seam run on the merge-base (roles.seam.spec.ts:524, connect ETIMEDOUT ::1:5473 on a dev server started on 5473 by an override config); the same test passed in the wave run an hour later. (2) F26-d’s Stage 12 pnpm e2e:gate on forge-et-fe 37f27bc6 (offboarding-reports.seam.spec.ts:593, connect ETIMEDOUT ::1:5173 / ECONNREFUSED 127.0.0.1:5173 at readUser, e2e/seam/support.ts:387), the only failure of 4415; machine load 0.62 peak with searchindexer and msmpeng (Windows Search and Defender) the busiest processes. The test is F5-h’s and touches nothing F26-d changed.
Also seen (2026-10-26, F3-a wave 1 overnight run, Windows): the browser-side refusal again — connect ECONNREFUSED 127.0.0.1:5173 in e2e/browser/attachments/attachments-attach.spec.ts:263, in the first full pnpm e2e:gate on forge-et-fe#296 (6db65d72), machine load 0.80 peak and 0.56 average. The test passed alone right after and in the next full run on the same commit. F3-a changes nothing in attachments.
Investigated on Windows, 2026-10-30 and 31 (Dinil’s overnight run; .forge/plans/bugs/2026-10-30-bug-groups-overnight-log.md): it happens in every full gate since the suite runs against the built app (forge-et-fe#365): 2, 4 and 3 tests in BUG-218’s three proof gates, and 2 in a gate on 12aff918 with a Gradle build beside it, always apiRequestContext (seam specs, and browser specs’ API calls), never a page load.
- The server accepts. A probe in another process opened a new TCP connection to
[::1]:5173every 200 ms through that last gate (about 13,000 connections, the slowest 0.12 s); in the 45 seconds around both failures, all 454 were fast. A first gate that night, onmain, saw no connection slower than 26 ms in 35,000. - Not a stalled worker. A Node client that starts a connect to
localhostwith Playwright’s settings and then blocks its event loop for 6 s still connects, to a server on::1or on127.0.0.1. - Playwright’s settings: the request context connects through Node’s
autoSelectFamilywithhappyEyeballsOptions(playwright-corecoreBundle.js):::1first, then127.0.0.1, at least 5 s per attempt.vite previewlistens on::1only, so the fallback is always refused. - Serving on
127.0.0.1is not the fix (forge-et-fe#369, a draft, not merged): with it, one gate (9eefdee2, Gradle beside it) failedconnect ECONNREFUSED ::1:5173thenconnect ETIMEDOUT 127.0.0.1:5173— the connect to the listening server timed out on the last address too. Its review also found that apnpm devstarted during a run would then take[::1]:5173and receive the suite’s requests. - Still to find: why a connect from the Playwright worker sometimes does not complete in 5 s while other processes connect at once. Next: a probe inside a Playwright worker (a test that opens raw connections in a loop beside the suite), and Windows’ own TCP counters (
netstat -s -p tcp,Get-NetTCPSetting) before and after a gate. Resolved: 2026-11-16 — fixed in forge-et-fe#455 (merged 2026-11-08 asf22ebc1e), the fix that closed BUG-357, which has the same signature (ETIMEDOUT ::1thenECONNREFUSED 127.0.0.1). #455 serves the browser gate’s app on127.0.0.1and calls it there, so no client tries::1first, and BUG-357’s investigation ended the failure in 20 of 20 requests three times. Dinil decided to close this bug on that fix rather than keep it open for two cases #455 does not cover: a connect that gets no answer at all (the draft forge-et-fe#369’s one gate), and runs outside the gate, which still calllocalhost:5173. The question of what remained was inowed.md, raised by forge-et-fe#455’s frontend review. The nightly full run onmainis the regression check; a recurrence of this signature, in or outside the gate, is filed as a new bug that names this one. Lesson: L-158, another instance — on Windowslocalhostis::1first, and a client that callslocalhostadds a connect attempt that a held-up worker can time out; BUG-196 added to its sources.
BUG-195 — Flaky test: ScanResultQueueIT’s AC-3 dead-letter row runs out its 60-second wait in a full build on a loaded machine (macOS)
Affects: F38-c (forge-et-be src/test/groovy/com/empireteas/attachment/service/ScanResultQueueIT.groovy:726, “AC-3: a body that is not JSON is not acknowledged, reaches the dead-letter queue, and claims nothing”) · Found in: python3 scripts/local-build-evidence.py on forge-et-be#255 (76145c3a), macOS, with a backend and an isolated compose stack running beside it and Firefox and the Docker VM busy, 2026-10-14 · Severity: low — one IT fails now and then, which fails the build record; the product is unaffected · Phase: 1
Repro: a full ./gradlew build on macOS while the machine is loaded.
Expected: the poison message reaches the dead-letter queue within the test’s 60-second PollingConditions wait.
Actual: SpockTimeoutError: Condition not satisfied after 539.25 seconds and 6 attempts — six polls in nine minutes, so the test’s own thread barely ran, not a queue that never redrove. Alone right after, ./gradlew test --tests ScanResultQueueIT passed 14 of 14 in 50 seconds; an earlier full build of the same branch passed it too.
Fix: not decided. Find whether LocalStack’s redrive or the JVM stalls under load (trace the queue’s receive counts and timestamps in a failing run), then give the wait a budget that names why, or make the redrive deterministic. A green re-run is not the proof (L-107).
Regression test: the spec passes inside repeated full builds on a loaded macOS machine.
Triage (2026-10-15, .forge/test-strategy.md § Flaky Tests):
- The record. The run was a stall, about 9 times its budget. Each of the 6 polls took about 90 seconds, but
dlqCarriesasks SQS withwaitTimeSeconds(1)and should return within about a second. About 90 seconds fits the AWS SDK’s defaults, three tries each ending at a 30-second socket timeout. So LocalStack, inside the Docker VM, stopped answering. A slow redrive or a starved test thread would not explain it. The machine was running other work beside the build: another backend, an isolated compose stack, Firefox and a busy Docker VM. The run predates the load check, so no load figure was recorded. - Forced load.
./gradlew test --tests 'com.empireteas.attachment.service.ScanResultQueueIT' --rerunwas run three times with 8 busy loops (one per core) beside it, at load averages up to 38 (about 4.7 per core). All three passed 14 of 14, and the AC-3 dead-letter row took 28.0, 27.8 and 29.0 seconds against its 60-second budget. There was no step to hold back: CPU load alone did not stall LocalStack. - Other evidence. Two full
./gradlew buildrecords on 2026-10-15 passed the spec: forge-et-be#261 (peak busy 0.67) and forge-et-be#264 (peak busy 0.77). - Outcome: not reproduced within the limit, and the original record names other work. This is the overload case of step 4.
Resolved: 2026-10-15 — closed as wontfix (machine overload), the lead’s decision (Tharinda). The evidence is the triage above: forced load did not reproduce it, the original record names other work, and two full builds passed it today. No change was made. The load check (forge-et-be#261, limit 0.65 since forge-et-be#264) now refuses a build on a machine that busy. A recurrence on a machine within the limit is filed as a new bug that links here, with the build record’s Machine load: line.
Lesson: L-162 (a failure that appears only under load is reproduced before it is called flaky). This is an instance where forced load did not reproduce it, so nothing was changed, as for BUG-149.
BUG-194 — Flaky tests: the size matrix’s WebKit 768×1024 lead states stall for minutes in a full gate on a loaded machine (macOS)
Affects: F40 and F11-f (forge-et-fe e2e/browser/size-matrix/size-matrix.spec.ts:216, the leads-* states, project size-768x1024-webkit) · Found in: pnpm e2e:gate on forge-et-fe#261 (8f8ae52), macOS, ET_E2E_WORKERS=4, against forge-et-be#255 on an isolated stack, with Firefox and the Docker VM busy (load average 20–30 on 8 cores), 2026-10-14 · Severity: low — the gate fails; the product is unaffected · Phase: 1
Repro: a full pnpm e2e:gate on macOS while the machine is loaded.
Expected: each state finishes within its 30-second test timeout.
Actual: 12 leads-* states at 768×1024 in WebKit ran 1.8 to 9 minutes each and failed (leads-lead, leads-lead-loading, leads-lead-error, leads-options-failed, leads-edit-stale, leads-notice, leads-new-form-errors, leads-edit-form, leads-lead-view-only, leads-lead-archive-forbidden, leads-lead-archive-refused, leads-lead-archive-stale). A run longer than the test timeout means the browser itself stopped responding, not a slow assertion. Run alone right after (-g "leads-" --workers=2), all 83 leads-* states passed in WebKit at 768×1024 in 1.5 minutes. The same symptom as BUG-129 on Windows (which also saw it in Chromium). An earlier gate on the same branch with 6 workers failed 65 tests across leads and mailbox specs the same way and was stopped.
Fix: not decided. As BUG-129: reproduce under forced load with a trace and find what stalls; then cut the work or give the worker a budget that names why. A green re-run is not the proof (L-107).
Regression test: the WebKit 768 project inside the whole suite on a loaded macOS machine, green across repeated runs.
Also seen (forge-et-fe#272’s gate, 2026-10-15): a pricing-set-cost-read-error timeout in WebKit 1024×768. It was not a stall, and its symptom is BUG-129’s, so it is filed as BUG-197.
Triage (2026-10-15, .forge/test-strategy.md § Flaky Tests):
- The record. The run was a stall: the states ran 1.8 to 9 minutes against a 30-second timeout. The machine was running other work beside the gate: Firefox, a busy Docker VM and another isolated stack, at a load average of 20 to 30 on 8 cores. The run predates the load check, so no busy share was recorded.
- Forced load. The
size-768x1024-webkitproject’sleads-*states were run withpnpm exec playwright test --project=size-768x1024-webkit -g 'leads-' --retries=0, twice. Each run had 8 busy loops (one per core) beside it, used the default workers, and ran against a backend on an isolated stack, at a load average of about 20 (2.5 per core). Both runs passed all 83 states, with no stall. The slowest state took 5 seconds. There was no late step to hold back: the symptom is a browser that stopped answering for minutes, not a race. - Other evidence. Two full gates on this Mac on 2026-10-15 passed every WebKit 768 lead state: forge-et-fe#272 (
4f8653b) and forge-et-fe#279 (c6dd163), with 3,701 passed each. - Outcome: not reproduced within the limit, and the original record names other work. This is the overload case of step 4. BUG-129, the Windows twin, is triaged on Windows by its owner.
Resolved: 2026-10-15 — closed as wontfix (machine overload), the lead’s decision (Tharinda). The evidence is the triage above: forced load did not reproduce it, the original record names other work, and two full gates passed the states today. No change was made. The load check (forge-et-fe#272, limit 0.65 since forge-et-fe#279) now refuses a gate on a machine that busy. A recurrence on a machine within the limit is filed as a new bug that links here, with the gate’s recorded load and stalls.
Lesson: L-162 (a failure that appears only under load is reproduced before it is called flaky). This is an instance where forced load did not reproduce it, so nothing was changed, as for BUG-149.
BUG-193 — A price correction can be dated after the next sale’s start, leaving the standard unpriced and the correction ended a day later
Affects: F26-g (forge-et-be src/main/java/com/empireteas/pricing/service/CorrectionRange.java, of, the close_from case; the range the round rows show and PriceCorrectionService enforces) · Found in: forge-et-be#248’s review (F26-g Wave 1, findings N5 and N6), 2026-10-13; recorded as a spec question in .forge/owed.md and decided by Dinil the same day · Severity: low — no wrong price is ever shown; a standard can be left with no price for some days · Phase: 1
Repro:
- A later round published early. Standard X’s price P belongs to sale 40, which is current. Sale 41 starts on the 19th, without X. Sale 42’s round is published early, also without X, which gives P an end date of the 26th. Because P now has an end, the next-sale limit no longer applies, and the Correct dialog accepts any date up to the 25th. Correct X from the 22nd.
- No later sale entered. P is open and no sale starts after today, so no limit applies. Correct X from a date two weeks away, then enter a sale that starts before that date.
Expected: a correction can never start after the next sale’s start, since the sweep ends every price of an older sale on that day. With no later sale entered, a correction can start only today.
Actual: in both cases, the sweep ends P on the next sale’s start, so X has no price until the correction’s date. The correction still belongs to sale 40, so the next sweep ends it after one day.
Fix: in CorrectionRange.of, for close_from, set correctBefore to the earlier of validBefore and the next sale’s start. With no sale after today, set it to tomorrow. The read model and PriceCorrectionService already share this range. Spec F26-g Rev 2 (FR-10) and Wave-1 contract Rev 5 say so. forge-et-fe’s seam-test helper that restates the rule (e2e/seam/price-expiry.seam.spec.ts) follows. No change to the screen, which takes its range from the server.
Regression test: CorrectionRangeSpec rows for both cases, and a PriceCorrectionIT case that refuses a date on or after the next sale’s start for a price that has an end.
Resolved: 2026-10-13.
- Backend: fixed in forge-et-be#249 (
6769adcf, tests6140d929, review fixes926c27b6…2b45ff2b), merged asacb3d56e. - Frontend tests: forge-et-fe#248 (
330693b,dc5a5e4,302ce59), merged as5304e86d. - Test evidence: 8 new or changed cases failed before the fix. With it, the full backend build passed (7,136 tests), and
pnpm e2e:gate --pushpassed against the fixed backend (3,420 passed). - Still open: the same gap reached by creating a sale later remains BUG-186’s.
Lesson: none — one-off: two edge cases of a date range found in review and settled as a spec revision. The limit’s own reason (“a correction must not outlive the next sale’s start”) now applies to every case, not only to an open price.
BUG-192 — A busy shadcn Button grows wider: its spinner is added beside the label
Affects: AD #39’s shadcn Button (forge-et-fe src/ui/shadcn/button.tsx), so F5-c (the user editor’s Save, since forge-et-fe#201) and every screen whose shadcn Button shows busy · Found in: e2e/browser/roles/edit-role.spec.ts AC-21 once BUG-191 moved the role editor’s Save onto the shadcn Button, 2026-10-14 · Severity: low · Phase: 1
Repro: on a role (with BUG-191’s save bar) or a user, change a field and press Save while the request is held; measure the button before and during the wait.
Expected: design-system §5’s busy state keeps the label and the width (F5-d AC-21), as src/ui’s Button does, and as the shadcn Button’s own comment says.
Actual: the spinner is a flex item beside the label, so the button grows by the spinner and the gap: the role editor’s Save role went from 89px to 111px.
Fix: the spinner sits in the button’s inline-start padding, absolutely positioned in a strip as wide as the padding (w-4, w-3, w-2, w-6 by size), at 14px or the strip less 2px a side, as src/ui’s Button places it (.spinnerSlot). An icon-only size has no padding and keeps the spinner in the flow.
Regression test: edit-role.spec.ts AC-21 measures Save role’s width before and while busy (within 1px); it failed on the shadcn Button before the fix and passes after.
Resolved: 2026-10-16 — fixed in forge-et-fe#262 (c4000d95). src/ui/shadcn/button.tsx sets the base class relative and places BusySpinner in an absolutely positioned strip as wide as each size’s inline padding (PADDING_STRIP); the icon sizes keep it in the flow. edit-role.spec.ts AC-21 passes in Chromium on the PR’s branch.
Lesson: none — one-off: the shadcn Button’s comment described the busy state src/ui’s Button has, and the first screen to show it busy (the role editor’s Save role) was the first test to measure it; no other Lesson: line or lesson in .forge/lessons.md names a shadcn part that does less than its comment says.
BUG-191 — On the user and role editors, Save and Discard are easy to miss and cannot be undone
Affects: F5-c (forge-et-fe src/features/users/UserEditorForm.tsx), F5-d (src/features/roles/RoleEditorForm.tsx) · Found in: UI/UX review, 2026-10-14 · Severity: low · Phase: 1
Repro: on a role, change a permission near the bottom of the page, then Save, or Discard. Do the same on a user.
Expected: the page shows that edits are unsaved and how many; Save and Discard stay reachable however far the page scrolls; the result names the record and what changed; a mistaken Discard can be undone.
Actual: the only feedback is a small type/meta line, “Role saved” or “Changes discarded”, at the left of the page under the title while the buttons are at the right; nothing marks the page or the changed cards as unsaved; on a long role the buttons scroll out of view with the header; and Discard cannot be undone.
Decision changed: BUG-159’s fix (a “Changes discarded” line in each editor’s live region, no confirm dialog) and the confirmation lines in the F5-c WI-2.2 and F5-d WI-1.2 / WI-2.2 plans. The F5-c and F5-d specs set no wording (F5-d AC-21 asks only that the editor shows the role as returned and that Discard restores it), so they need no revision. Nanndish decided the design on 2026-10-14; design-system Rev 61 records it as the §6 Save bar.
Fix:
- A save bar at the top of the editor’s column that stays in view while the page scrolls: “
unsaved changes to ” on the warning tint with Discard changes and Save role / Save user; “All changes saved” with Discard changes unavailable when nothing is unsaved (the user editor’s Save waits for a change; the role editor’s stays available, F5-d AC-20). The test ids roles-editor-save,roles-editor-discard,users-editor-saveandusers-editor-discardstay. - Each changed field, permission choice or tick carries an inline-start edge in the warning tone and a text note (“was Manage”, “added”, “removed”).
- Save and Discard confirm in a toast: “Saved
changes to ”; “Discarded changes to ” with Undo, which restores the discarded draft; the Undo toast stays until acted on or closed. The toast is shadcn’s Sonner ( sonner, approved by Nanndish on 2026-10-14) insrc/ui/shadcn/sonner.tsx, so the editors add noshadcnBackfill.jsonentry. Regression test: unit tests inRoleEditorRoute.test.tsxandUserEditorRoute.test.tsx: the bar’s count follows the edits, a changed card is marked, Save and Discard raise their toasts, Undo restores the draft and Discard sends no write; a browser check that the bar stays in view when a role is scrolled to its last card, and the page checks at 320px and 768px. Resolved: 2026-10-16 — fixed in forge-et-fe#262 (80e3b2a1, the cover kept off the header byce65f340).SaveBarandChangeMarkinsrc/ui/components/SaveBar/, the toasts throughsrc/ui/shadcn/sonner.tsxmounted once inmain.tsx;roleChangesFromanduserChangesFromcount and describe the edits. The unit tests named above pass in the full Vitest run; the role and user browser specs pass in Chromium. Lesson: none — one-off: a design decision from the UI/UX review, not a defect in how the editors were built; BUG-159’s line was what its plans asked for.
BUG-190 — The draft price dialog says “Nothing was changed” when it cannot tell whether the save was committed
Affects: F26-f (forge-et-fe src/features/pricing/DraftPriceDialog.tsx, SAVE_FAILED and its catch path) · Found in: the full review of forge-et-fe#245 (F26-g Wave 1), 2026-10-13 · Severity: low — misleading text; a retry is refused safely (STALE_VERSION or the server’s own check), nothing is written twice · Phase: 1
Repro: save a hand-entered draft price while the network drops after the request is sent, or while the server answers 2xx with a body the client’s parser refuses (ApiContractError), or a 5xx with no problem envelope.
Expected: the dialog says it could not confirm the save and offers a re-read, as EditSaleDatesDialog.tsx does through isWriteOutcomeUnknown (F26-g).
Actual: it shows SAVE_FAILED — “The draft could not be saved. Nothing was changed. Try again.” — although the write may have been committed, and the round is not re-read, so the table can show the old row.
Fix: branch on isWriteOutcomeUnknown(caught) before SAVE_FAILED, with the “couldn’t confirm” text and a Reload that re-reads the round, plus component tests for a network failure and a malformed 2xx body. F26-g’s review fixed the same gap in its own CorrectPriceDialog and NotInRoundDialog; this one is F26-f’s code, unchanged by F26-g apart from moving its figure fields out.
Regression test: those component tests.
Resolved: 2026-10-13 — fixed in forge-et-fe#250 (d10654e), merged as 6908626a. The draft dialog shows “couldn’t confirm” with Reload the round. Three new cases failed before the fix and pass with it.
Lesson: L-191. The check this Lesson line asked for found eight more write paths with the same gap. They are BUG-132 (filed as BUG-196), fixed in the same PR.
BUG-189 — Flaky test: the Standards book keyboard walk’s Enter on Add standard opened no dialog once under a full gate run (Windows)
Affects: F26-b (forge-et-fe e2e/browser/pricing/standards-book-a11y.spec.ts:93, “AC-31: every Standards book action works from the keyboard alone, with focus into and out of each dialog”, line 159) · Found in: F26-g Wave 1, Stage 12’s pnpm e2e:gate on feature/F26-g-wave-1 at 5129e0e, 2026-10-13 · Severity: low — a flaky test; the product is not affected as far as is known · Phase: 1
Repro: Windows 11, ET_BACKEND_PORT=8380 pnpm e2e:gate (3,405 tests, 8 workers, 24.7 min) against an isolated compose stack. Failed once in that run: tabTo reached Add standard, which was not aria-disabled, and after Enter getByRole('dialog', { name: 'Add standard' }) was not found within 5 s. The page snapshot shows the Standards book filtered to Al Rifai Trading and F26-g’s pricing warning above the tabs (a year-9000 sale another worker created). The same test on the same stack, alone, --repeat-each=4 --workers=2: 4 passed. It passed in F26-g’s O1 baseline and its four evidence runs.
Expected: the test passes however loaded the machine is.
Actual: one failure in one full run; not reproduced alone. Cause not found: useFocusAfterWrite on the page acts only when armed, which this walk does not do, and F26-g’s change to the shared Modal adds an optional headingRef only.
Fix: not decided — find where the Enter goes (log document.activeElement before the press and whether the button’s handler ran) under load, then wait on the condition the dialog needs rather than on the button’s state.
Regression test: the spec, passing in full gate runs on Windows.
Resolved: 2026-10-13 — fixed in forge-et-fe#250 (ecf80a6, 79732ce), merged as 6908626a.
- Cause: picking a client starts a lookup by id, and Add standard is blocked (“Checking the client in this link”) until it answers. Under load the test tabbed to the blocked button, and the answer then replaced it (BUG-014), so Enter went to
<body>. - Test fix: the walk now waits for
GET …/buying-standards/clients?id=and for the allowed button before tabbing, and asserts focus before Enter. - Product fix: BUG-014.
- Gate: full runs on
b39c599passed.
Lesson: none new — an instance of L-101 (a browser test waits for the condition the action needs) and of L-133 (BUG-014’s product side).
BUG-188 — The browser gate fails on every run: BUG-131’s Windows skip is not on the expected-skip list (Windows)
Filed as BUG-185 on feature/F26-g-plan and renumbered when main (which had its own BUG-182 to BUG-184) was merged in. The old id stays in forge-et-fe commits d25db43 and 0b130da, in the reason text of scripts/browser-gate/expected-skips.json’s BUG-131 entry (.forge/owed.md has the correction), and in forge-et-pwa#74’s title and branch (fix/BUG-185-shared-gate-test).
Affects: F40 / F5-h (forge-et-fe scripts/browser-gate/expected-skips.json; the skip is in e2e/browser/users/offboarding-keyboard-webkit.spec.ts, added by forge-et-fe#224, 3f4a01d) · Found in: F26-g Wave 1, Stage 12’s pnpm e2e:gate on feature/F26-g-wave-1 at 5129e0e, 2026-10-13 · Severity: medium — it fails the browser gate of every forge-et-fe PR run on Windows · Phase: 1
Repro: on Windows, pnpm e2e:gate on any forge-et-fe commit since 3f4a01d → skipped, and not on scripts/browser-gate/expected-skips.json: e2e/browser/users/offboarding-keyboard-webkit.spec.ts › Offboard a leaver by keyboard alone in WebKit (F5-h AC-17, live) › … and Browser gate: failed.
Expected: the skip the lead chose for Windows (BUG-131) is accepted by the gate; on macOS the test still runs and must pass.
Actual: the gate fails any skip not on expected-skips.json, and BUG-131’s fix did not add the entry (on macOS the test does not skip, so a macOS gate run passes).
Fix: add the test to expected-skips.json with a reason naming BUG-131 (an expected skip is allowed, not required, so macOS is unaffected) — forge-et-fe d25db43 on feature/F26-g-wave-1, its own commit (overnight F12 rule: a small, test-backed defect in old code). The gate’s own test src/test/scripts/browserGateRun.test.ts assumed the list was exactly the five laptop sign-in skips, so it now counts the real list and requires every non-laptop entry to name a bug (0b130da); that test is byte-copied into forge-et-pwa, so forge-et-pwa#74 re-copies it with its hash manifest and merges right after forge-et-fe#245 (Dinil’s decision, 2026-10-13).
Regression test: a pnpm e2e:gate run on Windows that passes with the skip counted.
Resolved: 2026-10-13 — fixed in forge-et-fe#245 (d25db43, the expected-skip entry; 0b130da, the gate test counts the real list) and forge-et-pwa#74 (6088e97, the byte copy and its hash; merged as 0a5418ef). Gate runs 2, 3 and 5 on Windows passed with the skip counted (6 skipped, 0 unexpected).
Lesson: L-190 (a file shared by copy with another repo makes its fix a two-repo change). The first cause — a skip added for one platform and checked by a gate run on the other — is one-off.
BUG-187 — Closing the not-in-round dialog while the summary re-read after Carry forward all is still in flight can drop focus to the page
Affects: F26-g (forge-et-fe src/features/pricing/WeeklyPricesTab.tsx, closeNotInRound, and RoundSummaryBar.tsx) · Found in: F26-g Wave 1, Step 5g, the fix 5129e0e’s own report, 2026-10-13 · Severity: low — a narrow timing window; focus is otherwise returned correctly · Phase: 1
Repro: on a round with standards not in it, open the not-in-round dialog, press Carry forward all n, and close the dialog (Esc, ✕ or Close) before the round summary’s re-read answers.
Expected: focus returns to the not-in-round link, or, when nothing is left and the link is gone, to the Weekly prices heading (WI-1.3’s plan).
Actual: focus first returns to the link, which is still drawn from the old summary; when the re-read then answers with notInRound 0 and removes the link, focus falls to body, because the close’s focus repair runs only once, at close.
Fix: re-run the repair when the summary bar’s link unmounts while it holds focus (or arm the repair until the post-carry re-read lands), with a component test that closes the dialog before the re-read resolves.
Regression test: that component test; the keyboard walkthrough’s not-in-round row.
Resolved: 2026-10-13 — fixed in forge-et-fe#250 (6736cd1; test rows 2b47cb3), merged as 6908626a.
- The fix: the close’s repair stays armed until the post-carry summary re-read lands.
- Tests: a keyboard row failed before the fix and passes with it. A pointer row says where focus is right after the close. A “moved” row proves the re-armed repair never takes focus the user moved.
Lesson: none — one-off: a focus repair armed once, before the read that removes its control has landed. L-133 covers the rule.
BUG-186 — A sale created while a correction is being saved can start before that correction, and its round then cannot be published
Affects: F26-c (forge-et-be src/main/java/com/empireteas/pricing/service/AuctionSaleService.java, create), with F26-g’s correction (PriceCorrectionService) · Found in: forge-et-be#248’s delta review (F26-g Wave 1), 2026-10-13 · Severity: low — it fails safely with a refusal, no wrong price is stored · Phase: 1
Repro: a Tea Room user saves a correction of a price whose next sale starts on day S (so the correction’s allowed range ends before S), and at the same moment someone creates a new sale starting between the correction’s effectiveFrom and S. AuctionSaleService.create takes neither the round lock nor the published-price lock and does not check for a later published price, so it can commit between the correction’s firstStartAfter(today) read and its commit.
Expected: a new sale cannot start inside a published price’s validity that the sale’s round would then have to replace before the price starts, or the correction is refused when such a sale appears first.
Actual: both commit. The new sale starts before the correction does; publishing its round later fails with 409 PUBLISH_CONFLICT (rolled back), and the user must withdraw the correction first. The same already happens on main for a sale created before a published later sale (F26-f).
Fix: not decided — options: have create take the published-price lock and refuse a start on or after the earliest published price’s start beyond it (F26-g FR-12’s rule for a date edit), or accept the refusal at publish time and say so in the help text. Decide with the F26-c/F26-g owner.
Also (found by forge-et-be#249’s review, 2026-10-13): the same gap without any race.
- Repro: a correction is accepted from +10 days while the next sale starts at +14. Later, a sale starting at +7 is created.
- Why it happens:
AuctionSaleService.createtakes no sale or published-price lock, and does not make theLATER_PRICE_PUBLISHEDcheck thatAuctionSaleDatesServicemakes for a date edit. - Result: the sweep ends the older sale’s prices at +7, so the standard is unpriced from +7 to +10, which is BUG-193’s symptom. BUG-193’s fix (FR-10 Rev 2) closes this only when no later sale exists.
- Fix: the first option above — the create refuses a start on or after a published price’s start beyond it, under the published-price lock — closes both.
Regression test: a concurrency test in the style of
PricingCalendarIT’s date-edit-vs-correction cases: a sale create committed while a correction waits. Status (2026-10-13): fix written, deferred by Dinil. forge-et-be#251 refused a create when a published price anchored to a sale that starts before the new sale’s start had a validity starting on or after that start, underlockPublishedPrices. forge-et-fe’s browser fixtures, which create sales starting today and three days ago while the same run publishes corrections from today, tripped it. Resolved: 2026-10-15 — fixed in forge-et-be#251 (e9693d1d…6f2f8154;2f0bdc75, replaced by49727ea7; review fixes4a66ff19,f6609becand7d1ddfcc), merged as3214324e. No frontend change: the New sale dialog already shows the refusal beside Start date (forge-et-fe#250). - The rule: every create takes
lockPublishedPricesfirst and is refused409 LATER_PRICE_PUBLISHEDwhen a published price anchored to a sale starting before the new start starts on or after the later of that start and tomorrow. A price that starts today or earlier never counts. - How it got there: the 2026-10-13 rule counted a price starting on or after the new start, and failed 4 pricing browser and seam tests, each creating a sale that starts today while a correction from today existed. A price cannot be withdrawn once it has applied to a day before today (AD #14), so counting a started price would refuse, for good, each later create starting on or before its first day.
2f0bdc75then limited the check to sales starting after today, but its review found that reopened the gap for a sale starting today or earlier against a correction dated after today (a correction may run up to the next sale). On 2026-10-15 Dinil chose to check every create and count only prices starting after today. - What remains: a correction from today does not refuse a sale created to start today or earlier. That sale’s round leaves the standard’s draft
later_price_publisheduntil tomorrow, while the correction keeps it priced, and a publish the next day publishes it. - Spec: F26-c FR-1 Rev 2; F26-g Wave-1 contract Rev 6 (§ Locks).
- Tests:
PricingCalendarITpins the refusal for a sale starting after today, today, tomorrow and before today, the tomorrow boundary, a correction from today with the next day’s publish, and both lock orders (40 tests, 0 failures);AuctionSaleControllerIT23 tests, 0 failures. The pricing seam and browser tests on forge-et-femain(a18d03c) against49727ea7: 292 passed, 0 failed. Before that, the fullpnpm e2e:gateagainst2f0bdc75had one unrelated failure, a users seam test refused by the dev server under load, which passed 20 of 20 alone; it is BUG-196.
Lesson: L-065 — the date edit already took the lock and the check; the sale create, the other writer of a sale’s start, did not.
BUG-185 — A tooltip closes while its trigger has keyboard focus when focus arrives within 200 ms of the pointer leaving
Filed as BUG-182 on feature/F26-g-plan and renumbered when main (which had its own BUG-182 to BUG-184) was merged in. The old id stays in forge-et-fe commit 8934a36 and its branch fix/BUG-182-tooltip-focus-close, merged with forge-et-fe#245.
Affects: FS-3 (forge-et-fe src/ui/components/Tooltip/Tooltip.tsx, the shared Tooltip) and every screen using it — found through F26-g’s Published · Corrected status (RoundStatus.tsx) · Found in: F26-g Wave 1, Stage 7 Step 5e, evidence run 3, 2026-10-13 · Severity: medium — an accessibility defect (the tooltip’s text does not show on keyboard focus) in a reachable state · Phase: 1
Repro: hover a tooltip trigger, move the pointer off it, and move keyboard focus onto the trigger within 200 ms (for example Tab from the control before it). Found by e2e/browser/pricing/round-correct.spec.ts:250 (AC-25 rendered), whose keyboard step came within 200 ms of page.mouse.move(0, 0) in one run of three; reproduced in a Vitest case beside Tooltip.test.tsx (hover, leave, Tab at once: open right after focus, closed 300 ms later with focus still on the trigger; the control case with a 300 ms wait stays open).
Expected: the tooltip shows whenever its trigger has keyboard focus (WCAG 2.1 SC 1.4.13 and the design system’s tooltip rule), however soon after the pointer left.
Actual: handlePointerLeave calls scheduleClose, which checks focus once and then starts a 200 ms timer (TOOLTIP_CLOSE_DELAY_MS) whose callback calls setOpen(false) without checking focus again. Focus inside those 200 ms calls the library’s state.open(true), but the controlled open is still true, so useControlledState does not call onOpenChange (Object.is on the same value), and the timer then closes the tooltip while the trigger keeps focus; it reopens only after a blur and a new focus.
Fix: in the close timer’s callback, check focus again before closing (or clear the hover close timer when the library reports a focus open), with the Vitest case above as the regression test.
Regression test: the Vitest case (hover, leave, focus at once → still open after 300 ms), and round-correct.spec.ts AC-25 rendered in the browser suite.
Resolved: 2026-10-13 — fixed in forge-et-fe#245 (8934a36, filed as BUG-182; the test moved to fake timers in 79cda0e), merged as f8f0709c. The Vitest case fails with the fix reverted and passes with it; round-correct.spec.ts passed in every later gate run.
Lesson: none — one-off: a library that skips onOpenChange when the controlled value does not change, with a close timer that checked focus only when it started.
BUG-184 — Flaky test: forge-et-pwa’s size matrix not-found state ran past its 30-second timeout once in WebKit at 320×640, with the page already rendered (Windows)
Windows only: seen only on Windows 11 and not reproduced on macOS (see Not reproduced on macOS below). Fix it and prove the fix on a Windows machine; a macOS run proves nothing here.
Affects: F40 (forge-et-pwa e2e/browser/size-matrix/size-matrix.spec.ts:103, state not-found from e2e/pageStates.ts, URL /f40-no-such-page) · Found in: pnpm e2e:gate on forge-et-pwa#71 (23df2a3), Windows 11, 2026-10-13 · Severity: low — one size-matrix state failed once, which fails the whole gate; the product is not affected · Phase: 1
Repro: pnpm e2e:gate on Windows 11 with 8 workers and nothing else building. The not-found state at size-320x640-webkit ended with “Test timeout of 30000ms exceeded” and no pending step in the call log. The page snapshot shows the not-found screen fully rendered: the banner, the level-1 heading “That page doesn’t exist.” and the Back to capture button. The other 125 tests passed in 2.6 min. The re-run on the same commit a minute later passed all 126 in 1.1 min.
Expected: the state’s page check finishes well inside 30 s at every size.
Actual: after the heading was visible, the rest of the test (the page title check and the shared page check: overflow, target sizes, axe) did not finish within 30 s, once.
Fix: not yet decided. First find which step stalls: run the state with --repeat-each=20 in WebKit at 320×640 and trace a failing run. A per-step timeout on the shared page check would name the step next time instead of the whole test timing out.
Regression test: the state passes in repeated WebKit runs at 320×640 (for example 20 repeats).
Resolved: 2026-11-13 — closed as wontfix (not reproduced), the lead’s decision (Tharinda). It failed once, on 2026-10-13, and has passed every run since:
- 20 of 20 on macOS (below);
- on Windows on 2026-11-02, 20 of 20 alone, 20 of 20 with every core busy, and 280 of 280 across the WebKit sizes (below);
- on Windows on 2026-11-13, 20 of 20
--repeat-each=20runs in WebKit at 320×640 and 112 of 112 across the whole size matrix in both engines, with no retries; flaky-report.py --days 14on 2026-11-13 (20 forge-et-pwa PRs and 19 nightly runs) does not list it.
No cause was found, and the original run had no load record, so test-strategy.md § Flaky Tests step 4 (machine overload) did not apply. The step the Fix: line asks for is in place: forge-et-fe#529 and forge-et-pwa#126 (2026-11-13) run each part of the size-matrix test and each phase of the shared page check as a named Playwright step that times out just before the test would. A recurrence now fails as “Step timeout … exceeded” at the step that stalled, in the first attempt’s error in the gate’s Playwright report, even when the retry passes. Under the flaky-report rule a recurrence is filed as a new bug that links here.
Lesson: L-162 — another failure that forced load did not reproduce, here closed as not reproduced with the lead’s agreement rather than as overload; added to L-162’s sources. Its new part is that a test timeout names no step, and the gate traces only the retry, so a stall that passes on retry leaves no record of where it stopped; the named steps with their own deadlines (forge-et-fe#529) fix that for the size matrix.
Not reproduced on macOS (2026-10-14, Tharinda’s claim released): on forge-et-pwa main (1bc03bb), not-found at size-320x640-webkit passed 20 of 20 alone (--repeat-each=20). The whole size matrix with --repeat-each=5 --workers=8 passed 160 of 160. The bug stays open for a Windows machine, where it was seen (tracker.md § Bug Tracking: a flaky failure is checked on the operating system that showed it).
Not reproduced on Windows (2026-11-02, Biseni): Windows 11, 12 cores, 16 GB, on forge-et-pwa main (5a984c4, Playwright 1.64.1), following test-strategy.md § Flaky Tests’ triage order. pnpm exec playwright test e2e/browser/size-matrix/size-matrix.spec.ts --project=size-320x640-webkit --grep not-found --repeat-each=20 --workers=2: 20 of 20 passed alone (1.1 min), and 20 of 20 with twelve busy loops beside it, one per core, the CPU at 100% (2.8 min). Every WebKit size project of the size matrix with --repeat-each=5 --workers=8: 280 of 280 passed (7.8 min), not-found at 320×640 five times among them. The same run with the Chromium projects added did not start: Windows refused to create more processes (fork: Resource temporarily unavailable) with about 3.7 GB of memory free, so the original gate’s full load (both engines, 8 workers) was not repeated. Triage step 4 does not apply: the original run names no other work on the machine and predates the gate’s load record, so the bug stays open as a defect to find. Next step: when it recurs in a Windows pnpm e2e:gate, the gate’s retry keeps a trace (playwright.config.ts TRACE) and its result.json keeps the load and any stall; read which step of the test was running from that trace before changing anything.
BUG-183 — Flaky test: the size matrix’s account-detail-contact-row-failed state fails intermittently in WebKit: the Make primary button is never stable enough to click (Windows)
Windows only: seen only on Windows 11 and not reproduced on macOS (see Not reproduced on macOS below). Fix it and prove the fix on a Windows machine; a macOS run proves nothing here.
Affects: F1-d (forge-et-fe e2e/browser/size-matrix/size-matrix.spec.ts:136, state account-detail-contact-row-failed, arranged in e2e/browser/size-matrix/arrange.ts) · Found in: pnpm e2e:gate on forge-et-fe#227 (bc2c6d4), Windows 11, 2026-10-13 · Severity: low — one size-matrix state fails now and then, which fails the whole gate; the product is not affected · Phase: 1
Repro: a full ET_E2E_WORKERS=6 pnpm e2e:gate on Windows 11, nothing else building, forge-et-be main on the local profile. At 768×1024 in WebKit the arrange step’s click on the row’s Make … primary button (account-contacts-make-primary) waited the whole 30 s test timeout: Playwright resolved the button, visible and enabled, and reported “element is not stable” on every attempt. The same state passed at every other size and in Chromium in that run.
Expected: the arrange step clicks the button once it has settled, at every size.
Actual: in WebKit at 768×1024 the button kept moving (or its box kept changing) for 30 s, so the click never happened and the state failed.
Fix: not yet decided. First find what moves: a transition or animation on the row, a layout that keeps reflowing at that width, or a scroll the arrange step causes. Then either stop the movement for the test (reduced motion, or wait for the row to settle as settleLeadPanels does for the lead view) or fix the layout if it never settles.
Regression test: the state passes in WebKit at 768×1024 in repeated runs (for example 20 repeats with --repeat-each).
Resolved: 2026-10-30 — fixed by BUG-218’s fix, forge-et-fe#365: the browser suite runs against a development build served by vite preview instead of the dev server, so each test’s fresh browser context loads some 70 files instead of 270 source modules, and Windows starts at most 6 browsers. Under the old load WebKit could not settle the row for 30 s; the row itself needed no change. The state passed 20 of 20 in WebKit at 768×1024 (--repeat-each=20 on ece00c5a), and in all three of BUG-218’s full gates in a row (a full ./gradlew build alongside), with no size-matrix test flaky.
Lesson: L-248.
Also seen (2026-10-13, F20-c’s frontend wave gate on af04c43): the same Make primary step, makeSecondPrimary, in the sibling state account-detail-contact-forbidden at 1024 × 768 in WebKit. The click went through, but the account-contacts-notice never appeared within 5 s.
- Run three times alone, that state failed once in six at that size.
- Run with
--repeat-each=10at both WebKit sizes, it passed 20 of 20 onmain(78e025e) and 20 of 20 on the wave against the same backend.
So the instability is load-related and comes before F20-c, not from it. The fix should cover both states.
Not reproduced on macOS (2026-10-14, Tharinda’s claim released): on forge-et-fe main (0ea8055), on an isolated backend stack, every account-detail-contact-* and pricing-import-* state at both WebKit sizes, with --repeat-each=10 --workers=8, passed 869 of 869. That run covers BUG-013 too.
A lead for whoever picks it up on Windows:
settleAccountPagewaits for the contacts and the standards, but not for the contact row’s lazily loaded Email filler (ContactActionsPlace,settleLazyFillerine2e/browser/accounts/detail-support.ts).- When that chunk loads late, it lands in the row’s wrapping action group beside Make primary and can move it.
- If the trace shows the row reflowing, wait for the filler (
settleLazyFiller(page, ACCOUNT_COMPOSE_FILLERS.contact)) before the click.
The bug stays open for a Windows machine, where it was seen.
BUG-182 — Flaky test: Leads inbox AC-14a fails intermittently: a lead-view read lands inside the window where the test counts Retry’s requests (Windows)
Affects: F11-b (forge-et-fe e2e/browser/leads/leads-inbox.spec.ts:416, “AC-14a (smoke): with the tags read failing, its select shows the error and Retry; the list, Country and the lead view still work; Retry…”) · Found in: pnpm e2e:gate on forge-et-fe#227 (bc2c6d4), Windows 11, 2026-10-13; the same test also failed in an earlier full run on 2026-10-12 · Severity: low — the test fails now and then, which fails the whole gate; the product is not affected · Phase: 1
Repro: a full ET_E2E_WORKERS=6 pnpm e2e:gate on Windows 11 with the CPU near 90 % (the gate’s browsers plus other desktop apps). The test asserts that pressing Retry on the failed tags list sends exactly one request, /api/v1/tags; it received ["/api/v1/tags", "/api/v1/leads/554/shares"].
Expected: the assertion counts only the requests Retry causes.
Actual: the open lead view’s sharing panel read its shares (/api/v1/leads/554/shares) late, inside the counting window, so the test saw one request too many.
Fix: not yet decided. Options: let the lead view’s reads settle before the counting starts (as settleLeadPanels does elsewhere), or count only requests to the list Retry is for. Not by widening the expectation to accept any extra read, which would hide a Retry that re-reads too much.
Regression test: the test passes in repeated runs under load (for example 20 repeats with --repeat-each while another suite runs).
Lesson: L-101 (a browser test waits for the page to settle) — an instance. The counting window opened before the lead view’s on-demand sharing panel had made its read.
Resolved: 2026-10-13 — fixed in forge-et-fe#250 (1632ea2), merged as 6908626a.
- The fix: before counting Retry’s reads, the test now waits for the whole lead view to settle:
settleLeadDetail, thensettleLeadPanels, with the sharing panel shown and not loading or refreshing. - What it also removes: a second hidden source in the same window, the Tags section re-reading the failed tags list.
- Assertion: still
toEqual([each.path]). - Gate: full runs on
b39c599passed. Also seen: F26-g’spnpm e2e:gate --pushon forge-et-fe#245 at0b130da, 2026-10-13 (Windows 11, 8 workers, 3,426 tests):["/api/v1/tags", "/api/v1/leads/98/shares"]; the Country case passed in the same run and the next full run passed. Filed there as a separate bug, then folded in here as a duplicate.
BUG-181 — In WebKit, a blocked control’s reason tooltip can close at once when tabbing to the control scrolls it into view
Affects: FS-3 (forge-et-fe src/ui/components/Tooltip/Tooltip.tsx, the shared tooltip on every blocked control, design-system §9); seen on F5-f’s Share… button (SharingPanel.tsx SHARE_BLOCKED_UNOWNED) · Found in: BUG-175’s review probe (forge-et-fe#242), 2026-10-13 · Severity: low — the reason is lost now and then, never the control · Phase: 1
Repro: a WebKit size-matrix run of leads-lead-share-unowned that reaches the aria-disabled Share… button with Tab from the top of the page (keyTo from e2e/browser/pricing/round-support.ts) instead of focusing it after it is in view: 20 repeats at each WebKit size, three workers.
Expected: the tooltip “A lead with no owner can’t be shared. Set an owner first.” shows whenever the blocked button has keyboard focus (design-system §9: disabled with a reason; WCAG 1.4.13, content on focus stays until dismissed or focus moves).
Actual: 7 of 40 runs had the button focused and no tooltip (3 at 1024 × 768, 4 at 768 × 1024). The Tab scrolls the detail column to the button, and react-aria’s useCloseOnScroll closes the tooltip when that scroll event arrives after the tooltip has opened (BUG-175’s cause). Chromium at 6× CPU throttling showed it 12 of 12 with Tab, so it has been seen only in WebKit (Safari).
Fix: not decided — a change to the shared Tooltip’s behaviour, for the lead. One option: ignore a close that arrives while focus is still on the trigger and the scroll is the one the focus caused (for example, a close within a frame or two of opening on focus); another: reopen on such a close while the trigger keeps focus. Either keeps react-aria’s close on a scroll the user makes.
Regression test: a Tooltip component test that opens on focus and fires a scroll on an ancestor right after, expecting the tooltip to stay; and the WebKit Tab probe above, green across repeated runs.
Resolved: 2026-10-13 — fixed in forge-et-fe#246 (merged as 78e025ee; the fix fec9f997, review fixes dc9e0aba and 935b2e6e), the lead’s choice of the first option in the form the browser showed it must take. Instrumenting Tooltip in WebKit showed the failure was worse than 7 in 40: the library opens the tooltip on focus while the button is still below the fold (top 1086px in a 1024px window), WebKit then scrolls it to 526px, and the scroll’s report closes the tooltip about 25ms later — every time; the old size-matrix check passed only because its toBeVisible caught that moment. Ignoring the close was tried and is wrong: the kept tooltip stays where WebKit first placed it and the page grows taller than the window. The fix: within two frames of a focus open, a close is followed by one reopen on the next frame, as a new tooltip element (a key) that react-aria places where the scroll left the trigger, unless a key or pointer was pressed or focus left the trigger; the reopen ends the window, disabling cancels a pending reopen, and unmounting removes the frame and the listeners. Evidence: Tooltip.test.tsx 22 passed, each guard shown to fail its row when broken; pnpm test 5,744 passed; the size-matrix state, now reaching Share… by Tab with no settling and checking the reason is inside the window, 70 of 70 at every size in Chromium and WebKit, and 11 of 20 failed in WebKit without the fix; the browser gate, the test record and an approve review record (frontend-reviewer) on the PR.
Lesson: L-162 (another bullet: a polling visibility check can pass on a transient state; assert the settled one, such as the overlay inside the window).
BUG-180 — forge-et-pwa’s secret scan reads every branch, as forge-et-fe’s did before BUG-179
Affects: FS-4 (forge-et-pwa .github/workflows/pr-validation.yml, the No secrets in any commit (gitleaks) step, line 180 on main as of this filing) · Found in: fixing BUG-179, 2026-10-12 · Severity: medium — no branch carries a finding today · Phase: 1
Repro: read the step: the checkout uses fetch-depth: 0, which fetches every branch, and the scan is gitleaks git --no-banner --redact . with no --log-opts, so it scans every ref.
Expected: each PR’s scan covers the commits HEAD reaches: the base branch’s history plus the PR’s own commits (lesson L-081).
Actual: a finding on any pushed branch fails every capture-app PR’s Fast lane, as BUG-179 did in forge-et-fe.
Fix: forge-et-fe#239’s change: SCOPE="--log-opts=HEAD" on the scan, preceded by the same self-test on a throwaway repo. The workflow is not on forge-et-pwa’s shared-with-fe list, so the change is made by hand.
Regression test: the step’s self-test, which fails when the scope is removed; reproduce locally with gitleaks 8.30.1 before and after.
Resolved: 2026-10-13 — fixed in forge-et-pwa#73, forge-et-fe#239’s diff applied as a patch (the two steps were identical before it); only the comment naming the bug differs. With comments removed, the step’s lines match forge-et-fe’s main (diff exit 0). Reproduced locally with gitleaks 8.30.1: the extracted self-test exits 0 with --log-opts=HEAD and 1 with --log-opts=--all (“the scope is not HEAD-only”); the scoped scan of the branch finds no leaks. All three app repos’ Fast lanes now scan only HEAD’s history, as the harness’s Repo hygiene does. CI change approved by Tharinda, 2026-10-13.
Lesson: L-081, another instance — the same unscoped step as BUG-179’s, in the third app repo; with this fix every repo’s scan follows the lesson’s second point.
BUG-179 — forge-et-fe’s secret scan reads every branch, so one branch’s false positive fails every PR’s Fast lane
Affects: FS-4 (forge-et-fe .github/workflows/pr-validation.yml, the No secrets in any commit (gitleaks) step) · Found in: forge-et-fe#236’s Fast lane, 2026-10-12 · Severity: high — every open forge-et-fe PR’s Fast lane failed · Phase: 1
Repro: push a branch with a gitleaks finding, then open any other PR. Seen on 2026-10-12: generic-api-key matched two test Idempotency-Key UUIDs (const KEY = '…' in sendEmailApi.test.ts and apiClient.test.ts) on #235’s branch feature/F20-c-WI-1.2-compose-dialog-and, and #232, #236 and #237, which do not contain them, failed with “leaks found: 2”.
Expected: a PR’s scan covers the commits HEAD reaches — the base branch’s history plus the PR’s own commits — as forge-et-be (forge-et-be#83) and the harness already scan (lesson L-081).
Actual: the checkout’s fetch-depth: 0 fetches every branch, and gitleaks git with no --log-opts scans every ref.
Fix: SCOPE="--log-opts=HEAD" on the scan, preceded by forge-et-be’s self-test on a throwaway repo. A CI change, approved by Tharinda on 2026-10-12. The two UUIDs themselves are false positives that #235 must still clear before it merges, because once on main they are in every PR’s history: an ignore entry on #235 (.gitleaksignore with the two fingerprints, or a scanner rule for UUIDs in tests) is Dinil’s change on his PR.
Regression test: the step’s self-test: a key on a branch HEAD does not reach must pass, and a key added then removed on HEAD’s history must fail.
Resolved: 2026-10-12 — fixed in forge-et-fe#239. Reproduced locally with gitleaks 8.30.1 on a checkout with #235’s branch fetched: the old command exits 1 (1294 commits, 2 leaks), the new one exits 0 (1204 commits, none). The self-test, extracted from the workflow, exits 0 with the scope and 1 without it (“the scope is not HEAD-only”). forge-et-pwa has the same step: BUG-180.
Lesson: L-081, another instance — its second point, “scope a repository scanner to the commits the PR brings”, was applied to forge-et-be only; the frontend’s step kept the unscoped form.
BUG-178 — A contributor that throws at the end of an OAuth flow is logged at WARN, though the catch calls it a bug
Affects: F20-a (forge-et-be src/main/java/com/empireteas/auth/config/CapturingTokenResponseClient.java, recordOutcome, the catch (RuntimeException e) at line 170 as of this filing) · Found in: backend-reviewer on forge-et-be#244 (BUG-174 and BUG-176), 2026-10-12; pre-existing code · Severity: low · Phase: 1
Repro: read the code. When an AuthorizationCompletionContributor throws, recordOutcome catches any RuntimeException, logs it with log.warn(…) and sends the browser to the failure redirect. Its own comment says a contributor that throws “is a bug in the contributor”. Since BUG-174, a KMS call that times out during a mailbox connect (ApiCallTimeoutException) reaches this catch too, as does a database fault in connectOrUpdate.
Expected: a fault is logged at ERROR with the exception last (coding-standards §8.3: “a broad catch (Exception) cannot pick the level correctly — it must distinguish expected domain exceptions … from everything else”), because ERROR is the level the alarms watch. An expected outcome a contributor signals by exception, if any, stays at WARN or below.
Actual: every contributor failure is a WARN, so a KMS or database outage that fails every mailbox connect raises no alarm.
Fix: log the catch at ERROR; if a contributor has an expected refusal it signals by exception, catch that type first at WARN. Check MailboxAuthorizationCompletionHandler.complete, which already turns MailboxOwnerChangedException into a refusal before it reaches this catch.
Regression test: a spec on CapturingTokenResponseClient with a contributor that throws, asserting one ERROR line naming the exception class and the failure redirect.
Resolved: 2026-10-12 — fixed in forge-et-be#245 (1fc818e9; review fix 8d52cac5), with BUG-177. A contributor did signal one expected refusal by exception: a mailbox that already has a live connection reached the catch as a DataIntegrityViolationException (MailboxConnectFlowIT pinned it at WARN). Its contract (AuthorizationCompletionContributor) already says a contributor must not throw to signal a refusal. So the refusal moved into the contributor, not into a type check in auth. MailboxAuthorizationCompletionHandler answers it with refusal() and a WARN naming the index, never the exception, but only when Hibernate names mailbox_connection_owner_live_idx or mailbox_connection_account_live_idx. The review found that matching SQLState 23505 alone would misreport a unique index added later, because the catch covers the whole connect transaction. Anything else propagates. CapturingTokenResponseClient now logs what escapes at ERROR, with the exception, and the contract’s Javadoc says so. Regression tests:
CapturingTokenResponseClientSpec: a contributor that throws gives one ERROR line with the exception class, and the failure marker.MailboxAuthorizationCompletionHandlerSpec: both live indexes give the marker; a 23505 on another index and a 23502 propagate.MailboxConnectFlowIT: the real duplicate row logs[constraint=mailbox_connection_account_live_idx], no fault line, and no account id.
With the main files reset to origin/main, the first and third fail (WARN, not ERROR; no refusal line), and so does the handler spec’s duplicate feature. /forge-review-pr with backend-reviewer: one Important (the SQLState-only match) and four Nits fixed in 8d52cac5, one Nit left as is (Spotless’s [: ]); the delta found nothing. ./gradlew build on the head 8d52cac5 (the Test record check): 6,883 tests, 0 failures, 2 skipped.
Lesson: L-149: its fourth point, that a bound turns a hang into a failure, which needs a path logged at a level an outage would be noticed by, as BUG-113 was. BUG-174’s bound sent the KMS timeout into this WARN catch. BUG-178 is added to its sources.
BUG-177 — The attachment download response prints its presigned URL in its default toString, and the banned-marker check does not look for url
Affects: FS-11, the file pipeline (forge-et-be src/main/java/com/empireteas/attachment/web/DownloadUrlResponse.java url; scripts/check-banned-markers.py) · Found in: BUG-176’s fix (forge-et-be, 2026-10-12), while probing which names the check should add · Severity: low · Phase: 1
Repro: run BUG-176’s probe, scripts/check-banned-markers.py with url added to SENSITIVE: it reports DownloadUrlResponse (url), plus AuthorizationOutcome (redirectUrl, the landing URL on the SPA, not a secret) and FilePipelineProperties (downloadUrlTtl, a duration).
Expected: a record holding a presigned URL overrides toString() to leave it out: its own Javadoc says the URL is “never logged”, and the backend’s rules say never to log a presigned URL (forge-et-be/CLAUDE.md, the file-pipeline convention; coding-standards §8.6). The check refuses the next one.
Actual: DownloadUrlResponse uses the record default toString(), which prints the signed URL, a read grant on one object for its lifetime. Nothing logs one today as far as review has looked; the first log line or exception message that includes the response would.
Fix: a redacting toString() on DownloadUrlResponse (<redacted>, as BUG-176’s records do) with a T1 spec; add url to SENSITIVE, and put AuthorizationOutcome and FilePipelineProperties in TO_STRING_EXEMPT with their reasons. Rerun the probe first: other records may have gained a url since.
Regression test: the T1 spec on the toString(), and a --self-test case for a url component that fails before the SENSITIVE change.
Resolved: 2026-10-12 — fixed in forge-et-be#245 (44acbc31; a review fix to an exemption reason in 8d52cac5), with BUG-178. DownloadUrlResponse prints <redacted> for its URL. The JSON body still carries it, because Jackson serializes the component; ScanOffUploadIT fetches it from the body. SENSITIVE gained url. The probe, rerun on main (e6221198), flagged the same three records as at filing. AuthorizationOutcome (an SPA landing URL) and FilePipelineProperties (a duration) went into TO_STRING_EXEMPT with their reasons. The review found that url also matches MailboxProperties.publicBaseUrl, which was already exempt, and extended that record’s reason. uri was left out: it flags only a redirect template. Regression tests: DownloadUrlResponseSpec fails with DownloadUrlResponse reset to origin/main. The UrlRecord --self-test case fails before url is added (UrlRecord.java gave 0 finding(s), expected 1) and passes after. ./gradlew build on the head 8d52cac5: 6,883 tests, 0 failures, 2 skipped.
Lesson: L-187: a presigned URL is a provider value named for its use, which a name-based check misses. BUG-177 is that lesson’s second source.
BUG-176 — Three mailbox records print Graph paging links in their default toString, and the banned-marker check misses them by field name
Affects: F20-b (forge-et-be src/main/java/com/empireteas/mailbox/service/BackfillPosition.java cursor; src/main/java/com/empireteas/mailbox/MailboxProvider.java MessagePage.nextCursor and DeltaPage.nextLink/deltaLink; scripts/check-banned-markers.py) · Found in: backend-reviewer’s delta review of forge-et-be#223 (BUG-102), 2026-10-10 · Severity: low · Phase: 1
Repro: on forge-et-be#223, the Fast lane’s “No banned markers in production code” step refused the new DeltaPosition record (“holds token — override toString() to leave it out”). The three records above hold the same kind of value — a Graph nextLink or deltaLink, carrying a $skiptoken or $deltatoken — and the check passes them, because its SENSITIVE list (scripts/check-banned-markers.py line 39 as of this filing) matches token but not cursor or link.
Expected: every record holding an opaque provider link overrides toString() to leave it out (coding-standards §2.1, §8.6), as the audit already marks backfillCursor and deltaToken sensitive; and the check refuses the next one.
Actual: all three use the record default toString(), which prints the link. grep finds no log call or exception message that passes one in mailbox/ today, so nothing leaks yet; the first log.debug("… {}", position) would.
Fix: a redacting toString() on each of the three (<redacted> or null, as DeltaPosition and MailboxTokenService.CachedToken do), with one T1 spec in the shape of DeltaPositionSpec; and widen the check — add cursor, or the exact names nextLink, deltaLink and nextCursor, to SENSITIVE. Run the script with each choice first to see what a bare link or cursor substring also catches, and exempt any non-secret field in TO_STRING_EXEMPT.
Regression test: the T1 spec on the three toString()s, and the check’s --self-test gaining a cursor/link case that fails before the SENSITIVE change.
Resolved: 2026-10-12 — fixed in forge-et-be#244 (8f6e2525; review fixes f25d3715), with BUG-174. The three records and a fourth the filing missed, the provider’s private GraphPage (the Graph response, with the same two links), print <redacted> for a set link, as DeltaPosition does. SENSITIVE gained cursor and link. The script was run over src/main with each candidate first: cursor and link matched only these four records, so TO_STRING_EXEMPT is unchanged. url matched a presigned URL, filed as BUG-177. Regression test: PagingLinkToStringSpec (4 features after review) fails 5 of 5 with the four main files reset to origin/main; --self-test with the new CursorRecord and LinkRecord cases fails before the SENSITIVE change (CursorRecord.java gave 0 finding(s), expected 1) and passes after. /forge-review-pr with backend-reviewer: no Blocker or Important. Its Nits on the spec and on two bound comments were fixed in f25d3715; the link substring was left as is. ./gradlew build on the head f25d3715 (the Test record check): 6,860 tests, 0 failures, 2 skipped.
Lesson: L-187 (new): a check that finds secrets by component name misses a provider value named for its use, such as cursor or nextLink. BUG-177 is a second instance; backend-reviewer’s toString() row points at the lesson.
BUG-175 — Flaky test: the size matrix’s leads-lead-share-unowned state fails intermittently in WebKit: the blocked Share button’s tooltip does not appear (macOS)
Affects: F5-f (forge-et-fe e2e/browser/size-matrix/arrange.ts:5889, the leads-lead-share-unowned after-step; the tooltip is SharingPanel.tsx’s SHARE_BLOCKED_UNOWNED) · Found in: the browser gate on forge-et-fe#230, 2026-10-12 · Severity: low — a flaky browser-gate test · Phase: 1
Repro: macOS, pnpm e2e:gate with three workers on forge-et-fe#230’s head (c9ade94), while two other agents’ test runs (Vitest, Gradle with Testcontainers) were loading the machine. The after-step focuses the aria-disabled Share… button with shareButton(page).focus(), checks it is focused, then waits 5 seconds for the tooltip “A lead with no owner can’t be shared”.
Expected: the tooltip shows on focus at every size, as the blocked-control pattern requires (design-system §9, disabled with a reason).
Actual: at 1024 × 768 in WebKit the tooltip was not found within 5 seconds; the button was focused. The same state passed at the other sizes and in Chromium in that run, and 17 of 17 in WebKit 1024 × 768 alone straight afterwards (--repeat-each=8 --workers=2).
Fix: not yet known. Reproduce under load first (L-162): whether the tooltip’s open delay, or how WebKit reports a programmatic focus() to the tooltip’s focus-visible check, is the slow step. If it is the test, wait on what opens the tooltip rather than lengthening the timeout; if the tooltip can fail to open on keyboard focus, it is a product defect.
Regression test: the state in the whole gate on macOS, green across repeated runs, and a WebKit run under a CPU throttle.
Resolved: 2026-10-13 — the test fixed in forge-et-fe#242 (ea2fcdd); the product side split out as BUG-181. The cause was neither the open delay nor focus-visible: the Share… button starts below the fold (top 1156px in a 768px window), focus() scrolls the detail column to reach it, and react-aria closes an open tooltip when an element holding its trigger scrolls (useCloseOnScroll). The browser delivers that scroll event on a later frame, so when it lands after the tooltip has opened, the tooltip closes with the button still focused. Scrolling the column by 1px after the tooltip opened closed it every time, in WebKit and Chromium; a forced window focus event did not (15 of 15 passed), and a 6× CPU throttle in Chromium did not (24 of 24). The state now brings the button into view and waits two frames before focusing (scrollIntoViewSettled, shared with BUG-151’s openAddTagList): 40 of 40 at both WebKit sizes with three workers, and the browser gate on the PR. A probe that reached the button with Tab and no settling missed the tooltip 7 times in 40 in WebKit, so a keyboard user can lose it too: BUG-181.
Lesson: L-162 (its sources gain BUG-175, and a bullet: the throttle may not reproduce a scroll-ordering race, and the probe that decides whether it is a product defect is the user’s own input — Tab, not focus() — in every engine the gate runs).
BUG-174 — The KMS client has no call timeout while a mailbox connect holds the owner’s row lock
Affects: F20-a (forge-et-be mailbox/KmsConfig.java, the KMS-wrapped refresh tokens), F5-c (the deactivation and email change that wait on the connect) · Found in: BUG-100’s fix (backend-reviewer on forge-et-be#235, 2026-10-11; pre-existing code) · Severity: low · Phase: 1
Repro: read the code. KmsConfig.kmsClient builds KmsClient.builder().region(…).credentialsProvider(…).build() with no ClientOverrideConfiguration, so no apiCallTimeout or apiCallAttemptTimeout: only the SDK’s defaults and retries bound a call. Since BUG-100, MailboxConnectionService.connectOrUpdate first locks the owner’s app_user row FOR SHARE, then calls cipher.encrypt, which asks KMS for a data key (KmsTokenKeyProvider.generateDataKey, one call per encrypt), all in one transaction. Cloud profiles only: under local and test the LocalTokenKeyProvider makes no call.
Expected: every outbound call made while database locks are held has explicit short timeouts, as GraphClientConfig gives the Graph client and RefreshTokenResponseClients the token refresh (BUG-101).
Actual: a KMS call that stalls holds the connect’s row lock. An administrator’s edit or deactivation of that user waits for it, and while it waits it holds the installation-wide access.manage advisory lock, so all user and role administration waits too.
Fix: set apiCallAttemptTimeout and apiCallTimeout, as named constants, either on the KMS client in KmsConfig (ClientOverrideConfiguration) or on each request in KmsTokenKeyProvider, the per-request shape S3ObjectStore uses for its head call. The values are the lead’s call, as BUG-101’s were. A unit test asserts the configured client’s timeouts. The bound itself can be proven only against real KMS: the register’s §B F20-a block has the row.
Regression test: a spec that asserts both timeouts on what the fix configures: the built client’s override configuration, or each request KmsTokenKeyProvider sends (against a faked KmsClient).
Resolved: 2026-10-12 — fixed in forge-et-be#244 (9de04546; review fixes f25d3715), with BUG-176. The lead’s values are set on the client in KmsConfig (ClientOverrideConfiguration), as named constants: apiCallTimeout 5 s and apiCallAttemptTimeout 2 s. Both GenerateDataKey and Decrypt get them. They match S3ObjectStore’s existence check and BUG-101’s 5 s refresh bound. The Javadoc states that 5 s is the most a stalled KMS call can hold the owner’s row lock. backend-reviewer checked this against SDK 2.54.3: the call timeout wraps the retries, and a connect makes one KMS call per lock hold. A timeout fails the connect to ?connect=failed, which the user sees. That path logs it at WARN, not ERROR (pre-existing, filed as BUG-178). The review also corrected MailboxTokenService’s worst case under the connection’s row lock: about 25 s, since the refresh holds it across a Decrypt and a GenerateDataKey. Regression test: KmsConfigSpec asserts both timeouts on the built client’s override configuration; with KmsConfig reset to origin/main, its two timeout features fail. Proving the bound against real KMS stays cloud-gated: the register’s §B F20-a row, open.
Lesson: L-149: the KMS client the SDK builds with its defaults is the “SDK defaults” case that lesson names. BUG-174 is added to its sources.
BUG-173 — Browser-gate tests that read every page of a list time out once the local database has grown from earlier runs
Affects: the browser gate’s suite in forge-et-fe, through tests that page through a whole list the suite itself keeps adding to: e2e/seam/leads-inbox.seam.spec.ts:1243 (F11-b AC-3, every lead page once per filter: status, country, customer type, lead source, owner and two tags), e2e/browser/leads/leads-writes.spec.ts:156 (F11-b AC-21, every page of the Owner filter’s user list, userByEmail) and e2e/seam/offboarding-reports.seam.spec.ts:544 (F5-h, users arranged on top of all existing users) · Found in: pnpm e2e:gate on forge-et-fe#227 (bc2c6d4), Windows 11, 2026-10-12 · Severity: medium — the gate fails for a reason unrelated to the PR, so a ready PR cannot get its Browser gate: line until the developer resets the local database; the product is not affected · Phase: 1
Repro: run pnpm e2e:gate repeatedly against one local database (forge-et-be docker compose up, data in forge-et-be/.local/postgres). Each run creates users, leads and accounts and removes none. On a quiet machine (no other build running, 6 workers) with 1,560 leads, 2,559 users and 662 accounts, the three tests above failed: AC-3 ran past its 120 s timeout, the offboarding test past 30 s; the same three had failed in two earlier runs on the same database, while about 1,000 other tests passed.
Expected: a gate test’s run time does not depend on how many earlier runs used the same database.
Actual: the three tests read every page of a list whose length is the database’s history, so they get slower with every run on every developer’s machine until they time out. After deleting .local/postgres (fresh migrations: 7 leads, 22 users, 2 accounts) the gate was re-run on the same commit.
Fix: decided by Tharinda, 2026-10-13: narrow what a test reads where the test itself walks the list, and have the gate stop before it starts when the local database has outgrown it. Only AC-3 walks a list in its own code: each of its filter checks reads every page of the manager’s leads. It now reads, per seeded lead it names, q=<name> with the filter, plus one first page of the filter alone to check that every row satisfies it, so its request count no longer grows with the database. The other two cannot be narrowed in the test. AC-21 waits for the Owner filter and the owner picker, which read every page of the user directory by design (BUG-109; Empire has about 20 users). The offboarding test walks nothing, so its time is the backend’s own work over thousands of users. So under pnpm e2e:gate (E2E_GATE=1) a Playwright globalSetup signs in as the all-permissions principal and counts the user directory and the leads list (page.totalItems). When either is over 1,000 it stops the run before any test, naming both counts and forge-et-be’s reset command (docker compose down && rm -rf .local/postgres && docker compose up -d, then ./gradlew bootRun). On 2026-10-13 a local database showed one full gate run adding about 200 users, 230 leads and 110 accounts, so the limit allows about four full runs from a fresh database before a reset. The tests failed at 2,559 users and 1,560 leads. A plain pnpm e2e is not checked. Not by raising the timeouts, which only moves the point where they fail.
Regression test: a unit test of the limit check (under, at and over the limit for each count, and the message naming the reset); AC-3 passing against the large local database; and a gate run on a database over the limit that stops before any test with that message.
Resolved: 2026-10-14 — fixed in forge-et-fe#252 (95c065e; review fixes f29b371 and ab1db02; main merged in at af9abe6 for BUG-135’s fix), merged as 5c7555b3. Evidence:
- The unit test of the limit check: 7 passed.
- AC-3 under
E2E_GATE=1: 4.1 s against a database of 641 active users and 468 leads. - With the limit set to 100 for one run, the gate run stopped before any test, printing the counts and the reset; a run without
E2E_GATEwas not checked. - The full gate on
ab1db02against a fresh database (compose.verify.yaml): 3,665 passed, 27 failed, 5 skipped. All 27 also fail onorigin/mainwithout this change, and are filed as BUG-133 and BUG-134. The PR carriesBrowser gate: skippedwith that reason, the lead’s decision. - An earlier gate run on a long-used local database also failed 8 tests that expect the IT Admin to hold only
access.manage. That database’sit_adminrole also heldemail.manage: stale local data, not a defect. - The test record on
af9abe6(macOS): 6,597 tests, 0 failed. /forge-review-prwithfrontend-reviewer: one Important and six Nits on the full review, one Nit on the delta. All fixed except one Nit left as is (no test of the globalSetup wrapper; the rule it applies is unit-tested). Lesson: L-147, another instance: run-unique fixtures grow every shared list, here past what a test that walks the whole list can read in time. The lesson gains a bullet (scope the read; the gate’s limit covers a screen that reads every page) and BUG-173 as a source.
BUG-172 — The test record’s own self-test runs past Vitest’s 20-second limit in every full pnpm test (Windows)
Affects: the local test record (forge-et-fe src/test/scripts/localTestRecord.test.ts:118, “the Fast lane documentation-only decision › passes its own self-test, git cases included”, which calls scopeSelfTest from scripts/fast-lane-scope.*; added by cee0d65 and da1cef0; forge-et-pwa carries a byte copy, where no testTimeout is set, so Vitest’s 5 s default applies and the self-test took 15 to 38 s alone) · Found in: a full pnpm test on forge-et-fe#201’s branch after merging main (9eb606b), Windows 11, 2026-10-12 · Severity: high — pnpm test:record runs the whole suite and records nothing when a test fails, and on Windows this test failed every full run, so no PR could get its required Test record check from a Windows machine; the product is not affected · Phase: 1
Repro: on Windows, run the full suite: corepack pnpm test. The self-test makes throwaway git repositories and commits (its “git cases”, one set under a signing config), each a separate git process run synchronously.
Expected: the test finishes inside the 20-second testTimeout (vitest.config.ts) whatever else the suite is running.
Actual: it timed out at 20 s in all three full runs on Windows 11 on 2026-10-12 (the first full run, 308 files and 377 s, then both pnpm test:record runs on forge-et-fe#201’s head 02fbcd0, 232 s and 233 s, 5,472 other tests passing). Run alone (corepack pnpm exec vitest run src/test/scripts/localTestRecord.test.ts) all 10 tests in the file pass, the self-test taking 10.0 s. Process start-up for git is slow on Windows, and slower still while Vitest’s other workers use every core.
Fix: give this one test its own { timeout: 120_000 }, as browserGateCli.test.ts (60 s) and browserGateBoundaries.test.ts (120 s) do for their git-heavy suites. Not by raising the global testTimeout, which would hide slow tests everywhere. Making fewer git calls would shorten the test but still leave it timed against an idle machine.
Regression test: the full pnpm test on Windows, which failed 3 of 3 runs before; the file’s own assertions are unchanged.
Resolved: 2026-10-12 — fixed in forge-et-fe#227 (1df3280; the comment reworded in 5478bce so it holds in both apps, and main merged in at bc2c6d4 to bring forge-et-fe#226’s version of the file), test only, and in forge-et-pwa#71, which re-copies the file from bc2c6d4 (82bd03f): every shared file matches forge-et-fe at that commit (check-shared-drift.mjs --live), and the manifest names it. The self-test carries its own 120 s timeout with a comment saying why; the global 20 s testTimeout is unchanged. The full pnpm test on Windows 11 passes with it (304 files, 5,320 passed, 1 skipped, 245 s), where the three full runs before it each failed on this test.
Lesson: L-125 (a test whose time limit assumes an idle machine fails on a loaded one) — another instance, here a test runner’s per-test timeout rather than a bound the test asserts; added to its source list.
BUG-171 — Flaky test: the size matrix’s leads-lead-tag-picker-many state fails intermittently in WebKit: the last option is 99% in the window
Affects: F11-f (forge-et-fe e2e/browser/leads/lead-detail-support.ts, expectShortTagList, added by BUG-123’s fix in forge-et-fe#224) · Found in: the browser gate on forge-et-fe#226, 2026-10-12 · Severity: low · Phase: 1
Repro: in forge-et-fe at main (97bb763), with the backend running, ET_BACKEND_PORT=<port> pnpm exec playwright test e2e/browser/size-matrix/size-matrix.spec.ts --grep leads-lead-tag-picker-many --workers=3.
Expected: the open Add tag list’s last option is wholly in the window at every size, as it was in forge-et-fe#224’s own gate (2e394b7, all passed).
Actual: toBeInViewport({ ratio: 1 }) on the last option reads a ratio of 0.9917 in WebKit. In a full gate run with three workers, while a backend build ran on the same machine, it failed at 768 × 1024 and 1024 × 768; with three workers and nothing else running, 768 × 1024 failed again with the same ratio; with one worker both sizes passed six times out of six. So the list fits, and something under load (the popover’s position, or the window’s height while WebKit settles) leaves the last option about a pixel short.
Fix: not yet chosen. Either the check measures the option against the popover and the window after the popover’s position settles, or it allows the sub-pixel shortfall; whichever it is, the check must still fail when ten rows do not fit (the 320 × 640 case BUG-123 found).
Resolved: 2026-10-12 — fixed in forge-et-fe#230 (4dc7ad0, review fixes c9ade94), the second option. expectShortTagList polls the last option’s and the line’s bounding boxes and asserts each reaches no more than half a pixel past the window, the tolerance its popover check already allowed; a failure prints the overshoot in pixels, and the popover box is read after the poll, from the settled layout. Ten rows that do not fit leave a whole row outside, so BUG-123’s case still fails. Proof: the whole browser gate on c9ade94, three workers, 3,175 passed, 5 skipped, including this state at every size in Chromium and WebKit (one earlier run failed an unrelated WebKit tooltip state under load, filed as BUG-175).
Lesson: L-186 (new — a browser check that something fits in the window allows half a pixel, and reports the overshoot).
BUG-153 to BUG-170 are every bug row of Nanndish’s manual test sheet for the Phase 1 screens built so far
(2026-10-08, forge-et-fe main at 16a2f07, forge-et-be at e60a5e4e; BUG-166 was added to the sheet on
2026-10-09). The sheet, with one screenshot per finding, is kept outside the repo; its Suggestions rows are
change requests, listed in .forge/owed.md. The bugs fall into four kinds, and each section says which:
- Breaks a written rule: BUG-153, BUG-155, BUG-156, BUG-160, BUG-163, and BUG-168 (closed by BUG-118’s fix).
- Built as a spec, contract or plan decided, and Nanndish decided to change that decision: BUG-154, BUG-157, BUG-158, BUG-159, BUG-161, BUG-162, BUG-164, BUG-165, BUG-166. Each section names the decision it changes.
- Built as the specs say, and waiting on the lead’s decision: BUG-167 and BUG-170.
- Closed
wontfix, because its feature is dropped: BUG-169, the top-bar search (F31 left V1 on 2026-10-11, PRD Rev 60).
These ids were claimed twice before on unpushed branches (from BUG-121, then from BUG-124) and took the next free
range each time main filed other bugs under the same numbers first; the fix branches still carry the earlier
ids until they are pushed.
BUG-170 — Sorting by name exists only on Accounts (and Leads’ Sort menu); Users, Roles and the Standards book cannot be sorted
Affects: F5-c (forge-et-fe src/features/users/UsersTable.tsx, fixed displayName,asc), F5-d (src/features/roles/RolesTable.tsx, server order by name), F26-b (src/features/pricing/StandardsTable.tsx, fixed standardNo,asc) · Found in: manual test pass, 2026-10-08 · Severity: medium (the tester’s) · Phase: 1
Repro: open Users, Roles and Pricing and try to sort a column.
Expected (the tester’s): the name column sorts on every list, as on Accounts.
Actual: Accounts sorts by Name, Customer type and Country from its column headers (F1-b FR-L3); Leads sorts by name through its Sort select (F11-b FR-L5, a queue rather than a table). Users, Roles and the Standards book have one fixed order.
Status: built as the specs say: F5-c, F5-d and F26-b specify no sorting, and shared table sorting is F32-a’s (Phase 2). Waiting on Tharinda’s decision (.forge/owed.md, change requests from the 2026-10-08 test pass). The backend already sorts users by displayName and email and standards by standardNo, name and grade; GET /roles takes no sort parameter.
Fix, if accepted: after BUG-118 merges, SortableHeaderCell on Users (Name, Email) and Pricing (Standard no, Name, Grade), with the sort in the URL as Accounts keeps it; Roles needs a backend sort parameter, or a client-side sort of its few rows.
Correction (2026-10-26): “BUG-118” here does not name a filed bug. main’s BUG-118 is the Combobox’s Show suggestions trigger, fixed on 2026-10-09 in forge-et-fe#200; the shared src/ui table this entry builds on was never filed under its own number. The nearest work is forge-et-fe#337 (design-system Rev 70’s list screens, a draft on 2026-10-26).
Regression test: clicking a header sends sort=<field>,<dir>, writes it to the URL and sets aria-sort on that header only; a second click reverses it; a reload keeps it.
Resolved: 2026-11-02 — fixed by forge-et-fe#355 (the UI revamp of the list screens, merged 0d4fe1b3, 2026-10-29), which gave every table sortable headers through SortableHeaderCell. Users sorts by Display name and Email through the list endpoint (UsersTable.tsx SORTABLE, the sort in the URL); Roles sorts every column in the browser, since its list is not paged (RolesTable.tsx, rolesSort.ts); the Standards book sorts Standard no, Name and Grade (StandardsTable.tsx). Columns that still cannot sort need their endpoint’s sort whitelist, which is the owed.md row “Backend sorting for the columns that cannot sort yet” (Nanndish). Regression tests are #355’s table tests (aria-sort on the sorted header, the sort in the URL). Found already fixed on 2026-10-29 while checking open bugs; no code was changed for this bug.
Lesson: none — one-off: a change request that a later redesign of every table met.
BUG-169 — The top-bar search does nothing
Affects: F5-b (forge-et-fe src/app/TopBar.tsx:43-49) · Found in: manual test pass, 2026-10-08 · Severity: high (the tester’s) · Phase: 1
Repro: click or type in the Search box in the top bar.
Expected (the tester’s): a search across the CRM.
Actual: nothing happens. The box is an icon and a “Search” label with no input and no handler.
Status: built as F5-b FR-9 / AC-9 says: the global search was put in its final place, inert, for F31 to build. F31 left V1 on 2026-10-11 (PRD Rev 60), so nothing will make it work, and the CRM keeps a search on each list screen instead (Users, Accounts, Leads, Pricing, Email — Pricing’s does not match client names by design, F26-b FR-R4; the Client filter does).
Fix: none. A search across the CRM will not be built, because F31 left V1.
Resolved: closed wontfix on 2026-10-12, Nanndish’s decision. Removing the inert control is a separate task: .forge/owed.md’s item Remove the inert top-bar search box from the CRM, which removes it from every route along with the frame.spec.ts assertions that expect it (forge-et-fe#201 removes it on the Users routes only).
Lesson: none — one-off: the control was built inert on purpose for a feature that was later dropped, and no code is wrong.
BUG-168 — Table headings look different from page to page
Affects: F5-c (Users), F5-d (Roles), F26-b (Pricing and the account’s Buying standards panel), F1-b (Accounts) · Found in: manual test pass, 2026-10-08 · Severity: low · Phase: 1
Repro: compare the header rows of the Users, Roles, Pricing and Accounts tables.
Expected: one header style, design-system §5 Table: micro/eyebrow type on surface/card-warm, as on the Users table (the test sheet’s screenshot).
Actual: Roles’ header row has no background; Pricing’s header cells are bottom-aligned; each table has its own CSS module.
Fix: BUG-118’s fix — one shared src/ui Table built from the Users table, which all five tables render through. Nothing separate is built: BUG-118’s fixing PR records its fix_pr here and closes this bug with it.
Regression test: BUG-118’s browser spec, which compares each table’s header cell with the Users table’s.
Correction (2026-10-26): no filed bug carries this fix. main’s BUG-118 is the Combobox’s Show suggestions trigger, fixed on 2026-10-09 in forge-et-fe#200, so nothing will close this bug as written. forge-et-fe#337 (design-system Rev 70’s list screens, a draft on 2026-10-26) moves the Accounts, Users, Roles and Pricing tables onto shadcn/ui; re-check the header rows once it merges, and either close this bug there or file the shared table under its own number.
Fixed as (2026-10-28): re-checked after #337 merged, measuring each table’s first header cell in Chromium at 1280px against the Users table’s (12.5px · 600 · no transform · rgb(44,33,23) on white · middle). Roles, Accounts, Contacts, Leads and Email templates matched. The Pricing standards book and weekly prices (bottom, 6px inline padding), auction sales (bottom), the account’s Buying standards panel (align-bottom) and the Settings tag tables (10.5px uppercase, rgb(110,95,78), transparent) differed; their header classes now centre the labels and the tag tables take the Users style, after which every table measures the same. The 6px inline padding on the book, the panel and weekly prices stays, because their body cells share it so ten columns fit. The reference is the Users header (design-system Rev 41), not §5’s type/micro row this entry names; design-system Rev 76 records it.
Merged with forge-et-fe#355 (2026-10-29): design-system Rev 74 moved every table onto one set of header and cell classes, src/ui/dataTable.ts, the Users table’s, which replaced the per-table header classes above. src/app/tableHeaderStyle.test.ts now checks that shared header (centred, secondary type at the body colour, semibold, on the card fill, sentence case) and that each table this bug named uses it and writes no header style of its own. Rev 74 keeps one exception, which this fix follows: the standards book and weekly prices tables wrap their labels and sit them on the bottom line, so ten and fourteen columns fit at 1024px.
Resolved: forge-et-fe#356, 2026-10-28 (Tharinda).
Lesson: L-242 (new) — a screen element every screen has, styled by each screen, drifts; BUG-163, BUG-164 and BUG-165 have the same cause.
BUG-167 — A role can hold “Publish weekly prices” while its Pricing access is None
Affects: FS-6 (the permission catalog, forge-et-be pricing/PricingPermissions.java, auth/service/RoleService.java refuseInvalidTarget), F5-d (forge-et-fe src/lib/permissionCatalog/catalogModel.ts, src/features/roles/PermissionGroupCard.tsx, permissionModel.ts) · Found in: manual test pass, 2026-10-08 · Severity: high (the tester’s) · Phase: 1
Repro: in the role editor, set Pricing to None and tick Publish weekly prices; save.
Expected (the tester’s): the publish permission cannot be held without access to the Pricing page, where publishing happens.
Actual: the save succeeds. The holder would have the publish permission but no Pricing entry in the navigation and no access to /pricing (it needs buying_standard.view).
Status: built as the specs say: FS-6 Appendix A.1 and F5-d FR-15 keep price.publish an independent tick, and the only dependency the catalog enforces is that an <entity>.all key needs its entity’s view or manage (AD #37). F26-f, now done, confirms the tester’s point: a price round is read only by buying_standard.manage, or by price.publish together with buying_standard.view (F26-f FR-2, NFR-3), so price.publish alone grants nothing the holder can use in the CRM. Decided 2026-10-14 by Tharinda: the catalog refuses the combination; the API refuses it with a 422; the editor clears and blocks the tick; a migration removes the grants of that shape already stored, each audited, and the startup consistency check then refuses one (PRD Rev 71, FS-6 Rev 16, F5-d Rev 3, design-system Rev 57).
Fix (as decided): a declared dependency — price.publish needs buying_standard.view or .manage — enforced in RoleService on save, grant and revoke with its own 422 code, checked by GrantConsistencyChecker, and published in the catalog; the role editor blocks the tick with its reason while Pricing is None, as it does All records, and choosing None clears it. Revise FS-6 Appendix A.1 and F5-d FR-7, FR-15. A migration removes price.publish from every role holding no buying_standard key, with one audit row each.
Regression test: a save of price.publish alone answers 422 and writes nothing; removing the last buying_standard.* key from a role holding price.publish is refused; the editor shows the tick blocked while Pricing is None.
Resolved: forge-et-be#255 and forge-et-fe#261 (2026-10-14).
- Backend. A catalog key can declare
requiresOneOf;price.publishrequiresbuying_standard.vieworbuying_standard.manage.RoleServicerefuses a grant, a save or the removal of the last requirement with422 REQUIRED_PERMISSION_MISSING(a generic code, since the rule is declared on the catalog), ordering a save’s writes so a requirement goes in before its key and out after it. The catalog refuses a requirement on an undeclared key, on the key itself, or on a key that itself needs another.GrantConsistencyCheckerreports a stored key without any of its requirements as a startup failure.GET /api/v1/permissionsanswersrequiresOneOf.V43__price_publish_needs_pricing_access.sqlremoves the stored grants of that shape, each with arole.grant_removedaudit row asSYSTEM(bug-167-migration). The local seed’sroles_edit_seamrole gainsbuying_standard.view. - Frontend. The role editor’s Publish weekly prices tick is blocked, like All records, while Pricing is None, with “Choose View or Manage on Pricing first” in its name and tooltip, and choosing None clears it. Tests that granted
price.publishalone grantbuying_standard.viewbeside it; the two role-editor journeys now assert the blocked tick in the browser. - Found in review (backend-reviewer): a
price.publish-only grant could still reach three API-only actions that checkprice.publishalone — the round withdrawal (POST /api/v1/auction-sales/{saleId}/withdraw), and a price’s withdrawal and correction (POST /buying-standard-prices/{id}/withdraw,/correct). The migration’s removal takes that away too; the CRM never offered it to such a holder. Tharinda confirmed on 2026-10-14 that removing that API-only access is fine. Lesson: none — one-off: a catalog key that is usable only beside another; the requirement is now declared in the catalog itself (requiresOneOf), so a later key of that kind gets the rule by declaring it.
BUG-166 — The role editor shows the role’s internal key (“Key: administrator”) under its name
Affects: F5-d (forge-et-fe src/features/roles/RoleEditorForm.tsx, the roles-editor-key line under the h1) · Found in: manual test pass, added to the test sheet 2026-10-09 · Severity: low · Phase: 1
Repro: open any role, for example Administrator, in the role editor.
Expected: the editor shows what a person edits: the role’s name and description and its permissions.
Actual: “Key: administrator” sits under the role’s name. The key is the role’s stable identifier, generated from the name the role was created with (lower case, other characters collapsed to _) and never changed by a rename (AD #34, F5-k FR-11). It is for the system, which refers to a role by it; a person has no use for it, and after a rename it no longer matches the name, which reads as an error.
Decision changed: F5-d FR-15 (“shows the role’s name, description and key (read-only)”) and AC-18 (“the key read-only”). Nanndish decided on 2026-10-09 to remove it from the screen; the API keeps returning key.
Fix: remove the roles-editor-key paragraph and its style from the role editor. Revise F5-d FR-15 and AC-18 in a ## Revisions entry. e2e/browser/roles/create-role.spec.ts, which checks the generated key through this line, reads it from the API response instead (GET /api/v1/roles/{id}); editor.spec.ts drops its key assertion.
Regression test: RoleEditorRoute.test.tsx — the loaded editor shows the role’s name and no text matching the role’s key or “Key:”; the browser row in editor.spec.ts asserts roles-editor-key is absent.
Resolved: 2026-10-12 — fixed in forge-et-fe#231, with BUG-158 and BUG-161. RoleEditorForm.tsx drops the roles-editor-key line and its .key style; the API still returns key. create-role.spec.ts checks the generated key on the create answer, and editor.spec.ts asserts the line is absent. F5-d spec Rev 2 revises FR-15 and AC-18. Regression test: RoleEditorRoute.test.tsx fails on the old editor and passes with the fix.
Lesson: none — one-off (a field the spec asked for that the tester found had no use on screen).
BUG-165 — Only the Pricing screen has a description under its title
Affects: F26-b (forge-et-fe src/features/pricing/PricingRoute.tsx:17-24), and the screens without one: F5-c (Users), F5-d (Roles), F1-b (Accounts), F11-b (Leads), F20-b (Email) · Found in: manual test pass, 2026-10-08 · Severity: low · Phase: 1
Repro: compare the screen headers.
Expected: one convention for every screen.
Actual: Pricing has a one-line description; no other screen has one. design-system has no screen-header rule; only F26-b’s plan asked for a description; the client-approved prototype has one under most screen titles.
Decision changed: none written — the gap. Nanndish decided on 2026-10-08: a description on every list screen.
Fix: a design-system §6 Screen header row (and Revisions entry): title at type/title, an optional one-line description at type/secondary in text/secondary beneath it, required on list screens; built as BUG-163’s PageHeader description. The wording for each screen follows the prototype where it has one (Accounts, Inbox, Users), and is agreed in the fix PR for the rest; Leads keeps its count beside the title as well (BUG-164).
Regression test: PageHeader.test.tsx renders the description linked to the heading; a component test per list screen asserts its description; the size matrix shows no header overflow at 320px.
Fixed as: PageHeader’s description, at type/secondary in text/secondary under the title and linked by aria-describedby, on every list screen (design-system Rev 76). Wording: Accounts and Email from the prototype; Users, Activities and Leads adapted from it; Roles, Email templates and the tag pages from their Settings card text; Pricing’s restored (94c0ce99 had removed it); Contacts proposed. The lead approved every line as written on 2026-10-29 (the table in forge-et-fe#356). The Leads count is no longer beside the title (fe328889 made it a visually hidden status line on 2026-10-23), so Leads has a description and no visible count. Tested by PageHeader.test.tsx and e2e/browser/shell/screen-header.spec.ts (each list screen’s description has text at type/secondary); the size matrix runs in the PR’s browser gate.
Resolved: forge-et-fe#356, 2026-10-28 (Tharinda).
Lesson: L-242 (new) — the same cause as BUG-163.
BUG-164 — The Leads title, count and New lead button sit on a baseline, so the count looks low
Affects: F11-b (forge-et-fe src/features/leads/LeadListPane.tsx:153-164, LeadListPane.module.css:30-60) · Found in: manual test pass, 2026-10-08 · Severity: medium · Phase: 1
Repro: open Leads.
Expected: the count reads as aligned with the title.
Actual: .titleRow aligns the 22px title, the 12.5px count and the 36px New lead button on a common baseline, so the count sits low against the title’s middle and the button sets the row’s height.
Decision changed: the F11-b WI-1.2 plan (“beside it on the baseline the count”), following the prototype. Nanndish decided on 2026-10-08 to centre it.
Fix: built on BUG-163’s PageHeader: the title and count centred on one line, and the action centred against them (align-items: center; no nudges, design-system §5).
Regression test: a browser check at 1280 and 768px that the vertical centres of #leads-heading, [data-testid=leads-count] and the New lead button are within 1px.
Fixed as: PageHeader’s title row is align-items: center, so the Leads title and New lead share one vertical centre, as every screen’s title and actions now do. The count beside the title is gone since fe328889 (2026-10-23, a visually hidden status line), so the check compares the heading with New lead only: e2e/browser/shell/screen-header.spec.ts found the centres 1.5px apart on the old code at both widths, and within 1px with the fix.
Resolved: forge-et-fe#356, 2026-10-28 (Tharinda).
Lesson: L-242 (new) — the same cause as BUG-163.
BUG-163 — Every screen title except Leads renders the browser’s default h1, not type/title
Affects: F5-c (forge-et-fe src/features/users/UsersScreen.tsx:236, UserEditorForm.tsx:419), F5-d (src/features/roles/RolesRoute.tsx:66, RoleEditorForm.tsx:387), F1-b (src/features/accounts/AccountListScreen.tsx:449, AccountFormPage.tsx:798, AccountFormRoute.tsx:429), F26-b (src/features/pricing/PricingRoute.tsx:19), F20-a (src/features/email/MailboxRoute.tsx:122), F20-b (src/features/email/InboxToolbar.tsx:60) · Found in: manual test pass, 2026-10-08 (reported as “the Leads heading does not match the other pages”) · Severity: medium · Phase: 1
Repro: compare the screen titles on Leads and on Users, Roles, Accounts, Pricing, Email and Mailbox.
Expected: design-system §3: type/title (22px / 600) for screen headers.
Actual: Leads is the one screen that is right. Nothing styles h1 globally (src/ui/styles/base.css styles only *, body and :focus-visible), so every other title is the browser default — about 26px bold, or smaller inside a <section> in some browsers — and Email, Mailbox and the editors also keep the browser’s top and bottom margins. The comments in RolesRoute.module.css:3, MailboxRoute.module.css:4 and InboxToolbar.module.css:3 say the heading’s type comes from base.css; it does not.
Fix: a src/ui PageHeader (title, forwarded heading ref, optional back link, description and actions; the h1 at type/title and text/heading, margin: 0, tabIndex={-1}), classified in src/test/interactionMatrix.tsx, and every screen moved onto it with its heading text, ids and refs unchanged (the browser suites and the focus-restoring hooks anchor on them). Not in base.css, which must stay byte-identical with forge-et-pwa. Correct the three comments. BUG-164 and BUG-165 build on it.
Regression test: a size-matrix assertion over every e2e/pageStates.ts state that main h1’s computed font size and weight equal type/title’s and its margins are 0; a static scan test that fails on an <h1 in src/features/** outside PageHeader.
Fixed as: src/ui’s PageHeader (back link, title, actions, optional description) and PageTitle (the h1 at type/title in text/heading, margin: 0, tabIndex={-1}, forwarding its ref), both classified in src/test/interactionMatrix.tsx. The list screens, Settings and its tag pages, Mailbox, Pricing import, the user and role views and the role editor use PageHeader; headers that carry more (the account page, the contact record, the user editor, the template view and editor, offboarding, the account form) keep their layout and use PageTitle for the h1. The message view’s subject heading, h1 or h2 by context, takes type/title in its module. The sign-in card and the src/app error surfaces keep their own heading styles. Heading text, ids and refs are unchanged, and the comments that credited base.css are corrected. Leads was not right either: its h1 measured 26px. The size-matrix assertion was narrowed to a fast spec over the list screens (e2e/browser/shell/screen-header.spec.ts, which failed on the old code) plus the scan, src/app/pageTitleOnly.test.ts (29 files listed before the move, none after).
Merged with forge-et-fe#355 (2026-10-29): design-system Rev 74 gave every record page PageHeading (a breadcrumb, and a back arrow beside the title) and set every record page’s title at twice type/body, 26px, bold. When the two met, the lead chose Rev 74’s size for every screen title over type/title. PageTitle now draws that size, record pages keep PageHeading with PageTitle as their h1, list screens keep PageHeader, and a list below a section (Users, Roles, Email templates, the tag pages, Pricing import) takes PageHeading’s breadcrumb through PageHeader’s breadcrumb, with the description lined up under the title. screen-header.spec.ts measures the new size.
Resolved: forge-et-fe#356, 2026-10-28 (Tharinda).
Lesson: L-242 (new) — a screen element every screen has, styled by each screen, drifts until one shared component and a scan test own it; BUG-164, BUG-165 and BUG-168 are the same cause.
BUG-162 — Lead text fields have no length limit anywhere, front or back
Affects: F11-a (forge-et-be V26__lead.sql:60 lead_source_detail TEXT, lead/LeadFieldRules.java, lead/web/LeadCreateRequest.java:61, LeadPatchRequest.java:65), F11-b (forge-et-fe src/features/leads/leadForm.ts:249-252, LeadFormDialog.tsx:416-434, 619) · Found in: manual test pass, 2026-10-08 · Severity: medium · Phase: 1
Repro: open New lead, paste several thousand characters into Lead source detail, save.
Expected: the form refuses a value over a stated limit on blur, with a message that names the fix, and the server refuses it with 400 VALIDATION_FAILED naming the field (test-strategy.md Always-Test: the server validation is the one that must exist).
Actual: the value saves. No lead text field has a limit in the column, the DTOs, LeadFieldRules or the form; the only length limits in the system are the role name and description (RoleFieldRules, 80/500).
Decision changed: F11-a and F11-b set no length rule. Nanndish decided on 2026-10-08 to add one.
Fix: a trimmed-length rule per lead text field in LeadFieldRules, which REST, intake, capture and migration all go through (as RoleFieldRules does, rather than @Size), and the same limits in validateLeadField with “Shorten the lead source detail to N characters”. The value of N for each field is chosen with the developer at fix time, after checking the longest value the F4-b migration writes (migration/normalise/TradeFairs.java:172 joins several fair names) and the capture app’s intake. A database CHECK is a schema change and is asked first. Update the F11-a field table in the same harness PR.
Regression test: LeadFieldRulesSpec rows at N and N+1; a controller IT where POST and PATCH with N+1 characters answer 400 naming the field; leadForm.test.ts and LeadFormDialog.test.tsx for the blur message, a blocked submit with focus on the field, and a server 400 shown on the field.
Resolved: 2026-10-12 — fixed in forge-et-be#242 (merged as a7fb2ee6; the rule and CHECKs in 8867dbc7, review fixes a876a8d8 and f7505af1, the migration renumbered to V41__lead_text_length.sql in afc5932f because forge-et-be#223 took V40) and forge-et-fe#232 (merged as 368472b1; the form in 1ead296, review fixes 62c25dc). Every lead text field has a limit, counted in code points after trimming the six ASCII white-space characters, exactly as each column’s named CHECK measures it: company name, contact name and job title 200, email 254, phone 50, website and lead source detail 500, country 100, inquiry 20,000. An over-limit value is 400 VALIDATION_FAILED with OUT_OF_RANGE on the field, and the form says “Shorten the beb401a (6,981 tests, 0 failures); the frontend browser gate on a32637c (3,175 passed, 5 skipped) and test record (5,353 passed); an approve review record on each PR (backend-reviewer, frontend-reviewer).
Lesson: none — a one-off. How the Pipedrive import treats an over-limit lead text is owed to whoever specs F4-b (.forge/owed.md).
BUG-161 — The Pricing client filter is blank when no client is chosen
Affects: F26-b (forge-et-fe src/features/pricing/StandardsToolbar.tsx:149-162) · Found in: manual test pass, 2026-10-08 · Severity: low · Phase: 1
Repro: open Pricing with no ?accountId.
Expected: the empty Client filter says what it shows.
Actual: the combobox is blank; the book lists every client’s standards.
Decision changed: none written — F26-b FR-U1 and AC-23 do not say what an empty filter shows. “Choose a client”, as the test sheet suggests, would wrongly hint that a choice is required, because empty means every client.
Fix: placeholder="All clients" on the Client Combobox. It states the filter’s value, not a format example, so design-system §5 Form fields gains a line allowing a filter’s empty-state value as its placeholder (a Revisions entry).
Regression test: PricingRoute.test.tsx — with no ?accountId the pricing-client-filter input shows “All clients”; choosing a client shows its name; clearing it shows “All clients” again and drops accountId from the URL.
Resolved: 2026-10-12 — fixed in forge-et-fe#231, with BUG-158 and BUG-166. The Client Combobox takes emptyValue="All clients", a new option that shows the value as the placeholder in text/secondary, not the faint colour a format example takes: the review found a faint placeholder (about 2.8 : 1) fails §2 and WCAG 1.4.3 for text a user must read, and the lead chose the darker colour. Design-system Rev 51 adds the exception to §5 Form fields. Regression tests: PricingRoute.test.tsx as above (fails without the fix on the missing placeholder) and Combobox.test.tsx for emptyValue and its style.
Lesson: none — one-off.
BUG-160 — List-screen search hints are inconsistent: placeholders on two screens, helper text on two, nothing on Email
Affects: F5-c (forge-et-fe src/features/users/UsersScreen.tsx:244-252), F1-b (src/features/accounts/AccountsFilters.tsx:97-105), F20-b (src/features/email/InboxToolbar.tsx:88-95, InboxToolbar.module.css:33) · Found in: manual test pass, 2026-10-08 · Severity: low · Phase: 1
Repro: compare the search fields on Users, Accounts, Leads, Pricing and Email.
Expected: design-system §5 Form fields: the label is above the control, placeholders show format examples only, and a hint is helper text under the control. Every list search is capped at layout/search-max.
Actual: Users (“Name or email”) and Accounts (“Account name”) put the hint in a placeholder; Leads and Pricing use helper text; Email’s Filter mail has neither and is not width-capped (nor is Pricing’s).
Fix: move the Users and Accounts hints into helper, give Email the helper “Subject or sender” (F20-b FR-9), and cap every list search at --et-layout-search-max. This is the opposite of the test sheet’s “add placeholders”, which would need a design-system revision. Sequence after BUG-118 merges: it edits UsersScreen.module.css, AccountListScreen.module.css and AccountStandardsPanel.tsx.
Correction (2026-10-26): “BUG-118” here does not name a filed bug: main’s BUG-118 is the Combobox’s Show suggestions trigger, fixed on 2026-10-09 in forge-et-fe#200. What this fix now has to wait for is forge-et-fe#337, which rebuilds the Users and Accounts list screens, and design-system Rev 70 (forge-et-harness#865), whose placeholder rule (Enter and the field’s name) differs from the Fix above; re-check both once they merge.
Regression test: a unit check that every Input size="search" under src/features sits in a Field with helper and passes no placeholder; a browser check that each list route’s search has non-empty helper text through aria-describedby and is no wider than layout/search-max.
Resolved: 2026-11-12 — the screens were already consistent on main, under a rule that reversed this entry’s Fix: line. Design-system Rev 41 and Rev 72 name a data-bar search box by a search glyph and a placeholder that repeats its aria-label, with no label above it and no helper text; Rev 83/84 set its width to nine space/8 steps (288px), and on Email to the list column under it; Rev 107 set the casing: the aria-label in sentence case (Search users), the placeholder the same words in Title Case (Search Users, through titleCase). forge-et-fe#201 (Users), #262 (Accounts, Pricing), #338 (Leads) and #343 (Email, Contacts) built the pattern. forge-et-fe#436 adds the regression test, src/app/listSearchNaming.test.ts, and recases the three aria-labels still in Title Case: Leads (Search leads), the Pricing round (Search standard, client or grade) and the Standards book (Search standards). For every search box under src/features, src/ui and src/app (a type="search" or role="searchbox" element, the older Input with size="search", and the FilterSearch and PickList wrappers), the test fails a box inside a Field, Label or label, one a label’s htmlFor names, one named by aria-labelledby, one without an aria-label or placeholder, one whose placeholder is not its aria-label in Title Case, one whose name does not start "Search " or is not in sentence case, and one that takes its name as a prop outside a wrapper’s own file; it also requires a search box on each of the eight list screens. It failed when Users’ placeholder was set back to “Name or email”, Email’s search was wrapped in <Field label="Filter mail">, Contacts’ SEARCH_LABEL was set to “Name or email” and Accounts’ search was given a <Label htmlFor>, and, before the recase, on exactly the three Title Case names. The width is checked by hand only (below): a unit test cannot measure layout, and the nightly size matrix checks overflow and target size, not a search box’s width. Checked in the browser against the local backend at 1024×768, before the recase: /accounts, /contacts, /leads, /settings/users, /pricing and /settings/email-templates each show a 288×40 search with no <label>; /email had no local mailbox, so its search is covered by the unit test only. The recase changed three aria-label strings and nothing on screen, so it was checked by the component tests, which find each box by its exact accessible name, not again in the browser. /forge-review-pr with frontend-reviewer approved forge-et-fe#436 on cf6a40d2 after three delta rounds. The Standards book search keeps its screen-reader description “Number, name or grade”, recorded in its component as deliberate; a description is not a name, so it is outside this bug.
Lesson: none — one-off: the screens were built before the data bar had one pattern, and the design-system revisions and shadcn moves that set the pattern also fixed them; the new test stops them drifting apart again.
BUG-159 — Discard on the user and role editors gives no sign that the changes were thrown away
Affects: F5-c (forge-et-fe src/features/users/UserEditorForm.tsx:377-382, 620-627), F5-d (src/features/roles/RoleEditorForm.tsx:366-373, 395-401) · Found in: manual test pass, 2026-10-08 · Severity: low · Phase: 1
Repro: on a user or a role, change a field and press Discard.
Expected: the screen says the changes were discarded, as Save (“User saved”) and Reload (“User reloaded”) already do (WCAG 2.1 SC 4.1.3, status messages).
Actual: the fields reset silently. On Users the button then becomes disabled under the pointer with focus left on it; on Roles Discard is always enabled, so pressing it with nothing changed does nothing visible.
Decision changed: F5-d FR-16 / AC-21 and the F5-c WI-2.2 plan say only that Discard restores the loaded values. Nanndish decided on 2026-10-08 to add the feedback.
Fix: a “Changes discarded” confirmation in each editor’s existing live region, focus moved to the h1 as Reload does (moveFocus('heading')), and Roles’ Discard disabled while nothing is unsaved, as Users’ already is. No confirm dialog: the leave guard already covers navigation, and the test sheet asked for feedback, not a confirmation.
Regression test: in UserEditorRoute.test.tsx and RoleEditorRoute.test.tsx — edit, Discard: the fields are restored, the live region reads “Changes discarded”, focus is on the h1, no write was sent; Roles’ Discard is aria-disabled before any edit.
Resolved: 2026-10-12 — fixed in forge-et-fe#236, with BUG-153 and BUG-156. Both editors set “Changes discarded” in their always-rendered aria-live="polite" region (users-editor-discarded, roles-editor-discarded), cleared by the next edit, and move focus to the h1 with moveFocus('heading'). Roles’ Discard is disabled={!unsaved || save.isPending}, as Users’ is; the save-in-flight half was not in the Fix: line and has its own unit row. Regression tests: the rows above (three failed before the fix) plus Discard is unavailable while a save is in flight. Checked in the browser on a role: after Discard the line shows, focus is on the h1, the name is restored and Discard is aria-disabled="true".
Lesson: L-133, another instance — Discard disabled itself under focus, a button that turns into a different state, with no focus target named and no announcement.
BUG-158 — The deactivation refusal names the user as “User ”, not by name
Affects: F5-c (forge-et-fe src/features/users/activationErrors.ts:37-42, 85-108, UserActivationCard.tsx:172-185; the text is from forge-et-be UserAdminService.java:149-152, 296-310) · Found in: manual test pass, 2026-10-08 · Severity: low · Phase: 1
Repro: deactivate a user who owns accounts or leads (or has direct reports).
Expected: the refusal names the person shown on the screen.
Actual: the panel reads, for example, “Nimal Wickrama was not deactivated / User 2 owns 1 account and 5 leads. Reassign them before deactivating this user.” The internal user id shows, against coding-standards §2 (no internal ids in API responses). The “Reference …” line under it is the request’s trace id, which coding-standards §3.2 requires on error text, and it stays.
Decision changed: the F5-c Wave-3 contract (.forge/plans/F5-c/F5-c-Wave-3-contract.md:56-62) fixed the detail text with the id, so that no name reaches the WARN log, and the client shows detail as the reason. Nanndish decided on 2026-10-08 that the screen words the reason itself.
Fix: frontend only. For all four refusal codes (USER_OWNS_RECORDS, USER_HAS_DIRECT_REPORTS, LAST_ADMIN_PROTECTED, SELF_DEACTIVATION_REFUSED), activationErrors.ts words the reason from the code and the display name it already receives (“Nimal Wickrama owns accounts or leads”), followed by the existing next-step line, deciding on code, never parsing or showing detail. The backend text and the log stay as they are. Keeping the counts would need structured members on the problem body, a contract change not made here. Record the change as a note on the Wave-3 contract.
Regression test: activationErrors.test.ts / UserActivationCard.test.tsx — given a USER_OWNS_RECORDS problem whose detail is “User 7 owns …”, the panel contains the display name and does not match /User \d+/; the same for the other three codes; e2e/browser/users/activation.spec.ts asserts the named reason on the live refusal and no User <id>.
Resolved: 2026-10-12 — fixed in forge-et-fe#231, with BUG-161 and BUG-166. activationErrors.ts words all four refusals from the code and the display name (“Nimal Wickrama owns accounts or leads”, “… has direct reports”, “… is the last active user who can manage users and roles”; the self refusal keeps its sentence), always followed by the next-step line, and never reads detail. The backend text and log are unchanged; the Wave-3 contract carries a note. Regression tests: activationErrors.test.ts, UserActivationCard.test.tsx and OffboardDeactivateSection.test.tsx (8 assertions fail on the old mapping); activation.spec.ts and offboarding-deactivation.spec.ts assert the named reason and no User <id> on the live refusal.
Lesson: none — one-off (the contract chose the id to keep names out of the log; the client now words the reason itself).
BUG-157 — Destructive triggers (Deactivate, Delete role, Disconnect) look like every other secondary button
Affects: FS-3 (forge-et-fe src/ui/components/Button/), F5-c (src/features/users/UserActivationCard.tsx:190-200), F5-d (src/features/roles/DeleteRoleAction.tsx:101-110), F20-a (src/features/email/ConnectedPanel.tsx:102-109), F1-b (src/features/accounts/AccountsTable.tsx:170-180, Archive), F26-b (src/features/pricing/StandardsTable.tsx:53, the Archive menu item) · Found in: manual test pass, 2026-10-08 · Severity: medium · Phase: 1
Repro: open a user’s Access card, a role, the Mailbox page, the Accounts list and the Standards book’s row menu.
Expected: a destructive action is marked as one before it is pressed.
Actual: each trigger is Button secondary (or a plain menu item); only the confirm inside the dialog is Button/danger.
Decision changed: F5-d WI-2.2 plan Decision 6 (and F5-c WI-3.2, which copied it): solid red only on the confirm, so the destructive action is not the loudest control beside Save. Nanndish decided on 2026-10-08 to mark the triggers, keeping that reasoning: a new outline style, quieter than the solid confirm.
Fix: a danger-outline Button family — surface/card fill, tone/danger border and text, the secondary family’s sizes and states — added to BUTTON_VARIANTS and Button.module.css, with a design-system §5 Button row and a Revisions entry; the Deactivate, Delete role, Disconnect and Archive triggers move to it, and the Standards book’s Archive menu item takes tone/danger text. The confirms keep solid danger. tone/danger text is 5.9:1 on white (design-system Accent & tones).
Regression test: Button.test.tsx and the interaction-matrix suite cover the new family; src/app/disabledFocusRing.test.ts lists it; DeleteRoleAction.test.tsx, UserActivationCard.test.tsx and MailboxRoute.test.tsx assert the trigger carries it; the size matrix’s axe and contrast checks pass.
Fixed as (retargeted, 2026-10-28): most of these screens had moved to shadcn/ui since this entry was written. The shadcn Button gains a destructive-outline variant (card fill, destructive border and text, a 10% wash on hover) and the older Button a danger-outline family; Deactivate on the user editor and Withdraw this round take destructive-outline, Delete role, Disconnect and Deactivate on offboarding (still on the older Button) take danger-outline, and the Accounts, Standards book and weekly-prices row menus’ Archive, Remove and Withdraw take the menu item’s destructive variant. Confirms keep solid danger. Three of those (offboarding Deactivate, Withdraw this round, the weekly-prices Remove and Withdraw) go beyond the triggers named above; they are the same kind on the same screens. Contrast measured in every theme: text and border 5.94:1 on the card, at least 4.69:1 under the hover and pressed washes, pinned by src/app/destructiveTriggerContrast.test.ts; e2e/browser/shell/destructive-triggers.spec.ts checks the rendered colours and axe on three screens. Design-system Rev 75 adds the row; F20-a Rev 9 revises AC-19. The other destructive triggers are listed in .forge/owed.md.
Resolved: forge-et-fe#356, 2026-10-28 (Tharinda).
Lesson: none — one-off: a design decision changed after the screens were built (F5-d WI-2.2 plan Decision 6 reversed on 2026-10-08), not a recurring defect.
BUG-156 — Modal titles have about twice as much space below them as above
Affects: FS-3 (forge-et-fe src/ui/components/Modal/Modal.module.css:63-80, Modal.tsx:123-131), seen first in F5-c’s Create user · Found in: manual test pass, 2026-10-08 · Severity: low · Phase: 1
Repro: open Create user (or any modal).
Expected: the title sits centred in its header row, with even space above and below.
Actual: the 36px ✕ button sets the header row’s height and align-items: flex-start pins the 22px title to the top, leaving about 33px under the title against 18px above it.
Fix: in Modal.module.css, give .title min-block-size: var(--et-layout-control-h-compact) with display: flex; align-items: center, so a one-line title centres on the ✕ and a wrapped title still starts at the top beside it. No margin or transform nudge (design-system §5, layout-first alignment).
Regression test: a browser row with Create user open: the title’s vertical centre equals the ✕ button’s (±1px), and the space from the title to the header’s bottom edge equals the space above it (±1px).
Resolved: 2026-10-12 — fixed in forge-et-fe#236, with BUG-153 and BUG-159. .title in Modal.module.css takes min-block-size: var(--et-layout-control-h-compact) (the ✕'s height), display: flex and align-items: center; no margin or transform. Every Modal title is a string, so the heading holds one text run, and a wrapped title still starts at the top beside the ✕. Regression test: e2e/browser/users/see-and-add.spec.ts › the Create user dialog header (BUG-156) measures the title’s text with a Range after the opening animation; before the fix its centre was 7 px above the ✕'s. Checked in the browser on Create user: title centre 114.95 against the ✕'s 115.25, 25.45 px above the title and 26.05 px below; New role and Delete role sit level too.
Lesson: none — one-off: the Modal header is the only header that pairs a one-line title with a taller fixed control under align-items: flex-start, and no other Lesson: line or lesson has this cause (BUG-010 was a translateY nudge; BUG-164 is a baseline alignment, still open). The layout-first rule in design-system §5 and forge-et-fe/CLAUDE.md already covers the fix.
BUG-155 — The Email and Mailbox screens: unbounded filter and card, and the retired row-hover tint
Affects: F20-b (forge-et-fe src/features/email/InboxToolbar.module.css:33, InboxList.module.css:23, 37), F20-a (src/features/email/MailboxRoute.module.css:13, ConnectedPanel.module.css:26) · Found in: manual test pass, 2026-10-08 (“not organized neatly”) · Severity: medium · Phase: 1
Repro: open /email and /email/mailbox on a desktop screen.
Expected: the filter capped at layout/search-max as on Users and Accounts; message rows hover in surface/card-warm (design-system §5 Table row; BUG-118 retires surface/hover as a row tint); the Mailbox page capped at the drill-in column width the other editors use (calc(2 * var(--et-layout-list-pane-width))), so Connected and Disconnect are not at opposite edges of the canvas.
Actual: Filter mail spans the canvas; rows hover in surface/hover; the Mailbox card spans the canvas (the test sheet’s screenshot shows Connected at the left edge and Disconnect at the right). The headings’ default size and margins are BUG-163.
Fix: max-inline-size: var(--et-layout-search-max) on .filter; the row hover token to --et-surface-card-warm; the Mailbox page a column capped and centred as BUG-154 sets for detail pages. Not in scope: the prototype’s two-pane inbox, which F20-b FR-10 / D-6 replaced with a flat list and F20-f builds, and the pager’s page-size select (F20-b WI-1.2 plan Decision 11).
Regression test: a browser spec on /email and /email/mailbox (connected) at 1440px: the filter is no wider than layout/search-max, a hovered row’s background is surface/card-warm, and the Mailbox column is no wider than twice layout/list-pane-width and centred.
Resolved: 2026-10-14 — fixed in forge-et-fe#257 (9ffe37c; review fixes 7f43358).
-
The fix.
InboxToolbar.module.css.filteris capped at--et-layout-search-max. That matches the Accounts list search and the top bar’s search; the Users search has its own width, so the Expected line above overstates it.InboxList.module.css.row:hoveris--et-surface-card-warm, as design-system §5 Table gives for a row.- The Mailbox page’s
.screentakes the BUG-154 detail-page column (capped and centred), so Connected and Disconnect sit inside it. - The Expected line’s reason, that BUG-118 retires
surface/hover, is not what BUG-118 is about. Design-system §1’ssurface/hoverrow now names the table-row and inbox-row exception (Rev 56).
-
Regression tests. Two rows in
e2e/browser/mailbox/inbox.spec.tsat 1440px, which run serially with that file’s other tests on its signed-in user:- the filter is no wider than the token, and a hovered row takes
surface/card-warm(with a pinned two-message list); - the connected Mailbox page is capped and centred. It is measured against
<main>'s client box bye2e/browser/shell/centredColumn.tsand found through a newmailbox-screenhook.
Both fail on
main’s CSS and pass with the fix. - the filter is no wider than the token, and a hovered row takes
-
Review.
frontend-reviewerfound two Importants, both fixed in7f43358: the centring measurement counted a layout scrollbar, and a comment named the wrong searches. -
Superseded for the Email inbox’s search by design-system Rev 84 (2026-11-03): with F20-i’s folder column, the inbox search spans the folder column, the gap and the list, wider than
layout/search-max. F20-i wave 2 rewrites the filter row ofe2e/browser/mailbox/inbox.spec.tsto check that rule. The hover row and the Mailbox page’s cap are unchanged, and every other list screen keepslayout/search-max.
Lesson: none — one-off: the screens were built before the design system had a detail-page rule, and the filter’s cap was missed in one screen. Rev 56 adds the rule, and design-system adherence is a frontend-reviewer dimension already.
BUG-154 — The user and account edit pages sit against the left edge of the canvas
Affects: F5-c (forge-et-fe src/features/users/UserEditor.module.css:47-50), F5-e (src/features/users/EffectivePermissionsSection.module.css:9-15), F1-b (src/features/accounts/AccountFormRoute.module.css:25-30, AccountFormPage.module.css:45-50) · Found in: manual test pass, 2026-10-08 · Severity: low · Phase: 1
Repro: open a user, or an account’s edit page, at 1280px or wider.
Expected: a capped detail column is centred in the canvas.
Actual: the column is capped at twice layout/list-pane-width (720px) and left-aligned, so the right of the canvas is empty.
Decision changed: the F5-c WI-2.2 plan capped the form column with a layout/* token and said nothing on placement; design-system §4 and §6 set no detail-page rule. Nanndish decided on 2026-10-08 to centre it.
Fix: a design-system §6 rule (and Revisions entry): a detail or edit page’s content is one column capped at twice layout/list-pane-width (the existing cap; a new token would need a forge-et-pwa change too, because tokens.css is shared byte for byte), centred with margin-inline: auto, its header and back link inside the same column. Apply it to the user editor (with the effective-permissions section) and the account edit page; the role editor keeps its full-width matrix.
Regression test: a browser row at 1280 and 1440px on /users/:id and /accounts/:id/edit: the column’s left and right gaps inside <main> are equal (±1px) and the back link’s left edge matches the column’s.
Resolved: 2026-10-14 — fixed in forge-et-fe#257 (45ea7d8; review fixes 7f43358).
- The rule. Design-system §6 gains the Detail page rule (Rev 56): an edit or settings page whose content is one column of forms or cards is capped at twice
layout/list-pane-width, centred in the canvas, with its back link and header inside the column. - Where it applies. The
.screenofUserEditor.module.css(the user editor, its Access card and effective permissions, and its loading, error and not-found frames),AccountFormPage.module.css(New and Edit) andAccountFormRoute.module.css(the edit route’s read states) is capped and centred withmargin-inline: auto. Noinline-sizeis set: each route’s section sits directly in<main>, which is a block container, so the auto width fills and the auto margins centre. - No change at the tablet floor or at 320px. At 768px and below the cap is about as wide as the content area, so centring changes nothing there.
- Regression tests.
e2e/browser/shell/detail-page-column.spec.tschecks/users/:idand/accounts/:id/editat 1280 and 1440px throughcentredColumn.ts: equal gaps within 1px inside<main>'s client box, the column within the cap, and the back link at the column’s left edge. Both fail onmain’s CSS and pass with the fix. - Visual check. Before and after screenshots of the user editor at 1440px.
- Pages the rule does not reach yet. The offboarding sub-view, the account page and the message view are still full-width and left-aligned. That is BUG-137.
Lesson: none — one-off: the plans capped the column and said nothing about placement, because the design system had no detail-page rule. Rev 56 adds it.
BUG-153 — The role editor (and any long screen) lets the whole frame scroll below its content
Affects: F5-b (forge-et-fe src/app/AppLayout.tsx:135, 162-172), F5-d (where it shows most: src/features/roles/EntityChoice.tsx:41, AllRecordsTick.tsx:87, DeleteRoleAction.tsx:132) · Found in: manual test pass, 2026-10-08 · Severity: medium · Phase: 1
Repro: open a role in the role editor and scroll to the bottom, then keep scrolling.
Expected: design-system §6 App frame: the frame stays fixed and only the canvas scrolls.
Actual: the whole page scrolls on past the content, taking the top bar with it, over empty canvas.
Fix: position: relative on <main>. The visually hidden legends and reasons are position: absolute, and with no positioned ancestor they are placed against the page, outside <main>'s overflow: auto, so each one extends the page’s own scroll height. The role editor has one hidden legend per entity, which is why it shows there first; Users, Accounts, Pricing, Email and Leads use the same pattern.
Regression test: a browser row on a new role (every entity at None), /users and a user’s page: document.documentElement.scrollHeight <= window.innerHeight and window.scrollY stays 0 after scrolling <main> to its end. No vertical page-overflow check exists in e2e/ today.
Resolved: 2026-10-12 — fixed in forge-et-fe#236, with BUG-156 and BUG-159. AppLayout.tsx gives <main> position: relative, so the hidden legends and reasons are placed inside the canvas. Regression test: e2e/browser/shell/frame.spec.ts › BUG-153, at 1280 × 600, asserts on each screen that <main> is longer than the window, the page is no taller than the window, <main> wheels to its end and window.scrollTo leaves the page at scrollY 0; the user-page row waits for Effective permissions, where most of its hidden text is. With the fix reverted the role editor (page 1132 > 600) and the user page (973 > 600) fail; the Users list passes either way (its hidden text sits near the top), so that row guards the screen against a later regression. Checked in the browser at 1024 × 768 on a new role: before, the page was 1165 px tall and scrolled 397.5 px past the frame; after, 768 px, <main> scrolled 839 px to its end and the page stayed at 0; a user’s page after: 768 px, <main> 1412 px, page at 0. A vertical page-height rule in the shared page check, for every page state at every size, is a decision in .forge/owed.md. Proof for all three bugs: the whole browser gate on 9d7a2ad, 3,179 passed, 5 skipped. The first run failed eight access tests only because the local database’s IT Admin role held an email.manage grant added by hand on 2026-10-07 to connect the CRM01 mailbox; with Tharinda’s agreement it was removed for the run and put back unchanged.
Lesson: L-176, another instance — the same escape at the app frame: an absolutely positioned descendant of an unpositioned scroll region is placed outside it, here making the page taller rather than wider.
BUG-152 — The certification list’s Retry, on the account and lead forms, unmounts itself when pressed, so a retry that fails again drops focus to the heading
Affects: F1-a (the account form), F12 (the lead form’s Tea requirements group) — the shared hook forge-et-fe src/lib/certifications/useCertificationOptions.ts:79-84, rendered by AccountFormPage.tsx and LeadTeaRequirementsGroup.tsx:268 · Found in: frontend-reviewer on forge-et-fe#223 (BUG-143’s fix), finding 3, 2026-10-11 · Severity: low — the read retries and a new Retry appears; keyboard focus moves away and the busy state never shows · Phase: 1
Repro: fail the certification read (GET /api/v1/certifications → 500) on either form, press Retry, and let the retry fail again.
Expected: Retry stays on screen, busy, until the retry settles; when it fails again focus stays on Retry; when it succeeds focus goes to the heading (L-183, coding-standards.md §3.2).
Actual: the list never held data, and TanStack v5’s refetch() returns such a query to pending and clears error (query.ts fetchState). So certifications.isError turns false at the press: Retry and its TraceLine unmount, busy={certifications.isFetching} never shows, and useFocusAfterWrite moves focus to the heading even when the retry fails again.
Fix: hold the failure in the shared hook until the read’s next result, as src/features/sharing/useRetryHold.ts does (matched on errorUpdateCount and errorUpdatedAt), so both forms are fixed at once.
Regression test: for each form, a retry that fails again keeps focus on Retry and shows it busy while in flight; a retry that succeeds moves focus to the heading.
Resolved: 2026-10-12 — fixed in forge-et-fe#223 (6d1e154). useRetryHold moved from src/features/sharing to src/lib/hooks so the shared hook can use it; useCertificationOptions holds the failure from the Retry press until the read’s next result and returns failure and retrying, and both forms render the failed state, the reference and the busy state from them. The account form’s helper line puts the failure before the pending state the refetch causes. Regression tests on both forms (AccountFormRoute.create.test.tsx, LeadTeaRequirementsGroup.test.tsx): a retry that fails again keeps the same Retry, busy while in flight and focused after; both failed with the hold removed. The old account-form test that expected focus to move to the heading after a second failure asserted the defect and was replaced; the existing success-path tests still show focus moving to the heading. The delta review found the busy state narrowed to the held retry, so a list that failed after it had loaded showed Retry never busy; fixed in bc2299f (busy whenever a re-read of the shown failure is in flight), with a test that failed before.
Lesson: L-183, another instance — added to its source list.
BUG-151 — Flaky test: the leads-lead-tags-none-left size-matrix state finds no tag list within 5 seconds at 1280 × 800 in Chromium, in a full browser gate (macOS)
Affects: F11-f (forge-et-fe e2e/browser/size-matrix/arrange.ts:4781, the leads-lead-tags-none-left arrange step) · Found in: the second pnpm e2e:gate on forge-et-fe fix/BUG-126-offboarding-account-lead-fixes at baf72cc (forge-et-fe#223), on a freshly reset isolated database, 2026-10-11 · Severity: low — a flaky browser test · Phase: 1
Repro: macOS, 8 cores, the whole gate at Playwright’s default 4 workers with nothing else running (3,165 passed, 1 failed, 27.0 minutes; load average 13 to 15). After leads-tags-add-trigger was clicked, getByRole('listbox') was not found within 5 seconds in size-1280x800-chromium. Run alone straight afterwards with --repeat-each=3: 12 passed.
Expected: the state passes the shared page check whatever the machine’s load.
Actual: the tag list opened later than the 5-second expect timeout under full-suite load.
Fix: not yet known. Each full run on this head failed a different lead-view test (BUG-149, BUG-150, this one), which points at load rather than any one test: compare L-175 (the same on Windows) and L-152 (run the gate with fewer workers). Reproduce with a throttled CPU (L-162) before changing the test.
Regression test: the size matrix inside the whole gate on macOS, green across repeated runs at the default worker count.
Cause (found 2026-10-12, L-162 probes): two timing faults in the test, neither in the product. (1) Playwright scrolls a control below the fold into view and clicks at once; the browser delivers that scroll event a frame later, after the click has opened the list, and react-aria closes an open list on any scroll around its input (useCloseOnScroll). (2) F5-g’s sharing panel is loaded lazily and arrives after the rest of the lead view; when it grows with the column scrolled, the browser’s scroll anchoring scrolls the column again, which also closes the list. A person scrolls long before clicking, so neither reaches a user. At a 4× CPU throttle, the bare click lost the list in 8 of 10 runs.
Resolved: 2026-10-12 — fixed in forge-et-fe#224 (cd4ee89), test only. lead-detail-support.ts gains settleLeadPanels (waits for the sharing panel to settle) and openAddTagList (settles the panels, brings the trigger fully into view, waits two frames, then clicks), and all seven Add tag trigger clicks in the suite use it. The same probe with the helper lost the list in 0 of 10 runs at the same throttle, and the tag states passed at all seven sizes (58 passed). BUG-123’s new axe test, which waits for the list long enough to see it close, failed 3 of 3 before and passes after.
Lesson: L-101 (a browser test waits for the page to settle, not for an element to appear) — another instance, with two new parts: a click that scrolls fires its scroll event after the click, and a lazily loaded panel shifts the controls below it; added to its source list.
BUG-150 — Flaky test: the leads-lead-match-picker-open size-matrix state finds no match dialog within 5 seconds at 768 × 1024 in WebKit, in a full browser gate (macOS)
Affects: F11-f (forge-et-fe e2e/browser/size-matrix/arrange.ts:1439 openPinnedMatch → e2e/browser/leads/lead-detail-support.ts:1076 openMatchDialog) · Found in: pnpm e2e:gate on forge-et-fe fix/BUG-126-offboarding-account-lead-fixes at baf72cc (forge-et-fe#223), against an isolated backend (et-bug-126-gate), 2026-10-11 · Severity: low — a flaky browser test · Phase: 1
Repro: macOS, the whole gate (3,164 passed, 2 failed, 5 skipped, 32.1 minutes); the backend’s bootRun and the compose stack were the only other load. After leads-customer-change was clicked, getByRole('dialog', { name: 'Change the account link' }) was not found within 5 seconds in size-768x1024-webkit. ET_BACKEND_PORT=18126 pnpm exec playwright test --project=setup --project=size-768x1024-webkit e2e/browser/size-matrix/size-matrix.spec.ts -g leads-lead-match-picker-open --repeat-each=3 straight afterwards: 12 passed.
Expected: the state passes the shared page check whatever the machine’s load.
Actual: the dialog’s appearance took longer than the 5-second expect timeout under full-suite load in WebKit.
Fix: not yet known. Reproduce with a throttled CPU (L-162) before changing the test; compare BUG-129 (WebKit lead states timing out under load on Windows), which may share the cause.
Regression test: the WebKit 768 project inside the whole gate on macOS, green across repeated runs.
Cause (found 2026-10-12): the dialog was never opened, not opened slowly: the failure reads “element(s) not found”. F5-g’s lazily loaded sharing panel arrives after the rest of the lead view and moves every control below it, so the click on Change could land where the button had been (BUG-151’s second fault). Reproduced: leads-lead-match-picker-open in size-768x1024-webkit at 8 workers failed 1 of 24 on the old opener. A Chromium throttle does not reproduce it (10 of 10 passed at 4×), and WebKit cannot be throttled through CDP.
Resolved: 2026-10-12 — fixed in forge-et-fe#224 (cd4ee89), test only: openMatchDialog calls settleLeadPanels before clicking Change. The same run with the fix: 48 of 48 passed.
Lesson: L-101 — another instance; added to its source list with BUG-151.
BUG-149 — Flaky test: the lead view’s AC-22 unreached-account smoke times out waiting for the lead heading in a full browser gate (macOS)
Affects: F1-d (forge-et-fe e2e/browser/accounts/account-lead-link.spec.ts:46, “AC-22 (smoke): a matched account the caller does not reach reads ‘An account you cannot open’, with no link”) · Found in: the same gate run as BUG-150, 2026-10-11 · Severity: low — a flaky browser test · Phase: 1
Repro: macOS, the whole gate as in BUG-150. settleLeadDetail waited 5 seconds for leads-view-heading to read the lead’s name; it still read “Lead”. The test routes the lead’s GET through page.route, fetching the real response and rewriting its account. Run alone straight afterwards with --repeat-each=3 (Chromium): 15 passed.
Expected: the lead view settles within the wait whatever the machine’s load.
Actual: under full-suite load the routed read, or the view’s settling, took longer than 5 seconds.
Fix: not yet known. Reproduce with a throttled CPU (L-162); if the routed route.fetch() is the slow step, the wait or the route is the place to change, not the product.
Regression test: the spec inside the whole gate on macOS, green across repeated runs.
Not reproduced (2026-10-12): --repeat-each=32 --workers=8 in Chromium, 32 of 32 passed; with a 6× CPU throttle set in the test, 16 of 16 passed. Unlike BUG-150 and BUG-151, its wait is on the lead’s own read (the heading reads “Lead” until the routed GET answers), not on a control below the sharing panel, so their fix does not cover it. Kept open with no known cause: if it recurs, record the gate’s load and whether the routed read had answered (a trace), before changing the wait.
Resolved: 2026-10-12 — closed as wontfix (not reproducible), the lead’s decision (Tharinda). Besides the runs above, the spec passed in two full browser gates on macOS the same day (forge-et-fe heads 5a33a58 and 80c2f98, 3,175 passed each, three workers). No change was made. A recurrence is filed as a new bug that links here, with the gate’s load and a trace showing whether the routed GET had answered.
Lesson: L-162 (a failure that appears only under load is reproduced before it is called flaky) — an instance where it was not reproduced, so nothing was changed.
BUG-148 — The capture app’s size matrix still marks a known failure on the whole test, not on the check kinds F40 Rev 5 requires
Affects: F40 (forge-et-pwa e2e/browser/size-matrix/size-matrix.spec.ts:36-38, which calls markKnownFailure(known.bug, known.reason) — test.fail() — for a whole page-state entry; e2e/pageStates.ts:29-33, whose KnownFailure has no check kinds; e2e/README.md describes the old form) · Found in: the acceptance-criteria walk of done features, F40 AC-20, 2026-10-11 · Severity: low — no capture-app page state carries a known failure today, so the gate hides nothing yet · Phase: 1
Repro: compare the two frontends on origin/main: forge-et-fe’s size-matrix.spec.ts (from forge-et-fe#148) builds a KnownFailureHere with kinds and expects exactly those check kinds to fail; forge-et-pwa’s still marks the whole test.
Expected: F40 FR-21 and AC-20 (Rev 5): in the size matrix a mark names the failing check kinds (an axe kind names the rule id), and a second kind failing, or a named kind passing, fails the test — in both frontends.
Actual: in forge-et-pwa, a marked entry passes the gate whatever else fails on that page at that size, because test.fail() accepts any failure.
Fix: port forge-et-fe’s form: kinds on KnownFailure (required in the size matrix), the KnownFailureHere resolution and expectPageClean’s kinds, the unit test that refuses a mark with no kinds, and the e2e/README.md text. Then F40 AC-20 can be ticked.
Regression test: the unit test forge-et-fe has for a size-matrix mark with no kinds, ported, and a page state marked with one kind that fails a second kind.
Resolved: 2026-10-12 — fixed in forge-et-pwa#67 (18aa602). KnownFailureKind and kinds on KnownFailure (e2e/pageStates.ts); the size matrix’s knownFailuresHere, which refuses a mark with no kinds and records a known-failure annotation, and expectPageCheck, which expects exactly the named kinds to fail; the matcher e2e/browser/size-matrix/knownFailureMatch.ts and its unit test src/app/knownFailureMatch.test.ts, byte for byte as in forge-et-fe; and the e2e/README.md text. e2e/support/pageCheck.ts already matched (check-shared-drift.mjs --live, exit 0). The unit test could not run before the port (the matcher did not exist) and passes after (6 passed).
Lesson: L-144 (a rule about “every” case needs a check that lists the cases) — F40 Rev 5’s mark was built in one frontend and only its count copied to the other; added to its source list.
BUG-147 — PipedriveDryRunTimingIT asserts F4-a AC-22’s 30-second bound inside the full build, the same wall-clock assertion BUG-130 removed
Affects: F4-a (forge-et-be src/test/groovy/com/empireteas/migration/PipedriveDryRunTimingIT.groovy:114, elapsedMs < BOUND_MS with BOUND_MS = 30_000L at :59) · Found in: backend-reviewer on forge-et-be#237 (BUG-130’s fix, finding 2), 2026-10-11 · Severity: low — not yet seen failing; F4-a WI-2.1 measured the run at 1,327 ms against the 30,000 ms bound · Phase: 1
Repro: not reproduced. The test times one dry-run exchange over the export-sized synthetic set and fails the build when it takes 30 seconds or more. Its own Javadoc (:45-46) says a slow CI runner could approach the bound without a regression.
Expected: the full suite fails only on what the code does, never on how loaded the machine is (lesson L-125); a response-time target is a timing-session reading, not a CI gate (test-strategy.md § Performance Baselines).
Actual: the time is asserted inside every full ./gradlew build, as BulkReassignStatementCountIT’s 5-second bound was until BUG-130.
Fix: as BUG-130: keep the row-count and summary assertions, print the elapsed time against the bound, and revise F4-a AC-22 so the bound is met by a timing-session reading (as F5-h Rev 1 revised F5-h AC-7). An F4-a spec revision, so the lead approves it.
Regression test: none can show the flake on demand; the proof is that no wall-clock assertion remains in the spec.
Resolved: 2026-10-12 — fixed in forge-et-be#237 (62069249), with BUG-130. PipedriveDryRunTimingIT keeps every row and summary assertion and prints the dry-run time and its margin against the 30-second bound instead of asserting it; F4-a Rev 8 revises AC-22 to match (approved by the lead). The spec alone on macOS: 1 of 1, 951 ms against 30,000 ms.
Lesson: L-125 — added to its source list.
BUG-146 — On a short screen the open lead’s details scroll off the top when the lead list scrolls
Affects: F5-g, F11-b (forge-et-fe src/features/leads/LeadsScreen.module.css, .detailColumn; the behaviour comes from BUG-145’s fix in forge-et-fe#222) · Found in: BUG-145’s review (forge-et-fe#222, frontend-reviewer Nit 6), 2026-10-11 · Severity: low — nothing is lost or hidden for good; the user scrolls back up to see the lead
Repro: /leads with leads and a lead open, in the two-pane layout at 1280 × 720. Scroll the page down to reach the list’s rows.
Expected: the open lead’s details stay in view beside the list while the user scrolls to another row.
Actual: since BUG-145, a list pane whose header plus its data-region floor (layout/list-region-block) is taller than the screen makes <main> scroll. The detail column is as tall as the screen and scrolls up with the page, leaving empty space beside the rows.
Fix: not decided — a UX change for the lead. One option: make .detailColumn sticky (position: sticky; top: 0; align-self: flex-start) with a maximum height of the visible area and its own scroll. Regression test: at 1280 × 720 with a lead open, scroll to the last row and assert the detail heading stays in the viewport. Recorded first in design-system Rev 45’s Scope line.
Resolved: 2026-10-12 — fixed in forge-et-fe#224 (b973c0c), the sticky column, as the lead chose on 2026-10-11 (“keep the details in view”). In the two-pane layout only, the screen row is at least the content height and grows with the list pane (a sticky box never leaves its containing block), and the detail column is sticky at top: 0 with <main>'s visible content height (100vh less the top bar and the space/7 gutter above and below), keeping its own scroll and keyboard stop. The single pane is unchanged. Regression test: e2e/browser/leads/leads-layout.spec.ts “BUG-146” — at 1280 × 720 with a lead open, scrolling the page to the last row leaves the column’s top where it was and its heading inside <main>, the column still scrolls on its own, and at 768 × 1024 the column is static.
Lesson: none — one-off: BUG-145 changed the page to scroll and its review named this side effect; the layout has no second sticky pane to keep in step.
BUG-145 — The lead list’s Shared with me switch pushes Sort onto a fourth filter row and squeezes the list region at 1280 × 720
Affects: F5-g (forge-et-fe src/features/leads/LeadFilters.tsx and LeadListPane.module.css, the filter grid) · Found in: F5-g Stage 7 Step 5g design vision pass (F5-g-WI-1.3-test-verify.md § Design vision pass), deferred in forge-et-fe#221’s review record, 2026-10-11 · Severity: medium — the populated lead list’s rows are reachable only by scrolling a region 24px tall at 1280 × 720
Repro: /leads with leads, at 1280 × 720 in the two-pane layout. The filter grid holds seven cells in two columns: Country, Buyer type / Lead source, Tag / Owner, Shared with me / Sort alone on a fourth row.
Expected: the populated list shows at least a few rows at 1280 × 720 without scrolling the header away.
Actual: the data region is 24px tall, at its floor, and sits below the bottom of the screen.
Cause (measured 2026-10-11, forge-et-fe origin/main after forge-et-fe#221, Chromium, leads-shared-with-me arrangement): the pane’s header alone is taller than 720px. It holds the title, the two header actions on their own row (F12), the status pills wrapping onto two lines, the search field with its hint, a four-row filter grid in which each field reserves a message line, and Clear filters. The extra filter row from the switch is one part of it.
Tried first: a CSS-only change that adds no row for the switch (Owner spans two grid rows; the switch and Sort stack beside it, DOM order unchanged). It shortens the filter grid from 327px to 288px, but at 1280 × 720 and at 320 × 640 the region stays at its 24px floor. Kept as part of the fix.
Fix (the lead’s choice, 2026-10-11 — a floor for the region): a new layout token, layout/list-region-block (24rem; design-system Rev 45), is the data region’s min-height. The pane is at least the column’s height and grows past it when the header plus that floor is taller, so <main> scrolls, as the pane already did when it had no rows; the region’s flex basis is 0, not 0%, so its rows never size the pane. Plus the no-extra-row change above. Measured in Chromium with the leads-shared-with-me arrangement: the region is 384px at 1280 × 720, 768 × 1024 and 320 × 640, and 649px at 1440 × 1400, with no horizontal page scroll at any of them. Regression test: e2e/browser/leads/leads-layout.spec.ts “BUG-145: …” — at 1280 × 720 the region is exactly the token’s height with its rows scrolling inside it, the switch sits beside Owner above Sort, and a row scrolls into view; at 1440 × 1400 the region reaches the column’s bottom and the page does not scroll.
Resolved: 2026-10-11 — forge-et-fe#222 (a6d6697, review fixes 2b80c51). The regression test fails with the pane CSS reverted (region 360px off its floor), with a 0% flex basis (3212.9px off — the rows size the region) and with Owner’s span removed (the switch 160px from Sort’s column), and passes with the fix (leads-layout.spec.ts, 18 passed in Chromium; the BUG-145 test also passes in a one-off WebKit run). forge-et-pwa takes the shared tokens.css byte copy in its own PR.
Lesson: none — one-off: the region now has a floor token, so a header that grows again (another filter, another action) grows the page instead of hiding the rows, and design-system Rev 45 names the token for the next two-pane list.
BUG-144 — The lead list’s leadGone re-reads the lead it has just dropped, with its tags and notes
Affects: F11-f (forge-et-fe src/features/leads/LeadsScreen.tsx, leadGone; LeadView.tsx’s LeadLoader) · Found in: F5-g WI-1.1 (plan ## Notes, “Finding — leadGone re-reads the lead it has just dropped”), 2026-10-11 · Severity: low — the extra requests’ answers are discarded once the selection clears; nothing wrong is shown
Repro: open a lead, then make it unreachable (archive it from another session, or remove the share that gave reach) and trigger any path that calls onLeadGone — the archive 404, a section’s 404, or the sharing panel’s re-read (F5-g). Watch the network panel.
Expected: after leadGone, no further GET /api/v1/leads/{id}, tags or notes read for the dropped lead.
Actual: leadGone calls clearSelection, removes the lead’s detail query, invalidates the lists and calls notify. The notify state update renders LeadsScreen before the router’s replace navigation lands, so LeadLoader renders once more with the same id; its useQuery observer re-creates the removed query and fetches the lead, and the Tags and Notes sections re-read too, before the view unmounts.
Fix: drop the lead’s query only after the view has unmounted — for example in the existing focusAfterGone effect, or cancelQueries then removeQueries once selectedId is null. Regression test: a LeadsScreen test that counts the lead’s reads after onLeadGone and expects none.
Resolved: 2026-10-11 — fixed in forge-et-fe#223 (03bfc09, 76102fc; review fix cf38709). One hook, src/features/leads/useDropLeadDetail.ts, drops a gone lead’s detail: the open lead’s query is removed (which also cancels a read in flight) in the effect that runs once the selection has moved away, and any other lead’s at once. LeadsScreen’s leadGone and useLeadForm’s leave both use it. leave had the same order and read the lead once more after an edit answered 404 or its read-back did; it was found while fixing this bug and fixed under it, since it is the same defect on the same lead. Regression tests: a 404 on archive sends no further read of the lead, its notes or the tags (LeadStatusAction.test.tsx; before the fix, 4 reads of the lead where 2 were expected); an edit that drops the lead sends no read of it after the PATCH beyond the read-back (LeadsScreenActions.test.tsx; before, 2 where 1, and 1 where 0). The review of #223 (frontend-reviewer) found two more things, fixed in cf38709: LeadView.test.tsx’s two section-404 tests now assert no further read of the lead (the third path into leadGone); and the 404 handlers run after an await, so leadGone and the hook compared the gone lead with the selection in the render that created them — archiving lead A and opening lead B before the 404 landed cleared B, a defect already on main. Both now read the selection on screen from a ref; the test “a 404 that lands after the rep has opened another lead leaves that lead open” failed with the old comparison (/leads?status=all where /leads/1043?status=all was expected). Three delta reviews followed, all on the same race. bc2299f: useLeadsUrlState’s select and clearSelection read the filters on screen when they run, so a 404 after a filter change keeps the new filter; and the stale-selection test asserts the gone lead’s cache is dropped. c5b1a54: both do nothing once the Leads screen has unmounted, clearSelection({ ifSelected }) returns to the list only while that lead is open (used by the lead form’s leave, which had no check), and isSelected replaces LeadsScreen’s own copy of the ref. 819b79d: leave arms the list’s focus fallback only when it returns to the list. Each has a test that fails without it.
Lesson: L-184 (new — remove a TanStack query only once no mounted component observes it), with a frontend-reviewer row. AccountFormRoute’s edit already dropped its account on unmount for this reason; the lead screen’s two call sites did not.
BUG-143 — The account form’s certification list shows no trace id and announces no Retry when the certification read fails
Affects: F1-a (forge-et-fe src/features/accounts/AccountFormPage.tsx, the certification group’s failure state) · Found in: F12 wave 2’s review of forge-et-fe#213 (frontend-reviewer), 2026-10-10 · Severity: low — the failure message and Retry are shown; the trace id a person quotes to support and the screen-reader announcement of Retry are missing
Repro: open Add account or Edit account with GET /api/v1/certifications answering an error (a pinned 500 in a browser test). The certification group shows “Couldn’t load certifications” and Retry.
Expected: as every other failed read in the CRM (the TraceLine convention, BUG-087), a TraceLine with the response’s trace id beside Retry, and Retry inside an always-rendered aria-live="polite" container so its appearance is announced.
Actual: neither. F12 added both to the lead form’s Tea requirements group (forge-et-fe#213, leads-form-certifications-trace and leads-form-certifications-live) but left the account form’s copy alone, because the PR did not change that screen’s failure state.
Fix: add the same TraceLine and live container to the account form’s certification group; both forms now build their certification options through the shared src/lib/certifications/useCertificationOptions.ts. Regression test: a component test with an id (trace shown) and without (absent), and the live container present in every state.
Resolved: 2026-10-11 — fixed in forge-et-fe#223 (7cc664e). The account form’s certification group now holds Retry and a TraceLine (account-form-certifications-trace) inside an always-rendered aria-live="polite" container (account-form-certifications-live), as the lead form does. Regression tests in AccountFormRoute.create.test.tsx: the live container in every state, and the trace line when the failed read carries a traceId, both failed before the fix; with no traceId, no line is shown.
Lesson: L-144 (a rule about “every” case needs a check that lists the cases) — BUG-087’s rule was again met in one form and not in its twin; added to its source list.
BUG-142 — Deleting a role resolves every active user’s permissions one by one, so it takes seconds once the database holds many users
Affects: F5-d (role delete, DELETE /api/v1/roles/{id}) and F5-k (the last-administrator guard): forge-et-be AdminAuthorityGuard.refuseIfLastAdminRoleDeleted, called from RoleService.delete. Specs that fail on it: forge-et-fe e2e/browser/roles/roles-a11y.spec.ts:172 (the “Deleted e2e/browser/roles/admin-journey.spec.ts:72 · Found in: the browser gate of forge-et-fe#201 (shadcn/ui, no ticket), Windows 11, 2026-10-09 · Severity: medium — every Roles delete spec fails on a local database that has collected many test users, so the browser gate fails on every forge-et-fe PR there; production, with Empire’s 10–20 users, takes a fraction of a second · Phase: 1
Repro:
- A local database that has served many browser runs. This one held 1,694 users, 1,480 of them active (
select count(*) filter (where is_active), count(*) from app_user;), nearly all created by tests. - Sign in through the stub login as
crm01@empire.lk, create a role withPOST /api/v1/roles, then timeDELETE /api/v1/roles/{id}.
Three deletes on forge-et-be main at 03404e8e took 7.79 s, 7.92 s and 7.52 s. In the backend log, the delete started by the toast spec for role 400 began at 17:45:43 and finished at 17:45:48.7, just after the spec’s 5 s wait for /roles had ended. The same spec fails 9 of 9 on a clean forge-et-fe main. Other Java processes were running at the time (Gradle daemons, and backend test suites in other worktrees), but only one backend served port 8080, and the 16:50 failure came before those suites started.
Expected: a role delete answers in well under a second whatever the number of users, and the delete specs pass.
Actual: refuseIfLastAdminRoleDeleted calls appUsers.findByActiveTrue() and, for each active user, holdsAdminAuthority → EffectivePermissionResolver.resolve(id), which reads that user’s permissions from the database, then keeps the users holding access.manage. It reads every user before deciding (a filter, not an anyMatch), so the time grows with the number of active users: about 5 ms each here. The guard’s other checks (refuseIfLastAdmin, refuseIfLastAdminSource) loop the same way but stop at the first administrator they find, so they stay fast as long as one is found early.
Fix: resolve per role, not per user (decided 2026-10-12). One grouped query counts the active users on each role (AppUserRepository, the active-only twin of countUsersPerRole); the guard then resolves those roles together with a new EffectivePermissionResolver.resolveForRoles, which reads all their grants in one query (RoleGrantRepository.findByRoleIdIn) and applies the same filter and manage ⟹ view expansion as resolveForRole, so EffectivePermissionResolver still decides what “effective” means, as the guard’s class description requires. Resolving each role with its own query was not enough: the local database above has 246 distinct roles held by active users, because test fixtures create a role per user. All three population checks use it: refuseIfLastAdmin (another holder exists when the active users on roles conferring access.manage number more than one), refuseIfLastAdminRoleDeleted (refused when the roles that confer the key to an active user are this role alone) and refuseIfLastAdminSource (allowed when another role confers the key to at least one active user). The checks run a fixed number of queries, whatever the number of users or roles. Rejected: a single query joining users to access.manage grants, which would be a second definition of “effective” next to the resolver; stopping at the first holder found, which still loads every active user and is still slow when every holder is on the deleted role. A test that seeds fewer users, or a longer wait in the specs, would hide a cost that grows with real data.
Regression test: a backend integration test that seeds a few hundred active users and counts the SQL statements a role delete runs, with the existing CapturingStatementInspector (the pattern of BulkReassignStatementCountIT); the count must not depend on the number of users.
Resolved: 2026-10-12 — fixed in forge-et-be#239 (dd71673, review fixes b10b0c8). AppUserRepository.countActiveUsersPerRole and EffectivePermissionResolver.resolveForRoles (one role_grant read, through RoleGrantRepository.findByRoleIdIn) replace the per-user loop in all three of AdminAuthorityGuard’s population checks; findByActiveTrue is removed. The old loop cost one grant query per active user, not two as the Actual line above says: findById was served from the persistence context. Regression test RoleDeleteStatementCountIT deletes an unheld role before and after adding 300 active users, each on a role of their own, and expects equal statement counts and one population read of each kind per delete; on the code before the fix it failed with 12 statements before and 312 after. Full ./gradlew build: 6,814 tests, 0 failures. Against the local database that showed the bug (by then 1,712 active users on 250 roles), forge-et-fe main (97bb763) roles-a11y.spec.ts and admin-journey.spec.ts passed 23 of 23, and three timed DELETE /api/v1/roles/{id} took 0.052 s, 0.045 s and 0.045 s.
Lesson: L-185 (a check over every user or record runs a fixed number of queries, never a per-item service call in a loop) — new in this fix, with a line in backend-reviewer’s N+1 item. No earlier bug had the same cause; BUG-123 grew with the data too, but in a browser list, which L-147 covers.
BUG-141 — The role editor sits against the left edge of the canvas
Affects: F5-d (forge-et-fe src/features/roles/RoleEditorForm.module.css .screen, RoleEditorRoute.module.css .screen) · Found in: UI/UX review, 2026-10-14 · Severity: low · Phase: 1
Repro: open a role (/roles/:id) at 1280px or wider.
Expected: the editor is one column capped at twice layout/list-pane-width and centred, with its back link and header inside it, as the user and account editors are since BUG-154 (design-system §6 Detail page).
Actual: the editor fills the canvas (984px of content at 1280px) from the left edge.
Decision changed: design-system Rev 56 kept the role editor at full width as a “wide grid” exception (BUG-154’s Fix: line: “the role editor keeps its full-width matrix”). The editor is one column of permission-group cards, which fits the cap with nothing overflowing. Nanndish decided on 2026-10-14 to centre it; design-system Rev 58 moves it into the rule.
Fix: max-inline-size: calc(2 * var(--et-layout-list-pane-width)) and margin-inline: auto on both .screen rules (the form, and the loading, not-found and error frames).
Regression test: e2e/browser/shell/detail-page-column.spec.ts gains the role editor at 1280 and 1440px through centredColumn.ts: equal gaps within 1px, the column within the cap, the back link at the column’s left edge.
Resolved: 2026-10-16 — fixed in forge-et-fe#262 (bec9c1f4). Both .screen rules take the cap and margin-inline: auto; detail-page-column.spec.ts passes in Chromium with the role editor added.
Lesson: none — one-off: Rev 56’s exception rested on the editor being a wide matrix, which it had stopped being; §6 names no other exception.
BUG-140 — The Roles list shows a Permissions column the UI/UX review removed
Affects: F5-d (forge-et-fe src/features/roles/RolesTable.tsx) · Found in: UI/UX review, 2026-10-14 · Severity: low · Phase: 1
Repro: open /roles.
Expected: the list shows each role’s name, description and number of users.
Actual: a fourth column shows the number of permissions the role holds, which does not help anyone choose a role: the permissions are read in the role’s editor.
Decision changed: F5-d FR-13 and AC-16 (and the F5-d WI-1.2 plan’s AC-16a, WI-3.1 plan’s AC-16) list the permission count. Nanndish decided on 2026-10-14 to remove the column; F5-d spec Rev 4 records it. GET /api/v1/roles keeps permissionCount (FR-2, AC-2 unchanged).
Fix: remove the column’s header and cell from RolesTable.tsx (and from the loading table’s placeholder cells).
Regression test: RolesRoute.test.tsx asserts the headers are Name, Description and Users and that a row has two data cells; roles-list.spec.ts and admin-journey.spec.ts read the same three columns in a real browser.
Resolved: 2026-10-16 — fixed in forge-et-fe#262 (b0e05aeb). The column’s header, cell and loading placeholder are gone from RolesTable.tsx; the tests named above pass.
Lesson: none — one-off: a design decision from the UI/UX review; the column was what F5-d FR-13 asked for.
BUG-135 — RichTextBody’s unit test fails: Escape does not close the link row, so pnpm test:record records nothing (macOS)
Affects: F20-c (forge-et-fe src/features/email/compose/RichTextBody.tsx; the test is RichTextBody.test.tsx:228, “removes a link, and Escape closes the row”) · Found in: pnpm test:record on forge-et-fe#252 (ab1db02), macOS, 2026-10-13 · Severity: high — the test record runs the whole suite and records nothing when a test fails, so no forge-et-fe PR can get its required Test record check · Phase: 1
Repro: in forge-et-fe at origin/main (a248849): pnpm exec vitest run src/features/email/compose/RichTextBody.test.tsx → 1 failed, 21 passed. The full suite on forge-et-fe#252: 1 failed, 6,595 passed, 1 skipped.
Expected: after Link opens the link row and Escape is pressed, the Link address field is gone.
Actual: the field (email-compose-link-address) is still in the document. It fails the same way on F20-c’s own head (3e98ae7), whose Test record passed on Windows (Windows_NT 10.0.26200, 6,511 tests, 0 failed), so the failure is on macOS only: the test or the link row’s Escape handling depends on the platform. BUG-134, Escape then Enter in the same dialog, may share the cause.
Fix: the test, not the product. A probe found focus on the editor (email-compose-body), not the link input, when Escape was pressed. Remove link calls the editor’s focus(), which Tiptap runs in a requestAnimationFrame (@tiptap/core commands/focus.ts, delayedFocus). The test clicked Link at once, so when that frame landed after the link row’s effect had focused its input, it took focus back. jsdom runs the frame on a timer, so whether it lands first depends on the machine and on the tests run before it: the test passed alone and failed after the file’s earlier tests. A person cannot click Link within one frame of Remove link, so users are not affected (frontend-reviewer checked the keyboard path too: Ctrl+K reaches the editor only after that frame). The test waits for the editor to hold focus after Remove link, asserts Link address has focus before Escape, and asserts Escape returns focus to the editor.
Regression test: this test passing in the full pnpm test on macOS and on Windows.
Resolved: 2026-10-14 — fixed in forge-et-fe#255 (165693b, review Nit 2b82ee9), merged as c2b40984. Before the fix the file failed 1 of 22 on every macOS run on main; with it, 22 of 22 in five runs. The test record on 2b82ee9 (macOS): 6,590 tests, 0 failed. /forge-review-pr with frontend-reviewer: no Blocker or Important. One Nit fixed; one left as is (no browser test opens the compose link row).
Lesson: L-101, another instance: a test acted before a deferred focus landed. The lesson gains a bullet on waiting for focus after an action whose focus move is deferred.
BUG-134 — Email AC-45 fails on main: after Escape closes Compose, Enter on Compose email does not reopen the dialog
Affects: F20-c (forge-et-fe e2e/browser/email/email-a11y.spec.ts:102, “AC-45 (keyboard): Escape closes an unchanged dialog with focus back on Compose email; …”) · Found in: BUG-173’s pnpm e2e:gate on forge-et-fe#252 (ab1db02), macOS, 2026-10-13 · Severity: medium — it fails the browser gate of every forge-et-fe PR · Phase: 1
Repro: in forge-et-fe at origin/main (a248849, F20-c wave 1 merged), with a backend on a fresh database: ET_BACKEND_PORT=<port> pnpm exec playwright test e2e/browser/email/email-a11y.spec.ts -g "Escape closes an unchanged" --repeat-each=5 --workers=1. 4 of 5 failed. On forge-et-fe#252’s head 2 of 3 failed, and it failed in both full gate runs that day (one on a long-used database, one on a fresh one).
Expected: the test opens Compose by keyboard, closes it with Escape (focus back on Compose email), then presses Enter and the dialog opens again.
Actual: the first open and the Escape pass; after Enter on Compose email, getByTestId('email-compose') is not found within 5 seconds (settleComposeDialog, e2e/browser/email/support.ts:1376). Running repeats in parallel also fails differently: two runs connect the same principal’s mailbox and the second stub-connect answers 500 on mailbox_connection_owner_live_idx.
Fix: not decided. First find whether Enter reaches the button and whether the dialog’s open is refused (for example a close still in progress), with a trace of a failing run; then fix the product if Enter is lost, or the test if it presses before the button is ready.
Regression test: the test passing 20 of 20 alone and in a full gate run.
Resolved: 2026-10-14 — fixed in forge-et-fe#256 (43b9815; review fixes d018924).
- The cause is in the product, and Enter is lost.
close()hides the native<dialog>and returns focus to Compose email at once.ComposeEmailProviderstayed busy until the dialog’scloseevent, which comes a task later. A busy browser handles the next key press before that task, soopen()ignored the Enter. - How it was proved. On macOS the test passed 20 of 20 on
mainalone, and 10 of 10 with the CPU throttled 8×. Holding thecloseevent back by 500ms in a copy of the test failed 3 of 3, with the bug’s own error (getByTestId('email-compose')element(s) not found). With the fix, the same run passed 3 of 3. - The fix. The dialog reports that it has started to close (a new optional
onClosing). The provider keeps an open made in that gap and runs it once the dialog has finished. Opens made while a dialog is loading or open are still ignored.composeEmail.ts’sOpenComposeEmaildoc says so. - The send-success path. It reports the close through
useEffectEvent, so the answer effect does not re-run when the callback changes. - Regression tests. Two rows in
ComposeEmailProvider.test.tsx, Cancel and a successful send. Each holds the jsdomcloseevent back, presses Open compose, then releases it. They assert the kept open shows,onClosedoronSentfires once, the sent toast is raised and focus returns to the opener. Both fail against the old provider. - Results after the fix. AC-45 passed 20 of 20 alone (
--repeat-each=20 --workers=1), and the email browser suite passed (56). - Review.
frontend-reviewerfound no Blocker or Important. Its six Nits are fixed ind018924.
BUG-034 had the same gap, and there it was fixed in the test, because a person cannot press within one task. On a busy machine, Chromium runs input ahead of that task, so the gap is longer.
Lesson: L-101, another instance (BUG-034’s cause). Its sources gain BUG-134, with a bullet: when the event that finishes a visible change comes a task later, input in that gap is the product’s to keep, not the test’s to wait out.
BUG-133 — The size matrix’s user-editor states fail in WebKit: the role select measures 1px by 18px, under the 24px target-size floor
Affects: F5-c (the user editor’s role field, forge-et-fe src/features/users/), through e2e/browser/size-matrix/size-matrix.spec.ts:216 (F40 FR-4, the shared page check) · Found in: BUG-173’s pnpm e2e:gate on forge-et-fe#252 (ab1db02), macOS, fresh database, 2026-10-13 · Severity: medium — 26 states fail the browser gate of every forge-et-fe PR; the effect on users is not yet known · Phase: 1
Repro: in forge-et-fe at origin/main (a248849) with a backend running: ET_BACKEND_PORT=<port> pnpm exec playwright test e2e/browser/size-matrix/size-matrix.spec.ts --project size-1024x768-webkit -g user-editor. Every user-editor state fails at 768×1024 and 1024×768 in WebKit; the Chromium sizes pass.
Expected: every user-editor state passes the shared page check at every size.
Actual: the check reports select "Sales RepresentativeIT Admin" is 1px × 18px, under the 24 × 24px floor (the role options read as one name). It looks like a hidden native select the role picker renders beside its visible control, which WebKit lays out at 1px wide.
Fix: not decided. Find which element it is. If it is a hidden form select that a user never reaches, hide it from the page check and assistive technology the way the component library intends; if a user can reach it, give it the target size.
Regression test: the user-editor states passing at both WebKit sizes in the size matrix.
Resolved: 2026-10-14 — fixed in forge-et-fe#256 (2f55f88; review fixes d018924).
- The element. It is the native
<select>that Radix Select renders beside its trigger when the Select is inside a form (SelectBubbleInput). It isaria-hidden, hastabindex=-1, and is styled with Radix’s visually-hidden styles (position: absolute, 1 × 1px,clip: rect(0, 0, 0, 0)). No user can point at it or reach it, and Radix already hides it from assistive technology. - Why only WebKit. WebKit gives a
<select>a minimum height, so the element measured 1 × 18px. The page check skipped the visually-hidden pattern only for a box of at most 1 × 1px. - The fix is test-side.
e2e/support/pageCheck.tsalso recognises the pattern by its clip:clip: rect(0, 0, 0, 0)on an absolutely or fixed positioned element, orclip-path: inset(50%)(Tailwind’ssr-onlyand the repo’s.visuallyHiddenrules). The product is unchanged. - Regression tests.
e2e/browser/support/page-check.spec.tsgains a row where clipped controls with 1 × 18px boxes pass; it fails without the fix. Three rows keeptarget-size: a 1 × 18px control not clipped, one clipped but not positioned, and one clipped to an area. - Before and after. Before the fix: 13 failed at 1024×768 in WebKit. After: 41 passed at both WebKit sizes (
-g user-editor). - The PWA copy.
pageCheck.tsis shared with forge-et-pwa by copy (scripts/shared-with-fe.json, L-190). The re-copy is in the BUG-184 PR, merged after this one.
Lesson: none — one-off: the page check recognised the visually-hidden pattern by box size alone, and WebKit’s minimum height for a hidden <select> broke that. No other Lesson: line or lesson has that cause.
BUG-131 — The WebKit offboarding keyboard test fails on every run: neither Tab nor Alt+Tab reaches a link in Playwright’s WebKit (Windows)
Affects: F5-h (forge-et-fe e2e/browser/users/offboarding-keyboard-webkit.spec.ts:51, “a leaver’s records, reports and deactivation are completed by keyboard alone, in WebKit”; the key choice is e2e/browser/users/offboardingJourney.ts and the spec’s top comment, F5-h WI-3.1 plan Decision 6) · Found in: F12 wave 2’s pnpm e2e:gate on feature/F12-wave-2 at 183d587, 2026-10-10 · Severity: medium — it fails the browser gate of every forge-et-fe PR run on Windows; the product is not affected
Repro: Windows 11, Playwright’s bundled WebKit. On pure main — forge-et-fe 344d2a0 and forge-et-be 86f5cee9 on the local profile on 8080, against an isolated compose stack on a fresh database: ET_BACKEND_PORT=8080 pnpm exec playwright test --project=chromium e2e/browser/users/offboarding-keyboard-webkit.spec.ts → 1 failed, 9 passed, every run (five runs, on main and on F12’s wave head, fresh and reused databases).
Expected: the walk reaches the user editor’s Offboard user link by keyboard and completes, as it does on macOS (F5-h wave 3’s gate, 06d9f5d).
Actual: Alt+Tab reaches the editor's Offboard user link within 60 presses fails. Logging document.activeElement after each press shows the same ten-stop cycle for both keys: the user menu, the sidebar toggle, the editor’s four fields, Save, Discard, Deactivate user, then <body>. No <a> is ever focused — the sidebar’s links are skipped too. The spec probes Tab first and falls back to Alt+Tab, which is Safari’s Option+Tab on macOS; Playwright’s WebKit on Windows does not treat Alt+Tab that way, so neither key reaches a link there.
Fix: test only, not decided — the owner of F5-h’s suite chooses. Options: find a key or a WebKit setting that makes Tab reach links on Windows; or keep the WebKit walk to macOS and state that in the spec and the plan’s Decision 6. Until then, F12’s wave 2 marks the test with markKnownFailure('BUG-131', …) only when process.platform === 'win32', so the gate passes on Windows and the test still runs, and must pass, on macOS. The fixing PR removes that mark.
Regression test: the same spec, run on Windows, passing without the mark.
Resolved: 2026-10-12 — the lead chose the second option on 2026-10-11: the WebKit walk runs on macOS only. forge-et-fe#224 (3f4a01d) replaces the Windows known-failure mark with test.skip(process.platform === 'win32', …), whose reason names this bug, so Windows reports the test as skipped with that reason and macOS still runs it and must pass it. F5-h’s Chromium keyboard walk still runs on every platform. F5-h WI-3.1 plan Decision 6 carries a note saying so. The regression test above does not apply: no Windows run can pass the walk.
Lesson: none — one-off: a Playwright WebKit difference on one platform, decided per test.
BUG-130 — Flaky test: BulkReassignStatementCountIT’s AC-7 rows miss their 5-second bound in a full ./gradlew build under load (Windows)
Affects: F5-h (forge-et-be src/test/groovy/com/empireteas/api/BulkReassignStatementCountIT.groovy, the two AC-7 rows “1,000 account records cost the statements 10 do, outside the audit inserts, in under 5 seconds” and the lead row) · Found in: F1-d Wave 1, the ./gradlew build on feature/F1-d-wave-1 at 4e6f5344 (which carries main with F5-h wave 2), 2026-10-09 · Severity: low — a flaky integration test · Phase: 1
Repro: Windows 11, ./gradlew build on forge-et-be (6,192 tests, 1 h 7 min), with an isolated compose stack and a running backend from another worktree on the same machine. Both AC-7 rows failed on large.millis < BOUND_MILLIS only: 5,944 ms for accounts, 6,480 ms for leads, against 5,000. Every statement-count assertion in the class passed. ./gradlew test --tests 'com.empireteas.api.BulkReassignStatementCountIT' on the same commit then passed 6 of 6.
Expected: the test passes however loaded the machine is.
Actual: a wall-clock bound that a loaded machine can miss. F1-d did not cause it: the lead row failed the same way, and F1-d changes nothing on the lead path.
Fix: not yet decided. L-125 says a test must not depend on how much work a real-time budget allows: keep the statement-count bound, which is the property AC-7 protects, and either drop the millisecond bound from the full suite or move it to a timing run that takes the machine lock.
Regression test: the full ./gradlew build on Windows, green across repeated runs.
Also seen, 2026-10-09: the record run for forge-et-be#225 on the same head (4e6f5344), with only the idle compose stack and a code reviewer reading files beside it, failed the accounts row alone at 5,232 ms (6,192 tests, 1 failed, 2 skipped). The lead row passed.
Also seen, 2026-10-09 to 2026-10-11 (BUG-100’s fix, forge-et-be#235, on a Windows 11 machine running other sessions’ builds): the full ./gradlew build on fix/BUG-100-mailbox-connect-deactivation-race failed both rows, at 8,101 ms (accounts) and 8,349 ms (leads) (6,143 tests, 2 failed, 2 skipped, 53 min). The class alone on that branch then failed only the accounts row (the lead row passed, at 4.2 s). The class alone on unchanged origin/main (894b1543), with other builds running, failed both rows, at 58 s and 36 s. A later full build on the branch passed both rows.
Renumbered: filed as BUG-126; main had taken BUG-124 to BUG-127 before this record merged, so it became BUG-130 (L-057). The F1-d commit docs(F1-d): record BUG-126, … and the forge-et-be#225 description written before the merge say BUG-126.
Resolved: 2026-10-11 — fixed in forge-et-be#237 (221de799), the first option of the Fix: line. The two AC-7 rows keep every statement-count assertion and print the 1,000-record time and its margin against the 5,000 ms NFR-1 bound instead of asserting it. The bound is read in a timing session on a quiet machine (test-strategy.md § Performance Baselines), so F5-h AC-7’s “completes in under 5 seconds locally” is now that printed reading rather than a test failure (F5-h spec Rev 1, approved by the lead on 2026-10-11). The spec run alone after the change, on macOS: 6 of 6, account 891 ms, lead 616 ms. It was not re-run on Windows: with no time assertion left, the rows no longer depend on the machine’s load.
Lesson: L-125 (a test that asserts what a time-bounded run got done fails on a loaded machine; remove the timing) — added to its source list.
BUG-129 — Flaky tests: two lead-view archive states time out in the WebKit 1024 size matrix under full-suite load (Windows)
Windows only: seen only on Windows 11; no macOS run is recorded as reproducing it. Fix it and prove the fix on a Windows machine.
Affects: F11-f (forge-et-fe e2e/browser/size-matrix/size-matrix.spec.ts:136, the leads-lead-archived and leads-lead-archive-done page states, project size-1024x768-webkit) · Found in: F1-d Wave 1, Stage 7 Step 5e, 2026-10-09 · Severity: low — a flaky browser test · Phase: 1
Repro: Windows 11. The forge-et-be F1-d wave branch (4e6f5344) on 8081, against an isolated compose stack (et-f1-d-wave1); ET_E2E_WORKERS=3 pnpm exec playwright test (every project, 2,464 tests, 1.2 hours) in the forge-et-fe worktree at 4314d8d. Run wi15-all-1: 2,457 passed, 5 skipped (the expected laptop sign-in skips), 2 failed.
Expected: each state passes the shared page check at every size, however loaded the machine is.
Actual: both states hit the 30-second test timeout inside the page check, with no real finding: one shows only Test timeout of 30000ms exceeded, the other a horizontal-scroll entry whose detail is the same timeout. Both states pass in every Chromium size and in WebKit at 768. Run alone, the five leads-lead-archive states in size-1024x768-webkit passed (14 passed with setup, 2 workers, 43 seconds). F1-d’s AC-22 link renders on these screens, but the failures are timeouts, not page-check findings, and the same states pass alone with the link present.
Fix: not yet known. Reproduce under forced load in WebKit and keep the trace; find the step that runs out (likely the WebKit page check on a heavier lead page). Then either cut the work, or give these states a longer budget that names why. A green re-run is not the proof (L-107).
Regression test: the WebKit 1024 project, run inside the whole suite on Windows, green across repeated runs.
Resolved: 2026-10-30 — fixed by BUG-218’s fix, forge-et-fe#365. The cause found on 2026-10-22 (below) holds: the run’s own browsers saturated the CPU. Fewer browsers alone did not stop it (4 browsers still had 5 flaky tests); what did was that each test’s fresh browser context no longer fetches and compiles the dev server’s some 270 source modules: the suite runs against a development build served by vite preview, and Windows starts at most 6 browsers. In BUG-218’s three full gates in a row (--full on ece00c5a, a full ./gradlew build alongside), the WebKit 1024 project was green in all three, both lead-archive states included.
Lesson: L-248.
Renumbered: filed as BUG-125; main had taken BUG-124 to BUG-127 before this record merged, so it became BUG-129 (L-057). Commits made before the merge say BUG-125.
Also seen (F26-g Wave 1, Stage 7 Step 5e, 2026-10-13, Windows 11, forge-et-fe WI-1.4 branch 7a87568 against the F26-g wave backend on 8380): the same failure in pricing states, one per run, a different state each time, in WebKit only: pricing-edit-dates-errors at 1024 (run 2, “waiting for element to be visible, enabled and stable” on the Edit dates button until the 30-second test timeout); pricing-import-sales-refresh-error at 768 (run 3, the same wait on pricing-import-sale-trigger); pricing-weekly-not-in-round-error at 1024 (a WebKit-only re-run of the pricing states, a horizontal-scroll entry whose detail is the 30-second timeout). Each passed when run again. Passing WebKit pricing states took a median of 6.1 s and at most 14.3 s, so the failure is a stall, not a slow page. The machine was loaded (other sessions’ backends and test runs; free memory fell to about 2 GB once). So the cause is not limited to the lead pages. Record: .forge/plans/F26-g/F26-g-WI-1.4-test-verify.md § O4.
Also seen (forge-et-fe#250’s gate, 2026-10-13, Windows 11, 8 workers, 3,452 tests, 79732ce), now in Chromium too:
- What timed out: four account size-matrix states at 1024×768 in one burst, at 24–33 s (
account-edit,account-new,accounts,account-edit-load-error), ande2e/browser/users/journey.spec.ts:164, the administrator’s journey (page.evaluate, the axe scan, until the 30-second timeout). - Re-runs: run alone on the same stack right after, all passed: the journey 10 of 10, the four states 13 of 13.
- So: the stall is not WebKit-only. A full gate run on a loaded Windows machine can push any state past 30 s.
Cause found (reported by Nanndish to the lead, recorded 2026-10-22; Windows 11, 16 cores): the full browser run uses every core on its own. Playwright starts 8 browsers at once by default (half the cores), and with the database containers and Windows’ own background work on top, the CPU sits at 96–100% for the whole run, with nothing else open. WebKit handles this worst: a WebKit test that takes 4–8 seconds on a quiet machine takes about 14 seconds in the full run, and the slowest up to 28, against the 30-second budget.
- Why tests fail: under that load WebKit sometimes stops drawing the page for several seconds. Playwright’s “is this element stable?” check and the axe scan both wait for the page to draw, so a test that meets one of those pauses runs out of its 30 seconds. This is why it is a different page each time (lead archive, pricing, accounts), why it shows only in a full run, and why the tests pass alone.
- What it is not: a defect in the CRM. The pages are correct; the machine has no spare CPU during the run.
- Next (in progress): the same run with 6 and with 4 browsers at once, to see whether the failures stop and how much longer the run takes; and catching the rarer case where a click waited more than 26 seconds.
- Likely fix: fewer browsers at once on Windows, which makes the run slower; the figures go to the lead. Two other options need the lead’s decision: giving WebKit tests more than 30 seconds, which this file allows but
.forge/test-strategy.md § Flaky Testsforbids; or excluding the worktrees folder from Windows Defender, which is a security setting.
BUG-128 — Flaky test: RowSelection.integration.test.tsx times out at 20 seconds in a full Vitest run under load (Windows)
Windows only: seen only on Windows 11; no macOS run is recorded as reproducing it. Fix it and prove the fix on a Windows machine.
Affects: F5-h (forge-et-fe src/ui/components/RowSelection/RowSelection.integration.test.tsx, “reads mixed, checked, then unchecked…”) · Found in: F1-d Wave 1, 2026-10-09 · Severity: low — a flaky unit test · Phase: 1
Repro: Windows 11, pnpm test (260 files, about 4,420 tests) on forge-et-fe branches carrying main at 80b5607 or later. Seen in two separate full runs the same day: the AC-22 fix’s first run on feature/F1-d-wave-1 (38fda7e), and WI-1.5’s author run on feature/F1-d-WI-1.5-e2e-account-detail. Each time it timed out at 20000 ms beside one other load-sensitive test (LeadFormDialog.test.tsx AC-26 once, LeadsScreenActions.test.tsx once), and both passed when run alone and on the full re-run.
Expected: the test passes however loaded the machine is.
Actual: under full-suite load it can exceed its 20-second budget.
Seen again, 2026-10-14 (F1-e WI-1.3, feature/F1-e-WI-1.3-tag-library at 920b1db7, base 4209b948): the test fails when nothing else loads the machine.
- A full
pnpm teston an idle Windows 11 machine failed it, beside 5 lead tests (LeadFormDialog.test.tsx×2,LeadsScreenActions.test.tsx×3) that pass alone. - Run alone, it timed out at 21.1 s, 20.4 s and 20.9 s.
- With
--testTimeout=60000it passed at 22.2 s; on theforge-et-femain checkout (1b6e7168) it passed at 17.7 s.
It sits near its 20-second budget even alone (25 user.click calls in jsdom), so it is a budget problem and not only a load problem.
Fix: not yet known. Reproduce with the file run as several parallel copies (as BUG-107 did), find the slow step, and cut it or give the test a budget that names why.
Regression test: the file run as parallel copies for three rounds with no failure.
Resolved: 2026-10-16 — fixed in forge-et-fe#229 (6a9f545; the review’s Nits in d7a5306 and 21566ac), test only. The cause was the case’s work, not the code: it clicked all 25 row boxes one at a time, each click re-rendering the table and the bulk-action bar (about 120 ms) and each getByRole lookup taking about 55 ms, so the case took 9.3 s alone on Windows 11, nearly half its 20 s testTimeout. It now ticks rows 1 and 2 (mixed), selects the page with the header box (checked), unticks row 25 (mixed, and no longer checked) and ticks it again (checked from a row change), then pages on (unchecked, the bar still counting 25): 6 clicks. The one-by-one walk moved to useRowSelection.test.ts (“reads page ‘all’ once every row of the page is ticked one by one…”), which also checks the bar offers Select all and reads “Selected 25 on this page”. The timeout was not raised (test-strategy.md § Flaky Tests). The runs, on Windows 11, 16 copies of the file in one Vitest run, three rounds: before, rounds 1–2 passed 80 of 80 and round 3 failed 13 (all 13 this case, “Test timed out in 20000ms”); after, 80 of 80 in each of six rounds (three before and three after merging main at 68c0be7). The case alone now takes 1.9–4.8 s. Review: frontend-reviewer, no Blocker or Important; its three Nits fixed in d7a5306, and the delta’s one Nit (a cast, now .at(-1)) in 21566ac.
Lesson: L-125 (a test whose time limit assumes an idle machine fails on a loaded one) — another instance, as BUG-172 was: a Vitest per-test timeout that an idle machine meets and a loaded one does not. Here the fix cut the test’s work rather than giving it its own timeout; added to L-125’s source list.
Renumbered: filed as BUG-124; main had taken BUG-124 to BUG-127 before this record merged, so it became BUG-128 (L-057). Commits made before the merge say BUG-124.
BUG-127 — Flaky test: the lead view’s “match dialog open” axe test scans while the dialog is still fading in, so it fails now and then (macOS)
Affects: F11-f (forge-et-fe e2e/browser/leads/lead-detail-a11y.spec.ts:210, “axe: a live lead with the match dialog open”; the dialog is src/features/leads/LeadMatchModal.tsx) · Found in: F5-h wave 3’s pnpm e2e:gate --push on feature/F5-h-wave-3 at 7f18e71, 2026-10-09 · Severity: medium — it fails the browser gate of any forge-et-fe PR now and then, in a file the PR need not touch
Repro: macOS, against an isolated stack (et-f5-h-wave3, forge-et-be 03404e8e on 18095), nothing else heavy running: ET_BACKEND_PORT=18095 pnpm exec playwright test --project=setup --project=chromium e2e/browser/leads/lead-detail-a11y.spec.ts -g "match dialog open" --repeat-each=10 → 1 of 10 failed. The same test passed in three whole-gate runs earlier the same day on the same code (330ff42, a9a0909 twice) and failed in the fourth.
Expected: the axe scan of the open match dialog passes every time.
Actual: axe reports color-contrast (serious) on the dialog’s Save button (data-testid="leads-match-save"). The test opens the dialog, waits for the two pickers to stop being busy, and scans at once; it never waits for the dialog’s open animation, and the file calls settleAnimations nowhere. The dialog animates in from opacity: 0, so a scan that lands during the fade measures the button partly transparent. This is the cause BUG-033 found in the confirm-dialog row; the settled colours pass.
Fix: test only, not decided in detail: call settleAnimations(page) (or wait for the dialog’s finite animations, as BUG-033’s fix does) before each expectNoAxeViolations in lead-detail-a11y.spec.ts that follows a dialog, popover or toast opening. No product change.
Regression test: the same test with --repeat-each=20 under a throttled CPU, failing before the fix and passing after.
Resolved: 2026-10-09 — forge-et-fe#206 (3ab3972). Each of lead-detail-a11y.spec.ts’s four axe scans now calls settleAnimations (from e2e/browser/users/support.ts) first; test only. The runs, on macOS against the isolated stack et-f5-h-wave3 with no CPU throttling: before, --repeat-each=10 failed 1 and --repeat-each=20 failed 2; after, --repeat-each=30 passed all 30. Whole gate on the head: Browser gate: passed · 3ab3972a20a2 · 2331 passed · 5 skipped · 0 known failures. Review: approve (frontend-reviewer: the dialog’s modal-in is a CSS animation getAnimations() reports, and the settled dialog is still scanned with the full rule set). Left as is: moving settleAnimations into e2e/support/ (a separate change, as pageCheck.ts is shared with forge-et-pwa).
Lesson: L-101 — a browser test waits for the page to settle, not for an element to appear; BUG-127 is added to its sources as another instance of BUG-033’s cause.
Renumbered: filed and fixed as BUG-125; main had taken BUG-124 and BUG-125 (F11-f’s wrap-up) before this record merged, so it became BUG-127 (L-057). forge-et-fe#206 — its branch fix/BUG-125-…, its commit fix(BUG-125): … and the describe comment in e2e/browser/leads/lead-detail-a11y.spec.ts — still says BUG-125. The comment was corrected to BUG-127 in forge-et-fe#223 (baf72cc); the branch and commit names stay as they were.
BUG-126 — After a refused deactivation, the offboarding screen’s Records section still says nothing is left while the header names the new record
Affects: F5-h (forge-et-fe src/features/users/OffboardDeactivateSection.tsx confirm(), which after a refusal re-reads only the offboarding summary; the Accounts and Leads sections of /users/:id/offboard) · Found in: F5-h WI-3.1 Run B, on feature/F5-h-WI-3.1-e2e-offboarding at a9a0909 against forge-et-be 03404e8e, 2026-10-09 (the plan’s Q-3) · Severity: low — the header’s blocking line and the blocked Deactivate user are correct, so nothing wrong can be done; only the list’s text is stale until the page is reloaded
Repro:
- Open
/users/<L>/offboardfor an active user L who owns nothing and has no direct reports. The header reads “Ready to deactivate” and the Accounts section reads “No accounts left to reassign”. - Without reloading, give L an account through the API (or from another browser tab).
- Choose Deactivate user and confirm. The server refuses with
422 USER_OWNS_RECORDS.
This is the flow of e2e/browser/users/offboarding-deactivation.spec.ts test (a); a one-off probe after its last assertion recorded {"emptyCount":1,"emptyText":"No accounts left to reassign","rowCount":0,"blocking":"Owns 1 account"}.
Expected: after the refusal, the Accounts section lists the account the header now counts, so the administrator can move it from the same screen.
Actual: the header reads “Owns 1 account” and Deactivate user is blocked with the records reason, but the Accounts section still reads “No accounts left to reassign” and shows no row, until the page is reloaded.
Fix: not yet decided. One option: on a USER_OWNS_RECORDS refusal, also invalidate the owned-record lists of each record type (the same keys a Change owner pass refreshes in offboardingQueries.ts). AC-15 asks only for the summary to be re-read, so this is a gap in the wave-1 design (WI-1.4 plan), not a regression.
Regression test: extend test (a) of offboarding-deactivation.spec.ts to assert the Accounts section lists the new account after the refusal, with a jsdom test beside WI-1.4’s refusal case.
Resolved: 2026-10-11 — fixed in forge-et-fe#223 (9be5302). After a refused deactivation, OffboardDeactivateSection confirm() now invalidates every owned-record list of the user, through a new prefix key offboardingKeys.ownedRecords(id), as well as the summary. Regression tests: a jsdom refusal case in OffboardDeactivateSection.test.tsx, which failed before the fix (Unable to find role="checkbox" and name "Select FIXTURE_Account 1"), and test (a) of offboarding-deactivation.spec.ts, which now asserts the Accounts section lists the new account; that file’s header no longer says the list is left unasserted.
Lesson: none — one-off: AC-15 asked only for the summary to be re-read, and the wave-1 design did exactly that; no other bug in bugs.md has a refusal handler that re-reads only part of the state the refusal reports.
BUG-125 — The Users search box lost characters from the middle of typed text once, under full-suite load
Affects: F5-c (forge-et-fe the Users list search box; test e2e/browser/users/journey.spec.ts:160, its searchFor helper at line 176) · Found in: F11-f WI-1.6’s whole-tree browser run, round 2 attempt 1, on forge-et-fe feature/F11-f-WI-1.6-e2e-lead-detail at 8e5ecb5, 2026-10-08 · Severity: low · Phase: 1
Repro: not yet reproduced. Seen once on Windows 11 while another session’s Testcontainers suite used a full CPU core: box.pressSequentially('f5c-wi41-browser-840df3a0'), then expect.poll(() => URL q).toBe(…) timed out with q = f5c-wi41-browsa0. The same spec passed on the next runs on that machine.
Expected: the search box, and the q it writes to the URL, hold every typed character.
Actual: the characters er-840df3 from the middle were lost. A controlled input whose value is rewritten from a debounced URL update could drop keystrokes typed between the read and the write; a test-side cause is also possible.
Fix: not yet known. First reproduce it with the CPU throttled (CDP Emulation.setCPUThrottlingRate, L-162) in a throwaway test; if it reproduces, the product loses typed text on slow machines and is fixed there, not in the test.
Regression test: the throttled probe, kept as a component or browser test once the cause is known.
Cause (found 2026-10-12, by reading the code against the recorded value): a product defect. UsersScreen resets the search box to the URL’s q whenever q changes and the box differs from it, so that Clear filters and Back move the box; but the debounce’s own write of q also changes it. When the router applied that write after more keys had been typed, the box went back to the older text and the newer keys were lost. The recorded q, f5c-wi41-browsa0, is exactly the box reset to f5c-wi41-brows followed by a0. A throttled browser probe did not reproduce it (0 losses in 39 runs at 6× and 12×), because it needs the navigation to commit late; a component test that holds the navigation with a route loader reproduces it every time.
Resolved: 2026-10-12 — fixed in forge-et-fe#224 (f02a19a). The screen keeps the last q it wrote, and that q arriving leaves the box alone; a q from elsewhere still moves it. Regression test in UsersScreen.test.tsx “keeps what was typed after a debounced write when the URL catches up late (BUG-125)”: the box held abc before the fix and holds abcdef after. The review of #224 added one case, fixed in aafaa7e: a filter change replaces a held debounced write, which then never arrives, so applyFilters now forgets it; a test that fails with the first fix’s code covers it. A filter change’s own write of q can land late in the same way, so since 2e394b7 applyFilters remembers it as this screen’s write when it changes q; its test fails without that.
Lesson: L-162 (a failure that appears only under load is reproduced before it is called flaky) — another instance, which also shows its limit: a race that needs a late router commit is not reproduced by throttling the CPU, but by holding the step that must land late (here a route loader); added to its source list.
BUG-124 — A plain (uncontrolled) Combobox blanks the text being typed when the page around it scrolls while its list is open
Affects: F26-b (forge-et-fe src/ui/components/Combobox/Combobox.tsx, every caller that does not pass inputValue — for example the Users editor’s manager picker, src/features/users/UserEditorForm.tsx) · Found in: F11-f WI-1.3’s review fix of the Combobox (e1733c7), 2026-10-08 · Severity: low · Phase: 1
Repro: in a component test, render a Combobox without inputValue, type a letter so the list opens, fire a scroll event on an element that contains the input, then type another letter: the input holds only the second letter (the frontend-implementer’s probe, recorded in the F11-f WI-1.3 plan ## Notes). In a browser on a slow machine, focusing a field below the fold scrolls its container after the first keystroke, with the same result (F11-f findings F-001).
Expected: typed text survives the list closing on a scroll.
Actual: react-aria closes an open list when any ancestor of the input scrolls (useCloseOnScroll) and commits as it closes; in uncontrolled mode its own resetInputValue puts the selected label (or nothing) back into the input. F11-f fixed the controlled (server-search) mode in Combobox.tsx (e1733c7) and left this mode alone, because a fix needs a change to the existing patches/react-aria@3.52.1.patch.
Fix: not yet decided — extend the react-aria patch so a scroll close does not reset the input, or move the remaining callers to the controlled mode, where Combobox.tsx already guards it.
Regression test: the component probe above, for the uncontrolled mode.
Resolved: 2026-10-12 — fixed in forge-et-fe#224 (e2c8c50), the second option, inside Combobox.tsx alone: with no inputValue, the component holds the text in its own state and runs react-aria in controlled mode, so the existing scroll-close guard covers every caller. No caller and no react-aria patch changed. The text follows the selection as the uncontrolled input did: the label from the first render, the new label (or nothing) when the caller changes the selection, and the label once the selected option arrives while the field is not focused. Two differences, neither relied on by any caller (UserEditorForm, CreateUserModal, RoundSalePicker, SaleStep were checked): picking the option already selected no longer calls onChange again, as the server-search mode already did; and a null commit (Esc, or leaving the field, with nothing selected or the selected option missing from options) restores the text instead of calling onChange(null) — clearing the text still calls it. Regression tests in Combobox.test.tsx: type, scroll, type keeps both letters (e before the fix, te after), and editing a selected label keeps the edit (Tetleyy before, Tetleyxy after); six more rows guard the text following the selection.
Lesson: none — one-off: F11-f fixed one mode of a two-mode component and filed the other here.
BUG-123 — The lead view’s tag suggestions list fails axe’s scrollable-region-focusable once the database holds enough tags for the list to scroll
Affects: F11-f (forge-et-fe src/features/leads/LeadTagsSection.tsx, the Add tag combobox’s “Suggestions” list box; test e2e/browser/leads/lead-detail-a11y.spec.ts:223, “axe: a live lead with the tag combobox open”) · Found in: F5-h wave 2’s browser gate on feature/F5-h-wave-2 at 121719c, 2026-10-09 · Severity: medium
Repro:
- Use an isolated stack that has served several full browser-gate runs (
et-f5-h-wave2, after two gates and two seam/browser runs). Each run’s lead tests create tags. - Run
ET_BACKEND_PORT=18095 pnpm exec playwright test --project=chromium e2e/browser/leads/lead-detail-a11y.spec.ts -g "tag combobox open" --repeat-each=5.
It fails 5 of 5. On the same stack’s first gate run, earlier the same day (0c03393), it passed.
Expected: the open suggestions list passes axe at any number of tags.
Actual: axe reports scrollable-region-focusable (serious) on the role="listbox" aria-label="Suggestions" element. Once enough tags exist for the list to overflow, it scrolls, and it holds no focusable element. A likely cause: the combobox keeps focus on the input and moves through the options virtually, so neither the list nor its options take focus. The cause is not confirmed.
Fix: not yet decided. One option is for the list box to be keyboard-scrollable, or for its options to be focusable; the other is to show a bounded number of suggestions. A test-only fix that seeds fewer tags would hide a real defect at production data sizes.
Regression test: the same axe test, run against a database with more tags than fit in the list.
Resolved: 2026-10-12 — fixed in forge-et-fe#224 (043eaf6), a bounded list, as the lead chose on 2026-10-11 (at most 10 suggestions). Add tag shows at most ten options, Create tag included and kept last; when more tags match, a line under the list says “Type to find N more tags”, and typing reaches the rest. Ten rows are taller than the list’s 40vh cap at 1280 × 720, so Combobox gains shortList (the list drops its height cap and is as tall as its options) and moreMessage (the line under the list, outside it, which the list names as its description). Regression tests: LeadTagsSection.test.tsx lists ten of 25 matching tags (all 25 before the fix); the axe test in lead-detail-a11y.spec.ts pins 25 tags and checks ten options, the line and no scrolling; and a size-matrix page state, leads-lead-tag-picker-many, checks the open list at every size. The review of #224 found two more things, fixed in aafaa7e and a97bcc4. First, the line was only the list’s description, which screen readers do not read while focus stays on the input, so a short-list Combobox now also says it from an always-rendered polite live region outside the popover. Second, ten rows did not fit the room the window leaves at 320 × 640: the last option ended outside the popover. Every other size fit, 1280 × 720 included. Add tag keeps “at most 10” from bp/tablet up and lists seven below it, and expectShortTagList now checks at every size that the last option and the line sit inside the popover and wholly in the window (all seven sizes pass). The live region has not been heard with a real screen reader; it follows the pattern every other announcement in the CRM uses. One manual check is still worth making with VoiceOver or NVDA (open the list, type one letter): react-aria announces its own “N options available” a moment later, and if that announcement interrupts the polite line, the line needs another place. The cap follows the window’s width (mq.tablet), while what overflows is height, so the fit is proven at the size matrix’s sizes only; a shorter desktop window outside them (for example a 1366 × 768 laptop with the field mid-window) is unchecked, and the user can still scroll the field up before opening the list. 2e394b7 makes the Combobox root position: relative, so the hidden region stays inside it (L-176).
Lesson: L-147 (run-unique browser fixtures grow the local database, so a test that reads a shared list fails later) — this one exposed a real defect at production data sizes rather than a test fault, and its fix pins 25 tags so the test no longer depends on what earlier runs left; added to its source list.
BUG-122 — The import screen’s keyboard test types a date month-first and moves selects with the arrow keys, so it fails where the region orders dates year-first (macOS)
Affects: F26-c (forge-et-fe e2e/browser/pricing/price-import-a11y.spec.ts:166, “AC-23: the whole import works by keyboard alone…”, the New sale step) · Found in: F5-h wave 1’s browser gate after merging main into feature/F5-h-wave-1 (abffef6), 2026-10-09 · Severity: medium — every developer whose Mac region orders dates year-first fails the browser gate on every forge-et-fe PR that contains this test · Phase: 1
Repro: macOS with the language English and the region Sri Lanka (defaults read -g AppleLocale prints en_US@rg=lkzzzz), on forge-et-fe main at 18d9d15; pnpm exec playwright test e2e/browser/pricing/price-import-a11y.spec.ts --project=chromium -g "AC-23" against a local backend. It fails every time (1 failed, 9 passed).
Expected: the keyboard walk enters 20 October 2026 into the New sale dialog’s Start date on any developer machine.
Actual: the test focuses the date input and types 10202026, which assumes the segments run month, day, year. Chromium on macOS lays out a date input’s segments from the system region whenever the system language matches the page’s language, so here the first segment is the year: the field holds 102020-02-06 and toHaveValue('2026-10-20') fails. The spec already sets locale: 'en-US' (line 71), which changes navigator.language and Intl but not the date input’s segment order.
With the date fixed, the same walk failed next on its column selectors: it moved each native <select> with ArrowUp/ArrowDown, which do not change a closed select on macOS (L-143), so a selector stayed on its old column.
Fix: in forge-et-fe#194, test only. dateOrder reads the browser’s date-part order once, from a scratch date input on a blank page in the same browser context (01, 02, 2003 typed: month first gives 2003-01-02, day first 2003-02-01, year first leaves it empty), and typeDate types each date in that order. typeAheadTo replaces arrowTo: it presses the wanted option’s first letter as many times as one-letter type-ahead needs to reach it, and each press counts as one re-check, as each arrow press did. A second type-ahead on one selector moves focus off and back first. The walk stays keyboard-only and keeps every assertion.
Regression test: the AC-23 test itself. On a Mac with region Sri Lanka it failed before the fix and passed after (the whole spec 10 of 10; AC-23 repeated three times, all passed).
Resolved: 2026-10-09 — fixed in forge-et-fe#194 (merged 2026-10-08 21:40 UTC as 23fd1b1c), test only, as the Fix above describes.
Lesson: L-143, a second instance, widened to native date inputs: their part order follows the operating system’s region on macOS Chromium, so a test reads the order rather than assuming month first.
BUG-121 — Every nav link logs “ child must be focusable” at drawer widths, because a disabled Tooltip still asks the library to check its trigger
Affects: F5-b (forge-et-fe src/ui/components/Tooltip/Tooltip.tsx, used by src/app/Sidebar.tsx’s NavRow; any other caller whose trigger cannot take focus while its tooltip is off) · Found in: the F5-h wave-1 browser run, 2026-10-08 — the Vite log of pnpm exec playwright test --project=chromium e2e/browser/users carried 52 copies of the warning · Severity: low (console noise in development builds; nothing on screen is wrong, and production builds skip the check)
Repro: run the frontend against a local backend, sign in, set the viewport to 768 × 1024 (the sidebar is then a closed drawer) and open /users. The console logs <Focusable> child must be focusable. Please ensure the tabIndex prop is passed through. once per nav link per mount (24 on one load in the browser pane).
Expected: a tooltip that is disabled asks nothing of its trigger, and the console stays clean.
Actual: Sidebar.tsx keeps each nav row’s Tooltip mounted but disabled outside the icon rail, and sets inert on the rail while it is a closed overlay. Tooltip passed disabled to the library’s TooltipTrigger but not to its Focusable, whose mount-time check (react-aria Focusable, development builds only) warns unless it is told it is disabled when its child fails isFocusable — and a link under inert fails it. The same Focusable also gave the disabled trigger tabIndex={0} and the library’s press handlers (onKeyDown, onPointerDown) it did not need; its focus handlers were already off, because useTooltipTrigger passes isDisabled to its own useFocusable.
Fix: pass isDisabled={disabled} to <Focusable> in Tooltip.tsx. Disabled, Focusable adds no tabIndex and none of the library’s press handlers, so the trigger keeps its own focusability and nothing changes for an enabled tooltip. The code comment in Tooltip.tsx still says “focus handlers” for these; the correction rides the next push to that file (forge-et-fe#189 review, Nit 1).
Regression test: Tooltip.test.tsx “asks nothing of a trigger under an inert region while disabled (BUG-121)” renders a disabled tooltip around a link inside an inert region and asserts no Focusable warning and no library tabindex. jsdom parses the inert attribute but has no inert property, which is what the library reads, so the test sets the property from a ref before the mount check runs; on the unfixed code it fails with the exact warning.
Resolved: 2026-10-08 — fixed in forge-et-fe#189 (0292e56): <Focusable isDisabled={disabled}> in Tooltip.tsx, with the regression test above. Checked in a running browser at 768 × 1024 on /users, with the sidebar a closed inert drawer: 24 warnings on one load without the fix, 0 in a fresh tab with it, and the first nav link carries no library tabindex. Found while running the F5-h wave-1 browser suite; fixed on its own branch at the lead’s request.
Lesson: L-024, another instance: a DOM-only test tier cannot see a headless library’s browser-only behaviour — here jsdom has no inert property, so Focusable’s mount check never fired in the unit tests.
BUG-120 — The AC-16 count tests compare the screen with a second read made after it, while other workers add leads
Affects: F11-b (forge-et-fe e2e/browser/leads/leads-inbox.spec.ts, the two AC-16 tests “the header count is the API total for the same filters” and “a page past the last shows Go to page N…”) · Found in: the BUG-117 browser gate (forge-et-fe#179), twice, 2026-10-07 · Severity: low — a flaky browser test · Phase: 1
Repro: run pnpm e2e:gate against a database with few new leads. The first test loads /leads, settles the list, then reads the new total with its own API call and expects the header to show it. Other workers create new leads in the same database in between, so the header shows the earlier total: “1 lead” against “3 leads” in one run, “5 leads” against “7 leads” in the next.
Expected: a test compares the screen with the answer the screen received (L-147).
Actual: the test fails whenever another worker creates a new lead between the screen’s read and the test’s own. The past-the-end test reads its count and totalPages the same way.
Fix: take the total from the screen’s own list response, not a later read.
Regression test: a probe that creates a lead between the screen’s read and the comparison fails the old test and passes the new one.
Resolved: 2026-10-07 — fixed in forge-et-fe#179 (caf605a), on the BUG-117 branch because it blocked that PR’s gate. A new support.ts helper, loadLeadsList(page, go), waits for the screen’s own list response, settles the list, and returns that response’s parsed list and URL; both AC-16 count tests compare with it, and the first also asserts its read asked for status=new. A temporary probe that created a lead in between failed the old shape 3 of 3 (for example “61 leads” expected, “60 leads” shown) and passed the new shape 3 of 3. leads-inbox.spec.ts passed 32 of 32, and the gate on caf605a passed with 1575 tests.
Lesson: L-147, extended: nor on a total read after the screen’s.
BUG-119 — The account form reads the certifications as one page of 100, in a second copy of the read BUG-114 fixed
Affects: F1-b (forge-et-fe src/lib/certifications/certificationsApi.ts listActiveCertifications, CERTIFICATIONS_PAGE_SIZE = 100; src/features/accounts/AccountFormPage.tsx the certification group and its Showing the first 100 certifications helper) · Found in: the frontend-reviewer pass on forge-et-fe#179 (the BUG-117 fix), outside its diff, 2026-10-07 · Severity: low · Phase: 1
Repro: read the path. listActiveCertifications sends active=true&size=100 and reads one page; its header says one page “lists the whole taxonomy (F1-a seeds far fewer)”. When page.totalItems is more than 100, the account form offers the first 100 and shows Showing the first 100 certifications (AccountFormPage.tsx:497).
Expected: a control filled from a paged list reads every page (L-151); the lead decided every page for the owner pickers (BUG-109) and the Tag filter (BUG-117). One read of the active certifications, not two.
Actual: a certification after the first 100 cannot be chosen on an account. The pricing feature’s copy of the same read (src/features/pricing/standardsApi.ts) reads every page since BUG-114, so the two copies now behave differently; the src/lib header already names moving the pricing copy here as a follow-up. No Empire database is near 100 certifications.
Fix: make src/lib/certifications/certificationsApi.ts read every page through readEveryPage (send page from page 2 on, pass the queryFn’s signal), remove the helper, and have the pricing form use this one read instead of its own copy, keeping BUG-114’s tests passing.
Regression test: unit rows for one page, two pages, a failing page 2 and an aborted signal; a component test on the account form where a page-2 certification is offered; the pricing tests unchanged in what they prove.
Resolved: 2026-10-07 — fixed in forge-et-fe#181 (f9017d2; review fixes 7894a66), on the lead’s go-ahead. src/lib/certifications/certificationsApi.ts listActiveCertifications(signal?) reads every page through readEveryPage, sending active=true&size=100 and page from page 2 on; useActiveCertifications passes TanStack’s signal. The account form’s Showing the first 100 certifications helper is removed. The pricing feature’s copy (listActiveCertifications, buildCertificationsPath, useActiveCertifications, key ['certifications','active']) is deleted: StandardFormDialog uses the shared hook with { enabled: open }, and its CERTIFICATION_NOT_ACTIVE handling invalidates the shared key family ['certificationOptions'], so both forms share one cache entry. This completes the follow-up F1-b WI-1.2 plan Decision 6 left open and removes the 100 cap its risk list named. Departure from the frozen F26-b Wave-1 contract: its GET /api/v1/certifications row says the pricing form sends size=100 and no other parameter; it now also sends active=true, which the endpoint already assumed when active was absent. BUG-114’s rows moved to the src/lib tests at the same strength. Regression tests: unit rows for one page, two pages, a failing page 2 and an aborted signal, hook rows for every page, enabling and a cancelled query, and an account form row where a page-2 certification is offered. Against the old code 6 failed; the other 3 cannot fail there. Vitest: 3463 passed. Not changed: the seam spec e2e/seam/standards-book.seam.spec.ts still reads the certifications without active; the request the form sends is covered by the endpoint’s declared active parameter and the live browser tests.
Lesson: L-151 (a control filled from a paged list reads every page), which names this bug. Reading every page also removed a second copy of the same read.
BUG-118 — A disabled Combobox gives its Show suggestions button the bare disabled attribute
Affects: the shared src/ui Combobox (forge-et-fe src/ui/components/Combobox/Combobox.tsx, the Show suggestions trigger near line 272) · Found in: F11-f WI-1.3 (RecordSearchPicker.test.tsx), 2026-10-07 · Severity: low · Phase: 1
Repro: render a Combobox with disabled set, then inspect the Show suggestions button beside the input.
Expected: the whole control is aria-disabled and stays in the tab order, as the component’s own note (lines 125–129) and design-system § Permission rendering require, so a keyboard or screen-reader user can reach it and read why it is unavailable.
Actual: the input is aria-disabled + readOnly, but the library’s isReadOnly renders the trigger button with the bare disabled attribute, which drops it from the tab order. The component’s test asserts only the input.
Fix: give the trigger aria-disabled and stop it opening the list when disabled, without the bare attribute; extend Combobox.test.tsx to assert the trigger has no disabled attribute. Until then F11-f’s RecordSearchPicker renders a blocked picker as a read-only Input instead of a Combobox (F11-f WI-1.3 plan, deviation 1), which can return to the Combobox once this is fixed.
Resolved: 2026-10-09 — fixed in forge-et-fe#200 (8f613b5; review fixes 2a39838). A new SuggestionsTrigger in Combobox.tsx reads the props the library hands the button through ButtonContext and, while the control is disabled, re-provides them without isDisabled and without any on… handler (today onPress/onPressStart open the list; stripping them all keeps a handler a react-aria upgrade adds from reopening it), and with aria-disabled; the ref, id, aria-expanded and labelling stay. A prop on the button alone could not do it, because the library chains its press handlers with the button’s own. Combobox.module.css gives the disabled trigger the disabled text colour and the not-allowed cursor, with no hover fill. Correction to the Actual line: the trigger is never in the tab order, enabled or not (the library sets excludeFromTabOrder, as the WAI-ARIA combobox pattern has it), so the defect was the bare attribute itself, not a lost tab stop; the input is the control’s one tab stop and was already aria-disabled. RecordSearchPicker still renders a blocked picker as a plain Input (its comment no longer cites this bug); returning it to the Combobox would change F11-f’s behaviour and is not part of this fix. Regression test: Combobox.test.tsx “disables the Show suggestions trigger without the bare attribute, and it opens nothing (BUG-118)”, which failed before the fix (“Expected the element to have attribute”) and passes after; the enabled-trigger test also asserts no aria-disabled; the review added Enter, Space, ArrowDown and screen-reader (detail 0 click) rows for the disabled trigger, of which all but ArrowDown fail when the library’s handlers are kept (ArrowDown guards a keyboard handler an upgrade could add). Vitest on 8f613b5: 4180 passed, 1 skipped; the Combobox and fields tests on 2a39838: 75 passed. Checked in Chromium and WebKit on a scratch page at 768 × 600, on 8f613b5: before, the disabled trigger had disabled and the pointer cursor; after, aria-disabled="true", no disabled, the not-allowed cursor, and a click opens no list, while the enabled trigger still opens it.
Lesson: L-188 (new — a component built on the headless library checks its disabled state on every part the library renders), with a frontend-reviewer dimension 20 pointer. BUG-121 (Tooltip) had the same cause and is in its source list.
BUG-117 — The Leads Tag filter reads the lead tags as one page of 100
Affects: F11-b (forge-et-fe src/lib/tags/tagsApi.ts listTags, TAG_OPTIONS_PAGE_SIZE = 100; src/features/leads/LeadFilters.tsx the Tag select and its Showing the first 100 helper) · Found in: the frontend-reviewer pass on forge-et-fe#177 (the BUG-114 fix), outside its diff, 2026-10-07 · Severity: low · Phase: 1
Repro: read the path. listTags('lead') sends entityType=lead&size=100 and reads one page. When page.totalItems is more than 100, the Tag filter lists the first 100 and shows Showing the first 100.
Expected: a filter filled from a paged list reads every page, or shows a cap the product decided (L-151). For the owner pickers and the Owner filter on the same screen the lead decided every page (BUG-109).
Actual: a tag after the first 100 cannot be filtered on. The cap comes from F11-b Wave-1 contract client-side fact 6 and the WI-1.2 plan’s D9, whose reason is that more than 100 active lead tags is not expected in V1. That is a premise about today’s count, not a product decision. tagsApi.ts is in src/lib and its header plans reuse by the account screens, so the cap would spread. No Empire database is near 100 tags.
Fix: unless the lead keeps the cap as a product decision in the F11-b spec, read every page through src/lib/api/everyPage.ts readEveryPage, as BUG-114 did: row 2 declares page, so send entityType, size and, from page 2 on, page. Pass the queryFn’s signal, remove the Tag helper, and record the departure from contract fact 6 in this entry’s Resolved: line.
Regression test: unit rows for one page, two pages, a failing page 2 and an aborted signal; a component test where a tag only on page 2 is offered in the Tag filter.
Decision: the lead, 2026-10-07. The Tag filter reads every page, as decided for the Owner filter in BUG-109. No cap.
Resolved: 2026-10-07 — fixed in forge-et-fe#179 (26e03de; review fixes 0d70246). listTags(entityType, signal?) reads every page through readEveryPage, sending entityType, size=100 and page from page 2 on, and useLeadTags passes TanStack’s signal. The Showing the first 100 helper, FIRST_HUNDRED and tagsTruncated are removed; the page size is now a private TAG_PAGE_SIZE. This departs from F11-b Wave-1 contract client-side fact 6 and the WI-1.2 plan’s D9, which list one page; the tagsApi.ts header says so. The e2e helpers readTags and readLeadTags still read page 1, which is enough to find a seeded tag because no browser test creates tags. Regression tests: unit rows for one page, two pages, a failing page 2 and an aborted signal, a cancelled-query hook row, and a LeadsScreen row where a page-2 tag is offered and filters the list. Against the old code 5 of the 7 failed; the other two cannot fail there and guard the one-page path and the signal. Browser gate on caf605a: 1575 passed, 5 skipped, 0 known failures, with ET_E2E_WORKERS=3 (L-152). The first gate run, on a machine at load average 42, failed 63 WebKit size-matrix tests on timeouts and the AC-16 count test; the second, with 3 workers, failed only the AC-16 count test, filed and fixed as BUG-120 on the same branch. The review found the account form’s certification read capped the same way, filed as BUG-119.
Lesson: L-151 (a control filled from a paged list reads every page), which names this bug.
BUG-116 — A cancelled or refused mailbox consent lands on the SPA’s home with no banner, or with an older connect’s result
Affects: F20-a (forge-et-be auth/web/ContributedAuthorizationLandingController.java:85-92; Spring Security 7.1.1 OAuth2AuthorizationCodeAuthenticationProvider.java:78-87) · Found in: the BUG-113 fix (forge-et-be#210), its fix agent and the backend-reviewer, 2026-10-07 · Severity: low · Phase: 1
Repro: start a mailbox connect from the Email screen, then cancel or decline on Microsoft’s consent screen. Entra returns ?error=access_denied&state=…. A state mismatch (invalid_state_parameter) fails the same way. Spring’s provider throws OAuth2AuthorizationException before it calls the token client, so CapturingTokenResponseClient (which records BUG-113’s failed outcome) never runs. The grant filter redirects to /authorized/microsoft-mailbox?error=access_denied&…. The landing controller finds no recorded outcome and sends the browser to the SPA base, logging at DEBUG.
Expected: a decision for the lead: whether a cancel should show the Email screen’s “We could not connect your mailbox. Nothing was changed. Try again” banner, or a cancel-specific message, or nothing. Either way it should land on the Email screen, never on an older connect’s result.
Actual: the user lands on the SPA’s home with no message. If an earlier connect recorded an outcome and the user never landed, this landing shows that older outcome instead.
Fix: after the lead’s decision, either (a) ContributedRegistrationRequestResolver records the contributor’s refusal URL as the default outcome when it starts a contributed flow, and the exchange overwrites it on success; or (b) the landing controller looks up the contributor for {registrationId} in its fixed map and, when the request carries error and no outcome is recorded, redirects to that contributor’s refusal URL (still checked by isWithinRedirectBase). Either way, clear any older outcome when a new flow starts. A UI change if the message changes: browser-validated.
Regression test: an IT through the real filter chain: a callback with error=access_denied lands on the decided URL, and an outcome left by an earlier flow is not shown.
Decision: the lead, 2026-10-07. A cancel, and every other refusal before the code exchange, lands on the Email screen with the existing “We could not connect your mailbox. Nothing was changed. Try again” banner (/email/mailbox?connect=failed). No new connect= value, no frontend change, no change to the F20-a redirect contract. One WARN with the error code only.
Resolved: 2026-10-07 — fixed in forge-et-be#211 (14a0e03; review fixes d2a414d). Option (b): with no recorded outcome, ContributedAuthorizationLandingController sends a callback that carries error, code or state to the contributor’s refusal() URL from its fixed map, still checked by isWithinRedirectBase, and logs Mailbox Consent Refused - [registrationId=…] [error=…]. The error code is logged only when it matches [a-z_]{1,64}, otherwise as not-an-error-code; a callback with a state logs authorization_request_not_found. ContributedRegistrationRequestResolver clears any recorded outcome when a permitted contributed flow starts (getSession(false)), so an older connect’s result is never shown. (a) was not chosen: a default outcome could not tell a refusal from BUG-113’s failed exchange, which already logs. Correction to the Repro: in Spring Security 7.1.1 a real state mismatch never reaches the provider’s invalid_state_parameter; the grant filter does not match the callback, which passes straight to the landing with its code or error and state. Both shapes land on the banner. The sign-in registration is unchanged. Accepted consequence, in the PR: Back after a successful connect can make Entra answer the old authorize URL with a fresh code for a used state; that callback now shows the failed banner, and “Nothing was changed” is true for it. Regression tests: 8 rows of MailboxConnectFlowIT failed before (7 landed on the SPA base, the earlier-outcome row showed connect=success) and pass after; a lower-case 65-character error code row fails when the limit is removed; EntraLoginIT rows pin the cancelled sign-in. Local build on d2a414d: 5466 tests, 0 failures, 2 skipped.
Lesson: L-149, extended. A refusal that fails before the bounded call never reaches the client that records the outcome, so the failure path covers it at the landing too.
BUG-115 — Flaky test: the AC-23 deactivated-owner browser test timed out once under load (macOS)
Affects: F11-b (forge-et-fe e2e/browser/leads/leads-writes.spec.ts:448, “AC-23, AC-21 (smoke): an owner deactivated while the form is open…”) · Found in: the BUG-112 fix’s regression runs (forge-et-fe#174), 2026-10-07 · Severity: low — a flaky browser test · Phase: 1
Repro: on macOS, with a forge-et-be main backend on port 8081 against a throwaway database holding more than 120 active users, run ET_BACKEND_PORT=8081 pnpm exec playwright test e2e/browser/leads --project=chromium --reporter=line with default workers. One run in four took 1.4m instead of about 50s, and in it this test hit the 30s test timeout with no assertion failure. The next run cleared test-results/, so the step it stopped on is not known. It passed alone three times (--repeat-each=3) and in the other three folder runs. Nothing else ran on the machine apart from that backend.
Expected: the test passes however loaded the machine is.
Actual: under load it can take longer than the 30s budget. It waits for every page of two directory reads through nextDirectoryRead (added by BUG-109, up to 10s per later page), deactivates a user live and waits for the re-read. That is the likely cost, but it is not yet shown.
Fix: reproduce under forced load (16 parallel copies of the file, as BUG-107 did), keep the error context, and find the step that runs out. Then either give the test a longer timeout that names why (test.slow()), or cut its waits. A green re-run is not the proof (L-107).
Regression test: the file run as 16 parallel copies for three rounds with no failure.
Resolved: 2026-10-07 — fixed in forge-et-fe#178 (50f6fb0; review fixes fc915d3). The cause was a race in the test, not a slow step. The Owner filter and the form’s owner picker share one user-directory query. When Edit was clicked while any page of the filter’s read was in flight, or before its answer was stored, the picker joined that read and sent no page 1 of its own, and nextDirectoryRead waited for one until the 30 s timeout. Before the fix, 3 of 16 parallel copies hung at that wait; with page 2 delayed 3 s (page.route) the old test timed out 3 of 3 and 4 of 4 times. The review found the same race in the AC-21 add-a-lead test, in a narrower window. A new support.ts helper, settleOwnerFilter, waits for every page of the filter’s read and for an option from it on screen before the form opens; both tests use it. AC-21’s exact picker comparison is now polled, because the picker shows the filter’s answer until the form’s own is stored. AC-23 also gets test.slow(): with the race fixed it still took up to 42 s on an 8-core Mac under 16 copies, every step completed. Forced proof (L-107): with each page-2 body held 2 s, AC-21 and AC-23 timed out 3 of 3 on main and passed 3 of 3 with the fix. Regression criterion, changed: AC-21 alone as 16 parallel copies for three rounds passed 48 of 48; AC-23 alone passed 45 of 48, the 3 failures all settleLeadsList’s 5 s first-load wait with 16 browsers loading at once. The whole-file criterion is not met on 8 cores, before or after the fix (39, 41, 26 and 52 failures in about 265 tests per round, at load averages of 33 to 80); every other timeout’s trace stopped inside a running step, none at a directory-read wait, so it is machine overload, not this defect. The leads folder passed 68 of 68 with default workers. Noted, not changed: the OWNER_NOT_ELIGIBLE test’s re-read wait (waitForResponse(directoryResponse())) accepts any directory page, so it can pass without proving the re-read; it cannot fail or hang.
Lesson: L-101, extended. A test that waits for the request an action sends must first let a shared query’s in-flight read end and show on screen, because the app joins that read and sends none.
BUG-114 — The pricing screens read the certifications and a standard’s costs as one page of 100
Affects: F26-b (forge-et-fe src/features/pricing/standardsApi.ts buildCertificationsPath, the certification checkbox group in StandardFormDialog; src/features/pricing/costsApi.ts row 8, page=1&size=100) · Found in: the frontend-reviewer pass on forge-et-fe#174 (the BUG-111 fix), outside its diff, 2026-10-07 · Severity: low · Phase: 1
Repro: read the two paths. buildCertificationsPath sends size=100 and nothing else, and the form offers that one page as checkboxes, with no cap shown. Row 8 reads a standard’s costs at page=1&size=100. costsApi.ts’s header says this makes the ids “the whole set”, and they are sent back as expectedCostIds.
Expected: a control or check filled from a paged list reads every page, or shows a cap the spec names (L-151).
Actual: with more than 100 active certifications, the ones after the first 100 cannot be chosen, and nothing says so. With more than 100 cost rows on one standard, the expectedCostIds set sent back is incomplete, so the server’s check that the client saw every cost no longer means that. Neither count is near 100 today.
Fix: read every page through src/lib/api/everyPage.ts readEveryPage (shared since BUG-111), honouring TanStack’s signal, for both reads. Correct costsApi.ts’s header. Check the costs contract (F26-b contract fact 7) before changing what expectedCostIds carries.
Regression test: unit rows for one page and two pages on each read, and a component test where a certification only on page 2 is offered.
Resolved: 2026-10-07 — fixed in forge-et-fe#177 (3b237e1; review fix 5fec7ea). listActiveCertifications and listCurrentCosts read every page through readEveryPage, and both hooks pass TanStack’s signal. The certification read sends size=100 and page from page 2 on; row 8 sends page=N&size=100. expectedCostIds is still every row id the list holds, as F26-b contract fact 7 requires; a set that shifts between pages is answered 409 COST_CHANGED, which re-reads both reads. This departs from the frozen contract’s wording only: row 120 says the form sends size=100 and nothing else, and fact 7 says row 8 is read at size=100. The costsApi.ts header and the two dialog headers are corrected. Regression tests: unit rows for one page, two pages, a failing page 2 and an aborted signal on each read, a component test where a page-2 certification is offered, and hook rows proving a cancelled query reads no further page. Against the old code 7 failed. Browser gate on 5fec7ea: 1300 passed, 5 skipped, 0 known failures. The review found the same shape in the Leads Tag filter, filed as BUG-117.
Lesson: L-151 (a control filled from a paged list reads every page), extended: a check whose ids are sent back as the whole set is the same case as a control, and finding one such read is the cue to look for the rest.
BUG-113 — A mailbox connect whose token exchange fails sends the user back with no message and logs nothing above DEBUG
Affects: F20-a (forge-et-be auth/config/CapturingTokenResponseClient.java, auth/web/ContributedAuthorizationLandingController.java; the Email screen’s connect result) · Found in: the backend-reviewer pass on forge-et-be#208 (the BUG-108 fix), outside its diff, 2026-10-07 · Severity: low · Phase: 1
Repro: connect a mailbox from the Email screen (/oauth2/authorization/microsoft-mailbox), and make the token endpoint fail the code exchange: a stall past the 5s bound (BUG-108), a 5xx, or an invalid_grant. OAuth2AuthorizationCodeGrantFilter redirects to /authorized/microsoft-mailbox?error=invalid_token_response&error_description=…. CapturingTokenResponseClient.getTokenResponse (~line 108) records no outcome when its delegate throws, so ContributedAuthorizationLandingController sends the browser to the SPA base.
Expected: a failed connect lands on the Email screen with the connect-failed message and a reference id (the ?connect= result the success path uses), and the backend logs it at WARN with the error code (never the code, tokens or assertion).
Actual: the user lands on the SPA’s home with no banner, so they cannot tell the connect failed, and the backend logs it at DEBUG only, so an Entra outage during connects leaves no trace in the logs.
Fix: record a failed outcome in CapturingTokenResponseClient when the delegate throws, map it in the landing controller to the existing connect-failed result, and log it at WARN with the OAuth2 error code and the correlation id. Check sign-in’s failure path does the same: it already redirects to /login?login=failed.
Regression test: a slice test where the token client throws OAuth2AuthorizationException(invalid_token_response). It asserts the redirect carries the connect-failed result, one WARN line names the error code, and no line or message contains the code, a token or eyJ.
Resolved: 2026-10-07 — fixed in forge-et-be#210 (9bfb4f2c; the review fix ce8f68d7). CapturingTokenResponseClient.getTokenResponse now catches OAuth2AuthorizationException, the one type the code-grant filter turns into an ?error= redirect. It writes one WARN, Mailbox Consent Exchange Failed - [registrationId=…] [error=<code>] [cause=<root-cause class | none>], records the contributor’s fixed refusal URL (/email/mailbox?connect=failed), and rethrows. The landing controller is unchanged, and the Email screen already shows that result as “We could not connect your mailbox. Nothing was changed. Try again”. The backend-reviewer found that sign-in’s own failed exchange was not logged either, so EntraLoginConfig.failureHandler() now writes one Entra Sign-In Exchange Failed - [error=<code>] [cause=…] WARN when the token client’s exception is in the cause chain. An unsignable assertion is excluded, because it already logs at ERROR. Neither line carries the error description, a message or the exception. Departs from the Expected line: no reference id is shown, because the frozen F20-a Wave-1 redirect contract allows only connect=failed. No audit event is recorded either: no data is written, the sign-in leg’s exchange failure records none, and AD #36’s list does not include it. The reviewer agreed with both. Regression tests: CapturingTokenResponseClientSpec and a MailboxConnectFlowIT row through the real filter chain (with the fix reverted, 4 of 7 and 2 of 14 failed, landing on http://localhost:5173 instead of …?connect=failed), and two EntraLoginIT rows for the sign-in WARN (2 of 18 failed before). The fakes put the code, client_assertion and eyJ into the error description, so logging it would fail them. A cancelled consent fails before this code; that is BUG-116.
Lesson: L-149 (an outbound call’s failure path is part of its contract: bound it, and make its failure visible) — BUG-113 is the failure half of BUG-108’s timeout, which turned a hang into a refusal that nobody saw. Added to L-149’s sources.
BUG-112 — Flaky test: a leads permissions browser test reads the newest lead’s name while another worker’s keyboard test is renaming it (macOS)
Affects: F11-b (forge-et-fe e2e/browser/leads/leads-permissions.spec.ts:103, AC-26 for the executive; e2e/browser/leads/leads-a11y.spec.ts:155) · Found in: the BUG-109 fix’s leads browser run (2026-10-07) · Severity: low — a flaky browser test · Phase: 1
Repro: on macOS, with a forge-et-be backend on port 8081 against a throwaway database (125 active lead.view users), and the two specs running in parallel workers, run ET_BACKEND_PORT=8081 pnpm exec playwright test e2e/browser/leads --project=chromium (log /private/tmp/claude-502/bug109-leads-before.log). It was seen once in three runs, while nothing else ran on the machine apart from that backend.
Expected: AC-26’s executive check passes however the workers interleave.
Actual: the executive opens the newest lead (the first row), and expected its heading “E2E lead keyboard … muxlxykuj4djk”; it got the same name plus " edited". leads-a11y.spec.ts:155’s keyboard test, in another worker, renames the newest lead it created. The executive’s test reads a lead it does not own, which another file changes.
Fix: make AC-26’s executive case open a lead its own test created (a run-unique company name, found with the list’s search), not the newest row, so no other file’s test can change it.
Regression test: the leads folder run in parallel three times with no failure; the executive case names its own lead.
Resolved: 2026-10-07 — fixed in forge-et-fe#174 (2a6db47; the review fixes ee6fe3b, 808dfc7). In leads-permissions.spec.ts, the executive’s AC-26 case now opens a lead its own test created: the all-permissions user creates E2E lead <runToken> executive, and the case finds it by search and opens it by id. The rep’s case creates its own two leads. The frontend-reviewer found the same race in the other file, so leads-a11y.spec.ts’s keyboard test now creates its own lead and reaches it with a typed search; it Tabs forward from the All pill, because the status pills come before the search box. The AC-14a option-failure cases in leads-inbox.spec.ts now open the rep’s own Swedish lead. No assertion was weakened. Regression runs on macOS, against a backend on a throwaway database, in parallel: before the keyboard-test fix, the leads folder passed 68/68 three times. After it, it passed 68/68 in three of four runs. The other run timed out one test in a file this fix did not touch, under load, filed as BUG-115.
Lesson: L-147 (a browser test never relies on what a shared list’s first page or newest row holds) — extended with this bug.
BUG-111 — The Users screen’s manager picker offers only the first 100 active users
Affects: F5-c (forge-et-fe src/features/users/usersApi.ts MANAGER_CANDIDATES_PATH = /api/v1/users?active=true&size=100, usersQueries.ts useManagerCandidates) · Found in: the BUG-109 fix (forge-et-fe#172, the fix agent and the frontend-reviewer), 2026-10-07 · Severity: low · Phase: 1
Repro: on a database with more than 100 active users, open a user in the Users screen and open the Manager picker.
Expected: every active user can be chosen as a manager. This is the same rule the lead set for the owner pickers in BUG-109.
Actual: the picker lists the first 100 active users by name; anyone after them cannot be chosen. The browser tests pin the read to avoid it (e2e/browser/users/support.ts pinManagerCandidates, edit.spec.ts:67, activation.spec.ts:80). No Empire user meets it at 10–20 users.
Fix: read every page, as listUserDirectory does since BUG-109 (pages 2 to totalPages in order, kept once per id, a failed page rejects the read, a bounded page count, TanStack’s signal). Share the paging helper rather than copying it if both can use one. Then revisit the pinned browser reads.
Regression test: unit rows for one page, several pages and a failing page, and a component test where a user only on page 2 is offered as manager.
Resolved: 2026-10-07 — fixed in forge-et-fe#174 (1ee46e6; 839671d, 38d0815). A shared helper, src/lib/api/everyPage.ts readEveryPage, now holds the paging BUG-109 built: page 1, then pages 2 to totalPages in order, kept once per id, a failed page rejecting the read, more than 50 pages refused, and TanStack’s signal honoured. listUserDirectory uses it with unchanged behaviour and tests, and listManagerCandidates reads every page through it. The bug was worse than filed: the user editor marks a current manager Inactive when the candidates read does not list them, so a manager past the first 100 showed as Inactive. That is fixed too. Regression tests: unit rows for one page, three pages, a failing page 2, an aborted signal, a user on two pages kept once, and the 50-page bound. Without the helper wired in, 5 failed. A component case covers a page-2 user who is offered and not marked Inactive. A pinned browser test in edit.spec.ts serves a 140-user, two-page answer and proves “Pinned Manager 140” is offered; against the old single-page read it failed (“the manager picker offers user 140, only on page 2”). Its read sequence allows a page-1 read that restarts after the dev server’s double mount cancels it (seen as ['1','1','2']). see-and-add.spec.ts now asserts the IT Admin option with no fallback. The users browser folder passed 58/58 against a database with more than 120 active users.
Lesson: L-151 (a control filled from a paged list reads every page, or says where it stops), which names this bug. The review found the same shape in the pricing reads, filed as BUG-114.
BUG-110 — Flaky tests: three Users manager-list browser tests time out clicking an option that keeps re-rendering, under full-suite load
Affects: F5-c (forge-et-fe/e2e/browser/users/edit.spec.ts:315 — “AC-7: choosing a report of the user as their manager is refused on the manager field”; e2e/browser/users/users-a11y.spec.ts:517 and :525 — the editor’s and the create dialog’s manager list at three sizes) · Found in: F1-b WI-1.3’s full Playwright run on forge-et-fe feature/F1-b-WI-1.3-e2e-accounts at ccea3f2, 2026-10-07 · Severity: low · Phase: 1
Repro: Windows 11. forge-et-be origin/main (a20f9469) on the local profile on 8080, against an isolated compose stack on a fresh database; pnpm exec playwright test (every project, 1550 tests, about 30 minutes) in the forge-et-fe worktree. All three tests failed in chromium (1542 passed, 3 failed, 5 skipped). Running pnpm exec playwright test e2e/browser/users/edit.spec.ts e2e/browser/users/users-a11y.spec.ts --project=chromium on the same head straight afterwards passed 41 of 41. An earlier size-matrix run on the same machine, made while a full pnpm test and a build ran beside it, also timed out widely in WebKit (the BUG-045 / BUG-079 pattern); this run had nothing else running.
Expected: chooseManager clicks the named manager option once the list has settled.
Actual: locator.click resolves the option, then reports “element was detached from the DOM, retrying” again and again until the 30-second test timeout: the manager list keeps re-rendering under load. By the run’s end the database held more than 80 users (the option’s data-key was 82), from the users and accounts specs’ own fixtures.
Cause (2026-10-08): the tests open the manager list before the manager-candidates read has landed. chooseManager (e2e/browser/users/edit.spec.ts) and the list-opening helpers of users-a11y.spec.ts wait only for the editor’s h1 and an enabled role select (expectEditorLoaded); the candidates read is separate (useManagerCandidates, UserEditorForm.tsx). Until it lands the options are No manager plus, when the user has one, their current manager appended last; when it lands, managerOptions is rebuilt and the list’s options are replaced, so an option Playwright resolved is detached (the retry loop until the 30 s timeout) or a click lands on the old node and does not take (the 2026-10-07 gate run on 9939951: “No manager” clicked, the field kept the old manager). Under load, and since BUG-111 made the read page through every page, the read lands later, which is why only full runs failed. The product already signals the state: the manager Combobox is busy={candidates.isFetching}, which renders aria-busy="true" on its input. A second cause, found while fixing the first: with the wait in place, chooseManager still lost about one choice in two full Users runs at two workers. A probe in the page recorded the sequence: the list opened fully inside the viewport, Playwright scrolled the page about 230 ms later as part of clicking the option (no page script called it), react-aria closed the open list on that page scroll without selecting, and the click arrived after the list had gone.
Fix: test-only, in two parts. (1) Before opening the manager list, every Users browser test that opens it calls readyManagerList (e2e/browser/users/support.ts), which waits for the manager input’s aria-busy="false" (a settled candidates read) and brings the field to the middle of the viewport. (2) chooseManager chooses the option with the arrow keys and Enter, waiting after each key for the input’s aria-activedescendant to move by one option, instead of clicking it; a key press does not scroll the page. No product change: a person opening the list early sees it fill in, and a person does not scroll between opening the list and clicking a visible option.
Regression test: a browser test that holds the manager-candidates read, opens the editor, releases the read, then chooses a manager and asserts the field’s value — it fails when the wait is removed (the click lands before the list is rebuilt) and passes with it; plus the three original tests green in a full run on Windows with ET_E2E_WORKERS=3.
Possibly related: BUG-106 (the leads suite fails once a local database holds more than 100 active users who can see leads) — both appear only after several suite runs on one local database.
Filed as BUG-109 and renumbered BUG-110: main’s BUG-108 and BUG-109 merged first (L-057).
Also seen on macOS, 2026-10-08: F5-h wave 1’s browser gate (pnpm e2e:gate, forge-et-fe feature/F5-h-wave-1 at 44256c4, against an isolated backend whose database had served two earlier full runs) failed users-a11y.spec.ts:525 at 320 × 640 in Chromium alone (1758 passed, 1 failed, 5 skipped). The symptom differs from Windows: candidateOption found no option because the list had already closed — the page snapshot shows no listbox, and the live region’s last line was “1 option available”, so the list opened with the pinned candidate and then closed. The same test, filtered and run 10 times on the same head straight afterwards, passed 39 of 39. F5-h changes no Users-list or manager-field code.
Also seen on Windows, 2026-10-09, in a second test: F1-d wave 1’s browser gate (pnpm e2e:gate, forge-et-fe feature/F1-d-wave-1 at 243f92b, 3 workers, against the F1-d wave backend on a database that had served several full runs) failed edit.spec.ts:319 (the refused-report test above, at its new line) with the same detached-option loop, and edit.spec.ts:289 — “AC-7: clearing a manager sends managerId null, and setting it again shows the manager in the list” — where the option click after the first save left the field at No manager (2496 passed, 3 failed, 5 skipped; the third was F11-f’s leads-drill-in at WebKit 1024, which passed 10 of 10 alone). Run alone, edit.spec.ts with 2 workers failed :289 again (16 passed, 1 failed); :289 alone with 1 worker failed on 98c7d56 — a head whose first gate had passed it — and passed on 243f92b. So :289 fails on both heads and is not caused by F1-d, which changes no Users code. A delta of this bug, not a new one.
A likely cause, seen 2026-10-08 (F11-f WI-1.6 run, Windows): edit.spec.ts:289 failed the same way, and its call log shows the step that closes the list: Playwright resolved the option, ran “scrolling into view if needed”, then reported “element is not visible” and “detached from the DOM”. react-aria closes an open combobox list when any ancestor of its input scrolls (useCloseOnScroll), so Playwright’s own scroll-into-view can close the list it is about to click. This also fits the macOS symptom above (the list opened, then closed). The failures grew with the number of users in the test database: at 242 active users four Users tests failed, and on a freshly reset database none did; the same spec failed on main too (F11-f findings F-005, L-165). A fix to try: scroll the editor so the field is in view before opening the list, and wait for the list to settle; or check the list’s scroll container, not the page, holds the option. Related: BUG-124 (the uncontrolled Combobox resets its text on that same close).
Resolved: 2026-10-08 — fixed in forge-et-fe#184 (19b81e2, 8f6daf7, 7a1ba02). readyManagerList waits for aria-busy="false" and centres the field, and every Users spec that opens the manager list calls it (edit.spec.ts, users-a11y.spec.ts, journey.spec.ts, see-and-add.spec.ts); chooseManager chooses by keyboard. The regression test, edit.spec.ts › “BUG-110: with the manager candidates read held, chooseManager opens the list only once it has landed, and the choice saves”, holds the read, records the input’s aria-busy each time a list opens, and saves the choice; with the wait removed it failed (“the manager list was never opened while the candidates read was in flight”). A component test in UserEditorRoute.test.tsx shows a choice made after the candidates land is kept; it passes on unchanged product code. On Windows, with nothing else running: the Users folder passed 59 of 59 in three runs at two workers (before the keyboard change, the same run failed one or two manager tests in some runs), and edit.spec.ts passed 54 of 54 with --repeat-each=5.
Finished on Tharinda’s machine on 2026-10-12, because Windows could not run the browsers: main merged in twice and two review nits applied (80c2f98: the held-read test deactivates its user). Evidence, macOS: the four Users specs repeated three times with three workers, 141 passed, 0 failed, on a database that had served several full runs; pnpm e2e:gate passed at 80c2f98 and at the merged head 5a33a58 (3175 passed, 0 failed each). The repeated Windows run the regression line names was not made.
Lesson: L-155 (new — a browser test opens a list only once its options have loaded, and chooses an option of an open react-aria Combobox list with the keyboard), with a row in .forge/test-strategy.md § Flaky Tests.
BUG-109 — The owner picker and the Owner filter offer only the first 100 users who can see the records
Affects: F11-b (the leads Owner filter LeadFilters.tsx, and the owner picker in LeadFormDialog.tsx), F1-b (the account owner picker, which reads the same directory with account.view), through the shared src/lib/userDirectory/ (userDirectoryApi.ts, useOwnerPicker.ts) and src/ui/fields/OwnerPicker.tsx · Found in: the BUG-106 fix (forge-et-fe#168), raised as a product question; filed by the lead’s decision (2026-10-07) · Severity: low · Phase: 1
Repro: on a database with more than 100 active users whose role holds lead.view, open the Leads screen’s Owner filter, or New lead / Edit lead’s owner picker. listUserDirectory reads GET /api/v1/user-directory?permission=lead.view&size=100, one page (USER_DIRECTORY_PAGE_SIZE = 100; the backend’s MAX_PAGE_SIZE is 100), ordered by display name.
Expected: a person can find and choose any active user who can own the record, however many there are (the lead’s decision of 2026-10-07).
Actual: only the first 100 by name are offered; anyone after them cannot be chosen as an owner or filtered by. useOwnerPicker adds the record’s current owner when the page leaves them out, but nobody else. userDirectoryApi.ts states the premise: “one page at size=100 lists the options (Empire has 10–20 users, client-side fact 6)”. It holds today, so no user meets this yet; it fails in local databases that the browser suites have filled (BUG-106).
Fix: to be decided with the plan, because it changes the picker’s interaction. One option is a searchable picker that sends the typed text as the directory’s q (debounced, as the list searches are) and shows the first page of matches, keeping the current owner. Another is reading every page when the first says totalPages > 1. Either way, F11-b’s spec (AC-13 and the owner picker’s ACs) and F1-b’s need a revision entry for the changed behaviour, and the design system’s picker pattern may change too. Browser-validated, as a UI change (.claude/rules/tracker.md § Bug Tracking, step 4).
Regression test: component tests on OwnerPicker and the Owner filter with a directory of more than 100 users, showing a user past the first page can be found and chosen. A browser test does the same against a database seeded with more than 100 users who can see leads.
Resolved: 2026-10-07 — fixed in forge-et-fe#172 (899eb2c, 5de0b6b; the review fixes cdb9c47, c5df01f). The lead chose to read every page (Option B), with no change to how the picker works. listUserDirectory reads page 1 as before; when totalPages > 1 it reads pages 2 to totalPages in order and returns one list, kept once per id. A failed page rejects the whole read, more than 50 pages (5,000 users) rejects after page 1, and a cancelled query stops before the next page (TanStack’s signal). “Showing the first 100” is removed from the owner picker and the leads Owner filter, because the list is no longer cut short; the Tag filter keeps its own. F1-b’s account screens read through the same hooks, so they get the fix with no change (F1-b spec Q-7); the note in .forge/plans/F1-b/F1-b-decomposition-plan.md § Notes and a comment on forge-et-fe#171 tell F1-b what to do on merging main. Regression tests: unit rows for one page, three pages in order, exactly 50 pages, more than 50, a failing later page, an aborted signal and q (before the fix, 9 failed, for example “expected [1 request] to deeply equal [3]”); component tests in LeadFormDialog and LeadsScreen with a user only on page 2; and a pinned browser test in leads-inbox.spec.ts that serves a 140-user, two-page directory and proves the Owner filter and New lead’s picker offer user 140, and that choosing them sends ownerId. Against the old single-page code that browser test failed with “the directory read’s page 2 of 2 did not arrive within 10000 ms”. AC-21 now compares the picker with every page the form read. In the browser, against a throwaway database holding 125 users who can see leads, the Owner filter offered 127 options (Anyone, Unassigned and 125 users) after reading …&size=100 and then …&page=2. Choosing the 126th option sent ownerId=7 and set ?owner=7, New lead’s picker offered all 125, and neither control showed “Showing the first 100”.
Lesson: L-151 (new — a control filled from a paged list reads every page, or says where it stops), with a frontend-reviewer row. BUG-111 has the same cause in the Users manager picker.
BUG-108 — The Microsoft sign-in and mailbox-consent code exchanges have no read timeout
Affects: FS-6 / F5-a (EntraLoginConfig), F20-a (ContributedOAuthClientConfig) · Found in: the backend-reviewer pass on forge-et-be#207, the BUG-101 fix (2026-10-07), outside that PR’s diff · Severity: low · Phase: 1
Repro: read the two authorization-code token clients: auth/config/EntraLoginConfig.java:238 (sign-in) and auth/config/ContributedOAuthClientConfig.java:187 (mailbox consent) use Spring Security’s default RestClient, which sits on HttpComponentsClientHttpRequestFactory with no read timeout. Point the Entra token endpoint at a server that accepts the connection and never answers.
Expected: every outbound HTTP call has a connect and a read timeout the code sets (L-149), as the refresh client has had since BUG-101’s fix (RefreshTokenResponseClients.tokenEndpointClient).
Actual: a stalled Entra token endpoint holds the sign-in or consent request’s thread with no bound. No database lock is held during either exchange, so nothing else waits on it, which is why this is low.
Fix: give both exchanges’ token clients the same RestClient that RefreshTokenResponseClients.tokenEndpointClient builds (the converters, the error handler and the timeouts), shared rather than copied. Test each one with a stalled loopback server, in the shape of RefreshTokenResponseClientsSpec.
Resolved: 2026-10-07 — fixed in forge-et-be#208 (9c2fd2fe; the review fix 4f5605e8). A package-private auth/config/TokenEndpointClients now holds BUG-101’s builder and its 3s connect and 5s read constants. It also offers authorizationCodeClient(), which EntraLoginConfig.entraTokenResponseClient() starts from. The PS256 private_key_jwt assertion converter is unchanged. The consent leg wraps that bean, so it is bounded too, and its no-token-client fallback builds the same client instead of Spring’s unbounded default. The backend-reviewer compared the requests over the wire: the bodies, Accept and Content-Type are identical to Spring’s default client’s. Only the transport differs (HTTP/2 over TLS, no redirect following), and the old client had re-POSTed the code and client credentials to a 307’s Location. Regression test (AuthorizationCodeTokenClientsSpec, a loopback HttpServer, 11 rows): the wired factory’s timeouts for sign-in, consent and the fallback; a 200 still parses, with grant_type, the code, client_assertion_type and a PS256 header whose x5t#S256 matches the certificate; invalid_grant still maps as before; and a stall fails as invalid_token_response after the 5s bound, with no code, client_assertion or eyJ in any message. Before the fix: 11 tests, 5 failed (the factory was HttpComponentsClientHttpRequestFactory; both stall rows hit the 30s method timeout). Before merge, a real Microsoft sign-in and mailbox connect on a laptop (local,entra-laptop, the branch’s head 4f5605e8) checked the HTTP/2 exchange with login.microsoftonline.com that the loopback specs cannot. CRM01 signed in (Identity Linked, auth.signed_in), and the mailbox connected (mailbox.connected, ACTIVE). The first token refresh then ran (mailbox.token_rotated) and the back-fill stored a message, with no WARN or ERROR in the log. The IT Admin role was given email.manage locally for the check. The JWKS fetch during id-token validation was checked and is bounded by the library at 30s (NimbusJwtDecoder’s RestTemplateWithDefaultTimeouts). The review also found that a failed mailbox connect shows nothing and logs only at DEBUG; that is BUG-113.
Lesson: L-149 (a library-built HTTP client has no timeout until the code sets one), which already names this bug; its source list now adds the fix.
BUG-107 — Flaky test: a BUG-083 cache test counts the data reads before the second user’s read has been sent, so it fails under load (Linux CI)
Affects: F5-d (forge-et-fe src/app/sessionQueryCache.test.tsx, the case “is empty at sign-out, and the next user in the tab sees none of the first user’s reads”, added by BUG-083’s fix in forge-et-fe#132) · Found in: F5-e wave 2, the Fast lane of forge-et-fe#165 (run 37518272525, Unit and component tests step, head 8e96ed0, 2026-10-07) · Severity: low — a flaky unit test; a red Fast lane holds a PR until it is re-run · Phase: 1
Repro: in CI, on GitHub’s Linux runner, the unit step took 348 s and the case failed at sessionQueryCache.test.tsx:179, expect(dataReads()).toBe(2) → expected 1 to be 2; the PR changed only e2e/ files. On macOS, in a checkout of forge-et-fe at 8e96ed0: pnpm exec vitest run src/app/sessionQueryCache.test.tsx passed 15 times out of 15 on an otherwise idle machine; run as 16 parallel copies, three rounds, it failed 2 times out of 48 with the same assertion.
Expected: the case passes however loaded the machine is.
Actual: after the second user signs in, the case waits for loading to render and then reads the data-read count at once. The second user’s query function can be called after that render, so under load the count is still 1 when it is read.
Fix: wait for the count, not just the render: await waitFor(() => expect(dataReads()).toBe(2)), or assert the count after the second user’s data has rendered. Check the file’s other cases for the same read-after-render shape.
Regression test: the case itself, run as 16 parallel copies for three rounds (as in the repro) with no failure.
Resolved: 2026-10-07 — fixed in forge-et-fe#168 (ab1e249). Both cases that read the data-read count right after loading rendered now wait for it with await waitFor(() => expect(dataReads()).toBe(2)): the BUG-083 case named here, and the BUG-085 background re-read case, which had the same shape. The flake did not reproduce before the fix (0 of 96 at 16 parallel copies, 0 of 32 at 32), so the race was forced instead (L-107). With the probe’s read delayed 50 ms in a throwaway checkout, both cases failed on main with expected 1 to be 2, and passed (5 of 5) with the fix. Unforced, 16 parallel copies × 3 rounds after the fix: 0 of 48 failed.
Lesson: L-107 (a green re-run does not prove a fix for a load-dependent failure; force the failure) — this bug added to its sources.
BUG-106 — The leads browser tests read the user directory as one page of 100, so they fail once a local database holds more than 100 active users who can see leads
Affects: F11-b (forge-et-fe e2e/browser/leads/support.ts readDirectory; leads-inbox.spec.ts’s owner filter; F11-b AC-13, AC-26b) · Found in: F5-e wave 2, the Stage 12 pnpm e2e:gate in the wave worktree on forge-et-fe#165 (head dda2594, 2026-10-07) · Severity: medium — any developer’s browser gate fails after several runs on one local database · Phase: 1
Repro: on macOS, a forge-et-be main backend (a20f9469) on the local profile, one isolated Postgres used by pnpm e2e:gate three times plus the effective-permissions files four times (258 active users by then, most created by the browser suites’ run-unique fixtures). pnpm e2e:gate in forge-et-fe failed 3 tests: leads/leads-inbox.spec.ts:138 (AC-13 — selectOption(String(me.id)) on the owner filter timed out: “did not find some options”), and leads/leads-writes.spec.ts:591 and :612 (AC-26b — “the user directory lists user 6”). On a fresh database the same head passed (1292 passed, 0 failed).
Expected: the leads tests pass on a local database however many times the suites have run on it, as forge-et-fe/e2e/README.md states (“The local database does not need resetting”).
Actual: readDirectory reads GET /api/v1/user-directory?permission=lead.view&size=100, the first page only, and the owner filter’s options are the directory’s first page of 100 (src/lib/userDirectory/userDirectoryApi.ts USER_DIRECTORY_PAGE_SIZE = 100; the backend’s MAX_PAGE_SIZE = 100). Every gate run adds fixture users whose roles hold lead.view, so a seeded user (user 6, the signed-in user) falls off that page.
Fix: in the leads suite, look each needed user up with the directory’s q search (by display name) instead of reading the first page, and choose the owner-filter fixture so its owner is on the first page — for example a run-unique user whose name sorts first. Decide separately whether the owner picker itself should search past 100 (Empire has 10–20 users, so it is not a product defect today).
Regression test: the three tests pass against a local database holding more than 100 active users with lead.view (create them through the admin API before the run).
Resolved: 2026-10-07 — fixed in forge-et-fe#168 (e092066, 6daa9a5; comments and the review’s two Nits in 337a724). readDirectory is replaced by searchDirectory, which uses the directory’s q search and reads every page. A test that picks a user from the owner filter or picker now creates a test-owned user named 0 E2E first owner <file> <token>. A digit sorts before every letter-led seeded and fixture name, so that user is on the first page. The test asserts the user is offered before choosing them, and removes them in a finally (their leads go back to the all-permissions user, then S2 deactivates them); each file’s afterAll sweeps up any its own crashed runs left. AC-13’s owner case runs in the all-permissions session; the AC names no role. AC-21 and three AC-23 tests depended on the same first page and were fixed the same way; AC-21 now compares the picker with the directory answer it received, plus the caller when that answer leaves them out (useOwnerPicker’s current owner). Regression run: a throwaway stack (compose.verify.yaml) with 110 extra active sales_rep users named Aaa BUG-106 seed …, which sort ahead of every seeded user. On main the leads folder (chromium) had 7 failures: the three named here, AC-21 and three AC-23. After the fix: 67 passed, twice. The full browser gate on 6daa9a5 against the same database: 1298 passed, 5 skipped, 0 known failures. Whether the owner picker itself should search past 100 users was a product question (Empire has 10–20 users). The lead decided on 2026-10-07 that it should, and it is filed as BUG-109.
Lesson: L-147 (run-unique fixtures grow the local database, so a test never relies on a list’s first page) — extended with the picker case and this fix.
BUG-105 — The served OpenAPI 3.1 document publishes every @Nullable record field as non-nullable
Affects: FS-10 (forge-et-be platform/web/NullableRecordComponentConverter.java, FS-10 finding F-001, L-019) — every endpoint whose response record has a @Nullable component · Found in: F5-e WI-1.1’s AC-12 run (2026-10-06, recorded in its plan’s Failed Approaches), raised again by backend-reviewer on forge-et-be#205 · Severity: low — no client is generated from the document yet · Phase: 1
Repro: on forge-et-be main (e2c4e350), boot on the local profile, sign in through the stub login and GET /v3/api-docs: the document is openapi: 3.1.0; UserResponse.manager (a @Nullable UserRef) is published as {"$ref": "#/components/schemas/UserRef"}, and MeResponse.User.managerId (a @Nullable Long) as a plain integer — neither admits null.
Expected: a @Nullable record component is published as nullable in the 3.1 spelling, so a client generated from the document types it as optional-or-null — the purpose FS-10 F-001 gave the converter.
Actual: the converter calls Schema.setNullable(true); the 3.1 serializer writes no nullable member (3.1 has no such keyword), so the mark is lost. OpenApiIT’s comment (“carries neither required nor nullable”) records the symptom without a test that fails on it. F5-e worked around it for one field (EffectivePermissionsOpenApiConfig.markImpliedByNullable, on forge-et-be#205).
Fix: in the converter, for each @Nullable component: a property with a type gets "null" added to its type set (["integer", "null"]); a $ref property is replaced by anyOf: [{$ref}, {type: "null"}], the 3.1 form for a nullable reference (a $ref cannot carry a type of its own); keep setNullable(true) for a 3.0 serializer. Schemas built here take the property’s spec version, so the 3.1 serializer writes their types. Once this and forge-et-be#205 are both on main, F5-e’s per-field workaround is removed (it becomes a no-op: the type set already holds null).
Regression test: an OpenApiIT feature that reads the served document: MeResponse.User.managerId’s type set holds integer and null; UserResponse.manager is anyOf the UserRef reference and {type: null}; a non-nullable sibling (UserResponse.displayName) stays a plain string.
Resolved: 2026-10-06 — fixed in forge-et-be#206 (6c65c8e0). The fix found a second cause: for a named model the swagger chain answers only a $ref (#/components/schemas/User) and keeps the properties in the context’s defined models, so the converter had never patched any record since FS-10; it also could not resolve swagger’s Jackson JavaType through ResolvableType. The converter now patches the defined model and reads the class from the JavaType. FieldUpdateSchemas (the three-state ids and numbers) carried the same 3.1 defect and is fixed in the same commit; LeadControllerIT and UserControllerIT now read those fields’ type as a set. Receipts: the new OpenApiIT feature EXIT=1 on the old converter (managerId missing null), EXIT=0 with the fix; the 13 test classes that read /v3/api-docs, plus *FieldUpdate*, EXIT=0. Nullable enums (fixed in the same PR, e0a158ca then 386fdfd9, on backend-reviewer’s reviews): a nullable enum property (for example LeadCreateRequest.customerType) got type: ["string","null"] while its enum did not list null, which JSON Schema 2020-12 still rejects. A first fix used a customizer @Ordered last; the delta review showed springdoc ignores @Order on a @Bean method (it sorts the customizer instances, and a method reference carries no order), so it worked only by bean registration order. The fix that stands is order-independent: OpenApiEnums (platform.web.openapi) sets a vocabulary and lists null when the property’s type set holds it — the converter wrote that type before any customizer runs — and the lead and account configs use it in place of their two copied helpers; the converter itself adds null to a Java enum resolved while the schema is built. OpenApiIT checks the whole document — an enum lists null exactly when its type admits null — and found two more on its first run (MailboxConnectionResponse.statusReason, Subscription.reasonCode); OpenApiEnumsSpec covers the helper. F5-e’s per-field workaround (EffectivePermissionsOpenApiConfig.markImpliedByNullable, forge-et-be#205) becomes a no-op and is removed once both PRs are on main.
Lesson: L-019 — a second instance: the converter written for that lesson shipped without the served-document assertion its corollary (1) asks for, and never worked. The lesson now states the rule for any code that shapes the OpenAPI document, and the 3.1 spelling.
BUG-104 — The Users list’s live region wraps the whole table, so a screen reader may read every row on each filter or page change
Affects: F5-c (forge-et-fe src/features/users/UsersScreen.tsx, the users-region div; spec AC-2, AC-14, NFR-2) · Found in: F5-c WI-1.2’s pre-PR review (2026-10-03, an Important deferred to the lead or the WI-4.1 screen-reader pass); raised as Q-1 of the F5-c WI-4.1 plan and filed by the lead’s decision (2026-10-06) · Severity: low · Phase: 1
Repro: on forge-et-fe main (a27552d, UsersScreen.tsx:251), the Users screen’s always-rendered <div aria-live="polite" data-testid="users-region"> contains the whole data region — the table with its rows, or the empty, no-matches, error or stale panel. Type a search term, change the status filter, or page: the region’s content changes, so a polite live region asks the screen reader to read it all, up to a page of rows (25 by default, 100 at most).
Expected: a list change is announced in one short line — the count and the state — as the email screen does (EmailRoute.tsx keeps the list region silent and announces through a visually hidden inbox-status paragraph, after the forge-et-fe#49 review) and the Leads list does (LeadListPane.tsx: “214 leads”, “Loading leads”, “No leads”). The table itself is not live.
Actual: the table-wide region (WI-1.2’s AC-2 criterion asked for the region to wrap the data, coding-standards §3.2) announces the rows on each change. Axe and the size matrix pass: the markup is valid; the cost is what a screen-reader user hears.
Fix: take aria-live off users-region and add a visually hidden one-line status (role="status" or aria-live="polite", test id e.g. users-status) that the six list states write to: the count on a populated list (“UsersScreen.test.tsx (its users-region row at :215 asserts the region is live) to assert the status line’s text per state and that the table region is not live. R-FND-002 makes F5-c’s list pattern the one later lists inherit, so the fix restates the pattern in UsersScreen.tsx’s header comment.
Regression test: UsersScreen.test.tsx rows per state on the status line’s text; the size matrix’s Users states still pass.
Resolved: 2026-10-07 — fixed in forge-et-fe#168 (363fa49, and the rule in forge-et-fe/CLAUDE.md, 2c7d723). users-region is no longer live. A visually hidden <p role="status" data-testid="users-status">, outside the card and rendered in every state, says “Loading users”, “ErrorPanel is role="alert" and announces itself, and repeating it would announce it twice. That follows EmailRoute’s inbox-status and LeadListPane, as the frontend-reviewer agreed on #168; the lead confirmed it on 2026-10-07. Regression tests (UsersScreen.test.tsx): the line’s text in each state, the same element from loading to rows, empty during a page change and on both failures, and no aria-live on the region or any ancestor. Before the fix, 6 failed; after, src/features/users passed 349 of 349. The size matrix’s Users states passed in the gate on 6daa9a5. coding-standards.md §3.2, forge-et-fe/CLAUDE.md and frontend-reviewer dimension 19 now state the list exception.
Lesson: L-150 (new — a live region around a list reads every row; when two screens depart from a written rule for the same reason, change the rule), with the frontend-reviewer dimension 19 row.
BUG-103 — Two Leads keyboard tests change a closed native select with Home and ArrowDown, keys that leave it unchanged (macOS)
Affects: F11-b (forge-et-fe e2e/browser/leads/leads-a11y.spec.ts, AC-28’s keyboard tests) · Found in: F5-c wave 4, the Stage 12 pnpm e2e:gate --push on forge-et-fe#160 (2026-10-06) · Severity: medium — every browser gate run on macOS fails on these two tests · Phase: 1
Repro: on macOS, against a forge-et-be main backend (248597d), in a checkout of forge-et-fe main (ab6c5c9): ET_BACKEND_PORT=<port> pnpm exec playwright test --project=chromium e2e/browser/leads/leads-a11y.spec.ts → 2 failed, 12 passed. :51 (“a filter by type-ahead …”) times out in changeAndSettle after pressing Home on the Country select to go back to “Any country” (the error snapshot still shows “Sweden” selected). :175 (“the pager — Next with Enter, Rows per page with an arrow key”) times out after ArrowDown on the “Rows per page” select (the snapshot still shows 25 selected).
Expected: AC-28’s keyboard tests pass on every machine the team runs the browser gate on.
Actual: on macOS, Chromium does not change a closed native <select> on Home or ArrowDown (it opens the picker or does nothing); on Windows and Linux it does, which is why forge-et-fe#157’s gate passed (Browser gate: passed · 262010957fca). Type-ahead changes it on every platform (the same test’s type('Sweden') works on macOS).
Fix: keep the steps keyboard-only (AC-28) but change the selects by type-ahead: type the “Any country” option’s label in place of Home, and type 50 in place of ArrowDown, with a comment saying why. Applied on feature/F5-c-wave-4 (forge-et-fe#160) by the lead’s decision, because the failure blocked that PR’s gate; F11-b’s owner is told.
Regression test: the two tests themselves, passing on macOS (and still on Windows / Linux).
Resolved: 2026-10-06 — fixed in forge-et-fe#160 (855dccc, the header comment in 58748b4 and e7d09e4): type('Any') and type('50'), with a Shift+Tab / Tab refocus before the second type-ahead on Country. On macOS: leads-a11y.spec.ts 14 passed, --repeat-each=3 24 passed, and the browser gate on e7d09e4 passed (1240 passed, 0 failed).
Lesson: L-143 — on macOS, Home and the arrow keys do not change a closed native <select>; change it by type-ahead. The rule is in .forge/test-strategy.md § Flaky Tests (forge-et-harness#479).
BUG-102 — The delta poll’s conditional link write matches a NULL that a reconnect restores, so a poll from before the reconnect can store the old mailbox’s links
Filed as BUG-084 and renumbered BUG-096, BUG-100 and then BUG-102: main’s BUG-084 to BUG-101 merged first, in three batches (L-057).
Affects: F20-b (forge-et-be/src/main/java/com/empireteas/mailbox/repository/MailboxConnectionRepository.java recordDeltaToken; MailboxDeltaPollJobHandler.storeLinks) · Found in: the fix of BUG-068, 2026-10-04 — the same cause in the delta poll · Severity: low · Phase: 1
Repro: a poll starts while the connection is ACTIVE and has no stored links (delta_token IS NULL, the first poll after a connect or a reconnect). While it pages, the connection lapses to RECONNECT_REQUIRED and the user reconnects, possibly to a different mailbox; the reconnect’s resetBackfill sets delta_token back to NULL. The poll then calls storeLinks(connectionId, null, links).
Expected: the poll from before the reconnect stores nothing; the next poll starts fresh against the reconnected mailbox, which is why resetBackfill clears the links (spec Rev 1, wave 2).
Actual: recordDeltaToken matches delta_token IS NOT DISTINCT FROM :expected, and both the stale poll and the reset row hold NULL, so the write lands and the next poll resumes from links the old access token fetched. NULLIF(:token, '') can also return the column to NULL from a stored value. Rare: it needs a lapse and a user’s reconnect inside one poll’s run.
Fix: the BUG-068 fix applied to the delta links: a delta_generation column (or the back-fill generation, if the two are always reset together), moved on by every write to delta_token and by resetBackfill, and recordDeltaToken comparing it in place of the stored value. A schema change: ask the lead first.
Plan (written before the code, 2026-10-09): the lead approved the schema change (confirmed by Nanndish on 2026-10-09).
- Its own counter, not
backfill_generation. The two are reset together, but every back-fill page also movesbackfill_generation, and the first back-fill runs while the poll does (every first connect), so a poll compared against it would lose its link write on every page and repeat its round. - A new migration (the next free version at merge time, V38 as of writing):
delta_generation BIGINT NOT NULL DEFAULT 0andCHECK (delta_generation >= 0). Existing rows get 0 and satisfy the check. - Every statement that writes
delta_tokenmoves the counter:recordDeltaTokensetsdelta_generation = delta_generation + 1and matchesdelta_generation = :expectedGenerationin place ofdelta_token IS NOT DISTINCT FROM :expected;resetBackfilladds one too. Thestatus <> 'DISCONNECTED'condition stays. MailboxConnectionmaps the column read-only, asbackfill_generation. The poll reads the generation with the links at its start and passes it tostoreLinks→MailboxIngestService.recordDeltaToken. The audit payload is unchanged: the links’ old value is still the one the poll read, because no write todelta_tokenleaves the generation unchanged.docs/data-model.md§9 gains theALTER;docs/audit/README.mdsays the counter is not recorded, as it does forbackfill_generation.- Tests: a repository IT (below), plus the concurrent-polls case in the shape of
MailboxBackfillGenerationIT;EmailMessageSchemaITpins the column and the CHECK;MailboxDeltaPollJobHandlerSpec,MailboxDeltaPollITandMailboxIngestAuditITupdated to the new signature. Regression test: a repository IT in the shape ofMailboxBackfillGenerationIT: a poll that read no links before aresetBackfillwrites nothing after it. Resolved: 2026-10-09 — fixed in forge-et-be#223 (5f6b0066; the migration comment corrected in32bcb40b, becausemainnow deploys to dev; the review’s two Nits fixed inedef03dd). The plan above was followed; the migration ships asV40__mailbox_delta_generation.sql(written as V38; renumbered in0f53124bafter mergingmain, which had taken V38 and V39). Its header states the rolling-deploy cases: a task still on the previous code writes the links, or resets them on a reconnect, without moving the generation, so until the deploy ends a new task’s poll can overwrite that write or store a pre-reconnect poll’s links./forge-review-pr --commentwithbackend-reviewer: a full review of32bcb40bfound no Blocker or Important and two Nits (that header left out the reconnect case; a Javadoc inMailboxAuditChangesstill said the compare matched the stored value), both fixed inedef03dd, which a delta review passed. The Fast lane then refusedDeltaPositionin its banned-marker step: the record’s defaulttoString()printed the links, which neither review had caught.0e4fc2b2overrides it to print<redacted>, withDeltaPositionSpec, and a delta review passed it. That review found the same defect in three older mailbox records, which the check misses by field name: filed as BUG-176.
- Regression tests:
MailboxDeltaPollIT“a poll overtaken by a reconnect stores nothing, though it started with no links as the reset row holds” drives the real poll, with the reconnect made through the stub-state and stub-connect endpoints while the inbox page is read. On unmodifiedmain(894b1543, only this test file copied in) it fails:deltaToken(connectionId) == nullreadinbox stub-delta:1:inbox:8/sentitems stub-delta:1:sentitems:6, the stale poll’s links. It passes with the fix.MailboxDeltaGenerationIT“a poll that read no links before a reconnect writes nothing after it” fails with the reset’sdelta_generation + 1removed (recorded == 0read1) and passes with it; its concurrent-polls case and its cleared-to-NULLcase pass either way.EmailMessageSchemaITpins the column and the CHECK;MailboxDeltaPollJobHandlerSpecchecks the write carries the generation and links the poll read. - Build: the eight affected test classes, 137 tests, 0 failures.
./gradlew buildon5f6b0066’s tree: 6146 tests, 0 failures, 0 errors, 2 skipped. The PR’s local build record is on forge-et-be#223. Lesson: L-141 — its second instance, now fixed; the lesson gains one rule from this fix (a run that owns a set of columns gets its own counter unless no other writer moves a shared one mid-run), and its grep now finds no compare-and-set on a stored value inforge-et-be.
BUG-101 — The mailbox token refresh to Microsoft has no timeout while it holds database locks
Affects: F20-a (the mailbox token), F5-c (the deactivation and email-change release) · Found in: F5-c wave 3 (backend-reviewer delta review on forge-et-be#201, 2026-10-06; pre-existing code) · Severity: medium · Phase: 1
Repro: read the refresh path: MailboxTokenService.accessTokenForBestEffort → TokenRefreshGrant (auth/config/ContributedTokenRefreshGrant) → RefreshTokenResponseClients, which builds new RestClientRefreshTokenTokenResponseClient(). Spring Security builds that client’s RestClient with no request factory, so no connect or read timeout; MailboxTokenService retries up to MAX_ATTEMPTS while it holds the connection’s pessimistic row lock and a pooled database connection.
Expected: every outbound call that a caller waits on, with locks held, has explicit short timeouts — as GraphClientConfig gives the Graph client (“Short timeouts, named because a caller depends on them”).
Actual: a refresh to Entra that stalls holds its row lock and pooled connection with no bound; during a deactivation or email-change release it also holds the installation-wide access.manage.guard advisory lock, so all user and role administration waits.
Fix: give the refresh-token client a RestClient with explicit connect and read timeouts (setRestClient(...) in RefreshTokenResponseClients), using named constants beside GraphClientConfig’s. A unit test asserts the configured factory’s timeouts; a slice test with a stalled stub server asserts the refresh fails within the bound.
Resolved: 2026-10-07 — fixed in forge-et-be#207 (bd667310; the review fix d1b0582a). RefreshTokenResponseClients builds the refresh client’s RestClient on JdkClientHttpRequestFactory, with CONNECT_TIMEOUT 3s and READ_TIMEOUT 5s. The first version used Graph’s 5s and 15s. The backend-reviewer showed that the read timeout bounds the whole exchange, and that MailboxTokenService makes three attempts under the row lock, so 15s gave a worst case of about 45s. With 5s the worst case is about 15s, and both Javadocs now state it. The lead confirmed the 3s and 5s values on 2026-10-07. setRestClient replaces Spring Security’s whole client, so the form converter, the token-response converter and OAuth2ErrorResponseErrorHandler are added back. The default client had been HttpComponentsClientHttpRequestFactory with no bound. Regression tests (RefreshTokenResponseClientsSpec, a JDK HttpServer on loopback): the configured factory’s timeouts; a 200 still parses and invalid_grant still maps to PermanentFailure through the production client; and a stalled endpoint gives TransientFailure within the bound. Before the fix, the stall row hung until “Method timed out after 30.00 seconds”. With the first version’s 15s bound, the production path then failed in 15.04s; with a short test bound, in about 0.53s. The connect timeout is asserted only as configured, because no address reliably refuses to connect on every machine. The authorization-code exchanges (sign-in and mailbox consent) have the same missing timeout outside this diff; that is BUG-108.
Lesson: L-149 (new — a library-built HTTP client has no timeout until the code sets one), with a backend-reviewer row.
BUG-100 — A mailbox connected while the user is being deactivated keeps syncing after the deactivation
Affects: F5-c (FR-7, AC-8), F20-a (the mailbox connection) · Found in: F5-c wave 3 (backend-reviewer on forge-et-be#201, review I1, 2026-10-06) · Severity: low · Phase: 1
Repro: user U is connecting their mailbox: the consent callback has passed callerWithMailboxAccess() and is exchanging the code with Microsoft. An administrator’s deactivation of U runs MailboxConnectionService.releaseConnectionsFor and commits before U’s connectOrUpdate writes. If U had a connection, the release locked and disconnected it, and connectOrUpdate’s findFirstLockedByOwnerUserIdAndStatusNot(…, DISCONNECTED) then finds none and inserts a new ACTIVE row; if U had none, neither side blocks and the release cannot see the uncommitted insert.
Expected: a deactivated user’s mail is no longer synced (F5-c FR-7).
Actual: U is deactivated but owns an ACTIVE connection, and the ingest and delta jobs keep reading the mailbox; nothing in mailbox reads is_active. The window is milliseconds and needs U to finish connecting at the moment of deactivation. BUG-094’s planned per-request check would not close it, because the consent request was already authenticated.
Fix: in connectOrUpdate, before the insert or update, read the owner’s is_active with a lock that conflicts with the deactivation’s UPDATE app_user — a new auth.access port method such as UserDirectory.lockActive(long) running SELECT is_active FROM app_user WHERE id = ? FOR SHARE — and refuse the connect for an inactive user. It works in either order, because the deactivation flushes its UPDATE before the release reads. A regression test commits a deactivation between the consent check and connectOrUpdate and asserts no ACTIVE connection remains.
Resolved: 2026-10-12 — fixed in forge-et-be#235 (2c0c1126, 68c920f1, e3c0c393, a3d71c8e). The review of the first commit found that an administrator’s email change, which also releases the user’s mailbox, races a connect the same way, so the fix covers both. connectOrUpdate(ownerUserId, ownerEmail, …) first reads its owner through a new auth.access port method, UserDirectory.lockActiveUser(long, String). It runs AppUserRepository.lockActiveByIdAndEmail (@Lock(PESSIMISTIC_READ), so FOR SHARE, on active = true and lower(email) = lower(:email)), is Propagation.MANDATORY so the lock lasts until the connect’s transaction ends, and returns a boolean only. ownerEmail is the email the caller signed in with, which an email change does not update in the session. A deactivated or renamed owner is refused with a new MailboxOwnerChangedException before anything is written; the consent callback turns it into ?connect=failed, and the local/test stub connect answers it as an unexpected error (no catalogued code for that one race). The method is named lockActiveUser, not lockActive as the Fix line suggested, to sit beside isActiveUser. Accepted: an edit or deactivation of a user who is connecting waits for the connect’s KMS data-key call and Graph subscription calls; the Graph calls are bounded by GraphClientConfig, and the KMS client’s missing timeout is BUG-174. Regression test: MailboxConnectOwnerChangeRaceIT forces both orders, for a deactivation and for an email change, each with and without a connection the user already had (8 cases), and asserts that the second change waited on the user’s row. With the fix all 8 pass. Against unchanged origin/main (894b1543), the 4 deactivation cases fail: the two with no earlier connection on outcome.secondWaited (nothing waited), the two with one after 30 s, because the second change waits only on the mailbox_connection row. Against the first commit alone (2c0c1126), the 2 “email change first” cases fail. RepositoryUserDirectoryIT calls lockActiveUser directly (a case-only difference matches; another email, a deactivated or unknown user does not; a call outside a transaction is refused). ./gradlew build on the head a3d71c8e (after merging main): 6,829 tests, 0 failures, 2 skipped. Earlier full builds hit BUG-130’s timing failure, which main has since fixed.
Lesson: L-156 (a read that decides an insert locks the row a concurrent writer changes) — this fix added to its sources. BUG-093 and BUG-094 are its remaining open instances.
BUG-099 — FormErrorSummary’s error links are 16px tall, under the 24px target-size floor
Affects: F26-b (the shared src/ui FormErrorSummary, used by the pricing dialogs), F11-b (the New lead and Edit lead form), F1-b (the New account and Edit account form, WI-1.2 — its form page states carry the same mark until this is fixed) · Found in: F11-b wave 1, the Stage 7 Step 5e size-matrix run of page state leads-new-form-errors (2026-10-05) · Severity: low · Phase: 1
Repro: on forge-et-fe feature/F11-b-WI-1.5-e2e-lead-inbox (c8288c8), with a backend running, pnpm exec playwright test e2e/browser/size-matrix -g leads-new-form-errors. The page state opens New lead and submits it empty, so the form shows FormErrorSummary with a link per error.
Expected: every visible control, the summary’s error links included, is at least 24 × 24 CSS px, or meets a WCAG 2.2 SC 2.5.8 exception (the CRM’s floor — PRD § Accessibility, test-strategy.md § Browsers & Viewports).
Actual: each link (a[data-testid="leads-form-errors-link-…"]) is about 212–257 × 16 px. The page check reports target-size at 768×1024, 1024×768, 1280×800 and 1440×900 in Chromium and at 768×1024 and 1024×768 in WebKit. At 320px the links wrap and pass. FormErrorSummary.module.css’s .link sets body line-height and no minimum block size, and the links sit in a list, not in running text, so the inline-link exception does not apply. F26-b’s SetCostDialog and StandardFormDialog use the same component; no page state shows their summaries, so the gap had not been checked before.
Fix: in src/ui/components/FormErrorSummary/FormErrorSummary.module.css, give .link a 24px minimum block size (for example display: inline-flex; align-items: center; min-block-size: var(--et-…), using the token the other src/ui controls use for the 24px floor) without changing its look, and add the summary’s links to the component’s interaction-matrix check. Regression check: the size matrix’s leads-new-form-errors state passes at every size with its known-failure mark removed.
Resolved: 2026-10-07 — fixed in forge-et-fe#168 (aa4f215). .link is now display: inline-flex; align-items: center; min-block-size: var(--et-space-7), the 24px token BackLink uses, and its look is unchanged. FormErrorSummary was already in the interaction matrix. A stylesheet test asserts the rule: before the fix it failed (expected … to match /display:\s*inline-flex/); after, it passes. The { bug: 'BUG-099', kinds: ['target-size'] } mark is removed from leads-new-form-errors in e2e/pageStates.ts. No F1-b account form-error page states exist on main yet. The browser gate on 6daa9a5 passed with 0 known failures (1298 passed, 5 skipped).
Lesson: L-148 (a link that is not in running text gets no inline-link exemption and needs a 24px target; only the browser gate measures it) — this bug added to its sources.
BUG-098 — The browser-gate hook lets a PR through when an if, for or while block that changes directory is piped
Filed as BUG-096 and renumbered BUG-098: two other open branches had also taken BUG-096, and one BUG-097 (.claude/rules/tracker.md, L-057). The PR comment on forge-et-harness#455 that reported it still says BUG-096.
Affects: F40 (forge-et-harness/.claude/hooks/guard-browser-gate.mjs, the bash parser) · Found in: the /forge-review-pr adversarial pass on forge-et-harness#455, the BUG-059 fix (2026-10-05) · Severity: medium · Phase: 1
Repro: from a forge-et-fe checkout, with no passing gate result, read if true; then cd <harness>; fi | cat; gh pr create -t x -b y (also for d in x; do cd <harness>; done | cat; gh …, and either with >/dev/null before the |) with the hook’s tokenize, findOccurrences and readOccurrence. Reproduced on origin/main (3dbf4c1) and with the BUG-059 fix.
Expected: gated — the block is a pipeline element, so it runs in its own subshell and its cd does not reach the gh pr create, which still runs in the fe checkout.
Actual: allow, as if it were a harness PR. The parser does not store an if, for, while or until block as a group: each of its parts (if true, then cd …, fi) is a command of its own, so the | after fi or done is recorded on the fi or done command, and the cd before it is counted as running in the same shell as the gh. A false pass, not a false refusal; the Fast lane description check still fails such a PR. The cause is the same kind as BUG-059’s (a separator after a compound command recorded on the wrong item), in code BUG-059’s fix did not change.
Fix (decided 2026-10-08 — parse, do not refuse): read if … fi and while/until/for/select … done as one group, as case … esac is read (BUG-044), so a following |, & or redirection applies to the whole block. Rejected: refusing a | or & after fi or done when the block holds a cd or GH_REPO, because finding the block’s extent needs the same parsing, and the refusal would also block a cd and a gh inside one piped block, which bash runs correctly.
Regression test: hook unit tests that the repro forms are gated from the fe checkout, and that if true; then cd /w/fe-wt; fi; gh pr create still resolves to the fe checkout.
Resolved: 2026-10-08 — fixed in forge-et-harness#522 (branch fix/BUG-098-piped-if-for-while-blocks). BashParser.parseList keeps a stack of the if, while, until, for and select statements open in the list; the matching fi or done in command position (or the } of a for/select { … } body) moves the statement’s items into one group of that type, so a following redirection is stored on it and a following | or & marks it, and runsApart keeps its cd or GH_REPO from reaching the gh. The parts inside are read as before, so an unpiped if true; then cd x; fi; gh … still counts the cd. The new group types join SAME_SHELL_GROUPS. An unclosed statement, a fi or done with nothing of its kind open, and a word right after fi or done are now refused as unparseable, as bash rejects them. Probing the fix found a third form of the same cause, already on main: a separator before a ! or time prefix was stored on the prefix’s own command, not on the group after it, so echo x | time { cd …; }; gh … and echo x | ! { cd …; }; gh … were allowed as a harness PR. A group, case or statement after a reserved-word prefix now takes the prefix’s separator (sepHere), which also makes false && ! { cd …; }; gh … refused as conditional, as false && { cd …; }; gh … already was. The F40 design record states both in FR-H1 and FR-H4, with a revision entry. New tests in guard-browser-gate.test.mjs (describe('if, while, until, for and select statements (BUG-098)')): 28 forms gated from the fe checkout (the repro forms with |, &, >/dev/null, 2>&1, <, a here-document, a pipe into the statement, nesting, GH_REPO, a { … } for body, and the prefix forms), 10 allowed forms (a cd before gh in the same shell, with or without a redirection, and a cd and a gh inside one piped block), 2 conditional refusals, the stored-group shape, the occurrence reads, and 10 unparseable forms. node --test .claude/hooks/tests/guard-browser-gate.test.mjs: 578 pass, 0 fail; with the parser put back to origin/main (7512f63d), 40 of the new tests fail and 538 pass. bash .claude/hooks/tests/guard-browser-gate.test.sh on Windows (Git Bash): 84 passed, 0 failed, AC-H6 skipped (no --fe-checkout). With real git from a forge-et-fe checkout, readOccurrence read if true; then cd <harness>; fi | cat; gh pr create -t x -b y, the for … done form, both again with >/dev/null, and echo x | time { cd <harness>; }; gh pr create -t x -b y as a harness PR (allow) on main and as forge-et-fe (gated) with the fix; if true; then cd <harness>; fi; gh pr create -t x -b y stays allow on both.
Lesson: L-095 — the review of BUG-059’s fix found this false pass by probing the parser adversarially, and probing this fix found the !/time prefix form; BUG-098 is added to its sources.
BUG-097 — ScanResultQueueGuard does not refuse a cloud shape with the SQS listener switched off
Affects: F38-c (forge-et-be/src/main/java/com/empireteas/platform/scan/ScanResultQueueGuard.java) · Found in: the design of BUG-070’s fix, 2026-10-05 · Severity: low · Phase: 1
Repro: start a cloud shape with scanning on (et.files.scan.provider absent or guardduty), et.messaging.scan-result-queue set, and spring.cloud.aws.sqs.listener.auto-startup=false.
Expected: the backend refuses to start, as it does when et.messaging.scan-result-queue is missing, because nothing would consume GuardDuty’s scan results.
Actual: it starts. auto-startup is one switch for every @SqsListener container, so the scan-result listener exists but never polls, and every upload stays PENDING and undownloadable — the silent failure the guard exists to prevent. BUG-070 is the same gap in SchedulerConfigGuard.
Fix: the same check BUG-070’s fix adds to SchedulerConfigGuard (refuse auto-startup=false unless et.environment is test), applied where ScanResultQueueGuard requires the queue; if both guards then carry the same check, move it into one shared helper in platform.
Regression test: a ScanResultQueueGuardIT row that a deployed cloud shape with scanning on and the listener off refuses to start, and that the test contexts still start.
Resolved: 2026-10-07 — fixed in forge-et-be#207 (d0d26156). BUG-070’s check moved into one helper, platform.messaging.SqsListenerSwitch, and both guards call it. SchedulerConfigGuard’s message is unchanged. ScanResultQueueGuard runs the check after its queue-present checks, and skips it when et.files.scan.provider=none, because nothing then waits on the queue. SchedulerConfigGuardIT’s deployed-label rows now remove the other guard’s bean definition first (support/WithoutGuard), because the two guards are unordered and both refuse that shape. Regression tests (ScanResultQueueGuardIT and the bare-factory spec): a full boot under dev, uat, prod and production with the listener off refuses to start; a test context starts; and the bare-factory rows mirror BUG-070’s. Before the fix: 29 tests completed, 11 failed. After: the guard specs, SchedulerConfigGuardIT, ScanOffStartupIT and the ArchUnit specs ran 56 tests, 0 failed.
Lesson: L-067 (a consumer’s activation conditions are what switch it off in cloud; its guard must cover every one), which already names this bug.
BUG-096 — The access browser specs scan the Pricing screen without waiting for its standards list to load
Affects: F26-b (FR-U1, the Standards book list), F5-k (AC-22, the route-guard page check) · Found in: F11-b, WI-1.5 plan authoring, reading e2e/browser/access/support.ts (2026-10-05) · Severity: low · Phase: 1
Repro: on forge-et-fe main (44ab20d), read e2e/browser/access/support.ts: SECTION_LOADING_TEST_IDS lists Email, Mailbox, Roles and Users only, and the comment above it says “The other screens (Accounts, Leads, Pricing) render no fetched content”. Then git grep -l -E 'apiClient|fetch\(|useQuery' origin/main -- src/features/pricing lists StandardsBook.tsx, standardsQueries.ts, standardsApi.ts and others, and StandardsBook.tsx:546 renders <LoadingSkeleton data-testid="pricing-standards-loading" …>. route-guards.spec.ts:236 calls waitForSectionSettled(page, route.heading) before its axe scan, which for Pricing waits on nothing. Not yet seen failing; found by reading.
Expected: the route-guard page check scans Pricing once its loading branch is gone, as it does for Email, Mailbox, Roles and Users (the comment’s own instruction: “When one of them starts fetching, add its loading test ids here, or the scan will run over its skeleton”).
Actual: the scan can run while pricing-standards-loading is on screen, so it checks a skeleton instead of the settled list, and its result depends on how fast the list loads.
Fix: add Pricing: ['pricing-standards-loading'] to SECTION_LOADING_TEST_IDS (plus any other loading id the Pricing route shows on first render — re-check git grep -n "data-testid=\"pricing-.*loading" -- src/features/pricing), and correct the comment so it names only the screens that still fetch nothing (re-run its grep; Accounts’ AccountStandardsPanel fetches, but check whether /accounts renders it before listing Accounts as fetching). Regression check: the access specs that visit /pricing (route-guards.spec.ts, and navigation-by-role.spec.ts / shell/a11y.spec.ts if they call waitForSectionSettled with Pricing) pass, with the scan now running after the skeleton is gone. F11-b WI-1.5 adds the Leads entry and corrects only the Leads part of the comment, so the two changes touch the same lines — whichever lands second merges the other.
Resolved: 2026-10-06 — fixed in forge-et-fe#157, the F11-b wave 1 PR (commit 71e8b10, a fix from its /forge-review-pr): SECTION_LOADING_TEST_IDS gains Pricing: ['pricing-standards-loading'] (the Pricing route’s only first-render loading id; the Set cost dialog’s is not on first render), and the comment names only Accounts as fetching nothing, with the grep that re-checks it (AccountsRoute.tsx imports nothing, and no route mounts AccountStandardsPanel). Regression check: the browser gate on that PR’s head 2620109 passed (1155 passed, 5 skipped, 13 known failures), the access specs that visit /pricing included.
Lesson: L-136 — the comment said Pricing renders no fetched content and listed the screens to wait for; F26-b made Pricing fetch and nothing updated the list. The fix replaces the dated sentence with the grep that re-checks it.
BUG-095 — The account list rows carry no LTV, so F1-b cannot show its Est. LTV column
Affects: F1-a (FR-A1, the account list contract), F1-b (FR-L2, AC-2) · Found in: F1-b, re-checking its plans against main after F5-f merged (2026-10-05) · Severity: medium · Phase: 1
Repro: on forge-et-be main, call GET /api/v1/accounts as any user who reaches an account that has an estimated LTV, then GET /api/v1/accounts/{id} for the same account.
Expected: each list row carries estimatedLtv (nullable) and ltvCurrency, as the detail record does, so F1-b’s list can show Est. LTV without computing anything or reading each record (F1-a spec Rev 7, FR-A1, AC-30; F1-b FR-L2, AC-2, NFR-5).
Actual: the detail record has both fields; the list row (account/web/AccountSummaryResponse.java) has ownerId, lastContactAt, dormant and sharedWithMe but no LTV. F1-a’s FR-A1 never listed it, so the omission was in the contract, not only the code.
Fix: add @Nullable BigDecimal estimatedLtv and String ltvCurrency to AccountSummaryResponse, and copy them from the entity in AccountResponseMapper.toSummary (the entity is already loaded; no query changes). Regression tests: AccountResponseMapperSpec asserts the summary copies both fields, including a null estimate; a list integration test asserts a row’s two values equal the detail record’s (AC-30), and fails before the change. The harness PR ticks F1-a AC-30.
Resolved: 2026-10-05 — fixed in forge-et-be#194 (5c45531). AccountSummaryResponse gains @Nullable BigDecimal estimatedLtv and String ltvCurrency, and AccountResponseMapper.toSummary copies both from the loaded entity; AccountCollectionDefinition is unchanged, so neither is sortable or filterable. ltvCurrency is always present — USD by default — even when the estimate is null (the F1-b list’s row type parses it as non-null). Regression tests: AccountResponseMapperSpec “the summary row copies estimatedLtv and ltvCurrency from the entity, a null estimate included” (USD, EUR and null rows) and AccountControllerIT “each list row carries estimatedLtv and ltvCurrency equal to GET /{id}, a null estimate as a present null (AC-30, BUG-095)”. Before the fix the two specs gave 53 tests completed, 5 failed; after it the account specs passed (mapper 19, AccountControllerIT 34, collection definition 5, shared-with-me 10, 0 failures), and the full ./gradlew build on 5c45531 passed with 5031 tests, 0 failures, 2 skipped (the local build record on #194). backend-reviewer: no Blocker or Important; its two Javadoc Nits were left as is (review record on #194).
Lesson: L-134 (new) — a list API written before its first screen is specced must be checked column by column against that screen’s spec. No other Lesson: line names this cause. The catch point is the screen spec’s review, not code review, so no backend-reviewer, frontend-reviewer or test-strategy.md row is added; a spec-reviewer row is left as a follow-up.
BUG-094 — A sign-in that completes while the user is being deactivated can leave them a working session
Affects: F5-c (FR-7, AC-8, NFR-4), FS-6 (FR-A2, sessions) · Found in: F5-c wave 3 (plan-reviewer on the WI-3.1 plan, review I1, 2026-10-05) · Severity: low · Phase: 1
Repro: user U is signing in (Microsoft, or the local stub). The login request reads U as active (IdentityResolver / StubLoginService); Spring Session writes the session row — or sets principal_name on the anonymous row made before the Entra redirect — in its own REQUIRES_NEW transaction at the end of the request. An administrator’s deactivation of U commits between those two steps: its DELETE FROM spring_session WHERE principal_name = <U> (F5-c WI-3.1) cannot see the row that is not yet written or not yet named.
Expected: after a deactivation, U’s next request is 401 (F5-c AC-8, NFR-4).
Actual: U’s session is saved after the delete, and no per-request check reads is_active, so U stays signed in until the 8-hour sliding timeout. The window is milliseconds and needs U to finish signing in at the moment of deactivation.
Fix: a per-request is_active check in the sign-in layer — for example in the per-request grant read (EffectivePermissionResolver), ending the session and answering 401 for a deactivated user — so the deactivation does not depend on catching every session row. A regression test drives the login’s session save after a committed deactivation and asserts the next request is 401.
Resolved: 2026-10-10 — fixed in forge-et-be#224 (b7b05562; 8f1b9b66 and the review fixes c4ab595c, 5fdf3433). A new auth.config.DeactivatedSessionFilter, added to the chain right after LogoutFilter, reads UserDirectory.isActiveUser on every request that carries a signed-in user (stub or Microsoft principal). For a deactivated user it invalidates the session (Spring Session deletes the row and expires the SESSION cookie), clears the context and lets the request go on as anonymous, so a route that needs a session answers 401 UNAUTHENTICATED and an open route still answers. It logs one INFO line and records a new security event, auth.session_ended (context.reason = user_inactive), through recordSecurityEventOrLog; an audit failure never keeps the user signed in. The Fix: line suggested the check in EffectivePermissionResolver; that class already denied a deactivated user, but it runs only where a permission is checked, so it could not turn GET /api/v1/me (which had answered 200) into a 401. The filter sits before the OAuth2 filters, so no sign-in or mailbox consent starts on such a session; BUG-100 is a different race and stays open. SecurityChainIT’s deactivation row, which pinned the old 403 from the permission check, now expects 401. Regression tests: DeactivatedSessionIT (the race cannot be driven over HTTP in a fixed order, so each row signs the user in, then sets is_active false with the session row left in place) and DeactivatedSessionFilterSpec (both principal types, the audit call and its failure). With the filter removed from the chain, DeactivatedSessionIT gave 7 tests completed, 5 failed (Status expected:<401 UNAUTHORIZED> but was:<200 OK>); with it, 7 passed. local-build-evidence.py on 5fdf3433: BUILD SUCCESSFUL in 25m 5s, 6154 tests, 0 failures, 2 skipped. /forge-review-pr with backend-reviewer: no Blocker; five Importants (stale documentation of the standalone security events, the filter’s order unpinned, no INFO line), all fixed across two delta rounds; review record approve on 5fdf3433.
Lesson: L-156 — the same cause, where the insert is Spring Session’s own write, so no lock can span it; the lesson gains a bullet (check the state where the row is used) and its backend-reviewer row a sentence, with BUG-094 added to its sources.
BUG-093 — A record assigned to a user while they are being deactivated can leave a deactivated owner
Affects: F5-c (FR-7, AC-18), F1-a (account owner, FR-O1), F11-a (lead owner), F5-h (the bulk reassign routes, FR-B1 — the same unlocked owner check, held for the whole request) · Found in: F5-c wave 3 (WI-3.1 plan, ## Risks, 2026-10-05) · Severity: low · Phase: 1
Repro: two administrators at the same moment: one deactivates user U, who owns no records; the other assigns an account or a lead to U. The owner check (UserDirectory → AppUserRepository.existsByIdAndActiveTrue, used by AccountService and LeadIntakeService for OWNER_NOT_ACTIVE) reads U as active with no lock, and the deactivation’s owned-records count (WI-3.1’s OwnedRecordCounters) cannot see the uncommitted assignment.
Expected: a deactivated user owns no record (AD #35; spec FR-7: deactivation is refused while the user owns any record).
Actual: if both commit, U is deactivated and owns the new record. The window is a few milliseconds and needs a hand-made assignment at the moment of deactivation. F5-h’s bulk reassignment, or a manager reaching the record, can move it afterwards.
Fix: make the owner check lock the owner’s row against the deactivation — for example existsByIdAndActiveTrue taking FOR SHARE (@Lock(PESSIMISTIC_READ) on a query form), so it conflicts with the deactivation’s row update: if the assignment locks first, the deactivation waits and its count sees the record; if the deactivation updates first, the assignment re-reads the user as inactive and is refused. BUG-090’s fix changes AppUserRepository in the same way, so the two may share a PR. A regression test drives the two transactions in a fixed interleaving (the LastAdminRaceIT pattern).
Resolved: 2026-10-12 — fixed in forge-et-be#241 (merged as a83f9910; the merge of main and its fixes in fa731373 and 0543e0d3). Every write that stores an owner (account create and update, lead intake and update, the bulk reassign routes) reads its owner through UserDirectory.lockActiveUser(long), which holds the user’s row FOR SHARE; a deactivation takes the row FOR NO KEY UPDATE before it counts the user’s records, so whichever comes second waits for the first. OwnerDeactivationRaceIT runs the race in both orders; OwnerCheckConfinementSpec limits the lock-free isActiveUser to reads that decide no owner write. The merge of main brought BUG-094’s DeactivatedSessionFilter (allowed, with its reason) and BUG-100’s MailboxConnectOwnerChangeRaceIT, whose wait detector now also matches the deactivation’s new lock. Evidence: the build record on 0543e0d (6,870 tests, 0 failures, 2 skipped) and an approve review record by backend-reviewer.
Lesson: L-156 (its source list gains BUG-093, and a bullet that the writer takes its lock before the read it decides on).
BUG-092 — A blocked “All records” box’s accessible name drops its visible label in real browsers
Affects: F5-d (role editor) · Found in: F5-d wave 3 (WI-3.1’s editor.spec.ts AC-18 test, Run B in Chromium, 2026-10-05) · Severity: low · Phase: 1
Repro: open a role in the role editor (/roles/:id) with None chosen for an entity that has an <entity>.all key (Accounts, Leads), and read the accessible name of its All records checkbox (roles-all-account) in Chromium’s accessibility tree (Playwright toHaveAccessibleName).
Expected: the name starts with the visible label and carries the reason — “All records Choose View or Manage first” (spec AC-19, design-system §9 Permission rendering; WCAG 2.5.3 Label in Name).
Actual: the name is “Choose View or Manage first”. AllRecordsTick.tsx sets aria-labelledby="<checkbox id> <reason id>" on the blocked box; the checkbox’s own id contributes nothing in Chromium’s name computation (an input referenced from its own aria-labelledby yields no text), so the wrapping <label>'s “All records” is lost. The component test (RoleEditorRoute.test.tsx) passes because jsdom’s name computation differs.
Fix: give the visible label text an id the box can reference — e.g. aria-labelledby="<label-text id> <reason id>", which needs Checkbox to accept an id for its label span (or AllRecordsTick to render its own clipped “All records” span with an id) — and remove the markKnownFailure('BUG-092', …) mark in forge-et-fe e2e/browser/roles/editor.spec.ts in the same PR, which then passes in Chromium. When removing the mark, put back in that test, before its name loop, await expect(leads.none).toBeChecked() and the aria-disabled="true" check, so the unmarked test proves the ArrowUp reached None (forge-et-fe#144 delta review).
Resolved: 2026-10-05 — fixed in forge-et-fe#145 (c6a41a4). Checkbox gains labelId on its visible label text, and AllRecordsTick names the blocked box aria-labelledby="<label text id> <reason id>"; the markKnownFailure('BUG-092', …) mark is removed from editor.spec.ts, whose test again checks that None was chosen and each box is blocked before asserting the name. Regression test: RoleEditorRoute.test.tsx “names the blocked box from its label text and the reason, never from the input itself (BUG-092)” fails on the old component (expected [ '_r_5_-all', '_r_5_-reason' ] to not include '_r_5_-all') and passes with the fix; the browser test passes in Chromium in the PR’s gate run. PermissionGuard uses the same self-reference for the buttons and links it wraps; their name is their content, so it is not affected.
Lesson: L-129 (an accessible name built with aria-labelledby is proven only in a real browser; reference the visible label’s element, never the control itself) — written from this bug at F5-d’s Reflect.
BUG-091 — The role editor’s save and reload errors and the create-user form’s errors show no reference id (BUG-087’s rule)
Affects: F5-d (role editor), F5-c (wave 1 create-user form) · Found in: F5-c wave 2 (frontend-reviewer on forge-et-fe#142, 2026-10-05) · Severity: low · Phase: 1
Repro: make a save in the role editor (/roles/:id) fail with a 5xx or a network error, or its Reload fail; or make a create in the Users screen’s Create user form fail with a 5xx, or its role or manager picker read fail.
Expected: every error state for a failed API read or write shows the request’s correlation id under the message, as TraceLine does on the Users and Roles read-error panels (BUG-087, forge-et-fe#134), so a person can quote it to support.
Actual: only the page-level read-error panels carry a TraceLine. RoleEditorForm’s save-failure summary and failed-Reload message, and CreateUserModal’s form-level failure and its two inline picker read failures, show the message with no reference. F5-c wave 2’s user editor had the same gap and fixed it in forge-et-fe#142 (users-editor-save-trace, -reload-trace, -roles-trace, -managers-trace).
Fix: in each of those error states keep traceIdOf(error) and render <TraceLine testId="…-trace" traceId={…} /> under the message (roles-editor-save-trace, roles-editor-reload-trace, users-create-trace, users-create-roles-trace, users-create-managers-trace), with component tests that show the id when the problem carries one and no line when it does not — the forge-et-fe#142 pattern.
Resolved: 2026-10-07 — fixed in forge-et-fe#168 (8450283). Four of the five states were already on main: 87ffb39 had added roles-editor-save-trace, users-create-trace, users-create-roles-trace and users-create-managers-trace, with tests. The failed Reload was left. onReload now resolves to {kind: 'reloaded', role} | {kind: 'failed', error}, as the user editor’s does. The form keeps the failed Reload’s error and shows roles-editor-reload-trace in place of the save’s line. Regression test (RoleEditorRoute.test.tsx): a failed Reload with an id shows the reference, and one with no id shows no line. Before the fix it failed with Unable to find … roles-editor-reload-trace; after, src/features/roles passed 186 of 186.
Lesson: L-144 (a rule about “every” case needs a check that lists the cases), which already names this bug as a source.
BUG-090 — A first Microsoft sign-in that races an administrator’s email change can keep the old identity link
Affects: F5-c (FR-4a, AC-17), FS-6 (FR-A3) · Found in: F5-c wave 2 (backend-reviewer on forge-et-be#188, 2026-10-05) · Severity: low · Phase: 1
Repro: user U has no user_identity row yet. The person who owns U’s old address signs in with Microsoft for the first time at the same moment as an administrator saves a new email for U: IdentityResolver reads AppUserRepository.findActiveByEmail(<old address>) with no row lock (auth/service/IdentityResolver.java:107) and then inserts the user_identity row (:116), while the edit (RepositoryUserAdministration.update) updates app_user.email and runs findAllByUserId after its saveAndFlush.
Expected: an email change always leaves U with no identity link (FR-4a, AC-17, WI-2.1 plan D6), so only a sign-in by the new address can link U.
Actual: the sign-in’s insert takes only a key-share lock on app_user (the foreign-key check), which does not conflict with the edit’s row update, and the edit’s findAllByUserId cannot see the uncommitted insert. If both commit, U keeps a link to the old address’s Microsoft account, and that person keeps signing in as U through findByProviderAndSubject. The window is a few milliseconds and needs a hand-made administrator edit at the same moment as that person’s first sign-in.
Fix: make AppUserRepository.findActiveByEmail take @Lock(PESSIMISTIC_READ) (FOR SHARE), so it conflicts with the edit’s row update: if the sign-in locks first, the edit waits and its later findAllByUserId sees and deletes the committed link; if the edit updates first, PostgreSQL re-checks lower(email) after the edit commits and the sign-in is refused. A regression test drives the two transactions in a fixed interleaving (the LastAdminRaceIT pattern).
Resolved: 2026-10-08 — fixed in forge-et-be#215 (6ada7e5e; the review fixes a153b159 and 994a6fb7). IdentityResolver reads the user through a new AppUserRepository.lockActiveByEmail, the same query with @Lock(PESSIMISTIC_READ), which Hibernate sends as FOR SHARE. The lock is on a new method, not on findActiveByEmail as the Fix line said, because StubLoginService.login calls findActiveByEmail in a read-only transaction and PostgreSQL refuses FOR SHARE there (ERROR: cannot execute SELECT FOR SHARE in a read-only transaction, checked on postgres:18-alpine). ActiveEmailLookupConfinementSpec keeps findActiveByEmail to the stub login. Regression test: IdentityLinkRaceIT holds one transaction open and runs the other, in both orders. Before the fix both cases failed (neither waited); after it both pass, and ./gradlew build on the head 994a6fb7 passed with 5419 tests, 0 failures. Not changed: a session the old address got before the email change stays valid, because an email change does not end sessions (F5-c’s design; only deactivation does).
Lesson: L-156 (new — a read that decides an insert locks the row a concurrent writer changes, or the writer’s clean-up misses the insert), with a backend-reviewer row. BUG-093, BUG-094 and BUG-100 are open instances of the same cause.
BUG-089 — The shared Combobox leaves the page behind its open list focusable while hiding it from screen readers, and its hidden Dismiss button is under the 24px floor
Affects: F26-b, and every screen that uses the shared Combobox (forge-et-fe/src/ui/components/Combobox/Combobox.tsx, on main since Foundation) · Found in: F26-b WI-1.5’s size-matrix run, /forge-deliver Stage 7 Step 5e (2026-10-02) · Severity: medium · Phase: 1
Repro: in forge-et-fe on feature/F26-b-WI-1.5-standards-book-e2e at 22bcde2, against the wave-1 backend on the local profile: playwright test e2e/browser/size-matrix, page state pricing-client-filter-open (the Pricing page with the Standards book’s client filter open and its list of three options showing). It fails on all seven size projects (size-320x640-chromium, size-768x1024-chromium, size-1024x768-chromium, size-1280x800-chromium, size-1440x900-chromium, size-768x1024-webkit, size-1024x768-webkit). No page state on main opens a Combobox’s list, so this is the first check of the open state.
Expected: with the list open, the shared page check is clean: no axe aria-hidden-focus violation, nothing focusable inside an aria-hidden region, and every control at least 24 × 24 px (.claude/CLAUDE.md § Quality Gates › Browser QA; design-system §9).
Actual: while the list is open, React Aria sets aria-hidden="true" on everything outside the combobox (the top bar, the toolbar’s search, the archived switch, Add standard, the table’s scroll region, every row’s Actions button, the sidebar toggle) but leaves those elements focusable and not inert, so axe reports aria-hidden-focus (serious) and the page check lists each element. The popover’s visually hidden button "Dismiss" (React Aria’s dismiss button) measures 16 × 6 px, under the 24 × 24 px floor.
Fix: in the shared Combobox, make the content outside the open list inert rather than only aria-hidden (or have React Aria do so), or keep the outside content reachable and drop the aria-hidden marking, whichever keeps the combobox’s own keyboard behaviour; and give the hidden Dismiss button a target-size exemption that names it as visually hidden (an e2e/support/axeExemptions.ts entry and the page check’s equivalent), or enlarge it. The owner of the shared component decides; F26-b does not change it.
Decision (2026-10-05, Nanndish): keep the outside content reachable and drop the aria-hidden marking. react-aria 3.52.1 and react-aria-components 1.21.1 are the latest releases, so no upgrade fixes it, and the component cannot change it: useComboBox calls ariaHideOutside([input, popover]) in an effect with no option to turn it off. So a pnpm patch of react-aria (patches/react-aria@3.52.1.patch) removes that one call, as the WAI-ARIA combobox pattern has it. Accepted cost: a screen-reader user can move off the open list into the page. inert was tried first and rejected: the patch passed shouldUseInert: true (the option the Modal popover uses), which cleared the page check, but Tab with the list open then lost focus to <body> in Chromium and WebKit (the library removes inert in an effect, after the browser has picked the next control), and the first click on another form field only closed the list, which failed e2e/browser/users/see-and-add.spec.ts (the Manager list in Add user). The Dismiss button needs no exemption: it is react-aria’s DismissButton, styled 1 × 1 px inside a clipped visually hidden wrapper, and the browser’s button padding and border make it 16 × 6 px. A rule in Combobox.module.css sets the popover’s button padding and border to 0, so it is the 1 × 1 px box the page check already skips as visually hidden. The gate’s lockfile reader (scripts/browser-gate/lockfile.mjs, byte-copied in forge-et-pwa) learns pnpm’s (patch_hash=…) snapshot segment, and patches/ becomes a gate input.
Regression test: the pricing-client-filter-open page state itself; on the fix, remove its knownFailures entry, and the size matrix must pass on all seven projects.
Attribution (2026-10-05, frontend-reviewer on forge-et-fe#136): pre-existing in main’s Combobox, not caused by F26-b. The wave’s Combobox change adds server-search props and aria-labelledby and does not touch the popover; react-aria-components renders the popover’s DismissButton unconditionally, and react-aria’s useComboBox calls ariaHideOutside (hide, not inert), in versions this wave did not change. main’s only other consumer, CreateUserModal, sits inside a showModal() dialog where the page is already inert, so no check saw it before.
Numbering: filed on the F26-b wave branch as BUG-077 on 2026-10-02; renumbered BUG-089 when main was merged in, because main had merged a different BUG-077 and BUG-088 was claimed on chore/F38-j-reflect (L-057).
Resolved: 2026-10-05 — fixed in forge-et-fe#150 (3c54a10, review fixes 6cb8d6b and fe73993), with the gate files byte-copied into forge-et-pwa#61, which merges after it. The unit suite also guards the patch: with the unpatched builds, 2 of the 28 Combobox tests fail. With the known-failure mark removed, pricing-client-filter-open failed on all seven size projects before the patch and passed on all seven after; a new standards-book-a11y.spec.ts test checks that Tab, Shift+Tab and one click leave the open client filter for the right control and that nothing behind it is aria-hidden or inert. pnpm patch-commit on Windows wrote a 1.4 MB patch that deleted thousands of the package’s files (it compared the wrong trees), so the patch was built with git diff --no-index over the three edited files instead.
Lesson: L-024 — another overlay primitive whose open state was first checked in a browser by a feature, not when the primitive was built; BUG-089 added to its sources.
BUG-088 — The backend test suite is at the test Postgres connection limit, so one more Spring context fails unrelated specs
Affects: FS-2, FS-9 (forge-et-be — src/test/groovy/com/empireteas/support/SharedPostgresSpec.groovy’s container, max_connections=300; every @SpringBootTest context’s Hikari pool, built by DatasourceConfig with Hikari’s default of 10) · Found in: F38-j WI-1.2, 2026-10-04 (F38-j finding F-001) · Severity: medium · Phase: 1
Repro: on forge-et-be main before the fix, add one more integration spec with its own context (distinct properties, no @DirtiesContext) and run the full ./gradlew build. On F38-j WI-1.2’s 8abd03a this failed 10 specs it did not touch (BuyingStandardAccountIdIT, the pricing, tag and timeline specs), each initializationError with FATAL: sorry, too many clients already. Sampling pg_stat_activity during a full build on 3621c8c showed a peak of about 293 connections.
Expected: a new integration spec’s context fits under the container’s connection limit, however many specs the suite has.
Actual: Spring keeps up to 32 contexts cached for the whole test JVM, each holding a 10-connection pool, so the suite’s demand can reach 320 against a limit of 300. The failure lands on whichever specs start their contexts last, not on the spec that added the context.
Fix: cap every pool in the test JVM only. src/test/resources/hikari-test.properties sets maximumPoolSize=5, and build.gradle’s test task sets hikaricp.configurationFile=/hikari-test.properties, which HikariConfig’s constructor loads from the test classpath before DatasourceConfig sets the connection details. A full cache then holds at most 160 connections. Production code and pools are unchanged.
Regression test: TestPoolBudgetIT asserts that a fresh HikariConfig and the application’s own DataSource bean both have a pool of 5, and that ContextCacheUtils.retrieveMaxCacheSize() × pool size + 50 headroom fits under the running container’s max_connections.
Resolved: 2026-10-04 — fixed in forge-et-be#181 (f0d1e98, final head d65b533). TestPoolBudgetIT fails all three features with build.gradle at main (pool size 10; 32 * 10 + 50 = 370 <= 300 false) and passes them with the fix. The full ./gradlew build passed on f0d1e98 (4644 tests) and on d65b533 (4645 tests), each with 0 failures and 2 skipped (the local build record on #181). backend-reviewer found no Blocker; its two Importants (name the spec an IT, and check the application’s pool rather than only a fresh HikariConfig) were fixed in d65b533, and the delta review found nothing new. Sampling pg_stat_activity every 3 seconds during that run, the Postgres containers seen peaked at 151 and 148 connections, against about 293 before the fix (another session was building at the same time, so one of the two may have been its run).
Lesson: L-030 — the same cause, a context cache times a per-context pool approaching the database limit; this fix applies the durable remedy L-030 names, and BUG-088 is added to its sources.
BUG-087 — The Users and Roles error panels show no reference number (correlation id)
Affects: F5-c, F5-d (forge-et-fe — src/features/users/UsersScreen.tsx read-failed and stale panels, src/features/roles/RolesRoute.tsx list error, src/features/roles/RoleEditorRoute.tsx load error) · Found in: the frontend-reviewer review of forge-et-fe#131 (the 2026-10-04 project audit, finding 17), 2026-10-04 · Severity: low · Phase: 1
Repro: make the users list read, the roles list read or a role read fail with a problem body that carries a traceId (for example a 500 INTERNAL_ERROR), and look at the error panel.
Expected: the panel is followed by the Reference <traceId> line, as the email screens’ error states show it (coding-standards §3.2: error text carries the correlation id, so a user who reports a failure gives support the request to look up).
Actual: the panels name the failure and offer Retry but show no reference, so a reported failure cannot be matched to its backend log lines.
Fix: render TraceLine (moved to src/ui by forge-et-fe#131) with traceIdOf(error) (src/lib/api/traceId.ts) after each of those panels, with test ids users-read-trace, users-stale-trace, roles-list-trace and roles-editor-trace; F5-c and F5-d each take their own screens in their next wave.
Regression test: per panel, a component test that the line shows the id when the problem carries one, is absent without one (asserted after the panel is visible), and is gone after a successful Retry — the MailboxRoute.test.tsx correlation-id rows are the pattern.
Resolved: 2026-10-04. Fixed in forge-et-fe#134 (1ab739a, final head 8e38754), which fixes all four panels in one PR rather than one in each feature’s next wave.
- The fix. Each panel is followed by
TraceLinewithtraceIdOf(error). The role editor shows the role read’s id, or the catalog read’s when the role’s error carries none. - Tests that failed before the fix. With the three screen files reverted, the six new component tests that expect the line all failed: 6 failed and 139 passed across the affected files.
- Other tests. Each panel’s absent-without-an-id case passes. The role editor’s two both-failed cases fail against a version that prefers the catalog’s id. The size matrix’s two Roles error states now answer with a 32-hex trace id, so the 320px, tablet and desktop checks cover the line. The users browser spec asserts both of its lines.
- Review.
frontend-reviewer’s first pass found one Important: the size matrix sent no trace id on the Roles states. It and every Nit were fixed, and the third delta review found none open. Lesson: none (one-off). The four screens were built whileTraceLinewas still private to the email feature, and forge-et-fe#131 moved it tosrc/ui.frontend-reviewer§ STACK-SPECIFIC CHECKS now flags any error panel for a failed request that has no reference line, so the next screen is checked for it.
BUG-086 — The role editor’s browser tab did not name the role
Affects: F5-d (forge-et-fe — src/features/roles/RoleEditorForm.tsx, src/app/RouteTitle.tsx) · Found in: the frontend-reviewer review of forge-et-fe#131 (the 2026-10-04 project audit, finding 1), 2026-10-04 · Severity: low · Phase: 1
Repro: open two role editors in two tabs (/roles/7, /roles/9).
Expected: each tab names its role — design-system §9 (Rev 36): on a record page the record’s name comes first once it has loaded (Tea Room — Roles — Empire Teas CRM).
Actual: every role editor’s tab read Role — Empire Teas CRM, and nothing let a screen pass its record’s name to the page title.
Fix: a record-title context in src/lib/recordTitle.tsx — useRecordTitle(name) publishes the loaded record’s name and clears it on unmount — read by RouteTitle, which puts it before the route’s own title; a record route’s handle.title is its section’s name (roles/:id → Roles). The role editor publishes the role’s saved name (its h1).
Regression test: RouteTitle.test.tsx (the record name first, the section title while it is not known, dropped on unmount) and the size matrix’s exact title on the four loaded role-editor page states.
Resolved: 2026-10-04 — fixed in forge-et-fe#131 (8371cc1). The four loaded role-editor page states assert Pinned role (page check) — Roles — Empire Teas CRM in the browser gate, which the code before the fix fails (it renders Role — Empire Teas CRM); the gate passed on 8722608.
Lesson: none — one-off: the record-title rule arrived with design-system §9 Rev 36 in the same audit, after the role editor shipped; RouteTitle and useRecordTitle now give every later record page the title by construction, and the size matrix asserts it exactly.
BUG-085 — A background session re-read that returns a different user leaves the mounted screens showing the first user’s data
Affects: F5-d, F5-k (forge-et-fe — src/lib/session/SessionProvider.tsx’s background re-read of GET /api/v1/me, F5-k FR-19; src/app/AuthenticatedGate.tsx; the TanStack Query cache installed by F5-d wave 1) · Found in: fixing BUG-083, 2026-10-04 · Severity: low · Phase: 1
Repro: in one browser profile, sign in as user A in tab 1 and open /roles. In tab 2, sign out and sign in as user B. Return to tab 1: the focus re-read of /me returns B.
Expected: tab 1 shows B’s session and only data read for B.
Actual: the re-read publishes B while the authenticated screens stay mounted. Clearing the cache (BUG-083’s fix) does not reach them: TanStack’s QueryCache.remove (query-core 5.104.1) destroys the query and notifies the cache’s listeners, not the query’s observers, so a mounted useQuery keeps its last result, A’s, until it re-renders and builds a fresh query. Unlike BUG-083 this reaches deployed environments: the other tab’s Entra sign-in changes the session cookie the two tabs share.
Fix: AuthenticatedGate keys the authenticated subtree (its PermissionProvider) by the signed-in user’s id, so a change of user remounts every screen below it and each useQuery starts against the cache SessionProvider has just cleared. Fixed in the same PR as BUG-083, whose clear it depends on.
Regression test: a component test that renders an authenticated screen reading through a query as user A, triggers the focus re-read answered by user B, and asserts that A’s data is gone and B’s read is made.
Resolved: 2026-10-04 — fixed in forge-et-fe#132 (85509ed, final head 088ef43). Filed in that PR as BUG-084 and renumbered when harness main filed BUG-084 first. src/app/sessionQueryCache.test.tsx’s BUG-085 case failed before the fix (the screen still showed secret of Ada after the re-read returned Ben). With SessionProvider’s clear but no key it still failed, and with the key it passes. Two same-user cases, a sign-in refresh and a focus re-read, assert that the cache is kept and nothing remounts (one data read); a mutant whose clear ignores the user id fails both. In Chromium on the local profile, tab 1 on /email as administrator.local, with its stub mailbox connected, kept that inbox on screen under the second user’s name after tab 2 signed in as all.permissions.test.local; on the final head it shows the second user’s connect screen. frontend-reviewer found no Blocker; its Importants (a same-user focus re-read case, and this renumbering) were fixed in 088ef43, and the delta review found the code clean.
Lesson: L-099 — every cache names the event that clears it. This bug adds that clearing a cache does not change a value a reader already holds, so the readers on screen must read again (here, by remounting them); L-099 is extended with it, and frontend-reviewer § STACK-SPECIFIC CHECKS gains the row.
BUG-084 — Flaky test: AuthenticatedGate’s unsafe-return-path case reads sessionStorage before the gate clears it
Affects: F5-a (forge-et-fe/src/app/AuthenticatedGate.test.tsx — “does not follow an unsafe kept value; it lands on the default page and clears it”, a BUG-054 regression test) · Found in: the Fast lane unit run of forge-et-fe#130 on head 0f51dca (run 37195614431), 2026-10-04 · Severity: low · Phase: 1
Repro: GitHub Actions ubuntu-latest runner, the whole vitest --run suite (112 files, about 220 s). The case failed at line 314: expect(sessionStorage.getItem(RETURN_PATH_KEY)).toBeNull() received "//evil.example/accounts". On macOS it passed 15 of 15 runs alone and 3 of 3 runs of src/app with three feature folders, and the full suite passed locally on the same head (1907 passed). forge-et-fe#130 does not change AuthenticatedGate or the session code.
Expected: the case asserts the stored value is cleared only after the gate has acted on it.
Actual: the case’s anchor is findByTestId('where') reading /, but the default page is / before and after the gate acts, so the anchor can resolve before the session read lands and the gate clears the value; the sessionStorage check then runs too early on a slow runner.
Fix: anchor on something the gate’s own action produces — wait for the value to be cleared (await waitFor(() => expect(sessionStorage.getItem(RETURN_PATH_KEY)).toBeNull())) after an anchor that proves the session read landed (the authenticated content), then assert the location is /.
Regression test: the changed case; it cannot be made to fail on demand, so record its fails-before as the CI run above and check it with a repeated loaded run.
Resolved: 2026-10-04 — fixed in forge-et-fe#132 (0d69278, titled test(BUG-083) because BUG-084 was filed on harness main while that PR was open). The same case also failed on macOS: after BUG-083’s fix and a merge of main, it failed in 2 of 3 full pnpm test runs at a load average of about 18, while unchanged main passed 2 of 2 at a similar load. The location anchor already proves the session read landed (WhereProbe renders only inside the authenticated outlet), so the case now waits for the clear itself: await waitFor(() => expect(sessionStorage.getItem(RETURN_PATH_KEY)).toBeNull()). The sibling case “follows a kept route once /me confirms the session, and clears it” had the same shape and waits the same way. After the change, the full suite passed in 3 of 3 runs at a load average of about 15, and twice more on the final head 088ef43.
Lesson: L-101 — the test checked as soon as an element appeared, while the effect that clears the kept route could still be pending. BUG-084 is added to its sources.
BUG-083 — The query cache survives a sign-out, so a second user in the same tab briefly sees the first user’s data (local and test stub sign-in only)
Affects: F5-d (forge-et-fe — the TanStack Query cache installed by F5-d wave 1; src/app/queryClient.ts, src/lib/session/SessionProvider.tsx, src/app/AuthenticatedGate.tsx) · Found in: the pre-PR review of F5-d-WI-1.2 and the review of forge-et-fe#128 (F5-d wave 1), 2026-10-04 · Severity: low · Phase: 1
Repro: on the local profile, sign in through the stub as a user who holds email.manage and access.manage (for example the default test principal) and open /email and /roles; sign out; in the same tab sign in through the stub as a different user and open the same screens.
Expected: the second user sees only data read for their own session.
Actual: the app-root QueryClient is never cleared when the session ends or the signed-in user changes, so cached reads from the first user’s session (the mailbox connection, the inbox list, the roles list, a role) can render for the second user until their own re-read lands. Deployed environments are not affected: Entra sign-in is a full-page redirect that reloads the app and starts an empty cache.
Fix: clear the query cache (queryClient.clear()) when the session ends or the signed-in user’s id changes, in the session layer, so no cached read outlives the session that made it. The clear goes in SessionProvider, not AuthenticatedGate (decided 2026-10-04, before the code was changed). SessionProvider is the one place every transition passes through: sign-out (signOut), expiry (the client’s 401 handler and the re-read’s 401), sign-in (refresh) and a re-read that returns another user (revalidate). It can clear the cache before it publishes the new state, so no render pairs the new session with the old cache. AuthenticatedGate sees the transitions only as re-renders, could clear only in an effect after the render that already read the cache, and is not mounted on /login between a sign-out and the next sign-in, so it cannot compare the two users. SessionProvider reaches the client with useQueryClient() from @tanstack/react-query, not from src/app, so the lib-is-a-leaf folder rule holds; it now needs a QueryClientProvider above it, which main.tsx already has. A refresh for the same user (LandingRoute picking up a new grant) does not clear.
Regression test: a component test that reads data as user A, signs out, signs in as user B, and asserts no read from A’s session renders and the cache is empty at sign-out.
Resolved: 2026-10-04 — fixed in forge-et-fe#132 (85509ed, c5987e1, final head 088ef43). While the PR was open, F5-c wave 1 (forge-et-fe#130) added QueryCacheReset, which cleared the cache in an effect once the session was unauthenticated. That fixed this bug’s sign-out and expiry cases on main, but could not fix BUG-085. By the lead’s decision, the clear stays in one place: c5987e1 removes QueryCacheReset, and SessionProvider clears before it publishes a sign-out, an expiry or a different user (F5-c WI-1.2 plan ## Notes). src/app/sessionQueryCache.test.tsx uses a client with the app’s settings. Before the fix, its sign-out and 401 cases failed with one query still cached, and with the cache check skipped, Ben’s screen after a sign-out and sign-in rendered secret of Ada; they pass with the fix. In Chromium on the local profile (administrator.local with its stub mailbox connected, then all.permissions.test.local in the same tab, with every mailbox and roles read held 4 s), the code before F5-c and this fix showed A’s connected inbox and the roles table to the second user during their own reads; the final head showed loading states. pnpm typecheck, pnpm check, pnpm check:boundaries and pnpm check:testids exit 0, and pnpm test passed on the final head (1908 passed, 1 skipped). frontend-reviewer found no Blocker; its findings were fixed in 088ef43 and the delta review found the code clean.
Lesson: L-099 — every cache names the event that clears it; F5-d installed a cache whose clearing event (the session ending, or the user changing) nobody named. BUG-083 is added to its sources, with BUG-085.
BUG-082 — Flaky tests: the two MailboxBackfillIT cases of BUG-077 fail in the full suite on a loaded machine (macOS)
Affects: F20-b (forge-et-be/src/test/groovy/com/empireteas/mailbox/MailboxBackfillIT.groovy) · Found in: the local full builds of forge-et-be#163’s head 705e3ef, 2026-10-04 · Severity: low · Phase: 1
Repro: run the full ./gradlew build on a Mac while other sessions run backend Testcontainers suites on the same machine (here five other backend suites at once, load average 40 to 59).
Expected: both cases pass whatever the machine’s load.
Actual: the first run, stopped by a one-hour limit before the suite finished, had failed both cases. “A back-fill continuation resolves the folders afresh too, on a task whose cache still holds the old ids” found rowCount(connectionId, inbox) == 0 where it expects 1 (message stub-message-8-121). “A run is bounded by time as well as pages: slow pages end it early, and a continuation finishes the job once” found continuedOnThisThread(...) [... [pages=1,] where it expects [pages=2,. The second run failed only the time-bound case (4,340 tests, 1 failure). Run alone (./gradlew test --tests com.empireteas.mailbox.MailboxBackfillIT), all 12 cases passed, and the next full build of the same commit passed (4,340 tests, 0 failures). forge-et-be#163 changes only how test helpers wait on SQS calls, which these cases’ assertions do not depend on. The symptoms are BUG-077’s, seen on a Windows laptop; BUG-081 is another mailbox IT that failed under load on a Mac.
Cause (confirmed 2026-10-04): both cases fail on the back-fill run’s time bound, which MailboxBackfillService measured with System.nanoTime() against half the visibility timeout (1 s in the mailbox ITs). The time-bound case is as BUG-077 describes. The continuation case is the same bound, not the ingest queue: backfill.run stores each message on the calling thread, and the run reads three pages from the case’s cursor, the last holding the inbox message. When the run passes 1 s after its first or second page, it hands the last page to a continuation on the queue and returns, and the case reads 0 rows. On an idle machine that run takes about 32 ms, so it needs a stall of about a second, which is why it failed only in the more loaded of the two builds. Forced on demand by arming the stub’s page delay (1100 ms in the time-bound case, 600 ms before backfill.run in the continuation case), the unfixed class failed both cases with the reported messages, [pages=1,] and rowCount(8, stub-message-8-121) == 0 after Continued … [pages=2,.
Fix: BUG-077’s second option. MailboxBackfillService reads the run’s time from the injected BusinessClock, and a new unit spec, MailboxBackfillServiceSpec, steps a MutableClock per page to pin where the run stops. MailboxBackfillIT’s time-bound case asserts that the run stopped early, after one or two pages (each stub page waits at least 700 ms, so two always outlast the 1 s bound), and its continuation case waits for the back-fill to complete and the queue to empty, then asserts folderCalls() >= 1, as the case before it already does.
Regression test: the two cases, run in a full build on a Mac under load (for example with a second full build running) without failing; the same forced page delays as above, which the fixed class passes; and MailboxBackfillServiceSpec.
Resolved: 2026-10-04 — fixed in forge-et-be#169 (bd8b030). With the forced page delays, the unfixed class failed both cases with the reported messages and the fixed class passed all 12. MailboxBackfillServiceSpec (5 cases) and MailboxBackfillIT (12) pass alone. Two full builds of bd8b030 on the Mac passed: ./gradlew build (10 min 12 s, load average 8 to 17) and ./gradlew cleanTest build --no-build-cache beside a second full backend suite whose run overlapped it throughout (19 min 7 s, load average 17 to 22), each 4,353 tests with 0 failures. In the second, the continuation case’s run took 304 ms, against 32 ms on an idle machine. /forge-review-pr’s backend-reviewer found no Blocker or Important; its four nits are applied in 6793ba3 (the time-bound case asserts pages in 1..2, the wall-clock trade-off is documented, the spec pins the conditional cursor writes) and two wording nits from the delta review in d62538d. The recorded full build of the final head d62538d passed with 4,353 tests and 0 failures, and its review record is approve. BUG-077, the same cases on Windows, stays open until a Windows build passes.
Lesson: L-125 — a test must not depend on how much work a real-time budget allows. BUG-081 wrote it, with this bug and BUG-077 among its sources. This change adds BUG-082’s fix to it: code that stops after a time budget reads an injected clock, so a unit spec steps the clock to pin where it stops. It also adds that to the lesson’s backend-reviewer row, and test-strategy.md § Flaky Tests cause 3 now points at the lesson.
BUG-081 — Flaky test: MailboxBackgroundAuditIT’s renewal case finds no subscription_renewed event in the full suite on a loaded machine (macOS)
Affects: F39 (forge-et-be/src/test/groovy/com/empireteas/mailbox/MailboxBackgroundAuditIT.groovy, the case “the renewal job writes mailbox.subscription_renewed as SYSTEM:mailbox-subscription-renewal (AC-10)”) · Found in: the first local build record for forge-et-be#164’s head 555afd5, 2026-10-04 · Severity: low · Phase: 1
Repro: run the full ./gradlew build on a machine that other sessions are also using for Testcontainers suites (here a Mac running about 18 other test containers at the same time). The case failed once in one full run; BUG-077’s MailboxBackfillIT case failed in the same run.
Expected: the case passes whatever the machine’s load.
Actual: at MailboxBackgroundAuditIT.groovy:219, rows.size() == 1 failed with rows empty: after runRenewalJob() there was no mailbox.subscription_renewed event for the connection the case had just set to expire in one hour. Run alone (./gradlew test --tests '*MailboxBackgroundAuditIT*' --tests '*MailboxBackfillIT*', 2026-10-04), all 8 cases of this class passed, and the next full build of the same commit passed (4,340 tests). PR #164 changed only documentation and a shell script, so the failure is not in code it touched. The nightly PR validation run on main passed on each of the five days to 2026-10-02. Cause: the handler gives MailboxRenewalService.renewDue a budget of half the visibility timeout, which the test profile sets to 2 s (MailboxIngestContextSpec), so 1 s. renewDueDeclared starts that clock before it runs the due query and checks it before every renewal, the first included, so a run whose query takes more than 1 s on a loaded machine reaches no subscription and records nothing. Other specs’ connections are not the cause: a surviving full-suite report from 2026-10-04 (MailboxBackgroundAuditIT’s JUnit XML) logs due=1, reached=1 for both of this class’s renewal-job runs, so the case’s own connection is the only one due. A run that skips a subscription for a deferred token or an exhausted limiter would log a WARN and does not depend on load, so it does not fit. The lapse case (“a lapse reached through the renewal job …”) runs the job the same way and has the same exposure. BUG-077 is the same kind of cause: a 1 s bound taken from the 2 s test visibility timeout, racing a loaded machine.
Fix: a test change only. The class’s runRenewalJob helper runs the job through MailboxRenewalJobHandler.handle until the case’s own connection is no longer due (renewed or lapsed), inside a bounded PollingConditions. This is what the service promises: a subscription a run did not reach stays due and the next run reaches it. Each attempt still enters through the handler, so the actor stays SYSTEM:mailbox-subscription-renewal, and rows.size() == 1 still proves one event. The renewal and lapse cases both use it. MailboxRenewalService is unchanged.
Regression test: a new case in MailboxBackgroundAuditIT holds LOCK TABLE mailbox_connection IN ACCESS EXCLUSIVE MODE on a second connection until pg_stat_activity shows the run’s due query waiting on it, then for the budget plus 500 ms, so the first run’s due query outlasts the budget every time (due=1, reached=0). It also checks that the stub made exactly one renewal call. It fails with the single-run helper and passes with the fix.
Resolved: 2026-10-04 — fixed in forge-et-be#168. Before the fix, the new case failed in both of two runs (./gradlew test --tests '*MailboxBackgroundAuditIT*BUG-081*'): rows was [], the original symptom, and the run logged due=1, reached=0. After the fix, all 9 cases of the class passed, the new case’s first run logged reached=0 and the next one renewed, and the PR’s local build record on its final head f763371 (with main merged in) passed the full suite (4,349 tests, 0 failures). backend-reviewer made the stall start only once the run’s due query is seen waiting, and capped the lock with lock_timeout; with the helper changed back to a single run, the reviewed case still failed (runs >= 2 with runs = 1, due=1, reached=0).
Lesson: L-125 — new: a test must not depend on how much work a real-time budget allows; the 1 s test budget (half the 2 s visibility timeout) is the trap. BUG-077 and BUG-082 are the open instances in MailboxBackfillIT, and backend-reviewer now flags a new or changed test that counts one run’s work against a sub-second budget.
BUG-080 — ScheduledJobRoutingIT reads the demo job’s processed_message row before the handler has committed it
Affects: F20-b (forge-et-be/src/test/groovy/com/empireteas/platform/scheduling/ScheduledJobRoutingIT.groovy, the row “three kinds on one queue each reach their own handler, and none reaches another’s”) · Found in: the pre-merge local build of forge-et-be#161 after merging main (148c9d2), 2026-10-04 · Severity: low · Phase: 1
Repro: on forge-et-be#161’s branch with main merged in, run ./gradlew test --tests 'com.empireteas.platform.scheduling.ScheduledJobRoutingIT'. It failed in the full build and again on its own (2 of 2), with recordedCount("demo-${run}") == 1 reading 0, while the demo handler had logged “Scheduled Job Consumed” for that key. On main alone the class passed (1 of 1).
Expected: the row passes whenever the handler records its key, whatever the timing.
Actual: the demo @SpringSpy appends the message to demoSaw before callRealMethod() runs, so the PollingConditions block can pass on demoSaw while the real handler has not yet committed its processed_message row. The row then read recordedCount once, outside any wait, and saw 0. The other specs that expect a recorded row (ConsumeAckIT, IdempotentReplayIT, ScheduledIdempotentReplayIT, ScheduledDispatchIT) wait for it inside eventually first; PoisonDlqIT reads it outside a wait, but only to assert 0 after the message has reached the DLQ, which a race cannot fail.
Fix: read the count inside its own PollingConditions(timeout: 20, delay: 0.2).eventually { assert recordedCount(…) == 1 }, with a one-line comment giving the reason. A test change only.
Regression test: the row itself: with the fix the class passed on the merged tree (2 of 2, exit 0) where it had failed 2 of 2 before.
Resolved: 2026-10-04 — fixed in forge-et-be#161, found by the pre-merge refresh the git conventions require; the same PR’s full local build record covers it.
Lesson: none — one-off: it is the only spec that expected a recorded row without waiting for it; the other four wait with eventually first, and no lesson or other Lesson: line names an assertion on a handler’s side effect made right after a spy saw the call.
BUG-079 — Flaky tests: browser tests that deep-link into a lazily loaded route time out on a loaded machine while the dev server transforms the route
Affects: F5-k (forge-et-fe/e2e/browser/access/route-guards.spec.ts), F20-a (e2e/browser/mailbox/connect-disconnect.spec.ts:127) and every browser spec whose first step is a deep link · Found in: F5-k wave 3, the wave-head pnpm e2e:gate (.forge/plans/F5-k/F5-k-WI-3.1-plan.md § Notes › Wave-head gate, on feature/F5-k-wave-3 until that branch merges), 2026-10-02 · Severity: low · Phase: 1
Repro: run pnpm e2e:gate in forge-et-fe while the machine is busy (load average 33–38 on 2026-10-02). route-guards.spec.ts › “a deep link to /email/mailbox as all.permissions.test.local@empireteas.example (holds email.manage) renders the Mailbox screen” failed once: the Mailbox h1 was not visible within the 5 s expect timeout, and the page still showed main: Loading… (src/app/RouteSkeleton.tsx). It passed on a re-run at normal load.
Expected: a deep-link test’s result does not depend on how busy the machine is.
Actual: under pnpm dev (Playwright’s webServer) Vite transforms a module the first time it is requested. The first spec to deep-link into a lazy route (src/app/router.tsx) asks for that route’s whole module tree, so it pays the transform cost inside its own 5 s wait. On a loaded machine that cost passed 5 s. The Chromium first-heading timeouts recorded under BUG-045 “Also seen” (Mailbox, Pricing, not-found, and the dev server’s “Importing a module script failed”) match this cause.
Fix: not a longer timeout (.forge/test-strategy.md § Flaky Tests). A new setup-project file, e2e/fixtures/routes.setup.ts, reads the lazy(() => import(…)) calls from src/app/router.tsx, opens the app shell, and imports each route module in the browser, which resolves only once the module’s whole tree has loaded. Vite keeps each transform for the life of the server, and every project that runs a browser spec depends on setup, so no spec pays the first transform. Only the test side changes; the built bundle is unchanged. Chosen over server.warmup.clientFiles in vite.config.ts, which warms in the background after the server starts and which Playwright does not wait for.
Regression test: the setup file itself. Proof by a forced failure (L-107): with a temporary Vite plugin adding 1.5 s to the transform of each src/features/ module (reverted after), connect-disconnect.spec.ts › “AC-22a/AC-12/AC-21” failed at line 127 with Timeout: 5000ms and main: Loading… when the setup file was absent, and passed with it.
Resolved: 2026-10-03 — fixed in forge-et-fe#122 (542bbcb; the frontend-reviewer fixes in f8c5aaf read every import() in router.tsx and fail the setup unless one module is found for each lazy() call, so a route written in another shape cannot drop out unnoticed). The forced failure above, then a full pnpm e2e:gate at both commits: 394 passed, 5 skipped, 0 failed. An earlier gate run at 542bbcb lost its local backend partway through and does not count.
Lesson: L-116 — new: under the dev server, the first browser test to reach a lazy route pays its transform; warm the routes in setup. It is the second bug with this cause (the Chromium recurrences under BUG-045), and frontend-reviewer dimension 2 now flags a lazy() the warm-up does not cover.
Also seen on the F26-b wave branch, before this fix reached it (2026-10-02): the F26-b Step 5e browser runs on forge-et-fe bad2c04 failed page states on /pricing with TypeError: Importing a module script failed. and timeouts under load, and the Stage 12 gate runs at f657c96 failed on timeouts while other sessions’ backend suites held the CPU; every one passed on re-run. F26-b filed it as its own BUG-078, which was dropped as a duplicate of this bug when main was merged into the wave branch (2026-10-04). Its records are in .forge/plans/F26-b/F26-b-WI-1.5-test-verify.md and the F26-b Decomposition Plan’s notes.
BUG-078 — Local and test queue provisioning waits on LocalStack with no time limit
Affects: FS-7 · Found in: the local build record for forge-et-be#147’s head 5b9d6b8, 2026-10-03 · Severity: low · Phase: 1
Repro: run the full ./gradlew build while two other backend suites run on the same machine (here three suites and two local stacks: four LocalStack and five PostgreSQL containers in a 7.7 GB Docker VM). A thread dump of the test JVM showed it starting a Spring test context and waiting in QueueProvisioner.createRedrivenQueue (QueueProvisioner.java:170, called from start() line 69), and no test result was written for more than a minute.
Expected: a provisioning call that gets no answer from LocalStack fails within a bounded time with a message naming the queue, so the suite fails fast instead of waiting.
Actual: each SQS call in QueueProvisioner (createQueue and getQueueAttributes, lines 141, 145 and 170) ends in CompletableFuture.join(). join() adds no bound of its own and ignores interrupts, so the wait lasts as long as the client’s own timeouts allow, which may be none. On this run the answers did come back and the build passed (4204 tests, 0 failures), but it took 24 minutes instead of about 10. A lost reply is bounded only by those client timeouts, which the code neither sets nor checks. QueueProvisioner is @Profile({"local", "test"}), so a laptop’s bootRun waits the same way; cloud environments do not use it (their queues are Terraform’s).
Fix (Tharinda’s decision, 2026-10-03: the first option): every SQS call in QueueProvisioner waits through one helper, await(future, queueName), which calls get with CALL_TIMEOUT, a constant of 3 minutes per call (Tharinda’s decision on review, raised from the 60 s first proposed). The value has to clear the client’s own limits. Spring Cloud AWS builds the client on netty-nio-client with SDK defaults, so a silent read gives up after about 30 s and is retried (the SDK’s documented defaults, not measured here). A call that stalls once and then answers can therefore take 30–65 s, and the slow run went over a minute with a context’s queue start-up unanswered on a run that passed. Three minutes is above the read timeout plus a retry, so only a reply that is truly lost trips it. The helper handles each outcome:
- No answer within the timeout: it cancels the future and throws an
IllegalStateExceptionnaming the queue and LocalStack. - SQS error: it reaches the caller as
join()delivered it, aCompletionExceptionwhose cause is the SDK’s exception, now with a message naming the queue. - Interrupt: it restores the thread’s interrupt flag and throws an
IllegalStateExceptionnaming the queue.
A package-private constructor takes the timeout so the unit spec runs in milliseconds; no configuration key was added. The SDK’s apiCallTimeout was not used. docs/messaging/README.md step 5 states the bound.
Regression test: QueueProvisionerSpec (T1) runs start() on its own daemon thread against a stubbed SqsAsyncClient and gives up after 20 s (5 s in the run before the fix shown below). A thread of its own is needed because join() ignores the interrupts Spock’s @Timeout sends: a first draft that relied on @Timeout(10) alone held the test JVM for 10 minutes on the unfixed code, until it was stopped. The spec has four cases:
- A DLQ create that is never answered.
- A main-queue create that is never answered (the DLQ answers).
- An interrupt while waiting.
- An SQS error, which must still reach the caller as the
CompletionException, with a message naming the queue.
On main’s code the first three fail with start() was still waiting on SQS after 5 s, and the fourth passes because it pins behaviour that already existed. With the fix all four pass (4 tests, 0 failures), again after the review changes.
Resolved: 2026-10-03, fixed in forge-et-be#161. The first commit 70cdfd9 used 60 s and passed ./gradlew build (4216 tests, 0 failures, 2 skipped). The review raised it to 3 minutes and named the queue on SQS errors in ffcd8e1 (a full ./gradlew build, 4216 tests, 0 failures); 8d53655 re-wraps a README paragraph, and the PR’s local build record names that head.
Lesson: L-124. Code that waits on a CompletableFuture from an SDK call to LocalStack or AWS uses get with a timeout, or a client built with apiCallTimeout. It never uses a bare join() or get() on a client with no call timeout. join() adds no bound of its own and ignores interrupts, so a test timeout cannot end the wait. The test helpers’ seven join() calls on SQS futures were the remaining instances; forge-et-be#163 replaces them with a bounded wait.
BUG-077 — Flaky tests: two MailboxBackfillIT cases changed for BUG-050 fail in the full suite (Windows)
Affects: F20-b · Found in: the local build record for forge-et-be#134’s merge head 9c42797, 2026-10-02 · Severity: low · Phase: 1
Repro: run the full ./gradlew build on a laptop that is busy (here a Windows laptop running Docker for the suite’s Testcontainers). The case “a run is bounded by time as well as pages: slow pages end it early, and a continuation finishes the job once” fails about once in the four full runs made on 2026-10-02.
Expected: the case passes whatever the machine’s speed.
Actual: continuedOnThisThread(...) is [... [pages=1,] where the case expects [pages=2,. The case arms a 700 ms delay per Graph page and relies on the run stopping once it is past half of the 2 s visibility timeout, at 1 s: two pages take 1.4 s, so the run stops after the second. That holds only while the first page plus its database and application work stays under 1 s, a margin of 300 ms; on a loaded machine the first page alone passed 1 s and the run stopped after one. The same commit passed the full suite in CI (gh workflow run "PR validation", run 37005717449). The case was last changed by 46dccd5 (BUG-050). It does not touch the code of forge-et-be#134, #135 or #137. A second case failed on the next run (2026-10-02, 21:25, nothing else running): “a back-fill continuation resolves the folders afresh too, on a task whose cache still holds the old ids” found rowCount(connectionId, inbox) == 0 where it expects 1 (message stub-message-8-121 not stored); it was also changed for BUG-050. Two full local runs on #134’s head 9c42797 failed in a row, on these cases only, while CI passed the same commit. Run alone (./gradlew test --tests '*MailboxBackfillIT*', on main’s mailbox code, 2026-10-02 21:30), all 12 cases pass, so they fail only inside the full suite: load or test order, not the code under test.
Fix: not decided. Two options: assert that the run stopped on the time bound (fewer pages than MAX_PAGES_PER_RUN, and at least one) rather than after exactly two; or make the time bound injectable so the case drives it with a fixed clock instead of real delays. The second removes the timing from the case. Update 2026-10-04: forge-et-be#169 fixes the same two cases for BUG-082, the Mac occurrence, with the second option; its section has the confirmed cause, which is the same for both cases here. This bug stays open until a full build on a Windows machine passes with that change.
Regression test: the case itself, run under load (for example with a second full build running) without failing.
For whoever checks it on Windows (Dinil, Biseni or Nanndish, once forge-et-be#169 has merged):
- On an up-to-date
mainofforge-et-be, run the full./gradlew buildtwice, at least once with Docker Desktop busy (for example a second backend suite running from another checkout). Both must pass with 0 failures. Run the class alone too:./gradlew test --tests com.empireteas.mailbox.MailboxBackfillIT --tests com.empireteas.mailbox.MailboxBackfillServiceSpec. - If either case fails again, read its run’s log lines before changing anything:
build/test-results/test/TEST-com.empireteas.mailbox.MailboxBackfillIT.xmlholds them, and the lines from the case’s own thread (Test worker) that startMailbox Backfillshow where each run stopped (Continued … [pages=N,) or finished (Completed). After the fix, a stop on the time bound no longer fails either case, so a new failure has a different cause: file it as its own bug with those lines. - To see the old failure on demand, as BUG-082 did (L-107): on the commit before forge-et-be#169, arm
stub.armPageDelay(Duration.ofMillis(1100))in the time-bound case, andstub.armPageDelay(Duration.ofMillis(600))just beforebackfill.run(connectionId)in the continuation case. Both fail. Onmainthe same edits pass. Revert them afterwards. - Windows-specific points: run Gradle from Git Bash or PowerShell in the repo root (
.\gradlew.bat buildin PowerShell); keep Docker Desktop running before the build starts, since Testcontainers needs it; and record the result withpython3 scripts/local-build-evidence.pyonly on a PR, never for this check alone. - When both builds pass, set this bug to
status: fixedwithowner: <your first name>intracker.yaml(insidetracker-write.sh begin…finish), and add aResolved:line naming the commit and the two builds’ test counts, plus**Lesson:** L-125., in a harness PR. Resolved: 2026-10-05 — fixed by forge-et-be#169 (merged ase7182f6), checked on a Windows 11 laptop with Docker Desktop, onforge-et-bemainat19416b71. Two full builds passed, each with 5,027 tests, 0 failures and 2 skipped:./gradlew buildbeside a second backend suite (./gradlew test --rerun-tasksfrom another checkout) whose run overlapped it,MailboxBackfillITincluded (41 min 32 s), and./gradlew build --rerun-taskswith nothing else running (34 min 6 s). In the loaded build, back-fill runs stopped on the time bound after one page (Continued … [pages=1,) and no case failed. Run alone,MailboxBackfillIT(12 cases) andMailboxBackfillServiceSpec(5) pass. An earlier full build in the laptop’s main checkout passed every test but failedspotlessGroovyCheck: 80 files in that working tree still had CRLF line endings, although.gitattributeschecks text files out as LF, so the two builds above ran in a fresh worktree of the same commit. Lesson: L-125 — this bug is one of its sources; the injected clock that forge-et-be#169 added for BUG-082 also holds on Windows under load.
BUG-076 — Boot’s form-content filter reads a signed-out PUT, PATCH or DELETE form body into the heap before authorization
Affects: FS-6 · Found in: BUG-042 (the backend-reviewer pass of /forge-review-pr on forge-et-be#139) · Severity: medium · Phase: 1
Number: filed as BUG-073, which harness#345 had already taken for another bug; renumbered BUG-076 on 2026-10-02. forge-et-be commit d35090a’s subject still says BUG-073.
Repro: with the backend running, open a raw connection and send a signed-out PUT /api/v1/anything with Content-Type: application/x-www-form-urlencoded, Content-Length: 10485760 and only the first 64 KB of the body.
Expected: a 403 straight away, without the body being read: CSRF refuses a signed-out write with no token.
Actual: no answer: the server waits for the rest of the body. Spring Boot’s FormContentFilter (on by default, ahead of Spring Security) parses a form body on PUT, PATCH and DELETE so it can expose it as parameters, and reads it whole into the heap with no size limit — thread dump on the laptop-mode backend, 2026-10-02: FormContentFilter.doFilterInternal → parseIfNecessary → FormHttpMessageConverter.read → StreamUtils.copyToString. Tomcat’s max-http-form-post-size limits only Tomcat’s own parameter parsing, not this read. A form POST and a JSON PUT were refused in 0.01 s. Medium rather than low: unlike BUG-042’s temporary files, this is heap, unbounded, and reachable by anyone, so a few large requests could exhaust the backend’s memory. Nothing is deployed yet.
Fix (Nanndish’s decision, 2026-10-02, shipped with BUG-042): spring.mvc.formcontent.filter.enabled: false in application.yml. Every write in the API takes a JSON @RequestBody (the PATCH endpoints on accounts, contacts and leads) or no body (DELETE /api/v1/mailbox/connection); nothing declares a form content type, and forge-et-fe’s API client sends JSON only.
Regression test: three SecurityChainIT rows (PUT, PATCH, DELETE) that send the 10 MB announced form body with only 64 KB and require the 403 within 5 s. All three got no answer on the unfixed config and pass with the fix; the account, contact, lead and mailbox endpoint ITs (33, 30, 45 and 27 tests) pass with the filter off.
Accepted, with a limit (the delta review of forge-et-be#139): a signed-out POST with an urlencoded body and a matching made-up CSRF cookie and header still has its body read by Tomcat when OAuth2AuthorizationCodeGrantFilter reads the parameters. That read is capped by server.tomcat.max-http-form-post-size (unset, so Boot’s 2 MB default), and Tomcat skips it when Content-Length is above the limit, so unlike BUG-042 and the bug above it is bounded. If it is ever lowered, use a small value such as 64 KB, never 0: Tomcat counts a multipart request’s form fields against the same limit, and 0 would empty PipedriveDryRunController’s importOwnerId.
Resolved: 2026-10-02 — fixed in forge-et-be#139 (d35090a), with BUG-042.
Lesson: L-088 — the same cause as BUG-042 and F4-a’s finding: a framework component that runs before sign-in reads the request body. L-088’s extension for BUG-042 names this filter too.
BUG-075 — Creating a lead with a receivedAt before year 1 is stored, or answers 500, instead of 400
Filed as BUG-063, renumbered BUG-069 and then BUG-075: main’s BUG-061 to BUG-072 merged first, in two batches (.claude/rules/tracker.md, L-057). The PR title uses the new number and the branch name keeps the old one; the code comments cite L-087 rather than a bug number, so a renumbering no longer touches the code.
Affects: F11-a · Found in: backend-reviewer on forge-et-be#135 (BUG-074), outside that PR’s diff; reproduced 2026-10-01 · Severity: low · Phase: 1
Repro: POST /api/v1/leads as a user holding lead.manage, with a body carrying "receivedAt": "0000-12-31T23:59:59Z", "0001-01-01T00:00:00+05:30" (year 0 in UTC) or "-1000000000-01-01T00:00:00Z" (Instant.MIN).
Expected: 400 VALIDATION_FAILED naming receivedAt, and no lead written, as for a receivedAt in the future.
Actual: the year-0 values answer 201 Created and the lead is stored with that date, and Instant.MIN answers 500, because PostgreSQL’s timestamptz cannot hold it. LeadController.createLead refuses only a receivedAt after now, so nothing bounds the past. Reproduced 2026-10-01 with three new LeadControllerIT rows against main (f67377b): expected:<400 BAD_REQUEST> but was:<201 CREATED> twice and but was:<500 INTERNAL_SERVER_ERROR> once. This is L-087’s rule (every LocalDate or Instant request value is range-checked before it reaches SQL), which this value missed; a search of origin/main on 2026-10-01 found no other date or time-stamp request value without a bound (LeadCreateRequest.receivedAt is the only one in a request body, and F26-a’s asOf is bounded).
Fix: in LeadController.createLead, beside the future check, refuse a receivedAt before 0001-01-01T00:00:00Z with an OUT_OF_RANGE entry on receivedAt (“must not be before 0001-01-01T00:00:00Z”). The future check already refuses anything after now, so the upper end needs no second bound. Decision (2026-10-01): OUT_OF_RANGE, as F26-a’s asOf answers: the value was read as an instant, and the year is checked after binding. The controller Javadoc’s list of create refusals names it. The check is on the REST path only, as the future check is: LeadIntakeService applies neither, by design (a capture device’s clock may run ahead), so the endpoint that first feeds the intake from outside (F36) must bound receivedAt itself (L-087).
Regression test: the three LeadControllerIT refusal rows above, each 400 VALIDATION_FAILED naming receivedAt with nothing written, and a test that the entry’s code is OUT_OF_RANGE and that 0001-01-01T00:00:00Z is stored as sent. The rows fail before the fix as shown.
Resolved: 2026-10-02 — fixed in forge-et-be#137. On main’s code the three rows failed (201 CREATED, 201 CREATED, 500 INTERNAL_SERVER_ERROR); with the fix LeadControllerIT passes, 49 tests, 0 failures, and ./gradlew build on head 3cfc0a8 passed with 3983 tests, 0 failures. /forge-review-pr 137: no Blocker; its wording nits applied, and Tharinda’s review on b343ea9 blocked only on the build record. After main was merged in, the head 5516ef5 passed the full CI suite (run 37005755988) and the local build record (4033 tests, 0 failures), with an approve review record; merged 2026-10-02 as cdd09db.
Lesson: L-087 — a receivedAt the rule’s “every LocalDate or Instant request value” covers but no check applied; BUG-075 is added to L-087’s sources. The F36 intake endpoint is where the next instance would come from (see Fix).
BUG-074 — A list filter on a date or time-stamp field answers 500, not 400, for a year the database cannot store
Filed as BUG-062, renumbered BUG-068 and then BUG-074: main’s BUG-061 to BUG-072 merged first, in two batches (.claude/rules/tracker.md, L-057). The PR title uses the new number and the branch name keeps the old one; the code comments cite L-087 rather than a bug number, so a renumbering no longer touches the code.
Affects: FS-10 · Found in: F26-a’s /forge-review-pr of forge-et-be#110 (F26-a finding F-007) · Severity: low · Phase: 1
Repro: a collection definition declares a date filter, rangeFilterable('startDate', FieldType.DATE), or a time-stamp filter, FieldType.INSTANT, and a client sends ?startDate=+9999999-01-01 or ?createdAtFrom=+300000-01-01T00:00:00Z.
Expected: 400 VALIDATION_FAILED with a TYPE_MISMATCH entry on the parameter, as for any other value that is not a valid date (docs/api/collections.md, the FieldType table: DATE accepts yyyy-MM-dd).
Actual: FieldType.parse reads the value with LocalDate.parse or Instant.parse, and both accept signed years of five or more digits. The value passes validation and reaches the query; PostgreSQL refuses it (its date type stops at year 5874897, timestamptz at 294276), so the request answers 500. No list endpoint declares a date or time-stamp filter yet, so nothing can reach this today; the first list screen that adds one (F1-b or F11-b) would. F26-a fixed the same gap on its own asOf parameter (forge-et-be#110); this is the shared copy.
Fix: in FieldType.parse, refuse a DATE value, or an INSTANT value read in UTC, whose year is outside 1–9999, by throwing IllegalArgumentException, so ListQueryResolver answers 400 with its existing TYPE_MISMATCH entry. The 1–9999 bound is L-087’s, and the one F26-a’s asOf uses. Decision (2026-10-01): TYPE_MISMATCH, not OUT_OF_RANGE. The declared format has a four-digit year (yyyy-MM-dd), so a value outside it is not of the type, and the resolver already turns every parse failure into that entry, so no new error path is needed. F26-a’s asOf answers OUT_OF_RANGE because its controller checks the year after Spring has parsed the date, which FieldType does not do. The FieldType table in docs/api/collections.md states the bound.
Regression test: ListQueryResolverSpec rows rejected with TYPE_MISMATCH: startDate=+9999999-01-01, startDateTo=+10000-01-01, startDateFrom=0000-12-31, startDateFrom=-10000-01-01, createdAtFrom=+300000-01-01T00:00:00Z, createdAtTo=+10000-01-01T00:00:00Z (the first instant past the range), createdAtFrom=0000-12-31T23:59:59Z, createdAtFrom=0001-01-01T00:00:00+05:30 (year 0 in UTC), and Instant.MAX and Instant.MIN to the second; accepted: 0001-01-01 to 9999-12-31, and 0001-01-01T00:00:00Z to 9999-12-31T23:59:59.999999999Z. The rejection rows fail before the fix, because the values are accepted.
Resolved: 2026-10-02 — fixed in forge-et-be#135. With FieldType.java put back to main, the first seven rejection rows failed (65 tests, 7 failures); with the fix, and the three rows the review added, ListQueryResolverSpec passes, 68 tests, 0 failures, and ./gradlew build on head b874f44 passed with 3991 tests, 0 failures. /forge-review-pr 135: no Blocker or Important; its four nits applied, and Tharinda’s review on d8b864e blocked only on the build record. After main was merged in, the head 5fa4b21 passed the full CI suite (run 37005735096) and the local build record (4041 tests, 0 failures), with an approve review record; merged 2026-10-02 as c8141e3. Time-stamp filters (INSTANT) had the same gap and are fixed by the same change.
Lesson: L-087 — this is the shared copy of the gap L-087 records for asOf, which F-007 named as a follow-up; BUG-074 is added to L-087’s sources.
BUG-073 — scripts/local-build-evidence.py cannot start Gradle (Windows)
Filed as BUG-061, renumbered BUG-067 and then BUG-073: main’s BUG-061 to BUG-072 merged first, in two batches (.claude/rules/tracker.md, L-057). The PR title uses the new number and the branch name keeps the old one; the script’s comment no longer cites a bug number, so a renumbering no longer touches the code.
Affects: FS-4 · Found in: F26-a wave 2, recording the backend build for forge-et-be#110 (F26-a finding F-006); reproduced again 2026-10-01 · Severity: medium · Phase: 1
Repro: on Windows, from a Claude Code session (its Bash and PowerShell tools set NoDefaultCurrentDirectoryInExePath=1), in a clean forge-et-be checkout, run python3 scripts/local-build-evidence.py --no-post (or without --no-post, on a branch with an open PR). From a terminal where that variable is not set, the script works.
Expected: the script runs gradlew.bat build and prints the record (or posts it on the PR).
Actual: it stops before the build with FileNotFoundError: [WinError 2] The system cannot find the file specified. main() starts Gradle by the bare name gradlew.bat (added in forge-et-be#101). When NoDefaultCurrentDirectoryInExePath is set, Windows stops searching the current folder for a bare program name, so the wrapper is not found. Checked 2026-10-01 on Windows 11 and Python 3.14.7: with the variable set, the bare name fails and .\gradlew.bat starts; with it unset, both start. The variable is set by the Claude Code tools, not at user or machine level. The team runs this script from Claude Code sessions, and every push to a ready backend PR needs it (.claude/rules/git-conventions.md, § The backend test suite runs locally), so on Windows the record could not be made without a workaround (putting the checkout at the front of PATH, as F26-a did).
Fix: start the wrapper by a path relative to the current folder. A small gradle_command() returns .\gradlew.bat when os.name == "nt" and ./gradlew otherwise, and main() uses it. The script already has to run from the repository root (its test-results path is relative). The backslash is needed because cmd.exe reads / as an option switch. The first version of the fix used the full path, and backend-reviewer found that it fails in a folder whose name holds an & but no space, because Python quotes an argument only when it holds a space and cmd.exe then treats the & as a command separator; the relative path never carries the folder name. Nothing else about the record changes. The self-test’s case count is now counted rather than typed, so it no longer needs editing when a case is added. forge-et-be#133 changes the same file: a test merge on 2026-10-01 conflicted only on that count line, which whichever PR merges second resolves by keeping the counted form.
Regression test: two --self-test cases. The first checks the command: .\gradlew.bat on nt, ./gradlew elsewhere. The second runs only on Windows and starts the wrapper itself: in a temporary folder whose name holds a space and an &, with a stub gradlew.bat that exits 7 and with NoDefaultCurrentDirectoryInExePath=1, the launch must return 7. Against the old bare name, both fail, the second with not started: [WinError 2]. On Windows, python3 scripts/local-build-evidence.py --no-post also stops with FileNotFoundError before the fix, from a Claude Code session, and runs the build after it.
Resolved: 2026-10-02 — fixed in forge-et-be#134. Against the old bare name the two new self-test cases fail (the launch case with not started: [WinError 2]); with the fix --self-test passes, 10 cases, with the variable set and unset. The fixed script itself ran ./gradlew build on Windows from a Claude Code session on head 3a73618 (3979 tests, 0 failures) and posted the PR’s record, and it recorded #135 and #137 the same way. /forge-review-pr 134: no Blocker; its two Important findings (the stated cause, a test of the launch itself) and both nits applied in 3a73618. Tharinda’s review on da5b4f6 blocked on the conflict with forge-et-be#133, resolved in 6b10c9f as the review asked (--self-test 15 cases on Windows). After main was merged in again, the head 9c42797 passed the full CI suite (run 37005717449) and the local build record (4029 tests, 0 failures; two earlier local runs on it failed only on BUG-077’s flaky cases), with an approve review record.
Lesson: L-046 — the script was run only on its author’s platform, where the variable is unset, so the Windows failure surfaced only when a Windows developer ran it from Claude Code; BUG-073 is added to L-046’s sources. Any script a Claude Code session starts on Windows should launch a program in the current folder by a path (.\name.bat), never by bare name.
BUG-072 — The “sync lapsed” sentence uses the type/meta text style instead of type/body
Affects: F20-b (forge-et-fe/src/features/email/SyncLapsedNotice.tsx, styled by syncIndicator.module.css:38-42) · Found in: the review of F20-b WI-2.2 (plan ## Notes, item 7), deferred to the developer since 2026-09-25; filed by the lead’s decision of 2026-10-02 · Severity: low · Phase: 1
Repro: let a mailbox’s sync lapse and open /email.
Expected: the sentence a user must read and act on is body copy, at type/body (design-system §2 and §4).
Actual: it is set at type/meta, the size for secondary details such as the last-synced time.
Fix: use type/body for the lapsed sentence, keeping text/secondary if the design system allows it at that size. A visual change: confirm it against design-system §4 first.
Regression test: a SyncLapsedNotice unit test that the sentence uses the body type token.
Resolved: 2026-10-03 — fixed in forge-et-fe#124, with BUG-071. A new .sentence rule in syncIndicator.module.css sets the lapsed sentence at type/body and keeps text/secondary; .meta stays for the last-synced time, the only text F20-b AC-26 puts at type/meta. Design-system §4 covers shape, space, elevation and motion and says nothing about type or colour; §3 makes type/body the working size, and §2 limits only text/muted and text/faint by size, so text/secondary stays (6.2:1 on surface/card, 5.7:1 on the tone/danger tint). The sentence is rendered by SyncLapsedBody, which the inbox notice and the connected panel share, so the panel’s lapsed sentence changes too, matching its reconnect reason (ConnectedPanel.module.css .reason). The new SyncLapsedNotice.test.tsx row reads the rule block of the class the sentence carries; on main’s code it failed (expected '…font-family: var(--et-type-meta…' to contain 'var(--et-type-body-face)') and with the fix it passes. In the browser (pnpm dev against ./gradlew bootRun, the mailbox lapsed with stub-renewal-failure, 1024px): inbox-sync-reason measured 11.5px / 15.525px / 500 on main’s code and 13px / 19.5px / 450 with the fix, colour rgb(110, 95, 78) both times; /email/mailbox’s mailbox-sync-reason also measures 13px / 19.5px. The review of forge-et-fe#124 added margin: 0 to .sentence and .meta, so the flex gap token alone sets the spacing: the browser’s 13px paragraph margins went, and the inbox notice card measured 151.5px before and 125.5px after, the panel card 344.0px and 318.0px, with the sentence unchanged at 13px / 19.5px. With the margins gone the .sync gap was the only space between the indicator row, the sentence and Reconnect, so the same review, by Tharinda’s decision, raised it from space/3 (6px) to space/4 (9px), design-system §4’s default gap: at 1024px the space above and below the sentence went from 6px to 9px, the inbox notice card from 125.5px to 131.5px and the lapsed panel card from 318.05px to 324.05px; the healthy panel (one row, 246.55px at 1024px and at 768px, where the row does not wrap) is unchanged.
Lesson: none — one-off: the F20-b WI-2.2 stylesheet gave the last-synced time and the lapsed sentence one shared .meta rule, and the wave review caught it (forge-et-fe#55, item 7) and deferred it as a visual change rather than missing it. No other Lesson: line or lesson records a sentence a user acts on set at a secondary-detail type token.
BUG-071 — A sync lapse is not announced through the /email live region
Affects: F20-b (forge-et-fe/src/features/email/EmailRoute.tsx:230) · Found in: the review of F20-b WI-2.2 (plan ## Notes, item 2), deferred to the developer since 2026-09-25; filed by the lead’s decision of 2026-10-02 · Severity: low · Phase: 1
Repro: with a screen reader on /email, let the mailbox’s sync lapse while the page is open.
Expected: the screen reader hears that mail has stopped arriving, as it hears other inbox status changes through the inbox-status live region.
Actual: the lapsed notice appears on screen, but nothing is announced.
Fix: put one short lapse sentence into the existing inbox-status polite region when the status becomes lapsed. A change to what screen readers hear: confirm the wording first.
Regression test: an EmailRoute unit test that a change to the lapsed state writes the sentence into the inbox-status region, and that the healthy state leaves it empty.
Resolved: 2026-10-03 — fixed in forge-et-fe#124, with BUG-072. The lapse has its own polite live line, inbox-sync-announcement: visually hidden, rendered in every state, empty on mount, and holding LAPSED_ANNOUNCEMENT (syncState.ts), “New mail is not arriving until you reconnect your mailbox”, exactly while the lapsed notice shows; one lapseShown condition in EmailRoute gates both, and the line empties when the lapse clears. inbox-status keeps the list’s sentence only, as before. A separate line, rather than a sentence added to inbox-status as the PR first did, because the list’s sentence replaces “Loading your mail” in that line right after a fast load, which can cut the lapse off before it is read, and some screen readers re-read a whole line on every page change (the review of forge-et-fe#124). The line relies on the connection read starting in loading on mount, so the lapse arrives as a change; a read answered from a cache on mount (the TanStack Query move, F5-c WI-1.2) would mount it already full and nothing would be announced. A lapse is detected only on a connection read — page load, or the re-read after Connect, Reconnect or Retry — because /email does not poll, so a lapse that happens while the page is open is shown and announced at the next read. The wording was approved by Tharinda on 2026-10-03. Four EmailRoute.test.tsx rows: the lapse is written into the line, which already existed empty before the connection read answered, and leaves it after Reconnect; a healthy mailbox’s line is empty; a page change leaves the line’s text node untouched and inbox-status never holds the lapse; and a lapsed mailbox whose list read failed still shows the notice and fills the line. All four fail on main’s code and on the PR’s first version (Unable to find an element by: [data-testid="inbox-sync-announcement"]) and pass with the fix. In the browser on a lapsed mailbox the line holds the sentence and inbox-status holds the list’s sentence.
Lesson: none — one-off: the screen already had the always-mounted polite line that coding-standards §3.2 requires, and the wave review caught the missing sentence (forge-et-fe#55, item 2) and deferred it because the wording needed a decision; nothing was missed by a check. No other Lesson: line or lesson records a status shown on screen with no announcement. The nearest note, BUG-026’s “a screen reader hears only the heading after the focus move”, is about where focus goes after a re-read, and was left to the developer as a UX choice.
BUG-070 — SchedulerConfigGuard does not refuse a cloud shape with the SQS listener switched off
Affects: F20-b (forge-et-be/src/main/java/com/empireteas/platform/scheduling/SchedulerConfigGuard.java) · Found in: the review of F20-b WI-2.1 (plan ## Notes, E4), deferred to the developer since 2026-09-25; filed by the lead’s decision of 2026-10-02 · Severity: low · Phase: 1
Repro: start a cloud profile with spring.cloud.aws.sqs.listener.auto-startup=false and a scheduled job configured.
Expected: the backend refuses to start, as it does when et.messaging.scheduler-queue is missing, because no scheduled job would ever run.
Actual: it starts, and timed work (subscription renewal, polling) silently never runs.
Fix: needs a design: the cloud-shape test contexts set auto-startup=false on purpose (WI-2.1 D4b), so the guard must tell a test context from a deployed one, for example by refusing only on the cloud profiles a deployment uses.
Fix design (decided 2026-10-05): a deployment activates no Spring profile — the image runs java -jar and a cloud shape is the default profile with ET_ENVIRONMENT set — so the profile route above has nothing to key on. Instead the guard tells them apart by et.environment: when the shape carries a scheduled job and the queue is set, it reads spring.cloud.aws.sqs.listener.auto-startup with Boot’s Binder (as Spring Cloud AWS’s SqsProperties binds it, so false, off, no and the environment-variable spelling all count) and refuses false unless et.environment is test. A deployment always names dev, uat or prod, so any other value, a new or misspelt name included, refuses. No new configuration key and no task-definition change. The cloud-shape test contexts that set auto-startup=false and label themselves dev change to test; no main code reads the value. Files: SchedulerConfigGuard.java, SchedulerConfigGuardIT.groovy, the cloud-shape ITs that set auto-startup=false with et.environment=dev, and the application.yml comment on the listener. The same switch silencing the scan-result listener is BUG-097, fixed separately.
Regression test: a SchedulerConfigGuardIT row that a deployed cloud profile with the listener off refuses to start, and that the test contexts still start.
Resolved: 2026-10-05 — fixed in forge-et-be#193 (81e68c39, review fixes 746bca8d, and the delta review’s two comment Nits in the merge of main, ae01dd58). SchedulerConfigGuard reads spring.cloud.aws.sqs.listener.auto-startup with Binder and, when it is false, refuses to start unless et.environment is test; the message names the key, SPRING_CLOUD_AWS_SQS_LISTENER_AUTO_STARTUP and the label it found ((not set) when there is none). application.yml and EtProperties now say the label test is reserved for test contexts and never used by a deployment. The 12 cloud-shape test files that switched the listener off with et.environment=dev now say test; EntraLaptopGuardIT keeps dev, because EntraLaptopGuard is PriorityOrdered and refuses first. Regression tests: SchedulerConfigGuardIT “a deployed cloud shape (et.environment=#label) … refuses to start” (dev, uat, prod, production) and four bare-bean-factory rows (the false/FALSE/off/no spellings refused under dev and allowed under test, both environment-variable spellings, no label at all, and the listener left on). On the first commit, ten rows failed on the old guard (17 tests completed, 10 failed) and all 17 passed with the fix; after the review fixes the spec has 18 rows, all passing.
Lesson: L-067 (a consumer’s activation conditions are what switch it off in cloud; its guard must cover every one) — extended with this bug and BUG-097, and a backend-reviewer row added.
Decision: reading et.environment in the guard, which coding-standards.md §5 otherwise discourages and the forge-et-be#193 review flagged as an Important, was approved by Tharinda on 2026-10-06 over a separate acknowledgement setting: a deployment runs with no profile, the cloud-shape tests cannot use the test profile without losing the cloud bean graph, and a flag would be a way to switch the check off that nobody sees. §5 now limits the exception to a startup guard comparing the label with test, so it does not become a pattern.
BUG-069 — The first Graph delta call fetches the whole mailbox page by page, with no filter or page size
Affects: F20-b (forge-et-be/src/main/java/com/empireteas/mailbox/GraphMailboxProvider.java, the first delta request) · Found in: the review of F20-b WI-2.1 (plan ## Notes, E3), deferred to the developer since 2026-09-25; filed by the lead’s decision of 2026-10-02 · Severity: low · Phase: 1 · cloud-gated
Repro: connect a large real mailbox (dev, once Graph is reachable) and watch the first poll.
Expected: the first round costs a bounded number of Graph calls.
Actual: the first delta call sends no $filter and no Prefer: odata.maxpagesize, so the first round can page through far more than the back-fill window needs. The first-round cost is already a row in cloud-gated-validation.md §B F20-b.
Fix: add Prefer: odata.maxpagesize (and a received-date $filter if Graph’s delta query allows it for messages) to the first delta call; measure on dev.
Regression test: a GraphMailboxProviderSpec row that the first delta request carries the header; the cost itself is measured on dev (register row).
Resolved: 2026-10-05 — fixed in forge-et-be#192.
- The filter. Graph’s message-delta documentation allows
$filter=receivedDateTime ge {value}on a round’s first request and keeps it in everynextLinkanddeltaLink. The first delta request now carries it with the back-fill window’s start, truncated to seconds as the back-fill page’s filter is. - The page size. Every delta request, the first and each followed link, sends
Prefer: odata.maxpagesize=50afterIdType="ImmutableId". Graph keeps a round’s query options in the link but not its headers, so the value is sent on every request. Back-fill paging is unchanged, because it already sets its size with$top. - The interface.
MailboxProvider.deltaMessagestakesInstant receivedFrom, read only when no link is given.MailboxDeltaPollJobHandlerpasses the window start its in-window check already uses. The check stays, because the bound is fixed when a round starts.LocalMailboxProvideraccepts the argument and ignores it. - The 5,000 limit. Graph documents that a filtered delta round returns at most 5,000 messages. Only the first round after a connect is backed by the back-fill, which pages the window once. A round the poll restarts later, after a refused or dropped link, is bounded by the cap alone, so in a folder with more than 5,000 messages in the window it may miss mail whose notification was also missed (before this fix a restart was unfiltered: slow but complete). Whether to narrow a restarted round’s filter or start a gap back-fill is a design question left open for the lead by the review of forge-et-be#192. That question, how the cap behaves, and whether Graph honours the filter, the page size and the immutable-id preference together are in the §B F20-b register row, which stays open for dev.
- A refused fresh round (the review of forge-et-be#192). The filter and page size make a 400 on a round’s first request possible. Before, the poll read any 400/404/410 as Graph refusing the stored link: with no link to drop, it logged
Mailbox Delta Link Discardedand counted no failure, so a rejected request would have stopped the poll quietly. A refusal now drops the link only when there was one; on a fresh round it fails that connection’s poll, where it is counted and, when every connection fails, logs theMailbox Delta Poll All FailedERROR line. - Tests. Two
GraphMailboxProviderSpecrows check the first request (window filter plus bothPrefervalues, in order) and a followed delta link (the page-size value, no added filter). With the request restored to$selectonly and no page size, both failed:Request URI expected:<…&$filter=r…, andPreferwas[IdType="ImmutableId"]against[IdType="ImmutableId", odata.maxpagesize=50]. With the fix, both pass. Two further rows show the back-fill’s first page and followed cursor carry the immutable-id value only. AMailboxDeltaPollJobHandlerSpecrow checks, with a seven-day window rather than the default, that a fresh round on each folder asks for mail from the window’s start. A second row checks that a 400 on a fresh inbox round, withsentitemsanswering, counts the connection as failed, logs the all-failed ERROR line and logs no discarded link; without the handler fix it failed at the ERROR check, and with it it passes. - Build.
./gradlew buildon the first commit,f783517c, passed with 5031 tests, 0 failures and 2 skipped; the local build record for the PR’s head is on forge-et-be#192.
Lesson: none — one-off: the F20-b WI-2.1 review found the unbounded first round (plan ## Notes, E3) and deferred it on purpose until it could be measured on dev, so no check missed it. The provider’s other Graph reads are already bounded ($top on the back-fill page, single items elsewhere), and no other Lesson: line or lesson records a list call left without a filter or page size.
BUG-068 — The back-fill compare-and-set has no generation column and no database CHECK
Affects: F20-b (forge-et-be/src/main/java/com/empireteas/mailbox/repository/MailboxConnectionRepository.java:215-217, migration V18) · Found in: the review of F20-b WI-2.1 (plan ## Notes, E2), deferred to the developer since 2026-09-25; filed by the lead’s decision of 2026-10-02 · Severity: low · Phase: 1
The entry was filed naming V20; the back-fill columns were added by V18__mailbox_connection_sync_state.sql (V20 is subscription_reason_code), corrected when the bug was taken.
Repro: read the back-fill claim in MailboxConnectionRepository: it compares against a NULL-expected value to claim a back-fill run.
Expected: the claim compares against a value that changes on every run, and the database refuses a state the code never writes.
Actual: the compare-and-set relies on a NULL-expected compare with no generation counter, and V18 has no CHECK constraint on the columns it relies on. A reconnect (resetBackfill) sets the cursor back to NULL, so a run that read its first page before the reconnect still matches expected = NULL afterwards and can take over the new back-fill. It is correct today only because nothing else returns the cursor to NULL while a run is in flight; a later change that reuses the claim could double-run a back-fill without the database noticing.
Fix: add a backfill_generation column, compare-and-set on it, and add a CHECK constraint in a new migration. Both are schema changes; the lead approved them on 2026-10-03, at low priority, so anyone may take the bug.
Plan (written before the code, 2026-10-04):
- A new migration (the next free version at merge time, V34 as of writing):
backfill_generation BIGINT NOT NULL DEFAULT 0,CHECK (backfill_generation >= 0), andCHECK (backfill_completed_at IS NULL OR backfill_cursor IS NULL)— a finished back-fill keeps no cursor, whichcompleteBackfillalready guarantees. Existing rows hold generation 0 and already satisfy both checks. - The generation is a version counter for the back-fill columns: every statement that writes them (
recordBackfillCursor,completeBackfill,resetBackfill) setsbackfill_generation = backfill_generation + 1, and the first two match onbackfill_generation = :expectedGeneration AND backfill_completed_at IS NULLin place of theNULL-expected cursor compare. A run reads the generation with the connection at its start and adds one after each write it wins. A reconnect bumps the generation, so a run from before it can no longer write. MailboxConnectionmaps the column read-only (as the other sync columns);MailboxIngestServiceandMailboxBackfillServicecarry the generation; the audit payload is unchanged (the cursor’s old value is still the one the run read, because no write to the cursor leaves the generation unchanged).docs/data-model.md§9 gains theALTERbeside V18’s.- Tests: a repository IT for the regression (below, plus a pre-reconnect run refused after
resetBackfill); the existingMailboxBackfillIT,MailboxBackfillServiceSpec,MailboxIngestAuditITandEmailMessageSchemaITrows updated to the new signatures and column. Regression test: a repository IT that two concurrent claims of the same generation let exactly one through, and that the CHECK refuses a value the code never writes. Resolved: 2026-10-05 — fixed in forge-et-be#195 (ca0d5b05, withmainmerged in asdf391bfeand the review’s Nit fixes in2236d433;/forge-review-prwithbackend-reviewerfound no Blocker or Important). The plan above was followed; the migration ships asV35__mailbox_backfill_generation.sql, becausemain’sV34__buying_standard_certification.sql(F26-b) merged first.
- Regression test:
MailboxBackfillGenerationIT“a run that read the back-fill before a reconnect writes nothing after it, though both expect no cursor” fails with the reset’sbackfill_generation + 1removed (recorded == 0read1: the stale run’s write landed) and passes with it. Its two concurrent-write cases pass either way.EmailMessageSchemaITpins the column and both CHECKs. - Build:
./gradlew buildonca0d5b05: 4569 tests, 0 failures, 0 errors, 2 skipped. The PR’s local build record on the final head2236d433: 5037 tests, 0 failures, 0 errors, 2 skipped. - Manual QA on the
localprofile: a first connect imports only the last 30 days; reload, sign-out and new mail add no duplicates; four reconnects in a row, one with mail that arrived while disconnected, re-import with no duplicates;backfill_generationmoved by exactly one per reset plus one per page each time; both CHECKs refuse the invalid values. Locally the stub mailbox gets its default messages only whenet.mailbox.public-base-urlis set, which the example local config does not say. Lesson: L-141 — a new lesson: a compare-and-set on a value another statement can set back to an earlier state lets a stale writer through, so compare a counter every write moves on, the reset included. BUG-102 is its second instance, in the delta poll;backend-reviewerchecks for it.
BUG-067 — After a re-consent, other backend tasks keep serving the narrower mailbox token until it expires
Filed as BUG-060 and renumbered BUG-067: main’s BUG-058 to BUG-066 merged first (L-057).
Affects: F20-a · Found in: the post-merge /forge-review-pr of forge-et-be#39 (BUG-016’s fix, merged with no recorded review), Important I-3, 2026-10-01 · Severity: medium · Phase: 1
Repro: with more than one backend ECS task (the deployed shape; one task locally cannot show it), let a user re-consent the mailbox to add calendar and contacts. The task that handles the re-consent clears its token cache. Have another task, or an SQS ingest worker, make a calendar or contacts call for the same connection within the token’s lifetime.
Expected: every task uses a token that carries the newly granted scopes as soon as the re-consent completes.
Actual: MailboxConnectionService.java:234 calls tokens.invalidate(connectionId), which clears the ConcurrentHashMap in MailboxTokenService (:108, :290) on that task only. Every other task keeps serving its cached, narrower access token until it expires, up to about an hour, which is BUG-016’s failure in a multi-task form. The class comment (:48-52) accepts per-task caching for refresh counts, not for stale scopes. The single-task race is covered by BUG-023’s row lock.
Fix (lead decision, Tharinda, 2026-10-02): store the scopes the token was issued with (success.grantedScopes()) in CachedToken, and refuse a cache hit that does not cover the scopes the caller needs, so a stale narrow token is refreshed on any task. Rejected: the scopes in the cache key (every caller must pass them), and a cache lifetime of a few minutes (more refreshes, and the window stays). Add a test with two service instances that counts provider calls (L-099).
Regression test: two MailboxTokenService instances sharing one connection: after a re-consent handled by one, the other’s next token for a wider scope is not the cached narrow token.
Resolved: 2026-10-03 — fixed in forge-et-be#154 (cfa66a0, with the review fixes 0041c04 and f05b070, and 98227e0 for the lead’s decision below). CachedToken keeps the normalised scopes its token was issued with and the connected_at of the consent it was issued under; a hit that does not cover the scopes the caller needs is a miss and refreshes, on whichever task holds it. A caller names its scopes through the new accessTokenFor(long, Collection<String>); the one-argument forms ask for MailboxScopes.REQUIRED, so no existing caller changed. The review (backend-reviewer) found that a fresh token still short of a needed scope would be refreshed on every request; the fix skips the refresh while a live token from the same consent is cached, so the provider is asked once per consent and token lifetime. Lead decision (Tharinda, 2026-10-03), on the reviewer’s escalation: a fresh token without a scope in MailboxScopes.REQUIRED moves the connection to RECONNECT_REQUIRED / SCOPES_CHANGED, as AC-10 does when the stored consent lacks one, instead of leaving it ACTIVE with mail silently failing; a token short only of a scope beyond the required set leaves the connection alone. Regression test: MailboxReconsentCacheIT “after a re-consent on one instance, another instance’s next token for a wider scope is not its cached narrow one” fails with the scope check removed (otherTaskRefreshes.get() == 2 read 1: the second instance served its cached narrow token) and passes with it; MailboxTokenServiceSpec adds the served and refreshed rows (the refreshed row also fails without the check), the scope-not-consented row, the no-second-refresh row and the required-scope SCOPES_CHANGED row. Proving it across real ECS tasks is a cloud-gated row (§B F20-a).
Lesson: L-099 — every cache names the event that clears it, and in-memory state is per ECS task. BUG-067 is its third instance (after BUG-016 and BUG-050) and is added to its source list.
BUG-066 — The mailbox “could not connect” banner comes back when Chrome restores the page from its back/forward cache
Affects: F20-a · Found in: the manual Chrome check for BUG-036 (forge-et-fe#99) · Severity: low · Phase: 1
Repro: in Chrome, on the built CRM (pnpm build, then pnpm preview; the dev server’s HMR socket keeps a page out of the back/forward cache), signed in with laptop Entra sign-in (F5-i). Open /email/mailbox?connect=failed: the banner “We could not connect your mailbox. Nothing was changed. Try again” shows and the URL becomes /email/mailbox. Type another site in the address bar, press Enter, then press Back.
Expected: the banner does not show. It reported a failed connect the user has already seen, and the URL no longer carries the marker.
Actual: the banner shows again. DevTools → Application → Back/forward cache says “Successfully served from back/forward cache”, and the URL is /email/mailbox with no ?connect=. MailboxRoute.tsx reads the marker into useState and removes it from the URL (F20-a AC-20), but nothing clears the state when Chrome restores the frozen page. This is the cause BUG-036 fixed on the login page (L-109). Found on the developer’s laptop, 2026-10-01; the screen also showed its “You do not have access to mailbox settings” panel, because that user’s role on the local database lacked email.manage, which is unrelated.
Fix: as useLoginMarker in LoginRoute.tsx: an effect in MailboxRoute that listens for pageshow and calls setNotice(undefined) when event.persisted is true, removing the listener on cleanup. Same accepted side effect as BUG-036: a Back from Microsoft’s consent page after “Try connecting again” also restores the screen without the banner (the button’s busy state already clears on the same event, useConsentRedirectBusy). Confirm with the developer before building, as BUG-036’s was. Two points from the second review of forge-et-fe#99 (L-109): check what the banner sits above, since the screen’s mailbox connection read also comes back frozen, and re-read it on the restore before dropping the banner if a stale state could show; and this would be the fourth copy of the pageshow/persisted listener (two in LoginRoute.tsx, one in useConsentRedirectBusy.ts), so consider one shared src/lib hook, which both features may import (AD #31).
Fix as decided (the developer, 2026-10-04, options 1A and 2A): (1) A restore re-reads the connection, then drops the banner. The connection read does come back frozen: src/app/queryClient.ts sets refetchOnWindowFocus: false, so nothing re-reads it on a restore, and the frozen body could offer Connect for a mailbox connected in another tab since, or show a connection that lapsed meanwhile. On pageshow with persisted, MailboxRoute calls the screen’s focus-restoring refetch (BUG-026: the settled panel stays up with its actions busy) and calls setNotice(undefined) once the re-read settles, whatever it returns, the order useLoginMarker uses. (2) One shared hook, src/lib/browser/usePageRestore.ts, calls its callback on pageshow with persisted and removes its listener on unmount; all four sites use it (useSignInRedirectBusy and useLoginMarker in LoginRoute.tsx, useConsentRedirectBusy.ts, and MailboxRoute), and their existing persisted: true / persisted: false rows show the move changes no behaviour.
Regression test: MailboxRoute.test.tsx rows: from /email/mailbox?connect=failed, dispatching new PageTransitionEvent('pageshow', { persisted: true }) inside act leaves the settled screen (its heading visible) with no mailbox-connect-failed panel; with persisted: false the panel stays. Both fail before the fix. A persisted: true restore also issues a second connection read and a persisted: false one does not; the shared hook gets its own unit test. Plus the manual Chrome check above on the built app, since Playwright’s Chromium disables the back/forward cache.
Resolved: 2026-10-05 — fixed in forge-et-fe#147, as decided above. With MailboxRoute.tsx put back to origin/main, four rows of MailboxRoute.test.tsx fail (the three BUG-066 restore rows, including a re-read that fails, and the BUG-024 retry row whose banner a restore now drops); with the fix all 64 pass. src/lib/browser/usePageRestore.test.ts covers the hook. Manual Chrome check on pnpm build + pnpm preview, laptop Entra sign-in as CRM01: after /email/mailbox?connect=failed, leaving for another site and pressing Back no longer shows the banner.
Lesson: L-109 — the same cause as BUG-036, on the mailbox screen. L-109 now names the shared usePageRestore hook as the way to clear a one-time message on a restore, and frontend-reviewer dimension 15 flags a pageshow listener of its own.
BUG-065 — Flaky test: the nav-permissions shown-side row can fail with “route.fetch: Test ended” under a parallel gate run (Windows)
Filed as BUG-059 and renumbered BUG-065: main’s BUG-059 to BUG-064 merged first (.claude/rules/tracker.md, L-057). The branch that filed it keeps the old number in its earlier commit messages.
Affects: F5-b · Found in: the BUG-036 pnpm e2e:gate run on forge-et-fe#99 · Severity: low · Phase: 1
Repro: a full pnpm e2e:gate run on forge-et-fe ead8b91, against a fresh local database, on a Windows laptop. The row e2e/browser/shell/nav-permissions.spec.ts › “AC-4: an entry renders when its key IS present (shown-side via the /me route-mock)” failed once; the serial row after it was skipped. The same spec passed 5 of 5 with --repeat-each=5 --workers=1.
Expected: the row passes on every run.
Actual: Playwright failed it with route.fetch: Test ended … while running route callback. mockPermissions (e2e/browser/auth/support.ts) answers GET /api/v1/me by calling route.fetch() and reshaping the body. A second /me read, such as the re-read on window focus, was still in that fetch when the test’s assertions finished, so the callback outlived the test. The product behaves correctly; this is a test race (the L-101 shape: an action outliving the state the test waited for).
Fix: end the test with await page.unrouteAll({ behavior: 'ignoreErrors' }), as Playwright’s message suggests, or have the helper wait for the in-flight /me reads to settle before the row’s last assertion. Test-only.
Regression test: the row itself, run under parallel load (the full gate, or --repeat-each with several workers) with no failure.
Resolved: 2026-10-03, fixed in forge-et-fe#123 (4a723f4) together with BUG-057, whose Resolved line has the detail.
- Fix: the change is in the helper, not in each test.
mockPermissions’ callback ignores the errors that mean its test has ended and rethrows any other, so no test needs to end withpage.unrouteAll. - Cause, corrected: the cause is the context closing when the test ends. A second
/mealone does not break the callback. - Evidence: the new
mock-permissions.spec.tsreproducesroute.fetch: Test ended.deterministically before the fix. This row passed 20 of 20 runs in each of Chromium and WebKit with 4 workers, and the full gate passed.
Lesson: L-119. A route callback that awaits the network can outlive its test.
BUG-064 — The browser-gate hook does not see a standard-input shell named by -, a stdin device path, a flag that takes an argument, or a shell outside its list
Affects: F40 (forge-et-harness/.claude/hooks/guard-browser-gate.mjs) · Found in: the second adversarial review of the BUG-044 fix (2026-10-01) · Severity: medium · Phase: 1
Repro: from a forge-et-fe worktree with no passing gate result, through .claude/hooks/guard-browser-gate.sh: echo 'gh pr create -t x -b y' | bash - (also | sh -, | bash -eo pipefail, | bash +o posix, | bash --rcfile /dev/null, | bash /dev/stdin, | dash), { bash -; } <<< 'gh pr create -t x -b y', bash /dev/stdin <<< '…' and source /dev/stdin <<< '…'. All were already allowed on origin/main.
Expected: refused — each shell reads gh pr create from standard input and runs it in the fe worktree.
Actual: allowed with no gh occurrence. The option reader takes - and the stdin device paths (/dev/stdin, /dev/fd/0, /proc/self/fd/0) as a script file; a flag bundle ending in o (-eo, -euo), +o/+O, --rcfile and --init-file do not consume their argument, which is then read as the script; dash, ksh, mksh, ash and busybox sh are not in POSIX_SHELLS (dash is installed in Git Bash); and source/. reading standard input is not checked. The piped forms contradict the hook README’s statement that … | bash is refused. The Fast lane’s description check still fails such a PR, so this is a false pass in the early warning, not a way past the gate.
Fix: in the option reader, treat a - operand and the stdin device paths as standard input; make any flag bundle containing o or O, +o/+O, --rcfile and --init-file consume the next word; widen the shell list to dash, ksh, mksh, ash and busybox sh; treat source/. of a stdin path as reading standard input.
Regression test: a refuse test per form above in .claude/hooks/tests/guard-browser-gate.test.mjs, each failing before the fix.
Resolved: 2026-10-05, fixed in forge-et-harness#457. The option reader takes a lone - as the end of the options, and a script operand of /dev/stdin, /dev/fd/0 or /proc/self/fd/0 as standard input; each o in a flag bundle, --rcfile and --init-file take the next word, and so does each O for bash and sh only (sh may be bash; in zsh -O is a plain flag and dash rejects it, so for other shells it takes none and a -c after it is still read — the PR’s review found ksh -O -c '…' and dash -O -c '…' allowed when every O took an argument); dash, ksh, mksh, ash and busybox sh are shells; source / . of a stdin path reads standard input and runs in the calling shell (in a pipeline, where bash runs it in a subshell, a cd it reads does not carry over). bash -- - and source - open a file named -, so they stay allowed. node --test .claude/hooks/tests/guard-browser-gate.test.mjs: 422 pass; the 40 new tests run against the unfixed hook: 30 fail, every repro form among them.
Lesson: L-095 — another false pass in the command parser, found by the adversarial review of the BUG-044 change rather than by the hook’s design cases; BUG-064 is added to its source list.
BUG-063 — Flaky test: the sidebar-collapse browser test can fail when a /me route handler is still running as the test ends
Affects: F5-b (forge-et-fe/e2e/browser/shell/sidebar-collapse.spec.ts, the mockPermissions helper in e2e/browser/auth/support.ts) · Found in: a pnpm e2e:gate run for BUG-045 (2026-10-01) · Severity: low · Phase: 1
Repro: one full gate run in forge-et-fe at 9059af6 (BUG-045 merged with main) failed [chromium] sidebar-collapse.spec.ts:67 “BUG-003: collapsed rail centres the logo and nav glyphs …” at 0 ms with route.fetch: Test ended raised inside mockPermissions’ page.route('**/api/v1/me', …) callback (e2e/browser/auth/support.ts:66); the serial file’s other two tests were then skipped. The next run at the same commit passed (371 passed). It has not been reproduced on demand.
Expected: a route handler still waiting on route.fetch() when a test ends does not fail a test.
Actual: the handler’s in-flight route.fetch() throws after the test it belonged to ended, and Playwright reports the error against the test running at that moment. The log also shows two “Importing a module script failed” errors from the Vite dev server near the end of the same run.
Fix: to decide once it recurs: Playwright’s own suggestion is await page.unrouteAll({ behavior: 'ignoreErrors' }) at the end of each test that calls mockPermissions (or in an afterEach in that file), so an in-flight handler cannot outlive its test.
Regression test: to be written with the fix; until then, record each recurrence here with the gate run’s head and the failing test.
Resolved: 2026-10-03, fixed in forge-et-fe#123 (4a723f4) together with BUG-057, whose Resolved line has the detail.
- Fix: the change is in the helper rather than an
afterEachper file.mockPermissions’ callback ignores the errors that mean its test has ended and rethrows any other. - Cause, corrected:
sidebar-collapse.spec.tsdoes not callmockPermissions. The failing callback was most likely left over from an earlier test in the same worker (landing.spec.tsandnav-permissions.spec.tsrun before it inshell/), and its rejection landed on the test running at that moment. - Regression test: the new
mock-permissions.spec.tsreproduces theroute.fetch: Test ended.rejection deterministically before the fix. - Evidence:
sidebar-collapse.spec.tspassed 20 of 20 runs in each of Chromium and WebKit, and the full gate passed.
Lesson: L-119. A route callback that awaits the network can outlive its test.
BUG-062 — The browser-gate hook does not read a cd or a standard-input shell behind a prefix command
Affects: F40 (forge-et-harness/.claude/hooks/guard-browser-gate.mjs) · Found in: the BUG-044 fix’s second implementation pass (2026-10-01) · Severity: medium · Phase: 1
Repro: from a forge-et-fe worktree with no passing gate result, through .claude/hooks/guard-browser-gate.sh: (1) echo 'gh pr create -t x -b y' | timeout 60 bash (also | "time" bash); (2) nohup cd /w/harness; gh pr create -t x -b y (also with sudo, env or exec in front of the cd). Both shapes were already allowed on origin/main.
Expected: (1) refused — timeout runs bash, which reads gh pr create from the pipe and runs it in the fe worktree; (2) refused — an external program cannot change the calling shell’s directory, so the gh still runs in fe.
Actual: allowed. (1) The stdin-shell check reads a shell only in command position, so a shell that is the argument of a prefix command is not seen. (2) The hook counts the cd behind the prefix as if the shell ran it. Shape (2) was not probed against real bash in the pass that found it. The Fast lane description check still fails such a PR.
Fix: commandStart reads each prefix command with its own options and operands — command (-p only, so command -v bash runs nothing), builtin, exec (-c, -l, -a NAME), and the programs nohup, env, sudo, timeout (its duration too), nice and time (a quoted "time" or a path, which bash runs as a program) — and returns the command they run, so a shell there is read as a stdin shell. A word like X=1 after a prefix command is an assignment only after env or sudo; after the others it is the command name, as bash reads it. A new readCommandPrefix also reports the first prefix that is a separate program, or exec; behind one, a cd is no directory change and an export GH_REPO=… or a bare GH_REPO=… sets nothing (nohup export GH_REPO=… and env GH_REPO=… were allowed by the same cause). command, builtin and the reserved time run in the calling shell, so a cd behind them still counts. Checked against bash 5.2 before the change: nohup, env, timeout, nice and exec in front of cd or export leave the shell unchanged; command cd, builtin cd and time -p cd change the directory. Reading every shell name in argument position is rejected: it would refuse git log | grep sh.
Regression test: hook unit tests that both repro shapes are refused, and that git log | grep sh and timeout 60 pnpm test stay allowed.
Resolved: 2026-10-05, fixed in forge-et-harness#459.
- Regression test: the
prefix commands (BUG-062)block in.claude/hooks/tests/guard-browser-gate.test.mjs. Againstorigin/main’s hook 26 of its 41 tests fail; the other 15 pass on both, since they cover commands that must stay allowed (git log | grep sh,timeout 60 pnpm test,command -v bash) and shapesmainalready read. With the fix the whole file passes (431 tests, the review cases below included). - Same cause, fixed with it:
nohup export GH_REPO=…harnessandenv GH_REPO=…harnessbefore agh pr createwere read as settingGH_REPO, so an fe PR was allowed as a harness one. - The docker hook reads commands through the same
commandStart, so it now seestimeout 60 docker run,sudo -u root docker runandnice -n 5 docker run, and, throughreadCommandPrefix, no longer takesenv X=…;ornohup export X=…;as setting a label variable. Six cases are added toguard-docker-resources.test.sh;timeoutand “sudowith options” leave its accepted gaps (.claude/hooks/README.md, the hook’s header,docker-resource-labels.mdRev 2). - From the PR’s adversarial review (fresh context, no Blocker): bash takes
timeand!as reserved words only while every earlier word was reserved, soX=1 time cd …andcommand time cd …leave the directory unchanged and… | X=1 time -o t bashreads the pipe;readCommandPrefixnow follows that rule.GH_REPO=… command;sets nothing and is no longer kept. Both false passes were onmaintoo. - Left as accepted gaps (README, the Fast lane CI check still fails such a PR): wrappers outside the list (
setsid,stdbuf,doas),sudo -s/sudo -iwith no command,env -C <dir>in front of a nested shell, andsudo -D <dir>in front ofgh. All four are allowed onmaintoo. - Shell suites on this machine (Windows):
guard-browser-gate.test.shpasses, 84 passed, 0 failed.guard-docker-resources.test.shas written cannot find its temporary workspace on Windows, onmainas well, so its cases were run through a copy that uses Windows paths and calls the Node helper directly: 88 passed, 0 failed.Repo hygieneruns both suites on Linux.
Lesson: L-095 — a shell-parsing guard misses the forms its author did not think of, so it is reviewed adversarially and its gaps are listed. BUG-062 is another instance and is added to its source list. What was new this time is that a shape was recorded as a bypass without being run in real bash. Running it here (bash 5.2) confirmed the expected results before the fix was written.
BUG-061 — The browser-gate hook does not read $(…) or backticks inside ${…} or $((…))
Affects: F40 (forge-et-harness/.claude/hooks/guard-browser-gate.mjs, lexDollar) · Found in: the adversarial review of the BUG-044 fix (2026-10-01) · Severity: medium · Phase: 1
Repro: from a forge-et-fe worktree with no passing gate result, through .claude/hooks/guard-browser-gate.sh: : ${x:-$(gh pr create -t x -b y >&2)}, and echo $(( $(gh pr create -t x -b y >&2; echo 1) + 1 )). Both are allowed on origin/main; the same shape in a case subject (case ${x:-$(gh pr create …)} in *) :;; esac) became reachable with BUG-044’s case support.
Expected: refused — bash expands the inner $(…) and runs gh pr create in the fe worktree.
Actual: allowed with no occurrence found: lexDollar treats ${…} and $((…)) as opaque text.
Fix (decided 2026-10-05 — parse, do not refuse): read the body of ${…} with readArith, the matched-pair reader $((…)) and $[…] already use, so every $(…) and backtick in it is parsed and quoted text is skipped when finding the closing }, as bash does. Rejected: refusing any ${…} that holds a substitution, because it would block ordinary commands such as echo "${x:-$(date)}".
Regression test: hook unit tests that the three repro commands are refused, and that echo ${HOME} and echo $((1 + 2)) stay allowed.
Resolved: 2026-10-05 — fixed in forge-et-harness#458. lexDollar’s ${ branch now calls readArith('{', '}', '${') in place of its brace counter. The $((…)) half was already read on main before this fix: BUG-044’s commit f2596911 (forge-et-harness#342) added readArith for arithmetic, so only the ${…} repros still passed. A side effect: a quoted } inside ${…} (${x:-'}'}, "${x:-"}"}") no longer ends the expansion early, matching bash 5.2. New tests in guard-browser-gate.test.mjs (describe('command substitutions inside ${…} (BUG-061)')): the three repros read as a gated forge-et-fe gh pr create; nine more ${…} forms (backtick, double-quoted, nested, #, /, an array index, a here-document body) each yield one occurrence; twelve allowed forms, a cd inside the substitution kept in its own shell, and three unclosed forms refused as unparseable. node --test .claude/hooks/tests/guard-browser-gate.test.mjs: 410 pass, 0 fail; before the fix, 15 of the new tests failed. bash .claude/hooks/tests/guard-browser-gate.test.sh: 84 pass, 0 fail (AC-H6 skipped, no --fe-checkout). Through the real hook from a forge-et-fe checkout, main’s parser found no occurrence in : ${x:-$(gh pr create -t x -b y)} and the case-subject repro; the fixed parser reads both and decides them. The /forge-review-pr adversarial pass then found a remaining false pass: inside double quotes or an unquoted here-document, bash runs a $(…) or backtick within single quotes inside ${…} ("${x:-'$(echo RAN)'}" prints 'RAN'), while still not closing the ${…} at a } in those quotes; the first fix skipped the quoted span. lexDollar and readArith now carry an inDq flag, set from lexDouble and scanUnquotedHeredoc, and in that context the span is scanned for substitutions; unquoted, single quotes still hide the text, as in bash. Four more occurrence tests and two more refusal and allow tests; the four failed before this change. The delta review of that fix found the same gap in arithmetic, already on main: bash reads an arithmetic body as if within double quotes, so $(( '$(…)' )), $[ '…' ], (( '$(…)' )) and a for (( … )) header run a single-quoted $(…) before the syntax error; tryDparen and $[ now call readArith with inDq. Five more occurrence tests, which failed before. node --test .claude/hooks/tests/guard-browser-gate.test.mjs: 421 pass, 0 fail.
Lesson: L-095 — the adversarial review of the BUG-044 change found this false pass, as L-095 says a command-parsing guard’s own tests will not. BUG-061 is added to its sources.
BUG-060 — The capture app’s error screens render no <main> landmark, so the shared page check cannot run on them
Affects: FS-1 (the capture shell’s error screens: src/app/RouteErrorElement.tsx, src/app/AppErrorBoundary.tsx, src/app/FailureSurface.tsx) · Found in: the /forge-review-pr pass on forge-et-pwa#44 (BUG-043), 2026-10-01 · Severity: low · Phase: 1
Repro: in forge-et-pwa at forge-et-pwa#44’s head, run expectPageClean(page) (e2e/support/pageCheck.ts) on the router’s error screen (block the capture route’s chunk, as the BUG-043 row in e2e/browser/shell/a11y.spec.ts does).
Expected: the screen has one <main> landmark, like every other page, so assistive technology can jump to its content and the shared check runs.
Actual: the check fails on “no <main> on the page”: the root route’s error element replaces AppLayout, which is what renders <main>, and FailureSurface renders a plain <div role="alert">. The 404 branch of the same element has the same shape. Axe does not report it, because its region and landmark-one-main rules are best-practice, not WCAG 2.1 AA.
Fix: to be decided — render the error screens’ content inside a <main> (in FailureSurface and the 404 branch, or once in a wrapper both use), then switch the BUG-043 row to the shared expectPageClean. Consider adding the error screens to the capture app’s page-state list so the size matrix covers them at every size (L-106).
Regression test: the BUG-043 row in e2e/browser/shell/a11y.spec.ts using expectPageClean(page), which fails today on “no <main>”.
Resolved: 2026-10-02 — fixed in forge-et-pwa#47 (first filed there as BUG-058 by the post-merge backfill, which had not yet seen this entry). FailureSurface and the 404 branch render inside their own <main>, and the BUG-043 row uses expectPageClean at phone, tablet and desktop widths. With the two source files put back to origin/main, the three route-error browser rows and three new unit assertions (strict getByRole('main'), covering the 404 branch the real router cannot reach) fail; with the fix they pass. Browser gate: 126 passed. The error screens are still not in the page-state list (L-106); the row’s hand-listed widths cover them.
Lesson: L-106 — the browser gate checks only the page states in its list, as for BUG-043.
BUG-059 — The browser-gate hook lets a PR through when a redirect follows a { … } or ( … ) group that changes directory
Affects: F40 (forge-et-harness/.claude/hooks/guard-browser-gate.mjs, the bash parser) · Found in: the BUG-044 fix’s own probe pass (2026-10-01) · Severity: medium · Phase: 1
Repro: from a forge-et-fe worktree, with no passing gate result, run { cd /w/harness; } >/dev/null | cat; gh pr create -t x -b y (also with 2>&1 in place of >/dev/null) through .claude/hooks/guard-browser-gate.sh. Reproduced on origin/main as well as on the BUG-044 branch.
Expected: refused — the cd runs inside a pipeline, in its own subshell, so the gh pr create still runs in the fe worktree.
Actual: allowed, as if it were a harness PR. A redirection after a group is stored as a separate empty command, so the | is recorded on that empty command rather than on the group, and the group’s cd is counted as running in the same shell as the gh. A false pass, not a false refusal. The Fast lane description check still fails such a PR, so a PR cannot merge this way, but the hook’s refusal is skipped.
Fix: to be decided — keep a redirection attached to the group it follows, so a following | applies to the group; add allow and refuse tests for { … } >x | y, ( … ) 2>&1 | y and { … } >x; gh ….
Regression test: hook unit tests that the repro (both redirect forms) is refused, and that { cd /w/harness; }; gh pr create from the harness stays allowed.
Resolved: 2026-10-05 — fixed in forge-et-harness#455. The bash parser stores a redirection or here-document that comes right after a { …; } group, a ( … ) subshell, a case … esac or a (( … )) on that item, so a following | or & marks the group and its cd or GH_REPO no longer reaches the gh. The tree walk reads a group’s redirections and here-documents before its body, as bash performs them, so a gh inside a substitution there is still found. The same cause also let through { cd …; } >x & gh …, { cd …; } <<EOF | cat, a GH_REPO set in a redirected group, and a redirected case (read as uncertain before, so refused, now read as running apart). A ( … ) subshell’s cd never reached the gh, so that form was not a false pass. node --test .claude/hooks/tests/guard-browser-gate.test.mjs: on origin/main the 7 new rows fail (the six refusal rows and the row that checks the redirection is stored on the group), 389 pass; with the fix all 396 pass (398 after the review added two rows pinning that a cd in a group’s body does not reach a gh in its redirection, as bash runs the redirection first), including the allow rows { cd /w/harness; }; gh …, { … } >/dev/null; gh …, { … } 2>&1; gh … and { … } >/dev/null && gh … from the fe worktree and the harness. With real git from a forge-et-fe checkout, readOccurrence read both repro forms as a harness PR (allow) on main and as forge-et-fe (gated) with the fix. bash .claude/hooks/tests/guard-browser-gate.test.sh on Windows (Git Bash, jq 1.8.2): 84 passed, 0 failed, AC-H6 skipped (no --fe-checkout).
Lesson: L-095 — another false pass in the command parser found by probing it adversarially rather than by its design cases; added to that lesson’s source list. The review of this fix found a second instance of the same cause in if/for/while blocks, filed as BUG-098.
BUG-058 — About 16 CRM mailbox browser tests fail when the local backend runs with et.mailbox.public-base-url set, and nothing says why
First diagnosed as a race with the connect-time back-fill; the frontend-reviewer on forge-et-fe#105 found the real cause, and #105 was closed unmerged. The Actual and Fix below are the corrected ones.
Affects: F20-b (the CRM’s mailbox browser and seam specs in forge-et-fe/e2e/) · Found in: the pnpm e2e:gate runs for BUG-045 and BUG-046 (2026-10-01) · Severity: medium · Phase: 1
Repro: on Windows 11, with forge-et-be at origin/main (6d26b86) on the local profile, run pnpm e2e:gate in a forge-et-fe checkout at origin/main plus an unrelated change (BUG-045 at 0c5d885, BUG-046 at 062f236). Both runs fail the same 16 tests (17 with one WebKit size-matrix row), all mailbox: mailbox.seam.spec.ts “a connected caller with no mail is 200 with items []”, connect-screen.spec.ts (BUG-015, AC-21), full-journey.spec.ts, inbox-a11y.spec.ts (AC-29 and AC-30, empty and empty-filtered), inbox.spec.ts (wave 1 verification, AC-21, AC-23 ×2, AC-24 ×2, AC-25b), and sync-indicator.spec.ts; 20 tests are skipped instead of the expected 5, because serial specs stop after a failure. The same suite passed eight full gate runs on 2026-09-29 against an older backend (03830cb).
Expected: the suite passes on a machine set up as e2e/README.md describes.
Actual: the backend in those runs was started from the developer’s local launch entry with --args='--et.mailbox.public-base-url=http://localhost:8080'. With that key set, MailboxConnectionService.attemptSubscription records an ACTIVE subscription on a stub connect and LocalMailboxProvider.seedDefaultMailbox seeds four “Stub back-fill …” messages, so the connected mailbox shows mail and “Receiving mail” at once. The mailbox specs are written for a plain bootRun, where a connect records subscription NONE and the inbox stays empty (“Waiting for mail”). The tests and the app are both correct; the setup was wrong, and nothing in the suite said so. The backend’s application-local.yml.example adds to it: it says that outside laptop mode setting et.mailbox.public-base-url “would change nothing”, which is no longer true of the stub path.
Fix: test-setup only in forge-et-fe, no assertion changed: e2e/fixtures/auth.setup.ts checks, after the sales rep signs in (not the default principal, whose connection mailbox.seam.spec.ts asserts at the same time from the seam project, which does not wait for setup), that a stub connect records subscription NONE, disconnects again, and otherwise stops the setup project (and every browser project that depends on it) with a message naming the key and the restart command; e2e/README.md § Running it locally states the prerequisite. Owed separately in forge-et-be: correct the application-local.yml.example comment.
Regression test: the setup project fails with the message against a backend started with the key, and passes against a plain bootRun; the full pnpm e2e:gate passes against the plain start.
Resolved: 2026-10-01 — fixed in forge-et-fe#107 (forge-et-fe#105, the race-based fix, closed unmerged). At b1d0125, pnpm exec playwright test --project setup against a backend started with the key failed “authenticate as sales rep” with the message (Received: "ACTIVE"), and passed 8 of 8 against a plain bootRun; the full pnpm e2e:gate passed 371, 5 skipped. The local launch file now carries a crm-be-e2e entry (plain bootRun) beside the developer’s own, for gate runs. Still owed in forge-et-be: correct the application-local.yml.example comment that says the key changes nothing outside laptop mode.
Lesson: L-108 — when tests fail on every branch, check how the local services were started before changing the tests.
BUG-057 — Flaky browser test: the /me permission mock can read a response the page has already dropped
Affects: F5-b · Found in: the browser gate on forge-et-fe#101’s head d7cec79, 2026-10-01 · Severity: low · Phase: 1
Repro: run the browser gate. Once in a while e2e/browser/shell/landing.spec.ts “AC-12: a zero-permission session lands on the fail-closed empty state” fails after under a second with apiResponse.json: Response has been disposed at e2e/browser/auth/support.ts:73.
Expected: the test passes on every run.
Actual: mockPermissions calls route.fetch() and then response.json(). When the page drops the request in between (a second /me while the first is in flight, or a navigation), Playwright disposes the response and the handler throws, which fails the test. The same spec passed 38 of 38 with --repeat-each=10 on the same head, so this is a race in the test helper, not in the app.
Fix: catch the disposed-response error in mockPermissions and let the dropped request go (route.abort() or return), or read the body with route.fetch() in a try and fall back to route.continue().
Regression test: landing.spec.ts with --repeat-each over many runs, plus a spec that triggers a second /me while the first is in flight.
Same cause as BUG-063 and BUG-065: all three are a mockPermissions callback (e2e/browser/auth/support.ts) still in route.fetch() when the request is dropped or the test ends. They are kept as separate bugs because each names a different failing test, but one change to the helper fixes all three.
Resolved: 2026-10-03, fixed in forge-et-fe#123 (4a723f4), shared with BUG-063 and BUG-065.
- Cause, corrected: the response is not disposed because the page drops the request. A probe showed that an abort, a navigation and a second
/mewhile the first is in flight all leaveroute.fetch(), the body read androute.fulfill()working. Playwright closing the test’s context when the test ends is what clears the fetched response (Response has been disposed) or rejects a fetch still in flight (route.fetch: Test ended.). - Fix: the callback, now the exported
permissionsRouteHandler, ignores only those gone-test errors and rethrows any other. - Regression test:
e2e/browser/auth/mock-permissions.spec.ts. Two of its rows close a context of their own with the reasonTest ended.and run the callback against it. Both fail with the try/catch removed, with exactly the two errors above, and pass with it. - Repeat runs:
--repeat-each=20overlanding.spec.ts,sidebar-collapse.spec.ts,nav-permissions.spec.tsand the new spec, with 4 workers, passed 249 of 249 in Chromium. A temporary Desktop Safari project gave the same result, 249 of 249, in WebKit. - Checks on
4a723f4:pnpm typecheck,pnpm check,pnpm testandpnpm e2e:gate --push(418 passed, 5 skipped) each exited 0. - Not changed:
pinConnectionReadinsync-indicator.spec.tshas the same shape, but itsafterEachalready runspage.unrouteAll({ behavior: 'ignoreErrors' }), so it was left as it was.
Lesson: L-119. A route callback that awaits the network can outlive its test, so it ignores the gone-test errors and rethrows the rest. BUG-057, BUG-063 and BUG-065 are its three instances.
BUG-056 — When “Try again” finds the service back but no session, keyboard focus drops to the page body
Affects: F5-a · Found in: the delta /forge-review-pr of forge-et-fe#101’s fix commits (finding 3), 2026-10-01 · Severity: low · Phase: 1
Repro: with the backend down, open /login, so the service-unavailable notice shows. Start the backend, then press “Try again” from the keyboard.
Expected: the notice gives way to the plain sign-in, and focus moves to “Sign in with Microsoft”.
Actual: the notice and the focused “Try again” button unmount together, so focus falls to <body> and a keyboard or screen-reader user starts again from the top of the page (src/features/auth/LoginRoute.tsx, the serviceUnavailable block, added by forge-et-fe#101).
Fix: when serviceUnavailable goes false while the retry button had focus, move focus to the Microsoft button. The shared Button forwards no ref today, so the fix needs one there first (or a ref on a wrapper). Not in forge-et-fe#101: the reviewer placed it outside the bounded delta pass.
Regression test: a unit test that a keyboard retry which settles unauthenticated leaves focus on “Sign in with Microsoft”.
Resolved: 2026-10-05 — forge-et-fe#118 (ce0ce9d; review fixes 993029f and 0951224; main merged in 6ffa9dd; the press that focused nothing a1c2b20; comments 6f042fc). useServiceRetry notes, at the press, whether “Try again” held focus or nothing did, and counts the checks that have settled; a layout effect then moves focus to “Sign in with Microsoft” only if the press qualified, “Try again” has left the page and nothing holds focus now. A failed check, or focus the user moved elsewhere while the check ran, is left alone. “Nothing did” covers Safari’s click and tap and Firefox’s click on macOS, which do not focus a button, and assistive-technology activations that move no focus; the review first deferred that case as a separate bug, and the author folded it into this PR instead (2026-10-05). The shared Button takes an optional ref (a plain prop in React 19). Fail-before: the regression test failed on the old code with focus on <body>, as do the pointer-press assertion added in review and the press-without-focus row on 6ffa9dd’s code; with the “has left the page” guard removed, the row for a press without focus whose check fails fails. Other rows pin that moved focus is not taken back and that a failed keyboard retry keeps focus on “Try again” (asserting the check ran). Manual round on 2026-10-05 (author, Chrome on Windows, the real backend stopped and restarted, on 0951224): the notice, a failed retry, the keyboard and mouse hand-offs, a signed-in retry and the other sign-in buttons all passed. A re-run on 692af5b (author, Chrome on Windows, /me blocked in DevTools) passed as well: a failed retry by keyboard and mouse, both hand-offs, focus moved during the check, a signed-in retry, the other sign-in buttons, tablet width, and a console click() that activates “Try again” without focusing it (as Safari does) with the service back and still down. Safari was not checked by hand (no Mac on the team’s Windows machines); the press without focus is pinned by unit tests that activate the button without focusing it.
Lesson: L-133 (new in this PR) — the same cause as BUG-026, a focused control removed by the state change its own press caused; frontend-reviewer dimension 18 checks for it.
BUG-055 — A menu whose only item is disabled opens with nothing focused, so the item and its reason cannot be reached by keyboard
Affects: F5-a · Found in: the delta /forge-review-pr of forge-et-fe#101’s fix commits (finding 2), 2026-10-01; the cause predates that PR · Severity: low · Phase: 1
Repro: press “Sign out” while the logout request is slow, then reopen the user menu from the keyboard while the sign-out is in progress. The menu holds one item, “Sign-out is in progress”, and it is disabled.
Expected: the disabled item takes focus, so its reason is read, as the component’s own doc says (“stays focusable so its reason can be read”, src/ui/components/DropdownMenu/DropdownMenu.tsx:12).
Actual: the menu opens with nothing focused (focusItem(ids[0]) gets undefined), and the arrow keys skip disabled items (DropdownMenu.tsx:97, 143, 352), so the item cannot be reached. forge-et-fe#101 first made a menu with only disabled items possible; the exposure lasts one network round trip.
Fix: include disabled items in the menu’s roving focus (the ARIA Authoring Practices allow it), or at least focus the first item when none is enabled.
Regression test: a DropdownMenu unit test that a menu with only disabled items opens with the first item focused and that the arrow keys reach it.
Resolved: 2026-10-02 — fixed in forge-et-fe#117 by the first option (Nanndish’s choice): every item, disabled ones included, is in DropdownMenu’s roving focus, and the existing click guard keeps a disabled item inert to Enter and Space. In a mixed menu the arrow keys now stop on a disabled item rather than skipping it, so the old unit test that asserted the skipping was replaced. With DropdownMenu.tsx put back to origin/main, four new unit tests fail (an all-disabled menu opens with its first item focused and the arrows reach every item; the arrows stop on a disabled item in a mixed menu; Enter and Space on a disabled item select nothing), and so does a new browser test in e2e/browser/shell/current-user-menu.spec.ts, which holds the logout request, reopens the user menu by keyboard and expects “Sign out. Sign-out is in progress” to be focused. With the fix they all pass.
Lesson: none — one-off: DropdownMenu is the only component whose items carry a disabled reason (grep -rln disabledReason src in forge-et-fe lists only it and CurrentUserMenu.tsx, 2026-10-02), and no other Lesson: line has this cause. The design-system §9 rule that a disabled control stays focusable already covers it; the menu’s own focus list had not followed that rule.
BUG-054 — The page a user was trying to reach is lost across the real Microsoft sign-in redirect
Affects: F5-a · Found in: the post-merge /forge-review-pr of forge-et-fe#21 (F5-a wave 1, merged 2026-09-18 with no review), run 2026-09-30 · Severity: medium · Phase: 1
Repro: signed out (or with the session expired), open a protected deep link such as /email/mailbox; on the login page, sign in with Microsoft through the real Entra redirect (laptop mode, F5-i, is enough — no cloud environment is needed).
Expected: after sign-in the user lands on the page they asked for, and after an expiry the “your session expired” notice shows and the page is restored (F5-a FR-3 / AC-5, FR-6 / AC-8).
Actual: the user lands on the default screen. The return route lives only in React Router location state (src/features/auth/LoginRoute.tsx:19-30), and startMicrosoftSignIn.ts:22-29 stores nothing before location.assign sends the browser to Microsoft, so the full-page redirect drops it. The stub sign-in never leaves the SPA, which is why FR-3, AC-5 and AC-8 pass there and nowhere else. The cloud-gated register row (.forge/cloud-gated-validation.md §B F5-a, “from-route preservation across the real full-page redirect”) treated this as a cloud-only item, but the missing carrier is app code.
Fix (lead decision, Tharinda, 2026-09-30): before the redirect, store the return path in sessionStorage — only a validated same-origin path (starts with a single /, not //, and is not /login), keeping its search and hash. After /me succeeds, read it once, remove it, and navigate there; an invalid or missing value falls back to the default screen. In progress on forge-et-fe fix/BUG-052-sign-in-post-merge-fixes, shared with BUG-052 and BUG-053.
Regression test: a unit test that starting the sign-in stores the validated path and that the post-/me routing restores it once and then clears it; that an off-origin or // value is ignored; and a laptop-mode sign-in from a deep link that lands back on it. The real Entra round trip on dev stays the register’s cloud-gated check.
Resolved: 2026-10-01 — fixed in forge-et-fe#101 (src/lib/session/returnPath.ts, startMicrosoftSignIn.ts, AuthenticatedGate.tsx). With startMicrosoftSignIn.ts and AuthenticatedGate.tsx put back to origin/main, four tests fail: the LoginRoute test that the route is stored before leaving for Microsoft, and the three AuthenticatedGate BUG-054 tests (restored once and cleared, followed only once, an unsafe value ignored). With the fix they pass, as do the returnPath accept and reject tables. The laptop-mode deep-link sign-in is run after the PR merges (cloud-gated-validation.md §B F5-a).
Lesson: L-059 — forge-et-fe#21 was a wave PR merged with no /forge-review-pr, so nothing checked the merge for it; the review gate design (.forge/design/harness/forge-review-gate.md, forge-et-harness#328) is the systemic fix, which makes the review record a required check on app-repo PRs.
BUG-053 — A failed login-page chunk or an error in the sign-in gate shows React Router’s raw error page with no way out
Affects: F5-a · Found in: the post-merge /forge-review-pr of forge-et-fe#21 (F5-a wave 1, merged 2026-09-18 with no review), run 2026-09-30 · Severity: medium · Phase: 1
Repro: keep a CRM tab open across a frontend deploy, then sign out or let the session expire, so the SPA loads the lazy LoginRoute chunk, whose old file name no longer exists. Or make the pathless AuthenticatedGate root throw while rendering.
Expected: a plain error screen that says the page could not load and offers a “Reload the page” action.
Actual: React Router’s built-in “Unexpected Application Error!” page, with a stack trace in development and no control to recover. Neither the /login route (src/app/router.tsx:30-36) nor the pathless AuthenticatedGate root (src/app/router.tsx:63-74) has an errorElement, so the error reaches the router’s default.
Fix (lead decision, Tharinda, 2026-09-30): give both routes an errorElement that shows an error screen with a “Reload the page” button (a full reload fetches the current bundle). In progress on forge-et-fe fix/BUG-052-sign-in-post-merge-fixes, shared with BUG-052 and BUG-054.
Regression test: a router test where the login route’s lazy import rejects and one where the gate throws: each renders the error screen with the reload action, not the default error page; and a browser check that the screen passes the gate’s size and axe checks.
Resolved: 2026-10-01 — fixed in forge-et-fe#101 (RouteErrorElement with a reload recovery on /login and the AuthenticatedGate root, src/app/router.tsx). With router.tsx and RouteErrorElement.tsx put back to origin/main, both routerErrorElements.test.tsx cases fail (a failed sign-in chunk, a throw in the gate); with the fix they pass. The browser gate measures the screen as the login-load-error page state.
Lesson: L-059 — forge-et-fe#21 was a wave PR merged with no /forge-review-pr, so nothing checked the merge for it; the review gate design (.forge/design/harness/forge-review-gate.md, forge-et-harness#328) is the systemic fix, which makes the review record a required check on app-repo PRs.
BUG-052 — A failed sign-out shows the user as signed out while the server session stays live
Affects: F5-a · Found in: the post-merge /forge-review-pr of forge-et-fe#21 (F5-a wave 1, merged 2026-09-18 with no review), run 2026-09-30 · Severity: high · Phase: 1
Repro: signed in, make POST /api/v1/auth/logout fail (the backend unreachable, a 5xx, or a CSRF refusal), then choose Sign out from the user menu. Reload the page.
Expected: sign-out either completes, or the user stays signed in and is told it did not (F5-a FR-5 / AC-7 clear the session store on 204, not on a failure).
Actual: the login page shows as if sign-out had worked, but the server session is still live. src/lib/session/SessionProvider.tsx:78-90 clears the local session state in a finally block whatever the logout request answered, and src/app/CurrentUserMenu.tsx:66-69 calls void signOut(), so the rejection is unhandled. The SPA cannot clear the SESSION cookie itself, because it is host-only on api.* (AD #33). On a shared PC, the next person to reload the page is signed in as the previous user, with no notice.
Fix (lead decision, Tharinda, 2026-09-30): clear the local session only when the logout request succeeds. On a failure, keep the user signed in and show “Sign-out didn’t complete. Try again.” A 401 from the logout request means the server session has already ended, so it counts as signed out. In progress on forge-et-fe fix/BUG-052-sign-in-post-merge-fixes, shared with BUG-053 and BUG-054.
Regression test: a SessionProvider test where the logout request fails keeps the session and surfaces the message; a 401 clears it; a 204 clears it; and a CurrentUserMenu test that the failure shows the message and leaves no unhandled rejection. Each fails before the fix.
Resolved: 2026-10-01 — fixed in forge-et-fe#101 (SessionProvider.signOut rejects on any failure but a 401 and changes nothing; CurrentUserMenu shows the message in an always-mounted polite live region and holds the item while a sign-out runs). With SessionProvider.tsx and CurrentUserMenu.tsx put back to origin/main, eight tests fail: the three SessionProvider tests (a 5xx, a network failure and a 403 keep the session) and the five CurrentUserMenu BUG-052 tests; with the fix they pass. The browser gate measures the message as the sign-out-failed page state.
Lesson: L-059 — forge-et-fe#21 was a wave PR merged with no /forge-review-pr, so nothing checked the merge for it; the review gate design (.forge/design/harness/forge-review-gate.md, forge-et-harness#328) is the systemic fix, which makes the review record a required check on app-repo PRs.
BUG-051 — Adding a role’s all key and removing its last view or manage key at the same time can leave the all key on its own
Affects: F5-l · Found in: WI-1.1’s pre-PR review of forge-et-be#125 (F5-l wave 1) · Severity: low · Phase: 1
Repro: a role holds lead.view and no lead.manage. Two requests reach RoleService at the same time: one calls addGrant(role, "lead.all"), the other removeGrant(role, "lead.view"). Each runs in its own transaction.
Expected: one of the two is refused with 422 ALL_KEY_WITHOUT_VIEW_OR_MANAGE (spec FR-5, FR-6), so the role never holds lead.all without lead.view or lead.manage.
Actual: each method reads the role’s stored grants, finds the rule satisfied, and writes. addGrant sees lead.view still present; removeGrant sees no lead.all yet. Nothing locks the role between the read and the write, so both commit and the role ends with lead.all alone. The grant-consistency check (FR-8) then reports it as a failure, and GrantConsistencyStartupCheck refuses to start the application on the next boot. No endpoint calls these methods yet — F5-d’s role editor is the first — so it cannot happen today; it becomes reachable when F5-d ships. A second variant (found by the backend-reviewer on forge-et-be#126): a role holding lead.view, lead.manage and lead.all, with removeGrant(role, "lead.view") and removeGrant(role, "lead.manage") at the same time — each sees the other key still present and passes, and no addGrant is involved. With more than one ECS task, GrantConsistencyStartupCheck then refuses every new task start (deploy, scale-out, replacement) until the row is removed by hand.
Fix: serialise grant writes per role. addGrant and removeGrant lock the role’s row (SELECT … FOR UPDATE, through a pessimistic-lock repository method on RoleRepository) before reading the role’s grants, so a second writer on the same role waits and then sees the first writer’s result. Writes on different roles do not block each other. The last-admin check in removeGrant runs under the same lock. To be fixed before F5-d exposes grant writes over HTTP; a fix/BUG-051-… branch on main, or inside F5-d if it is built first.
Regression test: a RoleServiceIT case with two threads and a latch that start addGrant(role, "lead.all") and removeGrant(role, "lead.view") together on a role holding only lead.view; exactly one succeeds, the other is refused with ALL_KEY_WITHOUT_VIEW_OR_MANAGE, and afterwards the grant-consistency check reports nothing. A second case does the same for two concurrent removals (lead.view and lead.manage on a role holding both and lead.all). Both fail without the lock.
Resolved: 2026-10-04 — fixed in forge-et-be#149 (Nanndish’s PR, brought onto main after forge-et-be#167 by the lead’s decision of 2026-10-04; it was to ship inside F5-d WI-1.1, lead 2026-10-03, but was committed after #167 merged). RoleRepository.findWithLockById (PESSIMISTIC_WRITE); addGrant, removeGrant, delete and F5-d’s save read the role only through that row lock before reading its grants; lock order: the last-admin advisory lock first, then the row. RoleServiceIT carries both variants (two threads, a latch, a held SHARE lock on role_grant released only when both sessions are blocked): exactly one call succeeds, the other is refused with ALL_KEY_WITHOUT_VIEW_OR_MANAGE, the grant-consistency check is clean. Variant 1 failed without the fix ([null, null] — both succeeded); variant 2 was already serialised by removeGrant’s advisory lock and fails only with that removed too, so it guards a future narrowing of that lock.
Lesson: L-065 — another check-then-write race on one aggregate; the fix locks every writer of the role (addGrant, removeGrant, save, delete), not only the pair the bug named.
Follow-ups (Nits from the backend-reviewer pass on #149, left as is because each would cost another full local build record; take them with the next change to RoleService): (1) the Javadoc in RoleRepository and RoleService says SELECT … FOR UPDATE, but Hibernate 7.4.5 emits FOR NO KEY UPDATE on PostgreSQL — it does not block a foreign-key reference; the last-admin lock covers user moves; (2) requireLocked’s Javadoc (“never a copy read before the lock”) is false for StubGrantService, which loads the role by key in the same transaction first — resolve its id with a projection, or narrow the Javadoc; (3) add a LastAdminRaceIT case: a concurrent addGrant waiting on a role that delete then removes answers 404.
BUG-050 — The ingest service’s excluded-folder cache is never cleared, so a reconnected mailbox can keep stale folder ids
Affects: F20-b · Found in: the /forge-review-pr pass on forge-et-be#120 (BUG-048) · Severity: low · Phase: 1
Repro: a mailbox is marked RECONNECT_REQUIRED / MAILBOX_NOT_ENABLED (BUG-048); IT then provisions a new Exchange mailbox for the account, and the owner reconnects without the backend restarting.
Expected: the back-fill and ingest exclude the new mailbox’s Drafts, Junk and Deleted Items (F20-b FR-3).
Actual: MailboxIngestService.folderCache holds the excluded well-known folder ids keyed by connectionId:address and is never evicted. The reconnect keeps the connection id and the address, so the cached ids are the old mailbox’s; a newly provisioned mailbox has new folder ids, so nothing matches and the exclusion fails silently until the process restarts. A soft-deleted mailbox that is restored keeps its ids, so the risk is limited to a newly provisioned one. The code predates BUG-048; BUG-048 made this path reachable.
Fix (lead decision, Tharinda, 2026-09-30, revised twice the same day after review — both halves): (1) clear the connection’s cached folders at the start of every back-fill run, on whichever ECS task picks up the job, before the run resolves the folders. A back-fill is several queued jobs, not one: MailboxBackfillService stops after MAX_PAGES_PER_RUN (20) pages or half the visibility timeout and queues a continuation (enqueueBackfillContinuation), which any task can pick up with a cursor already set. So clearing only when the cursor is reset would miss the continuations, and clearing in MailboxConnectionService.connectOrUpdate would reach only the task that served the HTTP reconnect, which merely enqueues the first run (enqueueBackfillAfterCommit). The cost is three folder look-ups per run of up to 20 pages. (2) Give every entry a time limit of one hour, so live notifications handled on other tasks refresh on their own. Store the time each entry was resolved alongside the folder ids in the existing ConcurrentHashMap, and treat an older entry as missing. Two traps: folders() stores with putIfAbsent, which never replaces an expired entry, so use put or compute; and the service has no clock, so inject BusinessClock (as AccountService does) and test the limit with a fixed clock. The limit is a constant in the service, not a configuration key, and no cache library is added (Caffeine is on the classpath only as an annotation-processor dependency, so using it would be a new runtime dependency — ASK FIRST). Considered and rejected: clearing in connectOrUpdate (the first version of this decision — it reaches only the task that served the reconnect), clearing only when the back-fill starts from a reset cursor (the second version — continuations on other tasks would still use stale ids for up to an hour), the time limit alone (up to an hour of wrongly ingested Drafts, Junk and Deleted Items), and a cache key carrying a value that changes on every reconnect (the most complete, but it may need a new column).
Regression test: a back-fill run — the first and a continuation — after the folder ids change stores the new ids, and a message in the new Drafts folder is excluded; and, with a fixed clock, an entry older than the limit is resolved again.
Resolved: 2026-10-02 — fixed in forge-et-be#138, both halves as decided:
- Clearing.
MailboxBackfillService.runDeclaredcallsMailboxIngestService.forgetFoldersfirst, in every back-fill job. - Time limit. Entries carry their resolution time and expire after
FOLDER_CACHE_LIFETIME(one hour, read through an injectedBusinessClock), and are stored withput. - Beyond the decision (1). The ids resolved before the claim are passed into it, so the check inside the claim’s transaction can never call Graph.
- Beyond the decision (2) (Nanndish, 2026-10-02, from the
/forge-review-prpass). The call sits before the “already complete” check, not after it, so it also runs in the job that stops at once. A re-consent of anACTIVEconnection queues that job, and a mailbox IT replaced would otherwise keep its old ids on that task for up to the hour. The decision said “before the run resolves the folders”; this clears in a job that resolves none. - Fails before the fix. With the fix set aside, the new
MailboxBackfillITcases stored the draft in the new Drafts folder (rowCount == 1): a reconnect after the folder ids change, a continuation run, and (on the first placement) the re-consent of anACTIVEconnection.MailboxIngestFolderCacheSpec, in its first two-case form, failed 2 of 2 with the expiry check removed and 1 of 2 withputIfAbsentrestored. The third case (the key prefix), added at review, fails with the:dropped from the prefix (1 of 3). - Passes with the fix.
./gradlew buildpassed on the PR head46dccd5, withmainmerged in: 4,006 tests, 0 failures, 0 errors, 2 skipped (the local-build record on the PR). That last commit is test-only, from the second/forge-review-prpass.MailboxIngestIThad asserted exactly one folder resolution during the connect-time back-fill. Now that every job resolves afresh, a back-fill that continues in a second job would break that, so the test now asserts that later ingests add none. - Manual check. Nanndish repeated the reconnect by hand on the local stub twice: before the review fixes and again on the reviewed code. Each time only the new mailbox’s inbox message was stored.
- Still owed. Behaviour across two real ECS tasks needs dev: register §B F20-b, “The excluded-folder cache across ECS tasks”.
Lesson: L-099 — every cache names the event that clears it, and an in-memory cache is per ECS task, so the clearing has to run where the work runs or the entry needs a time limit. BUG-050 is that lesson’s second instance (after BUG-016), and is already in its source list.
BUG-049 — The mailbox consent shows Microsoft’s account picker instead of using the signed-in account
Filed as BUG-045 and renumbered BUG-049: main’s BUG-043 to BUG-046 merged first (.claude/rules/tracker.md, L-057). Code comments and PR titles use the new number; the branch name keeps the old one.
Affects: F20-a · Found in: the laptop-mode connect test with CRM05@empire.lk · Severity: low · Phase: 1
Repro: in a browser signed in to more than one Microsoft account (for example CRM01 and CRM05), sign in to the CRM as one of them, open the Email screen and click Connect.
Expected: Microsoft uses the account the rep is signed in to the CRM with.
Actual: Microsoft shows its account picker, because the mailbox authorization request names no account. Picking a different account is refused at completion (F20-a FR-3, MailboxAuthorizationCompletionHandler), so no data is at risk, but the rep sees only the generic “connect failed” result and does not know why.
Fix: the authorization request for the mailbox registration carries login_hint=<signed-in user's email>, so Microsoft pre-selects that account and normally skips the picker. auth stays feature-neutral: OAuth2RegistrationContribution gains an opt-in hintSignedInUser, the mailbox registration sets it, and a SignedInUserLoginHint customizer adds the hint through the request customizer both redirect filters share (BUG-040’s). The completion’s same-identity check stays as it is (lead approval, 2026-09-29).
Regression test: MailboxAuthorizationAccessIT asserts the hint on the mailbox start (fails without the fix: no login_hint parameter) and its absence on the sign-in start; SignedInUserLoginHintSpec covers a signed-in user, nobody signed in, and a registration that did not ask.
Resolved: 2026-09-29 — fixed in forge-et-be#121. Without the fix MailboxAuthorizationAccessIT fails with no login_hint parameter; with it the access IT, SignedInUserLoginHintSpec and LaptopModeIT (the laptop-only chain: the consent carries the hint, the sign-in keeps its chooser and no hint) pass. An assumption, recorded at the review’s request: the hint is the CRM user’s email (app_user.email, matched at sign-in against the ID token’s email claim), while Entra matches login_hint against the sign-in name (the UPN). They are the same for Empire’s accounts, CRM01–CRM05 included; where they differ, Entra shows its username page pre-filled with a name it cannot use, which is recoverable and changes nothing about security, since the completion’s subject check still decides. If Empire’s UPNs turn out to differ from their mail addresses, hint preferred_username for an Entra principal instead, read through Principals. To confirm in the next live laptop-mode connect.
BUG-048 — A mailbox Graph cannot read shows as connected, and its sync fails silently
Filed as BUG-044 and renumbered BUG-048 (L-057); the branch name keeps the old number.
Affects: F20-a, F20-b · Found in: the laptop-mode connect test with CRM05@empire.lk, which then had no Outlook licence (it has a working mailbox from 2026-10-04 — environments-and-identity.md §5) · Severity: medium · Phase: 1
Repro: in laptop mode, sign in as CRM05@empire.lk and connect its mailbox from the Email screen.
Expected: the user is told the account has no mailbox the CRM can read, and what to do about it; no job retries a failure that cannot succeed.
Actual: the connect succeeds (mailbox.connected, connection ACTIVE, no status reason) and the first token refresh works. The connect-time back-fill’s first Graph call (GraphMailboxProvider.pageMessages) answers 404 MailboxNotEnabledForRESTAPI (“The mailbox is either inactive, soft-deleted, or is hosted on-premise”). Only a 404 on a single-message fetch is classified, so this one propagates, the job rolls back, the queue redelivers it three times (maxReceiveCount=3) and it lands in et-mailbox-ingest-dlq. The connection stays ACTIVE, no mail arrives, and nothing on the Email screen says why. A mailbox whose licence is removed after it was connected fails the same way on its next sync.
Fix: treat MailboxNotEnabledForRESTAPI as permanent. The Graph adapter raises a typed exception for that error code; the back-fill and the delta sync catch it, move the connection to RECONNECT_REQUIRED with a new reason MAILBOX_NOT_ENABLED, audit the change, and complete the job instead of rethrowing, so nothing is retried or dead-lettered. The Email screen shows the reason’s copy, saying the Microsoft account has no mailbox and to ask IT to enable Outlook, then connect again. Marking the connection, rather than refusing the connect, also covers a licence removed after connecting (delivery-team choice, 2026-09-29, after the lead asked for both bugs fixed). A new reason is a change to F20-a’s closed StatusReason vocabulary (its CHECK constraint needs a migration, and the Wave-1 contract lists the reasons), so the spec and contract get revision entries.
Regression test: a Graph adapter test that the 404 with that code raises the typed exception; a back-fill test that the connection ends RECONNECT_REQUIRED / MAILBOX_NOT_ENABLED and the job completes; a frontend test that the reason shows its copy; and the laptop-mode CRM05 connect repeated end to end.
Resolved: 2026-09-29 — fixed in forge-et-be#120 (backend, V30) and forge-et-fe#88 (the Email screen’s sentence). Backend: GraphMailboxProviderSpec raises the typed exception on each of the four mailbox reads and still propagates any other 404; MailboxIngestJobListenerSpec and MailboxDeltaPollJobHandlerSpec mark the connection instead of retrying; MailboxConnectionAuditIT records MAILBOX_NOT_ENABLED, one mailbox.reconnect_required as SYSTEM:mailbox-ingest, and changes nothing on a second call; MailboxStubEndpointsIT proves the Postgres CHECK accepts the value after V30. Frontend: the browser gate passed (349 passed, 5 skipped), and at 1280 px and the 768 px floor the reconnect-required screen shows the new sentence with no horizontal overflow. The live CRM05 connect is repeated once both PRs are on main.
BUG-047 — After connecting a mailbox the browser lands on the API’s raw /api/v1/me?continue
Filed as BUG-043 and renumbered BUG-047 (L-057); the branch name keeps the old number.
Affects: F20-a · Found in: the laptop-mode connect test with CRM05@empire.lk · Severity: high · Phase: 1
Repro: in a fresh browser session, open the CRM, sign in with Microsoft, open the Email screen and connect the mailbox.
Expected: after the Microsoft consent the browser returns to the SPA at /email/mailbox?connect=success (F20-a).
Actual: it lands on http://localhost:8080/api/v1/me?continue, the API’s raw JSON. The connection itself is saved. Every user’s first mailbox connect after signing in does this, in every environment. The cause is Spring Security’s default request cache: the login page’s unauthenticated GET /api/v1/me is refused with 401, and because the SPA’s API client sends no Accept header (the browser sends */*), the default matcher, which skips only requests asking for JSON, saves it as SPRING_SECURITY_SAVED_REQUEST (reproduced with curl -H 'Accept: */*'; with Accept: application/json nothing is saved). The saved request survives sign-in, because the session’s attributes move to the rotated session and the custom sign-in success handler never reads it. OAuth2AuthorizationCodeGrantFilter then prefers a saved request over the registration’s redirect URI (Spring Security 7.1.1, lines 267–274), so the mailbox consent’s completion replays it with ?continue.
Fix: configure http.requestCache(cache -> cache.requestCache(new NullRequestCache())) on the one filter chain. The SPA owns navigation, and the API never needs to replay a request after authentication; the chain’s shared RequestCache is what the OAuth2 client filters read (OAuth2ClientConfigurer, lines 260–289), so the one setting covers the mailbox leg. A side effect: an anonymous request refused with 401 no longer creates a session, since saving the request was what created one.
Regression test: an anonymous GET /api/v1/me with Accept: */* leaves no SPRING_SECURITY_SAVED_REQUEST in any session; and a mailbox consent completed in a session that first made that refused request redirects to /email/mailbox?connect=success, not to /api/v1/me?continue.
Resolved: 2026-09-29 — fixed in forge-et-be#119. Without the fix, MailboxConnectFlowIT’s round trip after a refused /me (signed in through the real login leg against the faked token endpoint) landed on http://localhost:<port>/api/v1/me?continue, and the refused request created a session; with it both pass, and 802 sign-in, session, CSRF and mailbox tests pass. The test harness had hidden the bug: the round trip always requested the stripped redirect URI instead of following the callback’s Location; it and MailboxBackfillIT’s now follow it.
BUG-046 — An empty skip line fails the description check with the generic “not in an accepted format” message
Affects: F40 (scripts/browser-gate/ci-check.mjs in forge-et-fe and forge-et-pwa) · Found in: F40-WI-2.2 live proof, case C12 (2026-09-29) · Severity: low · Phase: 1
Repro: on a ready PR that changes a gate input, set the description’s gate line to Browser gate: skipped — (nothing after the dash). forge-et-fe run 36574489098, forge-et-pwa run 36576833043.
Expected: the Fast lane fails (it does), naming the rule: a skip line needs a reason.
Actual: it fails with “the Browser gate: line is not filled in, or not in an accepted format (“Browser gate: skipped —”)”, the malformed-line text, so the author is not told that only the reason is missing.
Fix: in ci-check.mjs, when the line starts Browser gate: skipped and has no reason, say “a skip line needs a reason after the dash: Browser gate: skipped — browserGateCiCheck.test.ts.
Regression test: a ci-check unit test that an empty skip line fails with the skip-specific message.
Resolved: 2026-10-01 — fixed in forge-et-fe#103 and forge-et-pwa#45 (the two ci-check.mjs and browserGateCiCheck.test.ts copies byte-identical). The new tests failed 7 of 51 before the fix and pass 51 of 51 after, in both repos; full gates: forge-et-fe 371 passed at c366681, forge-et-pwa 124 passed at 02cc683. Review follow-up left open for a decision: a reason made only of punctuation (Browser gate: skipped — .) is still accepted, and refusing it would change the rule.
Follow-up decided (lead, 2026-10-02): refuse it. The post-merge review of forge-et-fe#103 and forge-et-pwa#45 (frontend-reviewer) found that a doubled dash (Browser gate: skipped — —), the likeliest typo, passed as a skip, so a PR could merge with a skip nobody explained; the rule (“a stated reason, visible on the PR”) already excludes punctuation. A skip reason must now contain a letter or digit, in both frontends’ rules.mjs and ci-check.mjs and in the harness hook’s SKIP_LINE_RE, changed together (forge-et-fe and forge-et-pwa fix/post-merge-review-nits; forge-et-harness chore/post-merge-audit-2026-10-02).
Lesson: none — one-off: one message for one malformed-line case, now covered by a test.
BUG-045 — Flaky tests: WebKit size-matrix rows on the mailbox and user-menu states time out once under load (Windows)
Affects: F40 (forge-et-fe/e2e/browser/size-matrix/size-matrix.spec.ts, WebKit projects at 768 × 1024 and 1024 × 768) · Found in: F40-WI-2.2 live proof (2026-09-29) · Severity: low · Phase: 1
Repro: on Windows 11, run pnpm e2e:gate in a forge-et-fe checkout while the machine is also busy (on 2026-09-29 the orchestrator was polling GitHub and a second gate had just run). At feature/F40-WI-2.2-throwaway-draft (96a8015, identical to origin/main apart from one comment line) the run failed 8 rows, all WebKit: mailbox-connect, mailbox-connect-failed and user-menu-open at 768 × 1024; email-not-connected, email-list, email-sync-lapsed, mailbox-connect-failed and mailbox-disconnect-dialog at 1024 × 768. Each timed out after 5 s in waitUntilReady (size-matrix.spec.ts:31) waiting for the state’s ready test id (for example mailbox-not-connected). The same commit passed the whole gate on the next run (348 passed, 281 s), and the neighbouring ready branch passed at 380 s.
Expected: the gate’s result does not depend on how busy the machine is.
Actual: once in the day’s eight full frontend gate runs, the WebKit rows that pin mailbox reads (arrange.ts) or open the user menu missed the 5-second ready wait. Not reproduced on demand; likely the pinned route or the menu rendering later than 5 s in WebKit under load.
Fix (lead decision, Tharinda, 2026-09-30 — a half-day timebox): spend at most half a working day re-running the WebKit size projects with --repeat-each on a loaded Windows 11 machine, where the failure was seen. If the timeout comes back and the ready wait is the cause, give waitUntilReady a longer timeout for the WebKit projects, or wait on the pinned read’s response before the ready check (test-only), and close as fixed. If it does not come back within the timebox, close as wontfix (not reproducible), with the commands run and their results recorded in this section; a recurrence is filed as a new bug that links here. Either way it stops blocking the Phase 1 seal.
Regression test: the WebKit size projects green across repeated full-gate runs on a loaded Windows machine.
Resolved: 2026-10-01 — fixed in forge-et-fe#104: waitUntilReady gives the WebKit size projects 15 s, choosing by the browserName fixture. Before: --repeat-each=5 on both WebKit tablet projects under load failed 4 of 213, all at the 5000 ms ready wait. The first version read testInfo.project.use.browserName, which is unset for the size projects, so the wait never changed; the frontend-reviewer found it, and the green runs made with it showed only a quieter machine. The corrected code was proved by a temporary forced failure (the not-found state’s ready heading pointed at a missing name, then reverted): size-768x1024-webkit reported Timeout: 15000ms, size-1440x900-chromium 5000ms. Full pnpm e2e:gate at 4f4fc02: 371 passed, 5 skipped. Only the ready wait is widened; the Chromium recurrences below stay open under this entry.
Lesson: L-107 — a green re-run does not prove a fix for a load-dependent failure; force the failure to see the change.
Also seen (2026-09-30, the BUG-035 gate runs in forge-et-fe#90, not yet investigated): the same 5-second waits failed twice on a loaded Windows machine, and the same commits passed when the gate ran alone.
- Commit
57473f9, the first run after the local backend was restarted: one Chromium row,connect-a11y.spec.ts:150“AC-28: connect failed is reachable and activatable from the keyboard”. After Enter,mailbox-connectedwas not visible within 5 s (connect-a11y.spec.ts:386). The row then passed 5 of 5 with--repeat-each=5 --workers=1. - Commit
44cf2dd, whileforge-et-pwa’s build and unit tests ran on the same machine: 9 rows. Five were Chromium mailbox rows:connect-screen.spec.ts:381,:413and:495,inbox.spec.ts:269andsync-indicator.spec.ts:487. Four were WebKit size rows:email-not-connected,email-listandemail-sync-lapsedat 768 × 1024, andnot-foundat 1024 × 768. All weretoBeVisible5 s timeouts. The same commit passed the whole gate alone next (349 passed).
So the load that trips the wait also reaches Chromium mailbox rows, not only WebKit, and not-found, which reads no mailbox.
- 2026-10-01, Chromium (not covered by the WebKit-only fix): commit
c366681(BUG-046, a regex rewritten with escapes), while a review agent ran tests on the same machine: 4 Chromium rows. Three were first headings not visible within 5 s (sync-indicator.spec.ts:705“Mailbox”,frame.spec.ts:106andresponsive-reflow.spec.ts:19“Pricing”), andresponsive-reflow.spec.ts:47failed withpage.goto: Page crashed. The dev server logged “Importing a module script failed” twice. The previous commit passed the full gate alone. Earlier the same day, the BUG-058 runs at1bac423failed two Chromium first-heading waits once (navigation-by-role.spec.ts:147,login-page.spec.ts:79) under the same kind of load. If the Chromium side keeps recurring on a quiet machine, it gets its own bug. Filed as BUG-079 on 2026-10-03: the first deep link into a lazy route paying the dev server’s transform.
BUG-044 — The browser-gate hook refuses any command with a case statement that mentions gh pr create/new/ready
Affects: F40 (forge-et-harness/.claude/hooks/guard-browser-gate.mjs, the bash parser) · Found in: F40-WI-2.2 live proof (2026-09-29) · Severity: low · Phase: 1
Repro: in a Claude Code session with the hook registered (harness main after forge-et-harness#285), run the Bash command case x in x) echo "the words gh pr ready appear here only as text";; esac.
Expected: allowed — no gh command runs; case … in pattern) … ;; esac is ordinary bash.
Actual: refused with “Blocked by Forge harness: this PR has no passing browser gate. Reason: the command cannot be parsed: a “)” has no matching “(””. The fast path matches the text, and the helper’s parser reads the case pattern’s closing ) as an unmatched parenthesis, so it fails closed. First met as a false refusal of the orchestrator’s own polling loop (a case "$j" in completed*) break;; esac wait loop with gh pr ready inside a quoted prompt). echo "\(a)" with the same text is allowed, so it is the case form, not escaped parentheses.
Fix (lead decision, Tharinda, 2026-09-30 — teach the parser case): teach the bash tokenizer the case … in … esac form (a pattern’s ) closes the pattern, ;; ends an arm), so the hook still refuses whenever it cannot parse a command. Rejected: allowing an unparseable command in which no gh word stands at a command position, because it weakens a guard that blocks when in doubt. Add refusal and allow tests, and, per L-095, get an adversarial review of the parser change and list what it still does not handle.
Regression test: a unit test in guard-browser-gate.test.mjs that the command above is allowed, and that a real gh pr create inside a case arm is still read and refused.
Resolved: 2026-10-01 — fixed in forge-et-harness#342. The parser reads case … in … esac and still refuses what it cannot parse. Two adversarial reviews hardened it: line continuations inside case/esac; a cd or GH_REPO in an arm made uncertain; time -p; coproc refused; a ( after a word refused outside bash’s forms; a stdin-reading shell with no input of its own refused everywhere; arithmetic ((…)), for ((…)), $((…)) and $[…] read as one unit, which also closes three false passes on main ((( x = 1<<2 )) hiding the next line as a here-document, and a gh inside $(( $(…) )) or $((cd x); …)). node --test .claude/hooks/tests/guard-browser-gate.test.mjs: 376 pass; the wrapper suite 84 pass. The head’s tests run against the first-pass hook: 26 fail. False passes the reviews found that are already on main are BUG-059, BUG-061, BUG-062 and BUG-064. Accepted, in the hook README: a shell with --version or --help next to a gh pr create is refused.
Lesson: none — one-off: a grammar gap in one parser, now covered by tests; the reviews’ wider false passes are filed as their own bugs.
BUG-043 — The capture app’s two error-screen buttons are under the 44px target floor
Affects: FS-1 (the capture shell’s error screens shipped with Foundation slice 1, as BUG-031 is filed) · Found in: F40-WI-1.2 plan review (2026-09-28) · Severity: low · Phase: 1
Repro: in forge-et-pwa on origin/main (4dfdbb8), make a route fail to load (so RouteErrorElement.tsx renders), or throw inside the app (so AppErrorBoundary.tsx renders), and measure the button’s box.
Expected: every interactive target in the capture app is at least 44 × 44px (design-system §9; the capture-app CLAUDE.md rule; PRD § Accessibility).
Actual: “Back to the dashboard” (RouteErrorElement.tsx) sets padding but no min-height and comes to about 39.5px at the 13px body size (about 41px once F40’s 14px step-up lands); “Reload the page” (AppErrorBoundary.tsx) has no min-height and no border and comes to about 39px at 14px. The BUG-031 shape. Neither screen is an FR-4 page state, so F40’s shared check does not see them. Measured from the CSS by the WI-1.2 plan author and reviewer, not yet in a browser.
Fix: give both buttons a 44px minimum height from the token. Both are styled inline (style={{ … }}), so add minHeight: 'var(--et-layout-control-h)' to each style object, as NotFoundRoute.tsx does for the 404 page’s button (BUG-031).
Regression test: a browser row that renders each error screen and asserts the button’s height is at least 44px (or F40’s shared check run on both screens).
Id history: filed on the F40 plan branch as BUG-035, renumbered BUG-037 when main took 035 and 036, and BUG-043 when main took 037 and 040 to 042.
Resolved: 2026-10-01 — forge-et-pwa#44. Both buttons take minHeight: 'var(--et-layout-control-h)' (44px), as NotFoundRoute.tsx does since BUG-031; nothing else on either screen changed. Regression tests: Vitest rows in RouteErrorElement.test.tsx and AppErrorBoundary.test.tsx assert the minimum height (3 of 5 failed with the fix undone, 5 of 5 pass with it), and a Playwright row in e2e/browser/shell/a11y.spec.ts blocks the capture route’s chunk so the router’s error screen renders, then measures “Back to the dashboard” at 360 × 780 (41px with the fix undone, at least 44px with it) with an axe scan and the no-horizontal-scroll check. “Reload the page” needs the app itself to throw, so only its unit test covers it. pnpm e2e:gate passed on the head (124 tests). Review (/forge-review-pr on forge-et-pwa#44, frontend-reviewer): pass-with-nits. Applied: the row waits for the error heading (the 404 branch shares the button name), adds the scroll check, and its comment is corrected. Not applied: the shared expectPageClean on this screen — it requires a <main>, which the error screens do not render; filed as BUG-060.
Lesson: L-106 — the browser gate checks only the page states in its list, so screens a user reaches only on a failure (error boundaries, the router’s error element) escape it; BUG-031 and BUG-060 are the same gap.
BUG-042 — The CSRF check parses an anonymous multipart body before authorization
Affects: FS-6 · Found in: F4-a (the /forge-review-pr pass on forge-et-be#112) · Severity: low · Phase: 1
Repro: with the backend running, send a signed-out multipart/form-data POST to any path with no X-XSRF-TOKEN header (for example curl -F file=@<large file> http://localhost:8080/api/v1/anything). The answer is a 403, but only after Tomcat has parsed the whole body.
Expected: a request with no CSRF header is refused without its body being read, so an anonymous caller cannot make the server buffer or store upload parts.
Actual: Spring Security’s CsrfTokenRequestAttributeHandler falls back to request.getParameter("_csrf") when the header is missing, and for a multipart request that makes Tomcat parse every part inside CsrfFilter, before authorization. Under the default file-size-threshold (0 B) each part goes to a temporary file, deleted when the request ends; under a raised threshold it would stay in heap (up to about 27 MB per request), which is why F4-a Decision 14 chose option C (F4-a spec Rev 2). Not verified by a live test yet: the reviewer found it by reading the code.
Fix (lead decision, Tharinda, 2026-09-30 — header-only CSRF token): (1) First run the one-off check: a signed-out multipart POST with no X-XSRF-TOKEN header, watching heap and the temporary directory, and record the result here. If the body is not parsed, stop and bring the result back to the lead before changing anything. (2) If it is confirmed, the CSRF request handler configured in SecurityConfig (today a plain CsrfTokenRequestAttributeHandler) resolves the token from the token’s own header (csrfToken.getHeaderName(), not a hard-coded name) and never falls back to the _csrf request parameter. Implement it as a subclass of CsrfTokenRequestAttributeHandler that overrides only resolveCsrfTokenValue, so the rest of today’s wiring is kept: setCsrfRequestAttributeName(null) (the eager cookie write, SecurityConfig.csrfTokenRequestHandler()) and the same handler set on CsrfAuthenticationStrategy. The SPA always sends the header, and grep -rn _csrf src/main finds no form posting the parameter (re-run the grep before changing the handler). Tests that send the parameter break: Spring’s SecurityMockMvcRequestPostProcessors.csrf() sends the token as _csrf unless .asHeader() is added, and LaptopModeIT.groovy uses .with(csrf()) (lines 226 and 229 as of 2026-09-30). Run git grep -n 'with(csrf())' src/test and change each to csrf().asHeader(). Keep everything else about the CSRF setup unchanged (the cookie name and domain from AD #33 / F5-j, the SPA’s header). Because this is security code, the PR also asks CI for the full suite (gh workflow run "PR validation" --ref <branch> in forge-et-be, .claude/rules/git-conventions.md), in addition to the local build record.
Regression test: it must be able to fail before the fix. MockMvc never parses a multipart body through Tomcat, so a MockMvc version passes on the old code too. Use either (a) a test of the handler itself: a request carrying a _csrf parameter and no header resolves to no token (fails before, passes after); or (b) a signed-out multipart POST with no header against a real embedded server (@SpringBootTest(webEnvironment = RANDOM_PORT)), asserting the 403 arrives without a temporary part file being created. (a) proves the handler no longer reads the parameter, but not that nothing else on the refused path parses the body, so with (a) also re-run the one-off live check from step (1) on the fixed build and record the result here; (b) proves both in one test.
Live check, step (1) (2026-10-02, Nanndish — confirmed, with a second cause): forge-et-be main at 4996d70, local profile, run with ./gradlew bootRun on port 8081 with server.tomcat.basedir set to a scratch directory so the part files could be watched. A signed-out 20 MB multipart/form-data POST to /api/v1/anything, sent at 2 MB/s, with no X-XSRF-TOKEN header: 403 after 10.2 s with all 20 MB uploaded; a part file upload_…tmp grew to about 20 MB in Tomcat’s work directory and was deleted when the request ended; heap used went from 88 MB to 93 MB (the default 0 B threshold sends parts to disk). The same request with a wrong header (X-XSRF-TOKEN: wrong), where the handler never reads _csrf, was also parsed whole (403 after 10.0 s, part file written). Thread dumps taken during each upload show two places that parse the body:
- The cause above:
CsrfFilter→CsrfTokenRequestHandler.resolveCsrfTokenValue→getParameter("_csrf")→Request.parseParts. - The refusal’s error page: a 401 or 403 leaves through
sendError; Tomcat forwards it to/error(StandardHostValve.status→ApplicationDispatcher.forward), andDispatcherServlet.checkMultipart→StandardServletMultipartResolver.resolveMultipart→getPartsparses every multipart request on any dispatch. So with the header-only handler alone, anyX-XSRF-TOKENvalue still made the server store the whole upload.
With spring.servlet.multipart.resolve-lazily=true (parts parsed only when a handler reads one) on the unchanged code, the wrong-header upload was refused in 0.14 s with 256 KB sent and no part file, and the no-header upload was still parsed whole (cause 1). Nanndish decided on 2026-10-02 to make both changes.
A third cause, found by the regression test (2026-10-02): with both changes, a signed-out upload carrying a matching made-up cookie and header (Cookie: XSRF-TOKEN=a, X-XSRF-TOKEN: a — the double-submit check compares the two, so any caller passes it) was still parsed whole before its 401. A thread dump of the test JVM showed Spring’s OAuth2AuthorizationCodeGrantFilter.matchesAuthorizationResponse calling request.getParameterMap() on every request that passes CSRF, to check whether it is a Microsoft reply. Two more OAuth2 filters read parameters on their own paths: the sign-in start (/oauth2/authorization/microsoft, DefaultOAuth2AuthorizationRequestResolver reads action) and the sign-in callback (/login/oauth2/code/microsoft, OAuth2LoginAuthenticationFilter reads code and state); a test row for each also hung on the unguarded build. These filters exist wherever the OAuth2 legs are configured — the test profile, laptop mode and every cloud environment — but not in plain local, which is why the live runs did not show them. Spring Security 7.1.1 gives them no request-matcher setting. The options were (A) ship the two changes and file the OAuth2 filters as a new bug, or (B) fix them here; Nanndish decided on 2026-10-02 to take (B).
The fix as built (forge-et-be branch fix/BUG-042-csrf-header-only): (1) HeaderOnlyCsrfTokenRequestHandler (a CsrfTokenRequestAttributeHandler subclass overriding only resolveCsrfTokenValue, reading csrfToken.getHeaderName()), set as the SecurityConfig.csrfTokenRequestHandler() bean with setCsrfRequestAttributeName(null) kept; LaptopModeIT’s three .with(csrf()) calls changed to csrf().asHeader() (grep -rn _csrf src/main found nothing). (2) spring.servlet.multipart.resolve-lazily: true in application.yml. (3) OAuth2MultipartParameterGuard: two filters, Hide just before OAuth2AuthorizationRequestRedirectFilter and Reveal just after AuthorizationFilter (first built after OAuth2AuthorizationCodeGrantFilter; moved by the review, below). For a multipart request only, Hide passes on a wrapper that answers no parameters while a request attribute is set, and Reveal clears the attribute, so every security filter from the first OAuth2 filter to the authorization decision sees no parameters and everything after them sees them again. Microsoft’s replies are GETs with code and state in the query string, never multipart, so sign-in and the mailbox consent are unchanged. The pair is placed by the filters’ registered order, so it is wired the same in every profile.
Regression tests: HeaderOnlyCsrfTokenRequestHandlerSpec (through the configured bean, a _csrf parameter with no header resolves to no token); OAuth2MultipartParameterGuardSpec (a multipart request has no parameters between the two filters and has them again after Reveal; a form-urlencoded, JSON or untyped request passes through as the same object; the attribute does not outlive the chain); five SecurityChainIT rows that open a raw socket to the embedded server, send the headers of a 10 MB upload with only its first 64 KB, and require the answer within 5 s with no part file in the context’s temporary directory — a server parsing the body waits for the rest and never answers. Results:
| Build | Handler spec | No header → 403 | Wrong header → 403 | Matching pair → 401 | Matching pair, sign-in callback → 302 | Matching pair, sign-in start → 302 |
|---|---|---|---|---|---|---|
| Old code | fails | fails (no answer in 5 s) | fails | fails | — | — |
| Handler change only | passes | fails | fails | fails | — | — |
| Handler + lazy resolution | passes | passes | passes | fails | fails | fails |
| All three changes | passes | passes | passes | passes | passes | passes |
(The two sign-in rows were added after the third cause was found, so they ran only on the last two builds.) On the fixed build without the guard, live (local profile): no header and wrong header were each refused in 0.04 s with 64 KB sent, no part file, heap 87 MB before and after.
Live check on the fixed build, laptop mode (2026-10-02): local,entra-laptop, so both OAuth2 legs are present as in a deployed environment. The partial 10 MB upload (headers plus 64 KB) answered, in 0.03 s or less each and with no part file: no header 403, wrong header 403, matching made-up cookie and header 401, the same to /login/oauth2/code/microsoft 302, the same to /oauth2/authorization/microsoft 302. Manual (Nanndish, the same build, against the Empire Entra tenant): sign-in as CRM01@empire.lk; sign-out; a refused sign-in as CRM02@empire.lk (“We couldn’t sign you in”); the mailbox connect, with the back-fill filling the inbox (it first showed an empty inbox because the copied application-local.yml lacked et.messaging.mailbox-ingest-queue — a local-config gap, not this fix); the five signed-out uploads with curl --limit-rate 2M and a 20 MB file, each answered at once with no part file; and a signed-in upload to the Pipedrive dry run sending only leads, answered 400 VALIDATION_FAILED for the missing parts, so a signed-in caller’s parts are still read. All as expected.
Review (/forge-review-pr on forge-et-be#139, with backend-reviewer): approve-with-fixes, no Blocker. Applied in 98956cd: Reveal moved from after OAuth2AuthorizationCodeGrantFilter to after AuthorizationFilter — SessionManagementFilter, ExceptionTranslationFilter and AuthorizationFilter saw the real parameters, and only BUG-047’s NullRequestCache kept a saved request from copying them; the guard spec now passes through an extra wrapper between the two filters and checks the attribute by its constant; the Javadoc says which filters have path matchers and that the hiding applies on the way down only; a comment pins the test’s announced size below max-request-size, above which the rows would pass on the old code too. After the move, the guard spec, the handler spec, SecurityChainIT, LaptopModeIT, EntraLoginIT, PipedriveDryRunControllerIT (a multipart form field read through the chain) and AttachmentPipelineIT all passed. Left as it is: an over-limit (413) test for a signed-in upload — a first attempt showed Tomcat resets the connection while the client is still sending, on the old code too, so the test cannot pin the 413; the gap is stated in the PR. The reviewer’s out-of-scope finding, FormContentFilter, was confirmed live and filed as BUG-076 (first numbered BUG-073), fixed in the same PR. Delta review (98956cd and d35090a, by --since 5075240; the merge 7ddda83 passed the merge test): pass, no Blocker or Important, each first-pass fix confirmed. Three Nits: a test pinning Reveal after AuthorizationFilter, and a PUT form row with a matching cookie and header, were left as they are (the BUG-042 rows cover the behaviour; Tomcat parses form bodies for POST only); the bounded form-POST read is recorded under BUG-076. Full verdict posted on the PR (issuecomment-5948441560); review record approve on 7ddda83, Forge review: success (issuecomment-5948448653).
Second review (/forge-review-pr forge-et-be#139, a full review of 7ddda83 with backend-reviewer, 2026-10-02): block, three Importants, fixed in d4e9230 at the lead’s request. (1) POST /api/v1/mailbox/notifications is open and exempt from CSRF, and its handler binds @RequestParam validationToken, which on a multipart request makes Spring parse the whole body: a signed-out multipart POST there still stored an upload. A multipart POST to that path is now denied ahead of its permitAll (Graph never sends multipart). (2) A signed-out form POST with a matching cookie and header made Tomcat read up to its default 2 MB form limit into the heap when OAuth2AuthorizationCodeGrantFilter asked for parameters; server.tomcat.max-http-form-post-size is now 64KB, above which Tomcat answers 413 without reading the body. (3) The form-content-filter bug’s number, BUG-073, was taken by harness#345; it is BUG-076. New SecurityChainIT rows: the notification path (401) and a 1 MB announced form POST (413) both got no answer on the old code and pass on the fix; a chunked upload row passes on both. The guard also hides getPart/getParts, and the part-file check compares the directory before and after the request.
Third review (/forge-review-pr forge-et-be#139, a delta of d4e9230 with backend-reviewer, 2026-10-02): one Important, fixed in 27decdd by the PR’s author. A signed-out GET with a multipart body to /actuator/health or /actuator/info (both permitAll) still made the server parse and store the upload: a GET needs no CSRF token, the actuator’s adapter reads the whole parameter map on every call, and Spring and Tomcat parse a multipart body on any method. Confirmed live: no answer within 5 s, and a thread dump showed Request.parseParts under StandardMultipartHttpServletRequest.initializeMultipart. One authorization rule, permitUnlessMultipart, now replaces permitAll on health, info and the notification path, and replaces d4e9230’s separate deny rule, whose own path matcher could have drifted from the permit’s. New SecurityChainIT rows GET /actuator/health and GET /actuator/info (401): with d4e9230’s SecurityConfig exactly these two fail (32 tests, 2 failures); with the fix all 32 pass. Nits fixed in the same commit: partFiles() resolves a relative multipart location against TEMPDIR, as Tomcat does; the form-post-size comment says a chunked form body is read up to the limit before the 413, and that only parts with no filename count against it.
Resolved: 2026-10-02 — fixed in forge-et-be#139 (d9d53bf, a comment-only 5075240, the review fixes 98956cd; BUG-076 in d35090a; main merged in as 7ddda83; the second review’s fixes in d4e9230; the third review’s fix in 27decdd). ./gradlew check on d9d53bf: BUILD SUCCESSFUL in 23m 48s, 4005 tests, 0 failures, 2 skipped. CI’s full suite on 5075240 (gh workflow run "PR validation", run 36975731815): Fast lane and Slow lane passed. The local build record on d4e9230: 4021 tests, 0 failures. CI’s full suite on d4e9230 (run 36991716578): Fast lane and Slow lane passed. The local build record on 27decdd: 4023 tests, 0 failures, 2 skipped. CI’s full suite on 27decdd (run 36999465864): Fast lane and Slow lane passed. The regression tests above fail on the old code and pass on the fix.
Lesson: L-088 — the same cause as F4-a’s finding (code that runs before sign-in parses a multipart body); this bug showed it is any parameter read before authorization, not only the CSRF one, and that a check in plain local misses filters that exist only in the deployed shape. L-088 is extended with both, and backend-reviewer’s auth-coverage checklist gains a row pointing at it.
BUG-041 — A trailing slash on et.security.login-redirect-uri breaks the sign-in and mailbox redirects
Affects: F5-a, F20-a · Found in: the /forge-review-pr of forge-et-be#116 (BUG-037) · Severity: low · Phase: 1
Repro: start the backend with ET_SECURITY_LOGIN_REDIRECT_URI=https://crm.empire.lk/ (or http://localhost:5173/ in application-local.yml) and refuse a sign-in, or complete a mailbox connect.
Expected: the backend refuses to start and names the key, the way it refuses a malformed et.security.seed-admin-email; a value it accepts always yields …/login?login=failed and …/email/mailbox?connect=….
Actual: the value is only checked as non-blank (EtProperties.Security, @NotBlank). The failure handler (EntraLoginConfig.failureHandler, since BUG-037) and MailboxAuthorizationCompletionHandler append their route to it, so a trailing slash gives https://crm.empire.lk//login?login=failed and //email/mailbox?connect=…. The SPA’s routes do not match a doubled slash, so the marker is lost again and BUG-037 comes back. ContributedAuthorizationLandingController.isWithinRedirectBase already allows for a trailing slash, so nothing else in the code rules one out. No environment sets the value yet (cloud register §D, ET_SECURITY_LOGIN_REDIRECT_URI), so nothing is broken today.
Fix: validate the shape at startup with a @Pattern on loginRedirectUri: an http or https origin only, with no path, trailing slash, query or fragment (no environment needs a path, and the SPA is served at the root). The message names the key and gives https://crm.empire.lk as the example. Refusing the value, rather than stripping the slash silently, keeps one spelling of the value across the task definition, the CORS list and the Entra reply URLs.
Regression test: a binding spec that binds https://crm.empire.lk/ and expects startup to fail naming et.security.login-redirect-uri, and binds https://crm.empire.lk and http://localhost:5173 and expects them accepted.
Resolved: 2026-09-29 — fixed in forge-et-be#117 (with BUG-040). EtPropertiesSpec gains seven refusal rows (a trailing slash on a cloud and a local origin, a query, a fragment, no scheme, another scheme, a path), each asserting the constraint’s own message, and five accepted values (the prod, dev and UAT origins, the laptop origin and the test origin). The first six rows failed without the @Pattern and pass with it; the path row was added when the review tightened the pattern to an origin only. Every value already in the repo passes. The §D ET_SECURITY_LOGIN_REDIRECT_URI row in the cloud register states the rule.
BUG-040 — Laptop mode cannot switch between the test accounts
Affects: F5-i · Found in: F5-i (the first laptop sign-ins with CRM01–CRM05) · Severity: low · Phase: 1
Repro: in laptop mode, sign in with Microsoft as one test account (say CRM05@empire.lk), sign out of the CRM, then press “Sign in with Microsoft” again to use CRM01@empire.lk.
Expected: Microsoft asks which account to use, so a developer can sign in as each of CRM01–CRM05 in turn, as forge-et-be/docs/auth/laptop-sign-in.md describes.
Actual: Microsoft signs the browser straight back in as the last account, with no prompt. Signing out of the CRM ends only the CRM session; Microsoft’s own session stays, and the authorization request carries no prompt, so Microsoft reuses it. The only workarounds are a private window or signing out at login.microsoftonline.com.
Fix: under the entra-laptop profile only, add prompt=select_account to the sign-in authorization request (registration microsoft), so Microsoft always shows its account chooser. The mailbox consent (microsoft-mailbox) is unchanged: it should use the account already signed in, and a consent as a different account is refused anyway. Dev, UAT and production are unchanged, so Empire’s users are not asked (lead decision, 2026-09-29). It is a @Profile("entra-laptop") bean, not a profile check in code (coding-standards §5), applied through DefaultOAuth2AuthorizationRequestResolver’s request customizer, because a copy of an already-built request keeps its old URL.
Regression test: LaptopModeIT asserts prompt=select_account on the sign-in start and no prompt on the mailbox consent start; EntraLoginIT (no laptop profile) asserts no prompt on the sign-in start.
Resolved: 2026-09-29 — fixed in forge-et-be#117. The customizer is applied where both redirect filters get their resolver (ContributedOAuthClientConfig.contributedRequestResolver()): a first version on the login leg alone was skipped, because the client leg’s filter answered the sign-in start, which LaptopModeIT caught. The laptop test fails without the fix and passes with it; LaptopAccountChooserSpec proves the parameter reaches the built authorization URL and is left off the mailbox consent. On a live laptop-mode backend, /oauth2/authorization/microsoft redirected with prompt=select_account. docs/auth/laptop-sign-in.md step 6 says so.
BUG-039 — Flaky test: the sync-indicator AC-30 browser test sometimes finds mail where it expects an empty inbox (Windows)
Affects: F20-b (forge-et-fe/e2e/browser/mailbox/sync-indicator.spec.ts, the AC-30 row “at 768px the lapsed notice at /email fits”) · Found in: F40-WI-1.5 orchestrator live run (2026-09-29) · Severity: low · Phase: 1
Repro: in forge-et-fe, with forge-et-be on the local profile, run the full pnpm e2e:gate on Windows. On 2026-09-29 (branch feature/F40-WI-1.5-size-matrix, 93f170a) the AC-30 row failed once: settleMailboxScreen(page, 'inbox-empty') (e2e/browser/mailbox/support.ts:479) timed out, and the page showed three “Stub back-fill …” messages under the lapsed notice. Run alone (--project chromium -g "AC-30" --repeat-each 3 --workers=1) it passed 3 of 3, and it passed in the next full gate run (22fa6fe).
Expected: the row passes on every run.
Actual: the row calls stubConnect, forces the renewal failure, then waits for the empty-inbox body. The stub’s first connect also back-fills a small default mailbox asynchronously (support.ts notes this at waitForListedSubjects), so when the back-fill lands before the check the inbox is not empty. It is a race in the test, not a product defect.
Also seen (2026-09-29, feature/F40-wave-1 at be5a9a9): the AC-14 smoke row of the same file (“Reconnect in the inbox notice clears the danger state”, sync-indicator.spec.ts:410) failed the same way at settleMailboxScreen(page, 'inbox-empty') (support.ts:479) and passed on the next full run. So the race is in the shared stubConnect → settleMailboxScreen(INBOX_EMPTY) pattern, not in one row: two of six full pnpm e2e:gate runs that day failed on it. With F40’s gate live, every frontend PR author will meet it.
Fix: let the row accept the listed-inbox body as well as the empty one (the notice’s layout is what AC-30 checks), or wait for the back-fill to settle before forcing the lapse. Test-only change.
Regression test: the row itself, green across repeated full-suite runs.
Resolved: 2026-09-29 — forge-et-fe#81 (7d1251d). Cause: on the local profile a connect never subscribes and queues its back-fill; forcing a renewal failure creates a subscription, which seeds the stub’s default mailbox, so the queued back-fill may or may not ingest it before the inbox is read. Only this file connects, forces the lapse and then expects an empty inbox. The four such rows (the wave-2 smoke, AC-14, AC-29, AC-30) now settle on whichever inbox body appears (settleInboxBeneathNotice, exactly one of inbox-empty / inbox-list); what each asserts about the notice is unchanged. After: sync-indicator.spec.ts --repeat-each 5 --workers=1 68 of 68, and the full pnpm e2e:gate passed (348). Review (/forge-review-pr on forge-et-fe#81): pass-with-nits, root cause confirmed against the backend; nits left as they are — a redundant toContain inside the helper, the smoke row not checking that the notice comes before the list (pre-existing), and the populated-list-at-768px case being certain only in full-journey.spec.ts:197.
BUG-038 — programmaticFocusRing.test.ts fails because it compares \-separated paths (Windows)
Affects: F5-b (the app frame’s focus ring, forge-et-fe d2a35b5 “drop the focus ring on script-focused page headings”) · Found in: F40-WI-1.1 implementation (2026-09-28) · Severity: low · Phase: 1
Repro: in forge-et-fe on origin/main, on Windows, run pnpm test (under Node 24). One test fails: src/app/programmaticFocusRing.test.ts. It fails the same way on an untouched tree (git stash, then the file alone).
Expected: pnpm test exits 0 on every developer machine (NFR-3 of F40; the team works on Windows and macOS).
Actual: the test builds or compares file paths with the platform separator, so on Windows it sees \ where it expects /. It passes on Linux CI, so no lane catches it.
Fix: normalise the paths in the test (for example path.posix / split(path.sep).join('/')) before comparing. Test-only change.
Regression test: the test itself, green on Windows and on CI.
Resolved: 2026-09-29 — fixed in forge-et-fe#87 (merged 2026-09-29; its title carries no bug id, so the tracker was reconciled on 2026-09-30). The stylesheet scan writes each path with forward slashes (path.split(sep).join('/')), as src/test/mutedTextBand.test.ts already did, at both places the file built paths that way; the PR records the test passing on Windows, and it is unchanged on Linux and macOS, where sep is already /.
Lesson: none — one-off: one test compared platform-separated paths while its sibling (mutedTextBand.test.ts) already normalised them; running pnpm test on Windows as well as macOS (F40 NFR-3) is where the next instance would show.
BUG-037 — A refused Microsoft sign-in never shows “couldn’t sign you in”
Affects: F5-a · Found in: F5-i (the first laptop sign-in with the real Entra tenant) · Severity: medium · Phase: 1
Repro: in laptop mode (F5-i), sign in with Microsoft as an account that has no active CRM user. Or, without Microsoft, send the backend a sign-in callback with an invalid state: curl -sD - -o /dev/null "http://localhost:8080/login/oauth2/code/microsoft?code=x&state=y" answers Location: http://localhost:5173?login=failed. Open that URL signed out.
Expected: the login page shows the single “We couldn’t sign you in” message (F5-a FR-8, AC-10), or “Sign-in is unavailable” for ?login=unavailable (AC-10b).
Actual: the browser lands on /login with no failure message. On a browser that has been signed in before, the page shows “Your session expired. Please sign in again.” instead, which is wrong: nothing expired. The backend’s failure handler (EntraLoginConfig.failureHandler) redirects to et.security.login-redirect-uri + ?login=…, which is the SPA’s root, /. / sits inside AuthenticatedGate, which answers the unauthenticated /me with <Navigate to="/login"> and drops the query string, so LoginRoute never sees the marker. With no marker, LoginRoute falls back to the prior-session “expired” notice (BUG-007). Each side was tested alone: EntraLoginFailureSpec, EntraLoginIT and LaptopModeIT assert the backend’s Location, and LoginRoute.test.tsx opens /login?login=failed directly, so no test followed the redirect from one into the other.
Fix: the backend sends both markers to the login route, <login-redirect-uri>/login?login=failed and /login?login=unavailable, the way the mailbox completion handler already sends its outcome to /email/mailbox?connect=…. /login sits outside the gate, so the marker reaches LoginRoute. No frontend change. A deployed environment serves /login as a deep link the same way it serves /email/mailbox. With this fix a real refusal does land on /login?login=failed, so BUG-036 (Back re-shows the message) becomes reachable in practice.
Regression test: EntraLoginFailureSpec, EntraLoginIT and LaptopModeIT assert the /login path in the Location (they fail before the fix, which redirects to the root); plus a browser check in laptop mode that the backend’s redirect shows the message.
Resolved: 2026-09-29 — fixed in forge-et-be#116. With the fix set aside the three specs ran 40 tests completed, 14 failed, every failure the redirect path; with it, all 40 pass. In laptop mode a fresh Chromium context followed the backend’s refusal redirect into the running frontend: before, it landed on /login with no message (or “Your session expired” with the prior-session marker set); after, on /login?login=failed with “We couldn’t sign you in. Please try again.” either way.
BUG-036 — Back after sign-out can re-show an old sign-in failure message
Affects: F5-a · Found in: the review of forge-et-fe#75 · Severity: low · Phase: 1
Repro: with the real Microsoft sign-in (a laptop Entra sign-in, F5-i, or a cloud environment), start a sign-in that fails, so the browser lands on /login?login=failed (or ?login=unavailable). Press “Sign in with Microsoft” again and succeed, then sign out. Press the browser’s Back button until the history entry /login?login=failed is reached. Reloading that URL shows the same thing.
Expected: the login page shows the failure message only on the return from the sign-in that failed.
Actual: the login page shows “couldn’t sign you in” (or “Sign-in is unavailable”) again. LoginRoute.tsx reads the marker from the query string on every render and never removes it, and “Sign in with Microsoft” is a full-page navigation, so the entry stays in the history. The local stub sign-in does not show it: on success the login page’s <Navigate replace> replaces that entry. This existed before forge-et-fe#75 for ?login=failed; #75 added the second marker, which behaves the same way. The message is misleading but reveals nothing.
Fix: when the login page reads a login marker, keep the message in component state and replace the current history entry with /login (router replace, keeping any preserved from state), so Back and reload no longer carry it.
Regression test: a LoginRoute.test.tsx case per marker that opens /login?login=failed (and ?login=unavailable), asserts the message and that the location no longer has the query; and a browser row in e2e/browser/auth/login-page.spec.ts that opens the marker URL, then reloads and sees no message.
Resolved: 2026-10-01 — forge-et-fe#99 (head 0921665). A useLoginMarker hook in LoginRoute.tsx reads the marker into state and replaces the entry with a clean /login (setSearchParams(…, { replace: true, state: location.state })), as MailboxRoute.tsx does for ?connect=. It strips only once the session is unauthenticated, only the two known values, and keeps the gate’s from state; the BUG-007 expiry precedence and #101’s service-unavailable precedence read the same state. Fail-before: 4 new LoginRoute.test.tsx rows and the 4 new reload and Back browser rows failed on main; all pass after. On the real path in laptop mode, the backend’s refusal redirect (302 → /login?login=failed) re-showed the message on reload and on Back before the fix, and showed it once after. pnpm e2e:gate passed on 73cece3 (375 passed, 5 skipped). Review (/forge-review-pr on forge-et-fe#99): no Blocker. Applied: the strip waits for an unauthenticated session; the authenticated row asserts the destination keeps no router state, since it passed with the guard removed; the Back row asserts goBack() navigated. Escalated to the developer: a page restored from the back/forward cache shows the message again from React state. Playwright’s Chromium disables that cache, so no browser row can see it. The back/forward cache (cfe8f5a, the developer’s decision): in real Chrome on the built app against the Empire Entra tenant, a refused sign-in (CRM02@empire.lk), then a successful one (CRM01@empire.lk), a sign-out and Back showed the message again with a clean URL, and DevTools → Application → Back/forward cache said “Successfully served from back/forward cache”. useLoginMarker now also clears the message on pageshow with event.persisted, as the sign-in button already does for BUG-027. Accepted: a Back from Microsoft’s page after a refused sign-in also restores the page without the message. Fail-before: the two new persisted: true rows (one per marker) failed with the alert still present; a persisted: false row keeps the message. After the fix the same Chrome steps showed no message, with the panel still reporting a back/forward-cache restore. On 269adc4 (the fix plus a merge of main at c85f6f7) the checks and pnpm e2e:gate passed (375 passed, 5 skipped). Second review (/forge-review-pr on 269adc4, posted on the PR): no Blocker; one Important finding, escalated and decided by the developer as option A. The restored page also keeps its frozen session state. After a refused sign-in on a browser that held a session, that state is reason: 'expired' (SessionProvider’s settleUnauthenticated), which the failure message hides; dropping the message alone showed “Your session expired” after a deliberate sign-out, against BUG-007 — a temporary unit row proved it. f456a11 makes a restore re-read the session (refresh(), as “Try again” does) and drop the message once the re-read settles: a sign-out since then stays silent, a live session leaves for the app, a failed re-read still drops the message. Fail-before: the three new rows (built on the test file’s renderRetryingLogin) failed; all pass after. Also applied: the restore rows assert no role="status" notice took the message’s place, and the shared session mocks resolve refresh as the real one does. Deferred to BUG-066: one shared pageshow hook. On f456a11, each with exit 0: pnpm check, typecheck, test (1448 passed, 1 skipped), check:boundaries, check:testids, build, check:bundle; pnpm e2e:gate passed (375 passed, 5 skipped). Chrome on f456a11 (2026-10-02; the built app on pnpm preview, the laptop-mode backend against the Empire Entra tenant): with localStorage et.hadSession set to 1 first (a browser that held a session and was not signed out), a refused sign-in (CRM02@empire.lk), a successful one (CRM01@empire.lk), a sign-out and Back showed no message and no “Your session expired”. The back/forward-cache panel said “Successfully served from back/forward cache”, and DevTools showed a new GET /api/v1/me answered 401 on the Back. The same steps without the localStorage line, and a Back from Microsoft’s page after a refused sign-in (no message, the button usable), passed too. Third review (/forge-review-pr on 269adc4..f456a11, posted on the PR): approve, no Blocker or Important; seven Nits, all applied in 0921665 (tests and the hook comment only, no change to what the user sees). Nit 1 was the developer’s decision, option (a), accepted: a re-read that fails (the network, or a 5xx, at the moment of the Back) leaves the frozen session state, so after a frozen expired the expiry notice shows even after a sign-out, and after a frozen unavailable the service notice shows; two rows pin it. The per-marker restore rows now start from a frozen expired session, so their no-notice check can fail; breaking the fix on purpose (dropping the message without the re-read) failed all five restore rows. On 0921665, each with exit 0: pnpm check, typecheck, test (1448 passed, 1 skipped), check:boundaries, check:testids, build, check:bundle; pnpm e2e:gate passed (375 passed, 5 skipped). The PR also filed BUG-065 (filed as BUG-059 and renumbered), a test race the gate hit once.
Lesson: L-109 — a one-time message held in React state comes back when Chrome restores the page from its back/forward cache, so it also clears on pageshow with persisted. The mailbox screen’s ?connect=failed banner has the same cause (BUG-066, fixed in forge-et-fe#147); frontend-reviewer dimension 15 points at the lesson. BUG-037 also involved the ?login= marker, but its cause was the backend dropping it, not the page keeping it.
BUG-035 — A disabled primary button looks almost the same as an enabled secondary button
Affects: F5-b · Found in: the review of forge-et-fe#73 · Severity: low · Phase: 1
Repro: in the Empire theme, put a disabled primary Button (aria-disabled="true") next to an enabled secondary one.
Expected: a user can tell at a glance that the primary action is unavailable, not that it is a secondary action.
Actual: since forge-et-fe#73 every disabled button uses one look, the same at rest and on focus (design-system §5 Disabled row, Rev 31): surface/inset fill #F7F1E6, border/control #E0DCCE, text/secondary #6E5F4E. An enabled secondary button is surface/card #FFFFFF, the same border, and text/strong #5C4C39. The two differ only by a faint cream fill and slightly lighter text, and the not-allowed cursor shows only to a pointer user. The accessible name carries the reason (§9), so screen-reader users are not affected. This follows from the lead’s 2026-09-28 choice of one disabled look; it is a design call, not a regression.
Fix (lead decision, Tharinda, 2026-09-30, extended the same day after review — a neutral grey disabled look, “option B2”, for buttons and fields alike): every disabled control that today takes the surface/inset fill — Button in every family, Input, Select, Combobox, and the unchecked box of Checkbox and Radio (“the fields’ disabled fill”) — takes a neutral grey fill #E4E4E0, and its text takes grey #66625A (4.8 : 1 on the fill); a button’s border is the same colour as its fill, so none shows. §5 keeps one disabled look for every control. The grey has none of the palette’s cream or brown, so a disabled control no longer reads as a secondary button or an ordinary field. Unchanged: no opacity; the not-allowed cursor; a Switch’s off track and a DropdownMenu item, which keep their current disabled treatment; the focus ring (accent/gold-deep, drawn on the surface around the control: 5.9 : 1 on surface/card, 5.1 : 1 on surface/canvas); and the reason carried in the accessible name (§9). Accepted: with the border matching the fill, a disabled button’s edge is faint (about 1.1 : 1 against surface/canvas, 1.3 : 1 against surface/card); WCAG exempts inactive controls, and the grey fill is the signal. Considered and rejected: the current look (wontfix), a dashed border on the current colours, a darker warm stone fill (#E6E0D6 — its cream hue still reads as part of the palette), stone fill plus dashed border (too heavy where several controls are disabled), and grey for buttons only (it would split §5’s single disabled look). The work, in one change across both frontends so they stay the same:
-
two new tokens, suggested
--et-surface-disabled(#E4E4E0) and--et-text-disabled(#66625A):- in
forge-et-fesrc/ui/styles/tokens.cssfirst, with neutral grey fallbacks, becausetokens.cssdeclares every token a component may use, andtheme-neutral.cssis deliberately partial and falls back to it. The rule is stated in the header ofsrc/test/neutralRender.test.tsx(a property only an app theme declares draws as nothing), but no automated test checks it — Vitest runs withcss: false— so check the fallbacks by hand; - then in both apps’
src/ui/styles/theme-empire.csswith the Empire values, the contrast stated in each comment as the other text tokens do; tokens.cssis a byte-identical copy inforge-et-pwa(scripts/shared-with-fe.json): re-copy it, then runnode scripts/check-shared-drift.mjs --update --from=<forge-et-fe commit>. The capture app’s gate compares againstforge-et-fe’sorigin/main, so merge theforge-et-fePR first, or run the capture app’s gate with--fe-ref <branch>;
- in
-
forge-et-fecomponents: the[aria-disabled="true"]rules inButton,Input,SelectandComboboxuse both tokens (a button’s fill and border both from--et-surface-disabled); the unchecked-box rule inCheckboxandRadio(.input[aria-disabled="true"]:not(:checked)) uses only the fill, because the box holds no text; their comments are updated; -
the capture app has no shared button or field components and no disabled control yet, so it gets only the tokens now, and its controls use them when they are built;
-
design-system.md§5 Disabled row updated (it names the fill assurface/insettoday), with a## Revisionsentry; -
the browser check (
pnpm e2e:gatein both repos) and before/after screenshots, per the UI-bug rule. Regression test:forge-et-fesrc/app/disabledFocusRing.test.tsasserts, in both themes, that each listed control’s disabled fill differs fromsurface/card,surface/insetandsurface/canvas(adding entries for theCheckboxandRadiounchecked-box rule, which the test does not cover for its fill today); that the disabled text is at least 4.5 : 1 on the disabled fill forButton,Input,SelectandCombobox; and that the focus ring stays at least 3 : 1 on the surfaces a control sits on; its header comment, which lists each control’s disabled look, is updated to match. It fails on the current look (the fill issurface/inset). Resolved: 2026-09-30 — forge-et-fe#90 and forge-et-pwa#40, as the extended decision above. -
Commits: forge-et-fe
44cf2dd(buttons),15a7f4c(review fixes),dbaa40f(fields),897832d(review fixes),734f023(mergesmain). forge-et-pwadffcc80,e2f2ac7(review fixes),9738af0(the fields extension). -
Merged 2026-09-30: forge-et-fe#90 as
5d8d2e1, then forge-et-pwa#40 as64808be. Before #40 merged,node scripts/check-shared-drift.mjs --livematched all six shared files against forge-et-femainat5d8d2e1. The gate on #90’s merged tree (734f023) passed on its third run (349 passed); the two runs before it each failed one different row on the 5-second visibility wait under load, the BUG-045 pattern. -
Merge order: merge #90 first. The capture app’s
tokens.csscopy names forge-et-fe897832d, and the live shared-file check in itspnpm e2e:gatecompares with forge-et-feorigin/main, so until #90 merges that gate needs--fe-refat #90’s branch. CI’s hash check does not depend on the order. -
The look:
Button(every family),Input,Select,Combobox, and the uncheckedCheckboxandRadiobox take--et-surface-disabledfor their disabled fill (#E4E4E0). The buttons and fields take--et-text-disabledfor their text (#66625A, 4.76 : 1). A button’s border takes the fill’s colour; fields keep their control border. The names and neutral fallbacks (#E4E4E7/#52525B) are intokens.css, copied to the capture app with its hash updated, andtheme-neutral.csshas#DCDFE4/#4B5563.Switchand theDropdownMenuitem are unchanged. -
Regression test:
disabledFocusRing.test.tsruns the grey checks over all six controls, in Empire, neutral and the no-theme fallback:- the fill is at least 1.2 : 1 from the fill the same control draws when enabled (read from its own rules; for a button, the white-filled families), and at least 1.08 : 1 from card, warm card, inset and canvas;
- the text clears 4.5 : 1 on the fill, and the ring clears 3 : 1 against it;
- a button’s border matches its fill, and the rule sets no
border:shorthand; - a checked, disabled
CheckboxorRadiokeepsaccent/gold(the base disabled rule sets no fill).
The Fix line notes that nothing checks the
tokens.cssfallbacks by default (Vitest runs withcss: false). This test’s no-theme layer resolves every name fromtokens.cssalone and holds the fallbacks (#E4E4E7/#52525B) to the same floors, so they are checked for these two tokens.Fail-before: 70 failed | 100 passed on
main’s rules. After: 172 passed; the unit suite has 1352 passed | 1 skipped. The capture app’sthemeContrast.test.tsholds its values to the same floors (1.2 : 1 fromsurface/cardandsurface/field): before,--et-surface-disabled is declared nowhere; after, 27 passed. -
Browser: live DOM on
/loginat 1280 × 860.- The disabled button went from inset fill
rgb(247,241,230)with the control border torgb(228,228,224)for fill and border, with labelrgb(102,98,90)at 4.76 : 1. - Disabled
Input,Select,CheckboxandRadiobuilt from the real classes all computed fillrgb(228,228,224); the fields’ text wasrgb(102,98,90). - On keyboard focus, a disabled button and a disabled input matched
:focus-visibleand drew thegold-deepring: 5.91 : 1 on the card, 4.64 : 1 against the fill. Opacity was 1, and the look was unchanged from rest. - Before and after screenshots were taken at rest, on focus, with a disabled primary next to an enabled secondary, and with each field enabled next to disabled.
- The disabled button went from inset fill
-
Gates: forge-et-fe
Browser gate: passed · 897832d06bd8 · 349 passed · 5 skipped · 0 known failures; forge-et-pwaBrowser gate: passed · 9738af0a795f · 123 passed · 0 skipped · 0 known failures(run with--fe-refat #90’s branch). Two earlier forge-et-fe runs failed on load, not on this change; they are recorded under BUG-045. -
Reviews (
/forge-review-pr), no Blocker in any:- forge-et-fe#90, first pass. Applied in
15a7f4c: contrast floors in place of plain inequality, theborder:shorthand check, and hue wording (the grey was called “cool”; it has a slight yellow cast). - The fields commit
dbaa40f. Applied in897832d: the checked-box row and the token comments (“a disabled control’s”, not “a disabled button’s”). Left as-is: an invalid and disabled field gets the grey only because the disabled rule comes later in each file, at the same specificity as the invalid rule. That was already true of the inset fill, so this fix did not introduce it. - forge-et-pwa#40. Applied in
e2f2ac7: the capture rows moved to the CRM’s floors, header wording, a re-wrapped comment, and the merge-order reason, which is corrected above. Lesson: none — one-off: Rev 31’s single disabled look reused an existing surface (surface/inset) that sits 1.12 : 1 from the white of an enabled secondary button and the cream of an enabled field. No otherLesson:line or lesson has that cause. The test’s floor against each control’s own enabled fill is where the next instance would show.
- forge-et-fe#90, first pass. Applied in
BUG-034 — The connect screen’s keyboard row presses Space before the dialog’s close has reached React
Affects: F20-a · Found in: the whole-suite run for BUG-028 (forge-et-fe#62) · Severity: low · Phase: 1
Repro: on a fresh local stack, run e2e/browser/mailbox/connect-a11y.spec.ts -g "AC-28: connected is reachable" with --workers=1. It fails on every run from forge-et-fe#56 (84a2c14) onward and passes on the F20-b merge before it (17a2d4c).
Expected: Enter opens the Disconnect confirm dialog, Esc closes it, and Space then opens it again.
Actual: the Space press opens nothing. Esc closes the <dialog> at once, but Modal hears about it only through the close event, which the browser dispatches one task later. Until then the open prop is still true, so a press inside that gap sets confirming to true (a no-op) and the late close event then sets it to false. Adding any delay before the press (even a console.log of the button’s state) makes the row pass. A person cannot press within one task of Esc, so this is a test race, not a product defect.
Fix: after Esc, the helper waits until the dialog’s content has unmounted (the close button’s count is 0), which is the point where React has processed onClose. No product change.
Regression test: the row itself, run five times serially with --workers=1, passing every time.
Resolved: 2026-09-27 — forge-et-fe#63 (fe744cd). expectKeyActivates waits after Esc for mailbox-disconnect-dialog-close to unmount, and first asserts that it is visible, so a renamed test id fails rather than silently removing the wait. Fail-before: 3 of 3 serial runs on main. After: 5 of 5, and the whole e2e/browser/mailbox folder passes (82). Review (/forge-review-pr on forge-et-fe#63): no Blocker; the visible-first assertion is the review’s fix.
BUG-033 — The confirm-dialog axe row scans while the dialog is still fading in
Affects: F20-a · Found in: the whole-suite run for BUG-028 (forge-et-fe#62) · Severity: low · Phase: 1
Repro: on a fresh local stack, run e2e/browser/mailbox/connect-a11y.spec.ts -g "open confirm dialog" with --workers=1. It fails on every run, on main and on the F20-b merge (17a2d4c).
Expected: axe finds no WCAG 2.1 AA violation in the open Disconnect confirm dialog.
Actual: axe reports color-contrast (serious) on the dialog’s Cancel and Disconnect buttons. The dialog animates in from opacity: 0 (modal-in, --et-motion-base), and the row scans as soon as the dialog is visible, so axe measures the buttons partly transparent. Once the dialog’s animations have finished, the same scan is clean (three of three runs), so the settled colours pass.
Fix: the row waits for every animation in the dialog to finish (getAnimations({ subtree: true })) before it scans. No product change.
Regression test: the row itself, passing on repeated serial runs. Before the fix it fails on each run.
Resolved: 2026-09-27 — forge-et-fe#63 (fe744cd). The row waits for every finite animation in the dialog to finish before the scan. Fail-before: 3 of 3 serial runs on main. After: 5 of 5. Review: skipping infinite animations (a busy spinner never finishes, so the wait would hang) is the review’s fix.
BUG-032 — Twelve auth and shell browser specs expect the pre-F1-a landing page
Affects: F5-a, F5-b · Found in: F20-b (the whole-tier runs recorded in .forge/plans/F20-b/F20-b-test-report.md; lesson L-068) · Severity: medium · Phase: 1
Repro: run the forge-et-fe Playwright suite against any local stack. Seven specs in e2e/browser/auth and five in e2e/browser/shell fail on main.
Expected: the suite asserts what the spec requires (F5-b FR-12 / AC-12): sign-in lands on the first nav destination the session holds.
Actual: F1-a granted the seeded e2e admin account.view, and on a fresh local stack GET /api/v1/me shows it also holds lead.view (checked 2026-09-27), so that admin holds all six nav keys and lands on /accounts. The suite still expects the older state. SUPERADMIN_LANDING in e2e/browser/auth/support.ts names /email, even though its own comment says the move to /accounts would come with F1. frame.spec.ts and nav-permissions.spec.ts expect four nav entries, with Accounts and Leads hidden. The product behaves as specified; the tests are stale. No PR lane runs this suite, so main never went red (L-068).
Fix: point SUPERADMIN_LANDING at /accounts and the Accounts heading. frame.spec.ts expects all six entries from the real session. nav-permissions.spec.ts’s hide-path test gets its partial session from the approved /me mock (every nav key except account.view and lead.view), because the real admin now has nothing hidden. That keeps the check that hidden entries at the top of the list leave no gap.
Regression test: the twelve rows themselves, passing on a fresh local stack and on the long-running one. Until F40 puts the suite in CI, a later change that moves the landing again shows up only when someone runs the tier.
Resolved: 2026-09-27 — forge-et-fe#61 (F5-k WI-1.2), which re-anchors the same ten specs on the new default user for AD #34: DEFAULT_LANDING names /accounts, frame.spec.ts expects all six entries, and Scenario B of signin-session-signout.spec.ts preserves /pricing rather than /accounts. That last point matters beyond the landing: /accounts had become the default, so Scenario B passed without proving that the route is restored.
- An independent fix was first written into forge-et-fe#63 and then removed from it, because the two PRs changed the same ten files and would have conflicted.
- Proof that the diagnosis is right: that independent fix brought the whole suite to 161 passed, on a fresh throwaway database and on the developer’s database that had been running for three days. Unchanged
maingave 147 passed / 14 failed on the fresh database (these 12 plus BUG-033 and BUG-034). - Worth adding to forge-et-fe#61 before it merges: a guard in Scenario B that the preserved route does not match
DEFAULT_LANDING, so a later landing move fails loudly instead of silently weakening the row again. - The six mailbox and seam rows that had also failed were a stale local
application-local.ymlwith noet.messaging.mailbox-ingest-queue, which F20-b added to the example. That is environment, not a defect;forge-et-fe/e2e/README.md § Running it locally(forge-et-fe#63) now covers it.
BUG-031 — The capture app’s 404 button is 40px tall, under the 44px touch floor
Affects: FS-1 (the capture shell) · Found in: the 2026-09-27 accessibility and responsive audit · Severity: low · Phase: 1
Repro: open the capture app at a phone width (360, 390 or 430px) and go to a path that does not exist.
Expected: every interactive target is at least 44 × 44px (design-system § The capture app).
Actual: NotFoundRoute.tsx’s button measures 135 × 40px. Axe is clean; the target-size floor is the capture app’s own usability rule.
Fix: give the button a 44px minimum height from the tokens, so no pixel literal is needed.
Regression test: a browser row at a phone viewport that asserts the button’s box is at least 44px tall; it fails before the fix and passes after.
Resolved: 2026-09-27 — forge-et-pwa#25 (1a65984). The button takes min-height: var(--et-layout-control-h) (44px). The new row in e2e/a11y.spec.ts runs at 360 × 780 and also asserts no horizontal scroll and a clean axe scan, because the 404 page had no scan before. Fail-before: the row failed on the old code (39.5 < 44). pnpm check, typecheck, test, check:boundaries, check:testids, check:shared and build all exit 0; playwright test 2 passed. Review (/forge-review-pr on forge-et-pwa#25): approve. Applied in 34a66d9: the new CLAUDE.md rule states the 14px capture body size, which the shell does not meet yet (base.css uses the shared 13px type/body), so the rule now says so and points at F40, which owns the fix.
BUG-030 — The login page scrolls sideways at 320px (it is 528px wide)
Affects: F5-a · Found in: the 2026-09-27 audit · Severity: medium · Phase: 1
Repro: open /login in a 320px-wide viewport, which is a 1280px desktop at 400% zoom.
Expected: the page reflows with no horizontal scroll (WCAG 2.1 SC 1.4.10; in scope since the lead decision of 2026-09-27, design-system Rev 28).
Actual: document.documentElement.scrollWidth is 528 at a clientWidth of 320. “Sign in with Microsoft”, the stub email input and “Continue as this user” sit partly off-screen.
Fix: the sign-in card takes max-width rather than a fixed width and shrinks with the viewport, with no raw pixel value.
Regression test: a browser row at 320 × 640 that asserts no horizontal scroll and every control inside the viewport; it fails before the fix and passes after.
Resolved: 2026-09-27 — forge-et-fe#62 (41ba1cc). Deviation from the Fix line above: the card’s width was not the cause. The brand and sign-in panels sit side by side, and the sign-in panel cannot shrink below its card plus padding (464px). A max-width alone leaves the page wider than 320px, and a CSS-only flex-wrap changed the tablet layout. Below bp/tablet the panels now stack: LoginRoute.tsx reads matchMedia(mq.tablet) and sets data-layout="stacked", because the repo forbids a raw pixel width in a CSS @media. Panel widths at 768px (304 / 464) and 1024px (543 / 481) are identical before and after. Fail-before: the new row in login-a11y.spec.ts failed on the old code (scrollWidth − clientWidth = 208).
BUG-029 — The mailbox page’s back link is 20px tall, under the 24px target floor
Affects: F20-a · Found in: the 2026-09-27 audit · Severity: medium · Phase: 1
Repro: open /email/mailbox at any size and measure the “Email” back link.
Expected: at least 24 × 24px (design-system §9 touch targets; WCAG 2.2 SC 2.5.8).
Actual: the BackLink measures 53 × 20px at every size, from 768px to 1920px.
Fix: in the shared BackLink component, a minimum block size of the 24px target from the tokens (so every back link is fixed, not only this one), with alignment unchanged.
Regression test: a component or browser row that asserts the back link’s box is at least 24px tall; it fails before the fix and passes after.
Resolved: 2026-09-27 — forge-et-fe#62 (41ba1cc). BackLink.module.css gains min-block-size: var(--et-space-7) (24px). The browser row is in connect-screen.spec.ts (AC-18). Fail-before: it failed on the old code (19.5 < 24).
BUG-028 — At tablet widths the closed sidebar is aria-hidden but its links stay focusable
Affects: F5-b · Found in: the 2026-09-27 audit · Severity: high · Phase: 1
Repro: sign in; set the viewport to 768 or 820px wide (the tablet tier, where the sidebar is an overlay); leave the sidebar closed; press Tab through the page.
Expected: nothing inside the hidden sidebar can take focus. Axe reports no aria-hidden-focus.
Actual: axe reports aria-hidden-focus (serious) on the aside on every signed-in page. The six nav links and the current-user button sit off-screen but stay in the tab order. Sidebar.tsx sets aria-hidden={isOverlay && !drawerOpen} and nothing else, so a keyboard user tabs into links they cannot see and a screen reader does not announce. The existing axe specs scan at desktop size only, which is why this was missed.
Fix: add inert alongside aria-hidden while the overlay is closed, which removes the subtree from the tab order and the accessibility tree. Opening the drawer removes both.
Regression test: a browser row at 768px that runs axe with the sidebar closed and asserts Tab never lands inside the aside, plus a component row that asserts inert is set when closed and absent when open. Both fail before the fix and pass after.
Resolved: 2026-09-27 — forge-et-fe#62 (41ba1cc, plus aa66aa1). Sidebar.tsx sets inert with aria-hidden while the overlay is closed. Fail-before: the Sidebar.test.tsx row and the responsive-reflow.spec.ts row both failed on the old code; the browser row’s Tab walk fails on its own (focus reached the hidden rail) before the axe scan runs. pnpm check, typecheck, test (790 passed), check:boundaries and check:testids all exit 0. The full Playwright suite gives 131 passed / 18 failed on the branch, and the same 18 fail on unchanged main against the same local stack, none involving these files. Review (/forge-review-pr on forge-et-fe#62): approve-with-fixes. Applied in aa66aa1: the Tab walk moved ahead of the axe scan so each is proven. Left as is: the open tablet drawer has no focus trap. That is a separate question, covered by F40’s keyboard walkthroughs.
BUG-027 — “Sign in with Microsoft” can start the Entra sign-in twice on a double press
Affects: F5-a · Found in: F20-b follow-ups (while fixing BUG-024, the same defect on the mailbox consent buttons) · Severity: low · Phase: 1
Repro: in a deployed bundle, press “Sign in with Microsoft” on /login twice in quick succession.
Expected: the first press starts the redirect and the button stays busy until the browser leaves the page.
Actual: LoginRoute.tsx wires the button straight to startMicrosoftSignIn, which calls location.assign. That returns before the browser leaves, and the button has no busy state, so a second press sends a second authorization request to Entra.
Fix: the BUG-024 rule. The button is busy from the press until the page is gone; it becomes usable again if the start throws, or on pageshow with persisted: true (the page restored from the back/forward cache after Back on the Microsoft page). The mailbox’s useConsentRedirectBusy lives in features/email, and features do not import each other, so the login page keeps its own small copy. Regression tests: component rows (a second press starts nothing; usable again after a back/forward restore; usable again when the start throws) and a browser row (the navigation answered 204 so the page stays; the button reads busy).
Resolved: 2026-09-25 — forge-et-fe#57 (691adf6). Fail-before: 2 component rows failed on the old code, and the browser row failed with aria-busy "false" in Chromium; all pass on the new code. pnpm check, typecheck, test (751 passed), build, check:boundaries, check:testids, check:bundle all exit 0. The e2e/browser/auth tier fails the same 7 specs with and without the change: the ones already red on main since F1-a’s landing change (L-068), none involving this button.
BUG-026 — The mailbox panel unmounts on every connection re-read, so keyboard focus is dropped
Affects: F20-a, F20-b · Found in: F20-b (forge-et-fe#55 review, finding 1; the connect-a11y flake) · Severity: medium · Phase: 1
Repro: on /email/mailbox (or /email with a lapsed sync), press Reconnect with the keyboard. The dev stub connect runs, then refetch re-reads the connection.
Expected: the panel stays on screen while it re-reads, and focus stays on, or returns to, a sensible place in it (WCAG 2.4.3 focus order).
Actual: useMailboxConnection’s refetch sets the result back to { state: 'loading' }, so the whole connected or reconnect panel is replaced by the loading render and then rebuilt. The focused button no longer exists, focus falls to <body>, and a keyboard or screen-reader user has to start again from the top of the page. The same unmount is the likely cause of e2e/browser/mailbox/connect-a11y.spec.ts:148 failing once in 11 runs: pressing Enter on Disconnect did not open the confirm dialog within 5 s, which fits a re-render that replaced the button between the focus and the key press.
Fix: during a re-read after the first load, keep the last settled result on screen and mark it as refreshing, instead of going back to loading (the behaviour TanStack Query gives by default, which F5-c or F1-b will move this hook onto). While refreshing, the panel’s action buttons are aria-busy and disabled. When the re-read settles, if focus was inside the panel and has been lost (the element is gone or focus is on <body>), move it to the panel’s heading (tabIndex={-1}). First load and error rendering are unchanged. Regression tests: a hook row (a refetch after a settled load never publishes loading), a component row (focus is on the heading after Reconnect re-reads), and a repeat run of the connect-a11y keyboard spec, confirmed in a running browser.
Resolved: 2026-09-25 — forge-et-fe#56 (eb8e8d8, plus 8b225d4 for the retry’s busy-while-refreshing). The hook keeps a settled connected/not-connected result with refreshing: true during a re-read; the first load and a retry from error still pass through loading. useRefetchRestoringFocus moves focus to the screen’s h1 (tabIndex={-1}) when a re-read removed the focused control — the panels have no heading of their own, so the screen heading is the target. Fail-before: 8 new unit/component rows failed on the old code. Browser: connect-a11y.spec.ts:148 ran 11 times for each of its four states with no failure (it had failed once in 11), and a live keyboard Connect and Disconnect on /email/mailbox left focus on H1 "Mailbox". Review (/forge-review-pr on forge-et-fe#56): no Blocker. Applied in cf5dd1c: rows for focus moved outside the screen (must not be stolen) and for a failed re-read (focus to the h1); the JSDoc states the scope element must be mounted on first commit; the e2e re-read helpers press Enter on a focused control rather than clicking. Left for the developer (UX): after a successful Disconnect the panel shows “Connected” with a busy Disconnect for one round trip until the re-read lands (TanStack Query behaves the same); and a screen reader hears only the heading after the focus move, not the new state.
BUG-025 — Subscription renewal treats a 401 or a non-HTTP error as a lapse, and asks the user to reconnect
Affects: F20-b · Found in: F20-b (forge-et-be#57 review, item E1) · Severity: medium · Phase: 1
Repro: a renewal run where Graph answers 401 (a rejected access token), or where the call ends in a non-HTTP exception such as IllegalStateException("Graph accepted the renewal but returned no expiry").
Expected: a one-off failure is logged and retried on the next renewal run. The subscription is marked lapsed only when Graph says it is gone (404/410), refuses it outright (other 4xx), or its expiry has actually passed (EXPIRED, which the renewal run already records).
Actual: MailboxRenewalService.classify falls through to REFUSED for everything it does not recognise, so a 401 and any non-HTTP exception mark the subscription FAILED with RENEWAL_FAILED. The inbox then shows the lapse notice and asks the user to reconnect, though the subscription may still be live. MailboxRenewalServiceSpec pins this (“an unusable answer” row).
Fix: developer decision 2026-09-25: retry on the next run. classify returns TRANSIENT for a 401 and for any exception that is not a RestClientResponseException. On a 401 the cached access token for the connection is also invalidated, so the next run asks for a fresh one rather than reusing the rejected token. 400/403 and other 4xx stay REFUSED; 404/410 stay GONE. The class Javadoc and scheduler-cloud-contract.md’s renewal row are updated to match. Regression test: MailboxRenewalServiceSpec rows for 401 and for a non-HTTP exception, each DEFERRED with no lapse recorded, plus the 401 row asserting the token is invalidated. The existing “an unusable answer” row moves from the lapse table to the new one.
Resolved: 2026-09-25 — forge-et-be#60 (4fa8fee). classify returns TRANSIENT for a 401 and any exception that is not a RestClientResponseException; a 401 also invalidates the cached token. The renewal row in docs/messaging/scheduler-cloud-contract.md now lists which failures lapse and which retry. Fail-before: 3 of 22 MailboxRenewalServiceSpec rows failed on the old code (an unexpected RENEWAL_FAILED lapse each); full ./gradlew check green after. Review (/forge-review-pr on forge-et-be#60): Important — a deferred non-HTTP failure logged only its class name, so a recurring programming error would go unexplained until EXPIRED; fixed in ab1177d (the exception goes to DEBUG after the WARN, as the token-deferral branch does). Nit — 408 Request Timeout now retries too (same commit, in line with the retry decision). Left as is: tokens.invalidate clears one instance’s cache only (documented in the Javadoc).
BUG-024 — Connect and Reconnect can start the Microsoft consent twice on a double press
Affects: F20-a · Found in: F20-b (forge-et-fe#55 review, finding 4 follow-up) · Severity: low · Phase: 1
Repro: in a deployed (non-DEV) bundle, press Connect on /email/mailbox (or Reconnect on the reconnect-required panel) twice in quick succession.
Expected: the first press starts the consent redirect and the button stays busy until the browser leaves the page.
Actual: NotConnectedPrompt.tsx and ReconnectRequiredPanel.tsx clear busy in a finally block after beginMailboxConnect. In a deployed bundle that call has just run location.assign, which returns before the browser leaves, so busy clears and a second press calls location.assign again — a second authorization request to Entra. The lapse notice had the same defect and was fixed in forge-et-fe#55 (32572d3).
Fix: the same pattern as SyncLapsedBody: busy clears only when import.meta.env.DEV (the stub connect re-reads in place) or when beginMailboxConnect rejects. Regression test: a row in each component’s test file that stubs DEV to false and asserts the button stays aria-busy and ignores a second press, confirmed failing against the old code.
Resolved: 2026-09-25 — forge-et-fe#56 (80f4a38, 8b225d4). Connect, Reconnect and the ?connect=failed banner’s “Try connecting again” (the same defect, found while fixing) clear busy only under import.meta.env.DEV or on a rejection. Fail-before: the new rows failed on the old code with aria-busy="false". Not covered here: the login page’s “Sign in with Microsoft” button (LoginRoute.tsx) had the same shape; it belongs to the sign-in feature and is BUG-027. Review (/forge-review-pr on forge-et-fe#56): F1 (Important) — in a deployed bundle a button stayed busy for good if the browser restored the page from the back/forward cache after the user pressed Back on the consent page; fixed in 9a86999: one shared hook, useConsentRedirectBusy, used by all four consent buttons (the lapse notice’s included), also clears busy on pageshow with persisted: true. Fail-before: 4 new rows failed on the old code. No timeout fallback for a user pressing Stop before the redirect lands (left as is: a timeout would re-allow the double press this bug fixes).
BUG-023 — Disconnect and re-consent answer 409 when a background token refresh saves the connection first
Affects: F20-a, F20-b · Found in: F20-b (WI-2.1 in-wave repair 2, ## Notes → “In-wave repair 2 — the same race on two user actions”) · Severity: medium · Phase: 1
Repro: connect a mailbox (the connect clears the token cache, so the back-fill job’s first token request always refreshes), then Disconnect, or re-consent, straight away. Also reachable later, roughly once per access-token lifetime, whenever the delta poll or a notification ingest refreshes the token.
Expected: the Disconnect or re-consent succeeds.
Actual: MailboxConnectionService.disconnect and connectOrUpdate read the row through liveConnectionFor with no lock, then save it with a version check. When a worker’s token rotation (MailboxTokenService.applySuccess, under the row lock) commits between that read and the save, the save matches no row and the user gets 409 STALE_VERSION (“The record was changed by someone else”). When the access token is not cached, Disconnect instead fails with 500: the token path’s own locked read hits the stale version inside the same transaction, the best-effort catch logs it, but the transaction is already rollback-only, so the commit throws UnexpectedRollbackException (found by the fix’s regression test, not in the original report). Offboarding’s releaseConnectionsFor fails the same way. No data is lost — the version check is what stops a stale refresh token overwriting the rotated one — but the user sees an error on a routine action. The two stub endpoints had the same defect and were fixed in WI-2.1 (be 081e271, 77d9387).
Fix: read the owner’s live connection with a pessimistic write lock at the start of both paths, before anything else loads the row: a new MailboxConnectionRepository finder with @Lock(PESSIMISTIC_WRITE) over the same predicate as findFirstByOwnerUserIdAndStatusNot. The lock must come first, because a locking query for an entity already in the persistence context returns the stale copy (WI-2.1 repair 3). disconnect and connectOrUpdate then serialise with the token path the way the stub endpoints now do; the token acquisition inside release runs in the same transaction and so re-enters the lock it already holds. releaseConnectionsFor (offboarding) gets the same treatment per row. Regression tests: an IT row for each path that holds the row FOR UPDATE in a second JDBC transaction, waits until the call is blocked on it (pg_blocking_pids), bumps version and commits, then asserts the call succeeds — the same shape as the stub endpoints’ rows — confirmed failing with 409 on the old code.
Resolved: 2026-09-25 — forge-et-be#60 (8a8728e). New locking finders findFirstLockedByOwnerUserIdAndStatusNot / findAllLockedByOwnerUserIdAndStatusNot are the first read in disconnect, connectOrUpdate and releaseConnectionsFor. Four new IT rows (Disconnect with and without a cached token, re-consent, offboarding release) failed on the old code — 409, 500, ObjectOptimisticLockingFailureException, UnexpectedRollbackException — and pass on the new. ./gradlew check EXIT=0, 1479 tests, 0 failures. In a live browser, Connect then an immediate Disconnect answered DELETE 204. Review (/forge-review-pr on forge-et-be#60): no Blocker; the locking is the first read on every path (OSIV off), Hibernate takes FOR NO KEY UPDATE, and no lock cycle with the workers is possible. Nits applied: the race ITs wait for the blocked call in cleanup() (63ee093); the offboarding finder’s Javadoc says the V13 index allows one live row per owner (00a6a75). Left as is: extracting one shared lock-race test helper (style), and MailboxTokenService’s pre-existing “a few hundred milliseconds” worst-case comment (not this diff). After origin/main was merged in, ./gradlew check EXIT=0, 1558 tests, 0 failures.
BUG-022 — Connecting a mailbox through the real Microsoft consent path checks no permission
Affects: F20-a · Found in: the 2026-09-24 access-model audit of forge-et-be main · Severity: medium · Phase: 1
Repro: on a cloud profile, sign in as a user whose role does not hold the mailbox key, and start the Microsoft mailbox consent from the Email screen’s connect action.
Expected: the connect is refused with 403, the same as reading or disconnecting a mailbox without the key (F20-a FR-3, FR-11).
Actual: MailboxAuthorizationCompletionHandler.complete creates the mailbox_connection without any permission check — neither it nor CapturingTokenResponseClient holds a PermissionChecker — so the connection is made. Reading and disconnecting are gated through MailboxConnectionService.callerWithMailboxAccess; the local stub connect is gated too. The effect is limited to the caller’s own mailbox, which is why it is medium, not high.
Fix: F5-k renames the key to email.manage (AD #34) and gates every mailbox path on it, the consent start and completion included, with a test that a role without the key is refused at both ends of the consent flow. Recorded here so the fix is tracked if F5-k’s plan is split.
Resolved: 2026-10-02 — forge-et-be#122 merged (written 2026-09-29; edc6b7b tests, 75da4a7 fix). F5-k wave 1 renamed the key but left both ends of the consent unchecked (re-checked on origin/main d24e752), so the gate moved out of F5-k WI-2.1 and shipped on its own; it meets F5-k AC-5. Start: OAuth2RegistrationContribution now carries requiredPermission, which MailboxOAuthRegistration sets to email.manage. ContributedRegistrationRequestResolver returns null for a signed-in caller without it, on both redirect filters. A null alone falls through authenticated() to a 404, so SecurityConfig adds a rule on each contributed authorization path that asks the same key. A signed-in caller without it gets the same 403 FORBIDDEN envelope as GET /mailbox/connection; an anonymous one still gets 401. Completion: CapturingTokenResponseClient checks each contributed registration’s key before its contributor runs, so auth gates both ends for any future registration too. MailboxAuthorizationCompletionHandler checks the key again as a second line (MailboxAuthorizationCompletionHandlerSpec, since over HTTP auth refuses first). Either way the callback lands on ?connect=failed with no MailboxProvider (Graph) call and no row. The Entra code exchange has already run by then; its tokens are local values, discarded with no copy kept. That covers a key removed while the consent screen is open. Checked by permission, never role (AD #26). Logged, not audited: no audit code covers a refused permission, and the other mailbox 403s are not audited (lead’s call). Regression: three new rows (MailboxAuthorizationAccessIT ×2, MailboxConnectFlowIT ×1) failed on edc6b7b: 23 tests, 3 failed, the start answering 302, both redirect filters producing a request, and the completion not refused. They pass on 75da4a7. ./gradlew check there found one more consequence: LaptopModeIT used a user with no grants for its signed-in consent start. That row now seeds a user holding the key, and gains a 403 row for one without it. origin/main was merged in twice, both sides kept. The first merge (BUG-040/041) changed the same resolver constructor; seven affected specs then passed, 200 tests, 0 failures. The second (BUG-043/044/045/049) gave the same record hintSignedInUser; its convenience constructor defaults only the hint, never the key. Seven affected specs then passed, 102 tests, 0 failures. Review (/forge-review-pr on forge-et-be#122, with backend-reviewer): no Blocker. Important 1: four backend docs still described the old behaviour, including laptop-sign-in.md telling you to connect as CRM01, whose IT Admin role holds no email.manage. Fixed, with the local grant step added. Important 2: the completion was gated only by the mailbox contributor, not by auth. The lead chose to add the auth-level check and keep the handler’s. Fail-before: the connect-flow BUG-022 row, now asserting auth’s refusal line, failed on the handler-only code (3 tests, 1 failed). Nits applied: an undeclared or duplicate consent key now stops the context at startup (ContributedOAuthClientConfigSpec); the resolver row asserts it ran against both redirect filters; the laptop doc’s SQL notes the scope column F5-l removes. A third merge of origin/main brought F5-l, which drops role_grant.scope (V31). LaptopModeIT’s seeded grant and the laptop how-to’s SQL no longer write it, so the scope note above is gone. Second review (/forge-review-pr at dd33ce7, with backend-reviewer), covering everything after the first: no Blocker, no Important. Under F5-l, checker.can(key) gives the same answer for email.manage at all four call sites. Nits: the missing-key branch logged “does not hold null”, now split into its own message (8beea63), and CapturingTokenResponseClientSpec covers both refusal branches. Left as is: no test proves the config passes catalog::isDeclared, but an undeclared key is never held, so the flow still fails closed. Live walk (2026-10-01, laptop mode against the real Empire tenant, backend at 1fde2a3; test-strategy § Microsoft Graph, L-100): signed in through Microsoft as CRM01, whose IT Admin role held no email.manage, GET /oauth2/authorization/microsoft-mailbox answered 403 FORBIDDEN in the problem envelope, with no redirect to Microsoft. With email.manage then granted to IT Admin in the local database, the Connect button’s real consent landed on /email/mailbox?connect=success: an ACTIVE connection with its refresh token stored encrypted, and mailbox.connected audited as USER. Not walked live: the key removed mid-consent, which MailboxConnectFlowIT proves, and which the §B F5-k register row keeps for dev. The first Connect attempt took the stub path and answered 404, because the laptop’s forge-et-fe/.env.local had lost VITE_ENTRA_LAPTOP=true. That was local configuration, not this fix. Pre-merge (2026-10-01): a fourth merge of origin/main (8463d58) brought forge-et-be#128, F5-k’s last-admin lock order. It merged with no conflict and shares no file with this fix. A third /forge-review-pr pass with backend-reviewer on that head found no Blocker, no Important and no Nit, and no interaction with the F5-k change. scripts/local-build-evidence.py on 8463d58: BUILD SUCCESSFUL, 3991 tests, 0 failures, 0 errors, 2 skipped.
Lesson: L-100 — the stub connect was gated and the real consent path was not, which is that lesson’s “the stub and the real path enforce the same rules”. The fix follows its other two points as well: the live walk on the real Entra path, and a comment on each test that reaches a step by a shortcut.
BUG-021 — The account list treats team scope as own, while opening a single account walks the reporting tree
Affects: F1-a · Found in: the 2026-09-24 access-model audit of forge-et-be main · Severity: medium · Phase: 1
Repro: give a manager’s role account.view at team; give one of their reports an account; call GET /api/v1/accounts as the manager, then GET /api/v1/accounts/{id} for the report’s account.
Expected: the list and the single read agree — the report’s account is in the list and opens.
Actual: the list omits it and the single read returns it. AccountRepository.visibleTo maps OWN and TEAM alike to owner_id = caller, while DefaultPermissionChecker.can walks the reporting subtree through TeamResolver. F1-a recorded the list side as an interim until F5-f (FR-O3, Q-11): narrower than the record check, never wider, so it leaks nothing — but the two disagree, and AD #35 (2026-09-24) requires one rule for both. No screen lists accounts yet (F1-b is not built), so no user meets it today.
Fix: F5-f (the ownership and sharing substrate) builds the one scope-and-share visibility rule in auth.access and moves both the account list and the account checks onto it, with a test that the list and the single read agree for own, team (shares included) and all, and that unowned records appear only at all.
Interim fix, decided 2026-09-25 (Biseni), ahead of F5-f: F5-f is an unassigned draft, and F1-b (the account list screen) would make the mismatch visible, so the list is brought into agreement with the single read now, without sharing, which does not exist yet. (1) Expose the reporting-subtree lookup through the public auth.access surface: a small interface implemented by the existing TeamResolver.subtreeOf, so account still imports nothing from auth.service. The single-record check’s own path, including withinTeam’s null-owner behaviour, is not changed. (2) AccountRepository.visibleTo resolves TEAM as owner_id IN (the caller's subtree, the caller included); OWN and ALL are unchanged; the contact list follows because it composes the account predicate, and account-less contacts stay visible only at all. (3) The local seed’s ('sales_manager', 'account.view', …) grant returns from all to team, matching recommended-roles.md again (WI-3.1 D13’s stated expiry). (4) Regression test: a manager at team lists a direct report’s and an indirect report’s accounts but not a peer’s, and the list and GET /accounts/{id} agree for own, team and all. It fails on the pre-fix tree. F5-f still adds the share arm and may replace the mechanism; this fix is the list half of its “one rule” for the case with no shares.
Resolved: fixed in forge-et-be#61. It adds auth.access.ReportingTree (implemented by TeamResolver) and changes AccountRepository.visibleTo and ContactRepository.visibleTo / findVisibleById so that team is owner_id IN (subtree + caller). DefaultPermissionChecker and withinTeam are untouched (empty diff). The local seed grant is back at team, with an idempotent upgrade UPDATE for laptops seeded before the fix. Regression: AccountControllerIT’s BUG-021 rows fail on the pre-fix tree (EXIT=1, 3 failed, all team: the list returned [20] while the single read opened [20, 21, 22], and the contact list was []) and pass after (34/34 at the final head 5c18552, including a team holder with no reports). ./gradlew check EXIT=0, 1,556 tests, 0 failures. Reviewed by /forge-review-pr on be#61: no Blockers. Its Important (a context test in RoleFixtureIT pinning that exactly one ReportingTree bean exists and that it is TeamResolver, so the list and the checker cannot diverge) and four Nits are applied in the same PR (5c18552): @FunctionalInterface dropped, one tree walk per ContactService request, an expiry note on the seed’s upgrade UPDATE, and a leaf team holder row. Approved F1-a documents that say “fixed by F5-f” (F1-a spec FR-O3/Q-11 text, the WI-1.2 plan) are left as the record of the plan at the time; the F5-f draft spec is updated to inherit the team arm.
BUG-020 — “Mailbox settings” on /email leads to a screen identical to the one it was clicked from
Affects: F20-a · Found in: F20-a (manual QA, local profile) · Severity: low · Phase: 1
Repro: sign in as any seeded user with no mailbox connected, open /email, and click the “Mailbox settings” link beneath the card.
Expected: a link into a sub-view arrives somewhere that offers something the origin did not.
Actual: /email/mailbox in the not-connected state renders the same NotConnectedPrompt card, with the same icon, the same sentence and the same “Connect Microsoft 365 mailbox” button. The only visible differences are the heading (Email → Mailbox), a ← Email back link, and the absence of the link that was just clicked. There are no settings there to reach. The screen reads as though the link deleted itself and nothing else happened.
Two things sharpen it: the two routes each run their own useMailboxConnection() with no shared cache (a deliberate scope decision — WI-1.2 plan Decision 8, lead-chosen option C), so the navigation refetches and flashes a loading skeleton between two identical cards; and the link is a bare anchor at the bottom of the screen with no grouping or label, so it reads as stray even before it is clicked.
Why it shipped this way: it is BUG-015’s fix. That bug was that /email/mailbox had no entry point anywhere in the app, and the fix restored one in the most contained way available. No acceptance criterion describes how a user reaches the sub-view — AC-18…AC-21 describe the screen’s four states and nothing else — so there was no stated requirement to design against, and design-system.md §8 still carries “Settings/mailbox connect — design at spec time” unresolved. The browser suite navigates by URL, so it cannot see this either.
Fix: render the link in the ACTIVE and RECONNECT_REQUIRED branches only, and not in the not-connected branch. In those two states the sub-view genuinely holds what /email does not — the status line, Disconnect, and the reconnect reason — so the link leads somewhere. In the not-connected state /email’s own Connect button already does the whole job, and the sub-view adds nothing. This keeps BUG-015 fixed on its own terms: the Email section still offers a way into its sub-view wherever that sub-view has content.
Deliberately not done here: moving the entry point to a settings affordance beside the Email heading. That is a design-system decision (§8’s unresolved line), not one feature’s, and F20-b replaces /email’s body with the inbox — so the placement question is better answered with the screen it will actually sit on. Recorded as an input to the F20-b spec, which is in review at the time of writing.
Note on BUG-015: its Fix line records the link as rendering “in the connected and not-connected branches”. This bug narrows that to connected and reconnect-required. BUG-015 stays fixed; the amendment is recorded on its own entry below so the two records do not disagree.
No spec revision: no F20-a acceptance criterion mentions the link, and AC-21 — “/email renders the same not-connected prompt component” — is still met, because the prompt is untouched. The change is to what renders beside it.
Regression: the not-connected row in EmailRoute.test.tsx flips from “offers it” to “does NOT offer it” and fails on the pre-fix tree (1 failed / 9 passed; 10 passed after), anchored on mailbox-not-connected so it cannot pass against an empty tree. The two connected rows are untouched and are what keep BUG-015 pinned. In the browser tier, connect-screen.spec.ts’s entry-point row walked the trip in the one state the link was removed from, so it now connects first and walks the trip BUG-015 was actually about — reaching Disconnect — and a new paired row asserts the link’s absence in the not-connected state. That row also lost its AC-18/ title prefix: AC-18 is the not-connected criterion, which the row no longer asserts, and AC-18 keeps its own dedicated row, so the prefix would have left a row claiming a criterion it had stopped proving. Found by /forge-review-pr on forge-et-fe#44 and fixed in the same PR.
Resolved: fixed in forge-et-fe#44. Browser-validated before the PR opened (local profile, backend :8080, dev server :5174, signed in as sales.rep.local@empireteas.example): /email with no connection renders the Connect card and no link; connecting brings the link back with href="/email/mailbox"; following it lands on the connected panel with the status line, the identity lines and Disconnect — a destination that genuinely differs from the origin. Gates: pnpm typecheck, pnpm test (64 files / 595 tests), pnpm check, pnpm check:boundaries, pnpm check:testids all EXIT=0.
BUG-019 — Mailbox meta text fails WCAG AA contrast on the connected panel
Affects: F20-a · Found in: F20-a (WI-3.1’s axe scan, first authoritative run) · Severity: medium · Phase: 1
Repro: connect a mailbox, open /email/mailbox, and run an axe scan over the connected (or reconnect-required) state.
Expected: no colour-contrast violation — spec NFR-6 / AC-28 require the automated axe scan to pass.
Actual: serious color-contrast on the address and display name. ConnectedPanel.module.css .meta used --et-text-muted, #8c7b66 on #ffffff = 4.09:1, at --et-type-meta-size 11.5px — normal size for WCAG, so 4.5:1 is required. Measured in a browser on the rendered card, not read off the palette.
Why nothing earlier caught it: forge-et-fe has no Vitest-side axe (@axe-core/playwright is Playwright-only, and a lead decision declined adding one), so WI-1.2 shipped under the accessible-mechanism assertions that stand in for a scan — dialog accessible name, status-indicator name matching its visible label, role="alert". Every one of those passes with a failing contrast ratio. The substitution was the right call and is recorded in the Decomposition Plan; this is its known blind spot, and Wave 3 is where it closes.
Fix: .meta uses --et-text-secondary (#6e5f4e, 6.16:1). One rule, scoped deliberately.
Resolved: fixed in forge-et-fe#41 — for this screen only. Merge order matters, and the intended order is #41 first: the same commit is also merged into the Wave-3 branch so forge-et-fe#42’s axe rows could run green, so if #42 lands first, close #41 as superseded and change this line and fix_pr to forge-et-fe#42 — otherwise this record cites a PR that never landed. Raised by /forge-review-pr on #42 (finding I-3).
The wider half — also fixed, in forge-et-fe#43. The question was whether --et-text-muted is usable at all on white at body size. Answer taken: the token keeps its value and its uses were corrected, because darkening it would change the icons too and retuning a palette colour is not a misuse correction. The audit found 11 real uses (the other grep hits are the three definitions and two prose lines): 5 moved to --et-text-secondary (AiProvenance .meta + .link, Combobox .optionDetail + .empty, Tabs .tab), 6 left as non-text where the 3:1 band applies and 4.09:1 clears it, and all 3 definitions re-commented so the band is unmissable at the point of declaration. A source-scan guard (src/test/mutedTextBand.test.ts) now fails if any rule pairs a body-size type token with muted, with the band computed from the type tokens rather than from copied numbers.
Two things that came out of the audit and are NOT closed here:
- Design-system §5 contradicts §2. §5’s Tabs row says “inactive =
text/muted”; §2 says never body-size meaning-bearing copy. The code followed §2, so §5’s Tabs row should be amended — a documentation change that alters no code. Owner: the design-system holder. Closed 2026-09-27 by design-system Rev 27, which amends the Tabs row and two more rows with the same problem that the audit did not name: §3’stype/micro(10.5px column headers and badges) and §5’s Table header. --et-text-faintis worse — about 2.7:1 — on placeholder text and the empty-state glyph. Raised with Dinil 2026-09-23 and deliberately not filed as a bug: a placeholder is meant to recede behind the entered value, so darkening it is a design trade-off rather than a defect to correct. Recorded here so the next person finds the reasoning instead of rediscovering the gap. The measurements, if it is ever picked up: entered text sits at 13.8:1 on the field, the placeholder at 2.7:1, and AA wants 4.5:1 — so a compliant placeholder would still be far lighter than a typed value.
BUG-018 — Re-consenting creates a second Graph subscription and orphans the first
Affects: F20-a · Found in: F20-a (pre-merge /forge-review-pr on forge-et-be#39) · Severity: medium · Phase: 1
Repro: connect a mailbox (a subscription is created and its handle stored), then re-consent — the F21-b / F22 scope-widening path. MailboxConnectionService.connectOrUpdate calls attemptSubscription unconditionally, and MailboxConnection.connect(...) does not touch the subscription columns.
Expected: a re-consent either reuses the live subscription or deletes it before creating its replacement.
Actual: a second subscription is created at Graph, and recordSubscription(...) then overwrites subscription_id and client_state with the new one’s.
Consequences: the previous subscription stays live at Graph with no stored handle, so nothing can ever delete it — not even a later disconnect — and it lapses only on its own expiry. Until then it keeps delivering duplicate notifications carrying a client_state that no longer matches the stored value, which F20-b’s receiver check will reject.
Why it is not in scope for the PR that found it: pre-existing from forge-et-be#37, contradicted by no written AC (FR-9 describes subscription creation, not supersession), and the fix is a genuine design choice — delete-then-create, or reuse the existing handle. Widening a two-defect fix PR to carry it is the scope creep that produced the original review gap.
Fix: not yet chosen. Decide delete-then-create versus reuse, then make attemptSubscription conditional on the connection not already holding a live subscription. Unreachable on the stub path today (LocalMailboxProvider mints a fake handle and no real subscription exists), so the regression needs the Graph adapter or a faked provider asserting the call count across two consents.
Resolved: fixed in forge-et-be#43 by delete-then-create (Dinil’s decision, 2026-09-23). attemptSubscription now supersedes the previous subscription before creating its replacement, using the access token the consent just produced rather than spending a refresh-token rotation for one already in hand. Reuse was rejected because it needs the liveness check and expiry extension that are F20-b’s renewal machinery.
The delete is best effort, deliberately — if it fails the replacement is still created, so the worst case is the pre-fix behaviour (one orphan) and a failed cleanup never costs someone a working mailbox.
Consequence F20-b depends on: “one connection → at most one live subscription” is not an invariant. That session was told directly, so its receiver treats an unattributable client_state as a benign, expected event rather than an attack, and its renewal job works from a single stored handle that is always current while an orphan it cannot see may exist at Graph until it lapses. A hard guarantee would need a reconciliation pass that lists subscriptions at Graph and deletes ones the CRM does not know — F20-b’s design decision, not this fix’s.
Accepted risk — the row lock is held across the Graph calls (post-merge review of forge-et-be#43, lead decision 2026-09-28). connectOrUpdate runs in one transaction that holds the connection’s row lock, and attemptSubscription’s delete and create are Graph calls made inside it. Moving them after commit would reopen the race this fix closes: two concurrent consents could each create a subscription. The hold is bounded by GraphMailboxProvider’s timeouts (5 seconds to connect, 15 to read, per call; two calls at most), and a consent is a rare, user-driven event at this team’s size, so the lock stays. Revisit if connects start timing out behind one another, or if the renewal job starts contending for the same rows.
BUG-017 — TokenRefreshResult.Success prints both tokens through the record default toString()
Affects: F20-a · Found in: F20-a (post-mortem /forge-review-pr on forge-et-be#37) · Severity: medium · Phase: 1
Repro: interpolate a TokenRefreshResult.Success into any log or diagnostic string — log.debug("refresh: {}", result) — and the access token and the rotated refresh token are both printed in full.
Expected: no token material renders, matching MailboxTokenService.CachedToken and TokenCipher.EncryptedToken, which both carry redacting overrides for exactly this reason (NFR-1, AC-26).
Actual: the record default prints every component.
Why it survived review: nothing in Wave 1 logs this type, so MailboxTokenSecrecyIT’s log-and-body capture has no path to drive. Success is the carrier a provider adapter constructs, so the exposure is code that did not exist yet — WI-2.1’s Graph adapter was being written when this was found.
Fix: a toString() override redacting both token components while keeping expiresIn and grantedScopes legible, plus TokenRefreshResultSpec asserting it directly and under string interpolation.
Resolved: fixed in forge-et-be#39 (with BUG-016).
BUG-016 — Re-consenting keeps serving the access token cached under the previous consent
Affects: F20-a · Found in: F20-a (post-mortem /forge-review-pr on forge-et-be#37) · Severity: high · Phase: 1
Repro: connect a mailbox, request an access token (which caches it), then re-consent — the path F21-b and F22 use to widen the grant to calendar and contacts. Request a token again.
Expected: the new consent supersedes any token cached under the old one, so the next request reaches the provider and returns a token carrying the new scopes.
Actual: MailboxConnectionService.connectOrUpdate wrote the new grant but never cleared MailboxTokenService’s cache, so the previous consent’s narrower access token kept being served until it expired — up to an hour.
Why nothing else caught it: granted_scopes is updated by the re-consent, so MailboxScopes.missingFrom finds nothing missing and no refresh is triggered; and the provider’s refusal of a scope the cached token never carried arrives as an authorization error, not a refresh failure, so it is not a state transition the token service reacts to. The stub ignores tokens entirely, so the Wave-1 suite is blind to it.
Impact: the one flow this breaks is the flow re-consent exists for. The user reconnects, is told it worked, and the added capability silently does not work.
Fix: tokens.invalidate(connectionId) after the flush in connectOrUpdate, plus MailboxReconsentCacheIT, which counts provider calls rather than comparing token strings — a cache hit and a cache miss differ only in whether the provider was asked.
Resolved: fixed in forge-et-be#39 (with BUG-017).
BUG-015 — /email/mailbox has no entry point anywhere in the app
Affects: F20-a · Found in: F20-a (post-mortem /forge-review-pr on forge-et-fe#37) · Severity: medium · Phase: 1
Repro: sign in with a connected mailbox and try to reach Disconnect, or with a connection in RECONNECT_REQUIRED and try to reach the reason. There is no route to either short of typing /email/mailbox.
Expected: the Email section offers a way into its own sub-view.
Actual: navConfig carries a single /email entry, the sub-view is not in the sidebar, and /email did not link to it. BackLink led out of the screen; nothing led in. The only other arrival is Wave 2’s ?connect= redirect.
Why it shipped: no acceptance criterion says “there is a link” — AC-18…AC-21 describe the screen’s four states, not how a user reaches it. Wave 3’s browser suite navigates by URL, so it would not have caught it either.
Fix: a “Mailbox settings” link on /email in the connected and not-connected branches, asserted by role and destination rather than by test id. Deliberately absent while loading or after a read failure, where the target would fail the same way.
Resolved: fixed in forge-et-fe#40.
Browser validation discharged 2026-09-23 (it was owed since the fix merged — .claude/rules/tracker.md requires a UI fix to be seen running, and the unit rows prove role, name and destination, not rendering). Walked on the local profile, backend :8080, dev server :5174, signed in as sales.rep.local@empireteas.example: from /email in the connected state the “Mailbox settings” link renders with href="/email/mailbox", and following it lands on the connected panel showing Connected, the mailbox address, the display name, the connected date and the Disconnect button — which is the specific journey this bug was about, since Disconnect was what a user could not reach. Done in the course of validating BUG-020, whose fix narrows the same link; the browser row connect-screen.spec.ts › “BUG-015: a user reaches the connect screen by clicking Mailbox settings on /email” now automates this same trip.
Amended by BUG-020: the Fix line above reads “in the connected and not-connected branches”. BUG-020 narrows that to the ACTIVE and RECONNECT_REQUIRED branches — in the not-connected state the link’s destination is indistinguishable from its origin, so it reads as a dead affordance. This bug stays fixed: the Email section still offers a way into its sub-view wherever the sub-view has anything to offer. Read the two entries together; this one is not rewritten, because it records what shipped.
BUG-013 — Flaky tests: the size matrix’s pricing-import states fail intermittently in WebKit under a full gate: a click on a visible, stable control waits out the test timeout (Windows)
Windows only: seen only on Windows 11 and not reproduced on macOS (see Not reproduced on macOS below). Fix it and prove the fix on a Windows machine; a macOS run proves nothing here.
First filed on F20-c’s branch as BUG-194; renumbered 2026-10-13 to the lowest free number (.claude/rules/id-numbering.md rule 2).
Affects: F26-c’s weekly import screen, as the size matrix arranges it (forge-et-fe e2e/browser/size-matrix/arrange.ts, the pricing-import-* states). Found in: pnpm e2e:gate on F20-c’s frontend wave head af04c43, Windows 11, 2026-10-13. Severity: low. One size-matrix state fails now and then, which fails the whole gate; the product is not affected. Phase: 1.
Repro: a full pnpm e2e:gate on Windows 11 with the default workers, with the backend on the local profile.
Expected: each arrange step’s click lands once its control is visible, enabled and stable, at every size.
Actual: two states failed in the gate run, in WebKit only:
pricing-import-create-unknownat 768 × 1024: the click on the sale option2026 / sale 9602 · …;pricing-import-new-sale-refusedat 1024 × 768: the click on Add sale.
In each case Playwright reported the element “visible, enabled and stable” and then waited out the 30 s test timeout. In the same run the dev server logged WebKit failing to load modules (“Failed to fetch dynamically imported module …/LoginRoute.tsx”, “Importing a module script failed”), which points to a load-related stall rather than a layout fault. Run alone, both states passed at both WebKit sizes three times each (--repeat-each=3, 6 of 6). F20-c changes nothing on this screen.
Fix: not decided. First find whether the stall is WebKit’s module loading under load (a pnpm dev server’s on-demand transforms), or the popover or list rendering under load. Then either warm the routes before the size matrix (as routes.setup.ts does for Chromium, BUG-079) or settle the screen before the click.
Regression test: both states pass in WebKit at both sizes in a full gate run, three runs in a row.
Resolved: 2026-10-30 — fixed by BUG-218’s fix, forge-et-fe#365. The stall was WebKit’s module loading under load, the first option above: each test’s fresh browser context fetched and compiled the dev server’s some 270 source modules. The suite now runs against a development build served by vite preview (some 70 files), and Windows starts at most 6 browsers. In BUG-218’s three full gates in a row (--full on ece00c5a, a full ./gradlew build alongside), both pricing-import-* states passed in WebKit at both sizes in all three, with no size-matrix test flaky.
Lesson: L-248.
Not reproduced on macOS (2026-10-14, Tharinda’s claim released): both pricing-import-* states passed in BUG-183’s run on forge-et-fe main (0ea8055): both WebKit sizes, --repeat-each=10 --workers=8, 869 of 869. The full gate on forge-et-fe#256 (d018924, 4 workers) also passed all 3696 tests. The bug stays open for a Windows machine, where it was seen.
BUG-012 — A filled record place whose filler renders nothing still draws its wrapper, adding space
First filed on F20-c’s branches as BUG-192; renumbered 2026-10-13 to the lowest free number (.claude/rules/id-numbering.md rule 2).
Affects: the record places of forge-et-fe:
src/features/leads/LeadPanelPlace.tsx, from F11-f;src/features/accounts/ContactActionsPlace.tsx, from F1-d;src/features/accounts/AccountPanelPlace.tsx, from F5-g.
Found in: F20-c Wave 1, in its WI-1.5 author dispatch (a reading) and confirmed by the orchestrator’s reading at the design pass on 2026-10-13. Severity: low (spacing only). Phase: 1.
Repro: sign in as a user whose role lacks email.manage (the seeded sales.view.test.local@empireteas.example or lead.view.test.local@empireteas.example), on forge-et-fe feature/F20-c-wave-1:
- open a lead, or a converted lead as any user;
- open an account whose contact has no email address.
F20-c’s fillers (LeadEmailPanel, AccountComposeAction, ContactComposeAction) return null in those cases, by design (F20-c FR-E1 to FR-E3).
Expected: each place component promises that an empty place renders “no wrapper and no space”:
- F11-f FR-S1 (“no heading, no empty box and no space”);
- F1-d FR-S1.
Actual: each place component treats a place as empty only when the composition gives it no filler. When a filler is given but renders nothing, the wrapper (leads-panel-compose, account-contacts-contact-actions, account-detail-place-actions) still renders with no content. In a flex column or row it takes a gap: space/5 in the lead view’s column, space/2 in a row. F20-c is the first feature whose fillers render nothing on condition, so it is the first to expose this. e2e/browser/accounts/account-detail-access.spec.ts:414 pins the empty contact group as present (toHaveCount(1)) and holding no button.
Fix: an empty wrapper takes no space, and stays in the page (.place:empty { display: none; } or equivalent in each place’s stylesheet). This keeps the AC-18 test’s assertions true, with a component test per place. Made in its own commit on F20-c’s wave branch, under the judgment rule for a small, test-backed defect in code outside the wave.
Resolved: 2026-10-13 — fixed in forge-et-fe#249 (174f98e), merged as a248849a.
- Each place’s stylesheet has
.place:empty { display: none; };AccountPanelPlacealso has.inline:empty. - Each place has a component test asserting the wrapper stays in the DOM, is empty, and is hidden by its rule. These tests failed before the CSS change (3 of 37).
account-detail-access.spec.ts:414still holds, and the final gate on3e98ae7passed.
Lesson: a place component’s “empty renders nothing” promise must cover a filler that renders null, not only a missing filler. F20-c was the first feature whose fillers render conditionally. One-off for now: no other place component follows this pattern.
BUG-011 — The shared page check’s axe scan hangs on a page holding a sandboxed frame without scripts
First filed on F20-c’s branches as BUG-182, a number main had meanwhile given to another bug, then renumbered to BUG-191, and on 2026-10-13 to the lowest free number (.claude/rules/id-numbering.md rule 2). The fix commit 2c502f8’s message still says fix(BUG-182).
Affects: the browser page check (forge-et-fe e2e/support/pageCheck.ts, scanAxe), used by every size-matrix state and expectNoAxeViolations · Found in: F20-c Wave 1’s browser run (Stage 7 Step 5e), 2026-10-13 · Severity: medium (a page state with such a frame cannot pass the browser gate) · Phase: 1
Repro: on forge-et-fe feature/F20-c-wave-1 with the backend running, pnpm exec playwright test e2e/browser/size-matrix --grep "email-message " — the HTML message view holds F20-c WI-1.3’s SafeMessageBody frame (iframe srcdoc, sandbox="allow-same-origin allow-popups allow-popups-to-escape-sandbox", no allow-scripts, by design).
Expected: the page check scans the page and reports its violations.
Actual: AxeBuilder.analyze() tries to run axe inside the frame, which refuses script, and stalls until the test’s 30 s timeout (browserContext.newPage: Target page, context or browser has been closed). Every state over that view fails at every size.
Fix: scanAxe excludes iframe[sandbox]:not([sandbox~="allow-scripts"]) from the axe context, so axe does not enter a frame that cannot run it. The rest of the page, the frame element itself included, is still scanned. Such a frame holds a sender’s HTML, not CRM interface, so nothing the CRM controls is left unscanned. Made in its own commit on F20-c’s wave branch (forge-et-fe 2c502f8), under the overnight judgment rule for a small, test-backed defect in code outside the wave. axe drops the frame element together with its content, so frame-title on it is asserted by F20-c’s own spec instead. pageCheck.ts is byte-shared with forge-et-pwa (scripts/shared-with-fe.json), whose copy must be re-synced once the wave merges (.forge/owed.md).
Resolved: 2026-10-13 — fixed in forge-et-fe#249 (2c502f8), merged as a248849a.
- Before the fix, the size matrix’s
email-messagestates stalled at every size; the run had 43 failures, these together with the target-size defect. - After it, the size matrix passed 2269 of 2269, and the final gate passed 3691.
- The
forge-et-pwacopy ofpageCheck.tsis still owed (.forge/owed.md).
Lesson: axe cannot enter a frame that refuses scripts, so a page check must leave such frames out of its scan and assert separately what it skipped (the frame’s title). One-off: SafeMessageBody is the only such frame.
BUG-010 — Top-bar search: the “Search” placeholder sits above the search glyph
Affects: F5-b · Found in: F5-b (manual QA, local profile) · Severity: low · Phase: 1
Repro: sign in; look at the inert global-search control in the top bar. Compare the vertical centre of the word “Search” against the vertical centre of the magnifier glyph beside it.
Expected: the placeholder’s visible text and the glyph read as sitting on the same line.
Actual: the text sits noticeably higher than the glyph.
Root cause: .search uses align-items: center, which centres each child’s box, not its
ink. “Search” has no descender, so its letters ride high in an 18.75px line box that still reserves
the font’s leading below the baseline — the placeholder reads as sitting above the glyph.
Fix (structural, no pixel nudge): text-box-trim: trim-both + text-box-edge: cap alphabetic
on .searchPlaceholder. This trims the font’s leading so the box becomes the cap-height-to-baseline
letters (~9.1px, down from the 18.75px line box), and align-items: center then centres the letters
themselves rather than the leading-padded line box. No magic value; it holds if the type scale
changes; and it uses no pixel literal, so TopBar.module.css’s token-only header (NFR-1/AC-14) is
satisfied outright. Guarded by @supports (text-box-trim: trim-both): a browser without it (Firefox,
as of writing) keeps plain line-box centring — a graceful sub-pixel offset, not a break.
(The first pass used transform: translateY(calc(var(--et-space-1) / 2)) — a 1px nudge. It was
replaced because a pixel nudge fights the line box and read slightly low against the glyph’s circle:
the optical sweet spot is sub-pixel, so no clean value existed. text-box-trim fixes the cause
instead of the symptom.)
Accepted trade-off: the magnifier’s handle weights the glyph’s ink low, so its circle sits
~0.67px above the icon’s box centre. No text-side fix can put the text exactly on the circle without
also nudging the shared Icon, so that residual is the glyph’s own geometry and is left alone — it
is below perceptual threshold at real size.
Regression test: forge-et-fe → e2e/browser/shell/frame.spec.ts, the BUG-010 case beside the
BUG-004/BUG-006 geometry checks. It asserts (1) the trim is applied (computed text-box-trim is
trim-both), (2) it took effect — the placeholder box height collapsed below its font-size (~9.1px
vs the 18.75px untrimmed line box), and (3) the trimmed letters centre on the glyph. Fails-before
(untrimmed: text-box-trim: none, box 18.75px) / passes-after. The browser tier is not PR-gated
(no lane runs pnpm e2e — test-strategy), so this guard runs in the orchestrator’s browser run;
here it was also run live against the running stack (frame.spec.ts 6/6 pass).
Resolved: fix + regression guard in forge-et-fe#36 (fix/BUG-010-topbar-search-alignment).
Validated live in a browser against the running stack; pnpm typecheck (src + e2e/), pnpm test
(56 files / 499 tests) and the e2e frame suite (6/6) all pass.
BUG-009 — Current-user chip shows the role key, not a display name; dummy users named after their roles
Affects: F5-a, F5-b · Found in: F5-b (manual QA, local profile) · Severity: low · Phase: 1
Repro: sign in; read the sidebar-footer current-user chip’s two lines.
Expected: line 1 a person’s name, line 2 the role’s human display label (“System superadmin”, “Sales Manager”).
Actual: line 2 shows the role KEY (superadmin, sales_manager); line 1 shows the role’s name (so for the superadmin account both lines are effectively “System superadmin”).
Root cause: /me returns role = the role’s stable key by design (Role.java key/name separation, AD #26) but exposed no display label for the UI; and the local seed set each dummy user’s display_name to their role’s name, so the person and the role collapse to the same words.
Fix: (1) /me additionally returns roleName (the role’s name) alongside role (the key) — additive, key untouched (forge-et-be#34); (2) the local seed gives the dummy users realistic person names distinct from their roles; (3) the FE shows roleName in the chip (follow-up FE change). Regression: MeEndpointIT seeds a role whose name ≠ key and asserts /me returns both distinctly (fails before — the field didn’t exist).
Resolved: 2026-09-21 — fixed and merged as forge-et-be#34 (0901d83) + forge-et-fe#35 (851687d). /me returns roleName (the role display label) alongside the stable role key — additive, key untouched (AD #26); the local seed gives the dummy users realistic person names distinct from their roles (none matching a real teammate); the FE chip renders roleName. The seeded system role’s own display was polished from “System superadmin” to “Super Admin” in the same round (forge-et-be#35, 9e5b62e, migration V12). Regression fails-before/passes-after: MeEndpointIT seeds a role whose name ≠ key and asserts /me returns both distinctly (the field did not exist before); the FE SessionProvider/CurrentUserMenu tests assert the chip shows roleName and the raw key never leaks. /forge-review-pr + backend-reviewer/frontend-reviewer clean (safe nits folded — additive roleName, assert key ABSENT not just label present, /me→session mapping asserted, /me contract doc synced); verified live (chip “Ravindu Jayasuriya” / “Super Admin”). Prod migrates only the superadmin role + bootstraps one admin — the dummy users/roles are local-seed-only (R__local_seed.sql), confirmed.
BUG-008 — Current-user menu: chevron points the wrong way, the dropdown overhangs the chip, and Sign out reads plain
Affects: F5-b · Found in: F5-b (manual QA, local profile) · Severity: low · Phase: 1
Repro: sign in; look at the current-user chip and open its menu.
Expected: the disclosure chevron points toward where the menu opens; the dropdown lines up with the chip below it; the single Sign out reads as an intentional control; the collapsed hover is clean.
Actual: the chevron was a fixed down arrow while the menu opens upward; the content-sized dropdown was narrower than the wide chip and overhung it; the Sign out item was a plain text row; and the collapsed (icon-only) hover drew an odd rounded rectangle with notches above/below the avatar.
Fix: single chevron up at rest, flipping to down when open; the chip is full-width (opt-in fullWidth on the shared DropdownMenu, default unchanged) so chip and menu align exactly; a LogOut glyph on Sign out; a faint token-derived resting tint on the chip; and a clean circular collapsed hover halo. Regression: e2e asserts the dropdown width == chip width (aligned) and the chevron transform flips rotate(180°)→rotate(0°) on open. (Renumbered from an in-flight BUG-007 that collided with the merged session-expiry BUG-007.)
Resolved: 2026-09-21 — fixed and merged as forge-et-fe#34 (4be0bc7). Chevron is a single glyph, up at rest and flipping to down when the (upward-opening) menu opens; the chip is full-width via an opt-in fullWidth on the shared DropdownMenu (default content-sized, unchanged) so chip and menu align exactly; a LogOut glyph on Sign out; a faint token-derived resting tint (color-mix on --et-surface-hover); a clean circular collapsed hover halo. Verified live in-browser (computed geometry: dropdown width == chip width; chevron transform rotate(180°)→rotate(0°) on open; collapsed halo circular and centred). /forge-review-pr 34 + frontend-reviewer clean (safe nits folded — opt-in fullWidth rather than changing the shared .triggerReset default, chevron addressed by data-testid not a DOM walk). The collision that forced the renumber is recorded as L-057.
BUG-007 — Session expiry is silent: the login page cannot tell an expired session from a first visit
Affects: F5-a · Found in: F5-a (lead-requested manual QA) · Severity: low · Phase: 1
Repro: local profile, signed in via the stub. Navigate to a non-default section (e.g. /users), delete only the SESSION cookie in devtools (leave XSRF-TOKEN), reload. The app re-bootstraps, GET /api/v1/me returns 401, and the gate redirects to /login — with nothing on screen saying why. Confirmed indistinguishable from a first visit by opening /login in a private window side by side: identical.
Expected: FR-6’s W-2 resolution (approved 2026-09-17): “preserve and restore the originally-attempted route; signal the expiry by routing to the login page with a ‘your session expired’ notice.” The route half works — re-authenticating lands the user back on the preserved route.
Actual: no notice, on either expiry path. A deliberate sign-out is also silent, which is correct and is a constraint on the fix, not a second defect.
Root cause (two parts). (1) The reason never reaches the login page. Since BUG-005, SessionState’s unauthenticated arm carries reason?: 'signout' and AuthenticatedGate reads it — but only to drop state.from on a voluntary sign-out. Every involuntary transition still sets a bare {status: 'unauthenticated'}: the cold-boot /me 401 (SessionProvider lines 51 and 84) and the global onUnauthorized handler (line 72) are indistinguishable from each other, so “expired” and “never signed in” collapse into one state. LoginRoute reads exactly one signal off the URL — ?login=failed, the Entra-refusal case (FR-8) — and nothing off router state except the preserved path. (2) On a reload the store starts empty, so an in-memory reason cannot survive regardless; BUG-005 records this same limitation for 'signout' (“a full page reload always resets to a reason-less state”). Distinguishing expiry across a reload needs something deliberately persisted, or a distinguishing signal on the 401 envelope.
Which path a user actually hits — only the reload one. Every authenticated screen in the build today (Accounts, Leads, Pricing, Users, Roles) is a placeholder that fetches nothing, and LandingRoute’s “Check again” button is unreachable for all eight seeded users (each holds at least one nav key). The only calls after sign-in are the bootstrap /me, logout and the stub login, so an expired session goes unnoticed while clicking around and surfaces only on reload. A fix that handles only the in-memory onUnauthorized path would therefore fix nothing a user can observe.
Why it shipped green: AC-8 does not encode the notice — it asserts clearing the session, preserving the route and landing back on it, all of which work. Implementation, review and the browser suite all checked against AC-8 and all passed; the FR clause had no acceptance criterion to hang on, so nothing downstream could see it. This is not an implementation miss. (The general form — an FR clause with no matching AC clause is invisible to every gate — is a candidate lessons entry, raised separately and not yet ruled on.)
Fix: extend the mechanism BUG-005 already built rather than adding a second one — a new reason: 'expired' on the unauthenticated arm, set where a 401 ends a session that was authenticated, with LoginRoute rendering one notice for that reason only. Because an in-memory reason does not survive the reload that is the only observable path, it needs backing by either (a) a persisted marker — record “this browser had a session” on successful authentication, clear it in signOut(), and on a cold-boot 401 treat marker-present as expiry; purely frontend, no contract change; or (b) a backend signal distinguishing “session presented and invalid” from “no session at all” on the 401 envelope — cleaner semantically, but it changes GET /api/v1/me, which the spec names a coordinated seam with FS-3’s permission guard (a coordinated revision of both artifacts, not a refactor). Constraint either way: a deliberate sign-out must stay silent, including BUG-005’s accepted case where reason persists across client-side navigation until the next authenticated refresh(). W-2 also owes the promotion of this behaviour to a cross-feature session-expiry convention in design-system.md now that F5-b has landed — do it in the same change.
Accepted limitation (BUG-007, mirroring BUG-005’s precedent). The prior-session marker is per-browser and persists across /login reloads: it is set when a session is established and cleared only on a voluntary sign-out, never on a 401. So after an involuntary expiry the “your session expired” notice re-shows on every subsequent /login reload until the user signs in or out; and on a shared browser a genuine first visit by a new user (after a prior user closed the tab without signing out) shows the notice once. Both are harmless — the marker can only ever add an informational notice, never grant access, change routing, or weaken auth enforcement — and clearing the marker on the 401 path was rejected because the two setState sites that both fire on a cold-boot 401 would then race to clear-and-read. This is the same shape as BUG-005’s accepted reason-persists-across-navigation limitation.
Regression proof (mandatory, fails-before/passes-after): a component spec asserting that an expired session renders the notice and that a deliberate signOut() does not; plus an assertion added to the existing browser expiry spec (e2e/browser/auth/signin-session-signout.spec.ts, “preserves the attempted route across a forced session expiry”), which today performs this exact repro and never looks for a message.
Resolved: 2026-09-21 — fixed and merged as forge-et-fe#33 (ea0e923); the cross-feature convention (design-system §9 Rev 21), the F5-a AC-8 notice clause (spec Rev 2) and the accepted-limitation note above landed in forge-et-harness#151 (7bbd087). The login page now shows a polite role="status" “Your session expired” notice for an involuntary expiry — carried across the only observable path (a reload) by a persisted per-browser marker set on /me 200 and cleared on sign-out — while a first visit and a voluntary sign-out stay silent and ?login=failed takes precedence. Regression fails-before/passes-after: the LoginRoute notice component test (fail-before demonstrated with the render disabled), the SessionProvider reason/marker tests across all three paths, and the browser assertions in signin-session-signout (Scenario B shows the notice, Scenario C stays silent). Adversarial frontend-reviewer pass: pass-with-nits, safe nits applied.
BUG-006 — Top-bar global-search affordance is undersized and cramped
Affects: F5-b · Found in: F5-b (manual testing, local profile) · Severity: low · Phase: 1
Repro: sign in; look at the top bar’s global-search control.
Expected: a search field that reads as the primary global-search affordance — a comfortable width spanning a good portion of the top bar (design-system §6 / the CRM prototype), the glyph and the “Search” placeholder comfortably spaced.
Actual: the control shrinks to its own content (~92px) and sits as a small pill; the glyph reads as crowding the label. .search in TopBar.module.css sets no width and no flex-grow, so it collapses to content width.
Fix: let .search grow within the top bar up to a capped maximum (flex: 1 + a new --et-layout-search-max token — token-only per AC-14), so it fills the available run to a sensible maximum instead of hugging its content. The control stays inert until F31 (FR-9/AC-9).
Resolved: 2026-09-21 — fixed and merged as forge-et-fe#31 (cda5ef6). .search grows to a capped --et-layout-search-max (420px, mirrored into both token layers); verified 420px in-browser (was ~92px). /forge-review-pr 31 pass-with-nits (safe fixes applied); browser 47 + unit 486 green.
BUG-005 — Explicit sign-out preserves the deep-link route, so re-login lands on the last page instead of the default section
Affects: F5-b · Found in: F5-b (manual testing, local profile) · Severity: medium · Phase: 1
Repro: sign in (lands on the first accessible section, e.g. Pricing); navigate to another section, e.g. Users; open the current-user menu → Sign out; sign in again → you land back on Users.
Expected: an explicit sign-out starts a fresh session, so re-login lands on the default (first accessible) section (lead decision, 2026-09-21). Deep-link return on an involuntary unauthenticated redirect — a typed protected URL reached while logged out, or a mid-session 401 expiry — is unchanged.
Actual: re-login returns to the page you signed out from. AuthenticatedGate stamps state.from = <current location> on every unauthenticated redirect, including the one an explicit sign-out triggers; LoginRoute then honours that preserved path after re-auth (preservedPath(location.state) ?? '/').
Root cause: the gate cannot distinguish a voluntary sign-out from an involuntary expiry — both surface as status: 'unauthenticated', so the route-preservation intended for deep-linking also fires on sign-out.
Fix: SessionProvider.signOut() sets the unauthenticated state with reason: 'signout' (a new optional field on the unauthenticated arm of SessionState); AuthenticatedGate omits state.from when reason === 'signout', so LoginRoute falls back to / → the first accessible section via LandingRoute. The 401/expiry paths leave reason unset, so they still preserve the attempted route (the migrated signin-session-signout Scenario B — land-back after a forced 401 expiry — stays green). Regression: a new test signs out from a non-default route and asserts re-login lands on the default, failing before the fix and passing after.
Accepted limitation (raised in /forge-review-pr 31 by frontend-reviewer): reason: 'signout' is an in-memory flag that persists across client-side navigations until the next authenticated refresh(). So after a voluntary sign-out, pressing browser Back to an earlier protected page (still in history) and then re-authenticating also lands on the default section rather than that page. This is consistent with the lead decision (“any explicit sign-out starts a fresh session → default landing”) and never weakens auth enforcement — the flag can only ever drop a preserved route, never grant access — so it is accepted rather than made one-shot. A full page reload always resets to a reason-less state, so genuine cold deep-links (a typed/emailed protected URL opened while logged out) still land back on the target.
Resolved: 2026-09-21 — fixed and merged as forge-et-fe#31 (cda5ef6). signOut() sets reason:'signout' and AuthenticatedGate drops state.from, so re-login lands on the default section; the involuntary 401-expiry path still preserves the route. Regression fails-before/passes-after (the gate unit test asserts from:none on voluntary sign-out vs the retained from:/accounts on expiry; e2e Scenario C lands on Pricing). The accepted Back-navigation limitation is recorded above.
BUG-004 — Sidebar brand-header divider does not align with the top-bar divider
Affects: F5-b · Found in: F5-b (manual testing, local profile) · Severity: low · Phase: 1
Repro: sign in; compare the 1px rule under the sidebar brand header with the 1px rule under the top bar.
Expected: the two horizontal rules sit on the same line — the sidebar brand header and the top bar are the same height, and the logo is vertically centred within the header.
Actual: the brand header is 73px tall while the top bar is --et-layout-topbar-height (64px), so the sidebar rule sits ~9px lower; the logo also reads slightly low within the taller header.
Fix: set .brand to block-size: var(--et-layout-topbar-height) with box-sizing: border-box and vertically-centred content (drop the block padding, keep the inline padding), so the header matches the top-bar height, the two 1px rules align, and the logo centres.
Resolved: 2026-09-21 — fixed and merged as forge-et-fe#31 (cda5ef6). Brand header height-matched to the top bar; verified in-browser (brand-header bottom 64px == top-bar bottom 64px, was 73 vs 64). e2e asserts the two rules share a bottom edge.
BUG-003 — Collapsed (icon-only) sidebar: logo and nav icons are not centred and the current-user block is clipped
Affects: F5-b · Found in: F5-b (manual testing, local profile) · Severity: medium · Phase: 1
Repro: sign in; collapse the sidebar to icon-only (the top-bar toggle at desktop, or a laptop-width viewport).
Expected: in icon-only mode the logo and each nav icon are horizontally centred in the narrow rail, and the current-user footer shows a centred avatar only.
Actual: the logo sits ~5px left of centre (centre 33px vs rail centre 38px) and each nav glyph ~4px left (centre 34px vs 38px); the current-user block keeps its full expanded width (avatar + name + role ≈ 162px) inside the 76px rail, so the name is clipped by the rail’s overflow:hidden.
Root cause: the icon-mode CSS hides the labels/wordmark with max-inline-size:0 but leaves the 9px flex gap beside the now-zero-width element, which shifts the visible glyph off-centre; and the current-user footer has no icon-mode treatment at all, so it renders at full expanded width in the narrow rail.
Fix: in icon mode set gap: 0 on .brand and .navLink (centres the logo and glyphs once the zero-width label no longer holds a gap); pass a collapsed flag from AppLayout (mode === 'icon') into CurrentUserMenu so it renders a centred avatar only — name and role hidden — when collapsed. Regression: browser-level assertions that in icon mode the nav glyph centre equals the rail centre and the current-user block does not overflow the rail.
Resolved: 2026-09-21 — fixed and merged as forge-et-fe#31 (cda5ef6). Icon-mode gap:0 centres the logo and nav glyphs; the collapsed footer is centred avatar-only (name/role hidden), so nothing is clipped. Verified in-browser: nav-glyph centre and avatar centre both equal the rail centre (offset 0; was ~4–8px left), user block within the rail. Review also caught the avatar being 8px off-centre — fixed by centring the icon-mode footer, with the e2e strengthened to assert it. e2e + unit green.
BUG-002 — managing_director (and any long-principal user) cannot sign in — session principal_name overflow
Affects: F5-a · Found in: F5-b (manual-testing setup, local profile) · Severity: high · Phase: 1
Repro: local profile (docker compose up -d + ./gradlew bootRun, seeded DB). Prime CSRF (GET /api/v1/me → 401), then POST /api/v1/auth/stub-login with X-XSRF-TOKEN and {"email":"managing.director.local@empireteas.example"} → HTTP 500. The other seven seeded users (superadmin, administrator, executive, sales_rep, sales_manager, taster, tea_room) return 204 and sign in fine.
Expected: every seeded (and real) user signs in — stub-login → 204 and GET /api/v1/me → 200.
Actual: 500. Backend log: DataIntegrityViolationException: ... UPDATE spring_session ... ERROR: value too long for type character varying(100).
Root cause: com.empireteas.auth.access.AuthenticatedUser is a record (long id, String email, String displayName) implementing only Serializable — it exposes no principal name. Both login paths build the UsernamePasswordAuthenticationToken with this record as its principal — StubLoginService and the shared IdentityResolver.principal(), which the real Entra path also uses. AbstractAuthenticationToken.getName() therefore falls through to principal.toString(), and Spring Session JDBC persists that string in spring_session.principal_name VARCHAR(100) (V7 migration). For managing_director the toString() is AuthenticatedUser[id=8, email=managing.director.local@empireteas.example, displayName=Managing Director] = 104 chars → overflow → 500. Shorter representations stay under 100, which is why only managing_director fails today.
Scope — stub-path-only (corrected during /forge-review-pr; the initial “Entra path is affected too” reading was wrong): the observed 500 is on the stub login path (UsernamePasswordAuthenticationToken with AuthenticatedUser as principal). The Entra path builds an OAuth2AuthenticationToken whose principal is AuthenticatedOidcUser extends DefaultOidcUser with nameAttributeKey="sub", so its principal_name is the Entra sub claim (a bounded ~44-char GUID) — never at overflow risk. Both paths build the same AuthenticatedUser, but that is not what governs principal_name, which is where the initial reading erred. The reason to still touch the Entra path is consistency, not overflow: fixing only the stub path leaves principal_name = app-user id on stub but = sub on cloud, a latent trap for any future session-lookup-by-principal feature.
Fix (two parts): (1) make AuthenticatedUser implement org.springframework.security.core.AuthenticatedPrincipal with getName() returning the id (Long.toString(id)) — a long is ≤ 19 chars, so it can never overflow principal_name, unlike the email/display-name which are TEXT/unbounded (the bug prompt’s email-first suggestion is refined to id here); this fixes the stub path. (2) override getName() on AuthenticatedOidcUser to return the same app-user id, so the Entra path persists the app-user id too rather than the sub — principal_name is then the app-user id on both paths (symmetric; safe because DefaultOidcUser’s OIDC machinery reads the token via getClaims()/getAttributes()/getIdToken(), not getName(), so login/logout/token handling are unaffected). (Widening the VARCHAR(100) column is the weaker alternative — it only raises the ceiling and leaves an opaque toString; not taken.)
Regression proof (mandatory, fails-before/passes-after — verified): LongPrincipalLoginIT (container) signs a long-email user in through the stub path → stub-login 204 + /me 200 (verified to 500 before the fix, green after); AuthenticatedUserSpec asserts AuthenticatedUser.getName() == the id (including a Long.MAX_VALUE id → 19 chars) independent of email/display-name; AuthenticatedOidcUserSpec asserts the Entra principal’s getName() == the app-user id, not the sub (the symmetric half). ./gradlew check green.
Resolved: 2026-09-21 — fixed and merged as forge-et-be#33 (c7c6557; commits 8804843 + c5a41b0). Both AuthenticatedUser and AuthenticatedOidcUser expose the app-user id as getName(), so principal_name is the bounded app-user id on both login paths. /forge-review-pr + backend-reviewer clean (approve-with-fixes → all applied); regression verified fails-before/passes-after; ./gradlew check green.
BUG-001 — Capture-app service worker: the predicate-drift guard its own comment claims does not exist
Affects: FS-1 · Found in: FS-2 (FS-1 plan-review calibration run) · Severity: medium · Phase: — (no delivery phase active)
Repro: in forge-et-pwa, add an extension to SHELL_EXTENSIONS in
src/sw/shouldCache.ts (e.g. .avif) and leave public/sw.js untouched. Run
pnpm test (green), pnpm build (green — scripts/validate-pwa.mjs passes).
Nothing reports that the tested predicate and the shipped predicate now disagree.
Expected: the build fails, or at minimum warns. public/sw.js lines 5–7 state
that the cache-scope decision is duplicated there “because a classic service worker
cannot import from the bundle — the two must stay in step, which is what
scripts/validate-pwa.mjs asserts on every build.”
Actual: validate-pwa.mjs asserts three unrelated things about dist/sw.js — that
the string /api/ still appears, that __APP_VERSION__ was substituted, and that the
cache name is versioned. It never compares the two predicates. The claim in the
shipped comment is false.
Not a behavioural defect today: the two predicates were diffed at filing and agree exactly — same guards, same order, same nine extensions. What is broken is the safeguard, and the comment asserting it exists is the dangerous half: it tells the next reader the drift is already covered. This is the vacuous-gate class FS-1 was built to avoid, in FS-1 itself — the unit test stays green while the shipped worker walks away from what it tests.
Fix: extract API_PATH_PREFIX and SHELL_EXTENSIONS from both files and assert
they are identical, plus assert the shipped worker still carries each structural guard
(non-GET, cross-origin, API prefix, extension-less document). State in the same place
exactly what the check does not prove — semantic equivalence of the two
implementations — so the replacement comment does not overclaim the way the original
did. Regression proof: plant the .avif drift above, watch the build fail, revert.
Resolved: 2026-08-23 — fixed and merged as forge-et-pwa#8 (commit c67e248).
validate-pwa.mjs now extracts API_PATH_PREFIX and SHELL_EXTENSIONS from both
shouldCache.ts and public/sw.js and asserts they are identical (failing if either
declaration goes missing), plus the four structural guards, and documents what it does
not prove. Regression demonstrated manually (plant the .avif drift → build fails →
revert). Follow-up, not blocking: an automated test feeding a drifted fixture pair to
validate-pwa.mjs and asserting a non-zero exit would make the guard permanent rather
than build-time-manual. The fuller fix belongs to F36, when the capture work reworks
the worker anyway: build public/sw.js as a Vite bundle entry so it imports the tested
shouldCache predicate directly — that deletes the duplication, this drift guard, and the
awkward cross-origin branch in one move. Recorded on the F36 row in features.md.