Owed, with owners
The engagement’s open follow-ups: each item is something the delivery team owes Empire, a teammate
or the harness, with the person who owes it. It was the **Owed, with owners.** list in
.claude/CLAUDE.md § Active Context, moved here so it is not loaded into every session.
Same rules as Active Context: current state only. An item that is done is removed here, and where
its history matters, a dated line goes to .forge/engagement-log.md. A row is removed only when
its work is done, checked against its source (the code, the PR, the AWS account, forge-et-iaac’s
log, the tracker), or when the lead marks it won’t-do or asks for it to go. Being tracked somewhere
else as well (a signals OQ, the cloud-gated register, environments-and-identity.md §4, a Trello
card, a design record) is not a reason to remove it. A defect may move to bugs.md and
tracker.yaml, and two rows may merge when nothing is lost. Why: rows removed while still open
were lost from the one list the team walks (the lead, 2026-10-16). At session start, reconcile
this file against the injected gate state together with .forge/active-context.md (.claude/CLAUDE.md § Boundaries).
-
Decide the capture app’s install card on desktop and after Not Now (Nanndish; raised 2026-11-20 by a read of
forge-et-pwasrc/lib/install/andsrc/features/settings/InstallCard.tsxagainst each platform, made while fixing BUG-526, BUG-527 and BUG-524). Three gaps, none a fault in the code as specified: (1) on Windows and macOS in Chrome or Edge the card says Add it to your home screen, which is a phone’s wording — the browser installs it as a desktop app; (2) Not Now hides the card for good in that browser (localStorageet.capture.installDismissed), with no way to bring it back, which is the likeliest reason a Windows tester no longer sees it (the others: the app is already installed in that browser, or Install was pressed and the browser’s dialog cancelled, which drops the offer until the page is reloaded); (3) Safari on macOS 14 and later can install the app (File › Add to Dock) but fires no install event, so no card shows there. iOS (one line, In Safari, tap Share, then Add to Home Screen) and Android Chrome (the browser’s own dialog) are correct as built; the iOS line is worth checking on an iPhone on iOS 26, whose Safari can keep Share behind its ⋯ menu. The deployed dev app’s manifest, icons and service worker all answer200, so nothing stops the offer there. No device was tested in this session: real installs on each platform are FR-CG1 item (g) of the cloud-gated register. Done when Nanndish decides which of (1) to (3) to change and the change merges, or records that they stay. -
Decide whether Claude may make a run sheet’s browser steps in the browser pane (Tharinda, as the lead; raised 2026-11-20 by F43’s dev run,
.forge/specs/F43-ai-quota-backlog-and-busy-answer-findings.mdF-002)./forge-deliverinvariant 4 and the run-sheet template give every browser step of a manual work item to the person, because a Claude session could not use the person’s signed-in browser. In the Claude Code browser pane it can: for F43-WI-5.1, Dinil signed in and watched, and Claude made the burst and the timed presses as page scripts (the record says so). If the lead allows it, a follow-up harness PR changes invariant 4 andforge-plan-author’s run-sheet case to say when Claude may make browser steps the person watches, and that the record names who made each. Done when the lead decides, and, on a yes, that PR merges. -
Move each laptop to Node 26 and pnpm 12 (Dinil, Biseni, Nanndish, each on their own machine; raised 2026-11-19 by the dependency upgrade the lead asked for): the app repos move from Node 24 and pnpm 10.15.0 to Node 26 and pnpm 12.22.0 (
.nvmrc,engines,packageManager), and pnpm’s project settings move from.npmrcand package.json’spnpmfield topnpm-workspace.yaml. Node 25 and later ship no Corepack, so each machine installs it from npm once. The session-start hook prints the fix commands on any machine whose Node major differs from.nvmrc(.claude/lib/toolchain-check.sh), and each app README’s steps 1 and 2 have the detail: install Node 26 with fnm,npm install -g corepack@0.36.0,corepack enable pnpm, thencorepack pnpm install --frozen-lockfilein every app checkout and worktree. Playwright 1.65.1 also needs a newer Chromium build: runcorepack pnpm exec playwright install chromium webkitonce in either frontend (the browser cache is shared). Done when each developer’s session start prints no toolchain warning. -
Redraw the capture app’s ✕ and Dismiss icons in
pwa-currentafter the shadcn move (Nanndish; raised 2026-11-21 by altrium-io/forge-et-pwa#201’s frontend review, Important 6): the shadcn ghost icon Button draws its icon inforegroundand fillsaccenton hover, where the older Button drewtext/secondaryandsurface/hover(.forge/design/ui/pwa-shadcn.mdD-3). That covers the sheet’s ✕ and the Ready to review notice’s Dismiss (row 4, screenshotsphone/30–30e,32,32aanddesktop/02-notifications.png), Sign Out?'s ✕ (row 3,phone/83,84), and the search box’s Clear search ✕ on Leads and Plan’s exhibitor list (row 6, altrium-io/forge-et-pwa#203:phone/11,13, and the Plan list screenshots that show a search). Row 7b (the lead drawer and its sheets) adds two more: the convert dialog’s link box is now a drawn box inprimary, not the platform’s check box (phone/25,25a,25d), and the country select’s invalid edge isdestructive, as the Input’s is, where it was the grey border (thelead-edit-details-server-errorsstate). Row 8 (the Plan campaign screens) adds the campaign sheet’s ✕, each campaign row’s Edit and Archive icons and Archive Campaign?‘s ✕ (phone/70,77,79), and widens the campaign form’s First Day and Last Day side padding from 9px to 12px, the Input’s, so they line up with Name (phone/75,75a,76,78; accepted by Nanndish on 2026-11-22). Row 9 (Plan’s exhibitors and the upload) adds the ✕ of Sort and Filter, Add Exhibitors and Check the Exhibitor List (phone/52,53,56,60,61), a Sort By radio drawn as a circle with aprimarydot where it was the platform’s radio inaccent/gold(phone/52,53), and Check the Exhibitor List’s ✕ shown unavailable while the import runs, as Cancel already was. Row 10b (the booth drawer) adds the drawer’s ✕ and ⋯ and its sheets’ ✕, and its More Actions menu: items fillaccenton hover and focus where they filledsurface/hover, and the menu fades in (the booth screenshots,phone/62onwards). Row 11 (the toasts, now Sonner) adds the toast’s motion (it slides up from its own height and fades in over about 0.4 s, and slides and fades out when it leaves, where it rose 6px and vanished), a swipe down to dismiss, the toast as a Tab stop with the focus ring (Alt+T reaches it), and the customer check’s Try Again as the shadcn outline Button. Redraw the icons in.forge/design/prototypes/pwa-current/and re-take those screenshots withcapture.mjs, or record in design-system § The capture app that the ✕ follows the shadcn look. Done when the prototype and the screenshots match the built app, or the design system records the choice. -
Move the lead page’s last tone-to-Alert helpers onto
ToneAlert(Nanndish; raised 2026-11-22 by the delta review of altrium-io/forge-et-pwa#211): row 10a addedToneAlerttoforge-et-pwasrc/ui/shadcn/alert.tsxand moved four identical helpers onto it.src/features/leads/lead/EditDetailsSheet.tsx’sMessageAlertis a fifth copy, andLeadActions.tsxandLeadWriteOutcome.tsxbuild the same Alert withliveworked out at the call site. Use<ToneAlert tone text live />in all three, in the next PR that touches them. Done when none builds the Alert fromalertForToneby hand. -
Remove the capture app’s native-dialog
closelistener (Nanndish; raised 2026-11-22 by row 10b, altrium-io/forge-et-pwa’s booth drawer move): since row 10b the app renders no native<dialog>, sosrc/features/notifications/useNotificationFeed.ts’sclose-event listener (which re-checked the held Ready to review notice when a native dialog closed) and the tests that dispatchcloseno longer do anything. Remove them, and check the held notice is still shown when a shadcn Dialog or Sheet closes. Done when nothing listens for or dispatchesclose. -
Correct the toast comment in the shared
tokens.css(Nanndish; raised 2026-11-22 by row 11 of the capture app’s shadcn move):src/ui/styles/tokens.css, a byte copy offorge-et-fe’s held bycheck:shared, still says the toasts are top-layer popovers. Since row 11 the capture app’s toasts are Sonner’s, which are not, and the CRM’s are Sonner’s too. Change the comment inforge-et-fefirst, then re-copy it toforge-et-pwaand runpnpm check:shared --update. Done when both copies say what the toasts are. -
Give the capture app a budget on everything it loads at startup (Nanndish; raised 2026-11-22 by row 11 of the capture app’s shadcn move,
.forge/design/ui/pwa-shadcn.mdD-6):forge-et-pwa’scheck:bundlemeasures the entry chunk alone, so a build that moves code into a chunk the entry preloads passes while the startup cost grows. Row 11 lowered the entry from 429.01 to 416.31 kB while the startup total grew from 523.76 to 558.66 kB raw (166.5 to 175.5 kB gzipped). Add a budget on the entry plus the chunksindex.htmlpreloads, raw and gzipped, tobundle-budget.jsonandcheck:bundle, set from the current figures. Done when a PR that grows the startup total past it failscheck:bundle. -
Decide whether the toast after Edit Details waits for the lead drawer to close (Nanndish; raised 2026-11-22 by the row 7b frontend review, Nit 8): after a save,
toastWhenClosed(LeadDrawer.tsx) shows Changes saved. while the drawer stays open. The toast is drawn above the drawer but sits in the app root, which the drawer makesinert, so its Dismiss cannot be pressed and a tap on it may reach the drawer underneath. A screen reader hears the drawer’s own status line. The older native-<dialog>drawer made the app root inert too, so this is probably not new. Check on a phone-size browser whether a tap on the toast reaches a drawer control; if it does, choose between holding the toast until the drawer closes and leaving it. Done when the choice is made and, if it changes the code, shipped. -
Move the email browser specs’ copied arranger helpers into one module (Tharinda; raised 2026-11-20 by forge-et-fe#595’s frontend review, Nit 9):
arrangerFor/arrangerCsrfnow have six copies,withConnectedUser/resetStubfour,attachLedgerfive andfillStubSignIntwo, acrosse2e/browser/email/*-smoke.spec.ts,live-inbox-e2e.spec.ts,e2e/browser/mailbox/auto-connect-smoke.spec.tsandtracking-scheduling-smoke.spec.ts. F20-i WI-6.1’s Decision 2 kept its wave from editing shared files. Oneforge-et-fePR addse2e/browser/email/arranger.tsand moves every caller to it. Done when that PR merges andgit grep -c 'function arrangerFor' origin/main -- e2ecounts one file. -
Replace the token-endpoint client’s
ResponseErrorHandlerbefore moving to Spring Framework 8 (Tharinda, at the Boot upgrade that brings Framework 8; raised 2026-11-20 by the review of forge-et-be#488): Framework 7.1 deprecates the wholeResponseErrorHandlerinterface for removal.forge-et-be’sTokenEndpointClientsstill hands Spring Security’sOAuth2ErrorResponseErrorHandlertoRestClientthrough it, as Spring Security 7.2 itself does. The compiler reports nothing yet, because the method called is not itself deprecated. When Spring Security ships a replacement, or the upgrade to Framework 8 removes the interface, the client moves to it. Done when noforge-et-becode usesOAuth2ErrorResponseErrorHandleror any otherResponseErrorHandler. -
Drop
--no-turbo-fast-api-callsfrom the capture app’spnpm devonce Node fixes the abort (Tharinda; 2026-11-19, found by forge-et-pwa#172’s selected browser run): on Node 26.13.1 the gate’s Vite dev server aborted with “Lazy deopt after a fast API call with return value is unsupported” while it sent a large module, and every later test failed. forge-et-pwa#172 starts Vite throughnode --no-turbo-fast-api-calls(package.jsondev; README step 4 says why). The Node fix, nodejs/node#67536, was open on 2026-11-19. When a Node 26 release carries it, setdevback tovite, raiseengines.nodeto at least that release, and run the capture app gate. The CRM’s gate serves a built bundle throughvite previewand did not abort; if itspnpm devaborts the same way on a laptop, giveforge-et-fe’sdevscript the same flag. -
Upgrade Terraform and the AWS provider in
forge-et-iaac(Shyamin Ayesh, the DevOps team; held on 2026-11-19 by the lead’s decision during the dependency upgrade): the workflows pin Terraform 1.12.1 (1.17.2 is out) and the overlays’ lockfiles holdhashicorp/aws6.69.1 (6.73.0 is out);forge-et-iaachas no Dependabot configuration. A merge todevrunsterraform applyon the dev account, so the DevOps team chooses when. Done when a PR todevmoves both, with its/forge-review-prverdict, or the DevOps team decides to stay. -
The hand-over meeting with Empire: what to tell and what to ask (Tharinda; the lead’s decisions of 2026-11-19 at the
owed.mdtriage). The Phase 1 hand-over and UAT drop 1 are one meeting, about Mon 30 Nov at Empire’s office: the demonstrations and UAT first, then Phase 2 scope if time allows (the row “The Phase 2 scope conversation with Empire”). Nothing is sent automatically; the lead or a teammate tells Empire by hand, in the meeting, by email or by WhatsApp. Each row named below stays open until its point has been told or answered, and is then removed. Points drafted for the lead:- To tell (one message or part of the meeting): (a) the answers to the 10 November asks — the capture app is in Phase 1; a follow-up for a customer email nobody has replied to is built (F44); the ERP link in V1 is the hand-off document only (F19, Phase 2); the wider brand and marketing demonstration is Kavindu’s to arrange (row “Give Empire the answers to its asks from the 10 November progress meeting”); (b) moved to the “Ask Kavindu before drop 1” point below (the lead, 2026-11-19); © an account briefing is prepared when someone asks for it and covers no past orders; (d) the three corrections to 16 October — the models do not learn from Empire’s data, each request carries that task’s own content from Empire’s own AWS account, and synced mail is stored; (e) how email works — only email linked to a lead, contact or account is kept, linked mail is readable by everyone who can see the record, and personal addresses can be blocked (sign-off item A-45); (f) the screens moved from the approved list-beside-detail prototype to tables with full-page records; (g) contact email addresses and phone numbers are plain text, and how each record is edited.
- To ask (one numbered list): how many trade shows a year Empire attends and plans (the cost annexe assumes six); OQ-62, newsletter consent, unsubscribe and sender (decided by the delivery team on 2026-11-19, PRD Rev 103; ask Empire to confirm the choices in Q10); OQ-68, whether reps want an email signature in the CRM; OQ-63, past orders on the account page; Empire IT’s Microsoft 365 maximum send size (the CRM allows 25 MB of files); whether reps send password-protected Office files. With the Phase 2 scope conversation: the Pipedrive questions (OQ-17, 22, 23, 25, 26, 28, 29, OQ-41) and a Pipedrive deals export.
- Ask Kavindu before drop 1, so UAT is set up in time (the lead, 2026-11-19; row “Tell Kavindu how UAT’s data works…”): which data UAT holds — (1) an empty system, (2) made-up records, or (3) Empire’s real Pipedrive data — and who signs in — (a) the test logins
crm01–crm05, whose changed passwords Altrium shares with them, or (b) the testers’ own Microsoft accounts. With own accounts,crm01(the seeded IT Admin) adds Kavindu, and Kavindu adds the others; orcrm01adds them all. Tell him the risks with the question: real Pipedrive data puts real customer addresses beside features that send mail (the price broadcast, scheduled sends, replies to unanswered email); an employee’s own mailbox connects at sign-in, so a reply sent from UAT is a real email to a real customer, tracked links in it stop working once UAT is removed, real mail is copied into UAT, and a mailbox connected to both UAT and production shows its follow-ups twice. The answer sets the acceptance criteria’s “Who signs in” and “With what data” (both left to Empire’s choice in the draft of 19 November). - Not in this meeting: Empire IT’s production asks (row “Empire IT asks”) wait until production is needed; the detailed sales-team presentation is replaced by UAT drop 1 (the lead, 2026-11-19). Done when every point has been told and every question asked.
-
Decide the fixes for the gaps found on the UI walk of dev’s lead pages (Dinil, 2026-11-09; checked against
forge-et-belead/service/LeadRanking*,LeadResearchIdentityandai/prompts/lead-ranking/v1/system.md). G1: a lead tagged Spam by F11-d still shows a priority such as P1 · Hot, because F11-g’s ranking does not read the triage tag (F11-g’s spec says nothing about spam), so the dashboard’s ranked-leads panel (F15-a) could list spam at the top. Recommended: no ranking and no priority shown while a Spam tag stands, ranked again when the tag is dismissed; a Duplicate keeps its rank. An F11-g spec revision, then the code. G2 — intended, Dinil, 2026-11-19: F11-g decision 3 ranks a lead whose research found nothing from the inquiry alone, because a genuine new buyer often has nothing online and F11-d’s spam check is the guard against fake inquiries; at most, Why this priority says research found nothing. G3: research that matched the wrong company raised lead #28 to P1 (research confidence 25%). Research already searches only on the company name or website domain and makes no search without one (F11-c decision 4, FR-A4;LeadResearchIdentity), so #28’s search text came from its company-name field; the fix is in ranking: pass the research confidence to it and treat research below a threshold as nothing found. An F11-g spec revision, then the code. G4: after Dismiss tag the lead page shows “Tag dismissed by, ” (F11-d FR-V2) without naming the tag, and the reason text stays under it with no label; name the tag (“Spam tag dismissed by …”) in TriageTagBlock.tsxand label the reason. An F11-d spec revision, then the code. Done when G1, G3 and G4 each have their revision and fix, or the lead says one is intended. -
Plan Phase 1’s UAT round as two drops, and tell Empire with the first which criteria come in the second (Tharinda; the lead’s plan of 2026-11-18). Drop 1: what is ready is released to UAT for Empire’s testing around Mon 30 Nov. How soon Empire starts and how long it takes is not known; the plan assumes about two weeks, to about Mon 14 Dec, past the SOW’s Phase 1 UAT end of Thu 10 Dec (
.forge/client/project-timeline.md). Drop 2: during that time the delivery team finishes the Phase 1 work still open and releases it to UAT as a new tag. It holds at least F20-f (criterion F3), F20-j (criterion F6) and F44 (criterion F7), which the lead wants in before production, plus any other Phase 1 feature not done by drop 1 (tracker.yaml,delivery_phase: 1).- Tell Empire, with drop 1, the criteria that arrive in drop 2, so their testers do not log
them as defects (
.forge/client/phase-1-acceptance-criteria.md). Say also that drop 2 runs past 10 Dec and alongside Phase 2’s build (target Fri 1 Jan 2027). - Decided by the lead, 2026-11-18: one
uatinstance serves both drops and Phase 2’s round, and is decommissioned after Phase 2 (AD #23 revised; F38-h spec Rev 11). Drop 2 is a new tag deployed onto it withdeploy-uat, so the hand-entered data and the testers’ records stay.
- Tell Empire, with drop 1, the criteria that arrive in drop 2, so their testers do not log
them as defects (
-
Decide how UAT sits between Phase 1’s UAT and Phase 2’s (Shyamin Ayesh decides; Tharinda explains the options to him; raised 2026-11-18). The lead wants UAT paused after Phase 1’s UAT and before production goes live, keeping its database so it can be revived for Phase 2’s round, and leans towards option 1. Every option is carried out by a person, by hand; nothing is triggered by AWS on a date (the lead;
.claude/rules/tracker.md § Release & UAT Lifecycle).- Stop the backend and the frontends. Set UAT’s ECS service to 0 tasks: no mail sync, no
sends, no AI calls, and Graph’s mail subscriptions lapse on their own within days. The CRM and
capture-app frontends are static files on S3 and CloudFront; with the backend stopped they load
but cannot sign in, so also disable both CloudFront distributions so nobody reaches a broken app.
The database keeps running with all data. Revive: enable the distributions, set the service back
to 1 task; testers may need to reconnect their mailboxes. Costs about the database, load balancer
and public IPs while paused (roughly $25–35 a month as an estimate). Do not stop the database
itself: AWS starts a stopped RDS instance again by itself after 7 days. A later
uat.yamlapply undoes this pause:forge-et-iaacbase/ecs/service.tfsetsdesired_count = 1with noignore_changes, andbase/cloudfront/distribution.tfsetsenabled = truewith no variable. So either apply nothing to UAT while it is paused, or add an overlay variable for both (a smallforge-et-iaacchange). - Snapshot, then remove. Take a final database snapshot, then
destroy. Almost no cost while paused. Revive: apply again restoring that snapshot, which needs a smallforge-et-iaacchange to restore from a snapshot, then deploy. More steps; mailboxes reconnect. - Keep it running. Simplest; full cost; testers’ mailboxes stay connected while production goes live (the overlap below).
- Decommission after Phase 1’s UAT. Destroy it and apply a fresh one for Phase 2’s round;
the checklist data is entered again and testers’ records are lost.
Before options 1, 2 or 4, testers disconnect their mailboxes on UAT. That also stops a revived
backend from sending scheduled emails that fell due while it was paused (F20-g FR-L3’s sweep
sends overdue ones; FR-L4’s re-check drops them when the mailbox is disconnected). Record the
choice and each pause and revive in
.forge/design/environments-and-identity.md§4. All four options fit AD #23 as revised 2026-11-18.
- Stop the backend and the frontends. Set UAT’s ECS service to 0 tasks: no mail sync, no
sends, no AI calls, and Graph’s mail subscriptions lapse on their own within days. The CRM and
capture-app frontends are static files on S3 and CloudFront; with the backend stopped they load
but cannot sign in, so also disable both CloudFront distributions so nobody reaches a broken app.
The database keeps running with all data. Revive: enable the distributions, set the service back
to 1 task; testers may need to reconnect their mailboxes. Costs about the database, load balancer
and public IPs while paused (roughly $25–35 a month as an estimate). Do not stop the database
itself: AWS starts a stopped RDS instance again by itself after 7 days. A later
-
Triage what Empire raises after UAT drop 1, then plan when each item is fixed (Tharinda; the lead’s rule of 2026-11-18). Record each issue or change request Empire raises as a UAT issue (
tracker.yamluat_issues:, releaseP1; a defect also goes tobugs.mdthrough the bug flow,.claude/rules/tracker.md § Release & UAT Lifecycle). Then plan how and when each is fixed: in drop 2, in a later UAT release, or as later work. UAT issues usually come before Phase 2 features; change requests are planned case by case (the standing rule:.claude/rules/tracker.md § UAT issue triage + routing, “Planning the fixes”). -
Tell Kavindu how UAT’s data works, and warn Empire’s testers about real mailboxes (Tharinda; raised 2026-11-18). Changed by the lead, 2026-11-19: Kavindu chooses. Before drop 1, ask him which data UAT holds (none, made-up, or real Pipedrive data) and whether testers use the test logins or their own mailboxes, telling him the risks of each; the question is drafted in the row “The hand-over meeting with Empire: what to tell and what to ask”, point “Ask Kavindu before drop 1”. If he chooses real Pipedrive data or real mailboxes, F38-h decision 5 and FR-8 and the UAT data checklist are revised, and real customer data in the dev account is weighed against AD #34’s “never a real person’s email outside production”, which as written covers the seeded administrator. The text below is the earlier position. The lead was unsure whether UAT should hold a clean slate, made-up data or Pipedrive data, because real customer data in UAT risks sending email to real customers while testing. Earlier position (2026-11-18): UAT holds made-up records entered by hand from
.forge/plans/F38-h/uat-data-checklist.md, with onlycrm01@empire.lktocrm05@empire.lkand inventedexample.comaddresses (F38-h spec decision 5, FR-8; AD #34: never a real person’s email outside production). Production gets the real Pipedrive data (the lead, 2026-11-18): the first load at the start of the parallel run and the final run after the data freeze (.forge/active-context.md§ Phase scope). Loading Pipedrive data into UAT would reverse AD #34 and put real customer addresses beside features that send mail: the weekly price broadcast (F37-a), scheduled sends (F20-g) and F44’s replies.- Recommended (Claude, 2026-11-18): under the rule of no questions to Empire before UAT, tell Kavindu the default rather than ask an open question: “UAT uses made-up data and test mailboxes. Your real Pipedrive data goes in on production at the start of the parallel run. Tell us if you would prefer something else.” A change he asks for is then ordinary work.
- The risk the data choice does not remove: Empire’s testers’ own mailboxes. The UAT data
checklist (Part 7) lets the IT Admin add Empire’s testers on UAT with their own Microsoft 365
addresses. Every recommended role but IT Admin holds
email.manage(Part 2), and a role withemail.manageconnects the user’s mailbox automatically at sign-in (PRDMailboxConnection, Rev 74). So a tester’s real mailbox joins UAT the first time they sign in, with no choice to skip it. From then on: the CRM’s inbox shows their real mail; a reply, an AI-drafted reply or a follow-up (F44) sent from UAT is a real email to a real customer; F20-f and F20-j copy real customer mail linked to UAT’s records into UAT’s database, in the dev account; and marking mail read in the CRM marks it read in Outlook. With made-up records few real emails match a record, so the copying is small, but replying from the live inbox is not limited by the data. - Decided by the lead, 2026-11-18: ask Empire before drop 1 whether their testers (the sales
team, Kavindu and management) sign in to UAT with their own Microsoft accounts or with the test
logins
crm01–crm05used now. The lead makes this one question an exception to the rule of no questions to Empire before UAT. With the test logins, Altrium hands out the passwords out of band (environments-and-identity.md§5). With their own accounts, the drop 1 message says plainly that anything they send from UAT is a real email, and asks them to untick Track Opens and Clicks on any email to a real customer (the next two points). - Tracked emails sent from UAT break after UAT is removed. A tracked send (Track Opens and
Clicks is ticked by default, F20-g FR-T3) rewrites each link to go through the backend that sent
it,
https://api.uat.crm.empire.lk/api/v1/email/track/…. Once UAT is paused or destroyed, a real customer who clicks a link in an email sent from UAT reaches no server, and the image no longer loads. Only matters if testers use their own mailboxes and email real customers. - Decided by the lead, 2026-11-18: no code guard; tell the testers and Kavindu the risks clearly. A recipient allow-list on UAT and dev, or a setting that turns tracking off on UAT, was offered and not chosen. The drop 1 message states plainly the risks of connecting their own mailboxes (real emails to real customers, tracked links that break later, real mail read in UAT) and of loading real Pipedrive customer data into UAT.
- The seeded administrator on UAT stays
crm01@empire.lk(AD #34, AD #40: never a real person’s email outside production). If Kavindu wants to test as administrator with his own account,CRM01adds him through the app’s Users screen and gives him a role holding the Users and Roles keys (AD #41); his address is never written in a harness file or a task definition (checklist Part 7). The lead agrees (2026-11-18): share theCRM01sign-in with Kavindu, out of band (environments-and-identity.md§5), so he can add his own account and others.CRM01is also the test mailbox used on dev, so whoever holds its password can sign in there too. Tell him in the drop 1 message, with the sign-in question. - Answered (Claude, 2026-11-18): a reply sent from UAT causes no problem when the same mailbox later connects to production. UAT and production are separate systems with separate databases, and production never reads UAT’s. The CRM sends through Microsoft Graph from the person’s own mailbox, so a reply sent from UAT sits in their Outlook Sent Items like any reply sent from Outlook. When they connect to production, F20-j’s 12-month past-mail scan links it to the matching production record (the real Pipedrive data), once, as it would an Outlook reply. UAT’s copy stays only in UAT’s database and is deleted with it after Phase 2.
- Two environments holding one mailbox at once (Claude, 2026-11-18, from the specs; to be
proved by F38-h WI-5.2’s check R-10, decision 11). The test mailboxes
crm01–crm05can be connected in dev and UAT at once (F38-h Q-3, decision 11); a real person’s mailbox can be connected only in UAT and production, since dev’s users are the test mailboxes (the UAT data checklist andenvironments-and-identity.md§5). In theory it is mostly harmless. Each environment keeps its own sign-in tokens, its own Graph change subscription (each with its own notification address, F20-a FR-9) and its own delta state in its own database, and Microsoft allows many subscriptions on one mailbox. What they share is the mailbox itself: read and unread marks, sent mail, and F20-i’s Empire CRM — All mail search folder, which each environment finds by name and records by id (F20-i D-5). Both would also act on the same incoming mail: each could link it (F20-f), run AI on it, or raise an F44 follow-up, so a tester sees duplicate tasks across the two and Bedrock is called twice. No email is sent unless a person sends it. Pausing UAT between the rounds (row above) removes the overlap until Phase 2’s round, when it returns if testers use their own accounts while production is live; disconnecting their mailboxes on UAT before they connect on production avoids it.
-
Fix the production go-live date against Empire’s first January 2027 trade fair (Tharinda; the lead’s plan of 2026-11-18). After UAT, if Empire will use the system, production is created and released, CRM and capture app together, around late December 2026 or early January 2027, with the real Pipedrive data imported (row above). The capture app moved into Phase 1 so that Empire can use it at its January 2027 trade fairs (
.forge/discovery/meeting-notes/2026-11-10-kavindu-crm-progress-update-demo.md), so production must be live before the first one.- Find the first fair’s date, and which fairs Empire means. No harness document records it as of 2026-11-18. Found on the web on 2026-11-18 (Claude, for Tharinda): Gulfood 2027 is listed for 15–19 March 2027 in Dubai (Dubai World Trade Centre and Expo City Dubai), not January; several exhibitor listings agree and one says the dates are confirmed, though no page from the organiser was among the results (Blueprint Exhibits, ProExpo, Portugal Global). Gulfood 2026 ran in late January, which may be why Empire said January. If Gulfood is the first fair that matters, production must be live by early March 2027, not early January. The search found no other January 2027 fair that is clearly Empire’s. Ask Kavindu which fairs the capture app is for and their dates; this is a schedule fact the go-live date needs, not a spec question.
- Start the production-account conversation now. When the account is created is not the
delivery team’s decision (row “The AI layer under Bedrock’s new-account limits”, item 4: Shyamin
Ayesh, Malinda Wellalage, Thilina Abeyrathna, Rukshan Salgado and Empire Teas). F38-i’s
production waves wait on it (
tracker.yamlF38-i notes). - Risk: Bedrock’s limits are per account, so a new production account starts at new-account quotas at go-live. That can slow the capture app’s card reading and research and the CRM’s lead AI during a fair. The same row’s items 1 and 2 (Haiku 4.5 everywhere, the backend’s pacing) are what keep them working; the earlier the account exists, the more usage history it has by then.
-
Correct the CRM Focus marker’s code comment (Dinil; raised 2026-11-20 by design-system Rev 150):
forge-et-fesrc/ui/components/LeadRankIndicators/LeadFocusMarker.tsxsays the marker’s design-system entry is the lead’s decision and not yet made. Dinil accepted the marker as built on 2026-11-20, and design-system §5 Focus marker (Rev 150) now describes it. Done when the comment points at that entry instead — in whichever PR next changes the file. -
Decide where the Priority panel’s inquiry-only note goes (Dinil; raised 2026-11-19 by F36-c-WI-3.4, plan Decision 7):
crm-currentdraws the Priority panel’s inquiry-only note before Why this priority; the built panel (forge-et-feLeadRankingPanel.tsx) puts it after Recommended Next Action — decide which order is right and change the other. The difference is older than F36-c wave 3 and was not changed by it. Done when the drawing and the built panel show the note in the same place. -
Add indexes for the capture app’s bell poll and To Review count when captured leads grow (Dinil, F36-c’s owner; raised 2026-11-18 by the review of forge-et-be#475, deferred there as fine at today’s volume — net p95 74 ms, W2-27): every user’s 20-second N-1 poll reads all of their
capture_notification_recipientrows through(user_id, read_at)(rows are kept, so they grow) and sorts byn.created_at(CaptureNotificationInboxRepositorywindow), andreadyToReviewCountscanscaptured_leadforstatus = 'to_review'with no index;CaptureTeamQueryscans bycampaign_idwhile row locks are held. The fix: an index on recipient(user_id, notification_id DESC)with the window ordered by id, a partial indexcaptured_lead (id) WHERE status = 'to_review'and one oncaptured_lead (campaign_id), in a new migration. Done when the indexes ship, or when a measurement on UAT data shows they are not needed. -
Decide whether
/forge-deliveropens the harness wave PR as a draft as soon as the wave branch is cut (Dinil, with Tharinda, whose design it would change; raised 2026-11-20 by F36-c’s Reflect,.forge/specs/F36-c-capture-research-notify-findings.md§ Reflect Q3 and F-006, L-311): Dinil decided on 2026-11-20 to make the Reflect’s three framework changes. The other two (forge-plan-authorandplan-reviewer: a change to an element’s tag or contents lists thee2e/files that select it, L-329;/forge-deliverStage 12: the drafts a wave’s plans hold for the orchestrator are written before the harness wave PR is marked ready, F-007) are made by the F36-c follow-up PR of 2026-11-20. This one was held out of it, because it reverses D-4 of.forge/design/harness/ids-on-pushed-branches.md, the lead’s design that replaced the same proposal (from F36-b’s Reflect) with “push a branch once it has taken an id” (.claude/rules/id-numbering.mdrule 1; Stage 0 already pushes the wave branch when it cuts it). That design lists what an early draft PR breaks: every status-checkpoint PR copies the wave branch’s ids and would failnext-id.py --check-pragainst it;/forge-pr-open --waveskips an open PR, so the draft keeps its placeholder title and body; and the manual-wave and app-only paths open no harness PR at that point. An early PR also changes only which branch renumbers on a clash, since in--check-pra branch with no PR gives way to an open PR (D-2). Check first why F-006’s BUG-482 clashed: F36-c’s wave branch took it at 07:44 on 2026-11-19 (76de72b23) and F36-d’s at 13:28 (fb335fd63); ifnext-id.pywas not run for one of them, or F36-c’s branch was not yet pushed, the fix is in how numbers are taken, not in when the PR opens. Done when Dinil and Tharinda keep D-4 (and this row is removed), or change it in a harness PR that also fixes those problems. -
Decide what follows F36-c’s failed
v2ranking runs (Tharinda, the lead, with Dinil as F36-c’s owner; raised 2026-11-20 by.forge/cloud-gated-validation.md§B F36-c item (k)): both livev2sets failed on Haiku 4.5 on dev (figures in the row). The focus never moved a priority the wrong way, but the model ranks many cases one level lower than the sets expect without a focus, so too few focus pairs match; the capture set also had two prose answers (BUG-518), a timeout and one unrelated focus that moved a priority, and the lead set had one subject ranked lower without research. Wave 3 is on dev; the first UAT or production release that carries it is held until a run passes. Choices: change the prompts and run again, revise the sets’ expected priorities where the model’s stricter answer is right (a spec revision of FR-RK6 and F11-g FR-A8), or wait for a stronger model. Done when the lead has decided and item (k) readsvalidated, or the release plan says otherwise. -
Say when a wave that touches only app repos writes its plans’ drafts for the orchestrator (Dinil; raised 2026-11-20 by the fresh-context review of the F36-c follow-up PR):
/forge-deliverStage 12’s new rule writes those drafts before the harness wave PR is marked ready (Stage 12b step 6), but a wave that touches only app repos has no harness wave PR to mark ready — its harness wave branch reachesmainthrough checkpoint PRs — and Stage 0’s line “Stage 12 opens it as the harness wave PR” does not matchforge-pr-open, which opens a harness wave PR only for a manual wave or one with a harness-onlysubWI. Such a wave’s plans can still hold drafts (F36-c WI-1.2’s## Noteshas one). Done when Stage 12 names the step at which such a wave writes its drafts, and Stage 0’s line agrees withforge-pr-open. -
Have
forge-et-be’s local build record run the Fast lane’s script checks (Dinil; raised 2026-11-20 by F36-c’s Reflect, finding F-005):scripts/local-build-evidence.pyruns./gradlew buildbut none of the Fast lane’s Python checks, so forge-et-be#484’s comment-pointer failure (FR-JB7without its document, fixed ine8995fc0) was found by CI after the PR was recorded and ready — a further instance of L-157. The record runsscripts/check-comment-pointers.pyand the other script steps of.github/workflows/pr-validation.yml’s Fast lane first, and refuses to record when one fails. Done when aforge-et-bePR makes it so, or the team decides L-157’s manual step is enough. -
Fold the capture app’s copied seam helpers into
e2e/seam/support/journey.ts(Nanndish; raised 2026-11-18 by the review of forge-et-pwa#162, N7):escapeRegExp,answersandtoastare written again inplan-screens.seam.spec.ts,review-screens.seam.spec.ts,plan-research-screens.seam.spec.tsandcapture-resend.seam.spec.ts, andjourney.tsnow holds them; WI-5.1’s plan (Decision 3) chose not to change those files. Done when each of those files imports them fromjourney.ts— done in whichever PR next changes each file. -
Give the capture app a browser-test feature map, so a selected run selects (Nanndish; raised 2026-11-18 by F36-a’s Reflect, finding F-002):
forge-et-pwahas noe2e/feature-map.json, sopnpm e2e:gate --selectedruns every browser test. forge-et-pwa#168, which changed one unit test’s time limit and a hash file, ran 7,688 browser tests for about 3½ hours on Windows.forge-et-fealready selects its browser tests from its map (e2e/feature-map.json,scripts/browser-gate/select.mjs;.forge/design/harness/fewer-nightly-reds.mdR-1). Done whenforge-et-pwahas a map, and a PR that changes no browser-test input runsnone selected. -
Load the compose dialog once at the top of
aiDraftFlow.test.tsx(Nanndish; raised 2026-11-18 by BUG-449’s triage and the review of forge-et-fe#583, L-116):src/app/aiDraftFlow.test.tsxrenders the realComposeEmailProvider, and its firstfindDialogpays the compose dialog’s firstimport()inside a wait raised to 10 s; BUG-449 measured that load at up to 10.3 s under load. Its AC-25 case also waits 10 s forLeadActionsPlace’s lazy action chunk. BUG-449 moved the same load to a top-levelawait loadComposeEmailLayer()inScheduledEmailsRoute.recover.test.tsx; the other files that mount the provider are already warm or never open the layer (review of forge-et-fe#583). Done whenaiDraftFlow.test.tsxloads both modules at the top and its waits are back to the default — in whichever PR next changes the file. -
Find out why lead 101’s AI triage on dev finished 8.5 minutes after the lead was created (Tharinda to assign; raised 2026-11-17 by F38-g-WI-4.4, F11-d.4,
.forge/plans/F38-g/F38-g-WI-4.4-record.mdR-7): lead 101 (RETEST 2026-11-16 4.4 F11-d Delta) was created at 04:26:49 UTC and never edited (version0); its triage readstate: donewithtag: DUPLICATEwithin a second, but itsanalysedAtis 04:35:14 UTC, when the model’s answer (spam) replaced the tag. Leads 100 and 102, created 15 s either side, finished at once. Read dev’s lead-triage logs andai_invocationrows for lead 101 between those times (a database and log read — not done by the retest, NFR-3) to see whether a throttled Bedrock call was retried or the job was redelivered (AD #13), and whether a second model call was made. -
Decide whether the API exposes the
Locationheader to the cross-site apps (Tharinda to assign; raised 2026-11-17 by F38-g-WI-4.4’s F11-a.1 setup): on dev,POST /api/v1/leadsfrom the CRM’s console answered201butLocationread asnullin the browser, because the CORS answer does not list it inAccess-Control-Expose-Headers(AD #33’s cross-site calls). Check whether any screen of the CRM or the capture app reads it; if none does, record that the body’sidis the contract and adjust the checks that expect it (F11-a.1, F11-a.2). -
Look at the lead page’s “A router only supports one blocker at a time” warnings (Tharinda to assign; raised 2026-11-17 by F38-g-WI-4.4): on dev (CRM
70fc7709), opening/leads/93logged that warning twice in the console, fromLeadPage.tsx:82andLeadLeaveGuard.tsx:89. Check whether two leave guards are registered at once on the lead page, and whether the unsaved-changes prompt still works on leaving the edit form. -
Walk F36-a’s Leads tab and lead drawer by keyboard alone (Nanndish; raised by F36-a-WI-4.2, AC-35, and by the review of forge-et-pwa#147, I9, 2026-11-17): the To Review and Completed tabs, both searches, the outcome chips, Show More Leads, a row into the lead drawer, Convert to Lead and its sheet, Reject with its reason and note, Move Back to Review, and closing the drawer, on
forge-et-pwafeature/F36-a-wave-4(forge-et-pwa#147). Done when the walkthrough is recorded in.forge/plans/F36-a/F36-a-WI-4.2-leads-review-plan.md § Progress, before the F36-a wave 4 PR is marked ready. -
Decide how orphaned prepared photos are cleaned up (Nanndish, as F36-a’s plan author; raised 2026-11-17 by the review of forge-et-be#457, N6): the convert (F36-a WI-4.1) and the card upload copy a card photo to
attachments/<ownerType>/prepared/<uuid>before the transaction (AttachmentService.prepare), and discard it when the transaction rolls back. If the JVM stops between the copy and the commit, or the commit’s outcome is unknown (the copy is then kept on purpose, L-213), the object has no attachment row and nothing removes it: storage only, no wrong data. An S3 lifecycle rule on theprepared/prefix is not the fix, because an attached photo keeps itsprepared/key (preparedKey, no copy at attach), so a rule would delete real attachments; the attachments bucket also keeps every version on purpose (the row above on bucket versions). The options are a timed sweep (EventBridge Scheduler → SQS → worker, AD #12) that deletesprepared/objects older than a day with noattachmentrow naming their key, or accepting the cost as small. Done when one is chosen and, if a sweep, it is a ticket. -
Decide whether a medium booth code may wrap in the booth drawer’s plate (Nanndish; raised by
frontend-revieweron forge-et-pwa#136’s fixdd7f75c, 2026-11-16): to keep the company readable at 320px with a booth code of up to 50 characters at the 24px title size, the booth plate now takes at most half the drawer’s row and its code wraps anywhere. A code of 8 to 10 characters (HALL4-A12,B12345678) therefore breaks across two lines at 320–360px, mid-token when it has no hyphen, though the company has room. The reviewer’s alternative keeps such a code on one line: drop the plate’smax-inline-size: 50%and give.companymin-inline-size: 5em(src/features/plan/booth/BoothHeader.module.css), which keepsbooth.spec.ts’s 320px test passing. Done when the layout is kept as it is or changed that way. -
Check the capture app’s camera states in a WebKit that has a camera API (Nanndish; raised 2026-11-14 by F36-a-WI-2.2): Playwright’s WebKit build on Windows has no
MediaDevices,MediaStreamorcanvas.captureStream, so the six capture camera page states (forge-et-pwae2e/pageStates.ts,engines: ['chromium']) run only in the Chromium size projects; the WebKit ones filter them out (playwright.config.tsgrepInvert,src/app/pageStateEngines.test.ts). Done when someone runs those states in a WebKit that has the API (the lead’s Mac nightly run, if its WebKit build has it, or Safari on a real iPhone over HTTPS, beside register item (e) of the F36-a block), records the result, and either lifts the restriction where the API exists or leaves it with the reason. -
Decide what the capture app says when the badge reader cannot load (Nanndish; raised 2026-11-14 by F36-a-WI-2.2’s pre-PR review, finding I-4): the QR decoder (
jsqr) is a lazily loaded chunk; when its download fails (a dropped connection at a trade show before the service worker has cached it), Badge QR keeps saying Point at the badge’s QR code and reads nothing. WI-2.2 makes the next switch to Badge QR try the download again, but the rep is not told. The design system’s rule is that a failure is never silent. Done when a sentence for that state (for example The badge reader couldn’t load. Check the connection, then switch to Badge QR again.) is drawn inpwa-current/and accepted, then built with its test — or when it is decided that the retry is enough, with the reason. -
Revise F36-a’s NFR-2 and AC-36 for a camera view that holds no control (Nanndish; raised 2026-11-13 by F36-a-WI-2.2’s plan review): NFR-2 and AC-36 ask for a focus ring for “every control on the camera view”, but the screens accepted on 2026-11-12 put Take Photo, Retake, Scan Again and Choose a Photo below the camera panel, and WI-2.2 built them there (design-system Rev 122). The one element over the camera view is the hint chip, whose text
forge-et-pwasrc/ui/styles/themeContrast.test.tsmeasures at 4.5 : 1 or more on its solid backing, with the toast’s Dismiss ring at 3 : 1. Done when the spec’s next revision says NFR-2 holds because no control sits on the camera view and AC-36 names the chip’s text and the toast’s ring as what is measured. -
Walk F36-a’s campaign controls by keyboard alone (Nanndish; raised 2026-11-14 by F36-a-WI-3.3, AC-35): choosing, creating, editing, archiving and unarchiving a campaign in the picker, and View Plan, on
forge-et-pwafeature/F36-a-WI-3.3-campaigns(forge-et-pwa#133) or the wave 3 branch once it is integrated. Done when the walkthrough is recorded in.forge/plans/F36-a/F36-a-WI-3.3-campaigns-plan.md § Progress(item 14), before the F36-a wave 3 PR is marked ready. -
Accept the capture app’s first-download budget of 410 kB, or move the new icons out of
icons.tsx(Nanndish; raised 2026-11-14 by F36-a-WI-3.3): mounting the active-campaign provider inAppLayoutbrought the campaign code into the app’s first download; WI-3.3 moved the one-campaign read into its own file, which tookindexfrom 414.24 kB to 403.15 kB (395.65 kB before WI-3.3), and raisedindex’s budget inforge-et-pwabundle-budget.jsonfrom 400 to 410 kB — a budget its plan let it raise only for the Plan chunk. Done when Nanndish accepts the 410 kB budget, or the new campaign icons move to a chunk of their own and the budget goes back to 400 kB. (Wave 3’sindexis 405.62 kB with WI-3.4 on top; wave 2’s own branch is back to 394.52 kB under the 400 kB budget, since forge-et-pwa#132 loads its send pass on first use — the same trick may bring wave 3 back under 400 kB.) -
Decide whether a toast may cover the last control at the foot of the capture app’s main area (Nanndish; raised 2026-11-15 by forge-et-pwa#132’s delta review of the shell change): the toast sits
space/5above the tab bar, over the last ~70px of<main>; with<main>scrolled to its end, its bottom-most control (for example the refused row a Try Again is about) is under the toast for its 4 s, longer while a pointer rests on it. The old sticky shell covered the same area, so this is not new. Done when Nanndish accepts it, or asks for<main>to gain bottom padding (orscroll-padding-bottom) the toast’s height plusspace/5while a toast shows. A route change resets<main>to its top on Back too, as Rev 125 says; Back keeping the position would be a separate request. -
Draw WI-3.4’s two remaining states in the capture-app prototype (Nanndish; raised 2026-11-15 by F36-a-WI-3.4, its finding F-5): Show More Exhibitors (deviation 8) and the plan regions’ offline and earlier-data lines (deviation 10) are built but not drawn in
.forge/design/prototypes/pwa-current/. Done when both are drawn and their screenshots taken, in the wave 3 harness change. -
Restore the two exhibitor archive-reason enums in F36-a wave 3 (Nanndish; raised 2026-11-15 by BUG-428’s fix, forge-et-be#436): F36-a-WI-3.1 described
ExhibitorResponse.archiveReasonandExhibitorArchiveRequest.reasonin prose becauseNullableRecordComponentConverterdropped any schema with a nullableallowableValuescomponent. Once #436 is onmainand merged intofeature/F36-a-wave-3(forge-et-be#435), put@Schema(allowableValues = …)back on both components and assert the enums,nullincluded, inCapturePlanOpenApiIT, as the F36-a wave 3 contract has them. Done when wave 3’s PR carries both. -
Walk F36-a’s plan list and booth drawer by keyboard alone (Nanndish; raised 2026-11-15 by F36-a-WI-3.4, AC-35): search, the status chips, Sort and Filter, Show More Exhibitors, a card into the booth drawer, the visit note and mark, Edit Details, Archive and Restore, and Add by Hand with Add Another, on
forge-et-pwafeature/F36-a-WI-3.4-plan-list-booth(forge-et-pwa#134) or the wave 3 branch. Done when the walkthrough is recorded in.forge/plans/F36-a/F36-a-WI-3.4-plan-list-booth-plan.md § Progress, before the F36-a wave 3 PR is marked ready. -
Walk F36-a’s exhibitor-list upload by keyboard alone (Nanndish; raised 2026-11-15 by F36-a-WI-3.5’s test-verify, AC-35; found missing from this file by F36-a’s Reflect, 2026-11-18): Upload Exhibitor List from the Add Exhibitors sheet and from the empty Plan panel, Check the Exhibitor List with its column matches, and the result toast, on
forge-et-pwamain. Done when the walkthrough is recorded in.forge/plans/F36-a/F36-a-WI-3.5-list-upload-screens-plan.md § Progress. -
Accept or undo the 500-column cap on an uploaded CSV row (Nanndish; raised 2026-11-15 by F36-a-WI-3.2’s pre-PR review): WI-3.2’s plan (Decision 7, Risk 5) says a CSV has no column cap, but a 5 MB row of one-letter cells is about 2.6 million columns, each listed in the preview’s
columnsand repeated in every sample row, so one upload could answer with a body of tens to hundreds of MB. WI-3.2 (forge-et-be#434, commitf20593f7) refuses a CSV row with a filled cell past its 500th column, with the contract’s existing CSV message — the cap a workbook row already has — and ignores empty cells past it, so trailing commas still read. Response shapes are unchanged. Done when Nanndish accepts the cap (WI-3.2’s plan Decision 7 then gets a revision entry), or asks for it to be its own refusal cause (a wave 3 contract revision) or to be removed. -
Decide whether the capture app’s empty Plan panel says Needs a connection. once or twice when offline (Nanndish; raised 2026-11-15 by F36-a-WI-3.5’s build, its finding F-1): offline, the nothing-exists panel shows the line under each of its two buttons (Upload Exhibitor List and Add by Hand), so it appears twice (forge-et-pwa#135). Done when Nanndish keeps it, or asks for one line under the pair.
-
Correct the frozen F36-a wave 2 contract’s
BadgeRead.contactEmailrow (Nanndish; found 2026-11-13 by F36-a-WI-4.3’s plan):.forge/plans/F36-a/F36-a-Wave-2-contract.mdsays a malformed email is refused at convert, but F11-a’sLeadFieldRuleschecks only the email’s length, and the frozen wave 4 contract agrees that the only intake refusal a test can trigger is a lead with neither company nor contact name. Done when the wave 2 contract gets a## Revisionsentry stating what convert does with a malformed email (and whether F36-a wants it refused, which would be an F11-a change), and WI-2.4’s and WI-4.3’s plans agree with it. The same revision also records three things forge-et-be#432 built that the contract does not state (its review record deferred them to the Decomposition Plan’s## Notes; added here by F36-a’s Reflect, 2026-11-18): the 256 KiB request bound answered413, the cap of 20 entries on anUNKNOWN_FIELDrefusal, and the photo read’sX-Content-Type-Options: nosniffandsandboxContent-Security-Policy headers.forge-et-bedocs/api/README.mdalready states all three. -
Move
FileTooLargeExceptionandUnsupportedFileTypeExceptiontoplatform(Nanndish; raised by the review of forge-et-be#432, Nit 5; found missing from this file by F36-a’s Reflect, 2026-11-18): both are incom.empireteas.attachment.serviceonforge-et-bemain, andcapture.servicethrows them too, so the capture feature depends on the attachment feature’s service layer for two shared refusals. Done when both live inplatformand the two features import them from there, with no change to the responses. -
Two capture-app test nits left for the next change to their files (Nanndish; found missing from this file by F36-a’s Reflect, 2026-11-18): (1)
forge-et-pwae2e/seam/support/plan.ts’s header does not listexpectCampaign,expectExhibitorandexpectUserRefamong its Phase 2 imports (WI-3.6’s pre-PR review); (2) no test covers a send pass that starts while the send-pass chunk is still loading inCaptureOutboxProvider(the delta review of forge-et-pwa#132). Done when the next PR that changes each file fixes it, or this row is closed as not worth a change. -
Decide whether the capture app must close the last cross-window sign-in gap (Nanndish; raised 2026-11-14 by forge-et-pwa#132’s delta review): the
SESSIONcookie is shared by every window of the capture app’s origin. Since #132’s fix, a send pass reads/mebefore its first send and stops when it names another rep, and a confirmed answer whosecapturedByis not the capture’s rep stops the pass and keeps the capture. One gap is left that the phone cannot close: if rep B signs in from another window between that/meread and a send, rep A’s capture is created under B, then kept and sent again later under A — a duplicate lead under B, which a reviewer rejects. It needs two windows of the app open on one phone with a sign-in in the second at that moment. Done when Nanndish accepts the gap as it is, or asks for a backend check (for example a header naming the expected user that E-1 refuses on a mismatch — a wave 2 contract revision). -
Decide whether Try Again shows on a capture whose photo is missing on the phone (Nanndish; raised 2026-11-14 by forge-et-pwa#132’s second delta review): a card whose photo the phone no longer has is kept as Refused with the phone’s own reason, The photo for this capture is missing on this phone. Remove it and capture the card again. Its row still offers Try Again, which re-checks, counts another attempt and stays refused, as every refused row’s does. Done when Nanndish keeps it for consistency, or asks for Try Again to be hidden for that reason (
forge-et-pwasrc/features/capture/NotSentList.tsx,refusal.code === PHOTO_MISSING_CODE). -
Decide whether a rep can untick rows before importing an exhibitor list (Nanndish; raised 2026-11-13 by F36-a-WI-3.5’s plan): the capture-app screens the lead accepted on 2026-11-12 (
pwa-current/, Check the Exhibitor List, screenshotsphone/60,61) draw a checkbox per row, but F36-a FR-P5 and the frozen wave 3 contract (U-2) import every row with a company. WI-3.5’s draft plan follows the spec and contract, drops the checkboxes, and redraws the screen before approval (its## Notes, What still blocks approval). Decided provisionally on 2026-11-14 by Claude under Nanndish’s overnight delegation: no row choice, as the spec and the frozen contract say; the redrawn screen is not put to the lead, because it follows the spec rather than changing it; WI-3.5 is built that way. Done when Nanndish confirms both in the morning — or, if a row choice is wanted, when the spec, the contract, WI-3.2’s and WI-3.5’s plans and the drawing are revised for it. -
Confirm that Enter in a one-line field does not send the capture app’s campaign and exhibitor forms (Nanndish; raised 2026-11-15 by the review of forge-et-pwa#136, finding B1): nothing decided what Enter does in the campaign form and the Add Exhibitor / Edit Exhibitor form, so return in Name created a dateless campaign and return in Company Name added the exhibitor before the booth and hall were typed. Decided provisionally on 2026-11-15 by Claude under Nanndish’s delegation: Enter in a one-line field is cancelled and the form is sent only by its button (the reviewer’s option (a), suited to a phone keyboard), recorded in the WI-3.3 and WI-3.4 plans’
## Notesand built onfeature/F36-a-wave-3. Done when Nanndish confirms it — or, if Enter should send the form, when the forms and their tests are changed back with a comment saying it is intended. A second choice of the same kind, in the same review (delta round 1, D4): the plan list’s no-match Clear now moves focus to Filter, not the search box, because focusing a text field after a tap opens the phone keyboard over the list just cleared; recorded in the WI-3.4 plan’s## Notes. Confirm both together. -
Confirm two choices in F36-a WI-3.4’s booth drawer (Nanndish; raised 2026-11-14 while reconciling WI-3.4’s plan with WI-3.3’s revision): (1) Save Note is
unavailablewhile the note is unchanged, with no line beside it saying why —forge-et-pwasrc/ui/Button.tsx’s rule asks the screen to say what is missing, and Claude judged the note field directly above to be that cue; (2) no design-system revision names the boothDraweras the capture app’s side form of §6’s CRMSheetpattern, which WI-3.3’s revision names for its bottom sheet. Both decided provisionally by Claude under Nanndish’s overnight delegation and recorded in.forge/plans/F36-a/F36-a-WI-3.4-plan-list-booth-plan.md § Notes. Done when Nanndish confirms them, or the plan (or the built drawer) is changed for either. -
Settle whether a wave PR into
mainneeds the full browser gate or only the selected run (Tharinda; raised 2026-11-13 on forge-et-fe#531):.claude/CLAUDE.md § Quality Gates › Browser QAsays a PR intomain, a wave PR included, needs no full gate and runs only the browser tests its change selects (PRD Rev 81, Rev 93). Butforge-et-fescripts/browser-gate/rules.mjsbranchNeedsFullRunandpullNeedsFullRunforce a full run for any…-wave-<N>branch (citingtest-run-cost.mdP-1, P-9), sopnpm e2e:gate --selectedon a wave branch ignores--selected. That costs 2–3 hours on Windows for a one-file test change. Done when the rule and the script agree, one way or the other, and the other is changed to match. -
Make
pnpm e2e:gate --selectedselect nothing for a change undere2e/smoke/only (Tharinda; found onforge-et-fe#525, BUG-417, 2026-11-13): the selected run falls back to the full browser suite when a changed file is imported by no spec, ande2e/smoke/*.smoke.tsfiles match no local Playwright project’stestMatch(playwright.config.ts,smokeConfig), so a smoke-only change asks for about an hour of browser tests that cannot run any of its code. Done when the gate’s selection treatse2e/smoke/**as selecting no spec and says so in its line, with a test. -
Move the three older copies of the constraint-name walk onto the
platform.persistencehelper (Tharinda; raised by the review offorge-et-be#415, F32-d wave 3, finding I2, 2026-11-12): the wave adds oneplatform.persistencehelper that reads the refusing constraint’s name from an exception chain, and uses it inSavedViewService. The same loop is still copied inemailtemplate/service/EmailTemplateService.java,mailbox/service/MailboxAuthorizationCompletionHandler.javaandcustomfield/service/CustomFieldDefinitionService.java, which that PR does not change (coding-standards §2 Reuse). Done when the three call the helper and their tests pass. -
Decide whether
RecordEditDialogshould report its close, asDeleteViewDialognow does (Tharinda, the lead; raised by the fix of F32-d’s Delete focus defect,forge-et-feed63a12d, 2026-11-12): a focus move made while a Radix modal is still closing does nothing, because the rest of the page isinertuntil its content unmounts (F32-d findings F-011).DeleteViewDialognow callsonClosedfromonCloseAutoFocusandViewActionsretries focus then. Save View, Save as New View and Rename use the sharedRecordEditDialog, which has no such signal; the WI-3.4 keyboard walkthrough saw all three put focus where they should, so this is a guard against a later timing change, not a known defect. Done when the lead decides, and the change, if any, lands with a test that emulatesinert. -
Move the CRM’s four copies of the dialog header into one
src/uipart (Tharinda, the lead, to assign; raised byfrontend-revieweron forge-et-fe#517, 2026-11-12):forge-et-fesrc/ui/overlays/RecordEditDialog,src/features/settings/SettingsDialog,src/features/notesAndTags/NotesDialogandsrc/features/priceEmail/PriceEmailDialogeach build the same header (the title row, the description, the ✕ in its.closeRow). BUG-408 had to be fixed in all four, and the BUG-156 browser test measures only the Create User dialog. One shared header part would let a single check cover every dialog. In the same change, movedateOrderandtypeDate(three copies:e2e/browser/custom-fields/values-support.ts,pricing/price-expiry-a11y.spec.ts,pricing/price-import-a11y.spec.ts) intoe2e/support/. -
Keep the full test suites under their CI caps (Tharinda, the lead; measured 2026-11-12 across the five repos’ runs since 2026-09-20): an ordinary PR’s Fast lane is flat (backend about 2.5 minutes, CRM 1.1, capture app 0.5, harness
Repo hygiene1.6), but the whole suites have grown: the backend’sSlow lanerose from 12.5 minutes (2026-09-30) to 47 (2026-11-11), against a 60-minute cap then and 90 since 2026-11-13, and the CRM’s from 15 (2026-10-16) to 24–28 (2026-11-11) against its 45. Decided 2026-11-13 (the lead): a Dependabot PR gets a local record like any other PR, so theSlow laneruns only on demand and the CI caps bound only a run someone asks for (test-run-cost.mdN-5; about 560 runner minutes went on Dependabot Slow lanes in the 30 days to 2026-11-12). Decided 2026-11-13 (the lead): the backend’s on-demandSlow lane— the rungit-conventions.mdsends people to for a migration or security code — is capped at 90 minutes, up from 60, so it does not time out as the suite grows (forge-et-be#420; only the minutes used are billed). Still to do: (0) when that run nears 90, decide between a larger GitHub runner (4 processors, 16 GB, paid per minute, set up by an org admin; two test JVMs do not fit the standard one, forge-et-be#411) and finding the slowest specs. (1) watch the nightly backend build, in two test JVMs since 2026-11-12 (test-run-cost.mdB-7; 31 minutes against 43 in one, under its 60-minute limit): a night whose load average is far above the Mac’s eight cores can lose one JVM at start-up (initializationErroron hundreds of specs, “Could not connect to Ryuk”), which is load, not a defect; on a night with no other load it is a bug. (2) The CRM’s Vitest suite grows about 2–3 minutes a week, which now lengthens the nightly run and the local records rather than a CI job: when it nears the nightly’s limit, split it with--shardor run more workers. (3) Look at thebe-buildandfe-vitesttimes in the nightly report every week or two. -
Give Empire the answers to its asks from the 10 November progress meeting (Tharinda, the lead;
.forge/discovery/meeting-notes/2026-11-10-kavindu-crm-progress-update-demo.md§3–§7). Each answer is decided and recorded; what is left is telling Empire, and the wider demonstration. (1) The capture app is delivered in Phase 1: agreed at the meeting, recorded on 2026-11-11 and 12 (PRD Rev 91); the row “Carry the capture app’s move to Phase 1 into every document” holds the rest. (2) Flagging emails nobody has replied to (Roshan; Kavindu believed it was already built, and nothing did it, checked 2026-11-12): decided 2026-11-12, F44, a new Phase 1 feature after F20-f (PRD Rev 92, acceptance criterion F7) — tell Kavindu and Roshan it is coming in Phase 1. (3) Automatic follow-up emails for stagnant leads (Sajeev De Silva): decided 2026-11-12 that no AI-drafted email is ever sent without a person approving it (the lead, confirmed with Rukshan Salgado, Thilina Abeyrathna and Malinda Wellalage; as the PRD already says). What he described is F25, outreach sequences (in the proposal’s V1 scope, Phase 2): reminders and follow-up steps scheduled on a lead, each email approved by a person before it sends. The meeting’s answer named bulk email (F37), so the answer to him names F25. (4) The ERP link: Kavindu answered “yes, once Empire has chosen its ERP”; decided 2026-11-12 that it is out of scope until Empire names the ERP replacing IFS and gives its details, and V1 gives only the handoff document (F19, Phase 2; OQ-2). (5) The demonstration for the wider brand and marketing team, and a date for implementation (Kavindu to arrange). Note for (2) and for “manage customers fully in the CRM without Outlook”: both depend on F20-f, mail linked to accounts, contacts and leads (Phase 1, not started, its spec a draft on 2026-11-12). Done when Empire has the answers to (2) to (4) and the demonstration in (5) is arranged. -
Decide whether the Leads and contact lists’ data bars take the account list’s two groups (Nanndish; raised 2026-11-19 by BUG-504’s fix): since BUG-504 the account list’s filters wrap in a group of their own and Columns / Save View keep to the first line’s end (design-system Rev 147). The Leads list is still one wrapping row: measured on 2026-11-19 at 1440px, its Columns and Save View sit on a second line for an Administrator, and Clear Filters can push them down. Done when the lists match, or the difference is recorded as intended.
-
Share one hook for the focus move after a Tag Try Again inside More Filters (Nanndish; raised 2026-11-19 by
frontend-revieweron BUG-504’s fix, finding 5):forge-et-fesrc/features/leads/LeadFilters.tsxandsrc/features/accounts/AccountsFilters.tsxeach move focus to the Tag trigger when a retry inside the portalled popover succeeds and leaves focus nowhere, written twice and already different (Leads drops the press from the settled query; Accounts compares the reads’ update time, becauseuseAccountTagFilter().retry()resolves with a time only). Left out of BUG-504’s PR because it changes the Leads filter, which that PR does not otherwise touch. Done when both use one hook insrc/ui/filtersbesideMoreFilters, with their focus tests unchanged. -
Check the two other popovers capped to Radix’s available height (Nanndish; raised 2026-11-20 by BUG-509’s fix, L-334):
forge-et-fesrc/features/activities/FollowUpDialog.module.css(line 31) andsrc/ui/fields/customFieldRenderers.module.css(line 57) cap a popover with--radix-popover-content-available-height, the cap that kept the Add Tag list from flipping above when it did not fit below. Done when each is checked at 320×640 and the smallest tablet size with its trigger near the window’s bottom edge, and either takes a fixed cap or is recorded here as meant to stay on one side and scroll. -
Draw the Email folder rail at the app’s width in the current CRM prototype (Nanndish; raised 2026-11-19 by BUG-507’s fix):
.forge/design/prototypes/crm-current/draws the collapsed folder rail’s tiles 56px wide, where the app’s rail is--et-layout-sidebar-width-collapsed(76px), so a 99+ badge, drawn by the prototype since BUG-507, would run about 7px past its tile there; the demo data never goes above 99. Done when the prototype’s rail takes the app’s width. -
Draw design-system Rev 107 in the current CRM prototype (Nanndish; raised by the change that made it, 2026-11-09):
.forge/design/prototypes/crm-current/still draws form labels intext/secondary, lower-case placeholders, the green hot dot and grey cold ring, the account page’s Buying Standards search inside the table’s card, and the Weekly Prices refresh line above the round summary. Done when the prototype draws the five changes asforge-et-febuilds them. -
Move the remaining hand-drawn section banners onto shadcn
Alert(Nanndish; raised by design-system Rev 110, 2026-11-11): Rev 110 makesAlertthe one banner component, andfeature/ui-polish-round-2movedErrorPaneland six notices onto it, but theseforge-et-fesection banners still draw their own box:PublishRoundDialog’spricing-publish-conflict,PriceEmailDialog’s.failureand.leftOutNotice,UserEditorForm.tsx(about line 524),EffectivePermissionsSection.tsx(249),LoginRoute.tsx(403, 409),CheckReport.tsx(108),CreateStep.tsx(99),SaleStep.tsx(284),ComposeEmailDialog.tsx(735, 1261, 1325),EmailRoute.tsx(106, 223),MessageView.tsx(62),TemplateView.tsx(256),TemplateEditorForm.tsx(803),LeadResearchPanel.tsx(84, 160),LeadRankingPanel.tsx(73, 143),LeadTriagePanel.tsx(37, 137),ComposeAiDraftButton.tsx(145, 217, 256),LeadFollowUpDraftPanel.tsx(261),RecordAttachmentsPanel.tsx(158) andAccountFormPage.tsx(1245–1279). Line numbers are as of that branch on 2026-11-11. An inline field error stays the field’s own message and is not covered. Done when each is anAlertin the variant its message calls for, or recorded here as kept and why. -
Decide whether every AI provenance source label takes sentence case (Nanndish; raised by design-system Rev 110 and F12 Rev 10, 2026-11-11): the inquiry extraction’s label is now Pasted inquiry,
, but the labels the same backend builder writes for the other AI features stay lower case — drafts email from … / inquiry, triage inquiry / lead fields, research public web search, ranking inquiry and research / inquiry only — because approved text writes them so (F11-h’s wave-1 contract’s source-labels table; design-system §5 and Revs 69 and 99). Done when either all take sentence case, with revisions to the specs, contract and design-system examples that quote them, or the mixed casing is recorded as intended. -
Decide whether AI Research should lead the lead page below
bp/laptop(Nanndish; raised by design-system Rev 110, 2026-11-11): Rev 110 puts the AI cards first on the lead page, but belowbp/laptopthe page is one column with the wide column’s cards first, so AI Research, first in the narrow column, comes after the participants place on a tablet;forge-et-fesrc/ui/recordPage’sRecordColumnscannot move a narrow-column card up. Done when it is decided to leave it, orRecordColumnsgains a way to order the cards in one column and the lead page uses it. -
Decide whether a dialog’s “is the user still here” check should ignore replace navigations (Nanndish; raised by the review of forge-et-fe#493, 2026-11-11):
forge-et-fesrc/features/pricing/StandardsBook.tsx(the Add standard dialog, F26-b Rev 9) andWeeklyPricesTabcompare the history entry (location.key) the dialog opened on with the current one before moving focus to the new row and writing the URL (L-180). A replace navigation also changes the key — the book’s search debounce (update({ q }, { replace: true })) or the past-the-end page clamp — so one that lands while the dialog is open leaves the new row unfocused and the URL as it was, though the user never left. Done when it is decided to keep that, or the check compares something only a push or a Back changes (the history index) and both screens use it. -
Finish naming BUG-403’s control-character rule in the note body Javadoc (Nanndish; raised by
backend-reviewer’s delta review of forge-et-be#403, 2026-11-13, left there because a Javadoc-only commit costs a new two-hour build record): inforge-et-be,lead/web/LeadNoteCreateRequest.java,account/web/NoteCreateRequest.javaandlead/service/LeadNoteService.javalist the note body rules (“required, not blank, at most 10,000 characters”) without theINVALIDrule for a control character other than a tab or a line break;account/web/NoteRequests.java’srequireBody@throwssays “holds a control character” without that exception; andlead/web/LeadNoteController.java’s class Javadoc has two lines over 120 columns. Done when the next PR that touches these files states the rule asNoteRequestChecksdoes and rewraps the two lines. -
Correct the form-label comment in
Field.module.css(Nanndish; raised by the wording review of forge-et-harness#1253, 2026-11-10):forge-et-fesrc/ui/components/Field/Field.module.css.labelsays the placeholder and helper text “stay intext/secondary”; the placeholder istext/faint(design-system §5 Form fields). Left out of forge-et-fe#478 because a stylesheet edit re-runs its whole test record. Done when the next change to that file rewords it. -
Decide whether the capture app’s placeholders take Title Case too (Nanndish, with Tharinda, whose decision Rev 105 was; raised by design-system Rev 107, 2026-11-09): Rev 107 put the CRM’s placeholders in Title Case;
forge-et-pwakeeps sentence case (Rev 105). Done when the capture app is recased by a PR of its own, or Rev 107’s exception is confirmed. -
Decide whether CI checks a frontend PR’s
Selected browser run:line (Tharinda, the lead; left open by.forge/design/harness/fewer-nightly-reds.mdQ-1, 2026-11-12): R-1 is checked today by the review (frontend-reviewerdimension 32,/forge-review-prstep 5). A CI check would mean the frontends’Fast lanereading the line on a PR intomainandscripts/browser-gate.mjsletting a selected run be pushed, which it refuses by design; both change workflow or gate code, so they need the lead’s approval. Done when the lead decides, and a yes is built. -
Fold the capture-app decisions of 2026-11-12 into the PRD and the F36 specs (Nanndish, with F36-a’s spec; raised 2026-11-12): the team’s discussion of 2026-11-12, drawn in
pwa-current(.forge/design/prototypes/README.md § The capture-app screens and flows agreed on 2026-11-12), changes what the PRD and the F36 stubs say. Items 1 to 4 are in the PRD since Rev 95 (2026-11-12), with FS-6 Rev 23 (capture.usein Appendix A), architecture Rev 88, the proposal’s post-issue note and the client pack (client-facing-prd.md§3.5,phase-1-acceptance-criteria.mdL1–L3,roles-and-permissions.md§4.7 and A-41); what is left of them is the specs’ part — F36-a’s spec states items 1 to 4 for capture, where a held capture is stored (and the.forge/design/data/entity-map.mdrow for it, whose Capture app (F36) paragraph says a capture submission is simply a lead and adds no CRM entity) and what converting needs besidecapture.use; F36-c’s spec states how research runs before conversion; F36-b’s states item 4 for card reading; andforge-et-bedocs/auth/recommended-roles.mdgains the key when F36-a declares it. Items 5, 6 and 8 are settled since PRD Rev 96 (2026-11-12): planning before a show is in Phase 1 (the lead’s decision, reported by Nanndish), built in F36-a (the Plan screens and data, items 5 and 6, as later waves) and F36-c (the research and ranking of every exhibitor, its expected time and running cost); both stubs say so, and the client pack (client-facing-prd.md§3.5,phase-1-acceptance-criteria.mdL8 and L9,open-questions.mdQ9) and a post-issue note on the proposal carry it. Item 9’s states were accepted as drawn by the lead on 2026-11-12 (reported by Nanndish). Also settled on 2026-11-12 (Nanndish): converting a captured lead needslead.managebesidecapture.use(F36-a’s stub); the “capture and feed first” fallback keeps the review (PRD Rev 96); and AD #37’s list of keys with no reach is left as it is, since FS-6 Appendix A already sayscapture.usehas none. Each needs its home: (1) review before the CRM — PRD F36 says captured leads land in the F11 triage intake; now a person converts each lead (creating the CRM lead) or rejects it, a rejected lead stays in the capture app with an optional reason from a set list and an optional note (the reason was made optional later on 2026-11-12, Nanndish), and nothing reaches the CRM unreviewed (PRD revision, F36-a and F36-c specs); (2) the converted lead’s owner is the person who converts it, with the capturing rep kept as captured by (AD #35’s one owner; F36-a); (3) a capture-app permission,capture.use(Use the capture app), giving full use of the app, a special key likeprice.publishthat ships fail-closed (AD #30) — add it to FS-6 Appendix A andclient/roles-and-permissions.md, and draw nothing finer; (4) one capture at a time — Scan a pile and choosing several photos are dropped, and a capture is sent only by Submit for Research (F36-a, F36-b); (5) an uploaded exhibitor list is read without AI (Nanndish, 2026-11-12, replacing that morning’s AI reading): the app reads the CSV or Excel file, matches each column by its heading against a fixed list of names per field, and the rep checks and corrects the match before importing; Company is required, and rows without one are skipped (the POC reads a list the same way; F36-a since PRD Rev 96). Every exhibitor on an imported list is researched (Nanndish, 2026-11-12): the POC’s Research the top N by fit setting, and the rough AI ranking of the whole list it needed first, are dropped — simpler, but every row now costs a research and the plan takes longer to finish; at Haiku 4.5’s quota of 10 requests a minute that F42 paces, 500 exhibitors need at least about 50 minutes at one AI call each, longer while leads are being triaged too, so the planning spec states the expected time. The larger cost is the research of each exhibitor before the show (score and reason): at the captured-lead research rate (about 2¢ plus a 0.7–0.8¢ web search,ai-provider-strategy.md§2) a 500-exhibitor show is roughly $10–15, against the annexe’s $26–31 a month of expected AI usage (§5), and neither the annexe norai-provider-strategy.mdcosts it as its own line — the planning spec states it; (6) campaigns archive and unarchive, Re-rank is dropped, and each booth on the plan takes visit notes and a visited mark; the Plan card shows no AI line — the short hook decided that morning was dropped later the same day (Nanndish), and the reason to visit is on the booth page — and a booth’s short stand code (S2-D03) is kept apart from its hall, so a card shows the code in full and cuts only the hall; the plan has a search, the status chips All, To Visit, Visited and Archived, and a sort (best fit by default, highest score first; booth number; company; recently added) and tier and hall filters; an exhibitor is added by hand with the company name only required, and website, booth number, hall, country and notes optional (the POC requires the name only, with website and booth optional), several can be added in a row (Add Another), is edited (a new company name or website researches it again), and is archived with an optional reason and restored; every campaign also takes its show’s first and last day (optional), from which the Plan screen shows it as upcoming, running or ended, and an ended campaign’s Plan screen leads with creating the next one (the README’s Plan puts the campaign first; today only an expo campaign has dates); (7) email — settled: the capture app sends no email, its notifications are in-app only (proposed by Nanndish, confirmed by Tharinda on 2026-11-12): AD #42, PRD Rev 94, architecture Rev 87, the proposal’s post-issue note of 12 November, the client pack and F36-c’s stub say so; what it leaves for the backend’s SES sender is the next row; (8) is exhibitor planning in V1? — OQ-55’s working position and F36-a’s spec keep the POC’s pre-show exhibitor CSV targeting out of V1, and no Gate 3 feature covers it, yet Tharinda showed it to Empire on 2026-11-10 and the decisions above draw it; Tharinda decides whether it is in scope and, if so, which feature holds it and its size, and records it against OQ-55 — settled by PRD Rev 96: in, in Phase 1, in F36-a and F36-c; its size is the lead’s row “Say whether planning before a show changes the price”; (9) the states drawn after the review (Nanndish, 2026-11-12; README § The capture-app screens and flows agreed on 2026-11-12) — a capture submitted with no connection is kept on the phone, listed as Not Sent on Capture and counted on Leads, and sent, oldest first, when the app is online again (F36-a, PRD Rev 64); the rep corrects what the card reader read, or the badge carried, with Edit Details before converting, a new company or website researching the lead again (F36-b); the existing-client check shows Already in the CRM (the account, its owner, whether the person is a contact, emails in the past year, the last contact), New to the CRM or Customer check pending, on the badge as soon as it is read and on the lead (F36-d); and nothing in the app opens the CRM (Open in CRM removed). Whether converting a lead whose company is already an account links the CRM lead to that account is settled by F36-d’s spec (D-11, FR-L1; approved 2026-11-13 under Nanndish’s overnight delegation, for his review): it does when the converter reaches the account and leaves Link the lead to this account ticked, and otherwise F11-d proposes the match in the CRM. Done when the PRD and the F36 specs state each, or the lead changes one. -
Say whether planning before a show changes the price, and cost it in the annexe (Tharinda; raised 2026-11-12 with PRD Rev 96): the lead put exhibitor planning in Phase 1 (F36-a and F36-c) on 2026-11-12. The proposal’s §4 does not name it, and its post-issue note of 12 November states the addition without a price or change-request statement, which is the lead’s to make. Researching every exhibitor is roughly $10–15 of AI and web search for a 500-exhibitor show at Sonnet 5.5’s research rate, and about $7.50 on Haiku 4.5, which every AI feature runs on for now (F36-c’s draft spec of 2026-11-13, NFR-4 and FR-DC3, which words the annexe line);
client/sow-cost-annexe.mdcarries it as its own line since 2026-11-16 (§6, with §5’s note and §6’s restated web-search row; F36-c-WI-1.6, F36-c’s wave 1); whether it changes the price is the lead’s decision alone. The capture app’s work by the Phase 1 build target (Fri 2026-11-20) also grows: none of the F36 specs is approved yet and F43 is at spec, so the capture-app progress review, proposed for 2026-11-20 and still to be confirmed with Empire, is where the reduced release can be chosen. The architecture’s capture-app sizing (about 2.5–3.5 weeks,architecture.md § Trade-Show Capture App) predates the addition, and the phase-1 acceptance criteria’s reduced-release paragraph defers L9 with L5, which follows from “research following” but was not asked of the lead: both are for the lead to confirm. Done when the lead has said whether a change request or price note is needed (the annexe half is done). -
Decide whether the annexe’s monthly web-search line rises to ≈ 0–3 (Tharinda; raised 2026-11-16 by F36-c-WI-1.6): F36-c’s planning searches take
client/sow-cost-annexe.md§6’s web-search row to about 4,600 searches a year, about $32 a year — about $2.70 a month — were every search paid, above §7’s line “≈ 0–2 — nil while the free credit covers the month”, whichclient/client-facing-prd.md§ 8 repeats. Every month stays inside Exa’s free credit at the planned volumes, so the expected figure is still nil. Either leave §7, or raise its line to ≈ 0–3 and the two totals’ upper ends by one (to ≈ 150 to 208 and ≈ 175 to 253) in the annexe and the client PRD, with a dated header line in each. Done when the lead has decided and any change is made. -
Run F36-b’s three remaining card-reading states in the capture app against a real backend (Dinil, with F36-b; raised 2026-11-12 by F36-b’s spec review; narrowed 2026-11-17 by F36-b WI-3.5).
- Still not run against a real backend: a card
readingfor over 10 minutes, a card read withcard-readswitched off (DISABLED), andnot_read/UNAVAILABLE. The local stack cannot produce them on demand: they need a clock it does not expose, a second backend start with the switch off, or six minutes of retries. - How they are covered now: F36-b’s spec (FR-H2; AC-7, AC-15, AC-16, AC-26) reaches them in the capture app’s browser tests with pinned
/api/v1/capture/leads/{id}responses, and proves them on the real backend in the backend’s integration tests (BusinessClock, the feature switch). - Now run against the real backend by WI-3.5’s seam specs (
card-read.seam.spec.ts,card-read-screens.seam.spec.ts;F36-b-WI-3.5-test-verify.md):read,not_read/NOTHING_READandnot_read/PHOTO_UNREADABLE(WI-3.5 Decision 13). - Done when such a run exists for the three left (for example a test-only clock or switch the local stack exposes, or a manual run on dev recorded in F36-b’s plan), or the lead decides the split is enough.
- Still not run against a real backend: a card
-
Confirm with Empire how many trade shows a year it attends and plans (Tharinda; raised 2026-11-13 by F36-c’s spec review, W-12): F36-c’s yearly cost and search figures (FR-DC3, NFR-4) and the annexe’s search volume assume six shows a year, from
lead-research-source-options.md§4; Empire has not given the figure. Done when Empire has said, and the annexe line F36-c adds carries the confirmed figure. -
Add F36-b’s strict card-read check one release after wave 2 (Dinil, F36-b’s owner; raised 2026-11-17 by the wave 2 review, forge-et-be#455, F36-b spec Rev 7):
V75__captured_lead_card_read.sqlaccepts a card capture with no card-reading state for one release, so the previous release can still insert card captures during a deploy (coding-standards.md§2 Persistence). In the release after the one that ships wave 2 — a later F36-b wave or a follow-up PR — a new migration sets any card capture still without a state tonot_read/UNAVAILABLE/read number 1 and adds the strict checks (a card capture has a state, a read number of at least 1 and a request time); the code’s “no state meansnot_read” handling can then go. Done when that migration is onmain. -
Correct the
et.ai.*comment on the F36-c model overrides (Dinil, F36-c’s owner; raised 2026-11-19 by the review of forge-et-be#485, a Nit left there because that PR is configuration only):forge-et-besrc/main/resources/application.yml’set.ai.*comment (about lines 302–311) says the task definitions “override capture-research the same way” as plan-research, which reads as Haiku 4.5’sus.profile, while dev sets capture-research and capture-ranking to the Global profile; it does not name the capture-ranking override, and its list of dev’s Global-profile features names only lead-triage, lead-research and lead-ranking.docs/ai/README.mdstates both correctly. Done when the comment matches F36-c FR-RS2 andtd.dev.json, in the next F36-c code PR. -
Keep F36-a’s and F36-b’s browser specs in step with what F36-c replaces (Dinil, owner of F36-b and F36-c, with Nanndish for F36-a; raised 2026-11-13 by F36-c’s Decomposition Plan review, pass 2 I2): F36-c makes Best fit the plan’s default sort, adds the tier and score to the plan card, changes the line under the plan’s title, adds research states to the To Review rows and changes the capture confirmation to Submitted · you’ll be told when it’s ready. F36-a WI-5.1 (its whole browser suite) and F36-b wave 3 (WI-3.2 to WI-3.6, numbered waves 4–5 before F36-b’s Decomposition Plan Rev 1) are written while F36-c is built, so a spec there may assert the old behaviour. Whichever side merges second updates the other’s specs in its own PR; each F36-c wave’s Stage 12 “has
mainmoved” step checks for them, and F36-a WI-5.1’s and F36-b WI-3.2 to WI-3.6’s plan authors read this row (F36-b’s wave-3 plans carry it, 2026-11-16). Done when F36-c’s wave 2 and both features’ E2E waves are merged withmain’s nightly run green. -
Add F36-b’s spec revision for the
CardReadEndedevent that F36-c wave 2 added to F36-b’s writer (Dinil, owner of F36-c; raised 2026-11-17 at F36-c wave 1’s Stage 12, narrowed 2026-11-19): F36-c uses both events —CardReadEnded(capturedLeadId, state)for every write that ends a card read (F36-c FR-TR3) and F36-b FR-E6’sCapturedLeadDetailsChangedwith causeEDITEDfor edits (FR-TR4) — and wave 2 shipped the event in F36-b’sCardReadWriter(forge-et-bec524d9dd, merged withforge-et-be#475). F36-c FR-DC2 asks for an F36-b## Revisionsentry for it in the shipping harness PR; none was added, and F36-b’s spec does not mentionCardReadEndedas of 2026-11-19. F36-c-WI-1.6’s plan holds the draft, written before FR-E6: correct it so it sits beside FR-E6 rather than contradicting it. Done when F36-b’s spec carries the revision (a reworded requirement in an approved spec, so one fresh-context reviewer checks it first,.claude/CLAUDE.md § Compounding Engineeringitem 6). -
Find why the backend test JVM’s heap keeps growing, so the next suite growth needs no raise (Dinil; raised 2026-11-17 by
backend-revieweron forge-et-be#456): the test heap went from 2g to 3g to 4g in under a month (BUG-247). Spring’s context cache keeps at most 32 contexts, so heap use should depend on each context’s size, not on the number of tests, unless closed contexts leak (listener containers, executors, static references). Take one heap dump on a failing run (-XX:+HeapDumpOnOutOfMemoryErroron the test task), or try a fixedspring.test.context.cache.maxSize. Done when the cause is named in BUG-247 and either fixed or the heap rule is changed with the lead’s agreement. -
Keep the capture app’s toasts open while the pointer is over them or focus is inside them, each on its own (Nanndish, with F36-a; raised 2026-11-14 by F36-c-WI-2.5’s plan review, pass 3 I2): F36-a’s
src/ui/toast/Toaster.tsx(onforge-et-pwafeature/F36-a-WI-2.2-capture-screen) keeps onepausedflag, set by pointer enter or focus and cleared by pointer leave or blur. Moving the pointer off a toast while focus is still inside it, or blurring while the pointer is over it, restarts the close timer. The design system’s toast rule, and F36-c’s notice (FR-NT8, which WI-2.5 builds with two flags), say it never closes while either holds. Fix the Toaster with two flags,pointerOverandfocusInside, and a test for each sequence. Done when F36-a’s Toaster pauses on either and its tests cover both. -
Decide whether a capture-app user may learn that a company outside their reach is a customer (Tharinda; raised 2026-11-12) — F36-d (option (a), 2026-11-12, under Nanndish’s overnight delegation) reveals nothing about an account outside the rep’s reach, so a rep without
account.allhears Not found among the accounts you can see for a colleague’s customer. Empire’s ask (PRD F36, Rev 23) is to know at the stand whether the organisation is a customer. Showing existence and activity to everycapture.useholder needs a refinement of AD #35/#37 signed off by the lead, a row inarchitecture-decisions-why.md, and a line inclient/roles-and-permissions.md; then F36-d is revised (.forge/specs/F36-d-existing-client-check-spec.mdFR-V2, D-8). Done when the lead decides. -
Keep one
emailtimeline row per linked email in F20-f’s spec (the F20-f owner; raised 2026-11-12 by F36-d’s spec review): F36-d’s “emails in the past year” counts the account’stimeline_eventrows of kindemail(F36-d FR-C1, W-2), and F20-f, which links mail to accounts, is a stub. F20-f’s spec states that each linked email writes oneemailtimeline row on its account. Done when F20-f’s approved spec says so, or says otherwise and F36-d FR-C1 is revised. -
Renumber the clashing
pwa-currentpictures 67, 68 and 69 (the owner of whichever lands second: F36-a’s WI-3.4 and WI-3.5, or F36-c’s PR #1381; raised 2026-11-14 by F36-d-WI-1.3’s plan): F36-a-WI-3.4’s plan addsscreenshots/phone/67-plan-no-match.pngand68-booth-archived.png, and WI-3.5’s adds69-upload-refused.png. F36-c’s open harness PR #1381 already commits67-plan-research-states.png,68-booth-not-yet-researched.pngand69-booth-research-failed.png. Done when the pictures and theirscreenshots/README.mdrows have distinct numbers onmain. -
Decide whether F36-d’s one-second read-to-result target (NFR-2, AC-27) needs a binding reading on dev (Nanndish; raised 2026-11-14 by F36-d-WI-1.4’s plan): F36-d’s spec measures it on the local stack only and registers nothing as cloud-gated (FR-CG1, AC-28).
.forge/test-strategy.md § Performance Baselines, however, says a spec with a response-time limit names a deployed environment for the binding reading and registers it until that environment exists. WI-1.4 records the local reading (desktop size, in theseamproject) with no pass/fail limit. If dev is needed, revise F36-d’s spec and add a.forge/cloud-gated-validation.md§B row. Done when Nanndish decides, and the spec and register agree with the decision. -
Give the capture app’s
Buttondesign-system §5’s busy spinner, then drop the check region’s own (Nanndish; raised 2026-11-16 by the delta review of forge-et-pwa#139, F36-d-WI-1.2):forge-et-pwa’ssrc/ui/Button.tsxmarksbusywitharia-disabled,aria-busyandcursor: progressonly, so a busy button looks like an idle one, against design-system §5 Busy / loading. That gap predates F36-d and also affects Not Sent’s buttons and Sign In’s Try Again. F36-d’s check region draws its own spinner in place of Try Again’s icon while a retry runs (src/ui/CustomerCheckRegion.tsx,eea926e). Done whenButtondraws the §5 spinner forbusy, with its tests, and the region’s own spinner is removed. -
Correct the “only an
account.allholder can getcontact” sentence in F36-d’s wave-1 contract § Reach and spec FR-V2 (Nanndish; raised 2026-11-16 by F36-d-WI-1.1’s pre-PR review): a caller withoutaccount.allwho reaches an inactive account is answeredcontact(linkable: false) for a contact on it — what FR-M2 step 1 and FR-V2’s reach rule give, revealing nothing the caller cannot already see — so the two sentences are narrower than the rule and the built code. The code is right and was not changed;forge-et-be’sdocs/api/README.mdwas reworded to today’s behaviour on the wave-1 PR (forge-et-be#439, the backend-reviewer’s Important). Done when a spec revision and a contract## Revisionsentry state the case (or Nanndish decides the code should refuse it, which is then a WI-2.1 change, and that README sentence changes with it). -
Draw the busy Try Again under an earlier answer on the lead’s page in
pwa-current(Nanndish; raised 2026-11-18 by F36-d-WI-2.3’s comparison with the drawings): the built region keeps the earlier answer and Check failed while the retry runs (F36-d WI-1.2 plan Rev 2), where the prototype shows Checking again…, or Checking… with nothing earlier. Draw it incrm-match.jsx(where="lead", the22elook with Try Again busy and its spinner, as44hdraws on Capture). Done when the drawing and its screenshot are inpwa-current. -
Bring the check region’s Try Again and the prototype together (Nanndish; raised 2026-11-18 by F36-d-WI-2.3’s comparison with the drawings):
pwa-current’s22dand22edraw Try Again as a pill indented to the text column, while the built region (forge-et-pwasrc/ui/CustomerCheckRegion.module.css.action, the capture app’sButton) puts the outline button at the box’s left edge, on Capture and on the lead’s page alike. Decide which one changes: a drawing change, or a design-system revision applied to both screens in one PR. Done when they agree. -
Draw the capture app’s convert-dialog link box as built in
pwa-current(Nanndish; raised 2026-11-19 by F36-d-WI-2.7’s comparison with the drawings):source/src/components/review-dialogs.jsxConvertDialogand screenshotsphone/25andphone/25adraw a 20px box, 10px from a medium-weight label; the built box isicon/md(18px),space/5(12px) from a regular-weight label, the radio rows’ values (design-system §9, The convert dialog’s In the CRM row, Rev 143). Done when the drawing and its screenshots match the build. -
Draw the convert dialog’s refused-link box and its disabled look in
pwa-current(Nanndish; raised 2026-11-19 by the delta review of forge-et-pwa#185, F36-d wave 2):source/src/components/review-dialogs.jsxConvertDialogand screenshotphone/25d-convert-link-refused.pngdo not draw two things built inforge-et-pwa3e3e872and1ee4994. (1) After a refused link, while the refused answer is still the newest, the In the CRM row shows with the box unticked (F36-d Wave-3 contract Rev 2);25dshows the refusal with the check’s new answer, and the drawing has the box ticked by default. (2) While the convert runs, the box’s row takestext/disabledand thenot-allowedcursor (design-system §9, The convert dialog’s In the CRM row, Rev 143). Draw both and re-take the screenshots withscreenshots/capture.mjs, or record them in design-system Rev 143’s Departs from the prototype list. Done when the drawing and its screenshots show both, or Rev 143 lists them as departures. -
Draw the convert-dialog link box’s checked colour as
accent/goldinpwa-current(Nanndish; raised 2026-11-19 by F36-d-WI-2.7’s comparison with the drawings):review-dialogs.jsxConvertDialoguses the primary button colour (accent-[hsl(var(--primary))]), while the built box takes the radio rows’accent/gold. Done when the drawing usesaccent/gold. -
Add a test that saves Edit Details with a changed company and watches the lead page’s customer check (F36-d AC-31) (Nanndish; raised 2026-11-19 by F36-d’s Reflect): AC-31’s states are built and tested through a reload and the page’s own re-read (F36-d-WI-2.3), and the check re-runs on any change to the company, email or website (
forge-et-pwa31f7ea0), but no test drives the Edit Details save itself; the Decomposition Plan left that trigger to F36-b. Done when aforge-et-pwatest (LeadDrawer.test.tsxore2e/browser/leads/lead-customer-check.spec.ts) saves Edit Details with a changed company, sees Checking again… with the earlier result kept, then the new result, and the spec’s AC-31 is ticked. -
Add the JSON reporter to the run command in the header of
forge-et-pwa’se2e/seam/customer-check-journey.seam.spec.ts(Nanndish; raised 2026-11-19 by F36-d’s Reflect, finding F-001, WI-2.8’s pre-PR review Nit 2): the header gives the journey’s command with thelistreporter alone, which keeps notestInfo.attachbody, so the log check after it has no values to search for (.forge/lessons.mdL-326). Done when the header’s command carries--reporter=list,jsonwithPLAYWRIGHT_JSON_OUTPUT_NAMEset outside the worktree, in the nextforge-et-pwaPR that touches the capture app’s seam tests. -
Have F36-a-WI-4.2’s author fix the empty brackets in Convert to Lead’s first line for a lead with no company (Nanndish, to pass on; raised 2026-11-16 by F36-d-WI-1.3): under
pwa-current’s?check=badge-lead-pendingthe second lead has no company, and F36-a’s unchanged first line reads Aigerim Seitkali () becomes a lead …. No picture shows it. Done when the line leaves out the brackets with no company, inpwa-currentand inforge-et-pwa’s convert dialog. Checked 2026-11-18:forge-et-pwais done —src/features/leads/lead/ConvertSheet.tsxonmainleaves the brackets out when the company is empty. Left:pwa-current, whosesource/src/components/review-dialogs.jsxstill draws({lead.company}). Tell the same author, with it, that WI-4.2’s re-capture of the lead’s page also re-renders F36-d’sphone/10b,22a–22fand25a–25c, which is expected (F36-d-WI-1.3 plan## Notes). -
Allow
blob:images in the capture app’s CSP (Nanndish, with the DevOps team; raised 2026-11-13 by F36-a’s Decomposition Plan review): the capture app’s Content-Security-Policy on dev isimg-src 'self'(environments-and-identity.md, applied by forge-et-iaac#14), which blocks F36-a’s card-photo preview and its photo read (ablob:URL). Aforge-et-iaacPR todevaddsblob:toimg-srcin dev, UAT and production, reviewed with/forge-review-prand approved by Shyamin Ayesh (.claude/rules/forge-et-iaac.md); F36-a’s FR-CG1 (f) checks it on dev. Done when the PR has merged and the dev console shows no CSP violation during a card capture. Update 2026-11-19: UAT now hosts the capture app with the sameimg-src 'self'(forge-et-iaac#23,overlays/uat/locals.tf; its review), and Phase 1’s UAT tests the capture app from Mon 30 Nov, so the fix is due before then, applied to UAT with auat.yamlapply. Update 2026-11-19 (later): forge-et-iaac#27 is open fromfix/capture-csp-blob-images(head95fa623). It changes the capture app’simg-srcto'self' blob:inoverlays/devandoverlays/uat, and leaves the CRM’s policy as it is. Its review verdict isapprove(record; one Nit, fixed). The merge’s apply should change one resource on dev, the capture app’s response-headers policy, in place. Production’s overlay is in Senuthmi Thimansa’s open forge-et-iaac#17, still withimg-src 'self'atoverlays/prod/locals.tf:62; a comment on #17 asks for the same change (the row “Bring forge-et-iaac#17 (overlays/prod) to F38-i WI-4.1’s change list before it merges”, item (b)). Next: Shyamin Ayesh approves and merges #27, anddev.yamlapplies it. Then the lead approves auat.yamlapply before Mon 30 Nov, and #17 carries the line. The row closes when all of that is done and the dev console shows no CSP violation while taking a card photo and while opening a card capture’s lead page. Once #27 is applied, the nextforge-et-pwaPR that touches them corrects the comments that still say the deployed policy allowsimg-src 'self'only:src/ui/brand/Logo.tsx:1-2,src/ui/brand/Logo.test.tsx:6-7,vite.config.ts:76andscripts/validate-pwa.mjs:89. -
Decide the capture app’s desktop layout beyond the centred column (Nanndish; raised 2026-11-13): F36-a WI-1.2 (overnight delegation, 2026-11-13) builds the design system’s centred column at every width (design-system §9; spec NFR-1) instead of
pwa-current’s desktop side menu, and redraws the prototype to match. Still to decide: whether the desktop gets a side menu after all (a design-system revision and the lead’s acceptance), and where F36-c’s notifications bell sits (the header, as built). Settled 2026-11-16: WI-4.2’s Leads tab shows no lead beside the list on desktop — a lead opens in the booth page’s drawer at every width (Nanndish, with Tharinda’s acceptance; drawn inpwa-current,desktop/01a-lead). Done when the lead decides the side menu and the bell. -
Scan capture photos, and handle
capture/keys inScanResultEventConsumer, before malware scanning is switched on (whoever switches scanning on — Tharinda, F38-j; raised 2026-11-13 by F36-a WI-2.1’s plan review): F36-a’s card photos are stored undercapture/in the attachments bucket with noattachmentrow (spec Rev 2, FR-C7). With scanning switched off in every environment this changes nothing today. If GuardDuty Malware Protection is turned on for the bucket, a result for acapture/key matches noattachmentrow while the object exists, soScanResultEventConsumerdefers it until the dead-letter queue takes it, and a capture photo would still be served without a verdict. Before scanning is switched on: decide how capture photos are scanned (a prefix scope, or a scan status oncaptured_lead), and make the consumer drop or apply results forcapture/keys. Done when both are decided and built, or scanning is scoped to excludecapture/with the reason recorded. -
The backend’s SES sender has no consumer after AD #42 (Tharinda, owner of F38; raised 2026-11-12): F36-c was to build the backend’s SES sender and prove a capture lead notification through it, and AD #42 removes that notification. The approved F38 specs still name it: F38 FR-5, FR-26 and AC-26 (the capture notification half), F38-e (decision 5 and its FR on the probe: “F36-c builds the backend’s sender and proves both mails through it”), F38-h (FR-5 and its dependencies: the sender setting’s first consumer is F36-c) and F38-i (the sender setting, first consumer F36-c); the register’s §B FS-5 SES row is updated to match (its update of 2026-11-12). Decide whether any V1 feature sends system mail through SES (F37-a sends through each user’s mailbox, by Graph), and so which feature, if any, builds and proves the sender; then revise those specs with revision entries (the capture halves are void either way). Done when the four specs and the register say who builds the sender, or that none is built in V1.
-
Agree the version files in
pwa-current/versions/(Tharinda, the lead; raised by Nanndish, 2026-11-11): the lead’s 2026-11-05 rule keeps one current copy per app, updated in place, with git history for earlier states (.forge/design/prototypes/README.md § The originals stay frozen). Nanndish added an exception for the capture app so the POC-themed version 1 stays beside the CRM-themed version 2 as a frozen file (§ Versions of the capture-app prototype). Done when Tharinda agrees the exception, or it is reverted to git history alone. -
Decide whether three Pricing dialogs may close while their write runs (Tharinda, the lead; raised by
frontend-reviewer’s review of BUG-376’s fix, 2026-11-09):SetCostDialog,StandardFormDialogandArchiveStandardDialog(forge-et-fesrc/features/pricing/) pass nocloseDisabled, so the ✕,Escand Cancel all close them while their write runs and its answer is never shown — BUG-376’s harm, by design rather than by a missed button (SetCostDialog’s own comment says so). Done when the lead says it is intended (this row goes), or it is filed as a bug. -
Watch the account and lead selection files’ other cases in the flaky report (Nanndish; seen while proving BUG-382, 2026-11-08): with one busy loop per core plus Docker and other sessions (100% CPU), “keeps the selection across the page, the sort and the page size” (both
AccountListScreen.selection.test.tsxandLeadsScreen.selection.test.tsx) hit the 20 s test timeout in 4 of 6 runs, and “a search clears it” (both files) missed its own 1 s wait once per set: it types seven characters and waits out the search debounce, likely cause 6 oftest-strategy.md § Flaky Tests(user.paste). Only under forced load beyond the records’ limit, so no bug yet (the flaky-report rule). Done whenflaky-report.pyshows either case flaky twice in 14 days (file a bug) or the lead drops the row. -
Watch the other Leads list cases that pick Sweden from Country (Biseni; raised by
frontend-reviewer’s review of BUG-372’s fix, forge-et-fe#490, 2026-11-11): five moreforge-et-fesrc/features/leads/LeadsScreen.test.tsxcases choose Sweden from the Country filter, whose popover renders about 250 countries (about 1.5 s a pick on an idle machine), without testing Country — the skeleton rows on a change, “a failed refresh keeps the rows, marked stale, with Retry”, and three focus cases on a list re-read after a failure (choose(user, 'Country', 'Sweden')onmain). Under two busy loops per core the three focus cases were the file’s slowest, about 8 s against the 20 s timeout; they passed in all three full runs that proved BUG-372. “a failed refresh keeps the rows…” is the Country filter’s own failed refresh and keeps Country; the other four could use a status pill or Lead Source, as BUG-372’s fix did. No bug yet (the flaky-report rule). Done whenflaky-report.pyshows one of them flaky twice in 14 days (file a bug, fixed as BUG-372 was) or they are changed. -
Decide whether design-system §9 states the card-heading count rule (Tharinda; raised by
frontend-revieweron forge-et-fe#447, 2026-11-08): the rule that a count in a card heading reads in parentheses, in its own span, in the secondary text colour, is written only in F1-d Rev 13, F3-a Rev 6 and F1-e FR-P2, so the next card has no rule to follow; the same.countstyle is copied in four CSS modules (AccountContactsPanel,RecordFollowUpsPanel,RecordNotesSection,RoleHoldersCard). Decide whether §9 gets the line, and whether it covers the Activities page’s group headings (Overdue 3, a Badge in anh2above a card); a sharedHeadingCountinsrc/uican follow. Done when §9 says it or the row is marked as not needed. -
Merge the browser suite’s duplicate keyboard and second-session helpers into
e2e/support/(Nanndish, owner of F3-a; raised by the frontend review of forge-et-fe#445, 2026-11-07, Nit 9): F3-a WI-4.1 addedchooseByKeyboard,keyToandwithUserPage/openUserPageine2e/browser/activities/support.ts, beside existing helpers for the same jobs —chooseByKeyboardine2e/browser/leads/support.ts,keyToine2e/browser/leads/follow-up-support.tsande2e/browser/pricing/round-support.ts, andsignedInPageinbriefing-support.tsande2e/browser/email/support.ts. Merge each into one shared helper ine2e/support/(WI-4.1’skeyTore-picks its direction on every press, which the others may want), and move the callers. Done when each job has one helper. -
Align the PRD capability matrix’s Activity row and the F3-a spec’s FR-A3 with PRD Rev 86 (Nanndish, owner of F3-a; raised by the fresh-context review of PRD Rev 86, 2026-11-07):
.forge/project-prd.md’s capability-matrix row “Activity (Task) / Calendar” gives the Sales Rep “C, R, U (own); mark done” with no delete and no assign, while F3-a’s FR-P1, FR-A6, FR-A9 and FR-A10 let any holder of the record’smanagekey who reaches it delete and assign, and “(own)” no longer fits an assigned follow-up; and the F3-a spec’s FR-A3 still quotes the PRD’s oldoverdue | today | upcoming → donelifecycle. Both reword approved text, so each needs a revision with its fresh-context review. Done when the matrix row and FR-A3 are revised, or the lead decides to leave them. -
Draw F3-a wave 3’s follow-ups in the current CRM prototype (Nanndish, owner of F3-a; raised by the frontend review of forge-et-fe#431, 2026-11-06):
.forge/design/prototypes/crm-current/crm-prototype.dc.htmlhas no Follow-ups card. (Add follow-up in both headers, before the primary action, was drawn with design-system Rev 110 on 2026-11-11.) Draw the cards in the built look: on the account page, the Follow-ups card, headed Follow-ups (n) (F3-a Rev 6), first in the Overview side column (rows with a title and Mark done, Show completed, the empty line No open follow-ups) and Add follow-up in the header actions before Send this week’s prices, the last button (design-system Rev 110); on the lead view, the same card after the summary and priority places in the Overview main column, and Add follow-up before Approve & send, the last button. Wave 3’s PR did not draw them because the rule that asks for it (.claude/rules/plans.md § Required Sections, 2026-11-05) came after WI-3.1’s plan was approved. Done when both screens in the current prototype matchmainafter the wave merges. -
Point the nightly run’s tracking addresses at its own backend (Tharinda, the lead, who runs the nightly job; raised by F20-g WI-1.8’s plan, 2026-11-11): F20-g wave 1 is on
main(since 2026-11-11), and itslocal | testsettings setet.mailbox.tracking-base-urltohttp://localhost:8080, but.claude/lib/nightly/run-nightly.pyboots the backend onBACKEND_PORT18080, so a tracked send’s links and image on the night’s backend point at the wrong port. The tests request tracking addresses by path, so they pass either way. Add--et.mailbox.tracking-base-url=http://localhost:18080toBACKEND_ARGS. Done whenrun-nightly.pycarries it. -
Draw F20-g Rev 4’s Scheduled changes in the current CRM prototype (Dinil, owner of F20-g; 2026-11-16, F20-g spec Rev 4 NFR-5, design-system Rev 130):
.forge/design/prototypes/crm-current/does not draw/email/scheduledin the inbox’s frame (the Email header with Mailbox Settings, Refresh and Compose, the empty data-bar row, the folder pane’s own card beside the table card), the Scheduled button in the not-connected prompt on/email(icon, name, waiting count and failed dot, after Reconnect), or the not-connected card above the list on/email/scheduled(Your mailbox is not connected. Reconnect it to send these emails. and the next-step button), whichforge-et-febranchfeature/F20-g-scheduled-accessbuilds. Draw them in the built look. Done when the prototype shows them as built. -
Decide whether the record page’s underlined tab bar moves onto the shadcn
Tabslinevariant (Tharinda, the lead; raised by the forge-et-fe#435 review, 2026-11-06): design-system Rev 101 adds thelinevariant for the Email list’s Focused / Other strip: list gapspace/6, tab block paddingspace/5, no inline padding, the selected tab’s bar drawn as aborder-width/selectedbottom border on the strip’s rule.forge-et-fesrc/ui/recordPage/RecordPage.module.css(its tab bar) draws another underline over the pill variant: gapspace/2, tab paddingspace/3byspace/4, the bar an inset shadow. Two looks for one pattern break design-system One look per pattern (L-254). Done when eitherRecordPageusesvariant="line"and its override is deleted, or a design-system revision records the record page’s bar as its own look, with its values. -
Split
LocalLiveMailboxafter F45 — now part of feature F46 (since 2026-11-14,.forge/specs/F46-split-files-over-size-limit-spec.md; Tharinda, the lead; decided at F45’s spec approval, 2026-11-12, its Q-2, and narrowed the same day by the file-size check’s Rev 1):forge-et-bemailbox/LocalLiveMailbox.java(1,085 lines, 889 code lines onbb9fe28e) is over the project’s limit of 800 code lines (.forge/design/harness/file-size-check.mdD-1) and is left out of F45, which splits only the three classes the lead named.mailbox/MailboxProvider.java, first named here, is 1,660 lines but 590 code lines — mostly the documented records the seam returns — so it is under the limit and is not split. F45 is done (2026-11-21), so the split can be planned now, with no change in behaviour, starting from F45’s helpers; it lowers and then removes the file’s entry inFileSizeSpec(design D-4). Done when the file is under the limit, or the lead marks this won’t-do. -
Split the busiest large CRM screens now, as a work item of its own — now part of feature F46 (since 2026-11-14, F46’s first wave,
.forge/specs/F46-split-files-over-size-limit-spec.md; Tharinda, the lead; raised by the cross-project code review of 2026-11-08; the lead decided on 2026-11-12 to do it now rather than inside each screen’s shadcn move-over PR):forge-et-fesrc/features/email/compose/ComposeEmailDialog.tsx(2,090 lines onmainon 2026-11-12, up from 1,636 on 2026-11-08; 40 non-merge commits in the 60 days to 2026-11-08),src/features/accounts/AccountFormPage.tsx(1,422 lines) andsrc/features/leads/LeadFormDialog.tsx(1,291 lines). No change is made without an approved plan, so the split is planned and reviewed first (/forge-deliver), with no change in behaviour. forge-et-fe#442, forge-et-fe#476 and forge-et-fe#508 (Tharinda, the compose dialog and the lead form) have merged; on 2026-11-12 one open draft still changed these files, forge-et-fe#493 (Nanndish, all three). The order, decided by the lead on 2026-11-12: forge-et-fe#508, a small fix, merges before the split starts (done, 2026-11-12); the split starts onmainafter #493 merges — never while #493 still changes the three screens, because each later push would then conflict with the moved code. #493 merged on 2026-11-13, so the split can start; on 2026-11-14 no open PR changed the three screens. On 2026-11-12 the lead decided not to involve Nanndish: #493 merges as it is, with no file-size work asked of him, and this split, with the row “Split the other source files over the limit”, does all the refactoring after it merges. All three screens are insrc/app/shadcnBackfill.jsonand.forge/design/ui/shadcn-backfill.md: a new file that receives code using an oldersrc/uicomponent failsshadcnOnly.test.tsunless it is listed there, and adding an entry isfrontend-reviewer’s decision (AD #39), so the plan says for each new file whether it moves to shadcn or takes an entry. The project’s file-size limit applies (800 code lines per file,.forge/design/harness/file-size-check.md): on26e4cd42the three hold 1,751, 1,151 and 1,058 code lines, and every file the split leaves must be at or under 800. Their current counts and progress are in.forge/design/harness/file-size-progress.md. Done when each screen is under 800 code lines, or the lead marks this won’t-do. -
Decide whether BUG-291’s held-loader tests need a flaky-test bug (Nanndish; raised by F3-a wave 2’s delta fixes, 2026-11-03): two of the regression tests forge-et-fe#372 added each failed once under load and passed alone, on Windows.
pricingUrlState.held.test.tsxwas flaky in forge-et-fe#371’s test record onea7d38e3.activitiesUrlState.held.test.tsx(“All in reach, then the Completed view before the first has rendered, keeps both”,expected 'idle' to be 'loading'in awaitFor) failed once in a targetedvitest run src/features/activitieson forge-et-fe#353’s2ad9fd54, and passed alone and in a re-run of the folder. Check the weekly flaky report (python3 .claude/lib/flaky-report.py); if either recurs, file the bug (.forge/test-strategy.md § Flaky Tests). -
Build the file-size check (Tharinda, the lead; asked for on 2026-11-12, after F45 and the CRM-screen split were decided, so that no source file grows that large again): the design is
.forge/design/harness/file-size-check.md(approved by the lead on 2026-11-12, revised the same day as Rev 1) — a limit of 800 code lines per source file in every app repo, with no exceptions, enforced by a unit test in each app repo with a list of the files still over it, each with a planned split, that only empties; a paragraph incoding-standards.md§1; and checks inplan-reviewer,backend-reviewerandfrontend-reviewer. Build it as its D-8 sets out: one harness PR, then one PR per app repo. Theforge-et-becheck PR merges after F45’s wave 1 (the lead, at F45’s Decomposition Plan approval, 2026-11-12), so its list is generated with F45’s three classes already under the limit (.forge/plans/F45/F45-decomposition-plan.mdNotes). Open PRs that would fail the check (scanned 2026-11-12 against each PR’s head;forge-et-beandforge-et-pwahad none): forge-et-fe#493 (Nanndish) takesfeatures/leads/LeadNotesSection.tsxfrom 489 to 857 code lines, a new file over the limit, and growsfeatures/notesAndTags/RecordNotesSection.tsxfrom 903 to 911; forge-et-fe#509 and forge-et-fe#513 (Tharinda, F32-d) growfeatures/accounts/AccountListScreen.tsxfrom 852 to 909 and 933. The lead decided on 2026-11-12 that these PRs may merge as they are, and the files they take over or grow are split by the row “Split the other source files over the limit”, which names them (design D-8 item 2). The same scan found #493 shrinking six listed screens (ComposeEmailDialog.tsxfrom 1,751 to 1,422 code lines). State on 2026-11-14: #493 merged on 2026-11-13, so theforge-et-fecheck PR no longer waits on it; as merged it did not shrink the screens (ComposeEmailDialog.tsx1,748 code lines onmain0f99a28f,AccountFormPage.tsxup to 1,170). #509 and #513 closed unmerged when their work merged in the F32-d wave PR forge-et-fe#522. Checked 2026-11-18: theforge-et-feandforge-et-pwachecks are onmain(src/app/fileSize.test.tsin each); theforge-et-becheck is not, and still follows F45 wave 1. Checked 2026-11-21: F45’s wave 1 merged on 2026-11-20 (forge-et-be#418), so theforge-et-becheck PR no longer waits on it. None of the four PRs (harness, three app repos) has been opened. The lead decided on 2026-11-14 to build the harness PR and theforge-et-feandforge-et-pwacheck PRs now, and to cover all nine frontend files over the limit (the CRM-screen split row and this row’s frontend files) with one new feature through/forge-deliver: wave 1 the three screens, wave 2 the other six. Every file over the limit, with its count and state, is in.forge/design/harness/file-size-progress.md. The day each app repo’s check merges, tell the team, and the author of every open PR in that repo that changes a source file mergesmainin and records again before merging (design D-9). Done when the three tests are onmain, or the lead marks this won’t-do. -
Split the other source files over the limit — now feature F46 (since 2026-11-14,
.forge/specs/F46-split-files-over-size-limit-spec.md, with the two rows above; Tharinda, the lead; decided on 2026-11-12 with the file-size check’s Rev 1, which holds every file to 800 code lines): a scan ofmainon 2026-11-12 (forge-et-bebb9fe28e,forge-et-fe26e4cd42) found 13 source files over the limit; 7 have their split planned elsewhere (F45, theLocalLiveMailboxrow, the CRM-screen split), and this row covers the other 6, plus one from an open PR (below):forge-et-bemailbox/service/EmailSendService.java(1,064 code lines), andforge-et-fefeatures/attachments/RecordAttachmentsPanel.tsx(1,018),features/activities/FollowUpEditBody.tsx(961),features/notesAndTags/RecordNotesSection.tsx(903),features/accounts/AccountListScreen.tsx(852) andfeatures/email/templates/TemplateEditorForm.tsx(805). Each is split by job with no change in behaviour, planned and reviewed through/forge-deliverlike F45:EmailSendServicenow that F45, which changed the same mailbox package, is done (2026-11-21), and the frontend files after the CRM-screen split, so the same reviewers see one pattern. Before planning, check each file’s open PRs (gh pr diff <N> --name-only) and its code-line count again. Each lowers and then removes its entry in the file-size test (design D-4). A seventh file, from an open PR: forge-et-fe#493 (Nanndish, open on 2026-11-12) takesfeatures/leads/LeadNotesSection.tsxfrom 489 to 857 code lines. The lead decided the same day that #493 may merge as it is, and this row splits the file once it does. The same holds for the growth inRecordNotesSection.tsx(#493) andAccountListScreen.tsx(forge-et-fe#509, forge-et-fe#513), which are already in this row. Those 7 files and the 7 planned elsewhere are all 14 files that break the limit, counted onmainand on the open PRs on 2026-11-12. A file that crosses the limit later is added to this row the day it is found, until the file-size check is onmainand refuses it (design D-4 rule 1). State on 2026-11-14: #493 merged on 2026-11-13; onmainLeadNotesSection.tsxis 857 code lines,RecordNotesSection.tsx911 andAccountListScreen.tsx933 (F32-d, forge-et-fe#522). Open forge-et-be#424 (Dinil, BUG-418) growsEmailSendService.javafrom 1,064 to 1,066. No new file has crossed the limit since 2026-11-12. The counts and progress of all 14 files are kept in.forge/design/harness/file-size-progress.md. Done when all seven are under 800 code lines, or the lead marks this won’t-do. -
Decide whether every dialog on the Rev 90 frame shortens its button labels below
bp/tablet(Nanndish; raised by the review of BUG-277’s fix in forge-et-fe#442, 2026-11-08): since BUG-277’s fix, below 768px the template editor shows Delete, Save, Create and, after a deleted template, Back, and the follow-up dialogs show Save, each keeping its full name for screen readers, because the full labels took a third row or spilled past the dialog at 320px. The other dialogs on the frame (Edit contact, New account, the role and user editors) keep Save Contact, Create Account, Save Role and Save User, so two looks for one pattern show at 320px (design-system One look per pattern, L-254). Either record the rule in a design-system revision (shorten only where the full labels would take a third row or spill), or shorten every framed dialog’s labels belowbp/tablet. Done when the design system states the rule and the dialogs follow it. -
Recase the row-selection labels that forge-et-fe#442 left in sentence case (Nanndish, owner of #442; raised by the review of #442’s merge of
main, 2026-11-08): Select all 60 accounts (src/ui/components/RowSelection/useRowSelection.ts) is a button that acts, so design-system §9 (Rev 90) makes it Title Case, and it shows on the account list, the lead list and the Deactivate dialog’s Records section; its count also has no thousands separator (“Select all 2350 leads”). The bulk bar’s name reads Account actions inOwnedRecordsList.tsxbut Account Actions and Lead Actions on the lists. (The search boxes’ names, once listed here, follow design-system Rev 107 since forge-et-fe#436: a sentence-casearia-labeland a Title Case placeholder, held bysrc/app/listSearchNaming.test.ts.) Done when the Select all labels and the bulk bars’ names read in Title Case with a formatted count, with the tests that assert them (*.selection.test.tsx,OwnedRecordsList.test.tsx,BulkActionBar.test.tsx, the Change Owner browser specs). -
State in design-system §9 the casing of a tooltip that repeats an icon-only button’s name (Nanndish; raised by the same review, 2026-11-08): §9 lists tooltips as sentence case, but a tooltip that repeats an icon button’s name, which §9 makes Title Case, now takes the name’s casing (Notes and Tags on a contact row, forge-et-fe#442; Collapse Folders in the Email folder column). Done when a design-system revision states the rule in one line.
-
Check whether
leadResearchComposition.test.tsx’s AC-28 rows need a flaky-test bug (Nanndish; raised by the review of forge-et-fe#413, 2026-11-05): “while the panel is pending” timed out at 20 s in 3 of about 13 local runs on Windows at 100% CPU, and “while the first read has failed” once did not find the taglistboxwithin the default second. Both passed on re-runs. Check the weekly flaky report (python3 .claude/lib/flaky-report.py); if either recurs, file the bug (.forge/test-strategy.md § Flaky Tests). -
Check whether the sharing card’s announcements are spoken while the Access dialog is open (Nanndish; raised by
frontend-reviewerin the review of forge-et-fe#408, 2026-11-05): the card’s live line (sharing-panel-live, “Shared with …”, “Removed …'s access”) sits outside the open Radix dialog, which hides the rest of the page from screen readers, so its announcements are most likely never spoken. Older than that PR. Confirm with a screen reader or the accessibility tree; if confirmed, file it as a bug. Done when checked and, if needed, filed. -
Decide whether
test-strategy.md § Flaky Testsnames the page budget as an exception to “never lengthen a timeout” (Tharinda, the lead; raised by BUG-329’s fix, 2026-11-03): BUG-329’s fix gives two waits after a navigation five seconds instead of Testing Library’s default second, asLeadsScreenActions.test.tsx’sPAGE_TIMEOUTand the composition tests’LAZY_PANELalready do. There was no race to fix: React Router renders the navigation in a transition, and the list screen took more than a second to render under load. L-256 states when such a budget is right (an outcome that is certain, proved by holding the slow step); § Flaky Tests still says never to lengthen a timeout. Either add the exception there, with a pointer to L-256, or say how such a wait should be fixed instead. -
Decide whether the client-facing PRD’s Sales Rep line changes “owns … activities” (Tharinda, the lead; raised by the F3-a-WI-4.1 plan review, 2026-10-27): F3-a makes a follow-up assigned, not owned (F3-a spec decision 8, FR-H1), and WI-4.1’s PRD revision rewords the internal PRD’s Sales Rep row to match.
.forge/client/client-facing-prd.md(about line 43) still says “Owns their leads, accounts, deals, quotes, activities”; it is issued to Empire, so whether it changes, and how (an edit or a post-issue note,.claude/rules/client-docs.md), is the lead’s call. Done when it is decided and, if changed, made. -
Check in Chrome that the mailbox connect-failed banner leaves on a back/forward return (Tharinda; raised by the delta review of forge-et-fe#382, 2026-11-02, L-109): forge-et-fe#382’s
14b35050clears the banner when Chrome restores a page from its back/forward cache, and Playwright’s Chromium disables that cache, so no automated test runs it in a browser. On the built app (pnpm buildthenpnpm e2e:app, notpnpm dev, whose live-reload socket keeps pages out of the cache), make a connect fail (the stub’sPOST /api/v1/mailbox/stub-connect-outcome, or laptop mode signed out of Microsoft), leave the CRM by a full-page navigation to another site, press Back: the banner is gone and the check was read again. Record the result on the PR or in WI-1.3’s plan, then remove this row. -
Stop
AuthorizationOutcomeprinting its redirect URL (Tharinda; raised by the delta review of forge-et-be#349, 2026-11-02, a Nit in F20-a code the PR did not change):forge-et-besrc/main/java/com/empireteas/auth/oauth/AuthorizationOutcome.java:15is a record whose defaulttoString()printsredirectUrl, which since F20-i carries the validatedreturnTopath and query — a CRM page’s query can hold a buyer’s name (coding-standards.md§8.6). No log line prints it today. OverridetoString()to printredirectUrl=<redacted>, as forge-et-be#349’s0724f547did forAuthorizationStartandPendingConnect; a small PR, or the next F20-i wave’s backend work item. -
Settle the small points left by the UI revamp preview (Nanndish; raised by its test updates and its review, 2026-10-30): (1) settled 2026-11-08 by Tharinda (BUG-306): the code follows F1-e and shows
—, forge-et-fe#446; (2)AccountFormPagekeeps a page layout, used when no dialog frame is passed, that nothing renders since editing became a dialog — remove it; (3) the lead’s Shared with card shows the owner as “—”, with a “—” circle, when the owner’s display name is not in the read; (4)UserActivationCard’sblockedReasonprop and its comment describe a block while the editor has unsaved edits, which the modal Edit user dialog makes impossible (the route always passesundefined) — remove both; (5)getRolePermissionKeys(src/features/users/usersApi.ts) readsGET /api/v1/roles/{id}beside the roles feature’s own read — move the read tosrc/liband use it from both (frontend-revieweron forge-et-fe#364, I13, left as a separate refactor). -
Decide whether a short window at 768px or wider should scroll the account page’s Buying standards tab (Tharinda, the lead; raised by BUG-294’s fix, 2026-11-02): from
bp/tabletthe standards table fills the page’s remaining height and scrolls in its card; on a window shorter than the table region’s floor the card collapses as it did at 320 before BUG-294. No supported size (768 × 1024 and larger) is that short. The CSS-only alternative grows the card to the bottom of<main>even with a few rows. Decide whether short heights must be covered; if so, file a bug. Done when decided. -
Add browser checks for the clauses the UI revamp preview’s spec revisions added (Nanndish; raised by the review of forge-et-fe#371, 2026-11-02): forge-et-fe#371 moved the existing browser tests onto the revamped screens, but no test, browser or component, checks these new clauses: (1) F1-d AC-4 and AC-6 (Rev 10), F1-b FR-C4 (Rev 11): the account summary strip’s primary contact (or None yet), the chosen tab kept in the address as
?tab=and a?tab=link opening that tab, and archiving and restoring from the header’s Archive account / Restore account (the only page-level archive path now); (2) F1-e AC-22 (Rev 9) and F11-f FR-V4 and AC-24 (Rev 5): Cancel with text asks Discard this note?, a body over four lines shows Show more, and frombp/laptopthe notes scroll inside their card, on the account and lead pages; (3) F1-c AC-11, AC-12 and AC-17 (Rev 3): the timeline’s filter chips (filtering and counts, a chip with no entries hidden, Show older entries with a chip chosen and where focus goes, the Noamong the entries shown. state, keyboard use), the entry list scrolling in its own focusable region frombp/laptop, and theweeks without contact line. Where to add them:e2e/browser/accounts/account-detail.spec.ts,account-notes.spec.ts,leads/lead-detail-notes.spec.ts,notes-tags-layout.spec.ts,account-timeline.spec.tsand an unmarked test besideaccount-timeline-a11y.spec.ts’s BUG-307 mark. The spec revisions’ “Not met” lines for these clauses are ticked when the checks pass. Done when they are merged. -
Decide five spec clauses the UI revamp preview’s code does not meet, which no browser test checks (Tharinda, the lead; raised by the update of
main’s browser specs after forge-et-fe#364, 2026-11-01): for each, the code or the spec changes. (1) F1-d FR-C7 and AC-16 (Rev 10): while a contact row’s write runs its reason replaces the menu as a status line; the code keeps the ⋯ trigger,aria-disabledand busy, with the reason in a visually hidden status line (forge-et-fesrc/features/accounts/ContactRow.tsx). (2) F5-g AC-15: the Shared with summary sentence attype/body; the code uses--et-type-secondary-*(src/features/sharing/SharingPanel.module.css.summaryText). (3) F5-g FR-D4 still says “the panel’s Share… turns to its disabled state”, which Rev 1 removed elsewhere — reword the spec. (4) F11-f FR-V3: a failed tag read shows its error with Retry inside the control; the code shows them under the tag list, outside the folded + Add tag chip (src/features/leads/LeadTagsSection.tsx). (5) design-system Rev 77 lists removing a note among the AlertDialog confirmations, while F1-e Rev 9 confirms inside the note (the account page does; the lead page usesConfirmDialog). The clauses a browser test does check are bugs instead (BUG-299, closed wontfix by F5-d Rev 9; BUG-300, BUG-304, BUG-305, BUG-306). Done when each is decided and the code or the spec changed. -
Let a harness-only work item hand any
.claude/edit to the orchestrator (Tharinda; raised by F38-i-WI-2.1, 2026-10-29): a harness-only work item may not edit.claude/**, and the hand-over through its plan’s## Notesis written for.claude/CLAUDE.mdonly (.forge/design/harness/harness-only-sub-work-items.mdD-3,/forge-deliverStage 12 “a wave with a harness-onlysubWI” step 2). F38-i-WI-2.1 changes the release rules (.claude/rules/release-and-branch-protection.md,.claude/commands/forge-release.md), so its plan (D-10) hands those two edits to the orchestrator on the wave branch, extending that route. Widen D-3, Stage 12 step 2,forge-worktree-upstep 2’s plan-defect rule and/forge-pre-pr-review’s harness mode to allow a## Noteshand-over of any.claude/file, or decide that such work is never a work item. Done when those files agree. -
Decide whether the app-level pages’ titles follow the record-title size (Tharinda; raised by
frontend-revieweron forge-et-fe#356’s merge with main, 2026-10-29): the lead chose on 2026-10-29 that every screen title is design-system Rev 74’s record-title size, and BUG-163’s fix applies it to every title insrc/featuresthroughPageTitle. Since design-system Rev 109 (2026-11-10) that size istype/title’s 24px at theh1’s bold (700), so the two now differ only in weight. Four titles outsidesrc/featuresstill usetype/titleat its own weight (24px, semibold):forge-et-fesrc/app/NotFoundRoute.tsx,src/app/FailureSurface.tsx,src/app/RouteErrorElement.tsxandsrc/app/NoAccessState.module.css;src/app/pageTitleOnly.test.tsscanssrc/featuresonly. Decide whether they follow; if they do, move them toPageTitle(or the same rule) and widen the scan. Done when they are moved or marked as staying. -
Close F38-e wave 3 (manual): re-run
/forge-deliver F38-eas soon as the record is pushed (Tharinda Rajapaksha, handed off 2026-10-26; this row added 2026-11-20, because the hand-off came before M1 wrote a row like this): the run sheets are.forge/plans/F38-e/F38-e-WI-3.1-plan.md(the alert path),F38-e-WI-3.2-plan.md(the SES probe) andF38-e-WI-3.3-plan.md(one request by its traceId, and the closing sweep), and each record is.forge/plans/F38-e/F38-e-WI-3.<n>-record.mdonfeature/F38-e-wave-3. Every AWS write in them is approved by the lead at the time. F38-f’s wave 3 waits on this wave’s merge (its gate). Done when the wave PR is open (M3 removes this row). (Update 2026-11-21: all three sheets were run on 2026-11-21 by Claude in the lead’s session and the records pushed. Two things are left before/forge-deliver F38-eM2: WI-3.1 R-2 (the no-change plan) fails on BUG-532 until itsforge-et-iaacfix merges and one manual run readsNo changes.; and WI-3.1 R-4 and R-6 need the four inbox reports (the row below). WI-3.3’s closing sweep (R-4, R-5) runs after both.) -
Collect the four inbox reports for F38-e WI-3.1 (Tharinda, from Dinil, Biseni and Nanndish; raised 2026-11-21): on 2026-11-21 the topic
et-dev-alertssent two test emails, both delivered by SNS to all four subscriptions — at 20:23 UTCALARM: "et-dev-f38e-test-alarm" in US East (N. Virginia)and at 20:42 UTCAWS Notification Message(the failed deployecs-svc/7652176713461752993). Each of the four developers reports to the lead the time each email arrived and, for the alarm, itsAlarm Name,State ChangeandReason for State Changelines, and for the failed deploy, itsservice,deployment,reasonandtimelines — never anything from the email’s footer. The lead adds them to.forge/plans/F38-e/F38-e-WI-3.1-record.mdR-4 and R-6 as attempt 2. Done when the four reports are in the record. -
Close F38-f wave 2 (manual): re-run
/forge-deliver F38-fas soon as the record is pushed (Tharinda Rajapaksha, handed off 2026-10-26): the run sheets are.forge/plans/F38-f/F38-f-WI-2.1-plan.mdtoF38-f-WI-2.6-plan.md, and each record is.forge/plans/F38-f/F38-f-WI-2.<n>-record.mdonfeature/F38-f-wave-2. Done when the wave PR is open (M3 removes this row). (Update 2026-11-21: all six sheets were run on 2026-11-21 by Claude in the lead’s session and every record is pushed; WI-2.3 is withdrawn by spec Rev 7. Left before/forge-deliver F38-fM2: WI-2.1 R-6b (lead ranking, 2 of 11 callsINVALID_OUTPUT) and R-6c (the capture app’s AI features), the lead’s decisions in the rows WI-2.1 added; WI-2.5 R-1, the install (the row below).) -
Install the capture app from dev for F38-f WI-2.5 (Tharinda, or anyone with Chrome; raised 2026-11-21): the browser pane Claude used cannot install a web app, so
.forge/plans/F38-f/F38-f-WI-2.5-plan.mdsteps 2.1–2.4 and 3–5 (install in a fresh Chrome profile, read the Network panel’s(ServiceWorker)column for the/api/requests, uninstall) are still to run; about five minutes. Add them to.forge/plans/F38-f/F38-f-WI-2.5-record.mdas R-1 and as attempt 2 of R-2 and R-4. Done when R-1 reads PASS. -
Decide how Empire’s full buying-standards list reaches the CRM in Phase 1 (Tharinda; raised by the F4-d spec, 2026-10-13): OQ-27 / client question Q5 asks where the list of about 35,000 standard numbers lives. The F4-d spec (Decision 1) keeps the Pipedrive migration to organisations and people, because the list lives outside Pipedrive, and F23-b’s spreadsheet import is Phase 2. Until the list is loaded, the weekly workbook import (F26-c) reports every standard not created by hand as an unknown standard number, so only standards the Tea Room enters (acceptance criterion G1) are priced. Once Empire answers Q5, choose a Phase 1 path — a one-off load by the delivery team, a small import feature, or entry by hand — and record it. Whoever loads the list also takes on two obligations approved specs gave F4-d: F26-a’s
setvalrule for migrated standard numbers (F26-a FR-S2) and the stored format of migrated standard numbers (F26-c W-2). -
Confirm who runs the production Pipedrive runs, and as which import owner (Tharinda; raised by the F4-d spec, Q-9, 2026-10-13 — before the first load is scheduled): the spec’s default is that Kavindu Anthony runs both production runs (the first load and the final run) under his own CRM user, which makes him the import owner, with a member of the delivery team running the dry run with him; beforehand his IT Admin role grants itself
settings.manage,account.manage,account.all,lead.view,lead.allandlead.manage(PRD F4 Rev 47; AD #34, audited). Confirm or change it. -
Issue the Phase 1 acceptance criteria to Empire (Tharinda): the draft is
client/phase-1-acceptance-criteria.md. Review it, settle its open points, and send it for Empire to agree. It was due by Mon 12 Oct, when Phase 1 starts on the client timeline, becauseclient/project-timeline.md§4 says each phase’s criteria are issued at its start; its header still reads “Not yet issued” on 2026-10-16. Reword H3 and C5 before issue (raised by the F3-a spec, approved 2026-10-13, decision 5): a completed follow-up on an account appears on the account’s timeline, and a follow-up on a lead writes no timeline row, because leads have no timeline. The draft’s H3 still says follow-ups appear on “the record’s timeline”, and C5 lists “follow-up tasks” on the account timeline where only completed ones appear; decide whether the change goes to Empire. Before issuing, also reword draft criterion I3 to match the F4-d spec (FR-R3, FR-W3): imported records belong at first to the import owner, normally the person who ran the import, and a Pipedrive organisation matched to an account already in the CRM keeps that account’s owner. Decided by the lead, 2026-11-19: the criteria are issued at the hand-over meeting (one meeting with UAT drop 1, about Mon 30 Nov at Empire’s office). Before then the delivery team (1) rewords H3, C5 and I3 and fixes whatever else in the draft is wrong now; (2) adds criteria for every completed Phase 1 feature the draft does not cover; and (3) adds each Phase 2 feature built early in Phase 1 once it is done (the row “The Phase 2 features started early in Phase 1”). The lead or a teammate sends it by hand; nothing is sent automatically. Done when it is issued. Update 2026-11-19: (1) and (2) are done in the draft of 19 November — C5, H1, H3 and I3 match the built CRM; E1 names the person-search switch (F11-i); C8 (several addresses per contact, F20-l) and F10 (email attachments, F20-h) are added; “How the criteria are tested” leaves the UAT data and sign-ins to Kavindu’s answer and lists F3, F6 and F7 for the second UAT delivery. F8, F9, G9 and Sections K, M and N already covered F20-g, F37-a, F34-c, F33-a and F32-d. Left before issue: (3) as each started-early feature finishes; the “Who signs in” and “With what data” lines once Kavindu answers; the second-delivery list re-read just before the meeting; and E8 depends on the row “Summarise migrated Pipedrive leads from their linked past email”. -
Summarise migrated Pipedrive leads from their linked past email, so acceptance criterion E8 holds (Tharinda to assign; the lead’s decision of 2026-11-19, replacing the row “Confirm E3’s rewording and choose E8’s path for migrated leads”, raised by the F11-d spec, decisions 4 and 7, FR-H2). E8 said every migrated Pipedrive lead is researched, ranked and summarised. Update 2026-11-20 (F11-j spec, approved by Tharinda): E8 now reads “every open lead migrated from Pipedrive has been researched and ranked, and every one with email from the buyer linked to it has been summarised from that email” (F11-j D-10). An open migrated lead with no inquiry text and at least one received linked email is summarised; a summarised lead that is later archived keeps its summary on screen but is not summarised again (D-1). Migrated leads carry no inquiry text (F4-a FR-M3, PRD Rev 44), so F11-d gives them no summary today (they still get the duplicate rule and a spam check on their fields). The lead chose to build it: once F20-j links a migrated lead’s past email, its summary is written from that email. It is F11-j (registered 2026-11-19, Phase 1, unassigned, blocked by F20-f): a feature of its own, not part of F20-j (Tharinda’s decision, F20-j spec D-6), because it needs only linked email, which F20-f gives too, and its AI calls need their own pacing (F42, F43); it revises F11-d FR-H2. Needed before E8 is shown at the first Pipedrive load. E3’s wording (a lead repeating an open lead is tagged; one from a known customer contact is shown with that customer, ready to link) was confirmed by the lead the same day. Done when migrated leads with linked past email get a summary on
main. -
Revise F20-j § Constraints “No AI work” for F11-j’s summaries (Tharinda, owner of F20-j; raised by the F11-j spec FR-H2, recorded 2026-11-20 at F11-j’s decomposition): F20-j’s scans still make no AI call. But linking a received email to a migrated lead queues F11-j’s summary (F11-j FR-J1), so a scan after a Pipedrive load queues one summary for each migrated lead it links mail to, paced at the background share (F11-j NFR-6, D-7). Write it in forge-et-harness#1579 while that spec PR is open, or as a dated revision of the F20-j spec after it merges. Done when the F20-j spec carries the revision.
-
F11-j: choose the principal that delivers the LINKING test mail (whoever implements F11-j-WI-2.2; the lead’s decision of 2026-11-21 at WI-3.1’s plan review, pass 2, finding I2). WI-2.2’s
ensureLinkingDelivered, WI-2.3’s browser check B1 and WI-3.1’s journey deliver thestub-notifyLINKINGfixture as the default all-permissions user and reset its live mailbox withPOST /api/v1/mailbox/stub-mailbox. That user’smailbox_connectionis owned byforge-et-fe’se2e/browser/mailbox/connect-disconnect.spec.ts(e2e/fixtures/authStates.ts), which leaves it disconnected, so in the nightly run the reset can answer404 MAILBOX_NOT_CONNECTED, and connecting first would change the row that spec asserts on. WI-2.2’s subtask 1 decides, on the merged base after wave 1, which principal delivers (preferably one no mailbox browser file claims, signing in on its own as thee2e/browser/email/*specs do, once linking is confirmed not to depend on which mailbox receives the mail) and records it in its plan’s Notes; WI-2.3 and WI-3.1 follow it. Done when WI-2.2’s plan records the choice and its helper uses it. -
Apply the UAT and production overlays with the
scheduled-email-sendschedule only together with a backend that includes F20-g wave 1 (Tharinda, owner of F38-h and F38-i; raised by F20-g-WI-2.1, 2026-11-09; F20-g Decomposition Plan## Notesitem 2). forge-et-iaac#19 adds the schedule tobase/scheduler, whose job list is the same in every environment, so the next apply ofoverlays/uat(#11) oroverlays/prod(#17) from adevthat holds it creates the schedule there. A backend without F20-g wave 1 refuses the unknown job kind every minute: each firing is received three times (the queue’smaxReceiveCount) and then dead-lettered, and is logged as a failed job, which production’sERRORlog alarm reports (register §B F11-g row 7). A release that carries wave 1 needs that apply, or its scheduled emails stay Waiting to Send. Before each UAT round’s apply and production’s first apply, check that the backend release includes F20-g wave 1, or decide what to do (a later apply, or abase/scheduleroption to leave a job out). The register’s §B F20-g row (a) has the read that proves it. Closes when both overlays have been applied with the schedule beside a backend that has wave 1. Update 2026-11-13: theuatoverlay was applied on 13 Nov withscheduled-email-sendENABLEDand no backend yet (F38-h spec Rev 8); F20-g wave 1 is onforge-et-bemain(#406), so the round’s-rctag needs only to be cut frommainafter it. The UAT half closes when that backend runs there. Update 2026-11-17: the UAT half is closed.v0.1.0-rc1(forge-et-bef392e8e, which has forge-et-be#406) has run on UAT since 16 Nov, beside the schedule (.forge/plans/F38-h/F38-h-WI-5.1-record.mdR-6). The production half stays open. -
Apply the UAT and production overlays with the
price-broadcast-sendschedule only together with a backend that includes F37-a wave 1 (Tharinda, owner of F38-h and F38-i; raised by F37-a-WI-2.1, 2026-11-15; F37-a Decomposition Plan## Notesitem 1). forge-et-iaac#20 adds the schedule tobase/scheduler, whose job list is the same in every environment, so the next apply ofoverlays/uat(uat.yaml; the instance applied on 13 Nov) oroverlays/prod(#17) from adevthat holds it creates the schedule there. A backend without F37-a wave 1 refuses the unknown job kind every minute: each firing is received three times and then dead-lettered, and is logged as a failed job, which production’sERRORlog alarm reports (register §B F11-g row 7); on UAT, ticks held inet-uat-schedulerbefore the backend starts are replayed into it. F37-a closed on 2026-11-18 with wave 1 onmain(forge-et-be#446,0a8ede23), so a release tag cut frommainafter that includes it; a tag cut before it, such asv0.1.0-rc1, does not. Before each UAT round’s apply and production’s first apply after #20 merges, check that the backend release includes F37-a wave 1, or decide what to do (a later apply, or abase/scheduleroption to leave a job out). The register’s §B F37-a row (a) has the read that proves it. Closes when both overlays have been applied with the schedule beside a backend that has wave 1. Update 2026-11-17: #20 is ondev, and UAT has not been applied since. UAT runsv0.1.0-rc1, which does not have F37-a wave 1. The lead decided that day to leave UAT as it is until the next-rctag. That tag is cut from amainthat has F37-a wave 1, anduat.yamlapply runs before its deploy (the row “Bring UAT up toforge-et-iaacdevwith the next-rctag”). Update 2026-11-19: the apply ran on 19 Nov (run 44464377454), before the tag, so UAT’sv0.1.0-rc1backend now dead-letters the schedule’s ticks every minute; the decision is in that row. The UAT half closes when the next-rcruns there. -
Apply the UAT and production overlays with the
bulk-email-sendschedule only together with a backend that includes F37-b wave 1 (Tharinda, owner of F38-h and F38-i; raised by F37-b-WI-2.1, 2026-11-19; F37-b Decomposition Plan## Notesitem 1). forge-et-iaac#29 adds the schedule tobase/scheduler, whose job list is the same in every environment, so the next apply ofoverlays/uat(uat.yaml, run by hand fromdev) oroverlays/prod(#17) from adevthat holds it creates the schedule there. A backend without F37-b wave 1 refuses the unknown job kind every minute: each firing is received three times, then dead-lettered and logged as a failed job, which production’sERRORlog alarm reports (register §B F11-g row 7). F37-b started early in Phase 1 and moves to Phase 2 if not done by the end of Phase 1’s UAT round (trackernotes), so a Phase 1 release tag may not include wave 1; the F37-a row above shows auat.yamlapply reaching UAT before its backend had the handler. Before eachuat.yamlapply and production’s first apply after #29 merges, check that the backend release there includes F37-b wave 1, or decide what to do (a later apply, or abase/scheduleroption to leave a job out). Register §B F37-b row (a) has the read that proves it. Closes when both overlays have been applied with the schedule beside a backend that has wave 1. -
Apply the UAT and production overlays with the
contact-sync-sweepschedule only together with a backend that includes F22 wave 1 (Tharinda, owner of F38-h and F38-i; raised by F22-WI-2.1, 2026-11-21; F22 Decomposition Plan## Notesitem 2). forge-et-iaac#30 adds the schedule tobase/scheduler, whose job list is the same in every environment, so the next apply ofoverlays/uat(uat.yaml, run by hand fromdev) oroverlays/prod(#17) from adevthat holds it creates the schedule there. A backend without F22 wave 1 refuses the unknown job kind every hour: each firing is received three times, then dead-lettered and logged as a failed job, which production’sERRORlog alarm reports (register §B F11-g row 7). A release that carries wave 1 needs that apply, or nobody’s Outlook contacts are read each hour there. F22 started early in Phase 1 and moves to Phase 2 if not done by the end of Empire’s Phase 1 UAT round (trackernotes), so a Phase 1 release tag may not include wave 1. Before eachuat.yamlapply and production’s first apply after #30 merges, check that the backend release there includes F22 wave 1, or decide what to do (a later apply, or abase/scheduleroption to leave a job out). Register §B F22 row (b) has the read that proves it. Closes when both overlays have been applied with the schedule beside a backend that has wave 1. -
Set
crm.bulk-email.postal-addressin production’s task definition (Tharinda, owner of F38-i; raised by F37-b-WI-2.1, 2026-11-19; F37-b Decomposition Plan## Notesitem 9). F37-b spec decision 12 sets Empire’s published company address as a plain task-definition value in every environment (AD #40);forge-et-be.aws/td.prod.jsondoes not exist yet, so F38-i adds the value when it creates that file. Until it is set, every bulk send in production is refused with422 POSTAL_ADDRESS_MISSING. Register §B F37-b row (f) has the check. Closes when production’s task definition carries it. -
Give
overlays/uatandoverlays/prodthe lead-ranking queue once forge-et-iaac#18 is ondev(Tharinda:overlays/uaton forge-et-iaac#11, andoverlays/prodas F38-i WI-4.1’s D-1 (21) on forge-et-iaac#17; raised by the review of #18, 2026-11-08, posted on the PR). #18 adds thelead-rankingqueue, its DLQ and thelead-ranking-sweepschedule tobase/sqsandbase/scheduler, so every overlay that calls them creates all three, but onlyoverlays/devgives its task role rights on the queue and outputs its name. Without the same lines, UAT and production would run the hourly sweep against a queue their backend cannot read or write, and no lead would be ranked there. In each overlay’smodule "ecs", addmodule.sqs.lead_ranking_queue_arntoconsumed_queue_arnsandsent_queue_arns, and add alead_ranking_queue_nameoutput (ET_MESSAGING_LEADRANKINGQUEUE), as #18 does for dev. Whichever of #11 and #17 merges after #18 mergesdevin and adds the lines; if one of them merges first, #18 (or a follow-up PR todev) adds them to that overlay. The row closes when both overlays carry them ondev. Update 2026-11-09: #18 merged and was applied on dev, so, at Tharinda’s request, the delivery team mergeddevinto both PRs and added the lines: #11 at037c55a(overlays/uat) and #17 at897752d(overlays/prod), each withfmt -check -recursive,init -backend=falseandvalidateexit 0 (Terraform 1.12.1 in Docker) and a comment on the PR. The row closes when both PRs merge intodev(13 Nov: #11 merged, so theoverlays/uathalf is done; the applieduatinstance haset-uat-lead-rankingand its DLQ). Separately, whoever writesforge-et-be’std.uat.json(F38-h) andtd.prod.json(F38-i WI-9.1) setsET_MESSAGING_LEADRANKINGQUEUEto that environment’s queue (et-uat-lead-ranking,et-prod-lead-ranking) andET_AI_FEATURES_LEADRANKING_MODELas for lead triage, as forge-et-be#375 did intd.dev.json; without the queue key the backend starts healthy and ranks nothing (register §B F11-g row 1). Update 2026-11-11: Shyamin Ayesh is holding #17 until production’s infrastructure is needed (the row below), so theoverlays/prodhalf closes only when #17 merges. -
Decide whether UAT and production serve the OpenAPI document (Tharinda; seen in dev’s startup log on 2026-10-20, and listed as a low startup warning in
.forge/audits/2026-10-13-aws-dev-account-audit.md). Every backend start logs SpringDoc’s WARN that/v3/api-docsis enabled by default. On dev,/v3/api-docs,/swagger-ui.htmland/swagger-ui/index.htmlanswer 401 to a signed-out caller (read 2026-10-20 19:55 UTC), so the document sits behind sign-in. The decision: keep it behind sign-in everywhere, or switch it off in UAT and production withspringdoc.api-docs.enabled=false(and the Swagger UI with it), which is a new task-definition key for those environments (.claude/CLAUDE.md § Boundaries › ASK FIRST). Decide before the first UAT deploy (F38-h). Decided for UAT by the lead, 2026-10-23 (F38-h spec decision 9): off — UAT’s task definition setsSPRINGDOC_API_DOCS_ENABLED=false(F38-h FR-5, AC-7), the one new key. The backend usesspringdoc-openapi-starter-webmvc-api, which serves no Swagger UI, so there is no UI key to set; the/swagger-uipaths answering 401 on dev are the security filter, not a UI. Decided for production by the lead, 2026-10-27 (F38-i spec decision 9): off as well — production’s task definition sets the same key (F38-i FR-9, AC-9). Dev still serves the document behind sign-in. UAT’s half is built:forge-et-be.aws/td.uat.jsonsets the key (forge-et-be#425, merged 2026-11-14). Shown on UAT, 2026-11-19 (F38-h WI-5.2, record R-5):/v3/api-docsanswered 404 to a signed-inCRM01on 2026-11-17 atv0.1.0-rc1. The row stays until F38-i builds production’s. -
dev’s apply changes the SES configuration set on every run (Tharinda, as F38’s lead, with Shyamin Ayesh; found 2026-11-19 in the apply of forge-et-iaac#24):
dev.yamlruns 44462431613 (7e3481a, 05:03 UTC) and 44488259851 (8b48548, 08:31 UTC) each plan and applymodule.ses.aws_sesv2_configuration_set.thisdelivery_options.max_delivery_seconds0 → null in place, whilebase/seshas not changed since0eee623(F38-a) and the lock file since37a0c2e. So AWS reports 0 and the code sends nothing, and the plan never settles. It is harmless (0 and unset both mean SES’s default), but every apply lists one change that is not the PR’s, which hides a real one. Find what changed (a provider default, or the setting changed outside Terraform) and settle it — setmax_delivery_secondsexplicitly inbase/ses, orignore_changesit with a comment. Done when adev.yamlrun with no code change plans “No changes”. Update 2026-11-19, evening: the plan-only run of forge-et-iaac#17’s head (44569651738, 16:45 UTC) planned the same change and nothing else, so it is still there after run 44488259851’s apply.overlays/prod(#17) calls the samebase/ses, so a fix there settles production too;overlays/uathas nosesmodule. Update 2026-11-19, later: forge-et-iaac#28 (open) setsignore_changes = [delivery_options]on the configuration set, with a comment; its plan-only run (44572051073, lead-approved) no longer lists the configuration set. Done when #28 is applied and the nextdev.yamlrun plans only its own changes. -
Remove
svc-harness’sspike-invoke-gatewaypolicy once forge-et-iaac#28 is applied (Tharinda, the lead; raised 2026-11-19 by #28). On 19 Nov Dinil decided to leave the AgentCore gateway and this policy in place, since neither has a standing charge (environments-and-identity.md§4, thebase/agentcoreline). The same day the lead asked for the gateway to be removed with the audit fixes, and #28 removes it fromdev. Once #28’s apply has run, the policy allowsbedrock-agentcore:InvokeGatewayon a gateway that no longer exists. The policy is not Terraform, so it is removed by hand:aws iam delete-user-policy --user-name svc-harness --policy-name spike-invoke-gateway. This is an AWS write, so the lead approves it first. Then update F38-a’s “What was made by hand” row in §4 and the F38-a spec’s temporary item. Done whenaws iam list-user-policies --user-name svc-harnessno longer lists it. -
Fix the smaller findings of the 14 Nov review of
forge-et-iaacdev(Tharinda; the review of the code merged ondevwithout a review, 2026-11-14; its two Important findings are fixed by forge-et-iaac#28; dev’s database guards are optional in dev, F38 Rev 18, and are in the “[ET] Misc tasks” row). The findings still open, each a Nit:base/ecs/execution_role.tf: the AWS-managedAmazonECSTaskExecutionRolePolicyallows ECR pulls and log writes on*. Replace it with an inline policy scoped tocrm/backendand the log group (ecr:GetAuthorizationTokenstays*), added before the attachment is removed.base/security-groups: the database group allows all egress, which it never needs (egress = []). The load balancer’s group allows all egress but needs only 8080 to the backend’s group (a separateaws_vpc_security_group_egress_rule, to avoid a cycle).base/ses/records.tf: DMARC isp=nonewith norua=, so no reports arrive and production inherits it.base/ses/identity.tf: the configuration set has no event destination, so bounces and complaints reach only the suppression list. Add one forBOUNCEandCOMPLAINTto the alerts topic when system mail lands.base/cloudfront/distribution.tf:PriceClass_100is fixed inbase/, so production has no Asian edges. Make it a variable, and setPriceClass_200in prod.base/ecs/capacity_providers.tf:FARGATE_SPOTis registered but unused. Use it on dev only, or remove it.base/kms/keys.tf: themailbox-tokenkey has noprevent_destroy. Renaming itsfor_eachkey would schedule dev’s key for deletion. Document the recovery (aws kms cancel-key-deletion), or protect the key in the dev and prod overlays only.base/security-groups/rds.tf: turning ondatabase_ingress_from_backend_onlychanges the group’s description, which replaceset-dev-rdsand fails the apply part-way. Make the description independent of the flag before the flag is used.CLAUDE.mdline 9: “CI runs the same checks” is wrong, sincedev.yamlvalidates onlyoverlays/dev.
Raise them as one PR to
dev, or record here, with its reason, each one the team decides to leave. -
Bring UAT up to
forge-et-iaacdevwith the next-rctag (Tharinda, owner of F38-h; Senuthmi Thimansa for theuat.yamlruns; the lead’s decision of 2026-11-17).- What changed on
dev. Since UAT’s last apply (13 Nov,b175dae),devgained thecard-readqueue and its DLQ (forge-et-iaac#22, F36-b), theprice-broadcast-sendschedule (#20, F37-a) and lead research’s 270 s visibility timeout (#21, F36-c). UAT runsv0.1.0-rc1, which has neither F37-a wave 1 nor F36-b wave 2 (WI-5.1 record, difference 12). - Why not apply now. An apply now would start a schedule whose job kind rc1 refuses every minute.
- What to do. Leave UAT on rc1. When the next
-rctag is cut, cut it from amainthat holds F37-a wave 1 and F36-b wave 2. Runuat.yamlplan, thenapply, withserviceticked, before pushing the tag. Then deploy the tag, so the backend that reads the new queue and job kind starts after they exist. Theuat.yamlplan will no longer sayNo changes.until that apply runs. - The card-read key in
td.uat.json: fixed by forge-et-be#464 (2026-11-17). It setsET_MESSAGING_CARDREADQUEUE=et-uat-card-read; the gap was found by F38-h WI-5.2’s record, R-2. Because of it, the apply must run before the deploy: a queue name with no queue fails the deploy. Check F36-c’s two research queues and theplan-researchmodel override (register §B F36-c item ©) the same way, once they reach UAT. - Done when UAT has been applied from a
devthat holds #20, #21 and #22, beside a backend with both waves. - Update 2026-11-19: the apply ran before the tag. Senuthmi Thimansa ran
uat.yamlfromdevat7e3481afor the capture hosting (plan run 44463955499, apply run 44464377454:14 added, 3 changed, 0 destroyed), which also brought #20, #21 and #22, so UAT now has thecard-readqueue and theprice-broadcast-sendschedule while it still runsv0.1.0-rc1. What the 17 Nov decision avoided now happens: the backend refusesprice-broadcast-sendevery minute and each tick is dead-lettered (et-uat-scheduler-dlq: 341 messages at 11:03 UTC;UnknownScheduledJobKindExceptionin/ecs/uat/backend). No data is at risk; the DLQ and the log fill until the next-rcdeploys. Decision for Tharinda: (a) leave it until the round’s-rc(about 30 Nov) and purgeet-uat-scheduler-dlqjust before the round (an AWS write), or (b) disable UAT’sprice-broadcast-sendschedule by hand now (an AWS write, the DevOps team’s; the nextuat.yamlapply turns it back on), or © push the next-rcearlier than the round’s start. Recommended: (a), since nothing breaks and every option needs one write. Decided by Tharinda (lead), 2026-11-19: (a) — the cost is under USD 0.20 until 30 Nov (measured 11:30 UTC: 180 log events and about 1.2 MB an hour; the schedules and SQS calls sit inside the free tiers). F38-h-WI-6.1 emptieset-uat-scheduler-dlqjust before the round, with the lead’s approval at the time. Changed by Tharinda (lead), later on 2026-11-19: push an earlyv0.1.0-rc2frommainin all three app repos as soon as WI-5.2 has finished, so UAT runs a backend that knows the job (the failures stop then, and the DLQ is emptied after that deploy) and the capture app’s whole UAT path is proved before the round; the round itself then starts on the next free-rc(v0.1.0-rc3, about Mon 30 Nov). Claude runs it in Tharinda’s session, with Tharinda approving each write (the apply, each tag push, the purge): (1)uat.yamlplan, then apply withserviceticked, which brings forge-et-iaac#24 and anything else ondev, and, if forge-et-iaac#25 has merged by then, first check thattd.uat.jsononmainnames its queues (register §B F36-c item ©); (2) choose in each repo amaincommit the last scheduled nightly run passed, the backend’s with acrm/backend:build-<SHA>image (.claude/rules/release-and-branch-protection.md§ Release Process); (3) push the annotated tag in all three repos; (4) read the threedeploy-uatruns, run the smoke suite againstuat(only if the tag carries the smoke suite’suattarget — the row “Build the smoke suite’suattarget inforge-et-fe”; otherwise that run waits for rc3), and check the capture app (version meta tag, manifest type, a sign-in, a signed-in API call from the capture host); (5) emptyet-uat-scheduler-dlq. Record it inenvironments-and-identity.md§4. Still to do before the tag: forge-et-iaac#24 (lead triage’s 150 s timeout, merged 08:31 UTC) and anything merged todevafter it reach UAT only with anotheruat.yamlplan and apply, run before the-rcpush as above. The row is done when that-rcruns on UAT.- Progress 2026-11-19 (Claude, in Tharinda’s session; paused when the session’s tokens ran out). Tharinda approved every write in steps 1–5 in advance, and told Claude to run them without asking and to merge the harness PR that records them. Step 1 done:
uat.yamlplan run 44559155469 atdev8b48548showed one change,et-uat-lead-triage’s visibility timeout 30 → 150 s (forge-et-iaac#24); apply run 44559352768 at 15:51 UTC:0 added, 1 changed, 0 destroyed. Decided by Tharinda (lead), 2026-11-19: tag rc2 after the scheduled nightly run of 2026-11-20, on the commits it passes, not on the 19 Nov night’s commits: on that night the capture app’s browser gate did not run (it passed only in a hand run, on2f7bf77), and its CRM commitcc97f8cpredates the smoke suite’suattarget (forge-et-fe#591). To resume: (a) read the 2026-11-20 scheduled comment on the harness issue labellednightly-log(#981); a red suite or no scheduled run means no tag; (b) check that F38-h-WI-5.2’s R-10 second half (at or after 18:27 UTC on 19 Nov, record onfeature/F38-h-wave-5) has been read, because a UAT deploy restarts the backend under that check; © planuat.yamlagain and apply only ifdevgained changes (forge-et-iaac#25, #26 and #27 were open at the pause; #25 also needs thetd.uat.jsoncheck in step 1); (d) then steps 2–5 above and the record inenvironments-and-identity.md§4. - Progress 2026-11-20 (Claude, in Tharinda’s session): no tag, because the 20 Nov night was red. The scheduled comment on #981 (posted 23:21 UTC on 19 Nov) shows two red suites. The CRM browser gate failed 8 tests at fe
c77a7f6(forge-et-fe#593): the Columns menus’ keyboard specs,views-across-users,select-tab-after-chooseandleads-focus-marker-tooltipat both WebKit tablet and desktop sizes. The capture app browser gate failed 2 seam tests at pwa9ddaddd(forge-et-pwa#189):capture-journeyandreview-screens, both on the convert to a CRM lead. Every other suite passed, at beb33e11b, fec77a7f6and pwa9ddaddd. Checks (b) and © were made, both reads: R-10’s second half was read (WI-5.2 record, 18:41–19:05 UTC on 19 Nov, merged in harness #1567), andforge-et-iaacdevwas still at8b48548, where UAT was last applied, with #25 to #29 open. The six queue names intd.uat.jsononmainall exist in UAT. forge-et-fe#591 (the smoke suite’suattarget) was still a draft, so rc2’s step 4 has no smoke run againstuatunless it merges first. No AWS write, workflow dispatch or tag push was made. Next: once bothnightly-redissues are closed, tag rc2 on the commits of the first scheduled night that passes every suite. Resume from (a) above with that night’s date. UAT stays onv0.1.0-rc1, andet-uat-scheduler-dlqkeeps filling until then (the 19 Nov decision (a) covers the cost). - Progress 2026-11-20, later (Claude, in Tharinda’s session): both
nightly-redissues are closed. All 10 failures were out-of-date tests (BUG-498 to BUG-502,.forge/bugs.md); each failed again when re-run alone at the night’s commits and passed on its fix. The fixes merged in forge-et-fe#596 and forge-et-pwa#190, and forge-et-fe#593 and forge-et-pwa#189 were closed with the re-run results. Next: the scheduled night of 2026-11-21 is the first that can pass every suite with the fixes. Resume from (a) above with that date.
- Progress 2026-11-19 (Claude, in Tharinda’s session; paused when the session’s tokens ran out). Tharinda approved every write in steps 1–5 in advance, and told Claude to run them without asking and to merge the harness PR that records them. Step 1 done:
- What changed on
-
Build
forge-et-pwa’sdeploy-uatand the capture app’s UAT hosting for Phase 1’s round (Tharinda, F38-h’s owner; for Phase 1’s round since the lead’s decision of 2026-11-12, first decided for Phase 2 by the lead, 2026-10-23, F38-h spec decision 6, F38 Rev 24): Phase 1’s round deploys only the backend and the CRM, because the capture app is Phase 2 scope (F36) andforge-et-pwahad no dev deploy. Before Phase 2’s round:forge-et-pwa’sdeploy-uat(from itsdeploy-dev, F38-d wave 5, with the release vulnerability check and the version check), theuatoverlay’s capture-app bucket, distribution, certificate andcapture.uat.crm.empire.lkrecords, the capture origin in UAT’s CORS list, and the capture bundle’s version inforge-et-fe’s pre-deploy check (F38 FR-35, FR-36). F38-i builds it (spec FR-18 and AC-18, approved 2026-10-27, decision 7); the row closes when that work merges. Changed 2026-11-11: the capture app is Phase 1 scope from that day (features.md § Delivery Plan), so Phase 1’s UAT round needs the capture app too — Tharinda decides whether this work, and F38-h’s decision 6, move to Phase 1’s round. 2026-11-12: Section L of the Phase 1 acceptance criteria now tests the capture app in Phase 1’s UAT environment (PRD Rev 91), so this work is needed before Phase 1’s round (UAT from Mon 30 Nov). Decided by the lead, 2026-11-12: F38-h builds the capture app’s UAT hosting for Phase 1’s round, and F38-i does the same for production. Next: revise F38-h decision 6 (on top of harness PR #1276, which revises F38-h and is open) and F38-i FR-18, then build it. Update 2026-11-13: #1276 merged on 11 Nov, and forge-et-iaac#11 merged on 13 Nov, which meets F38-i wave 7’s gate (“F38-h wave 3 is merged, so the capture hosting is added to it”); theuatinstance applied that day has no capture-app bucket, distribution or certificate. Still next: the F38-h decision 6 and F38-i FR-18 revisions, which say whether wave 7 moves into F38-h; the hosting then goes on the running instance with auat.yamlapply (serviceticked once the service exists, never unticked on a running instance). Update 2026-11-19 (evening): built and mostly merged; the harness is revised. Senuthmi Thimansa built it from this row, not from a plan, and set it up the same day: forge-et-iaac#23 (the hosting; approved by Shyamin Ayesh, merged 05:02 UTC, applied to UAT byuat.yamlrun 44464377454), forge-et-be#474 (the capture redirect and CORS origin intd.uat.json, merged 06:27 UTC), forge-et-pwa#169 (deploy-uat, merged 06:28 UTC),forge-et-pwa’suatGitHub Environment andforge-et-fe’sCAPTURE_URLvariable (environments-and-identity.md§4). The harness records it as F38-h spec Rev 12 (FR-13, AC-18; decision 6 reversed), F38-i spec Rev 4 (FR-18 and AC-18 retired, waves 7 and 16 withdrawn) and F38 Rev 36; the release rules now tag all three repos together. The capture app first deploys with the next-rctag: since F38-h spec Rev 15 that is the earlyv0.1.0-rc2, pushed once WI-5.2 has finished, and the round starts onv0.1.0-rc3, about Mon 30 Nov. Reviews: #23 reviewed after the merge (posted; clean, two findings — theblob:row above and theprice-broadcast-sendrow); #474 and #169 reviewed after the merge (comments posted; clean, one Nit on #169). #474 and #169 merged with noForge revieworTest recordstatus, through the ruleset bypass, and their descriptions do not say so (.claude/rules/git-conventions.md§ Branch Protection); the comments ask Senuthmi to add the line. Still open:- forge-et-fe#582 (ready; review record
block, three Importants): the harness revision it waited on has merged (forge-et-harness#1533); it still needs aTest recordon its head and its capture wait changed to compare the full tag (item 3). Then a delta review and the merge, before the round’s-rctag. - The lead’s answer to #582’s question — should a failed capture run block the CRM deploy? Recommended: yes, as F38 FR-35 asks (F38-h FR-2’s Rev 12 bullet is written that way); the lead’s choice of item 3 keeps that block meaningful on every
-rc. - Decided by Tharinda (lead), 2026-11-19: the waits compare the full release tag (F38-h spec Rev 14). Every candidate of a round carries
0.1.0, so a wait onX.Y.Zpassed against the previous-rc’s backend or capture app. The work is the next row, “Make the UAT version waits compare the full release tag”. - AC-18 is recorded by F38-h-WI-6.1 at the round’s first
-rcdeploy.
Done when #582 has merged and AC-18 is recorded.
- forge-et-fe#582 (ready; review record
-
Make the UAT version waits compare the full release tag (Tharinda, F38-h’s owner, or the DevOps team if they take it; the lead’s decision of 2026-11-19, F38-h spec Rev 14). Due before the round’s first
-rc(about Mon 30 Nov). Three changes, each reviewed with/forge-review-pr:forge-et-be:deploy-uatwrites the release tag (github.ref_name) into the revision it registers, as it fills the image digest, and/actuator/inforeports it (for example as aninfo.*value read from that environment variable). The image is promoted from dev, never rebuilt, so the tag cannot be built into it. It is one new task-definition value in UAT, approved with the decision.forge-et-pwa:deploy-uatwrites the tag into the bundle it builds (a meta tag besideapp-version, from a build-time variable that is not aVITE_setting).forge-et-fe: both waits indeploy-uat— the backend’s (existing) and the capture app’s (forge-et-fe#582) — compare that tag withgithub.ref_nameinstead ofX.Y.Z.
Added 2026-11-19 (the review of forge-et-fe#591, finding 7): the smoke run against
uathas the same gap. Its checks 4 and 5 compareX.Y.Zwith the release tag’s version, as FR-11 asks, so a run withv0.1.0-rc3also passes against an rc2 backend or CRM bundle left in place by a failed rc3 deploy. Once change 1 has merged, check 4 onuatcan compare the backend’s reported tag withET_SMOKE_RELEASE_TAG(a smallforge-et-fechange toe2e/smoke/checks.smoke.tsande2e/smoke/versions.ts). The CRM bundle carries no tag, so check 5 stays onX.Y.Zunless the CRM’sdeploy-uatbuild also writes the tag. Not needed for the row to close.Done when the three have merged and the round’s first
-rcshows both waits matching the tag (F38-h-WI-6.1, AC-18). -
Give the UAT data checklist’s email contact a test mailbox that exists (Tharinda, F38-h’s owner; found by F38-h WI-5.2’s record R-10, 2026-11-19).
crm04@empire.lkhas no Exchange mailbox or licence (environments-and-identity.md§5, corrected). The checklist’s contactTEST Jonas Brandt(A-1) uses it, and criteria F3, F4 and G8 send mail to him, so in the round those emails would bounce.CRM04is also Part 3’s Tea Room tester on UAT, which needs no mailbox. Two ways to settle it: Empire IT givesCRM04an Exchange licence (an IT request, not a requirements question), or the checklist moves the contact to another test mailbox that is notCRM01(noemail.manageon UAT) orCRM05(the smoke account only). Before UAT’s data is entered again. Also left by the check: a dead mailbox connection forCRM04on dev (connection 18,RECONNECT_REQUIRED, no mail and no subscription), because the Mailbox page offers only Try Again in that state and no Disconnect. Harmless; it goes ifCRM04is licensed and reconnects, or a developer deletes it on dev. -
Change the five test accounts’ passwords after Phase 1’s UAT round (Tharinda, with Empire IT, who hold the
empire.lkaccounts; raised by F38-h WI-5.2, 2026-11-19). The lead shared the passwords ofCRM01–CRM05in a Claude Code session on 2026-11-19, so that Claude could finish WI-5.2’s mailbox check unattended if Microsoft asked for a sign-in. Claude typed them only into Microsoft’s password field and wrote them to no file, log, record or commit, but they are in that session’s transcript on the lead’s laptop. After the round ends (10 Dec), ask Empire IT to change all five and share the new ones only through the password vault, not in a chat. Done when all five are changed. -
Build the smoke suite’s
uattarget inforge-et-fe(Tharinda, F38-h’s owner; found by F38-h-WI-5.2’s record, R-7, 2026-11-19). F38-h FR-11 and WI-4.2’s success criteria ask for it, but it was never built: forge-et-fe#541, WI-4.2’s PR, changed only.github/workflows/deploy-uat.yamland says the target “is left for a follow-up”; WI-4.2’s plan subtasks 5–7 are unticked. Todayscripts/smoke/run.mjsrefuses every target buthttps://dev.crm.empire.lk, atv0.1.0-rc1and onmain(read 2026-11-19 12:38 UTC). The work is WI-4.2’s subtasks 5–7: accepthttps://uat.crm.empire.lk, check the version against the release tag, and leave the dev run unchanged. It is aforge-et-fePR with/forge-review-pr, the test record and the selected browser run. Due before the earlyv0.1.0-rc2tag (the “Bring UAT up toforge-et-iaacdev” row, step 4). If it has not merged by then, rc2 goes out without the smoke run againstuat, and that run (WI-5.2’s R-7, AC-13’suathalf) waits forv0.1.0-rc3. Decided by Tharinda (lead), 2026-11-19: Claude builds it now, in a session of its own in Tharinda’s name, separate from the session running WI-5.2, so that it merges before rc2. Done when it has merged and the smoke run againstuathas passed on a tag that carries it.- Update 2026-11-19: built as forge-et-fe#591. The command is
ET_E2E_TARGET=https://uat.crm.empire.lk ET_SMOKE_RELEASE_TAG=<tag> pnpm e2e:smoke, run from a checkout of the tag (or ofmainonce it carries #591) with Chromium installed. Auatrun needs the tag and a dev run refuses one; onuat, checks 4 and 5 compare with the tag’s version without-rc<N>, with noghread. The smoke account, role andSMOKErecords must exist onuat(.forge/plans/F38-h/uat-data-checklist.mdParts 2–4). Still to do: the smoke run againstuaton the first tag that carries #591. - Update 2026-11-20: #591 merged (
738249ef, 2026-11-20 04:00 IST), with its review record and test record onf2146788. Its selected browser run had 4 failures, the AC-20b Columns keyboard walks, which fail the same way onmainin the 2026-11-20 nightly (BUG-498, from forge-et-fe#586). Still to do: the smoke run againstuaton the first-rctag cut from738249efor later.
- Update 2026-11-19: built as forge-et-fe#591. The command is
-
Decide how F11-d’s duplicate rule treats the final Pipedrive run (Dinil, since 2026-11-19 by the lead’s reassignment at the
owed.mdtriage; raised by the review offorge-et-be#279, 2026-10-20; must be settled before the final migration run, F4-b): spec FR-D2 counts only leads received earlier and FR-D3 never re-checks a lead after triage, while imported leads keep Pipedrive’s created time asreceived_at. An inquiry entered in both systems during the parallel run is then left untagged on both copies: the live lead was triaged before the import, and the imported copy has no earlier lead. Options: (a) when a lead is triaged, also queue the open leads received after it with the same email or company key; (b) compare an imported lead bycreated_atinstead ofreceived_at. Either is a spec revision of F11-d FR-D2/FR-D3 and a small backend change. -
Decide whether F11-d’s company match drops two legal-form words (Dinil, since 2026-11-19 by the lead’s reassignment at the
owed.mdtriage; raised by the review offorge-et-be#279, 2026-10-20): spec FR-D1 drops one trailing legal-form word, so Sri Lanka’s common(Pvt) Ltdnormalises to… pvtandEmpire Teas (Pvt) Ltddoes not matchEmpire Teas. Dropping up to two trailing words (or treatingpvt ltdas one form) is a spec revision of FR-D1 and a one-line change to the company expression inLeadTriageRepository. -
Two SQS follow-ups left by forge-et-be#282 (Dinil, or whoever next touches the messaging code; raised by its review, 2026-10-20, L-124): #282 bounds every SQS send a request or listener waits on to five seconds (
BoundedSqsSend), which settles the stalled-request problem forge-et-be#279’s review raised. Still open: (1) the lead-triage sweep keeps sending after a failure —LeadTriageSweepServicesends up toet.lead-triage.sweep-batch(200) leads in a row, so while SQS does not answer one run can take 200 × 5 s, far past the scheduler message’s 30 s visibility timeout, and the message is delivered again; stop the run after the first failed send (or an interrupt) and leave the rest to the next hourly run — a behaviour change to F11-d’s sweep (FR-J7), so Dinil decides, with Tharinda if the spec changes; (2) abandoned sends hold connections — a send whose wait ended keeps its connection until the SDK’s own timeouts, and the@SqsListenerlong polls share that client’s pool; a client-wideapiCallAttemptTimeoutwell above the 10 s poll wait (about 20–25 s) would bound them without cutting off receives. -
The Phase 2 scope conversation with Empire (Tharinda; not yet held, planned for the Phase 1 hand-over, which is one meeting with UAT drop 1, about Mon 30 Nov at Empire’s office, if time allows after the hand-over demonstrations and the UAT discussion — the lead, 2026-11-19; the Pipedrive deals export and the Phase 2 deal questions are raised there too): go through the Phase 2 features with Empire to agree what they need most, what they do not need, what can be skipped or deferred, and anything they want added. Phase 2 does not fit its build target as it stands (R-ARCH-001), so the conversation is to reduce it; anything added goes through a change request (
client/project-timeline.md§4). The issues Empire raises in Phase 1’s UAT usually come before Phase 2 features (the lead, 2026-11-18; the triage row above), so the meeting should also allow for them. Bulk email (F37-a, F37-b) was added to Phase 2 on 2026-10-09 (PRD Rev 57), and Tharinda told Kavindu that day that this meeting, held at the hand-over or earlier, decides which Phase 2 features Empire does not need: that is where the compensating descope for F37 is agreed and recorded in the proposal header. F37-a’s open questions are not taken to Empire: the delivery team decides them at spec time (the lead, 2026-11-13, no questions to Empire before UAT; the F37-a spec’s## Open Questionsholds a suggested starting point for each). -
Write the rest of the user guide before Phase 1’s UAT round starts (Biseni, from 2026-11-18, agreed with the lead; Tharinda started it on 2026-11-13). Due before the round opens on Mon 30 Nov, not at handover: an Empire tester who was not at the walkthrough meetings would otherwise report as defects things that work as designed. Proposal §9 promises Empire “a concise user guide and an administrator guide, delivered as part of handover”.
.forge/client/user-guide-weekly-prices.mdis the first chapter, a draft; each later one is.forge/client/user-guide-<area>.md. Still to write: accounts, leads, email, follow-ups and the capture app, each describing the CRM as built on UAT. Written in simple English (.claude/rules/client-docs.md, Simple English); the lead asked for this explicitly. Done when every Phase 1 area has a chapter and the guide has been given to Empire’s testers before the round. -
Write the administrator guide before handover (Tharinda): the other half of proposal §9’s promise. It covers users, roles, tags and email templates, describes the CRM as built, and follows the same file pattern and simple-English rule as the user guide. Due by handover, as before; it can follow the user guide’s chapters.
-
Give Empire one way to report problems found in UAT (Biseni, from 2026-11-18, agreed with the lead): Empire’s testers report every problem through one channel, not WhatsApp messages. Biseni prepares an Excel sheet before the round opens on Mon 30 Nov, and confirms the method with Empire on the round’s first day (this is a UAT-day question, so the rule of no questions to Empire before UAT is kept). The sheet’s columns match what triage needs (
.forge/release/_TEMPLATE-uat-issues.md): who reported it and when, the screen, the steps, what was expected, what happened, a screenshot, and how serious the tester thinks it is. The delivery team triages each row intotracker.yamluat_issues:asUAT-NNNand its prose into.forge/release/P1/uat-issues.md(.claude/rules/tracker.md§ UAT issue triage + routing). The sheet is the intake only, never the record. Done when the sheet exists, Empire has agreed to use it, and the team has agreed who triages it and how often during the round. -
Carry the capture app’s move to Phase 1 into every document (Nanndish, who owns F36; Tharinda for the Statement of Work and the UAT and production work; recorded 2026-11-11): the delivery team and Empire agreed at the 10 November progress meeting that the capture app, F36-a to F36-d, is delivered in Phase 1, to Phase 1’s dates (
features.md § Delivery Plan,tracker.yaml). Done on 2026-11-12 (PRD Rev 91): the meeting note (§4, §6, §8), the PRD, AD #20 and its archive entry,architecture.md § Trade-Show Capture App, the phase titles and themes intracker.yamlandfeatures.md, the client pack (client-facing-prd.md§3 and §3.5,project-timeline.md§1 and §2,phase-1-acceptance-criteria.mdSection L), a post-issue note on the proposal,.claude/rules/client-docs.md,team-guide.md,guides/tech-stack-reference.md,design/ai/lead-research-source-options.mdandforge-et-pwa’s README. Still to do: (a) the UAT and production work that follows, in the row “Buildforge-et-pwa’sdeploy-uat…” above, and with itarchitecture.md § CI/CD Pipelines(deploy-uatfor the capture app “comes with Phase 2’s round”),.claude/rules/release-and-branch-protection.mdand.claude/commands/forge-release.md(forge-et-pwatagged “from Phase 2’s round”),guides/test-runs-explained.md(“first released from Phase 2’s round”),test-strategy.md(the real-phone pass “in Phase 2’s round”), and the F38-b, F38-e, F38-h and F38-i specs and plans, each changed by the revision that moves the work (2026-11-19: done for UAT — F38-h spec Rev 12, F38-i spec Rev 4, F38 Rev 36, architecture Rev 92, the release rules and/forge-release;test-runs-explained.mdandtest-strategy.mdalready said Phase 1. Left in (a): the F38-b and F38-e specs, and F38-i’s production half, which its own waves carry); (b) the F36 spec bodies (Nanndish, as each spec is written); © F43’s spec, which lists F36-b and F36-c among the Phase 2 features that wait on it (Tharinda, F43’s owner); (d)forge-et-pwa’ssrc/features/capture/CaptureRoute.tsxcomment (“F36 (Phase 2)”, replaced by F36-a’s first PR) and itsvitest.config.tscoverage thresholds, set “at the Phase 1 → Phase 2 boundary” (Nanndish, in F36-a’s first PR: the thresholds switch on with it, the lead’s decision of 2026-11-12, now that the capture code lands in Phase 1); (e) Empire’s confirmation of the capture-app review on Fri 20 Nov, which the Statement of Work holds during Phase 2 (proposed by the lead, 2026-11-12; Tharinda, with Empire;project-timeline.md§ The capture application). Done when (a) to (e) are done. -
The Phase 2 features started early in Phase 1: add each finished one to the Phase 1 acceptance criteria, and move back what is not done (Tharinda; decided 2026-11-06, Gate 3 Run 4, PRD Rev 84; the rule for finished ones decided 2026-11-18): the lead’s rule (2026-11-18): every started-early feature finished in Phase 1 goes into
.forge/client/phase-1-acceptance-criteria.md, with its PRD revision and the client pack carried in the same change, as PRD Rev 90, Rev 92 and Rev 100 did. Done so far: F34-c (Section K, Rev 90), F33-a (Section M, Rev 92), and on 2026-11-18 F20-g (F8 and F9), F37-a (G9) and F32-d (Section N), all Rev 100. F22 moved back to Phase 2 on 2026-11-11, and F20-k moved there from Phase 1 with it. Still open: F23-d (export of the account, lead and contact lists), taken back into Phase 1 by the lead on 2026-11-15 and not done; F37-b (bulk email to a saved contact list), taken into Phase 1 by the lead on 2026-11-19 and not started; and F20-k (email block lists, after F20-f) and F22 (contact sync, after F20-i), taken back into Phase 1 by the lead on 2026-11-20 and not started. When one is done, its/forge-reflectrun adds its criteria and carries the move through the documents (step 5e). The lead’s rule (2026-11-19): each stays in Phase 1 through Empire’s Phase 1 UAT round (drop 1 about Mon 30 Nov, drop 2 during the testing), so it can go out in either drop. If one is not done by the end of that round, setdelivery_phase: 2on it and carry the move through the documents (the phase-boundary check,.forge/checklists/phase-boundary-checklist.md§3). F38-i keeps its own rule (its spec Rev 3). F36-a to F36-d are firm Phase 1 scope (the row above), not started-early items. The Phase 2 scope conversation (the next row) should know which are already built. Done when F23-d, F37-b, F20-k and F22 are each in the acceptance criteria or back in Phase 2. -
Drop
role="status"from the remaining loading skeletons (Tharinda, to assign; raised byfrontend-revieweron forge-et-fe#510, 2026-11-12): the record places’ own status lines speak in every state, but the loading skeletons insrc/features/notesAndTags/SectionSkeleton.tsx,src/ui/skeletonLines.tsx,src/features/timeline/AccountTimelinePanel.tsxandsrc/features/attachments/RecordAttachmentsPanel.tsxare stillrole="status"with anaria-label, so two status regions exist while loading (the skeleton’s announces nothing). Give themaria-busyonly, as forge-et-fe#510 did for the custom fields, and update the tests that find them by role. Done when no skeleton is a status region. -
Ask Empire to confirm the newsletter choices, Q10 / OQ-62 (Tharinda): OQ-62 was decided by the delivery team on 2026-11-19 in F37-b’s spec authoring, under the lead’s rule of 2026-11-13 (PRD Rev 103; the resolved row is in
project-prd-history.md): each newsletter goes from the sender’s own mailbox, to every contact on the list who has not unsubscribed, with Empire’s postal address and an unsubscribe link that opens a page with one button, and unsubscribed or bounced addresses are left out of later sends.client/open-questions.mdQ10 andclient/client-facing-prd.mdSection 3.4 state the choices for Empire to confirm. The hand-over meeting’s list (the row “The hand-over meeting with Empire”) asks Empire to confirm them. Done when Empire has confirmed them, or asked for a change at UAT that is recorded as a UAT issue. -
Ask Empire OQ-68, the email signature (Tharinda; raised by the F20-c spec, 2026-10-12; open in
project-prd-signals.md): email sent from the CRM carries no signature, because Microsoft Graph neither adds nor exposes the Outlook one. Ask whether reps want a signature setting in the CRM; a “yes” is a small follow-up feature. -
Ask Empire OQ-63, past orders (Tharinda; open in
project-prd-signals.md): PRD F1 promises past orders from the ERP on the account page, and V1 has no ERP feed. -
OQ-49’s remaining point: §11’s “staging” wording (Dinil; open in
project-prd-signals.md). (Who receives the AWS Budgets alerts left this row on 2026-10-15: the lead decided the DevOps team and Empire Teas arrange the budgets and their recipients, and the delivery team only records them — F38-a FR-11 for dev, F38-i for production; F38 Rev 13. Since 18 Oct (F38 Rev 18) billing settings are Shyamin Ayesh’s, Kavindu Anthony’s, Malinda Wellalage’s and Thilina Abeyrathna’s, and the delivery team neither asks for nor tracks them.) -
Measure typing in the other list screens’ search boxes, and move their text into the box where it is slow (Biseni; raised by BUG-261’s fix, forge-et-fe#437, 2026-11-06): BUG-261 found that the Leads screen held its search box’s text, so each key press re-rendered the whole screen and typing a 17-character search took about 7 s in WebKit on Windows (lesson L-273). The Users, account list, contacts and email-template screens hold their text the same way (
queryTextinUsersScreen.tsx,AccountListScreen.tsx,ContactsScreen.tsxandEmailTemplatesScreen.tsxonmain4a91ed9c). Time typing in each in WebKit against a bare input on the same page, and file a bug for each that is slow. -
Correct the F1-c spec’s How the rows are made paragraph (Biseni Perera, author of the F1-c spec; raised by the F1-c-WI-1.3 plan review, 2026-10-19): the paragraph says no route on
mainwrites rows with a chosen pastoccurred_atand that routes writeleadrows dated now, butPOST /api/v1/leadsaccepts a pastreceivedAtand anaccountId(forge-et-belead/web/LeadCreateRequest.java), and a lead matched to the account gets itsleadrow dated at thatreceived_at(lead/service/LeadTimelineEntries.java). Correct it in a wording-only spec revision; no requirement changes. The WI-1.3 plan now uses the route for AC-1’s same-instant tie (.forge/plans/F1-c/F1-c-WI-1.3-plan.md § Notes, Same-instant rows). -
Email from F1-f’s contacts page and from contacts with no account (Biseni, owner of F1-f, assigned by the lead on 2026-11-19 at the
owed.mdtriage; raised by the F20-c spec, 2026-10-12): neither F20-c nor F1-f builds it. A row action Email on the contacts list would avoid opening a Modal over F1-f’s contact view; a contact with no account also needs a send-reach rule (it is reached only throughaccount.all, AD #37). -
Tick F1-f AC-12 once its tests pass on
main(Biseni, owner of F1-f; raised by BUG-389’s fix, 2026-11-10): AC-12 carries a Not met note until forge-et-fe#381’sContactsScreen.test.tsxandcontacts-list.spec.tspass onmain. The refreshing test failed there until forge-et-fe#487 (BUG-389) fixed its row lookup behind the open archive confirmation. Once a nightly full run onmainafter that merge passescontacts-list.spec.ts, tick AC-12 with a spec revision entry (.claude/rules/specs.md) and remove this row. -
Decide how F1-e’s contact Modal guards a close over unsaved text (Biseni, owner of F1-e and F1-f; raised by F1-f-WI-1.2’s final plan review, 2026-10-12; restated 2026-10-14):
src/ui/components/Modal/Modal.tsxonmainhas F20-c’s optional before-close hook,onCloseRequest, which F1-f WI-1.2 adopted in place of its own. From F1-f WI-1.3’s plan Rev 1 (AD #39), F1-f’s contact view is a shadcn Dialog with its own close guard and uses noModal. F1-e-WI-1.4 Decision 22 guards its own contact Modal another way: it lets the close happen and reopens the Modal. Once both features are onmain, decide whether F1-e’s Modal moves toonCloseRequestor, under AD #39, to a shadcn Dialog with the guard F1-f’s view uses, so the CRM is left with one close-guard mechanism. -
Add the contact view’s two place-driven page states, now that F1-e fills its places (Biseni, owner of F1-e and F1-f; raised by
frontend-revieweron the F1-f wave, 2026-10-16): the contact view’s refresh failed mark and its unsaved-text question are reachable only when a place’s component reports a change or unsaved text, so F1-f’s size matrix has noPAGE_STATESentry for either. When F1-e’s Tags and Notes fill the places (whichever of F1-e and F1-f reachesforge-et-femainsecond), addcontacts-view-refresh-failedandcontacts-view-unsavedtoe2e/pageStates.tsande2e/browser/size-matrix/arrange.ts. F1-e reachedmainsecond (forge-et-fe#284 fills the places and addscontacts-view-notes-tagsonly), so these two states go in F1-e’s follow-up PR with the nine Nits below. -
Move the other two boolean-filter helper copies onto
ListQuery(Biseni, owner of F1-f; raised by/forge-review-pr forge-et-be#267, 2026-10-16): F1-f’s wave PR addedListQuery.singleBoolean(String)andListQuery.withoutFilters(Set<String>)inforge-et-be’scom.empireteas.platform.queryand moved the contacts list onto them, butAccountService.sharedWithMeFilter/withoutSharedWithMeandLeadService.booleanFilterkeep their private copies, because that is other features’ code. Move both onto the shared helpers and delete the copies; in the same change, makesingleBooleanthrowIllegalArgumentExceptionwhen the named filter is not aBOOLEANfield, with oneListQuerySpecrow (the review’s Nit). -
Move every modal and every success message onto design-system Rev 90 (Nanndish; decided 2026-11-03 on the UI/UX designer’s recommendation): Rev 90 gives every modal one look (a centred card, only the body scrolling, the buttons kept in view) and three sizes (
layout/dialog-sm|md|lg, 420, 560 and 800px), and makes every success a toast. Inforge-et-fe, one PR per screen or area, as the shadcn backfill goes: add the threelayout/dialog-*tokens totokens.cssand the sizes tosrc/ui/shadcn/dialog.tsxandalert-dialog.tsx; moveRecordEditDialog,ConfirmDialogand the dialogs that set their own width in a CSS module of their own (the<DialogContent>/<AlertDialogContent>callers undersrc/features) onto them; move the inline success messages to Sonner — the contacts notice (src/features/accounts/ContactsNotice.tsx, an inlineAlertin an always-mounted status element chosen in F1-f WI-1.3 Decision 31 before Sonner was in the repo; this settles the row raised by F1-f-WI-1.3’s plan Rev 1 on 2026-10-14 for Tharinda to decide, andToastalready maps to Sonner insrc/test/shadcnScan.ts), the notes sections’ “Note added” / “Note saved” / “Note deleted” (src/features/notesAndTags/RecordNotesSection.tsx,src/features/leads/LeadNotesSection.tsx),ActivityActionNotice,useLeadsNotice, the account page’s “Account saved”, and the in-houseToaston the Users list and the compose-email dialog (.forge/design/ui/shadcn-backfill.md). A failure’s inline message with Retry and a sticky notice stay inline (Rev 90). Each move keeps the success announced to a screen reader: a toast raised in the render that closes a dialog may land while Radix still holds the page’s Toaster regionaria-hidden(the “Hear Template saved and User saved” row). Each PR that moves a screen whose approved spec states the old frame adds a revision to that spec (F11-b, F20-c, F20-d, F5-d, F26-d; Rev 90 § Scope). Done when no modal sets its own width and no screen confirms a success inline. In progress, 2026-11-05:forge-et-febranchfeature/modal-frame-rev90adds the three tokens, thesizeoption of the shadcnDialogContentandAlertDialogContent(src/ui/shadcn/dialog-frame.ts), and movesRecordEditDialogFrame(New account, Edit account, Edit contact, the user and role editors, the template editor and its two questions, and the Access dialog) andConfirmDialogonto them, with F5-d Rev 12 and F20-d Rev 6 and the current prototype’s matching dialogs; the self-sized dialogs and the success messages are still to move. The same branch moves forge-et-fe#407’s short-window fix for BUG-277 onto the Rev 90 frame: only the body scrolls, and below a 400px window (320px until forge-et-fe#442, BUG-277) the frame tightens its padding, gaps and margins. Still to do on the New Account screen’s own Rev 90 PR: its description (NewAccountDialog.tsx’shelper) is three sentences where §5 Modal asks for one, and it now stays in view under the title rather than scrolling (forge-et-fe#419 review, finding 10). When BUG-277 closes (with forge-et-fe#442), a design-system revision adds to §5 Modal the short-window spacing the code now has: below a 400px window the card’s padding isspace/4above and below andspace/5at the sides, its gapsspace/3and its marginsspace/4, in place ofspace/7andspace/6(RecordEditDialog.module.css,data-height="short"; forge-et-fe#419 delta review, N-5). Edit Lead (forge-et-fe#422) and its leave question take the frame and thelgandsmsizes in forge-et-fe#419. Still to move, from that PR’s review: the in-houseModalcallers that setwidth(New Lead andInquiryExtractionDialog, bothwide, 720px), and the<AlertDialogContent>callers with nosize(src/features/leadResearch/ResearchPersonBlock.tsx, the lead research Remove question on the lead page beside Edit Lead’ssmquestion, andsrc/features/email/ConnectedPanel.tsx). Built, 2026-11-07:forge-et-febranchfeature/ui-rev100(one PR for the whole CRM, Nanndish’s decision of 2026-11-06, in place of one PR per area) moves every inline success tonotifySuccess(src/ui/notifySuccess.ts) and puts the self-sized dialogs on the frame, with the spec revisions on harness branchdocs/ui-rev100-specs-prototype. Still on the olderModalafter it, each for a reason given in its spec revision or the PR: New Lead andInquiryExtractionDialog(the extraction-to-form focus hand-off relies on the native dialog’s close-then-open order), the compose window (the provider relies on the native dialog’s close event, BUG-134;src/uiModalhas nolg),ArchiveContactModal(a busy state, three refusal panels and anonClosedfocus return),CreateUserModalandChangeManagerDialog(their react-ariaCombobox/OwnerPickerlists portal outside a Radix focus trap;ChangeOwnerDialogmoved onto the frame with F32-c-WI-1.2),CreateRoleModalandDeleteRoleAction(their tests assume the native<dialog>), and the pricingDraftPriceDialog(itsComboboxlist would be cut off by the frame, the row below). Done when those move and this row’s other conditions hold onmain. In progress, 2026-11-09: forge-et-fe#476 moves every one of those dialogs off the olderModal(onlyPriceEmailDialogkeeps its own shadcn frame, already Rev 90’s look) and stops the sized shadcnDialogContenthiding overflow, so a Combobox list in a dialog is no longer cut off. -
Recase the CRM to design-system Rev 90’s Title Case (Nanndish; decided 2026-11-04 and, for field labels, 2026-11-05, on the UI/UX designer’s recommendation): §9 now puts every title (a screen’s, a card’s or section’s, a modal’s, a tab’s, a legend, an empty state’s heading), the navigation, every button and menu item that acts, a standalone link, a table’s column headers, every field label (on forms and in read-only views) and an icon-only button’s accessible name in Title Case, with two listed word sets (short words left in lower case, particles capitalised after a verb); placeholders, tooltips, helper text, messages (a field’s error included) and the values shown (statuses, tags, product lines, filter pills, select options, value menus) stay sentence case. In
forge-et-fe, one PR per screen or area, each recasing its screen whole (§9 Scope and the move over), or with the modal move above where a PR touches the same screen: the route titles (handle.title, whichdocument.titlereads), the Settings cards and their pages (src/features/settings/tagSettingsTypes.ts’s Account tags and the rest), the page, card and dialog titles, the tabs, every button and menu item, each list’s column headers, the form and read-only field labels (the label lists insrc/features/accounts/accountForm.ts,src/features/leads/leadForm.ts,LeadTeaRequirementsGroup.tsxandsrc/features/leadResearch/LeadResearchPanel.tsx, which say “sentence case” today, and every other form’s) and icon-only buttons’aria-labels. A field’s error message that names the field keeps sentence case, so an error that is built from the label must lower its case.src/features/leads/leadsListParams.ts’sSORT_FIELD_WORDSserves both the column headers and the sort options (“Buyer type A–Z”), which stay sentence case, so split it first. Add one test helper that applies §9’s two word lists, and make it replacenavConfig.test.ts’s Title Case regex (which rejects the lower-case words and codes) rather than sit beside it. The tests and comments that state the old rule, fromgit -C forge-et-fe grep -niE 'sentence.case|title.case' src e2e:src/features/leads/LeadsTable.test.tsx,src/features/email/MailboxRoute.test.tsx,src/features/users/UserActivationCard.test.tsx,UserEditorRoute.test.tsx,UsersScreen.test.tsx,src/features/settings/SettingsTagsRoute.test.tsx,src/features/leads/LeadTeaRequirementsGroup.test.tsx,e2e/browser/accounts/accounts-layout.spec.ts(F1-b AC-2),e2e/browser/email/templates-list.spec.ts,e2e/browser/mailbox/connect-screen.spec.ts,e2e/browser/mailbox/support.ts,e2e/browser/shell/frame.spec.ts;src/app/tableHeaderStyle.test.tsonly renames its test (it asserts nouppercaseclass, still true); the comments insrc/ui/dataTable.ts,src/ui/components/{Modal,Tabs,DropdownMenu,SortableHeaderCell,Toast},MailboxRoute.tsx,RolesRoute.tsx,CurrentUserMenu.tsx,NotesDialog.tsx,SettingsDialog.tsxandStandardsTable.tsx. A test that finds a control by its name or label breaks on recasing in unit tests (Testing Library matchesnameandgetByLabelTextcase-sensitively) and may not in browser tests (Playwright’s default does not). Each PR adds a revision to a spec whose requirement or acceptance criterion states the casing rule, for the screens it recases: F1-b, F1-d, F1-e, F1-f, F11-b, F11-d, F11-f, F12, F20-a, F20-b, F20-d (grep -rliE 'sentence.case' .forge/specs). Done when the helper runs over every route title,navConfig, the Settings tag titles, each table’s column-header list and each form’s label list, and thatgit grepfinds no comment or test that cites the old rule. The capture app takes the same rule (design-system Rev 105, Tharinda, 2026-11-08); its row is below. In progress, 2026-11-05: the helper isforge-et-fesrc/test/titleCase.ts(titleCaseProblems, with its own tests), and it replacesnavConfig.test.ts’s regex, on branchfeature/modal-frame-rev90; no screen is recased yet. Built, 2026-11-07:forge-et-febranchfeature/ui-rev100recases every CRM screen whole, in one PR (Nanndish’s decision of 2026-11-06), with atitleCaseProblemstest per area over its label and column-header lists, and the spec revisions on harness branchdocs/ui-rev100-specs-prototype. Remove this row when that PR merges and thegit grepabove finds no test or comment citing the old rule. -
Redraw the capture app’s two opaque icons on the CRM theme’s background (Nanndish; raised 2026-11-13 by F36-a WI-1.2, forge-et-pwa#127): the installed app takes the CRM theme’s colours (Nanndish’s decision of 2026-11-13), so
theme-crm.csschanged the background to--et-background(#fcfaf4), and forge-et-pwa#127 moved the manifest’sbackground_color(#fcfaf4) andtheme_colorand thetheme-colormeta tag (#ffffff, the header’s--et-surface-card), held to the tokens bysrc/ui/styles/captureStyles.test.ts.public/apple-touch-icon.pngandpublic/icons/icon-maskable-512.pngstill have the old#f4eee2baked into their opaque background (their corner pixels, checked 2026-11-13;icon-192.pngandicon-512.pngare transparent and need nothing). Regenerate both from the logo on#fcfaf4(forge-et-pwa/CLAUDE.md§ Design tokens). Done when both icons’ background is#fcfaf4. -
Check the capture app’s Sign In after Back in Chrome’s back/forward cache (Nanndish; raised 2026-11-13 by the forge-et-pwa#128 review, Important #1):
SignInRouteresets its?login=notice and the busy Sign In With Microsoft when the browser restores the page from its back/forward cache (src/lib/browser/usePageRestore.ts), which the unit tests prove with a syntheticpageshowonly. Run it by hand onpnpm previewin Chrome (the dev server is not cached): open/sign-in?login=failed, press Sign In With Microsoft, go Back from Microsoft’s page, and check that the page comes back from the cache (DevTools › Application › Back/forward cache) with no notice and the button usable. Done when the check is recorded on the PR or in F36-a’s plan. -
Build and draw the capture app in Rev 90’s casing (the author of each F36 PR in
forge-et-pwa; decided by Tharinda, 2026-11-08, design-system Rev 105):forge-et-pwafollows §9 as the CRM does. The shell part is done in F36-a-WI-1.2 (2026-11-13): the route title is Page Not Found (src/app/router.tsx,e2e/pageStates.ts), the buttons are Reload the Page and Back to Leads (Leads is the home tab;src/app/AppErrorBoundary.tsx,src/app/RouteErrorElement.tsx,src/app/NotFoundRoute.tsx), the tests find them by those names (src/app/routerChunkLoad.test.tsx,src/app/AppErrorBoundary.test.tsx,src/app/RouteErrorElement.test.tsx,e2e/browser/shell/a11y.spec.ts), and design-system Rev 120 records that the shell’s failure headings are messages, in sentence case; the sign-in, No Access, tab and Settings screens are built and drawn in this casing. Each F36 screen is built in Title Case for titles, navigation, buttons and menu items that act, standalone links, column headers, field labels and icon-only buttons’ names, with placeholders, helper text, tooltips, messages and values in sentence case. The current capture-app prototype (.forge/design/prototypes/pwa-current/pwa-prototype.html) still shows the POC’s sentence case (Delete lead, Archive campaign, All leads); the PR that builds a screen recases that screen in the prototype too. Each F36 spec cites §9 rather than restating the rule. As the CRM’s row does, add a casing check likeforge-et-fe’ssrc/test/titleCase.tsonceforge-et-pwahas its first label lists. Done when every built capture-app screen and its prototype screen are in this casing. -
Check the capture app’s iOS install message on an iPhone and an iPad (Nanndish; raised 2026-11-19 by the forge-et-pwa#186 review, findings 5 and 9; narrowed by F36-a Rev 7 the same day): Settings’ install card shows one line on iOS, In Safari, tap Share, then Add to Home Screen. (
forge-et-pwasrc/features/settings/InstallCard.tsx,IOS_INSTALL_MESSAGE; F36-a Rev 7, design-system §9 Shell chrome, which says the wording is checked against the iOS and iPadOS Safari current at the build). The tests use Safari user agents in Chromium and WebKit; nobody has followed it on a real device. On an iPhone and an iPad running current Safari, open the capture app and check that tapping Share, then Add to Home Screen, installs it as the line says. Done when the line is confirmed or reworded, with the result recorded on the PR that changes it or in F36-a’s spec. -
Choose one look for the capture app’s invalid controls (Nanndish; raised 2026-11-19 by the BUG-510 review, finding 4): an invalid field is drawn three ways. The built capture app changes only the border’s colour, at 1px (
forge-et-pwasrc/ui/TextInput.module.css, the invalid rule);pwa-currentdraws a 2pxborder-destructive(source/src/components/lead-detail.jsx,exhibitor-forms.jsx; screenshotphone/24-lead-edit-details-error.png); the CRM draws the danger border with atone/dangerbackground fill (forge-et-fesrc/ui/components/Input/Input.module.css). Design-system §5 Form fields and validation, Errors, says only “tone/dangerborder on the control”, with no width or fill. Decide whether the capture app takes the CRM’s fill or the prototype’s 2px, then state it in §5 with a design-system revision, and bring the capture app’s controls andpwa-currentto it. Done when the three agree. -
Decide the capture app’s input text size for iOS Safari (the author of the first F36 PR with a text field in
forge-et-pwa; raised by design-system Rev 109, 2026-11-10): iOS Safari zooms the page when a focused input’s text is under 16px. The capture app’s running text is 14px (capture.cssraisestype/secondaryandtype/metatotype/card-title’s size;type/bodyis 14px since Rev 109), so its fields will zoom on focus on an iPhone. The type scale has a 16px size (type/section), so one option is to set the capture app’s input text to 16px incapture.css; the other is to accept the zoom. Never stop the zoom withmaximum-scalein the viewport tag, which blocks pinch zoom (WCAG 1.4.4). Done when decided and, if 16px, built with the first text field. -
Bring the current capture-app prototype onto the 12–24px type scale (Tharinda, who took over the type-scale PRs from Nanndish — forge-et-pwa#116 and forge-et-harness#1266; raised at the hand-over, 2026-11-13): design-system Rev 109 leaves
prototypes/pwa-current/unchanged; it was written when that folder was a copy of the POC freeze. Since thenmainhas given it asource/folder and saved versions for F36-a, and on 2026-11-13 it still set sizes such as 9, 10.5, 15, 19, 22, 28 and 34px (grep -oE 'font-size:\s*[0-9.]+px' .forge/design/prototypes/pwa-current/pwa-prototype.html). Decide with the F36-a author whether the prototype and its source take the six sizes now or as each F36 screen is built, record the answer in a design-system revision (Rev 109’s Scope points here), and tell the F36 authors that a capture-apph1takessrc/ui/PageTitle.module.css(forge-et-pwa#116). Done when the prototype follows the scale or a design-system revision says when it will. -
Put a leading icon on every button in the CRM (Nanndish; design-system Rev 100, decided 2026-11-06): every button with a text label, dialog buttons and the save bar included, takes a leading
aria-hiddenicon from the concept registry, one glyph per concept. Do it once in the shared parts where a button’s concept is known (ConfirmDialog, the dialog frame’s Cancel, the save bar’s Save and Discard), then screen by screen, with the Rev 90 recasing PR where a PR touches the same screen. Add the concepts the registry lacks. Done when a test fails a labelledButtonwith no icon insrc/featuresandsrc/app, and nothing it lists remains. Built, 2026-11-07, onforge-et-fefeature/ui-rev100: the concepts are insrc/ui/icons/registry.ts, the shared parts draw theirs, and every area’s labelled buttons lead with one. Still to do: the enforcing test, and a decision on the concepts the implementers asked for and replaced with the nearest one —dismiss,link,unlink,replyAll,markRead/markUnread,reset,goToPage,openEmail,later,makeMain,changeManager(drawn withreassign),withdraw(drawn withretire) andcarryForward(drawn withnext). (keepEditingis no longer asked for: since design-system Rev 110 Cancel and the safe answer in a question take no icon.) -
Make inquiry extraction’s text agree with the control-character rule (Dinil, owner of F12; raised by
backend-revieweron forge-et-be#409, 2026-11-11): since BUG-403, a lead’s one-line text fields refuse a control character (tab, U+0085, U+2028 and the rest) with400 INVALID, butInquiryExtractionRules.text()(forge-et-belead/service/InquiryExtractionRules.java) strips only the ends, so a value the model returns with an inner tab prefills the form and the save is then refused naming that field. Either replace such characters with a space intext(), or move the value tounmatched. Done when the extraction never prefills a value the create refuses. -
Confirm or reverse the decisions taken overnight on forge-et-fe#442 (Nanndish; taken 2026-11-07 under his instruction to finish the PR without stopping for questions): design-system Rev 103 (a success inside a dialog the user stays in is also said in the dialog’s polite line when the toast region cannot be heard; sticky notices stay inline; Reload uses
refresh, the safe answer usescancel— reversed by design-system Rev 110, 2026-11-11, which gives the safe answer no icon; Sign in with Microsoft), F11-f Rev 11’s place for F3-a’s Follow-ups (straight after the AI summary, asmainhad it), F26-g FR-18’s visible line for skipped standards, the/settings/users/:id/offboardaddress redirecting to the user page, the Leads and Contacts notices sharing onesrc/uinotice (StickyNotice, whose Dismiss is an icon-only button beside the message, so the Leads notice’s Dismiss is no longer an outline button under it), and the dialogs left on the olderModalwith their reasons (the Rev 90 row above). Done when each is confirmed or changed. -
Decide whether a click outside the follow-up dialog may close it while it has unsaved edits (Nanndish; from forge-et-fe#442’s review fixes, 2026-11-07): design-system Rev 90 §5 Modal says a click outside does nothing while a dialog has unsaved edits, but the follow-up dialog (
forge-et-fesrc/features/activities/FollowUpDialog.tsx, now onRecordEditDialogFramethrough itsonInteractOutsideprop) still asks to close, as it did before the PR; it was kept to change behaviour least. Done when decided and, if it changes, fixed with a test. -
Draw Rev 90’s casing, toasts and dialog sizes and Rev 100’s button icons across the rest of the current prototype (Nanndish; from
forge-et-fefeature/ui-rev100, 2026-11-07): that branch recases every CRM screen, leads every labelled button with an icon, moves every success to a toast and sizes the dialogs, but the current prototype (.forge/design/prototypes/crm-current/) was changed only for Rev 100’s three screens (the tag tables, the Deactivatedialog, the lead page) in the same delivery. Every other screen there still shows sentence-case titles, buttons, column headers and labels, buttons without icons, and inline success lines. Done when the prototype’s screens match the built app on these four points ( .claude/rules/plans.md § Required Sections). -
One shared
InertWhileOpenmodule inforge-et-fesrc/ui/shadcn(Nanndish, author ofsrc/ui/shadcn/select.tsx, or Dinil, whichever starts first; raised by F1-f-WI-1.3’s plan Rev 1, 2026-10-14, and byfrontend-revieweron forge-et-fe#286, 2026-10-16):select.tsxkeeps a localInertWhileOpenthat addsinertto what Radix marksaria-hiddenwhile its list is open. F1-f’s adapteddialog.tsx(forge-et-fe#287, WI-1.3 Decision 23) and F20-d’salert-dialog.tsx(forge-et-fe#286, F20-d overnight log O-9) each carry a copy, so three copies are onmain. Move them onto one module that all three import, in a smallforge-et-fePR. Its tests must still prove that only the elements it made inert are released, and that<main>'s scroll lock is restored. -
One record combobox in
forge-et-fe(Nanndish; raised byfrontend-revieweron forge-et-fe#296, F3-a wave 1, 2026-10-18; deferred by the developer that day):src/features/activities/RecordPicker.tsxis a near-copy ofsrc/features/accounts/AttachAccountPicker.tsx— the same keyboard handling, focus after retry and Popover wiring (sincemain’s BUG-209 fix, both leave the retry hold to the sharedusePickerData). Move one combobox intosrc/ui(orsrc/lib/accountPickerbesideRecordPickerData) with the kind, field id,requiredandonBluras props, and put both callers on it, in a small PR whose existing tests stay unchanged; give the empty list arole="listbox"soaria-controlspoints at it. Whether both pickers later move to shadcnCommandwaits on approving thecmdkdependency (ASK FIRST). -
One list pager in
forge-et-fe(Nanndish; raised byfrontend-revieweron forge-et-fe#296, F3-a wave 1, 2026-10-18; deferred by the developer that day):src/features/activities/ActivitiesPager.tsxis the fourth copy of the list pager, afterusers/UsersPager.tsx,accounts/ContactsPager.tsxand the one inemail/templates/EmailTemplatesScreen.tsx; they differ in test ids, the label and an optional size select. Extract aListPagerintosrc/ui(atestIdPrefix, a label, an optional size slot) and move all four onto it in a small PR. -
One record gate for child records in
forge-et-be(Nanndish, or whoever next adds a child of an owned record; raised bybackend-revieweron forge-et-be#273, F3-a wave 1, 2026-10-18): three copies of the same gate now exist — find the record’s owner,can(<view key>)else404,require(<manage key>)else403, the “<Noun> <id>was not found” message:auth.service.OwnedRecordGates(shares and assignee candidates),lead.service.LeadChildAccess(lead notes and tags) andactivity.service.ActivityRecordAccess(follow-ups). Each callsPermissionChecker.canwithOwnershipContext.of, so reach itself is not restated. Publish anauth.accessport forvisible/requireManageable, haveOwnedRecordGatesimplement it, and move the other two onto it in a small refactor PR with their existing tests unchanged. -
Keep caller text out of the follow-up create body’s
toStringinforge-et-be(Nanndish; raised by the forge-et-be#327 review fix, F3-a wave 2, 2026-10-28):- What is wrong:
activity.ActivityFieldRules.Input.toString(), used on create, printstypeIdandassigneeIdraw.ActivityCreateRequestbinds them asJsonNode, so a string a caller sends there would appear in a log line. - Why it is not in #327: that PR fixed the same thing for the edit body (
PatchInput,ActivityPatchRequest) by printing a value only when itisIntegralNumber(). The create body is wave 1 code, already onmain, which #327 does not change. - Fix: apply the same
number(...)helper toInput(andActivityCreateRequest.toString, if it prints them), with a unit test that a string id does not appear. Nothing logs these records today, so it is latent.
- What is wrong:
-
Check whether the shadcn AlertDialog returns focus on close without
onCloseAutoFocus(Tharinda to assign; raised byfrontend-reviewer’s delta review of forge-et-fe#314, 2026-10-20):src/ui/shadcn/alert-dialog.tsx’s comment says Radix returns focus to what held it before the dialog opened. In a jsdom test of a dialog opened by a plain button (openheld in state, noAlertDialogTrigger), focus went to<body>on close instead. The one caller,TemplateEditorForm.tsx, sets focus itself throughonCloseAutoFocus={restoreFocus}, so no screen is affected today. Check it in a browser withoutonCloseAutoFocus: if focus is lost there too, file a bug (it may come fromInertWhileOpenreleasing the page after Radix’s return), fix it or correct the comment, and add the assertion tosrc/ui/shadcn/alert-dialog.test.tsx. -
Choose the WebKit focus key in the size matrix’s
leads-lead-share-unownedstate before its Tooltip moves to shadcn (whoever moves the lead Share… button’ssrc/ui/components/Tooltipto the shadcn Tooltip; raised byfrontend-reviewer’s delta review of forge-et-fe#317, 2026-10-21; L-168):forge-et-fee2e/browser/size-matrix/arrange.ts(the state, near line 7433) reaches Share… with plain Tab and Shift+Tab throughe2e/browser/pricing/round-support.tskeyTo, in both WebKit size projects. It passes on a Mac with keyboard navigation off only because react-aria’sFocusablegives the button an explicittabIndex, which WebKit’s plain Tab still reaches; a Radix Tooltip adds none, so the state will then fail on such a Mac, as BUG-241’s rows did. In the same change, havekeyTotake aFocusKeyand the WebKit projects passe2e/support/focusKey.tswebkitFocusKey’s, or reach Share… another way. -
Follow-ups whose assignee no longer reaches the record (Tharinda — at the Phase 1 review; raised by the F3-a spec, Q-4, 2026-10-13): an active assignee who loses reach to the record (the record is reassigned outside their reporting line, or the share that gave them reach is removed) keeps the follow-up but no longer sees it, and the record owner’s list does not pick it up; a follow-up on an unowned lead whose assignee is deactivated lands in nobody’s Mine list, and only
lead.allholders see it. Both stay visible on the record’s Follow-ups panel and under All I can see. Accepted as a Phase 1 limitation (.forge/specs/F3-a-follow-up-tasks-spec.mdQ-4). Decide at the Phase 1 review whether a reassignment or unshare should offer to move the record’s open follow-ups to the new owner. -
A dev-environment reading for F3-a’s response-time limit, NFR-4 (Nanndish; raised by the F3-a-WI-1.1 plan review, 2026-10-18): F3-a NFR-4 sets 300 ms at p95 for
GET /api/v1/activitiesand/counts, measured on the local Postgres, and FR-H3 says nothing in F3-a is cloud-gated..forge/test-strategy.md § Performance Baselines(L-084) says a laptop reading is evidence only and the binding reading is taken in a deployed environment. The developer decided both apply: WI-1.1’s local reading binds its PR (a breach blocks it until explained), and a dev-environment reading is also owed. Write F3-a spec Rev 2 revising NFR-4 and FR-H3 so the dev reading is a.forge/cloud-gated-validation.md§B row for F3-a, written by WI-4.1 in place of its “nothing cloud-gated” line; dev has served the backend since 2026-10-09. -
Ask Empire IT for their Microsoft 365 maximum send size (Tharinda, client contact; raised by the F20-h spec, 2026-10-13): F20-h limits an email’s files to 25 MB in total, which fits Exchange Online’s default 35 MB after encoding (decision 2). If Empire’s tenant sets a lower limit, F20-h’s limit is lowered by a spec revision.
-
Make the email timeline row’s write atomic when F20-f adds its producer (whoever builds F20-f; raised by F26-d’s wave review, forge-et-be#271, 2026-10-17): F26-d writes the timeline row keyed
('email_send', <send id>, 'email')asTimelineEntryService.existstheninsert(orrewriteSummary). With a second producer of that key (F20-f’s Sent Items link), a concurrent insert failstimeline_event_source_key, which nothing translates, so the price-email record answers 500. F20-f’s spec makes the write an upsert (INSERT … ON CONFLICT (source_type, source_id, kind) DO UPDATE SET summary = …inTimelineEntryService) or retries once. -
Keep F37-b’s bounce step on F20-f’s header-only read (whoever plans F37-b, or F20-f WI-1.1 if F37-b lands first; raised by the review of F20-f’s Decomposition Plan, 2026-11-19): F20-f’s ingest makes a header-only first read and stores only linked mail (F20-f spec D-4 Rev 1, FR-5). A non-delivery report links to no record, so it is no longer stored. F37-b’s bounce step (F20-b Rev 8, F37-b FR-D1) must therefore run on that header read, and its added
MESSAGE_SELECTfields must be part of it. If F20-f wave 1 is onforge-et-bemainwhen F37-b’s plan is written, F37-b’s plan puts its step there; if F37-b is there first, F20-f WI-1.1’s plan keeps it. Done when both are onmainand a bounce in the stub is still recorded with the header read only. -
Add a publish-only command to the browser gate (Dinil; raised by F26-d’s Stage 12, 2026-10-17, findings F-004):
forge-et-fe’sscripts/browser-gate.mjscan push a passing result only at the end ofrun --push. When that push failed during a connection drop, the stored result could be published only by hand, withcovers --head HEADandgh pr edit. Apublishcommand that pushes the branch and writes the stored line would make that a supported step. Make it inforge-et-feand re-copy it toforge-et-pwa(scripts/shared-with-fe.json). -
Decide whether a send left at
SENDINGshould stop holding its files (Tharinda, lead decision; raised by the review of forge-et-be#262, F20-h Wave 1, 2026-10-15): a send whose task stopped mid-send (a deploy, a crash) staysSENDINGuntil someone repeats its key, because F20-c marks a row stale only on a repeat. While it isSENDING, its files count as in use: the user’s Remove gets409 EMAIL_ATTACHMENT_IN_USE(hidden by FR-C7) and the hourly clean-up skips them, so they are never deleted, against FR-R3’s 24-hour intent. The fix both reviewers suggest is to treat aSENDINGrow past its own stale bound (created_at + deadline_seconds + 20 s, the expressionmarkStaleuses) as not in use, inEmailSendFileRepository.existsSendingNamingandfindCleanupCandidates. It changes FR-R3 and FR-U6’s “stillSENDING”, so it is a spec revision; if accepted, it is a small F20-h follow-up. -
Decide whether the server checks an attachment’s name against its type (Tharinda, lead decision; raised by the same review, 2026-10-15): the extension check of FR-C3 is client-only, so through the API a file whose content is a PDF can be sent named
invoice.hta, and the recipient’s mail client opens it by its extension. Names may also hold C1 control characters (U+0080–U+009F) and bidirectional overrides such as U+202E, which can makefdp.xlsxread asxslx.pdf. Refusing a name whose extension does not belong to the sniffed type, and Unicode format (Cf) and C1 characters, is cheap but changes the F20-h Wave 1 contract’s step 3, so it needs a spec revision. -
Check in F20-h’s real sends that Graph does not page a message’s attachment list (Dinil, owner of F20-h; raised by the same review, 2026-10-15):
GraphMailboxProvider.listAttachmentsreads one page and ignores@odata.nextLink. If Graph pages a large message’s attachments, a forward’s removals on a later page are refused400and the kept originals are under-counted. Confirm with F20-h WI-2.1’s real forward of a message with attachments, or from Microsoft’s documentation, and follownextLinkif it can page. WI-2.1’s forward (2026-10-16) read a two-attachment message in one page, which does not test paging; Microsoft’s documentation is the remaining way to answer it. 2026-11-20, read: Microsoft’s documentation does not rule paging out. The List attachments page says the method “supports the OData Query Parameters” and gives no page size; the Graph paging guide says default and maximum page sizes differ by API and that a client follows@odata.nextLinkuntil it is absent. So the fix stands:listAttachmentsfollowsnextLink, with a test of a two-page answer — a smallforge-et-bechange (F20-h), no spec change. -
Ask Empire whether reps send password-protected Office files (Tharinda, client contact; raised by F20-h-WI-1.1’s plan, 2026-10-13): Office saves an encrypted file in the older container format, so F20-h refuses it as it refuses a legacy
.xls(spec decision 1). If reps send password-protected price lists, accepting them is a spec revision. -
Ask Kavindu what the undecided trade-fair words mean (Dinil, owner of F4-b; carried from F4-a spec Rev 5, the F4-b tracker notes and F4-b spec W-4, 2026-11-06): in the Pipedrive Trade Fairs column,
Hotel,LK,SL,Belgium,EMP,Seoul,T,EK,ChamberandDistributionare not recognised fairs (F4-a Rev 5); until Kavindu answers they are imported as typed and flagged in the dry-run report (F4-a FR-N2). The F4-b tracker notes also leave undecided whether nationality words (Indian, Bulgarian) andTrinidadare countries; confirm which column they come from first — in the country column F4-a FR-N1 applies (the country is left empty and listed), in the trade-fair column FR-N2. Ask before the first production run; an answer that maps a word to a fair, a source or a country is an F4-a mapping change. -
Record a lead’s conversion as a
statuschange, so the import treats a converted lead as changed in the CRM (whoever builds F11-e; raised by the F4-b-WI-1.2 plan and its review, 2026-11-06): F4-b spec FR-13 says a lead a person converted is changed in the CRM, and FR-15 counts a person’s event only when it changes a mapped field. No conversion path exists onmain(convertedis only a vocabulary value);LeadStatusTransition’s javadoc says F11-e adds conversion as a transition. If F11-e’s conversion event recordsstatusin itschanges, a later Pipedrive run leaves a converted lead alone; if it records anything else, the run would overwrite the converted lead. F11-e’s spec states it. -
The cost annexe’s dev figure is about half of what dev costs (Tharinda — decide before the annexe is quoted to Empire again; 2026-11-18: the annexe’s §4 UAT figure was restated on dev’s measured cost, about $9–14 a week and $100–155 for the one instance kept from Phase 1’s round to the end of Phase 2’s, AD #23 revised; the dev and production rows are still to do):
.forge/client/sow-cost-annexe.mdputs dev at about $34 a month, but dev as built runs at about $62 (measured for 10–12 Oct, read on 14 Oct), because it has a load balancer the annexe left out, the backend runs around the clock, and public IPv4 addresses (about $11 a month in dev, $7–11 in production) are in neither environment’s figure (environments-and-identity.md§4 “What dev costs”). The off-hours schedule that would bring dev to about $40–45 (F38 FR-16) is deferred until development is complete (the lead, 2026-10-14 — F38 Rev 10, and a post-issue note on the proposal), so the annexe’s dev and production rows are to be re-derived on the figures as built. The annexe is client-facing, so a changed figure is stated as a dated revision, never edited silently. UAT too (F38-h spec NFR-1, 2026-10-23): §4 prices a round at about $15–20 for two weeks without a load balancer; Phase 1’s round runs about three weeks (F38-h decision 7) with a load balancer and three public IPv4 addresses, about $40 at list price. F38-h’s rehearsal measures the actual figure and adds it here, so the annexe is decided once for dev, UAT and production. (2026-11-19, F38-h spec Rev 13: the rehearsal no longer measures UAT’s cost — the cost-allocation tag is the billing owners’ to turn on if they want it, the lead’s decision; this row’s UAT figure stays the list-price estimate.) Production too (an estimate at list price, 2026-10-27, during F38-i’s decomposition; nothing is measured): about USD 100–105 a month before AI, against the annexe’s USD 90. The annexe leaves out public IPv4 (about USD 11) and the KMS keys and CloudTrail storage (a few dollars), and counts USD 5 for GuardDuty, which is off. Dev’s database, read inforge-et-iaacbase/rdson 2026-10-27, is the module defaultdb.t3.microon gp2 storage, not the annexe’st4g.micro. The lead decided not to build the Multi-AZ, WAF and cross-region backup options (+USD 30, +15 and +8; F38-i decision 2 in the engagement-log entry of 2026-10-29, “F38-i’s spec revision”). October measured (F38-i FR-14, read 2026-11-02): dev used USD 51.26 in October (a part month, from 9 Oct; no credits or tax) and runs at about USD 63 a month as built (25–31 Oct × 30), against the annexe’s USD 34; each line is compared inenvironments-and-identity.md§4 “What dev cost in October 2026”, which also notes that annexe §3’s static-hosting row and §6’s capture-app line may both count the capture app’s hosting. -
When development is complete, decide what happens to dev (Tharinda): scale it down, schedule it off outside working hours (F38 FR-16, deferred by Rev 10), or decommission it. It costs about USD 62 a month on Empire Teas’ bill while it runs around the clock.
-
Stop the note views handing the
Noteentity to the web layer (Biseni; raised by F1-e WI-1.1’s pre-PR review, 2026-10-14):RecordNoteService.NoteView(F1-e, account and contact notes) andLeadNoteView(F11-f, lead notes, onmain) each carry theNoteentity out of the service layer, against.forge/design/coding-standards.md§2 (“No entity leaves the service layer”). F1-e copied the lead shape on purpose (WI-1.1 plan Decision 6), so the two are fixed together: a field-only view record built in the service, with no change to either API’s response. The lead notes’ edit and delete (F11-f Rev 13,forge-et-befeature/polish2-lead-notes, 2026-11-11) keptLeadNoteView’sNotefield and added an editor name beside it, so it is now done separately; agree the timing with Dinil, who owns the lead notes. -
Fold F1-e’s two account-child gates into one (Biseni; deferred from the forge-et-be#265 review, 2026-10-15):
account/service/NoteParentAccess.java(notes, WI-1.1) andaccount/service/RecordTagAccess.java(tags, WI-1.2) enforce the same reach rule for account and contact children, line for line, withNoteTargetandTaggedRecordas twin enums. A third copy isContactService.loadVisible/requireManage/contextOf. They agree today; the reviewer checked contact reach through the account, the parent’s shares, andaccount.allfor a contact with no account. Owed: one package-privateAccountChildAccesswith oneACCOUNT/CONTACTenum carrying the note codes, the tag codes and the tag entity type, used by both services, with no behaviour change. Deferred because the two work items were built in parallel and this is a refactor with no defect behind it (WI-1.2 plan## Notes1). -
Move the record picker (the Add tag combobox) into
src/ui(Biseni; raised by the forge-et-fe#284 review, 2026-10-16): F1-e’ssrc/features/notesAndTags/AddTagPicker.tsxand F1-f’sAttachAccountPicker(forge-et-fe#269) are two hand-built ARIA 1.2 comboboxes on the shadcn Input and Popover, against the rule of two. Owed: whichever of the two PRs merges second moves the shared picker intosrc/uiand points both features at it. F1-e merged second, and forge-et-fe#284 merged without the move to avoid another full browser cycle, so it is F1-e’s follow-up PR. -
Put F1-e’s two loading states on the shared
SkeletonLines(Biseni, owner of F1-e; raised byfrontend-revieweron forge-et-fe#348, 2026-10-27): forge-et-fe#348 addedsrc/ui/skeletonLines.tsx, the §5 loading state on shadcn’s Skeleton (a busyrole="status"stack of body-line bars, the barsaria-hidden, the pulse stopped under reduced motion), for the Email screen and the message view.src/features/notesAndTags/SectionSkeleton.tsxandsrc/features/settings/SettingsSkeleton.tsxare two earlier copies of it. Point both features atSkeletonLinesand delete the copies, keeping their test ids and labels, in a smallforge-et-fePR. -
Show the selected row in Windows forced-colours mode (Tharinda to assign; raised by the review of forge-et-fe#348, 2026-10-27):
RowSelection’s selected rows mark selection with anaccent/gold-deepinline-start border on a hover fill. In forced-colours mode the browser paints every row’s border in one system colour and drops the fill, so the selected row looks like the others; onlyaria-current(or the checkbox) still says which it is. The Email list’s open message has, since design-system Rev 84, the sidebar’s treatment (a rounded fill and a 3px::beforebar), which forced-colours mode also drops. For the Email list and the sidebar, one option needs no@mediarule (which the email feature’smailboxTokens.test.tsforbids): draw the bar as a border on the pseudo-element (inline-size: 0; border-inline-start: var(--et-border-width-selected) solid var(--et-accent-gold-deep)), which forced colours paints in a system colour on the selected row only (frontend-revieweron forge-et-fe#388, N-6). The sidebar’s current item has, since forge-et-fe#373 (design-system Rev 78), a white fill, abox-shadowring and a 3px bar drawn as a::beforebackground; forced-colours mode drops all three, leaving only its semibold label andaria-current. The Email folder column’s current folder has, since design-system Rev 91 (forge-et-fe#403), the open message’s look, with the same forced-colours gap; since Rev 95 it is on the folder’s row (FolderColumn.module.css.row[data-selected]::before), and the collapsed folder pane’s current tile has it too (.tile[data-selected]::before). Fix it once for all five, for example with a@media (forced-colors: active)rule that gives the selected row a thicker orHighlight-coloured edge, and check it in a browser with forced colours emulated. -
Finish the current CRM prototype’s 2026-11-06 catch-up (Nanndish; left by the catch-up to
forge-et-fe4a91ed9c,.forge/design/prototypes/README.md):.forge/design/prototypes/crm-current/crm-prototype.dc.htmldoes not yet draw these built parts: the lead list’s and lead page’s stale-data bands; the CRM below laptop width (the sidebar’s automatic rail at laptop width and its tablet drawer, and the Email screen’s narrower layouts); and offboarding’s notices for a caller who can only view a record or cannot reach it. The Settings › Lead tags page also lists three lead tags while the Leads screens add a fourth, Pipedrive: Hot; make the two lists one. Done when each is drawn, or the lead decides it is not worth drawing. -
Measure the sidebar’s colours under the revamp layer (Nanndish; raised by
frontend-revieweron forge-et-fe#373, 2026-11-02):Sidebar.tsxputsdata-et-theme="revamp"on the sidebar on every route, butsrc/app/sidebarContrast.test.tsmeasures only the Empire and neutral themes.revampContrast.test.tscovers the selected entry’s glyph and bar onsurface/card; the count badge on the selected entry under revamp (about 12 : 1 when worked out by hand) is asserted nowhere. Either extend the sidebar test to the revamp theme (itstheme()resolver needscolor-mixsupport) or add the badge pair torevampContrast.test.ts, in a smallforge-et-fePR. -
F1-e frontend: the nine Nits left by the forge-et-fe#284 delta review (Biseni; 2026-10-16). Left so as not to re-run the two-hour browser gate before merge. None changes what a user sees today. Fix them in one small follow-up PR (
.forge/plans/F1-enotes carry the review):src/ui/shadcn/radio-group.tsx: an arrow key can callonValueChangetwice with the same value. Click only after keyup, and add a held-key test.- The
radio-group.test.tsx“Tab or a pointer” row cannot fail as written. e2e/pageStates.tsaccount-detail-notes-leave: passACCOUNT_DETAIL_DIALOG_OPEN.- The
tagSettingsQueries.tscomment (“a tag list is one page”) overstates the premise. - The
tagSettingsErrors.tscomment: Retire and Restore re-read before writing, so only the editor’s Save gets the409. RecordTagsSection.tsxWRITE_UNCONFIRMED: change it to “are being read again”.- Under a contextual denial, Add tag stays operable and a pick is refused silently. This is latent, because no
ContextualBlockwraps it. NotesDialog.tsxEsc handling: also requirecurrentTarget.contains(target)for nested dialogs. This is latent, because there is no nesting today.RecordTagsSection.tsxremove with an unknown outcome: clearfocusAfterRemovewhen the re-read still shows the tag.
-
Promote the active-by-default list filter to
platform(Biseni; deferred from the forge-et-be#265 review, 2026-10-15):tag/service/TagService.defaultToActiveis the fourth copy of one helper, afterCertificationService,AccountServiceandContactServiceonmain, against the coding-standards Reuse rule. Owed:ListQuery.withDefaultFilter(name, value), or the same inplatform.query, used by all four, in one backend PR with no behaviour change. -
State in
.claude/rules/specs.mdhow a later feature supersedes part of a done spec (Tharinda — a harness change; raised by the F42-WI-1.3 plan review, 2026-10-27): step 2 says a revision also updates the spec’s body, but when a later feature supersedes part of a done spec the engagement writes the revision entry only — “nothing in this body; FR-… is superseded in part by” — and leaves the body as built (F20-a Rev 6, F20-b Rev 2, the F20-l set, and F42’s F12 / F11-d / F11-c entries, decided 2026-10-27 in the F42-WI-1.3 plan’s Q-1). Add one line to step 2 naming that form, its reason (the body stays the contract the done feature was built and accepted against, and the rule lives once, in the later spec), and that the body update applies only to a spec whose own feature changes its requirements. The change rewords a rule, so it gets the fresh-context reviewer of .claude/CLAUDE.md§ Compounding Engineering item 6. -
State whether a read-only harness row counts under
forge-worktree-upstep 2 (Tharinda — a harness change; raised by the F20-i-WI-1.3 plan review, 2026-10-29): step 2 calls aforge-et-harnessrow beside an app-repo row a plan defect, butforge-plan-authorrequires every WI on a contracted seam to list the frozen wave contract in## Files to Modifyas read-only (the F20-i WI-1.1, WI-1.2 and WI-1.3 plans do), which matches the rule as written. Say in step 2 that a row marked read-only is not a touched repo, so provisioning does not abort on it. Done when step 2 andplan-reviewer’s check say the same. -
Say where the selected browser run’s backend gets the email smokes’ bucket overrides (Tharinda; raised by
frontend-revieweron forge-et-fe#595, 2026-11-20, F20-i wave 6):pnpm e2e:gate --selectedruns the email smokes, which need the backend started withrefill-per-minute=60000andinteractive-reserve=0(F20-i WI-2.6’s verify bucket, also given to the nightly runner); neitherforge-et-fe’sCLAUDE.mdnore2e/README.mdsays so, so a developer’s selected run can fail on429s. Add one line toe2e/README.md(and theCLAUDE.mdbrowser-gate note) in a smallforge-et-fePR. Done when it merges. -
Decide whether the browser gate checks that a list’s scroll region still shows rows (Tharinda — a harness decision, from F5-g’s findings F-005):
pnpm e2e:gate’s shared page check (e2e/support/pageCheck.ts, shared by both frontends) catches horizontal overflow, clipped controls, target size and axe, but not a scroll region squeezed until no row shows. BUG-145 was that case on the lead list at 1280 × 720, and only F5-g’s design vision pass found it. BUG-145’s fix gave the lead list a floor (layout/list-region-block), so that list is covered; a later two-pane list, or a list that does not use the token, is not. Options: a page-check rule that fails a scroll region holding rows shorter than the token; a check infrontend-reviewerthat every list pane uses the token; or no change. Decide, then make the change in one PR in both frontends, or record “no change” here. -
The Pipedrive questions for Empire: what the exports cannot answer, and a deals export (Tharinda, when the lead chooses to raise them — not in a progress update): OQ-17, OQ-22, OQ-23, OQ-25, OQ-26, OQ-28 and OQ-29 to Empire, and OQ-41 in conversation at a meeting. Empire gives no Pipedrive admin access or walkthrough, only exports (
client/dependencies.md§1.3). At the same time, rewriteclient/open-questions.mdPart 2, its intro and the “Pipedrive review session” row of its timing table, which still ask for access and a session. OQ-25 (templates) and OQ-17 (languages) name Phase 1 features (F20-d, F12); both were waived for those features’ specs (OQ-17 for F12 on 2026-10-07, OQ-25 for F20-d on 2026-10-15), so the answers change data or later work, not those features. Ask with OQ-25 whether templates need text that varies by recipient (raised by the F20-d spec, Q-1, 2026-10-15): a greeting in another language, or text shown only for some customer types. F20-d’s templates fill in only fixed merge fields; a “yes” is a follow-up feature. Ask for a Pipedrive deals export before Phase 2’s deal features: it answers OQ-1, OQ-8, OQ-12, OQ-13 and OQ-30, which block F2-a, F2-c, F7-a, F8-a, F18-a and F24. -
Change requests from the 2026-10-08 manual test pass (Tharinda to decide each; raised by Nanndish, whose bugs from the same pass are BUG-153 to BUG-170 in
bugs.md). None is in a spec, so each is new scope or a spec revision:- More filters, starting with a role filter on Users, and a review of which filters each list screen needs.
- A “Clear all” control on each filter bar.
- A length limit on every text input. BUG-162 covers the lead fields; the rest (accounts, users, standards) needs the same rule field by field.
-
Decide whether
pnpm test:recordmeasures coverage once FS-4 FR-J3’s thresholds bind (Tharinda — at the Phase 1 boundary, with the threshold flip): since 2026-10-12 the frontends’ PR CI no longer runs the unit tests, so the v8 coverage report comes only from the nightly run onmain, and a PR that lowers coverage would not fail before it merges. Choose whether the local record runsvitest --run --coverageand reports the numbers, or the nightly run stays the only coverage gate (.claude/rules/git-conventions.md § The test suites run locally, and the PR records them;vitest.config.tsin each frontend). -
The Phase 1 boundary check (Tharinda, or a developer Tharinda names — before Phase 1 seals, which is at the end of Empire’s Phase 1 UAT round, not before drop 1 (the lead, 2026-11-19;
.claude/rules/tracker.md§ Delivery Phase Rules): build due 2026-11-20, milestone Fri 27 Nov, hand-over and UAT drop 1 about Mon 30 Nov). Drop 2 carries F20-f, F20-j and F44 during Empire’s testing, and the started-early features F23-d, F37-b, F20-k and F22 are decided at the end of that round, each moved to Phase 2 if not done (the lead, 2026-11-19; F20-k and F22 taken in 2026-11-20). The session-start reminder for this check keeps showing until it is recorded, through the UAT round. Run.forge/checklists/phase-boundary-checklist.mdand record it in.forge/engagement-gate-runs.mdasPhase 1 boundary check — <date>; the seal step and/forge-releaserefuse without it, and sessions are reminded from 2026-11-06. The same for Phase 2 before it seals (build due 2027-01-01, milestone Fri 8 Jan 2027). -
Correct the scheduler role’s trust policy in
forge-et-be’s scheduler cloud contract (Tharinda; raised by F38-b WI-1.1’s plan review, 2026-10-17).docs/messaging/scheduler-cloud-contract.md§ The IAM shape limits the role’saws:SourceArnto the schedule ARNs, but AWS setsaws:SourceArnto the schedule group’s ARN and says not to scope it to a schedule (EventBridge Scheduler User Guide, Confused deputy prevention), so a role built from the contract refuses every AssumeRole. F38-b’s Terraform uses the schedule group’s ARN (F38-b spec Rev 2, FR-4). Change the contract by a smalldocs:PR toforge-et-bemain. -
Plan and run Phase 1’s system verification before its UAT round (Tharinda to assign; the plan is
.forge/test-strategy.md § System Verification to V1 Sign-off, 2026-10-17): before Phase 1 is handed over (Fri 27 Nov; UAT 30 Nov – 10 Dec): first everything onmainretested on dev (F38-f, F38-g); then, on the round’s UAT environment, a regression cycle by role with evidence under.forge/release/; every Phase 1 acceptance criterion walked as it will be demonstrated, on desktop and tablet in Chrome, Edge and Safari; an exploratory pass by role; the live Microsoft, SES and Bedrock paths; the migration rehearsed on Empire’s real export; and the non-functional rows due in the round. Assign who runs which part, and when. The mailbox answer, 2026-11-19 (F38-h WI-5.2, record R-10, F38-h decision 11): harmless — each environment keeps its own subscription, delta state and stored mail — soCRM02–CRM04may sign in on UAT while connected on dev. Not exercised: a renewal of two subscriptions on one mailbox, which is due only within 24 hours of expiry (7 days after connecting). -
F38-g’s gap retest: once wave 4’s last walk is recorded, list what merged since 22 Oct and retest it by flows (Tharinda; raised 2026-11-02, the lead’s decisions that day, F38-g spec Rev 3, FR-7, AC-8): F38-g’s retest (wave 4, Biseni Perera and Dinil Gamage, from 27 Oct) walks the features on
mainon 2026-10-22, while development continues — new features, screen reworks and bug fixes have merged since, and more will merge before the walks end, so nothing is evaluated before then. Risk: without this step, behaviour that merged after a feature was walked reaches Phase 1’s UAT round untested on dev. At wave 4’s close: read every PR merged tomainin the three app repos from 2026-10-22 to that day from GitHub (not the tracker, which lags), mark each change covered by a retest result or in the gap, and write about eight to ten end-to-end flows covering the gap — no per-criterion checklists, nothing retested from the start, development not stopped — as wave 5’s run sheets (WI-5.1, WI-5.2, about a day each for two people). Merges after that day go to the next round’s regression pass. The order, what was found on 2026-11-02 and a first outline of the flows:.forge/plans/F38-g/F38-g-decomposition-plan.md§ Notes, “Wave 5”. Meanwhile, for the walkers of groups 4.3–4.5: where a screen changed after 22 Oct, follow the check’s intent on the new screen and note “step adapted:” in the record. Done when wave 5’s PR is open. -
A Content-Security-Policy on both frontends (Tharinda): the other security headers —
Strict-Transport-Security,X-Content-Type-Options,X-Frame-Optionsand a referrer policy — are in F38-b’s forge-et-iaac#3 (e35554d, a CloudFront response-headers policy; F38 Rev 20, F38-b spec Rev 7; the lead, 2026-10-18), applied on dev on 20 Oct when #3 merged (curl -sI https://dev.crm.empire.lk/shows all four, 21 Oct). The CSP was left out on purpose and is not part of F38-d (the lead, 2026-10-18): the CRM shows HTML email bodies insrcdocframes, which inherit the page’s CSP, so it must keepstyle-src 'unsafe-inline',img-src data:and thesrcdocframes working (the register’s §B F20-c CSP row, (e)). Add it to the same response-headers policy once a draft is tested on dev against real HTML mail, and check it in the regression cycle. (Update 2026-11-02: forge-et-iaac#14 merged and was applied on 2026-10-28, andcurl -sI https://dev.crm.empire.lk/returns the CRM policy below and HSTS (read 2 Nov). Left: the signed-in check on dev against real HTML mail, then the check in the regression cycle; the row closes when both are recorded.) (Update 2026-10-27: forge-et-iaac#14 is open, ready for review (Shyamin Ayesh, Senuthmi Thimansa and Dinil asked), not merged. The lead chose to enforce both policies on dev straight away, with no report-only step.base/cloudfronttakes an optionalcontent_security_policy(default none, so theuatoverlay in forge-et-iaac#12 needs no change until F38-h gives UAT its ownapi.uat.*value).overlays/dev/locals.tfbuilds one policy per app fromhttps://api.dev.crm.empire.lk. CRM:default-src 'self'; script-src 'self'; style-src 'self' 'unsafe-inline'; img-src 'self' data:; font-src 'self'; connect-src 'self' https://api.dev.crm.empire.lk; frame-src 'self'; frame-ancestors 'self'; form-action 'self'; base-uri 'none'; object-src 'none'. Capture app: the same, exceptstyle-src 'self',img-src 'self',worker-src 'self'; manifest-src 'self', noframe-src, andframe-ancestors 'none'. What each app loads, read on 27 Oct from both liveindex.htmlfiles, every script and stylesheet the dev sites serve (136 CRM files, and the capture app’s files assw.jslists them), and the source onmainofforge-et-feandforge-et-pwa:- Scripts: one module script per app from
/assets. No inline script,evalornew Functionanywhere. - Styles: stylesheets from
/assets. In the CRM, Radix, react-aria and the scroll lock also add<style>elements at run time, so it needs'unsafe-inline'on top of the email frames. The capture app adds none, and React’sstyle={}props are set through the DOM, which a CSP does not block. (Update 2026-11-22: no longer true. Since its move to shadcn (altrium-io/forge-et-pwa#201, row 4), Radix’s scroll lock in the capture app adds a<style>at run time too, and Sonner (#213, row 11) adds its own;style-src 'self'blocks both. #213 bundles Sonner’s stylesheet from/assets, so the toasts keep their look; the scroll lock is BUG-543, whose fix is'unsafe-inline'in the capture app’sstyle-src, as the CRM has.) - Fonts: Inter, from
/assets/*.woff2. Images: the logo (/assets/*.webp) and favicons, from each app’s own host, plusdata:images inside HTML email. - Network calls: only the API. Uploads (email attachments, price import) go to the API as multipart requests, never straight to S3. The capture app makes no API call yet, and its service worker fetches only from its own host.
- Frames: only the CRM’s sandboxed
srcdocemail frames. They have the page’s origin (allow-same-origin), and each carries its own stricter policy (MESSAGE_FRAME_CSPinforge-et-fesrc/ui/components/SafeMessageBody/safeSrcDoc.ts). - Navigation: Microsoft sign-in and connecting a mailbox are
location.assignredirects, which a CSP does not govern. No form submits natively. No other host is contacted. - Later: when the capture app gains its camera,
img-srcwill needblob:for photo previews (the POC usescanvas.toBlob).
Still to do, after #14 merges: check the headers with
curl -sI; in the CRM, open a real HTML email (the register’s §B F20-c item (e)), send with an attachment, use the editor, a price import, dialogs and pickers, and sign-in, sign-out and mailbox connect, all with no CSP violation in the console; in the capture app, load it, register the service worker and reload it offline. Then record the result inenvironments-and-identity.md§4, close (e), and add the CSP to the regression cycle. UAT’s and production’s values come with F38-h and F38-i.)Update 2026-10-29: #14 is merged — approved by Shyamin Ayesh, merged into
devby the lead on 28 Oct (876d634, 16:18 UTC) and applied bydev.yamlrun 40724734100 (“Plan: 0 to add, 2 to change, 0 to destroy”: the two response-headers policies, updated in place). Checked on 29 Oct:curl -sIonhttps://dev.crm.empire.lk/,/leadsandhttps://capture.dev.crm.empire.lk/shows each app’s policy, as written above, with the four other headers; in a browser, without signing in, the CRM’s sign-in page loads with its font and styles and makes its API call, and the capture app loads and its service worker is active, with no CSP violation in either console (environments-and-identity.md§4, the #14 bullet). What is left is what needs a signed-in person: in the CRM asCRM01, open a real HTML email (register §B F20-c item (e)), send with an attachment, use the editor, a price import, dialogs and pickers, sign out and in, and connect a mailbox, all with no CSP violation in the console; and the capture app reloaded offline. Fold these into the F38-g retest or the next mailbox walk on dev rather than a session of their own. Then close (e), remove this row, and add the CSP to the regression cycle. UAT’s and production’s values still come with F38-h and F38-i. UAT’s is decided before forge-et-iaac#11 merges, which now carries theuatoverlay (the row “Finish forge-et-iaac#11, F38-h’s waves 1 and 3”). Update 2026-10-29, later: UAT’s value is set on #11 (c1ef354): dev’s CRM policy built fromhttps://api.uat.crm.empire.lk. UAT has no capture-app hosting, so it has no capture policy. It is served once F38-h’s rehearsal applies the instance, and the signed-in check above is repeated there as part of UAT’s regression cycle. Production’s value still comes with F38-i. - Scripts: one module script per app from
-
Keep the harness in step with
forge-et-iaac(whoever works on F38, or Tharinda — standing, until F38 is done). The DevOps team pushes todevdirectly, so changes land with no PR to notice. At the start of any F38 work, and whenever Shyamin or Senuthmi report a change, readgit -C ../forge-et-iaac log origin/devsince the last commitenvironments-and-identity.md§4 names, and update §4 (what is applied, open PRs), the register’s “Start here” steps and A-5, and Active Context’s open-risks sentence (.claude/rules/forge-et-iaac.md). Last reconciled atb175dae(2026-11-13: since876d634five PRs merged intodev, each applied by its owndev.yamlrun and none a direct push — #15 and #16 (4 Nov), Dinil’s #18 (9 Nov) and #19 (12 Nov), and #11 (13 Nov, merged by Senuthmi Thimansa after Shyamin Ayesh’s approval); Senuthmi then applied theuatinstance withuat.yaml(runs 43376819951 and 43378306661, F38-h spec Rev 8); the open PR is #17, at814beb1; the app repos’.awsand.github/workflowschanged since 26 Oct only through the delivery team’s PRs, none by the DevOps team;environments-and-identity.md§4, the register andactive-context.mdupdated the same day). Before that at876d634(2026-10-29: sincec728e06bthree delivery-team PRs merged intodev, each applied by its owndev.yamlrun — #10 (5401a88, 26 Oct, documentation only, run 40282639877), #13 (a77370d, 28 Oct, the task role’s Haiku 5.5 grant, run 40635870998) and #14 (876d634, 28 Oct, the frontends’ Content-Security-Policy, run 40724734100); no direct push; open PRs #11 and #12, F38-h’s; re-read later the same day: #12 merged into #11’s branchfeature/F38-h-wave-1(1c09397), so #11 is the one open PR anddevis unchanged). Before that atc728e06b(2026-10-26, F38-b WI-4.1: the merge of the delivery team’s forge-et-iaac#9, applied bydev.yamlrun 40271292253; every commit ondevafter0eee623is a delivery-team PR, noforge-et-iaacPR is open, and the app repos’ deploy files have changed since 13 Oct only through PRs; re-read 26 Oct after that:devunchanged, and the one open PR, forge-et-iaac#10, changes only itsREADME.md). Before that at5e078c0d(2026-10-22: the merge of the delivery team’s forge-et-iaac#5, the capture app’s hosting, applied bydev.yamlrun 39740047356; before it0eee623, Shyamin Ayesh’s direct push of 21 Oct addingbase/ses, applied by run 39553084687, andd230b133, the merge of forge-et-iaac#3, applied by run 39475850266), withforge-et-be’s andforge-et-fe’s deploy files changed since 14 Oct only through PRs, none by a direct DevOps push (re-read 22 Oct, 05:00 UTC; after #5 merged, no open PR anddevthe only branch) (Shyamin Ayesh’s last direct push isforge-et-bedfc9be15, 13 Oct;git -C ../<app-repo> log origin/main -- .aws .github/workflows). -
Review the GitHub Actions workflows the DevOps team wrote (Tharinda; asked by the lead, 2026-10-17; carried by F38-d, whose spec — drafted 2026-10-18, in review — owns the deploy pipelines, parent FR-19 – FR-23, and makes this review its first work (FR-10; the lead, 2026-10-18); UAT’s and production’s pipelines are F38-h’s and F38-i’s, whose specs point here too). Shyamin Ayesh pushed them without a PR or a review:
forge-et-be.github/workflows/build.yaml(with.aws/td.dev.json),forge-et-fe.github/workflows/build.yaml, andforge-et-iaac.github/workflows/dev.yaml(F38-b’s forge-et-iaac#3, merged on 20 Oct, changed its triggers, apply step, concurrency, token permissions, timeout,fmtcheck, lock-file and state-lock options and checkout version). Read each withgit -C ../<repo> log origin/<trunk> -- .github/workflows .aws, and review it in full, not only for the items below: against the team’s standards (.forge/design/coding-standards.md§7,architecture.md § CI/CD Pipelines, AD #22, AD #24, AD #38, the F38 parent spec’s FR-19 – FR-23); for mistakes and things missed; for inaccuracies and inconsistencies between the three repos and with the harness documents that describe them; and for improvements. At least: what triggers each workflow and from which branches; secrets and credentials (static keys or OIDC, what reaches the logs); pinned action versions; a concurrency group andtimeout-minutes; whether a deploy waits for a stable service and runstrivy image; immutable full-SHA image tags; separate build and deploy jobs; anything that applies or deploys from a branch other than the trunk; and the task definition’s values (.aws/td.dev.json: CPU architecture,stopTimeout, health check,JAVA_TOOL_OPTIONS, every register §D value and placeholder). Record the findings inenvironments-and-identity.md§4 (its item 6 lists F38-d items already known) and raise each fix with Shyamin or as a PR of our own. (Update 2026-10-19: the review was done on 2026-10-18 (UTC) by F38-d-WI-1.1 — findings DR-1 – DR-35 inenvironments-and-identity.md§4 “F38-d — the dev deploy pipelines, checked”, each with its home: F38-d WI-2.1, WI-2.2, WI-3.1 and WI-4.1 for those inside FR-1 – FR-9, items 8–10 of “DevOps follow-ups from F38-b’s review of forge-et-iaac” fordev.yaml, and the four rows below for the rest. The row stays until F38-d’s later waves have made the fixes and WI-7.1’s closing sweep (AC-10) checks them; F38-h and F38-i point here for the UAT and production pipelines.) (Update 2026-10-26, F38-d WI-7.1’s closing sweep: every finding is fixed, withdrawn or homed — DR-19, DR-23, DR-27 and DR-28 in their rows below, DR-29 – DR-32 in “DevOps follow-ups from F38-b’s review of forge-et-iaac”;environments-and-identity.md§4 “F38-d — the dev deploy pipelines, checked”, AC-10. What is left here is F38-h’s and F38-i’s: builddeploy-uatanddeploy-prodfrom the reviewed dev workflows, carrying those findings; the row closes when both are built.) -
Delete or replace the commented-out release job in
forge-et-be’sbuild.yaml(Tharinda; raised by F38-d-WI-1.1, 2026-10-19): the end of the file (lines 209–234 onmain, read 2026-10-26) holds a commented-outreleasejob, whichcoding-standards.md§7.5 says to delete; it also names older action majors (checkout@v4,configure-aws-credentials@v4) and the short-SHA tag. The spec leaves the release job to F38-i, so decide whether F38-d WI-2.1 deletes it now (F38-i then writesdeploy-prodfrom the dev job) or F38-i replaces it.environments-and-identity.md§4 “F38-d — the dev deploy pipelines, checked”, DR-19. (Update 2026-10-26, F38-d Reflect: F38-d closed with the job still there, so the choice is now between deleting it in the next PR that touchesbuild.yaml(F38-i then writesdeploy-prodfrom the dev job) and leaving it for F38-i to replace.) -
Bring forge-et-iaac#17 (
overlays/prod) to F38-i WI-4.1’s change list before it merges (Tharinda, as WI-4.1’s owner; Senuthmi Thimansa makes the changes, or agrees to a delivery-team commit; raised by the review of #17, 2026-11-05, and the change list posted on 2026-11-06 and 2026-11-08). WI-4.1’s plan (.forge/plans/F38-i/F38-i-WI-4.1-plan.mdD-1, items (1)–(21)) is the target state. State on 2026-11-14: a full review of #17’s head3fc1add(comment, verdictblock; it replaces the PR’s earlier review comments) finds every D-1 item met (Senuthmi’s814beb1, withdevmerged in at3fc1add), everybase/default leaving dev and theuatinstance unchanged, andfmt,validatein all three overlays and gitleaks clean with Terraform 1.12.1. Update 2026-11-18: Senuthmi pushede590a2b(16 Nov), which adds production’screate_serviceand builds the alerts ARN with the fixed service name, the review’s one open Important; it is not yet reviewed.devhas since gained #20 and #21 (base/only) and #22, whose card-read queueoverlays/prodmust wire (both queue lists ofmodule "ecs"and an output, as inoverlays/dev), and #17 conflicts again inREADME.md. Update 2026-11-19:dev(now8b48548) also gained #23 (the capture app’s hosting inoverlays/uat) and #24 (the lead-triage queue’s timeout,base/only), whichoverlays/proddoes not wire. Open forge-et-iaac#25 (F36-c) adds theplan-researchandcapture-researchqueues and thecapture-research-sweepschedule tobase/; once it is ondev, the nextdevmerge into #17 also addsmodule.sqs.plan_research_queue_arnandmodule.sqs.capture_research_queue_arnto both queue lists ofmodule "ecs"and the outputsplan_research_queue_nameandcapture_research_queue_name(Dinil’s comment on #17, 2026-11-19; the register’s §B F36-c items (a) and (b)). Update 2026-11-19, later: at Tharinda’s request the delivery team pushed4e2f973, which mergesdevat8b48548, andbba8a28, which wires the card-read queue (comment). A delta review ofe590a2b,4e2f973andbba8a28then posted anapproverecord (comment): thecreate_serviceImportant is fixed, and the merge changes nothing for dev oruat. Update 2026-11-19, evening: with the lead’s approval, the plan-onlydev.yamlrun ofbba8a28(run 44569651738, apply skipped) planned one change, the SES configuration set’smax_delivery_seconds0 → null, whichdev’s own applies also make (the row “dev’s apply changes the SES configuration set on every run”), and nothing of #17’s (comment). Left before the merge: (a) if #25 merges first,devmerged in again with its two queues wired, then another delta review and the plan-only run again on the new head; (b) the capture app’simg-srcinoverlays/prod/locals.tf(line 62,'self'only) gainsblob:, as forge-et-iaac#27 does foroverlays/devandoverlays/uat— Nanndish asked Senuthmi for it on #17 (comment, 2026-11-19, 13:28 UTC; the row “Allowblob:images in the capture app’s CSP”), and that push also needs a delta review and the plan-only run again; © Shyamin’s go-ahead, which waits until production’s infrastructure is needed (Shyamin, 2026-11-11). Nine nits on the PR are optional: seven from the3fc1addrecord and two from the delta. Until it merges, everybase/change merged intodevthatoverlays/prodmust wire is added on #17, and each push to #17 is reviewed again. The row closes whendevcarries #17. -
Decide how F38-i waves 5, 8 and 9 go ahead while forge-et-iaac#17 is held (Tharinda, as F38-i’s owner; raised 2026-11-11, when Shyamin Ayesh said #17 merges only once production’s infrastructure is needed). The tracker has wave 5 (production’s alarms, WI-5.1) wait on wave 4, because it edits
overlays/prod; wave 8 (removing dev’s three Entra secrets, WI-8.1) because it also drops their ARNs fromoverlays/prod; and wave 9 (the threedeploy-prodworkflows, no trigger) becausetd.prod.jsonusesoverlays/prod’s names (.forge/plans/F38-i/F38-i-decomposition-plan.md, the seam table and the notes on wave 8). With wave 4 held, all three wait until production is needed. The alternatives:- leave them waiting, so F38-i’s remaining code work and the production apply come together just before go-live;
- or change the Decomposition Plan so that wave 8 leaves
overlays/prodto #17 (one more change-list item), wave 9 builds against the output names WI-4.1’s D-1 (9) fixes, and wave 5 is either committed to #17 (with Senuthmi Thimansa’s agreement) or kept waiting.
Wave 8 also waits on F38-h waves 3 and 4. The row closes when the tracker’s
depends_onfor waves 5, 8 and 9 reflects the choice, or the lead decides to keep them waiting.Decided by the lead, 2026-11-19: re-plan waves 8 and 9 — wave 8 leaves
overlays/prodto #17 (one more item on #17’s change list), wave 9 builds against the output names WI-4.1’s D-1 (9) fixes, and wave 5 keeps waiting on wave 4. Left: the F38-i Decomposition Plan revision and the tracker’sdepends_onfor waves 8 and 9, by whoever next works on F38-i. -
Settle three things before production’s first apply (Tharinda, for F38-i WI-12.1’s plan; raised by the review of forge-et-iaac#17, 2026-11-05, outside WI-4.1’s change list). (1) The first apply’s order:
overlays/prod’saws_ecs_service.backendnames thebackendtask-definition family, which onlydeploy-prodregisters, anddeploy-prodneeds the repository and cluster the same apply creates, so a plain first apply fails on the service. Options:base/ecs’screate_service = falsefor the first apply, thendeploy-prodregisters the revision and a second apply creates the service (forge-et-iaac#11, built for UAT for the same reason — F38-h spec Rev 7, 2026-11-11, which replaced itsbootstrap_task_definition; with it off,base/ecs’sservice_nameoutput is null, andoverlays/prod’smodule "alerts"buildsbackend_service_arnfrom it in a string, which Terraform refuses to plan — so that line takes the name only when it is not null, orbase/alertsaccepts no service, raised by the review of #11 on 2026-11-11), a first apply with-targetthat leaves out the service and its alarms, or a revision registered by hand. 2026-11-13: the review record on #17 asks Senuthmi to take the first option on #17 itself —overlays/prodgains acreate_servicevariable and builds the alerts ARN asservice/${module.ecs.cluster_name}/backend— or to say on the PR that WI-12.1 does it. 2026-11-16: Senuthmi took the first option on #17 ine590a2b(not yet reviewed); once it is reviewed and merged, WI-12.1’s plan only follows the order #17’s description states — apply withcreate_service = false, rundeploy-prod, apply again. (2) The zone’s apex: before the apply, read thecrm.empire.lkzone for an existingA,AAAAorCAArecord atcrm.empire.lkor aTXTat_dmarc.crm.empire.lk; production’s Terraform creates all four, and the production DNS role must allow them (spec FR-3). (3)rds.force_sslfor AC-4:base/rdssets no parameter group, so production usesdefault.postgres18, where RDS turnsrds.force_sslon by default from PostgreSQL 15; confirm it against AWS’s documentation for the engine version, read it back after the apply (aws rds describe-db-parameters --db-parameter-group-name default.postgres18 --query "Parameters[?ParameterName=='rds.force_ssl']"), and if it is off, add abase/rdsparameter-group option with dev’s behaviour as the default. (4) An organization trail (added 2026-11-13 by the review record on #17): before the apply, the DevOps team reads the management account for an organization CloudTrail trail that already covers member accounts (aws cloudtrail describe-trailsthere,IsOrganizationTrail). If one exists,overlays/prod’smodule "audit"trail is a second paid copy of the same events, and WI-12.1’s plan decides whether production keeps it. The row closes when WI-12.1’s plan settles all four. -
Decide whether
dev.yamlvalidatesoverlays/uatandoverlays/prod(Tharinda, the lead — a CI change; raised by the review of forge-et-iaac#17, 2026-11-05):dev.yamlrunsfmt -check -recursiveover the whole repository butinitandvalidateonly inoverlays/dev, so abase/change that breaks theuatorprodoverlay passes CI and is found only when that environment is next applied. Oneinit -backend=falseandvalidatestep per overlay would catch it; neither needs AWS credentials. Decided by the lead, 2026-11-19: yes — aforge-et-iaacPR todevadds the two steps, reviewed with/forge-review-prand approved by Shyamin Ayesh. Done when that PR has merged. Update 2026-11-19: forge-et-iaac#26 (Tharinda) adds a single step todev.yamlthat loops over every overlay exceptoverlays/dev, rather than one step per overlay. It runs after dev’svalidateand beforeplan, so a broken overlay stops the run before dev is applied. Each overlay runsinit -backend=false -lockfile=readonlyandvalidate. The loop reports every overlay that fails, the step has the AWS credentials blanked, andoverlays/prodis picked up when #17 merges. The PR also updates the repository’sREADME.mdandCLAUDE.md. A full review withiaac-reviewer, followed by a delta review of the nit fixes, posted anapproverecord on2d716cd(comment). It waits for Shyamin’s approval. Validation does not catch a queue that is added tobase/sqsbut missing from an overlay’s queue lists, because that is still valid Terraform; that wiring stays a review item. When #26 merges, the same day: drop “(dev.yamlchecks onlyoverlays/dev, …)” from.claude/rules/forge-et-iaac.md’s “Check before review” bullet and say to validate every overlay; add the new step to thedev.yamlsteps in.forge/design/environments-and-identity.md(§4); then close this row. -
Decide whether the Entra tenant and client IDs stay in Secrets Manager (Tharinda; raised by F38-d-WI-1.1, 2026-10-19):
forge-et-be.aws/td.dev.jsonreadsET_SECURITY_ENTRA_TENANT_IDand_CLIENT_IDassecretsreferences (sincedfc9be15), although both are non-secret identifiers (§2.1) and register §D names only the private key as coming from Secrets Manager. Each costs about USD 0.40 a month and cannot be read from the task definition. Keep them (and note it in §D, so UAT and production follow the same pattern) or move them to plain values in F38-d WI-4.1.environments-and-identity.md§4 “F38-d — the dev deploy pipelines, checked”, DR-23. (Update 2026-10-26, F38-d Reflect: WI-4.1 kept them assecretsreferences and F38-d is closed; the decision is still open, and a move to plain values would now be a PR of its own, before F38-h writestd.uat.json.) Decided by the lead, 2026-10-28: the tenant ID, client ID and certificate become plain task-definition values in every environment, and only the private key stays a secret; F38-i builds it (the engagement-log entry of 2026-10-29, “F38-i’s spec revision”, item 12). The row closes when dev’s task definition andbase/secretscarry the change. (Update 2026-11-06, F38-i wave 3: forge-et-be#374 moves dev’s three values to plainenvironmentvalues in.aws/td.dev.json; the row stays open until that PR merges and deploys and F38-i wave 8 removes the three secrets fromforge-et-iaacbase/secrets.) (Update 2026-11-06, later: #374 merged and deployed asbackend:111, and the dev smoke run passed —environments-and-identity.md§4, “Dev’s Entra values as plain values”. What is left is wave 8’s removal of the three secrets.) -
Correct three
forge-et-bedocuments that say production reads the seeded administrator’s email from Secrets Manager (Tharinda; found by the backend review of forge-et-be#374, 2026-11-06, outside that PR’s diff):docs/auth/README.md(the seed-admin placeholder paragraph, lines 604–606 onbb5f873e),docs/migrations.md § Placeholders(lines 100–103) and the repo’sCLAUDE.md(line 256) say production reads the address from Secrets Manager; AD #40 makes it a plain task-definition value in every environment, production included. Fix them in the PR that writestd.prod.json(F38-i WI-9.1), or in any earlierforge-et-bePR that touches those files. The row closes when the three say “a plain task-definition value”. Checked 2026-11-18 onforge-et-bemain:CLAUDE.mdno longer says it;docs/auth/README.md(now line 707) anddocs/migrations.md(now line 108) still do. -
Decide whether the two frontends publish their source maps (Tharinda; raised by F38-d-WI-5.1’s pre-PR review, 2026-10-25): both
forge-et-feandforge-et-pwabuild withsourcemap: true(vite.config.ts), and their deploy workflows’sync::assetsstep uploadsdist/assets/whole, so the.mapfiles are public ondev.crm.empire.lkandcapture.dev.crm.empire.lk. On dev this is harmless and helps debugging. Decide before the first production deploy (F38-i) whether production keeps them public, excludes them from the upload (--exclude "*.map"onsync::assets), or builds without them. The decision applies to both workflows, which stay the same shape (F38-d FR-4). Left unchanged in forge-et-pwa#90 because the file is the plan-approved D-1 text. Decided by the lead, 2026-10-28: production builds them and does not upload them (--exclude "*.map"on each frontend’sdeploy-produpload); dev and UAT keep them public (the engagement-log entry of 2026-10-29, “F38-i’s spec revision”, item 17). The row closes when F38-i’sdeploy-prodworkflows carry it. -
Email the team when a backend deployment fails on dev (Tharinda, with F38-e; raised 2026-10-20 by F38-d spec Rev 3): the dev deploy no longer waits for the rollout, so a deploy that the ECS circuit breaker rolls back leaves the workflow run green and dev on the previous version. An EventBridge rule on
ECS Deployment State Change/SERVICE_DEPLOYMENT_FAILEDfor thebackendservice, sent to the alert SNS topic, tells the team. It is not in forge-et-iaac#3 because the topic and its recipients are F38-e’s (F38 FR-25), which has not started; F38-e’s spec carries it as FR-5 (decided by the lead on 2026-10-23; it goes to the four developers). The row leaves when F38-e’s AC-4 has proved it on dev. Until then, a failed dev deploy shows in the ECS console and as the old version still answering. (Update 2026-10-25: built and applied — forge-et-iaac#7 (F38-e WI-1.1) merged on 23 Oct; the ruleet-dev-backend-deployment-failedisENABLEDand the topic’s four email subscriptions are confirmed, read 25 Oct. No deploy has failed since, so AC-4 is not proved yet: the first of F38-d WI-6.1’s deliberate failed deploys, or the first real failure, proves it.) (Update 2026-10-26, F38-d WI-6.1: two deliberate failed deploys ran,ecs-svc/4268833645685348749(failed 14:05:56 UTC) andecs-svc/1719102153073437308(14:17:24 UTC); one subscriber received the email for the first and showed it in the session. F38-e WI-3.1 reads whether all four developers got both; register §B F38-e.) (Update 2026-11-21, F38-e WI-3.1: a deliberate failed deploy (ecs-svc/7652176713461752993) published once to the topic and SNS delivered it to all four subscriptions, 0 failed — register §B F38-e “AC-4”. The row leaves when the four inbox reports are in the record: the row “Collect the four inbox reports for F38-e WI-3.1”.) -
Look at the four liveness calls that got no answer within 5 s on 26 Oct (Tharinda; raised by F38-d WI-6.1, 2026-10-26): of 1,048 liveness calls from a laptop in Colombo during four rolling deploys, four ended
000(no answer within 5 s) — 13:32:20, 13:36:15, 14:06:51 and 14:12:11 UTC — one of them (13:36:15) while no deployment was running, and the load balancer recorded no 5xx and no target connection error at any of them. Most likely the laptop’s connection, not the drain; check with a second client (CloudShell or a monitor inus-east-1) during F38-f’s or F38-g’s walk before treating it as a defect. Evidence:environments-and-identity.md§4 “F38-d — the dev deploy pipelines, checked”, AC-2, AC-7 and AC-8. -
Write F38-h WI-4.1’s harness records (its plan’s D-11) in wave 4’s harness PR (Tharinda, F38-h’s owner; the main session that dispatches WI-4.2). State on 2026-11-14: forge-et-be#425 (WI-4.1’s
deploy-uatandtd.uat.json) merged as8e744b01, with anapprovereview record on its headb48e7e7(the delivery team pushed the on-mainskip forv0.0.0-rc*tags and four nits with Tharinda’s agreement) and aTest record; WI-4.2’s plan was re-reviewed and approved as its Rev 1 on 2026-11-13, so WI-4.2 can be dispatched now. Left: D-11 (a), (b), (d)–(f) and (i), (l) — thearchitecture.md§ CI/CD Pipelines rows and paragraphs and its revision entry, the deploy-user permissions, and theActuatorSpecversion row — and © from WI-4.2’s diff; (g) and (h) are done (the OpenAPI row narrowed to production, the Exa credit row removed, 2026-11-14). Done when those records are written. -
Answer F38-i’s three questions for the DevOps team before its plan (Shyamin Ayesh with Tharinda; raised at F38-i’s spec review, 2026-10-27): Q-5, where the imported
crm.empire.lkzone lives in Terraform (the root, its state and the management-account identity that runs it); Q-6, how break-glass access to the production database works (the mechanism, the read-only login, the off switch, the approval and time limit, the logging); Q-7, whether they agree to a new pull user in the development account that can only read and pull fromcrm/backend. Each answer is written into F38-i’s spec as a revision before its plan (.forge/specs/F38-i-production-go-live-spec.md§ Open Questions). Updated 2026-10-28 (the lead): Q-6 is dropped (no database tunnel or break-glass access in any environment) and Q-7 is settled (the image copy reads dev’s ECR with dev’ssvc-githubkey; no pull user) — the engagement-log entry of 2026-10-29, “F38-i’s spec revision”, items 4 and 6. Q-5 stays, asked of Shyamin as the lead approved on 2026-10-28; the import itself runs only when production is created (F38-i wave 10), and the root is written now (wave 4) and only validated. Spec Q-5’s “the management account has no state bucket” is wrong:et-ops(management account,us-west-2) holds dev’s state and UAT’s planned state. 6 Nov: still unanswered; asked again in a comment on the Trello production card on 5 Nov. F38-i wave 6 (overlays/dns) is gated on the answer being recorded in the spec. The question as sent:For F38-i (production), the
crm.empire.lkzone needs to be in Terraform before production’s DNS records are written. This is code only for now; nothing is imported or created until Tharinda says production is needed. (1) Root and state. We propose a new long-livedforge-et-iaacroot,overlays/dns, holding only the imported zone (never destroyed), with its state inet-opsatdns/terraform.tfstate, like dev’s and UAT’s. Is that OK, or where would you put it? (2) Who runs it.dns-devallows only dev’s names, so it cannot import the zone. Which management-account identity runs the import and that root’s later plans: a new hand-made role (for exampledns-zone, with Route 53 rights on this zone only, assumable by the CI user), or you yourself? (3) Production’s records. A new roledns-prodin the management account, made by hand likedns-dev, allowing onlycrm.empire.lk,api.crm.empire.lk,capture.crm.empire.lk, their certificate validation records and the SES DKIM, MAIL FROM and DMARC records, assumable by production’s CI user. Is that OK?The row closes when Shyamin’s answer to Q-5 is written into F38-i’s spec as a revision, before wave 4’s plan.
-
Decide whether the CRM’s and the capture app’s buckets expire old file versions (Tharinda; raised by the review of forge-et-iaac#11, 2026-10-29).
base/s3turns versioning on foret-dev-spaandet-dev-capture, and UAT’s and production’s buckets will be the same, with no lifecycle rule. Each deploy uploads new hashed files and its--deletesync removes the old ones, so every replaced or deleted file stays as a non-current version, billed as storage, forever. The amount is small, a few MB per deploy, but it grows with every merge tomain. One option is a lifecycle rule inbase/s3that expires non-current versions after, say, 30 days and removes expired delete markers; that still leaves a month in which a deploy can be rolled back by restoring versions. The attachments bucket is different and is not part of this row: a deleted attachment must stay recoverable, so no rule may expire its versions (the register’s §B row “(a) The attachments bucket on dev, versioned, with no lifecycle rule…”). A Terraform change goes by PR todev, and its plan shows only the two lifecycle resources added. Decided by the lead, 2026-11-19: yes, later, before production — the rule expires non-current versions after 30 days, added with the production work (F38-i). -
Give the attachments card a message for a file scanning cannot open, before scanning is switched back on (whoever switches scanning back on; raised by the F33-a spec, FR-H4 (a), 2026-11-06): while scanning is on, a file GuardDuty cannot open — a password-protected PDF, reported
UNSUPPORTED— staysPENDINGfor good (F38-c spec decision 2). The record’s Attachments card then shows it Being checked indefinitely; it cannot be downloaded, and can only be deleted and uploaded again without its password. Certification PDFs are often password-protected. While scanning is off, as in every environment since 2026-10-09, the case cannot arise. Before scanning is switched on, decide the message the card shows (or a rule at upload), build it inforge-et-fesrc/features/attachments, and decide whether to raise it with Empire in.forge/client/open-questions.md. -
Decide the procedure for erasing a record’s file for good on a data-protection request (Tharinda, the lead; raised by the F33-a spec, FR-H4 (b), 2026-11-06): the proposal keeps attachments for the life of the system, with deletion on request under the data-protection process (§A.6.3), and FS-11 FR-M4 leaves that deletion to the features that own attachments; it is not built. A user’s Delete on a record’s file, as on an email upload, keeps the row and every S3 version behind a delete marker (F33-a decision 3, F20-h decision 3), so an erasure would purge every version (
ObjectStore.deleteAllVersions) by an audited operator step with no screen. Decide the procedure and who may ask for it, for record files and email uploads together; building it is a work item of its own. -
Give the Fast lane’s Trivy cache a bounded key in
forge-et-be(Tharinda; raised by the review of forge-et-be#275, 2026-10-19):pr-validation.yml’strivy-db-${{ runner.os }}-${{ github.run_id }}key saves a new ~95 MB entry on every full Fast-lane run — 39 entries, about 3.6 GB of the 8.94 GB used against the repository’s 10 GB cache limit (gh cache list, 2026-10-19) — so the Gradle caches are evicted sooner. A daily key, asbuild.yaml’s image scan used until forge-et-be#286 removed that scan, keeps one entry a day. A CI change, so the lead approves it; checkforge-et-feandforge-et-pwathe same way. -
The DevOps team’s UAT setup, now that Empire has said yes to a UAT environment (10 Nov, recorded in F38-h spec Rev 8) (Shyamin Ayesh; F38-h spec § Constraints “Waits on the DevOps team”, FR-4, FR-9 step 1; read by F38-h WI-5.1’s readiness check, which stops if one is missing): the DNS role in the row above;
svc-github’s access key held as Environment secrets in theuatGitHub Environment offorge-et-beandforge-et-fe(corrected 2026-11-03:forge-et-iaacneeds none, because itsuat.yamlon forge-et-iaac#11 runs in that repository’sdevEnvironment, and F38-h FR-4 defines theuatEnvironment in the two app repos only); and theEnvironmenttag turned on as a cost-allocation tag, so UAT’s cost shows on its own (billing configuration, which the billing owners hold). The key, read 2026-10-26 (secret names only):AWS_ACCESS_KEY_IDandAWS_SECRET_ACCESS_KEYare held in thedevGitHub Environment offorge-et-be,forge-et-fe,forge-et-pwaandforge-et-iaac(no repository-level secrets;forge-et-iaacalso has an emptyprodEnvironment), and no repo has auatEnvironment yet. A job reads only its own Environment’s secrets, anddeploy-uatand theuatapply/destroy workflow run inuat(which admits onlyv*tags, spec FR-4), so the same key is needed there too, not a new one; GitHub never shows a saved secret’s value, so only the key’s holder can add it.forge-et-pwaneeds none until Phase 2’s round. The delivery team creates eachuatEnvironment, with its tag rule and variables, in F38-h wave 4 with the lead’s approval (Decomposition Plan § Notes); the secrets go in after. On Empire’s yes, in order: (1) the lead records the yes as an F38-h spec revision containing “Empire confirmed the temporary UAT environment on”, which opens wave 5’s gate; (2) the DevOps team’s three items above; (3) F38-h’s waves 1, 3 and 4 merged, and F38-g atdone(wave 5’s other gate); (4) UAT’s alarm set decided, if the round is to carry one (the row “Decide the UAT and production alarm sets”); (5) wave 5, the rehearsal (apply, deploy, destroy), then wave 6, the round’s apply around Fri 20 Nov (2026-11-19, F38-h spec Rev 13: no rehearsal destroy and no re-apply — the running instance is the round’s, and it is destroyed by hand when no longer needed). On a no, the Empire row above says what is revised. Nothing is needed before Empire’s answer: F38-h builds theuatcode first and creates nothing in AWS (the lead, 2026-10-26). On 2026-10-26 the lead messaged Shyamin about forge-et-iaac#11 (F38-h wave 1, under review): that the work is code only until Empire’s yes, that UAT is a separate temporaryoverlays/uatin the dev account reusing dev’s Entra secrets, SES identity,crm/backendrepository, Exa key andsvc-githubkey, and these three asks. The same status is a comment on the Trello card “[ET] UAT environment — temporary, one per acceptance round”. Done when all three are in place, checked by WI-5.1’s readiness read. Update 2026-11-13: (1) the yes is recorded (F38-h spec Rev 8); (2) the DNS role is done (the row above); thesvc-githubkey in theuatEnvironments waits for wave 4 to create them (none exists in any repository, read that day); theEnvironmentcost-allocation tag cannot be read from the dev account (“Linked account doesn’t have access to cost allocation tags”), so the billing owners confirm it; (3) waves 1 and 3 merged on 13 Nov — wave 4 is next, after its two plans are re-reviewed (the row “Re-review F38-h’s WI-4.1 and WI-4.2 plans”), and F38-g is not yetdone; (5) the rehearsal’s first apply has already happened — Senuthmi Thimansa applied the instance on 13 Nov, and the lead keeps it up (spec Rev 8; the row “Theuatinstance runs with no backend until F38-h wave 4”). Update 2026-11-17 (F38-h WI-5.1 R-1, R-2): two of the three items are done and their rows are removed. Thedns-uatrole works: both certificates areISSUED.svc-github’s key is in theuatEnvironment of bothforge-et-beandforge-et-fe(names only, set 16 Nov). Only theEnvironmentcost-allocation tag is left. The development account still cannot read it (AccessDeniedException), so it waits on the billing owners’ answer to R-2’s request. -
The
uatinstance runs with no backend until F38-h wave 4 (Tharinda, F38-h’s owner; Senuthmi Thimansa for theuat.yamlruns; raised 2026-11-13, F38-h spec Rev 8). Senuthmi applied step 1 of the three (withserviceunticked) on 13 Nov, beforedeploy-uatandtd.uat.jsonexist, and the lead keeps the instance up (about USD 1.3 a day idle;environments-and-identity.md§4, “Theuatinstance”). Until wave 4 is built: (a) step 3 (uat.yamlapply withserviceticked) must not run beforedeploy-uathas registered UAT’s task definition —base/ecs/service.tfstarts the service from the newest revision ofet-uat-backend, and none exists yet, so the apply would fail at the service; tell Senuthmi, and say who starts eachuat.yamlrun from now on (the run sheet’s person, F38-h WI-5.1, or the DevOps team on the lead’s word); (b) the eight schedules areENABLEDand their ticks collect inet-uat-scheduler(249 at 10:01 UTC on 13 Nov, four-day retention); when the backend first starts it replays the ones still held, which AD #13 makes safe on an empty database — purging the queue first is an AWS write the lead decides; © the round’s release must hold F20-g wave 1, sincescheduled-email-sendis already scheduled there —forge-et-bemainhas it since #406, so any-rctag cut frommainnow does (the row “Apply the UAT and production overlays with thescheduled-email-sendschedule …”); (d) the rehearsal’s destroy and three-step re-apply still run once before the round starts on Mon 30 Nov (spec Rev 8) (2026-11-19: no longer — spec Rev 13, no rehearsal destroy). Closes when wave 5’s rehearsal is recorded. Update 2026-11-13, later:deploy-uatis forge-et-be#425 (open, review recordblockon one fix); once it merges and theuatEnvironment holds the key (the row “Addsvc-github’s key to theuatGitHub Environments”), the first-rctag registerset-uat-backendand stops, and step 3 can run. Update 2026-11-17 (F38-h WI-5.1 record R-1, R-6): (a) to © are done.v0.1.0-rc1registeredet-uat-backend:1before step 3’s apply created the service on 16 Nov. The backend replayed the held ticks, soet-uat-schedulerheld0messages on 17 Nov and no purge was needed. The release has F20-g wave 1. (d) is left: the destroy and three-step re-apply are WI-5.3’s, and this row closes with it. -
Move build.yaml’s AWS targets into the dev Environment (Tharinda to assign; the DevOps team may take it; F38-h spec Rev 9, the lead’s decision on forge-et-be#425’s review, 2026-11-13;
forge-et-pwa’sbuild.yamlhas the same literals, and the commented-outreleasejob inforge-et-be’s can go in the same PR):forge-et-be’s andforge-et-fe’s.github/workflows/build.yamlstill name their AWS targets as literals (DR-28) — the cluster, service, family and Region inforge-et-be, the bucketet-dev-spa, distributionE2RVC3893O10TJand the workflow-level Region inforge-et-fe. In one laterci:PR per repo, each reads them from itsdevEnvironment’s variables, set first with the lead’s approval, asdeploy-uatreads theuatEnvironment’s (F38-h FR-4, decision 8). F38-h-WI-4.2’s plan D-2 (as approved on 2026-10-27) is a design to start from forforge-et-fe; F38-h-WI-4.1’s plan D-2 forforge-et-be. Each is a CI change the lead approves. Left out of F38-h wave 4 so that wave changes nothing in how dev deploys. Done when both dev workflows read their targets from thedevEnvironment andenvironments-and-identity.md§4 DR-28 readsfixed. -
DevOps follow-ups from F38-b’s review of forge-et-iaac (Shyamin Ayesh, if he chooses; raised by the review of forge-et-iaac#3, 18 Oct; optional, and not on the rewritten “[ET] Misc tasks” card — the lead, 18 Oct; kept for reference and for F38-i to weigh for production): (1) restrict the
forge-et-iaacdevenvironment’s deployment branches todev, becausedev.yaml’s branch rules live in the workflow file, which GitHub runs from the triggering commit, and the environment holdssvc-github’s administrator key (gh api repos/altrium-io/forge-et-iaac/environments/devshowed no protection rules on 18 Oct); (2) pass the subnets’ Availability Zones as a variable instead ofdata.aws_availability_zonesby index, so a change in that list cannot replace the subnets (check the plan shows no change); (3) pin the database instance to the tasks’ zone withavailability_zone; (4) inoverlays/dev/providers.tf, arequired_version, the unusedrandomprovider removed (only after checking the state holds norandomresource), andupper(var.project)without the interpolation; (5) anaws:SourceAccountcondition on theecs-taskstrust policy; (6)drop_invalid_header_fieldson the load balancer and a deny-non-TLS statement on the SPA bucket (the response-headers policy with HSTS is done in forge-et-iaac#3,e35554d, F38-b spec Rev 7); (7) pinning the workflows’ actions to commit SHAs; (8) whetherdev.yamlposts theterraform planon each PR and puts the apply behind a GitHub Environment approval, the two F38 FR-20 rules the delivery team does not build (F38 Rev 20; F38-d recordsdev.yamlagainst them, FR-5) — part of the repository’s conventions, still on hold; (9) raised by F38-d-WI-1.1, 2026-10-19: the other FR-20 rulesdev.yamldoes not meet at #3’s head0ff63c5— one job holds the key forfmt,planandapply(Terraform’s plan needs the key too, so the nearest form is a plan job and an apply job joined by the saved plan file), and nopaths-ignorefor workflow-only changes; (10) raised by F38-d-WI-1.1, 2026-10-19:actions/checkout@v4,configure-aws-credentials@v4andhashicorp/setup-terraform@v3are older majors than the app repositories’ and thanarchitecture.md § CI/CD Pipelines › Cross-cutting’s list, andforge-et-iaachas no.github/dependabot.ymlto keep them current (#3 itself chosecheckout@v4) (update 2026-10-29: since GitHub deprecated Node.js 20, everydev.yamlrun warns that these three are forced onto Node.js 24 (run 40724734100). forge-et-iaac#11 (0742cdd) movesdev.yamlanduat.yamltocheckout@v7,configure-aws-credentials@v6andsetup-terraform@v4, with the lead’s request to fix what the review found; the versions are current once #11 merges. Adependabot.ymlis still not there). Items 1, 8, 9 and 10 areenvironments-and-identity.md§4 “F38-d — the dev deploy pipelines, checked”, DR-32, DR-29, DR-30 and DR-31. None blocks F38-b. -
Decide whether the attachments bucket needs a CORS rule (Tharinda; raised by F38-b’s review of forge-et-iaac#3, 18 Oct): only if the CRM frontend reads a signed download URL with
fetchor XHR rather than navigating to it;forge-et-femainhas no such consumer yet. Decide when the frontend download feature is specified, and add the rule inbase/attachmentsthen. -
The other “[ET] Misc tasks” items (Shyamin Ayesh; for reference, reviewed as the work goes on, mostly after the items above; the card showed as “[ET] Misc tasks part 1” from 18 Oct, and by 21 Oct it was renamed “[ET] Misc tasks”, its description replaced by the list of what forge-et-iaac#3 and #4 added, and moved to Done, so the dev items below are on no card — each was done, left out of dev or made optional, as noted — and its production items are on “[ET] Production environment and go-live”): the account safety basics (CloudTrail, account-wide S3 Block Public Access, a password policy, root access management, alternate contacts — F38 Rev 13 left them out of dev, so they matter for production); Terraform state (dev’s stays in
et-ops, F38 Rev 13; production’s bucket is F38-i’s); the hand-made resources (the delivery team’s list is in §4 “What was made by hand”; confirming thatdns-devchanges onlydev.crm.empire.lknames is F38-i’s);forge-et-iaachousekeeping — nothing left for the DevOps team (the personal owner tag stays, the lead, 18 Oct; apply only fromdev,.gitignore, the lock file,actions/checkout@v4, 14-day dead-letter retention and expiring oldbuild-*images in ECR were done by F38-b’s forge-et-iaac#3, applied on 20 Oct; PRs #1 and #2 were already closed on 17 Oct); sharing keys only through the password vault; and, when production is built (F38-i), its account and who can reach it — including that no production role trustssvc-harnessor any other dev principal — production SES access, Bedrock quotas (or the request template), read-only personal logins with MFA, narrowed deploy and Terraform users with their 90-day key replacement, branch and tag rules on its GitHub Environments, the database fixes optional in dev, and 14-day backups (§4 items 7–14; F38-i). (Dropped for dev by the lead, 18 Oct, F38 Rev 18: replacing thesvc-githubkey every 90 days. Not the delivery team’s: AWS Budgets and the cost allocation tags — Shyamin Ayesh, Kavindu Anthony, Malinda Wellalage and Thilina Abeyrathna handle billing.) -
Sonnet 5.5 and Haiku 5.5 refused for the AWS account — the Bedrock quotas (Shyamin Ayesh raises them — the lead, 14 Oct, on Trello since 14 Oct, the card with the plan below as its comment of 26 Oct, in In Review on the board’s export of 3 Nov and moved to Blocked by the 3 Nov update; Dinil follows up for F12): both models have 0 tokens a minute in dev, so every call is refused, while Haiku 4.5 works at 10 requests a minute (register §C-5). On 15 Oct Shyamin filed six increases at the AWS defaults; AWS closed all six within half an hour without raising any quota (read 16 Oct; still 0, 0, 0, 0 and 10). Its replies never reached him because no email address was set up for them; he has set one up and will resubmit (16 Oct). On 21 Oct AWS refused three more requests, each at the default (Haiku 4.5’s Global requests per minute, then the Global tokens per minute of Haiku 5.5 and of Sonnet 5.5): “We are unable to approve your request at this time.” Sent 22 Oct: after AWS’s emailed refusal of the Sonnet 5.5 request (case 179257474300142, the same wording), Shyamin sent AWS the lead’s reply, posted on the Trello card on 21 Oct. The reply asked for Sonnet 5.5 and Haiku 5.5 at 400,000 tokens a minute on both their quotas, or any non-zero value, and Haiku 4.5 at 100 requests a minute (the row “The Bedrock request quota for Claude Haiku 4.5”). 22 Oct: Shyamin’s request “again on 21 Oct” was the two Service Quotas requests of 09:24 and 09:25 UTC, both refused; at 11:52 UTC every quota was unchanged. The quotas to read when it arrives are Sonnet 5.5’s Global
L-31AB82D0(the backend and WI-2.1) and Haiku 5.5’s cross-RegionL-47520EB7(WI-2.1); the Service Quotas request of 21 Oct for Haiku 5.5 was the Global one,L-F26CA8D1(environments-and-identity.md§4 item 4, the 22 Oct update; every case number in its AC-4 subsection). Until a quota is raised, nothing that uses the 5.5 models can be proved on Bedrock, and F38-a WI-2.1 waits (environments-and-identity.md§4 item 4). Re-read 25 Oct assvc-harness: no change — Haiku 4.5 at 10 requests a minute on both its quotas (L-CCA5DF70,L-E5084BBA), Sonnet 5.5’s GlobalL-31AB82D0, Haiku 5.5’s GlobalL-F26CA8D1and cross-RegionL-47520EB7all at 0; Service Quotas shows no request after 21 Oct, and AWS’s answer to the reply has not been reported. Re-read 26 Oct: the same five values, and still no request after 21 Oct. AWS’s answer, reported 26 Oct (an AWS Support email that resolves the case; the email names no case number): no increase. AWS says the newer models, Sonnet 5.5, Haiku 4.5 and the 5.5 models among them, need “a consistent usage and spend history” on the account over a few billing cycles; until then a new account has lower requests-per-minute quotas than the defaults Service Quotas shows. AWS reviews the account automatically and lifts the limits once usage and spend are enough, after which increases can be requested again. It suggests building that history by using the Bedrock models already available. Also 26 Oct: Shyamin filed two more Service Quotas requests at 09:52 and 09:53 UTC, for Opus 5.5’s cross-Region tokens a minute (L-A4430697, 30,000,000; Opus 5.5 is not in the design) and Sonnet 5.5’s cross-Region tokens a minute (L-94A31E46, 6,000,000), cases 179300836700331 and 179300838800048; AWS closed both at 10:08 UTC. Read afterwards assvc-harness: the five quotas above unchanged, andL-A4430697andL-94A31E46at 0. Next step, asked of Shyamin Ayesh in the lead’s message of 26 Oct: file no further quota requests until AWS lifts the limits, because each one is closed unchanged; the team re-reads the quotas weekly instead (the next row) and tells him when to request. So no 5.5 model can be called in this account for an unknown period, probably past Phase 1 UAT; every register item that waits for “once its quota is raised” has no date, F38-a’s two model calls included (now the register’s §C-5 row, F38-a Rev 4). What the team does about it is the next row. 11 Nov (the lead): a new Trello card, “[ET] Bedrock: set up a quota request template in the AWS management account for prod, and request a quota increase for dev …”, for Shyamin Ayesh, asks for two things. (1) The Service Quotas request template in the management account that F38-i decision 28 asks for (no row here tracked it before), holding dev’s six requests of 15 Oct, at the AWS defaults, because Service Quotas refused any lower value. The template files these requests automatically for each account created after it, production included. It changes no existing account, and AWS still reviews each request, so production should still expect new-account limits, probably Haiku 4.5 at 10 requests a minute on each route and the 5.5 models at 0 (the scan row “Scan production’s Bedrock setup” below). (2) One more dev request, an exception to the 26 Oct “no further requests”: a Support “Service limit increase” case opened from the management account and naming dev, at the 22 Oct reply’s figures. A quota request can only raise a quota, so neither step can lower dev’s quotas or remove its model access. Done when the template exists and the dev case’s number and AWS’s answer are recorded inenvironments-and-identity.md§4 item 4. Read 19 Nov assvc-harness: on 17 Nov two Service Quotas requests for Haiku 4.5 at 10,000 requests a minute (cases 179490479200086 and 179490496300492, both routes) were closed within about 20 minutes with no change; every quota is as on 26 Oct, and dev’s history shows no Support case at the 22 Oct figures, so step (2) is still open (§4 item 4, the 19 Nov update). -
The AI layer under Bedrock’s new-account limits (raised by AWS’s answer of 26 Oct, the row above; owners per item below). Sonnet 5.5 and Haiku 5.5 stay at 0 and Haiku 4.5 at 10 requests a minute until AWS judges the account’s usage and spend enough. Decided by the lead, 26 Oct:
- Haiku 4.5 is V1’s model wherever Sonnet 5.5 or Haiku 5.5 is planned, in every environment, until AWS lifts the limits (
.forge/design/ai/ai-provider-strategy.md§5 row 10). It is done by each feature’s model key in the environment’s task definition, so the code’s defaults stay and moving back is configuration plus the feature’s evaluation run (§4c). Dev already sets lead research, email drafts and the account briefing to Haiku 4.5 (the register’s §D model rows), and F38-h’s spec gives UAT the same overrides. - The backend paces its Bedrock calls to stay under the account’s requests-per-minute quota, so a burst waits instead of being throttled and opening the shared circuit breaker (the row “The Bedrock request quota for Claude Haiku 4.5”). Built by F42 (
.forge/specs/F42-bedrock-call-pacing-spec.md). - No escalation through Empire Teas: the lead does not expect Empire Teas to have an AWS account manager or support connection, and will tell them of AWS’s answer instead.
- When the production account is created is not the delivery team’s decision. The lead expects the end of November or December; Shyamin Ayesh, Malinda Wellalage, Thilina Abeyrathna, Rukshan Salgado and Empire Teas decide it. The limits are per account, so production starts with no Bedrock history and at the same or lower limits at go-live; items 1 and 2 are what keep the AI features working there.
Left to do:
- Prove the pacing on dev (Tharinda; F42 built it —
.forge/specs/F42-bedrock-call-pacing-spec.md, a follow-up of F12): each backend task paces every Bedrock attempt in memory to its share of the quota per inference profile (et.ai.pacing.limits, register §D), keeps two of each minute’s slots for calls a person waits on, hands background work that gets no slot back to the hourly sweep asRATE_LIMITEDwithout counting a lead attempt, and no longer lets a throttle open the circuit breaker (ai-provider-strategy.md§4). Items 1–3 of the register’s §B F42 block were proved on dev on 2026-10-29, item 5 (a burst during a deploy’s two-task overlap) on 2026-11-02, and item 4’s dev values were read. What is left is item 4’s UAT and production reads; production’s per-task rows go with its overrides (the line “Give production the Haiku 4.5 overrides”). - Tell Empire Teas of AWS’s answer (Tharinda): the 5.5 models are not available on their account for now, V1 runs on Haiku 4.5 until AWS lifts the limits, and the limits lift as the account is used.
client/dependencies.md§1.1’s Bedrock row says the quotas are “raised by Altrium during environment setup”, which AWS has not allowed. Done in the harness on 26 Oct: that row now carries a dated note, and the proposal header has a post-issue note of 26 Oct stating the interim model; telling Empire Teas directly is still the lead’s. - Give production the Haiku 4.5 overrides (whoever takes F38-i): F38-i’s spec, approved 2026-10-27, lists them (FR-9, AC-9), and F38-i builds them (F38-i-WI-9.1); its production task definition carries the same model keys as dev’s for as long as the limits apply. With them go the pacing rows (the F38-i spec’s input “Production’s pacing rows”; FR-9 re-derives
td.prod.jsonagainst register §D, which lists them): eachet.ai.pacing.limitsrow set to production’s quota for that model divided by its backend tasks, rounded down, restating every row (F42 FR-H4 item 4; register §B F42 item 4). The Sonnet 5.5 row stays a placeholder of at least 1 while no feature uses that model, whose quota is expected to be 0 in production (F38-i W-6), and no feature is moved to a model whose quota is 0; where a model’s share is at or below the background reserve (2 by default), the task definition also setsET_AI_PACING_BACKGROUNDRESERVEbelow every row, or startup refuses. - Scan production’s Bedrock setup before its task definition is written, and match the configuration to it (whoever takes F38-i, with Tharinda; asked by the lead on 2 Nov). Production is a separate account with its own new-account limits, so it may differ from dev (UAT does not: it runs in the development account and shares dev’s setup, AD #4). The code supports any of the cases below by configuration alone, once F43 has made every route ready (
.forge/specs/F43-ai-quota-backlog-and-busy-answer-spec.md§ In Scope); the scan confirms which case production is in and that nothing needs code.- Access for the scan.
svc-harnessand every other dev identity have no access to production (F38-i FR-1: people use IAM Identity Center with MFA and time-limited roles, no IAM user and no long-lived key). The DevOps team adds a read-only permission set for the delivery team in Identity Center (Bedrock, Service Quotas, IAM, ECS and Organizations reads only; neversecretsmanager:GetSecretValue), and the scan runs from a person’saws sso loginsession, with Claude reading through it only when that person is present. Ask for it when F38-i’s account work starts (Shyamin Ayesh). - What the scan reads, for Claude Haiku 4.5, Sonnet 5.5 and Haiku 5.5: each model’s access (
aws bedrock get-foundation-model-availability: agreement and authorization); which of its two routes exist and areACTIVE(aws bedrock list-inference-profiles:us.andglobal.); each route’s quotas (aws service-quotas list-service-quotas --service-code bedrock: requests and tokens a minute — the dev codes are inenvironments-and-identity.md§4 item 4, and production’s may differ); the backend task role’s Bedrock grant (all six routes; production’s is written by F38-i WI-4.1’soverlays/prod/locals.tf, and its read is register §B F43 item 4’s production part); any Organizations policy refusing Regions (F38-a Q-2), which must allow every Region of a route in use; and the deployed task definition’s AI model keys and pacing rows against those reads (register §D). - How the reads set the configuration, all in
td.prod.json(F38-i FR-9), no code:- Haiku 4.5 has a quota above 0 on both routes: the calls a person waits for on
us., the background features onglobal.(F43 point 4), each pacing row at that route’s quota divided by the backend tasks. - Only one Haiku 4.5 route has a quota: every feature on that route, one pacing row, and the reserve keeps 2 slots for people (F42).
- Neither Haiku 4.5 route has a quota: no AI feature can run, so each is switched off (
ET_AI_FEATURES_<FEATURE>_ENABLED=false) and people use the manual path; this is a go-live risk to raise with the lead and Empire Teas before launch, and the Haiku 4.5 request (F38-i FR-2) is chased. - A 5.5 model gains a quota: a feature may move to it only after its readiness work and evaluation run (
ai-provider-strategy.md§4c, §4d), on whichever of its routes has the quota. - A route’s tokens-a-minute quota is lower than its pacing row can use: the backend paces requests only, never tokens, and no token pacer is built unless production’s quotas need one (decided by the lead, 2026-11-05). That route’s pacing row is set instead to its tokens-a-minute quota divided by the tokens of the largest call it carries, as AWS counts them against that quota for that model (input and output; read AWS’s counting rule for the model when the scan runs), taking each feature’s largest call from the
AI Call Completedlines on dev, and divided again by the backend tasks. Haiku 4.5’s dev quota of 5,000,000 tokens a minute per route never binds at 10 requests a minute (environments-and-identity.md§4 item 4); a call that does go over is refused by Bedrock as a throttle, which a person sees as the busy answer and background work leaves to the hourly sweep (F42, F43), so nothing breaks while the rows are corrected. A model outside the three, or a route with no row yet, is configuration too: its price row and pacing row go in the task definition (ET_AI_PRICES_<n>_*andET_AI_PACING_LIMITS_<n>_*, each list restated in full —forge-et-bedocs/ai/README.md), and its route in the task role’s Bedrock grant (forge-et-iaac). If the scan finds anything these cases do not cover, that is a code change, raised as a bug before go-live. Each pacing row also assumes a fixed number of backend tasks: if production runs more, or scales them automatically, the rows are divided again by hand. The backend refuses to start on a mismatched setting (a feature model with no price or pacing row, a reserve not below every row), so a wrong task definition fails at deploy, not in use.
- Haiku 4.5 has a quota above 0 on both routes: the calls a person waits for on
- Done when the reads, the case chosen and the task definition’s values are recorded in
environments-and-identity.md§4 and register §B F42 item 4 (production’s values).
- Access for the scan.
- Re-read the quotas every week (Tharinda, or Claude in a session for the lead; AWS does not promise to say when it lifts the limits —
ai-provider-strategy.md§4d):aws service-quotas list-service-quotas --service-code bedrockassvc-harness, a read that needs no approval; when a value rises, record it inenvironments-and-identity.md§4 item 4 and AC-4, and the DevOps team then requests the quotas the backend uses (L-CCA5DF70;L-E5084BBA, Haiku 4.5’s Global quota, which F43 moves the background features onto — added 2026-11-03, as the 22 Oct reply asked for both Haiku 4.5 quotas;L-31AB82D0; andL-F26CA8D1, the Global quota of Haiku 5.5, whose profile was chosen on 27 Oct, for its call in the register’s §C-5 row “Sonnet 5.5 and Haiku 5.5 answer one call each in dev” and for the backend once a feature moves to it) at the 22 Oct reply’s values. Last read: 3 Nov, no change — Haiku 4.5 at 10 requests a minute onL-CCA5DF70andL-E5084BBA, Sonnet 5.5’sL-31AB82D0and Haiku 5.5’sL-F26CA8D1andL-47520EB7at 0, and no Bedrock quota request after the two of 26 Oct (bothCASE_CLOSED). Read again 6 Nov: no change (the same six quotas, Sonnet 5.5’sL-94A31E46also at 0), and no request since 26 Oct; on the Trello Bedrock card the same day. - Make a switch a configuration change only (
ai-provider-strategy.md§4d). Three changes, each with its full detail in its own record; the lead, 26 Oct: tracked, and fixed when the team takes them up — none blocks Phase 1, all three are needed before a feature moves to a 5.5 model: (a) BUG-248 (bugs.md), Sonnet 5.5 rejects the forced tool — fixed in forge-et-be#316 on 26 Oct; its live proof is register §B F12 row 10; (b) the row “Haiku 5.5 readiness in the backend” below; © the task role’s grant on all six routes — dev’s applied (forge-et-iaac#13, #16), UAT’s on forge-et-iaac#11, production’s on forge-et-iaac#17 as F38-i WI-4.1’s plan names it; its row closed on 2026-11-06 (engagement-log.md, that date). - Dinil’s Haiku-only evaluation (Dinil;
.forge/design/ai/model-choice-evaluation-plan.md, a proposal of 22 Oct, agreed as the approach by the lead on 4 Nov: Haiku 5.5 is evaluated first for every feature, and Sonnet 5.5 stays only where Haiku 5.5 fails): whether every AI feature can run on Haiku 5.5, with no Sonnet. It runs on Bedrock once the limits lift (its P-2), not before: the lead declined its D-2 on 26 Oct — no Anthropic API account billed per token, for cost; the team’s Claude plan seats are for people using Claude and do not give the backend an API key, so they cannot run the backend’s evaluation runner. What can be done before the limits lift: P-1 and P-4 (the two backend changes above), P-5 (the runner’s repeated and blind-paired runs) and the larger evaluation sets (§3), all without AWS. Still open for the lead: D-1 and D-3 (D-4 decided 4 Nov: the lead signs off each result). Its result sets each feature’s model when the switch happens. - Confirm the P1–P3 rubric with Empire (Tharinda; F11-g spec Q-1, added 2026-11-05). The PRD names three priority levels but no criteria; F11-g’s prompt ranks P1 for a tea buyer with a concrete need or clear signs of importing at scale, P2 for a plausible buyer whose need or scale is unclear, P3 for little sign of a fit (spec FR-A2). F11-g builds with it; a change is a new prompt version. Done when Empire’s answer is recorded in the spec.
AWS’s suggestion to use other Bedrock models to build spend is not taken without the lead: AD #19 keeps the AI layer on Claude, the spend is Empire Teas’, and any Bedrock call needs the lead’s approval (
.claude/CLAUDE.md § Boundaries › ASK FIRST). - Haiku 4.5 is V1’s model wherever Sonnet 5.5 or Haiku 5.5 is planned, in every environment, until AWS lifts the limits (
-
Haiku 5.5 readiness in the backend, so a feature can be switched to it by configuration (no owner yet; the lead, 26 Oct: tracked here and done when the team takes it up; not needed for Phase 1, needed before any feature moves to Haiku 5.5; raised by
model-choice-evaluation-plan.mdP-3 and P-4; the third switch change, BUG-248, was fixed in forge-et-be#316). Code as read onforge-et-beorigin/mainat25e61c91, 26 Oct:- Done (F43, 2026-11-04):
application.ymlcarries a price row and a pacing row for both Haiku 5.5 profiles, at AWS’s Bedrock price read on the day the rows were written (F43-WI-4.1), so setting a feature’s model to Haiku 5.5 no longer stops the backend; startup still refuses a model with no row. - Done (BUG-342, forge-et-be#361, 2026-11-04): a reply with no tool call and a stop reason of
refusal,content_filteredorguardrail_intervenedends the callINVALID_OUTPUTat once, unretried, and its WARN line nameskind=REFUSED. The lead chose this over a new outcome value, which would have needed a migration and changes in every feature. Which stop reason Haiku 5.5 actually sends is register §B F12 item 11. - Thinking blocks are already handled.
toolAnswerscans every content block for the named tool, so a response that begins with thinking blocks is read correctly. Done: forge-et-be#316 adds theBedrockAiProviderSpeccase that keeps it so (reasoning blocks before the tool call, read in one attempt).
- Tests: a
BedrockAiProviderspec for a refusal response, on the stub, with no AWS (the price row’s binding test came with item 1). - Live proof: one call on dev once Haiku 5.5’s quota is above 0 (register §C-5), with the lead’s approval of the call.
- Ticket: item 1 was F43 (forge-et-be#360) and item 2 BUG-342 (forge-et-be#361); both are done.
- Done (F43, 2026-11-04):
-
Correct the F38-f and F38-h plans that expect lead triage and lead research on the
us.profile (Tharinda, as F38-f’s and F38-h’s owner; raised by F43-WI-4.2’s grep, 2026-11-04): since F43’s wave 4, dev’s task definition setsET_AI_FEATURES_LEADTRIAGE_MODELandET_AI_FEATURES_LEADRESEARCH_MODELtoglobal.anthropic.claude-haiku-4-5-20251001-v1:0, and UAT takes the same from register §D..forge/plans/F38-f/F38-f-WI-2.1-plan.md(R-1 and the step that expects “the three_MODELlines set tous.anthropic.claude-haiku-4-5-20251001-v1:0”) and.forge/plans/F38-f/F38-f-WI-2.2-plan.md(the step that expects every feature’sAI Call Completedline on theus.id) are run sheets with the person, and.forge/plans/F38-h/F38-h-WI-4.1-plan.mdlists UAT’s “three AI overrides” on theus.id. Done when each plan states the four model keys as register §D gives them, or its record shows it was run before F43’s wave 4 deployed. -
Decide whether
lead-ranking’sINVALID_OUTPUTon a lead with no inquiry is a defect (Tharinda, as the lead, to decidebugs.mdor not; F11-g’s owner Dinil to fix; raised by F38-f-WI-2.1 on dev, 2026-11-21, record.forge/plans/F38-f/F38-f-WI-2.1-record.mdR-6b): both leads the run created with no inquiry text (140 and 141, the research evaluation cases) endedAI Call Completed … feature=lead-ranking, outcome=INVALID_OUTPUT … attempts=2(invocations 569 and 572), each attempt loggingBedrock Attempt Failed … kind=INVALID_OUTPUT NO_TOOL_USE stopReason=end_turn, so the lead’s Priority panel reads “The AI ranking isn’t available for this lead right now.” on Haiku 4.5’s Global profile. The other nine rankings of the run succeeded. Done when the lead decides, and on a defect, when itsBUG-NNNis filed. -
Exercise the capture app’s five AI features on dev (Tharinda, as F38-f’s owner, to decide where the run goes; raised by F38-f-WI-2.1, 2026-11-21, record R-6c):
forge-et-bemainhascard-read,plan-research,plan-ranking,capture-researchandcapture-rankingbeside the five features of F38-f’s spec. WI-2.1 exercisedlead-ranking(it runs on every CRM lead) but not these: they run only through the capture app,card-readneeds a card photo the browser session could not upload, and the plan features need a show campaign with exhibitors. Either a run sheet (for instance with F38-f-WI-2.5, the capture-app check) exercises them with three inputs each fromcard-read-v1.json,capture-ranking-v2.jsonand the lead-research set, recording model, latency, cost and a quality note per feature, or the lead records that F36’s own dev runs (register §B F36-b, F36-c) stand for them. Done when the lead decides and that run, or the decision, is recorded. -
Judge four AI quality findings from F38-f-WI-2.1’s dev run against the features’ evaluation bars (Dinil, owner of F11-c, F11-d, F11-g and F11-h; raised 2026-11-21, record R-3 to R-6b): (1)
lead-rankinggave P1 to a lead whose inquiry tells it to quote USD 3 per kg and a 20% discount (lead 145, the email-draftlead-injectioncase), where the ranking set’sprompt-injectioncase expects P3; (2)lead-researchfor the invented “Tierra Foods” matched a real UK firm and citeduk.linkedin.compages titled “TierraSphere”, a differently named firm, for two facts (lead 144), the kind of other-firm finding F11-c’s evaluation forbids; (3)lead-triageflagged two leads as spam because their Lead Source Detail held a test label (leads 141 and 142), so a source note a person types can decide the flag; (4) everyemail-draftgreets the contact by first name only (“Dear Anna”), where each case’schecks.contactNamegives the full name (whetheremail-draft’s scorer accepts a first name was not checked). All ran on Haiku 4.5 (dev’s overrides). Done when each is accepted as is, or filed as a defect or an evaluation-set change. -
Production’s account and sign-in (Shyamin Ayesh, answered from his work; on the “[ET] Production environment and go-live” card since 18 Oct, first the “[ET] Misc tasks” card’s production item): whether a production account exists, who can reach it, and production’s sign-in model (
environments-and-identity.md§3–§5; F38-i). Dev’s sign-in is settled: the team usessvc-harness’s access key (14 Oct). -
Decide whether an account’s owner may show as “—” when the owner’s role has no account permission (Tharinda to judge; seen in F38-g’s retest on dev, 2026-10-28, WI-4.1 record R-6, F5-c.10):
CRM02(Admin 1) createdRETEST 2026-10-28 F5-c ownedwhile holdingaccount.manage; after its role went back to one with onlylead.view, the Accounts list read asCRM01shows that account’s owner as “—”, not “Admin 1”. The owner’s name appears to come from the user directory filtered byaccount.view. Either it is intended (the directory lists only people who can be offered as owners) and the list should still name an existing owner, or it is a defect to file inbugs.md. Not recorded as a failed retest check. -
Folding in the 2026-10-01 meeting with Shyamin Ayesh (Tharinda): the record is
.forge/discovery/meeting-notes/2026-10-01-aws-setup-shyamin.md. The region is decided —us-east-1, as AD-O-9 already says (AgentCore Web Search is not offered inus-west-2); the move was applied toforge-et-iaac’sdevon 2026-10-05, and PR #1 and PR #2 are overtaken by what was pushed todev(both closed by 17 Oct, their changes ondev). Still to fold into the F38 specs and the register: the account structure and the meeting’s D-1 to D-3 decisions. (The developers’ and Claude’s access in dev and the pipelines were folded in on 14 Oct — F38 Rev 9–11,environments-and-identity.md§4.) -
Automated deployment (Tharinda — FS-1 cloud half): the backend ECS rolling-update pipeline and the two frontend S3 + CloudFront pipelines are designed in
.forge/design/architecture.md § CI/CD Pipelinesand are built by F38-d; no manual ECS or task-definition console step is planned. The backend’s first version exists since 2026-10-09 and the CRM frontend’s since 2026-10-13 (each abuildworkflow written by Shyamin Ayesh ahead of the F38-d spec; how they differ from F38 FR-20 isenvironments-and-identity.md§4 rows 9 and 15 — the frontend’s staged S3 sync with cache headers and its one-deploy-at-a-time rule were added by forge-et-fe#259, BUG-136); the capture app’s has deployed every merge toforge-et-pwamain(except a workflow-only push) since 2026-10-25 (build.yaml, F38-d wave 5, forge-et-pwa#91). -
The cloud bring-up, F38-a to F38-j (F38-i is tracked in Phase 1 while its code waves are worked, and moves back to Phase 2 before Phase 1 seals unless it is done — F38-i spec Rev 3, 2026-11-02) (Tharinda owns F38-a to F38-j, F38-i included —
tracker.yaml; the requirements are the F38 spec; F38 has no tracker row of its own since 2026-10-11): the AWS account has existed since 2026-09-29 — Empire Teas’ own, with administrator IAM credentials given to Shyamin Ayesh, who is writing the Terraform and the environments with Senuthmi Thimansa inforge-et-iaac(in dev the team usessvc-harness’s access key rather than a login per developer — 14 Oct). Since 2026-10-09 the backend runs on dev (https://api.dev.crm.empire.lk, inus-east-1), deployed on every merge toforge-et-bemain, and since 2026-10-13 the CRM frontend runs athttps://dev.crm.empire.lk, deployed on every merge toforge-et-femain, andCRM01can sign in (register §A-5 closed); the capture app’s shell has run athttps://capture.dev.crm.empire.lksince 2026-10-25, and uploads and the mailbox have their AWS resources (forge-et-iaac#3, applied 2026-10-20) and their task-definition values (forge-et-be#300, 2026-10-22) but are not yet proved by use (environments-and-identity.md§4). The Statement of Work’s Foundation date, Fri 16 Oct 2026 (environments and deploy pipeline), is not fully met; the lead decided on 15 Oct not to raise it with Empire Teas separately. F38 retests everything with code onmainend to end on dev (F38-f, F38-g). Until AWS lifts the account’s new-account limits, every AI feature runs on Haiku 4.5 in every environment (the lead, 2026-10-26,ai-provider-strategy.md§5 row 10), and there is no direct Anthropic API for local development (§5 row 7). The order of work is the “Start here” checklist at the top of.forge/cloud-gated-validation.md. -
The vulnerability findings under AD #38 (Tharinda): scans report on every PR and never block it, the dev deploy runs none (ECR scans each image on push — AD #38 as restated 2026-10-20), and
TRIVY_GATE_EXIT_CODEstays0. The backend’s Tomcat and Jackson findings were fixed by version overrides inforge-et-bePR #145, and the overrides were removed byforge-et-bePR #311 (Spring Boot 4.1.2, whose BOM ships the same or newer versions); a later finding the BOM lags is fixed the same way, and the override removed on the Spring Boot upgrade that catches up. The check that stops a UAT or production release is built withdeploy-uat(F38-h, F38 FR-35) anddeploy-prod(F38-i). The RC-1 clean-baseline attestation at Phase 1 handover (FS-4 spec Rev 3) stays. -
Empire IT asks (Tharinda — agenda in
.forge/design/environments-and-identity.md§5; only Kavindu Anthony has Azure access; the lead, 2026-11-19: asked when production is needed, by hand — an email, a meeting or a WhatsApp message — never sent automatically; the drafted message below stays ready). Done 2026-09-24: both certificates uploaded and the three mailbox and two laptop reply URLs registered. Still to ask: the production administrator’s email for the first-admin migration (AD #34 — production cannot start until Kavindu confirms it and it is stored in Secrets Manager, F38-i; never in git), a re-read of the granted Graph permissions (User.ReadBasic.All), and the certificate renewal routine (§5 agenda items 4, 5 and 7). The rest of the certificate plan (where UAT reads the non-prod key from) is to settle with Shyamin Ayesh (§2.3). 2026-10-28: production’s administrator email is now a plain value in production’s task definition (forge-et-be.aws/td.prod.json), not a secret (the lead; the AD row refining AD #34 in F38-i’s spec revision), so it is in git once confirmed. Drafted 2026-10-28 for Tharinda to send to Kavindu Anthony (the lead asked for it; nothing is sent by the delivery team; nothing is needed before production is created, so the dates are “before production is set up”):Hi Kavindu — a few things we will need from Empire before the production CRM is set up (we expect that around the end of November or in December; we will confirm the date). Nothing is needed today, but early answers help.
- The first administrator’s email. Production starts with one administrator account, which then creates everyone else’s. Which Empire Microsoft 365 address should it be? It is stored in the system’s configuration, so it should be a person or a shared admin mailbox that will stay in use.
- A mailbox for email-authentication reports. We publish a DMARC record for
crm.empire.lk, which tells other mail servers how to treat mail that claims to come from the CRM. Those servers send short daily reports to a mailbox you choose; which address should receive them? - Who receives production alerts. If the system has a problem (it stops answering, errors appear, a background job stops), an email goes to our team, Shyamin, Thilina, Malinda and you. Should anyone else at Empire receive these?
- Your own account for web search. Lead research uses a web-search service called Exa. During development it runs on Altrium’s free account, which cannot be transferred, so Empire needs its own (sign-up at exa.ai, free to start) and to share the key with Shyamin through the password vault.
- At go-live: we will ask you to remove the two “laptop” sign-in addresses from the CRM’s app registration in Azure, which were only for development. Two things for your information: production’s database is encrypted, and it runs in a single availability zone, as the proposal described. If you later want automatic failover to a second zone (about USD 30 a month), a web application firewall (about USD 15) or a backup copy in another AWS region (about USD 8), these can be added; just let us know.
-
Record the answers from the 6 Oct 2026 call with Shyamin Ayesh (Tharinda — needs the lead’s notes of the call): §4’s “Open with the DevOps team” list was drafted from the call’s agenda and Shyamin’s later work, but what he answered on the call itself was never written into the harness.
-
Decide which model F11-h’s drafts use (Dinil, with Tharinda; F11-h spec decision 2; once Sonnet 5.5 and Haiku 5.5 can be called on dev — it was due before Phase 1 UAT, which the 26 Oct decision below replaces):
email-draftdefaults to Sonnet 5.5 (AD #19,ai-provider-strategy.md§2), and dev runs it on Haiku 4.5 through an override while Sonnet 5.5 is refused. After the evaluation run on dev (F11-h spec FR-H3 item 2) — on Haiku 4.5 now, and on Haiku 5.5 and Sonnet 5.5 once their quotas are raised — decide whether drafts stay on Sonnet 5.5 or move to a Haiku model. A move needs the lead’s sign-off on AD #19 and a post-issue note on the proposal header (L-001), because the proposal’s §A.3.2 puts drafts on Sonnet. Meanwhile, decided by the lead on 26 Oct: drafts run on Haiku 4.5 in every environment, production included, until AWS lifts the account’s new-account limits (ai-provider-strategy.md§5 row 10; the proposal’s post-issue note of 26 Oct states it). This row’s lasting choice waits for Sonnet 5.5 to be callable, which AWS gives no date for. -
Decide whether AI source dates show the year outside the current year (Tharinda, with Dinil; raised by the review of
forge-et-fe#327, F13, 2026-10-22): design-system §9 says a date not in the current year carries its year, but F13 spec FR-V3 and the frozen Wave 1 contract pind MMMwith no year for the briefing’s open-item sources and itssource.label(as F11-h’s and F12’s labels do), so a dormant account’s sources from earlier years readfrom: lead, 24 Junwith no year. F13 shipped the spec’s form (overnight standing answer 4: spec over design system). Either record it as an accepted §9 deviation in a design-system revision, or show the year on both sides — a contract revision, the backend’sBriefingSourceLabeland the frontend’sformatInstantDayMonth, and the same for F11-h’s and F12’s labels. -
Stop held-request browser checks racing the 12-second client timeout (whoever next touches the AI panels’ browser support; raised by the review of
forge-et-fe#327, F13, 2026-10-22): F13’sbriefingAxeStatesand the size matrix’saccount-detail-briefing-loading/-refreshingarrangements, like F11-h’s held-draft states (aiDraftPins'hold'), hold the request with the panel’s real 12-second client timeout running, so a page check slower than that runs on the unavailable state; the F13 suite re-checks the busy state after each scan and fails loudly, but the size-matrix arrangements would pass on the wrong state. Freeze the page clock (page.clock.install()/pauseAt) in those arrangements, or record the dependency in.forge/test-strategy.md § Flaky Tests. -
Promote the AI confidence clamp to one shared helper (Dinil; raised by the review of
forge-et-be#296, F13, 2026-10-22): the same 0–100 clamp of the model’sconfidencenow exists three times —emaildraft/service/AiDraftService.java(clampConfidence),account/service/BriefingOutput.java(clampConfidence, a byte-identical copy F13-WI-1.1 accepted to stay in scope) andlead/service/InquiryExtractionRules.java(theBigIntegerform). Move one version intoai(adding it toFeatureSlicesSpec.AI_SEAM_TYPES) orplatform, call it from the three places, and delete the copies. -
Turn off the backend IT client’s automatic re-send of a
503(whoever next touchesforge-et-be’s test support; raised by F13-WI-1.1, 2026-10-22):RestTestClientover Apache HttpClient 5 re-executes a request that answers503once, after one second, aPOSTincluded (log:request will be automatically re-executed in 1 SECONDS (exec count 2)), so every503test inforge-et-bereaches its endpoint twice and writes twoai_invocationrows. Tests that read only the newest row (AiDraftControllerIT) hide it; F13’sAccountBriefingDisabledITandAccountBriefingLogHygieneITaccept one or two calls because of it. Build the IT clients with retries off (inLoginSupport) and tighten those two tests to exactly one call. Browsers do not re-send, so the product is not affected. forge-et-be#299 does the first half (2026-10-22):LoginSupport’s clients, and every spec’s own client built on it, use the JDK client, which never re-sends, andErrorLoggingITasserts exactly one WARN. Still to do: tighten the twoAccountBriefing*ITfeatures to exactly one call and drop their comments about the Apache client re-executing a POST. -
Tell Empire that account briefings are prepared on request and cover no past orders (Tharinda; raised by the F13 spec, FR-H2, 2026-10-21): the client-facing PRD (revision line of 21 October 2026, §3.3), the draft Who Can Do What (Where the AI fits) and the proposal’s post-issue note of 21 October 2026 now say an account briefing is prepared when someone who may edit the account asks for it, not in the background, and that it does not cover past orders (OQ-63’s default). Send the change with the next client update.
-
Add open follow-ups to the account briefing’s input (Dinil, owner of F13; raised by the F13 spec, decision 7 and FR-H3, 2026-10-21): the briefing finds open items only in the timeline, because no task table was on
mainwhen F13 was specified. F3-a has shipped (closed 2026-11-08), soGET /api/v1/activities?accountId=<id>&view=openis onmain(forge-et-be, wave 1, forge-et-be#273): add the account’s open follow-ups (title, type, due date, bucket) to the input in a new prompt version,account-briefing/v2, with its evaluation cases. -
Add a typed-in last order to the account briefing if Empire chooses OQ-63 option (a) (Dinil, owner of F13, after Tharinda’s answer from Empire): F13 was released by waiver on OQ-63’s default (b), with no order history (F13 spec decision 1). If Empire wants the manual last order field, F1-d’s revision adds the field and an F13 spec revision adds it to the briefing’s input.
-
Decide whether the AI circuit breaker is kept per provider now that features use different models (Tharinda, the lead; raised by Dinil from the forge-et-be#276 review, F11-h wave 1, 2026-10-19): F12 builds one breaker per provider (
ai-provider-strategy.md§4;AiCallRunner,AiCircuitBreakerinforge-et-be), andMODEL_UNAVAILABLE,PROVIDER_ERRORandTIMEOUTall count towards opening it. F11-h’semail-draftruns on Sonnet 5.5, which the account refuses today (register §C-5), while F12’sinquiry-extractruns on Haiku 4.5. So on any environment without the Haiku override for drafts (the item above), five draft presses in a minute open the sharedbedrockbreaker, and inquiry extraction then answersCIRCUIT_OPENuntil it closes. Options: key the breaker by provider and model (an access or quota refusal is about one model, not the provider), stop countingMODEL_UNAVAILABLE, or keep it and record the coupling indocs/ai/README.md. A change refines F12 FR-B5 and the strategy’s §4. -
Decide whether the inline actions-list renderer moves to
src/ui(Tharinda, the lead; raised by Dinil from the forge-et-fe#301 review, F11-h wave 1, 2026-10-19):forge-et-fenow has three copies of the same inline actions-list place —AccountPanelPlace’s inline mode,ContactActionsPlace, and F11-h’sLeadActionsPlace(the first in a different feature, made by F11-h WI-1.2 plan Decision 2 to mirrorAccountPanelPlace).frontend-reviewerasks for one generic place insrc/ui(components, props, group name, test-id prefix andonError, aroundPlaceErrorBoundary) used by all three. Deferred because the plan approved the copy and the refactor changes other features’ components; F3-a adds Add follow-up to the lead’s list next. -
Decide what the requests-per-minute targets for Sonnet 5.5 and Haiku 5.5 become in production (Tharinda, when F38-i’s spec is written; found by F38-a-WI-1.1’s reads, 2026-10-15): Service Quotas has no requests-per-minute quota for either 5.5 model, only cross-Region and Global tokens per minute (
environments-and-identity.md§4 “F38-a — the development account, checked”, AC-4). For dev the lead decided on 2026-10-15 that F38-a AC-4 is met by recording that no such quota exists; the lead writes that as the O-1 line in R-2 when running the WI-2.1 run sheet (.forge/plans/F38-a/F38-a-WI-2.1-plan.md). Production’s targets (F38-i, and the management account’s quota request template) raise the same question. -
Move the Exa search account to Empire before production (Tharinda, with Kavindu Anthony; the lead’s decision of 2026-10-20,
lead-research-source-options.md§11): Altrium holds the free Exa account F11-c is built and tested on. Exa’s API licence is non-transferable, so before F11-c’s search runs in production, Empire signs up for its own Exa account (on the free credit, or with Empire’s card if a paid plan is needed) and the production key is set from it in Secrets Manager; Altrium’s account then stays for dev, so each environment has its own month of free credit (lead-research-source-options.md§11). If Empire declines Exa, this row is replaced by the switch to the provider Empire accepts. The row leaves when production’s key is Empire’s. -
Empire’s agreement in writing and Exa’s data processing agreement (Tharinda, with Kavindu Anthony): F11-c’s Option 2 conditions asked for Empire’s agreement in writing and Exa’s data processing agreement covering an environment’s Exa account (Altrium’s for dev, which UAT also uses; Empire’s for production). Kavindu’s agreement on 10 Nov was spoken, so ask him to confirm it by email, and put the DPA in place for each account. They do not gate the person search: the switch is on in every environment, with no precondition (Dinil, 2026-11-18; F11-i decision 8, spec Rev 3), and Empire agreed to this approach on 2026-11-10. Done when both are in place for dev and for production.
-
E8 and the free credit — when the production migration runs (Tharinda; raised by the F11-c spec, FR-H2 ©, Q-3): F11-c researches the 692 migrated leads through its hourly sweep within a backlog allowance of 900 searches a month, so all of them take about one month at one search a lead — about two months at two, the default since F11-i shipped on 2026-11-18 (the person search on, for every lead with a contact name). E8 says every migrated lead is researched at Phase 1 acceptance; that holds only if the production migration runs early enough, or a card is added to Empire’s Exa account for that month (about $5–10,
lead-research-source-options.md§12). -
Grant
settings.lead_research.managein each environment (Tharinda, the lead, for UAT and production; whoever holds Roles on dev): F11-i shipped (2026-11-18) the Lead research Settings page’s key held by no role (AD #30; F11-i FR-K1), so nobody can change the person-search switch until a holder of Roles grants it on the Roles page to the roles that should. Done when dev, each UAT instance and production have a role holding it. -
Confirm whether the cost annexe’s yearly web-search count includes the person search (Tharinda, who rewrote the annexe on 2026-11-16):
.forge/client/sow-cost-annexe.mdgives about 4,600 searches a year; F11-i makes two searches a lead with a contact name the default. If 4,600 counts one search a lead, raise it (about 5,800 at two, bylead-research-source-options.md§11’s figures plus the show plans) and the “every search paid” figure with it (about $41 instead of $32); the monthly figure, nil inside Exa’s free credit, does not change (F11-i spec Rev 1). Done when the annexe is confirmed or corrected. -
Approve running
heavy-run.py --self-testinRepo hygiene(Tharinda, the lead, who approves CI changes; raised 2026-11-18 by the one-machine-lock change, L-315):.claude/lib/heavy-run.pycarries a--self-test, but addingpython3 .claude/lib/heavy-run.py --self-testto.github/workflows/repo-hygiene.yml’s “Script self-tests pass” step is a CI change, which needs the lead’s approval (.claude/CLAUDE.md § Boundaries). Done when the line is added, or the lead decides not to. -
Decide whether to merge the two cross-process test locks into one module (Tharinda, the lead, who decides; raised 2026-11-18 by
frontend-revieweron forge-et-fe#574, its finding I4): F11-i’se2e/fixtures/leadResearchSettingLock.tsis a second copy of F37-a’se2e/fixtures/priceBroadcastSweepLock.ts(F11-i-WI-1.3 plan D4 chose the copy so as not to edit F37-a’s file), and the copies already differ: only the new one breaks a lock that has had no holder file for 5 s, so the price-broadcast lock keeps that gap. The fix the review proposes is onee2e/fixtures/processLock.tswith both modules as thin wrappers that keep their names and directories, and a run of the price-broadcast browser specs. Done when the lead decides, and, if the answer is yes, that change has merged. -
Decide whether lead research moves to Haiku 4.5 in production (Dinil, with Tharinda; raised by the F11-c spec review, decision 6, Q-5):
lead-researchdefaults to Sonnet 5.5 (AD #19,ai-provider-strategy.md§2), with dev on Haiku 4.5. After the F11-c evaluation set runs on dev — on Haiku 4.5 now, on Sonnet 5.5 once its quota is raised (spec FR-H3 items 5 and 7) — decide whether research stays on Sonnet 5.5 or moves to Haiku 4.5; a move needs the lead’s sign-off on AD #19. Meanwhile, decided by the lead on 26 Oct: research runs on Haiku 4.5 in every environment, production included, until AWS lifts the account’s new-account limits (ai-provider-strategy.md§5 row 10; the proposal’s post-issue note of 26 Oct). This row’s lasting choice waits for Sonnet 5.5 to be callable, which AWS gives no date for. -
Make the comment-pointer checks catch
Decisions <n>andD-<n>pointers (Dinil; raised by the frontend and backend reviews of F11-c’s wave PRs, 2026-10-22):forge-et-fe’sscripts/check-comment-pointers.mjs(byte-shared withforge-et-pwa) matches neither the pluralDecisions 3–5nor a bare contract pointerD-1, andforge-et-be’sscripts/check-comment-pointers.pyaccepts a bareD-nwhen the class Javadoc names the contract elsewhere; both F11-c wave PRs passed the checks with pointers that named no document, which only the reviewers caught. Extend both checks (with their self-tests) so such a pointer must name its document in the same comment, re-copy the frontend script toforge-et-pwain a companion PR, and fix whatever existing pointers the stricter check then reports. Under way (2026-10-23): forge-et-be#308 (merged 2026-10-23) makesD-n,Decisions n–mandfact npointers name their document in the same comment, with a ticket or — for a bareD-nonly — a.mdpath or the entity map, and covers the row below as well; forge-et-fe#336 and forge-et-pwa#89 (both merged 2026-10-23) do the same in the frontends. Still owed now that they have merged: the check reads only added lines, so the bare pointers already onmainstay until someone edits them — inforge-et-be, about 29 comment lines (mostlyD-nwith no document) inmailbox/*,AccountTimelineController,EmailTemplateController,LeadNoteController,LeadTriageView,BuyingStandardService(Spec),BuyingStandardClientCollectionDefinition(Spec),TimelineEvent, theContactListAccount,ContactPrimaryRule,LeadAiDraftandLeadTeaFieldsITs,R__local_seed.sqlanddocs/proofs/pricing/01-schema.sql; inforge-et-fe, about 248 lines in 133 files (mostlycontract D-nandcontract client-side fact nwhose ticket is only in the file’s header comment; pricing, leads, notes and tags, accounts and the AI drafts carry most); inforge-et-pwa, none. Dinil’s decision (2026-10-23): fixed as each line is next edited, not in a sweep PR — the PR that edits such a line adds its document there, since the Fast lane then flags it. The row leaves when the three checks are onmainand those comments name their documents. -
Correct three statements made to Empire on 2026-10-16 (Tharinda, with Kavindu Anthony): that the AI models keep learning from Empire’s data, that no identifiable data is sent to the model, and that synced mail uses no database space. None matches the design or the issued proposal (§A.6.1): the models are not trained on the requests, and Anthropic never sees them (AWS’s Bedrock “Data protection” page), each task’s own content (an email, the lead’s fields, the thread and account context) is sent from Empire’s own AWS account, and every synced message is stored. The table is
.forge/discovery/meeting-notes/2026-10-16-kavindu-progress-update.md§6. Where requests are processed is settled (OQ-15, closed 2026-11-04: any commercial AWS Region; Q18 answered and closed), so this message need not ask it. -
Build the email design into the F20-i, F20-f, F20-j and F20-k specs (whoever takes each feature; F20-f was split into these four on 2026-10-17,
.forge/project-prd-history.mdRev 74). Every decision of 2026-10-16, 2026-10-17 and 2026-10-26 is in.forge/design/architecture.md§ Mailbox integration (Rev 62, Rev 63, Rev 74) — only linked email stored, in full with attachments; the live inbox (50 a page, at most 100; folders, categories, unread, search, an “All mail” view, threads, Focused/Other, a new-mail notice); automatic and manual linking and unlinking, with their permissions; Create lead from this email (F12 spec Rev 1); past mail back 12 months from every connected mailbox, on a new record, a mailbox connect and a bulk import, with a simple full-period scan; block lists; connecting at sign-in, its edge cases and Disconnect in settings, with F20-a’s mailbox screen kept for Try again and Reconnect and a failed connect never affecting sign-in; a contact’s other email addresses, all matched; a converted lead’s email linked to its new records — and each stub’s## Contextsays which parts are its own. The test approach, non-functional checks included, is.forge/test-strategy.md§ Microsoft Graph (Rev 21). The row leaves when the four specs are approved. -
Handle the review points the 2026-10-17 reviews of the email design raised, in the F20-i, F20-f, F20-j and F20-k specs (whoever takes each feature). Decided points are in
architecture.md§ Mailbox integration (previous row); these are for the specs to work out: reply, reply-all and forward from the live inbox need a source, because F20-c takes a storedemail_message(FR-O1, FR-P3) and an unlinked message no longer has one (F20-i settles it: the Graph id, or storing the message on reply), and F20-h’s forward-with-attachments has the same gap; existing connections whose grant lacks the calendar and contacts scopes are reconnected automatically on the next load, not through F20-a FR-6’s button (check first whether Microsoft already returns every consented scope, as F5-i found); F20-j’s jobs get a dead-letter queue and an alarm, like F20-b’s ingest (settled by F20-j spec Rev 2: the dead-letter queue stays; the signal is each failed delivery’s ERROR log line, which production’s backend-error alarm reads — F38-i decision 26); audit events on link, unlink and block carry no email content, so a data-protection deletion is not undone by the audit trail, and a connection released by a lost permission writes one; F20-k owns the blocked-address check inside F20-f’s linking service, which F20-j inherits; each spec cites.forge/design/data/entity-map.md(theemail_messagerow and the new link, attachment and block-list entities); the touch-points in F5-a (the sign-in redirect) and F5-k / F5-d (role changes) are named; if Phase 1 runs short, threads, Focused/Other, the new-mail notice and the category filter are the first to move to Phase 2; F11-h’s shipping PR adds lines to the F20-f stub (F11-h spec FR-H2, FR-H3, AC-33 in PR #740, merged 2026-10-19; the lines are in the F20-f stub’s## Context), so it adds them to the stub as rewritten on 2026-10-17, and F11-h’s decision 7 / FR-D2 reword the after-F20-f case, now that a lead’s linked email, colleagues’ included, is stored and readable by everyone who reaches the lead (noted on PR #740, 2026-10-17). Also from the first review: the “All mail” view cannot be a plain list of/me/messages, which includes Deleted Items and Junk, and Graph refuses most folder filters combined with date ordering (InefficientFilter), so use one Outlook search folder per mailbox over the shown folders, or open on the Inbox; the per-record past-mail search uses Graph’s$searchonparticipants, which returns at most 1,000 results per query, sorted by sent date, so a burst of new records groups several addresses into one query; the live inbox must meet the PRD’s 2-second screen target, so fetch only the list’s fields, load the next page early, and let interactive requests go before background jobs within Graph’s limit of four concurrent requests per mailbox, background jobs backing off on a 429; live-rendered HTML bodies are sanitised (a security-review item); linking, unlinking and blocking are writes with audit events (AD #36). The cloud-gated items are rows incloud-gated-validation.md(F20-f, F20-i, F20-j). The row leaves when the four specs are approved. -
Call F20-j’s past-mail scan after a bulk import (whoever builds the later of F20-j and the import it serves — F4-d, F4-b, or a spreadsheet import): once the import has written its records, it starts F20-j’s per-mailbox scan against the imported addresses (
.forge/design/architecture.md§ Mailbox integration). F4-d shipped without it (done 2026-10-26; its spec does not mention F20-j’s scan), so F20-j, built after it, makes the call for F4-d’s runs. F4-b’s spec (FR-20, 2026-11-06) also leaves the call to F20-j, which F4-b is built before: F20-j finds a run’s created leads by the run id, whichmigration.appliedand everylead.createdof the run carry asmigrationRun, and scans theircontactEmailaddresses. -
Fill the organisation email block list before the go-live mailbox scans (Tharinda, as the lead, or whoever runs go-live; F20-k spec FR-12, D-5, approved 2026-11-20): before the Pipedrive migration’s first production run and before users first connect their mailboxes in production, an administrator adds the known non-sales domains and addresses — suppliers, couriers, service firms, banks — to the organisation block list, so F20-j’s first scans do not link months of their mail. It needs, first, F20-k released to production and a role granted
settings.email_block_list.manage, which ships held by no role. If F20-k is not in production by then (it moves to Phase 2 if not done by the end of Empire’s Phase 1 UAT round, while F20-j ships in Phase 1), the lead decides at go-live planning whether the first scans wait for F20-k or whether mail wrongly linked meanwhile is unlinked one email at a time with F20-f’s Unlink. Done when the list is filled in production before those scans, or the lead has decided. -
Build a bulk removal of a blocked address’s past email only if Empire asks at UAT (Tharinda, as the lead, decides; F20-k spec D-5, approved 2026-11-20): F20-k’s blocks look forward only, and a wrongly linked email is removed one at a time with F20-f’s Unlink. A bulk removal was left out as rarely needed; it would need a background job, a count before removal, and an exception to AD #35’s reach rule recorded as a new AD (the remover may not reach every linked record). Done when Empire’s Phase 1 UAT round ends with no such request, or the work is specified as ordinary work.
-
Test that F20-f’s sweep and F20-j’s scans skip a blocked address (whoever writes the plan of the work item that reaches
mainsecond; F20-k Decomposition Plan § Acceptance Criterion Coverage, decided by Tharinda 2026-11-21; F20-k spec FR-8, AC-6, AC-26). F20-k’s blocked-address check sits inside F20-f’sMailLinkRules.decide. Its current form keeps its signature and reads the block lists itself, so every caller applies the check whatever the merge order. An overload takes a snapshot of the lists, so that a batched caller reads them once per batch (spec NFR-1). The callers outside F20-k are F20-f WI-3.1 (Link to Record) and WI-4.1 (the sweep), and F20-j WI-1.2 (the scan), WI-2.2 and WI-3.1. The owner of each move and test is decided when each side’s wave PR merges, withmainmerged in first:- AC-6b (the sweep): if F20-f WI-4.1 is on
mainwhen F20-k’s wave-2 PR merges, F20-k WI-2.2 moves the sweep to the overload and adds the test. Otherwise F20-f WI-4.1 does both before its own wave PR merges. The test checks that a stored F20-b row whose only match is a blocked address is deleted by the sweep, not linked. - AC-26 (F20-j’s scan): the same rule between F20-k WI-2.2 and F20-j WI-1.2. A scan skips a blocked address (an address entry and a domain entry, personal and organisation) and links the message’s unblocked addresses. F20-j’s WI-2.2 and WI-3.1 are moved to the overload on the same terms, with no AC test of their own.
Done when both tests are on
mainand every batched caller uses the overload.
- AC-6b (the sweep): if F20-f WI-4.1 is on
-
Give
MailboxSeamGuardSpec.FakeMailboxProviderevery operation F20-f adds to the mailbox seam (Tharinda, owner of F20-f, or F20-f WI-1.1’s implementer; raised by the review of F20-j’s WI-1.2 plan, 2026-11-21). The test classFakeMailboxProviderinforge-et-beMailboxSeamGuardSpec.groovyimplements MailboxProvider, so every new abstract seam operation needs an override there or the spec does not compile. F20-f WI-1.1’s plan addsfetchMessageContent(its Decision 1) and does not list that file. F20-j WI-1.1 Rev 1 adds its own two operations there. Done when F20-f wave 1’s PR compilesMailboxSeamGuardSpec(an added override, or the operation given a default in the interface). -
F37-b’s bounce step covers the past-mail scan’s bulk claim, whichever of F37-b WI-1.5 and F20-j WI-1.2 merges second (whoever writes the second of the two PRs; raised by the F20-j WI-1.2 plan, 2026-11-21, its subtask 9(b)). F37-b WI-1.5 sends bounce reports to
NonDeliveryReportListeneron every ingest path, the back-fill included. F20-j WI-1.2 turns the back-fill into the 12-month scan, which claims each page’s unmatched messages in bulk (IdempotencyGuard.claimAll) without storing them, so a bounce report on that path would be claimed and never reach F37-b’s step. If F37-b WI-1.5 is onmainfirst, F20-j WI-1.2’s subtask 9(b) routes the scan’s reports to the bounce step; if F20-j WI-1.2 is first, F37-b WI-1.5 adds the scan to its ingest paths. F20-j WI-1.2 also removespageMessages, which carries F37-b WI-1.5’s message-class expansion ($expand=singleValueExtendedProperties($filter=id eq 'String 0x001A'), its D-2), so the PR that merges second also adds that expansion to the scan’s header-page request (F20-j WI-1.1’sGraphPastMailReads), with itsGraphPastMailReadsSpeccase, and makes the stub’s header page (LocalPastMailReads) keep a stored message’s class. Done when a test onmainshows a non-delivery report seen by the scan reachesNonDeliveryReportListener. -
Tell Empire Teas how email works, before go-live (Tharinda, in the same message as the Correct three statements row): only email linked to a lead, contact or account is kept; the mailbox connects at sign-in for the roles that hold the mailbox permission; any email with a known buyer address in a connected mailbox is linked, back 12 months, and readable by everyone who can see the record — sign-off item A-45 in
client/roles-and-permissions.md§4.4 asks Empire to confirm it and to tell its staff; personal addresses can be blocked. The client-facing PRD (revised 17 October),client/dependencies.mdand the draft Phase 1 acceptance criteria (§F) already say so. Optional, for Kavindu’s ask: a Customer mail view of linked email next to the live mailbox. -
The AI on/off switch and monthly ceiling need a Phase 2 owner (Tharinda): the administrator switch for each AI feature “without a deployment” and the monthly hard ceiling that turns AI off are a V1 commitment in the proposal (§A.3.2; F38 W-10), and no feature builds them — F12 builds only configuration read through one resolver. The lead decided on 2026-10-08 that it is not in Phase 1 (F12 spec Q-1). Give it a Phase 2 feature, or get Empire’s agreement to drop it. (The one-click Create lead from this email is built by F20-f in Phase 1 — F12 spec Rev 1, answering its Q-5.)
-
F20-c —
rel="noopener noreferrer"on an SVG link in an email body is not proved in a browser (Dinil): WI-1.3’sSafeMessageBody(forge-et-fe#237) gives a kept SVG<a>linktarget="_blank"andrel, and opens it withallow-popups-to-escape-sandbox; no test shows every browser honoursrelon SVG links (the impact is bounded: the CSP and the link-scheme allow-list still apply). WI-1.5 assertswindow.openerisnullfor the HTML link only. Decide: add an SVG link to a stub fixture and assert it, or drophreffrom SVG links insafeSrcDoc. Raised by WI-1.3’s pre-PR review, 2026-10-12. -
Decide whether the CRM takes
react-aria-components1.22 (Tharinda, or whoever reviews the Dependabot PR that proposes it).forge-et-feruns 1.21.1, pinned~1.21.1by forge-et-fe#339, which usesreact-aria3.52.1, the versionpatches/react-aria@3.52.1.patch(BUG-089, the combobox fix) is written against; 1.22.1 (published 2026-10-09) usesreact-aria3.53.1, so taking it means checking whether 3.53.1 still needs the BUG-089 change, remaking the patch for 3.53.1 if it does, and running BUG-089’s tests inpnpm e2e:gate, in the PR that widens the pin. Held out of the 2026-10-25 dependency update for that reason. AD #39 givesreact-aria-componentsno new use, so leaving it at 1.21.1 until the backfill removes it is also an option; the row closes on either decision. -
F20-c — hidden or disabled email controls on archived and converted records (Dinil, with the lead Tharinda). The spec decision needs confirming or revising.
- What ships: on an archived account or contact, or a converted lead, F20-c’s Compose email and Email controls are hidden, as F20-c FR-E1, FR-E2, AC-28 and WI-1.2 plan D-20 say (“hidden, never disabled”).
- Why it conflicts: design-system §9’s permission table says a control blocked by its context (rather than by a missing permission) is shown disabled with its reason.
- The choice: confirm the spec’s choice, or revise the spec (a dated Revisions entry) and have a follow-up show the controls disabled with a reason.
- Raised by
frontend-revieweron forge-et-fe#249, 2026-10-13, as a Nit that needs a lead decision; no code change on that PR.
-
F1-e’s notes composer: give the blocked reason through
PermissionGuard, not an innerTooltip(Biseni, owner of F1-e, before WI-1.4 is built). Since BUG-014 (forge-et-fe#250),PermissionGuardwraps its child in its own tooltip even when the action is allowed. A guarded child that is itself a<Tooltip>, as F1-e WI-1.4 plan Decision 17 and its subtask describe for Add note, therefore logs “<Focusable> child must forward its ref”.- Follow
LeadNotesSectionafterf7c7ce5: a localPermissionProviderresolver turns the blocked state (“Write a note to add it”, “The note is being added”) intodenied-contextualwithunavailableReason, with one guard around a plainButton. - Update the plan’s Decision 17 wording to match.
- Follow
-
Two frontend fixes from the 2026-10-04 audit (Tharinda — one forge-et-fe PR plus the matching forge-et-pwa copy):
- F-08 —
forge-et-fenow has the generic hook,src/ui/useMediaQuery.ts(used by the account header,useTabletLayoutand the compose dialog), but two places still subscribe to media queries by hand with their ownuseSyncExternalStore:src/app/useSidebarLayout.tsandsrc/features/auth/LoginRoute.tsx(a third,src/features/leads/useLeadsLayout.ts, was removed by forge-et-fe#338 on 2026-10-26;git -C forge-et-fe grep -ln useSyncExternalStore origin/main -- srcon 2026-10-26 lists these two, the generic hook and the tests, andsrc/features/leadTriage/leadTriageQueries.ts, which subscribes to no media query). Rebuild them on the hook, and list it inshared-with-fe.jsonif the capture app’s copy is to stay byte-shared (coding-standards §0 rule of two). - F-09 — the literal lengths are tokens now (
layout/login-card,layout/message-max,layout/skeleton-block, design-system Rev 37; forge-et-fe#148 and the forge-et-pwa audit PR), and the token scan reads every inline style. Left: the CRM shell’s100vh(inline inAppLayout.tsxandBootstrapGate.tsx, a recorded exception of the scan; also in some CSS modules —git -C forge-et-fe grep -ln 100vh origin/main -- src) could become100dvhafter checking it in the browser gate’s WebKit tablet sizes.
- F-08 —
-
The shadcn/ui backfill (AD #39) (Nanndish, and each developer whose PR leaves a screen part-moved): move the screens listed in
.forge/design/ui/shadcn-backfill.mdover, one PR per screen, no ticket. The Users screens’ remaining parts come next: the create-user dialog and the editor’s dialogs (shadcn Dialog), the manager picker (Command in a Popover, replacing the react-aria Combobox), the toasts (Sonner), then shadcn’s Sidebar for the app frame. -
Tell the team about the shadcn-only test (Nanndish):
src/app/shadcnOnly.test.tsfailspnpm teston non-shadcn UI thatsrc/app/shadcnBackfill.jsondoes not list (AD #39). A PR or local branch started before it does not run it until it mergesmain, so each author mergesmainin before their test record and converts or lists what the test reports (.forge/design/ui/shadcn-backfill.md § How a PR in flight is handled). The three frontend PRs open on 2026-10-12 (forge-et-fe#226, #200, #184) have all merged since; what is left is telling the team, which is not recorded as done. -
The shadcn token gaps left open by forge-et-fe#201’s review (Nanndish; accepted as follow-ups by Nanndish on 2026-10-11, not by the lead; forge-et-fe#201 merged without them on 2026-10-13):
- The semantic token names in
forge-et-fesrc/ui/styles/theme-revamp.css(--et-foreground,--et-card,--et-border,--et-border-widthand the rest) have no neutral fallback intokens.css, so the shadcn components draw nothing with no theme imported, andneutralRender.test.tsxdoes not coversrc/ui/shadcn.tokens.cssis byte-shared withforge-et-pwa(check:shared), so the fix needs a matching capture-app PR (the capture app follows shadcn/ui since AD #43; itstheme-crm.cssdeclares the names its shadcn components read). src/ui/styles/shadcn.cssstill uses Tailwind’s own shadows (shadow-xs,shadow-sm,shadow-md, plain black) and Tailwind’s 4px spacing scale, not--et-shadow-*and the §4 spacing tokens. Mapping the shadows onto--et-shadow-cardand--et-shadow-menuchanges how cards, inputs and the Select list look; moving the Users screen’s lengths (left-3,pl-9,h-10,min-w-24,h-4,min-h-6) onto--et-space-*changes their sizes. Decide whether to map them or to record the shadcn primitives’ own geometry as an AD #39 exception. Tailwind’s default colour palette is already removed (--color-*: initial).
- The semantic token names in
-
Browser and unit test failures seen on 2026-10-09 while checking the first shadcn PR (Nanndish — log each in
.forge/bugs.mdor fix it):e2e/browser/roles/roles-a11y.spec.ts“the Deletedtoast passes the page check” fails on a clean mainevery time (9 of 9 runs, three sizes): after the delete is confirmed the page stays on the role. The same intermittent symptom appears indelete-role.spec.tsandadmin-journey.spec.tsunder load.e2e/browser/users/edit.spec.tsAC-7 (choosing a manager in the react-aria picker) andjourney.spec.tsfail in most full Chromium runs of the Users folders and pass when run alone.see-and-add.spec.tsAC-4/AC-9 failed the same way once (the manager picker did not offer IT Admin within 5 s, with about 1,200 test users in the local database) and passed alone. BUG-110’s fix (forge-et-fe#184, merged 2026-10-12, after these runs) makes every Users spec that opens the manager list wait for it; a full Chromium run of the Users folders confirms whether these are gone.- Vitest once failed to start test files (“Failed to start forks worker… Timeout waiting for worker to respond”) while a browser run used the same machine; the same files passed with
--maxWorkers=2.
-
The review gate’s last acceptance checks (Tharinda, who sends the drafted steps to a teammate;
.forge/design/harness/forge-review-gate.md, implementation note of 2026-10-02 on the Dependabot and release-branch checks): AC-7’s last half, a person’s push onto a Dependabot PR getting no automatic status (AC-7’s success half and AC-8 passed on forge-et-be#190, implementation note of 2026-10-05); AC-10, the release-branch probe, run by a teammate with write access and no admin (Tharinda to send the drafted steps), or else checked when the firstrelease/*branch is cut. Still unrun on 2026-10-16. -
The Docker naming rule on Windows (Dinil, or another Windows developer): AC-8 of
.forge/design/harness/docker-resource-labels.md(its § Remaining, step 3). In Git Bash and again in PowerShell, run AC-1’scd <worktree> && docker run -d --name et-… --label et.workdir="$PWD" …case through the hook, then AC-6’s first case throughbash .claude/lib/cleanup.sh --test-workspace <dir> --min-idle-minutes 0, and confirm the script removes the container once its folder is gone. The path comparison was tested only on macOS. -
Real-backend tests to add when F11-e (convert) ships (whoever specs F11-e, Phase 2 — unowned as of 2026-10-16, Tharinda until then). No route converts a lead yet, so tests that need a converted lead insert it directly or pin it. F20-c: AC-5 and AC-30 create a converted lead by a direct insert; re-run them through F11-e’s conversion route (
.forge/lessons.mdL-166). Offboarding: F5-h-WI-1.5’s browser test (e2e/browser/users/offboarding.spec.ts) gives the leavernew,approvedandarchivedleads — F11-h added the last two through its approve route and F11-f’s archive route (F11-h spec FR-H4, AC-32, forge-et-fe#301). When F11-e ships, it adds aconvertedlead the same way, so spec AC-11’s “every status” is proved end to end on a real backend; until then it is met by four tests together (lead decision 2026-10-08, F5-h Q-1;.forge/plans/F5-h/F5-h-decomposition-plan.md § Notes). When F11-e ships a route that converts a lead, the sharing panel’s controls on a converted lead are also proved on a real backend, ine2e/browser/sharing/, because F5-g tests them over a pinned record only (F5-g AC-4). -
How a restarted mail delta round is bounded (Tharinda — a design decision from the review of forge-et-be#192, BUG-069): a delta round the poll restarts after a refused or dropped link is filtered to the back-fill window and capped by Graph at 5,000 messages, and the back-fill does not cover it. Decide whether to narrow the restart’s filter (for example to
last_synced_atless a margin), start a gap back-fill, or accept the gap; the dev measurement incloud-gated-validation.md§B F20-b (which messages a capped round returns) informs it. -
Hear Add tag’s “Type to find N more tags” with a screen reader (Tharinda, or whoever next works on the lead view’s tags — from the review of forge-et-fe#224, BUG-123): open Add tag on a lead with more than ten matching tags under VoiceOver in Safari (and NVDA if a Windows machine is free), then type one letter. The line is said from a polite live region; react-aria announces its own “N options available” a moment later. If that announcement cuts the line off, move the count into react-aria’s announcement or the input’s description, and record what was heard in
.forge/bugs.mdBUG-123. -
Hear Template saved and User saved after an edit dialog closes, with a screen reader (Nanndish — Nit 14 of
frontend-revieweron forge-et-fe#381): a save in Edit template or Edit user calls the toast and closes the dialog in one render, so the toast may be inserted while Radix still holds the page’s Toaster regionaria-hidden. Save once in each under NVDA on Windows (and VoiceOver in Safari if a Mac is free) and note whether the toast is announced. If it is not, show the toast once the dialog has closed (TemplateViewonceeditingturns false, andUserViewlikewise). Edit lead has the same timing since forge-et-fe#422 (frontend-reviewerNit 10 on it):LeadFormDialog.tsxraises “Saved N changes to” and then closes the dialog. Check it in the same pass, and move its toast with the other two if they move. -
Start each record edit dialog’s focus on its first field (Nanndish — an out-of-scope note from
frontend-revieweron forge-et-fe#422):RecordEditDialogFrame(forge-et-fesrc/ui/overlays/RecordEditDialog.tsx) leaves Radix to focus the first tabbable element, which is the ✕, while design-system §5 Modal says focus starts on the first field. It affects every dialog on the frame: Edit account, Edit contact, Edit user, the role editor, Edit template and Edit lead. Decide whether the frame moves first focus to the first field of its body (for example anonOpenAutoFocusthat focuses the body’s first control), and make the change in one PR across the dialogs, with a test for each. -
Three known-failure test follow-ups, both frontends (Tharinda — Nits 3 to 5 of
frontend-revieweron forge-et-pwa#67, BUG-148’s fix; each aforge-et-fechange copied toforge-et-pwawhere the file is shared): (1)src/app/knownFailureMarks.test.tsstill exempts the size-matrix specs from the literal-bug-id rule (FROM_THE_LIST) and says they pass the id tomarkKnownFailure, which none does since forge-et-fe#148 — drop the exemption and the sentence; (2)src/app/knownFailureMatch.test.ts’s “each name at least one kind” row runs only over the realPAGE_STATES, which hold no marks, so it cannot fail — pull the rule into a pure function and add synthetic rows (no kinds, plainaxe), as the marks test does withbadListMarks; (3)forge-et-pwalacks two thingsforge-et-fehas: thee2e/README.mdparagraph “A marked test holds only the assertion its bug breaks” (L-131) and the size matrix’s check that the engine matches the project name. -
Build the load check for test runs (Tharinda):
.forge/design/harness/load-check-and-flaky-tests.md(approved 2026-10-15). The browser gate and the backend and frontend test records refuse to start on an overloaded machine and record the load during the run, and flaky-test bugs are triaged by recorded load, with overload-only failures closedwontfix. Built and merged on 2026-10-15: forge-et-fe#272, forge-et-pwa#79, forge-et-be#261, and the rules (H-1 to H-5) in forge-et-harness#703. Still owed: (1) run each on Windows (design AC-5) —pnpm testin forge-et-fe (itsmachineLoad.test.tsruns the real Get-Counter command), andpython scripts\local-build-evidence.py --self-testplus anET_LOAD_LIMIT=0.01refusal with--no-postin forge-et-be — Nanndish (deferred here by Tharinda, 2026-10-15, from forge-et-be#261’s review, finding I-2b; assigned to Nanndish by Tharinda the same day); (2) the Windows flaky-test bugs are triaged under it on Windows by Nanndish: BUG-013, BUG-128, BUG-129, BUG-183 and BUG-184 (BUG-196 was closed on 2026-11-16, fixed by forge-et-fe#455). BUG-194 and BUG-195 were closedwontfixon 2026-10-15. The limit was set to 0.65 on 2026-10-15 (design Q-1; forge-et-fe#279, forge-et-pwa#81, forge-et-be#264). -
Build the rest of Phase 2 of the test-run cost design: W-3, V-2, H-3 and M-6 (Tharinda, the lead):
.forge/design/harness/test-run-cost.md§ Where this stands. The selected build and test records and the browser gate’s removal from PRs intomainwere switched on on 2026-11-02, by the lead’s decision that day, without waiting for seven green nights (S-4 as revised; PRD Rev 81). Meanwhile the lead reads the weekly flaky report and the trend (python3 .claude/lib/flaky-report.py; F-2, H-6). -
Keep the nightly full run reliable, now that it is the only full run (Tharinda, the Mac’s owner; raised 2026-11-02 with the lead’s decision to take the browser gate off PRs into
main): the lead’s Mac stays on, plugged in, logged in and open overnight, with nothing left on the gates’ ports 15173 and 15174 or the CRM gate’s backend port 18080 (the repeating 01:55 wake is set:pmset -g sched). A developer’s backend on 18080 and a load reproduction’s busy loops cost the CRM gate of 13 November; the runner now waits up to 30 minutes for 18080, and a hand run closes the red issue once the repo is green (forge-et-harness#1351, #1359;.claude/rules/test-record.md, L-293). A closed lid cost the nights of 31 October, 1 November and 2 November, and a 1% battery the night of 3 November (the Mac hibernated at 02:56, 18 minutes into the CRM gate;test-run-cost.md, revision lines of 2026-11-02 and 2026-11-03). A missed or red night is fixed before new work starts (git-conventions.md § Pull Requests). Decided 2026-11-03 (the lead): it stays on the lead’s Mac, with no always-on machine. Decided 2026-11-03 (the lead): a missed or interrupted night is run again by hand the next morning (python3 .claude/lib/nightly/run-nightly.py --now,git-conventions.md § Pull Requests); the session start’s missed-night line now says so. Each browser gate needs its own port free (15173 for the CRM, 15174 for the capture app, since forge-et-fe#399 and forge-et-pwa#108 merged on 2026-11-04); apnpm devor a hand-run Playwright spec on 5173 no longer stops a gate. (The runner’s own leftovers are stopped since 2026-11-02; another worktree’s are not.) Keep it on mains power as well: on 2026-11-02 a 4% battery made it hibernate in the middle of the CRM gate. The database and LocalStack ports no longer clash (the nightly stack has its own since 2026-11-02). A held port 5173 stopped a gate on 28 October, 2 November and 3 November (an ad-hoc Playwright run’s preview server, stopped by hand that day), which is why the gates have had ports of their own since 2026-11-04 (test-run-cost.md, revision of 2026-11-04). -
Decide whether to build an
/et-nightly-triageskill (Tharinda, the lead; raised 2026-11-03): sorting the night of 3 November’s failures by hand took a few hours — re-running each alone on currentmain, sorting it into an environment problem, a stale test, a real defect or a flake, and finding or filing its bug (the steps are in.forge/guides/test-runs-explained.md § The nightly full run). A skill would read thenightly-redissue and~/.et-nightly/logs/<date>/, re-run each failure, sort it, find or file the bug and hand back a fix plan; the fixes still go through PRs, records and review. On 2026-11-06 the guide’s step 2 (re-run only the failing tests, at the night’s commits, against a backend started as the night starts it) sorted that night’s one CRM gate failure as machine load in minutes (forge-et-fe#430); those commands are what such a skill would run. -
block-dangerous-bash.shrefuses a safe command shape (whoever next changes the hooks; found 2026-11-03): an inline script that writes files only in app-repo worktrees is refused when the same command names a.claude/path elsewhere. It was worked around by splitting the command. Narrow the write check to the paths the script writes, with hook tests. -
Small leftovers from 2026-11-03 (whoever next edits each file):
forge-et-be’sscripts/local-build-evidence.pyheader paragraph (line 12) and two comment blocks in.github/workflows/pr-validation.yml(around the Spotless choice and the Slow lane) want re-wrapping, left because a re-wrap needed another full build record; and.forge/bugs.mdhas headings out of id order (BUG-315 above BUG-323, BUG-322 below BUG-292, and four older pairs), left because other sessions were editing the file. -
Check a backend response change against the frontend’s specs before merge (Tharinda, the lead; raised 2026-11-03 by BUG-324):
/forge-review-prstep 2 checks aforge-et-bePR against the frontend tests that assert a refusal it turns into an acceptance (L-200); forge-et-be#342 (AD #41) changed what/meand the role read return — a key with a dotted resource, a list of six keys in ascending order — and four CRM browser specs that assert those failed the next night. Recommended: widen step 2 to any backend PR that changes a response’s fields, values or order, by agit grepof the frontends’e2e/for the changed endpoint or key before the record. -
Widen
forge-et-be’s always-run set, now that the build record uses the selection (whoever next changesscripts/select-tests.py;backend-reviewerNit 6 on forge-et-be#348, deferred 2026-11-02): two specs read whole-repo data in waysREADS_FILESdoes not see —TimestampEntityShapeSpecscansfindCandidateComponents('com.empireteas'), andUnauditedWriteCheckSpecreadsdocs/audit/README.mdthroughAuditCodeDocumentSpec.DOCUMENT. Each has a backstop (a new entity needs a migration, which is shared code; the audit document’s own spec is selected for a change to it) and the nightly run catches both. Add specs that scan the root package ('com.empireteas'besidefindCandidateComponentsorClassFileImporter) and specs that name a file-reading spec’s class to the S-7 set. In the same change, say inbuild.gradle’s test-cache comment that a selected record’s filter is its own cache input (Nit 8; the matchingpr-validation.ymlcomment is a CI change for the lead). -
Name how many related test files a selected test record ran (whoever next changes
scripts/local-test-evidence.mjs;frontend-reviewerNit 6 on forge-et-fe#379, deferred 2026-11-02): the selection line names the changed files and the always-run tests that read files as text, but not how many test files Vitest found related to the change, so a change no test imports reads as tested. Add the count (the JSON report’s files less the always-run set) to the comment, and keep “(selected)” in the carried description. -
Move the shell parser out of
guard-browser-gate.mjsand delete its PR-decision code (whoever next changes the hooks; raised 2026-11-02): the browser-gate hook was unregistered and its bash wrapper deleted when the gate left PRs intomain, butguard-browser-gate.mjsstays because.claude/hooks/lib/shell-parse.mjsre-exports its parser forguard-docker-resources.sh(.claude/hooks/README.md). Move the parser intolib/shell-parse.mjs, delete the decision code and the gh stub it alone uses, and keep the parser’s tests in.claude/hooks/tests/. -
Run the test-run tooling on Windows (Nanndish, the Windows machine; raised 2026-10-27): what was built on 2026-10-27 has run only on macOS. On Windows: (1) start a
pnpm test:record --no-postinforge-et-fe, then a second one in another terminal, and check the second waits naming the first, starts when it ends, and thatGet-Process powershellshows the wake-lock process during a run and not after (design M-5, M-2); (2) the same withpython scripts\local-build-evidence.py --no-postinforge-et-bebeside a frontend record, which checks thetasklistliveness check; (3)pnpm test:selected --listandpython scripts\select-tests.py --listprint sensible choices with Windows paths; (4) M-4: compare the load record’sbusyfigures with Task Manager over one record (one Windows record read “busy 1.00 at start, 1.00 peak, 0.17 average”, an average below both); (5) after Q-6 and M-1, the before-and-after figures of the build record, the test record and the gate on one commit, for the design’s § Measurements (AC-4). File a bug for anything that fails. -
Answer Q-6: Dev Drive or Defender exclusions on the Windows machines (Nanndish, with whoever approves machine security;
.forge/design/harness/test-run-cost.mdQ-6, M-1; raised 2026-10-25): decide between a Windows 11 Dev Drive for the repositories, the pnpm store,~/.gradleand Docker’s data (Defender keeps scanning there, in its performance mode), or Defender exclusions for the same folders and thejava,nodeanddockerprocesses, and whether either needs Altrium’s security approval; then measure a build record before and after on one commit, for the design’s § Measurements. Nanndish measured the causes on 2026-11-02 (Defender at about 80% of a CPU, NTFS small-file costs, Docker through a named pipe into WSL 2, the search indexer, one test JVM — the design’s revision line of that date). Recommended: a Dev Drive first; the.wslconfigcap, removing the workspace from search indexing and a quiet machine on mains power can be done now. Parallel test JVMs are B-7, in Phase 3 after B-2 and B-3. -
Should a failed-chunk error page also offer Back to the dashboard? (Tharinda, a UX decision; raised by
frontend-revieweron forge-et-fe#280, BUG-139, Nit 4.) Since BUG-139, a chunk that failed to load offers only Reload the page, in place of the whole app (BUG-200). That is right after a deploy. But Chromium gives the same message when the network is down, and then a reload does not help and nothing else leads back into the app. The options:- Keep as is: a reload only.
- Two actions: a reload first, with Back to the dashboard as a second action.
- Add a line of copy: for example “The CRM was updated while this tab was open.”
If it changes, it is a small change in
src/app/RouteErrorElement.tsx. -
Record the rest of the 21 Oct demo’s UI decisions (Nanndish; the demo of 2026-10-21, design-system Rev 70): Rev 70 records the first set — tables for every list screen but Email, the ⋯ row menu, the whole row opening the record in view mode and the menu’s Edit opening it in edit mode, centred record pages with their own Edit action, every create flow in a modal, Enter
placeholders with no example values, Settings as a page of links with the email templates, Users and Roles moved there, and the sidebar order. The placeholder wording for select, date and picker controls was decided by Nanndish on 2026-10-27: Select (design-system Rev 74), built on forge-et-fefeature/ui-revamp. The demo decided more that is still to be written down; each addition goes into Rev 70 while it is unmerged, or a later revision, and intoshadcn-backfill.md § Target layouts. -
Decide each list screen’s third row action (Nanndish, with the UI/UX engineer; design-system Rev 70): every table row’s ⋯ menu holds View, Edit and one action specific to the screen, and none is decided yet — Leads, Accounts, Contacts, Pricing (the demo’s example was Archive), Users, Roles. Decide each in that screen’s plan or PR, and record it in its
shadcn-backfill.md § Target layoutsrow. -
Revise the approved specs the UI restructure contradicts (whoever builds each change; design-system Rev 70): F11-b (the two-pane leads list, and the first lead opened on arrival), F11-f (the lead detail’s column order), F20-d (email templates reached from the inbox header), F1-e (Settings › Tags as tabs), F1-b and F1-d (the account’s New page at
/accounts/new, which becomes a modal), F1-f (the contact view as a pop-up, which becomes the contact’s own page at/contacts/:id), F20-d again (the New template page and a template opening straight into its editor), F5-b (Users and Roles in the sidebar), F5-c and F5-d (the users and roles lists in the sidebar, and their records opening straight into the editor), and every list screen’s spec where a row opens its record only from the name (F1-b, F1-f, F26-b, F5-c, F5-d); a spec that fixes a placeholder’s text is revised in the PR that changes it. Each gets a## Revisionsentry in the PR that builds its change, never a silent edit (.claude/rules/specs.md). Done for forge-et-fe#337 on 2026-10-26: F1-b Rev 8, F1-d Rev 7, F1-e Rev 6, F1-f Rev 3, F5-b Rev 3, F5-c Rev 9, F5-d Rev 5, F20-d Rev 1 and F26-b Rev 6 (forge-et-harness#865). Left: F11-b and F11-f, with the Leads table; F5-b again when the sidebar is reordered; and any spec a placeholder change touches. (F1-d Rev 8 makes the account page a centred column, 2026-10-28.) -
Write the spec revisions for Rev 70’s first changes, which are being built without a spec and plan (Nanndish; his decision of 2026-10-23): the whole-row click, the ⋯ menu with View, Edit and each table’s existing row actions, view and edit modes for the user, role, contact and email-template pages (edit at
/<section>/:id/edit, and a successful save returning to view mode, the account page’s included), the contact page, and the New account and New template modals are built onforge-et-febranchfeature/ui-standardisation, and the Leads table onfeature/shadcn-leads-table, with no spec or plan, by the developer’s decision. The row above lists the approved specs each one contradicts; their## Revisionsentries are written before either branch’s PR is marked ready, and each PR body says that no spec or plan covers it and links here. In the same change,shadcn-backfill.md § Target layoutsmarks each screen the PR moves (the Accounts, Contacts, Email templates, Users, Roles and Pricing tables, the user, role, contact and template pages, the account form, Settings and its tag pages), as step 7 of How a screen is moved over asks (forge-et-fe#337 review).feature/ui-standardisation(forge-et-fe#337) done on 2026-10-26: its spec revisions and backfill rows are in forge-et-harness#865 (the row above). Left: the Leads table’s (feature/shadcn-leads-table). -
Decide whether contact names on the account page open the contact page (Nanndish, with the UI/UX engineer; design-system Rev 70): the account page’s contacts list shows each contact’s name as plain text, while the Contacts list’s name opens the new contact page at
/contacts/:id. Left as plain text for now by Nanndish’s decision of 2026-10-23; decide whether the name becomes a link to the contact page, and record the answer inshadcn-backfill.md § Target layouts. -
Decide whether the two contact dialogs share one shell (Tharinda, raised by
frontend-revieweron forge-et-fe#374, 2026-11-02):forge-et-fesrc/features/accounts/NewContactDialog.tsx(the contacts page’s New contact, F1-f Rev 7) repeatsContactFormModal.tsx’s dialog shell — the title row with ✕, the scrolling body, the footer, the latest-props ref, theonClosedeffect and the no-close-while-saving rule. The F1-f new-contact plan’s Decision 1 kept them apart so F1-d’s Add contact and its tests stayed unchanged. Decide whether to extract a shared contact-dialog shell in a later change that may touch F1-d. -
Decide what a record page’s back link returns to (Nanndish, with the UI/UX engineer; design-system Rev 70): the contact page’s ← Contacts opens the Contacts list with its search and filters cleared, while the browser’s Back returns to the list as it was left; the account page’s ← Accounts keeps the list’s view. Kept as built for now by Nanndish’s decision of 2026-10-23; choose one behaviour for every record page’s back link and apply it to each.
-
Decide whether the issued proposal gains a note on the Settings permissions (Nanndish, with Tharinda, whose decision it is; AD #34 refined 2026-10-26, AD #41): the issued proposal’s post-issue note of 24 September lists user and role administration and system settings as separate permissions. Decide whether a later post-issue note says that each Settings page is now granted on its own (
.claude/rules/source-trust.md). Narrowed 2026-11-18: the client documents already say it —client/client-facing-prd.md§2 (“the pages under Settings, one permission for each”) andclient/roles-and-permissions.mdnote 4. Done when the lead decides, and a yes is written. -
Write the spec revisions for the per-page Settings keys, which are built without a spec and plan (Nanndish; his decision of 2026-10-30, with Tharinda’s agreement to AD #41): each Settings page has its own key (AD #41), built on
forge-et-beandforge-et-febranchfeature/settings-page-access. The approved specs that namesettings.manageas one key — FS-6 (its Appendix A and FR-P11, the last-admin rule), F5-c, F5-d (the role editor’s Settings row), F5-e, F5-h, F1-e, F20-d, F4-a and F4-d — each get a## Revisionsentry before either PR is marked ready, never a silent edit (.claude/rules/specs.md), and each PR body says that no spec or plan covers it and links here. In the same change, the PRD’s capability matrix and its user-administration paragraphs (settings.managein.forge/project-prd.md, with a revision under.claude/rules/prd.md),architecture.md,design-system.md(a revision for the role editor’s Settings row) anddocs/audit/README.md’s label list say the same. The other live statements of the single key, found by the fresh-context review of forge-et-harness#1041, are revised in the same way: the specs F5-b (the Settings entry shown and guarded onsettings.managealone; it now shows to a holder of any page key), F4-b (the migration run, now both the Users and Roles keys), F38-f (/actuator/loggers, now both keys), F38-h AC-10 (the seeded IT Admin on a fresh instance, now every page key), F38 and F38-d (what/mereturns), F7-d (grantingquote.approve_sub_floor, nowsettings.roles.manage), F32-c, F11-f and F5-l (the run-precondition caller check, now both keys);.forge/test-strategy.md(the AD #26 negative case, the last-admin invariant at all four points, the fixture IT Admin, the privilege-escalation row);.forge/design/data/entity-map.md,record-sharing-model.md(the effective-permissions viewer, nowsettings.users.manage) androles-and-permissions-key-map.md;.forge/design/ui/shadcn-backfill.md(the Settings entry);.forge/design/prototypes/feature-detail.md(last-admin protection);.forge/guides/tech-stack-reference.md;.forge/cloud-gated-validation.md(the §B F5-k seed-admin row, which expects/meon dev to returnsettings.manageand now returns every page key, and the production import owner’s role), with a## Logline; this file’s import-owner row;.forge/plans/F38-h/uat-data-checklist.md(the Administrator role’s keys); and the F38-g WI-1.1 plan’s smoke-account check, which namessettings.manageamong the keys it refuses (confirm it is an allow-list of the C-3 keys, or add the six page keys to it). -
Decide which key governs configuration that has no Settings page (Tharinda; raised by the fresh-context review of forge-et-harness#1041): AD #34 put “all configuration” under
settings.manage, and the PRD capability matrix still says so; AD #41 replaces it with one key per Settings page and says a later page adds its own key, but does not say what governs configuration with no page — the AI feature switches (F12), the certification list (F1-a, F9) and pipelines (F24, Phase 2). Custom fields are settled: the F34-c spec (approved 2026-11-08, FR-9) gives them a Settings page with its own key,settings.custom_fields.manage, as AD #41 provides for a later page, covering F34-a’s record types too. The admin email block list is settled the same way: the F20-k spec (approved 2026-11-20, D-3) gives it the Email Block List Settings page with its own key,settings.email_block_list.manage. As built, no endpoint for such configuration checks a Settings key. Decide whether each gets its own key when its page is built (AD #41’s rule for a new page), and if so add that clause to the AD #41 row and revise those specs. -
Decide whether the CRM needs a notification bell (Tharinda, the lead; raised by the review of
.forge/design/prototypes/crm-current/against the PRD, 2026-11-05): the original prototype (crm/) has a bell in its top bar, but no PRD feature or spec delivers in-app notifications, and the built CRM has no top bar. The current prototype leaves it out. If it is wanted, it needs a PRD change and a feature; if not, nothing changes. -
Tell Empire the CRM’s screens are being restructured (Tharinda; design-system Rev 70): the prototype Empire approved shows leads as a list beside the selected lead’s detail, and the CRM now moves to tables, with each record on a full-width page of its own (design-system Rev 77, agreed by Tharinda on 2026-11-02). Say so at the next client touchpoint, so a change to an approved screen is stated rather than found at UAT (
.claude/rules/source-trust.md). -
Backend sorting for the columns that cannot sort yet (Nanndish; raised by the UI revamp,
forge-et-fefeature/ui-revamp, 2026-10-27, design-system Rev 74): every table’s header now sorts, frontend only, and only the columns the list endpoint already sorts by, or every row of a list the page holds whole (src/lib/params/clientSort.ts). The rest keep plain headers until their endpoint’s sort whitelist takes them, a backend change with an endpoint contract revision for each: Leads — Status, Owner, Labels; Accounts — Est. LTV, Last contact, Status, Owner; Contacts — Job title, Account, Phone; Users — Role, Manager, Status; Email templates — Subject; the Standards book — Client, Type, Held certifications, Price / kg, Valid until, Price state; the weekly round — Grade and the eight figure columns (Previous to Final); Auction sales — every column but Starts. Decide with the lead which of them users need, then add them per list. -
Add a background export if a list outgrows the export limit (Tharinda; the lead’s decision of 2026-11-15, F23-d spec FR-E7): a list export is built in its request, up to 10,000 rows (
et.export.max-rows), with no queue or stored file; above the limit the user is asked to narrow the filter. Empire’s largest list, its contacts, is about 5,000. If a list nears the limit (anEXPORT_TOO_LARGErefusal that cannot be narrowed, or F23-a’s deal lists growing past it), design the background path — an SQS job, the file in S3 behind an FS-11 signed URL, its expiry and a “ready” notice — as a new feature with its running cost named againstsow-cost-annexe.md. -
Sort the offboarding page’s owned-records list (Nanndish; raised by the UI revamp, 2026-10-27): its headers stay plain, because
GETowned records (forge-et-fesrc/features/users/offboardingApi.ts, F5-h) takes no sort parameter. Sorting needs an F5-h contract revision and the backend’s sort whitelist; the frontend sorts only the direct-reports list, which the page holds whole. -
Move the Account and Lead edit forms onto shadcn’s Input, Select and Checkbox (Nanndish; seen on 2026-10-27 when the edit pages took the save bar, design-system Rev 74): both forms still use the older
src/uicontrols, so their fields are about 68px tall against the Contact and User editors’ 56px, on the same edit-page layout. Moving them makes the four edit pages match; each form’sshadcnBackfill.jsonentries shrink with it (shadcn-backfill.md). -
Decide whether New account asks only for the required fields (Nanndish, with the UI/UX engineer; raised 2026-10-27 with the edit-on-the-record’s-page rule, design-system Rev 74): the dialog holds the whole account form (about 30 fields, the certification checklist included). The alternative asks for the name, country and customer type, then opens the new account’s page to fill in the rest. Left as it is until decided.
-
Add Rows per page to the Email templates pager (Nanndish; seen on 2026-10-27): the shared
ListPagershows it on Users, Accounts, Contacts, Leads and the Standards book, and the templates list leaves it out because its page size is fixed. Needs the list to take asizeparameter, as the others do. -
Review three choices the edit pages made (Nanndish, with the UI/UX engineer; built on
forge-et-fefeature/ui-revamp, 2026-10-27, design-system Rev 74): (1) the lead’s edit page hides the page’s other sections (panels, customer, tags, notes) while editing, because the notes section has its own leave guard and the router honours one blocker at a time, while the user editor keeps its sections below the form — settled 2026-11-05 (Nanndish): design-system Rev 96 edits the lead in a dialog over its page, which stays on screen with its sections; (2) the Account and Lead edit pages show no per-field change marks, which the Contact and User editors do (the account form’s checkbox groups do not fit the mark); (3) saving an account shows no toast, because the account page it returns to already says “Account saved” — decided 2026-11-03 (Nanndish): a toast, design-system Rev 90; the change is in the “Move every modal and every success message onto design-system Rev 90” row. Keep or change (2). -
Show that a tag’s Retire or Restore is working (Nanndish; built on
forge-et-fefeature/ui-revamp, 2026-10-27): the tag’s row reads the tag first, and the inline button used to show busy while it did; in the ⋯ menu, which closes on the press, nothing shows it. Decide whether the row shows a busy state (as the contacts table’s Restoring…) or the wait is short enough to leave. -
Put the email template editor on the save bar (Nanndish; found on 2026-10-28 while writing design-system Rev 74): the template editor is already the template’s own page, as the edit-on-the-record’s-page rule asks, but it keeps its own Save, Cancel and Delete under the form instead of the §6 Save bar at the top that the account, contact, lead, user and role editors use. Move it onto the save bar, or record why it differs.
-
Add acceptance criteria for two UI-revamp requirements (Nanndish; from the same review, 2026-10-28): F11-b FR-C1 (the New lead dialog at the
widewidth, at most two columns, footer Cancel then Create lead; since design-system Rev 90 the dialog islgwith Cancel then Create lead at the right and no footer band, so the criterion asserts Rev 90’s frame) and F11-f FR-V1 Rev 4 (the lead page’s header: breadcrumb, back arrow beside the title, the meta line under it) have no acceptance criterion. Add one each, as a spec revision, with the browser test that proves it. -
Tell Empire that contact email addresses and phone numbers are plain text, and how records are edited (Tharinda, at the next client touchpoint, with the row “Tell Empire the CRM’s screens are being restructured”; built on
forge-et-fefeature/ui-revamp, 2026-10-27, F1-f Rev 4, design-system Rev 74 and Rev 77): the contacts table, the contact page and the account page’s contact list no longer open the computer’s mail or phone app, or the CRM’s compose, from an address or a number; and Edit on an account, a contact, a user or a role opens a dialog over that record’s page, while a lead and an email template are edited on their own page in edit mode, with a save bar. -
Move the account list’s data bar onto
DataBarControls(Nanndish; deferred by the review of forge-et-fe#338, 2026-10-23): forge-et-fe#338 moved the Leads list’s data-bar controls (the labelled select, the switch and a failed read’s Retry) tosrc/ui/filters/DataBarControls;src/features/accounts/AccountsFiltersstill holds its own copy of the same three. Move it onto the shared one oncefeature/ui-standardisation(forge-et-fe#337, which changesAccountsFilters) has merged, so the two branches do not conflict. -
Fix the comments that name removed Leads files (Nanndish; deferred by the review of forge-et-fe#338, 2026-10-23, kept out of that PR because the files belong to other branches): forge-et-fe#338 removed
LeadListPane,LeadRowanduseLeadsLayout, and these comments still name them —src/app/AppLayout.module.css:8,src/app/programmaticFocusRing.test.ts:9,src/features/email/InboxList.module.css:33,src/features/email/MessageRow.tsx:13,16,src/features/email/MessageView.tsx:56,214,src/features/email/MessageView.test.tsx:543,559andsrc/features/accounts/AccountFormRoute.create.test.tsx:579. Point each at what replaced it (LeadsTable,LeadPage,src/ui/recordRow.tsx) or drop the reference. -
Check the Leads notice after a back/forward-cache restore in Chrome by hand (Nanndish; deferred by the review of forge-et-fe#338, 2026-10-23;
.forge/lessons.mdL-109): forge-et-fe#338 carries the Leads notice in history state (Lead created on the new lead’s page, no longer available on the list) and clears it on apageshowwithpersisted: true(usePageRestore), with component tests. Playwright cannot produce a real back/forward-cache restore, so check once in Chrome: show a sticky notice, go to another site, come Back, and confirm the notice is not shown again. -
Decide whether a list table pins its actions column (Tharinda; raised by the review of forge-et-fe#338, 2026-10-23): at 1280 × 800 the Leads table is 32–69px wider than its card, so the ⋯ column scrolls out of view at a desktop size; the Users table has the same layout. Design-system §6 Tables has the table scroll inside its card both ways, so this is as designed, and keyboard users reach the menu (the browser scrolls a focused cell into view). Pinning the column (
position: sticky; inset-inline-end: 0frombp/tabletup, as BUG-232 set for the pinned first column) would change every list table, so it is a design-system decision. -
Run
check-test-stats.py --self-testinRepo hygiene(Tharinda — a CI change, so the lead approves it; found by the documentation review of 2026-10-26):.claude/lib/check-test-stats.pyhas a--self-testmode that no workflow runs, the one exception.claude/lib/README.mdnames. Add it to the “Script self-tests pass” step of.github/workflows/repo-hygiene.yml, and in the same change drop the exception from.claude/lib/README.md. -
Move the five older data-bar search boxes onto
FilterSearch(Nanndish, or whoever next changes one of the files; deferred by the review of forge-et-fe#343, 2026-10-26, because that PR did not change them): forge-et-fe#343 addsFilterSearchtosrc/ui/filters/DataBarControls.tsx— the search glyph inside a 40px shadcnInput— and uses it on the contacts page and the/emailfilter.AccountsFilters.tsx,LeadFilters.tsx,pricing/RoundToolbar.tsx,pricing/StandardsToolbar.tsxandusers/UsersScreen.tsxeach still write their own copy of the same markup; replace each withFilterSearch, keeping its test id and accessible name. -
Prove on dev that a remote image in a received email is never requested (F20-c AC-35) (Dinil, or whoever next walks F20-c on dev; raised by F38-g WI-4.2’s F20-c.8, 2026-11-12): Outlook on the web embedded the picture pasted into the test message (a
data:image/pngrow, no request to any web host), so the walk shows the subject as literal text and the link opening withwindow.openernull, but not that a linked image is left unrequested. Send a message whose HTML holds an<img src="https://…">(from a mail tool that keeps the address, or a Graph send), open it in/email, and check the network panel shows no request to that host. Remove the row when it is proved or when the lead marks it won’t-do. -
Prove on dev that choosing a template clears an AI draft’s chip and confidence (F20-d AC-26 to AC-31, the F20-d.9 AI part) (Dinil, or whoever next walks F20-d on dev; raised by F38-g WI-4.2’s F20-d.9, 2026-11-16): on the F20 lead (
/leads/13) AI Draft is offered, but Haiku 4.5 returned no draft (INVALID_OUTPUT,NO_TOOL_USE, invocation 377) because the lead’s inquiry holds only a test string, so the dialog fell back to writing by hand. Use a lead whose inquiry reads like a real buyer’s, get an AI draft in Compose Email, then choose a template and check the AI chip and confidence go. Remove the row when it is proved or when the lead marks it won’t-do. -
Fold the two older copies of
ListQuery.singleBooleaninto it (whoever next changesAccountServiceorLeadService; found by the delta review of forge-et-be#391, 2026-11-09, outside that PR’s diff):forge-et-beAccountService.sharedWithMeFilter(AccountService.java:734) andLeadService.booleanFilter(LeadService.java:430) are line-for-line copies ofListQuery.singleBoolean, the duplication coding-standards §2 Reuse asks to remove. Callquery.singleBoolean(<name>)in each and delete the copy; the account and lead list ITs cover both.