Engagement Gate Runs
Append-only audit log for the three pre-implementation gates: PRD readiness, architecture probe, decomposition.
Format: newest entry on top. One
## Gate N Run Mblock per run. Updated automatically by/forge-prd-check,/forge-arch-probe,/forge-decomposein full mode (skipped indry-runmode).Why this file exists: the engagement-level artifacts (PRD, architecture, backlog) don’t have a natural lifecycle-metadata section like specs do (
## Revisions). This file is their shared audit trail. Per-feature gates (spec review, plan review, adversarial review, security review) record their findings inside the per-feature spec / plan artifacts and do not appear here. Tool-call-level gates (hooks) are ephemeral and do not appear here.Audit storage principle: see
.claude/rules/tracker.md“Gate Audit Protocol” section.
Gate 3 Run 4 — Reslice — 2026-11-06 (Tharinda)
- Mode: reslice apply, scoped to F32-a, F23-a, F33 and F34-a at the lead’s request, with three whole features moved to Phase 1 in the same session.
- Runner: Claude under Tharinda’s direction. Verdict: the lead chose the split and the move on 2026-11-06.
- Why now: Phase 2 does not fit its 24 December build target on its own (R-ARCH-001), and Phase 1 is due on 20 November. The lead asked which Phase 2 work could start in Phase 1’s time without bringing in deals or the capture app. Three Phase 2 features need neither: F20-g (tracking and scheduled send), F22 (contact sync) and F37-a (the weekly price broadcast). Four more need deals only because they also cover deals: list views (F32-a), list export (F23-a), attachments (F33) and custom fields (F34-a). Their account, contact and lead parts have no open blocker.
Feature Sizing Audit
| Feature | Cross-layer spread | Substrate-slice consumers | Capability clusters | Capability density | Tripped? | Outcome | Justification (if kept-as-one) |
|---|---|---|---|---|---|---|---|
| F32-a | yes — a saved-view table, the filter builder, every list screen | F23-a, F32-b, F34-b, F37-b | 2 — the engine on the existing lists · the Phase 2 lists | over 5 | yes (S2: the existing lists wait on F2-c only because the deal list is in the same feature) | cluster-cut (by record type) → F32-d, F32-a | — |
| F23-a | yes — an export job, a signed download, the export action | — | 2 — the export on the existing lists · on the Phase 2 lists | 3 | yes (S2, through F32-a) | cluster-cut (by record type) → F23-d, F23-a | — |
| F33 | yes — link tables, the panel, the allow-list, scan states | — | 2 — the store on accounts, contacts and leads · the Phase 2 records | over 5 | yes (S2: blocked by F2-c for the deal panel only) | cluster-cut (by record type) → F33-a, F33-b | — |
| F34-a | yes — definitions, a Settings page, JSONB columns, form fields | F34-b, F23-b | 2 — definitions on accounts, contacts and leads · the Phase 2 records | over 5 | yes (S2: blocked by F2-c for the deal form only) | cluster-cut (by record type) → F34-c, F34-a | — |
Applied Splits
- F32-a → F32-d (list views and saved filters on accounts, leads and contacts, with the contact views F37-b sends to) and F32-a (the deal, quote, sample and buying-standard lists, narrowed in place). F32-d takes the next letter on F32 (F32-c is taken;
.claude/rules/id-numbering.mdrule 4). - F23-a → F23-d (export of the account, lead and contact lists) and F23-a (export of F32-a’s lists, narrowed in place). F23-d is the next letter (F23-c is dropped but keeps its letter).
- F33 → F33-a (attachments on accounts, contacts and leads; F33’s stub renamed) and F33-b (deals, quotes, samples and product briefs). F33 had no letter, so its own tracker row is removed, as F38’s was on 2026-10-11; PRD F33 stays the parent.
- F34-a → F34-c (custom field definitions and forms on accounts, contacts and leads) and F34-a (deals, quotes, buying standards and samples, narrowed in place).
- Moved to Phase 1 whole, not split: F20-g, F22 and F37-a.
- Placement: F32-d, F23-d, F33-a and F34-c, and F20-g, F22 and F37-a, are tracked in Phase 1 to start early. Phase 1’s own work, fixes and release come first; each moves back to Phase 2 before Phase 1 seals unless done (as F38-i). F33-b and the narrowed F32-a, F23-a and F34-a stay in Phase 2. No owner is assigned.
- Dependencies: F32-d ← F1-b, F11-b, F1-f, F5-g; F23-d ← F32-d; F33-a ← F1-d, F1-f, F11-f; F34-c ← F1-d, F1-f, F11-f (all done but F32-d). F32-a ← F32-d, F2-c; F23-a ← F23-d, F32-a; F33-b ← F33-a, F2-c; F34-a ← F34-c, F2-c. F37-b’s saved lists come from F32-d instead of F32-a. Every new edge runs within Phase 1 or from Phase 1 to Phase 2.
- Preconditions: each split feature at phase backlog, spec a stub, no plan, no review annotation — all passed.
- Writes:
tracker.yaml(the new and narrowed features, the three moved, F37-b’sblocked_by, R-ARCH-001’s note,setup.decomposition.reslice_runsandlast_gate_run);features.md(Delivery Plan, graph, lanes, rows);project-prd.md§F23, §F32, §F33, §F34 andproject-prd-history.mdRev 84; the stubs created, renamed and narrowed, with revision entries in the approved specs that named the old owners (Rev 84 lists them).
Refused
None.
Gate 3 Run 3 — Reslice — 2026-10-26 (Tharinda)
- Mode: reslice apply, scoped to F20-f at the lead’s request, in the session that recorded the email decisions of 2026-10-26.
- Runner: Claude under Tharinda’s direction. Verdict: the lead chose the split on 2026-10-26.
- Why now: the lead decided that a contact can hold several email addresses, all matched (
design/architecture.md§ Mailbox integration, Rev 74). The work — a contact table change, the contact API and screens, the Pipedrive migration’s mapping and F11-d’s proposed match — has no email dependency and must land before the first production run of the migration, while F20-f is blocked by F20-b, F1-c, F11-f and F20-i.
Feature Sizing Audit
| Feature | Cross-layer spread | Substrate-slice consumers | Capability clusters | Capability density | Tripped? | Outcome | Justification (if kept-as-one) |
|---|---|---|---|---|---|---|---|
| F20-f | yes — new storage, backend jobs, inbox actions, and now a contact schema change with its screens and the migration | the other addresses are read by F20-f, F20-j, F11-d and the migration | 2 — linking and the timeline · a contact’s addresses | over 5 | yes (S1, S2) | substrate-cut → F20-l | — |
Applied Splits
- F20-f → F20-f (linking and the timeline, unchanged in id) and F20-l (several email addresses per contact), Phase 1, priority high, owner none. F20-l takes the next letter on F20 (
.claude/rules/id-numbering.mdrule 4). - Dependencies: F20-l ← F1-a, F4-d (both done); F20-f gains F20-l. F20-j and F20-k stay behind F20-f.
- Preconditions: phase backlog, spec still a stub, no plan, no review annotation — all passed.
- Writes:
tracker.yaml(F20-l, F20-f’sblocked_by,setup.decomposition.reslice_runsandlast_gate_run);features.md(rows, graph, lane C);project-prd.md§F20 and §Domain Model, andproject-prd-history.mdRev 77; the F20-l stub created and the F20-f stub narrowed.
Refused
None.
Gate 3 Run 2 — Reslice — 2026-10-17 (Tharinda)
- Mode:
/forge-decompose --reslice, dry-run then apply, scoped to F20-f at the lead’s request. - Runner: Claude under Tharinda’s direction. Verdict: the lead accepted the proposed cut on 2026-10-17.
- Why now: the email design decided on 2026-10-16 and 2026-10-17 (a live inbox; only linked email stored; automatic and manual linking; past mail for new records, connected mailboxes and imports; block lists; connecting at sign-in —
design/architecture.md§ Mailbox integration, Rev 62, Rev 63) made F20-f several features’ worth before its spec was written.
Feature Sizing Audit
| Feature | Cross-layer spread | Substrate-slice consumers | Capability clusters | Capability density | Tripped? | Outcome | Justification (if kept-as-one) |
|---|---|---|---|---|---|---|---|
| F20-f | yes — new storage (links, attachments in S3, block lists), backend jobs, several screens | none schema-only (F1-c and F11-g need linking itself) | 4 — live inbox · linking and timeline · past-mail jobs · block lists | well over 5 | yes (S1, S3, S4) | cluster-cut → F20-i, F20-f, F20-j, F20-k | — |
Applied Splits
- F20-f → F20-i (live inbox and connecting at sign-in), F20-f (linking and the timeline), F20-j (past-mail linking), F20-k (block lists), all Phase 1, priority high, owner none. The ids follow
F<n>-<letter>(.claude/rules/id-numbering.md): F20-f keeps its id, so the authored specs citing it for linking (F1-a, F1-c, F1-d, F11-f, F11-g, F20-c, F26-d) need no retargeting — FR-7.a precondition 5 warned on them and was cleared on that basis. - Dependencies: F20-i ← F20-b; F20-f ← F20-b, F1-c, F11-f, F20-i (the live inbox ships before F20-f stops storing unlinked mail); F20-j ← F20-f; F20-k ← F20-f. The only tracker downstream of F20-f, F20-e, is dropped and keeps its edge.
- Preconditions: phase backlog, spec still a stub, no plan, no review annotation — all passed.
- Writes:
tracker.yaml(three entries, F20-f’sblocked_by,setup.decomposition.reslice_runs);features.md(rows, graph, lane C);project-prd.md§F20 (one sentence) andproject-prd-history.mdRev 74; the F20-f stub narrowed and the F20-i, F20-j and F20-k stubs created.
Refused
None.
Gate 3 Run 1 — 2026-09-17 (Tharinda)
Mode: full (/forge-decompose), run by Claude under Biseni’s direction after two dry-runs the same day (gate3-dry-run-2026-09-17.md, -rev2.md — held outside the repo, shared with the team). Outcome: ✅ pass. Verdict given by Tharinda (lead), relayed by Biseni on 2026-09-17 with the team’s answers to the eighteen review questions below. Gate 3 has no risk-acceptance path; no failure was open at the verdict.
Preconditions: Gate 1 passed-with-risks (2026-07-07), Gate 2 passed-with-risks-and-spikes (2026-08-06, no spike open), foundation sealed 2026-09-16 (the precondition entry below). PRD at Rev 31 when the dry-runs ran; Rev 32 (the F4 “deal history” correction, finding 3) lands in this run’s PR.
Slicing principle chosen: by capability, cut vertically, phased by the lead’s instruction — and cut to the size one developer can build and one reviewer can verify end to end in a short cycle (the Gate-3 standing instruction, CLAUDE.md § Active Context). Substrate siblings (F1-a, F2-a, F11-a, F20-a, F26-a) are kept where they unblock the other lanes and are verified by API tests plus their first consuming screen. Rejected: by module, by user journey, one feature per PRD row, and the first dry-run’s own coarser leaning (cut only for parallelism or phase placement).
Feature count: 89 — 37 in Phase 1, 52 in Phase 2 — from 31 PRD-level candidates (F1–F26, F31–F34, F36). The first dry-run proposed 58; the lead’s finer-cut instruction (Q1) took it to 89. Eleven splits are marked borderline (◐) in features.md; the team accepted all of them (Q16) — merging every one back gives 78. F37 (bulk email, PRD Rev 28) stays a stretch row in features.md and is not in the tracker’s features:.
R-ARCH-001 restatement. The risk’s scope basis was “30 features plus the full-scope capture app”. 89 features is a re-slicing of that same basis, not scope growth: every sibling maps to a PRD paragraph already inside the basis, and the four items that did grow the basis since it was last stated — AD #30/#32 configurable roles, the F36 additions of Rev 23, Rev 30’s research pass, Rev 31’s price build-up — are each recorded against the risk where they landed. The cycle-day band itself is not re-computed here: forge-estimate’s baselines are calibrated (v0.34.0, 2026-07-24, source: org) but measure agent cycle time from other projects without human queue time, and under the parallel-delivery assumption they return 1.2–1.8 wall-clock days per feature — a number that cannot govern a three-developer schedule. The band is owed a re-run with per-feature wave × WI vectors once the first Phase 1 specs are wave-decomposed; /forge-set-dates sets the phase dates meanwhile.
Delivery phases (two, matching the client’s two UAT rounds):
| Phase | Title | Status | Features |
|---|---|---|---|
| 1 | Users, accounts, leads, email and weekly prices | in-progress, started: 2026-09-17 |
37 — F5-a…h · F1-a/b/c/d/e · F13 · F11-a/b/c/d/f/g/h · F12 · F3-a · F20-a/b/c/d/f · F4-a/b/d · F26-a/b/c/d/f/g |
| 2 | Deals, tea workflows, capture app and parity | locked |
52 — F2-a/b/c/d · F11-e · F3-b · F17-a/b · F24 · F18-a/b · F6-a/b/c · F7-a/b/c/d/e · F8-a/b · F9 · F10 · F14 · F15-a/b · F16 · F19 · F20-e/g · F21-a/b · F22 · F32-a/b · F23-a/b/c · F33 · F34-a/b · F4-c · F26-e · F31 · F25-a/b · F36-a/b/c/d/e/f |
current_phase: 1; every due_date null until /forge-set-dates. Two Phase 1 placements were taken against the lead’s Phase 1 / Phase 2 rule by team decision with the lead’s confirmation (Q7-B, Q8-B, Q17-A): F3-a (follow-up tasks on leads and accounts — the parallel run would otherwise have leads with no reminders, the number-one stated Pipedrive pain) and F20-d (email templates, seeding a predefined weekly-prices template so F26-d does not hand-roll its body). F12 (AI extraction) is Phase 1 as the first AIProvider consumer (Q4).
The team’s answers (2026-09-17, Biseni relaying; Q17 and the verdict are Tharinda’s): Q1 finer cut · Q2 split F21, F32, F25, F5-c (editor/viewer) · Q3 propose splits for F3, F8, F17, F34 · Q4 F12 Phase 1 · Q5 one drafted follow-up in Phase 1, cadences F25 Phase 2 · Q6 P1–P4 with a hot/cold label · Q7 lead-only activities in Phase 1 · Q8 templates in Phase 1 with a predefined weekly-prices template · Q9 OQ-51 default = admin pre-creation · Q10 the phasing divergence from proposal §6 is confirmed · Q11 no past-orders panel, new OQ · Q12 fix the F4 “deal history” sentence in this PR · Q13 SearchProvider spikes after Gate 3, before the F11-c spec · Q14 no owners at Gate 3 · Q15 PRD Rev 31 folded into F26-a/F26-c · Q16 keep all 89 · Q17 both Phase 1 pulls confirmed · Q18 Dinil owns the SearchProvider spikes.
Checklist walk (decomposition-checklist.md):
| Section | Result | Notes |
|---|---|---|
| Slicing — principle stated / applied uniformly | ✅ ✅ | Above; every row cut by the same test |
| Coverage — every in-scope item mapped / no scope creep | ✅ ✅ | One gap found and resolved as OQ-63 (finding 1). F5-f/F5-g are PRD Rev 29; F20-d’s seeded template is inside F20’s committed “starter set” |
| Separation — no double ownership / shared concerns in foundation | ✅ ✅ | Shared surfaces given explicit owners: dashboard (F17-a shell; cards by F6/F7/F18), deal drawer (F2-c; panels by F2-d/F8/F9/F10/F19), timeline (F1-a model, F1-c rendering; rows by F1-e/F3-a/F11-f/F20-f), list visibility (F1-b/F11-b own/all, F5-f team, F5-g shares), lead detail (F11-f; panels by F11-c/g/d/h), account detail (F1-d; panels by F1-e/F1-c/F26-b/F13) |
| Dependencies — graph drawn / no cycles / critical path / parallel lanes / foundation first | ✅ ✅ ✅ ✅ ✅ | features.md § Dependency Graph; walked edge by edge in both phases; yq check on the written tracker: no dangling blocked_by, no Phase 1 → Phase 2 edge |
| Sizing — FR-1 on every candidate / one-run heuristic / oversize split | ✅ ✅ ✅ | Table below; reslice mode not used |
| Per-feature spec readiness — stubs follow the template / self-contained | ✅ ◐ | 89 stubs from _TEMPLATE-spec.md with Context, In/Out of Scope and Constraints pre-filled; self-containment is what /forge-deliver’s spec authoring adds |
| Delivery phases — defined / themed / all assigned / order respects graph / one in-progress / demonstrable | ✅ ✅ ✅ ✅ ✅ ✅ | Phase 1 demos as sign in, manage users and roles, migrated accounts and leads with AI research, ranking and drafted follow-ups, follow-up tasks, email and templates in the CRM, weekly prices published and sent |
Findings and dispositions:
- PRD F1 promises “past orders via ERP” and nothing feeds it in V1 (OQ-2: manual handoff, no ERP→CRM read). F1-d renders no orders panel; OQ-63 raised against F1-d with a manual last-order field as the fallback to put to Empire (Q11-A).
- The Phase 1 / Phase 2 split diverges from proposal §6 and
project-timeline.md§2 — deals, Kanban, activities-on-deals, dashboard and search move to Phase 2, so the Phase 1 UAT round and the parallel run cover users, accounts, leads, follow-up tasks, email, templates and weekly prices, with Pipedrive staying the deal record until Phase 2. Confirmed by the team (Q10-A). This PR records it three ways: a post-issue note on the proposal header (never an edit to the issued text, L-001), aproject-timeline.md§2 revision, and the line to Empire below for Tharinda to send (client-contact rule,team-guide.md). - PRD §Functional Surface F4 still said “deal history” against §Scope’s fresh-start decision (2026-09-13). Corrected in this PR — body edit +
project-prd-history.mdRev 32 (Q12-A). SearchProviderprovider must be chosen before the F11-c spec, now Phase 1. Two spike items indesign/ai/lead-research-source-options.md; owner Dinil, after this gate and before the spec (Q13-B, Q18-B).- OQ-51 gates F5-c. Default recorded: admin pre-creation — the record exists, the person signs in with Microsoft; no SES dependency on F5 (Q9-A).
- F20-e (
info@) and F36-f (same-day email) are conditional on OQ-54 and OQ-60; isolated as their own features so the answers hold up nothing else. - Custom fields in the migration: F4-a/F4-d/F4-b migrate typed fields in the parallel-run import; F4-c brings activities and custom fields in the final migration after the data freeze (timeline §3).
- PRD Rev 31 landed between the two dry-runs; the four-figure price build-up is in F26-a and F26-c. The
forge-et-be/docs/data-model.md§2 revision stays Tharinda’s (AD #31). - The first dry-run reported
forge-estimateas uncalibrated (baselines: null); the file onmainis calibrated. Noted; the forecast caveat above is the consequence.
Line to Empire (draft for Tharinda): “We have finalised the build order. The first release for your testing (early November) covers signing in with Microsoft, users and roles, your accounts and leads brought over from Pipedrive with AI research and ranking, follow-up tasks, email and templates inside the CRM, and the weekly price publishing and per-customer send. Pipelines, deals and the Kanban boards follow in the second release (late November) — until then Pipedrive stays your record for deals. This is a re-ordering of the same scope, not a change to it.”
Owed after this gate: /forge-set-dates for the two phase dates (lead); the SearchProvider spikes (Dinil, before the F11-c spec); the line to Empire (Tharinda); the R-ARCH-001 band re-run once the first Phase 1 specs are wave-decomposed; FS-15’s three framework follow-ups and the FS-7 SYSTEM principal owner (unchanged, carried from the foundation review).
Writes made by this run (one harness PR): 89 stub specs under .forge/specs/; .forge/features.md regenerated; tracker.yaml features: (89), delivery.phases (2), current_phase: 1, setup.decomposition, setup.status: ready; this entry; the PRD ## Gate Status row; the PRD F4 paragraph + project-prd-history.md Rev 32; project-prd-signals.md OQ-63 + the Blocks re-mapping to sibling ids; the proposal header post-issue note; project-timeline.md §2; CLAUDE.md § Active Context.
Feature Sizing Audit
S1 cross-layer spread · S2 substrate-slice consumers (schema-only) · S3 capability clusters after cohesion · S4 raw capability density. ↔ = borderline S2 (consumers exist but are few or later-phase). Outcomes as accepted at the verdict.
| Feature | Cross-layer spread | Substrate-slice consumers | Capability clusters | Capability density | Tripped? | Outcome | Justification (if kept-as-one) |
|---|---|---|---|---|---|---|---|
| F1 Accounts | yes | yes — F26, F20-f, F4-d, F36-d, F2-a, F11-a, F22 | 2 | 7 | yes | substrate-cut → F1-a; cluster-cut → F1-b, F1-d, F1-e, F1-c | — |
| F2 Deals | yes | yes — F3-b, F24, F18, F17, F31, F6, F7, F8, F10, F11-e, F19 | 2 | 7 | yes | substrate-cut → F2-a; cluster-cut → F2-b, F2-c, F2-d | — |
| F3 Activities | yes | ↔ F18, F25, F17 | 1 | 5 | yes | cluster-cut → F3-a, F3-b (by linked entity and phase) | — |
| F4 Migration | yes | no | 3 | 7 | yes | read-write-cut → F4-a, F4-d; cluster-cut → F4-b, F4-c | — |
| F5 Auth/users/roles | yes | no | 4 | 8 | yes | cluster-cut → F5-a, F5-b, F5-c, F5-d, F5-e, F5-f, F5-g, F5-h | — |
| F6 Samples | yes | ↔ F17, F33, F1-d | 2 | 7 | yes | cluster-cut → F6-a, F6-b, F6-c | — |
| F7 Quotes | yes | yes — F19, F17, F14, F31 | 3 | 10 | yes | cluster-cut → F7-a, F7-b, F7-d, F7-c, F7-e | — |
| F8 Briefs | yes | ↔ F19 | 1 | 4 | yes | cluster-cut → F8-a, F8-b | — |
| F9 Cert check | no | no | 1 | 2 | no | kept-as-one | not tripped |
| F10 Export docs | yes | no | 1 | 3 | yes (S1) | kept-as-one | one rules engine and its checklist; no cut boundary |
| F11 Leads | yes | yes — F4-b, F36-a, F20-f, F5-f, F17, F31 | 4 | 11 | yes | substrate-cut → F11-a; cluster-cut → F11-b, F11-f, F11-c, F11-g, F11-d, F11-h, F11-e | — |
| F12 Extraction | no | no | 1 | 2 | no | kept-as-one | not tripped; AIProvider builder |
| F13 Briefing | no | no | 1 | 1 | no | kept-as-one | not tripped |
| F14 Quote assistant | no | no | 1 | 2 | no | kept-as-one | not tripped |
| F15 AI dashboard | no | no | 1 | 3 | no | cluster-cut → F15-a, F15-b | — (finer-cut instruction, not a trip) |
| F16 Brief assistant | no | no | 1 | 1 | no | kept-as-one | not tripped |
| F17 Dashboard | no | no | 1 | 6 | yes (S4) | cluster-cut → F17-a, F17-b | — |
| F18 Alerts | ↔ | no | 1 | 3 | no | cluster-cut → F18-a, F18-b | — (finer-cut instruction, not a trip) |
| F19 ERP handoff | yes | no | 1 | 3 | yes (S1) | kept-as-one | one document, one checklist, one seam |
| F20 Email | yes | yes — F21-b, F22, F20-b, F20-c | 4 | 10 | yes | substrate-cut → F20-a; cluster-cut → F20-b, F20-f, F20-c, F20-d, F20-g, F20-e | — |
| F21 Calendar | yes | no | 1 | 4 | yes (S1) | read-write-cut → F21-a, F21-b | — |
| F22 Contact sync | ↔ | no | 1 | 2 | no | kept-as-one | not tripped |
| F23 Import/export | no | no | 2 | 5 | yes (S4) | read-write-cut → F23-a, F23-b; cluster-cut → F23-c | — |
| F24 Pipeline config | no | no | 1 | 4 | no | kept-as-one | not tripped |
| F25 Sequences | yes | no | 1 | 6 | yes | cluster-cut → F25-a, F25-b | — |
| F26 Standards & pricing | yes | yes — F9, F7-a, F2-d | 4 | 12 | yes | substrate-cut → F26-a; cluster-cut → F26-b, F26-c, F26-f, F26-g, F26-d, F26-e | — |
| F31 Search | no | no | 1 | 2 | no | kept-as-one | not tripped |
| F32 List views | yes | no | 1 | 5 | yes | cluster-cut → F32-a, F32-b | — |
| F33 Attachments | yes | no | 1 | 3 | yes (S1) | kept-as-one | one upload path on FS-11, one list per record |
| F34 Custom fields | yes | ↔ F23-c, F4-c, F32 | 1 | 4 | yes | cluster-cut → F34-a, F34-b | — |
| F36 Capture app | BE + FE, no CRM schema | no | 4 | 9 | yes (S4) | cluster-cut → F36-a, F36-e, F36-b, F36-c, F36-d, F36-f | — |
Retargeting (FR-3): every schema-only consumer of F1, F2, F11, F20 and F26 points at the substrate sibling, so S2 resolves under the strict FR-2 rule for all five substrate cuts. The per-edge table is in the second dry-run report §2; the written blocked_by lists in tracker.yaml are the record.
Seam builders (R-FND-001): AIProvider — F12 · SearchProvider — F11-c · MailboxProvider — F20-a · AuctionPriceProvider (no-op) — F26-a · ErpHandoffService — F19. Workflow primitives (AD #29): RoundLifecycle — F6-a · ChecklistEngine/GateRule — F7-b · StaleClock — F18-a · tag comparison — F9.
Gate 3 precondition — Manual foundation review — 2026-09-15 (Tharinda)
Mode: manual (.forge/specs/foundation/README.md step 4 — one session over every foundation spec and plan; no command drives this step).
Outcome: ⚠️ pass-with-risks. Verdict given by Tharinda.
Trigger: every non-cloud foundation slice done — FS-1, FS-2, FS-3, FS-4, FS-5, FS-6, FS-7, FS-9, FS-10, FS-11, FS-14 (FS-12/FS-13 dropped before start by AD #29). FS-8 is at Status: draft and was reviewed beside this walk by lead decision D3 (2026-09-15); it is not a precondition.
Scope and method. The walk was thematic, not one pass per slice: each slice had already carried a lead approval, a multi-pass pre-PR review and a Reflect, so the layer no per-slice review could see was the cross-slice one — the coding-standards.md §6 invariants, the hand-offs one slice promised to another, and the accumulated substrate load. Inputs: the eleven specs, the plans’ ## Progress and ## Notes, the eleven findings files, the FS-8 draft, the checklists, lessons L-028/L-035/L-041 and source-trust.md. Every tree-state claim below was verified against origin/main of the three app repos (forge-et-be fbe664a, forge-et-fe 0641bfb, forge-et-pwa b366792), not taken from the plans. Settled before the walk and not raised as findings (lead, 2026-09-15): D2 — first-administrator provisioning is owned (F5’s spec + the FS-1 cloud half); D4 — the Node pin is settled inside the FS-8 spec; and the recorded cloud-gated work — the FS-1 Terraform bootstrap, the FS-5/FS-7/FS-9 cloud halves, FS-11’s GuardDuty provider (AD-O-4), the RDS engine pin and app-role search_path hand-off, FS-6’s cloud smoke-login, certificate thumbprint check and CSRF cookie host question, and the Tomcat CVEs on gradle.lockfile (a Boot patch bump owed before the Trivy gate flips at Phase 1→2). /council is not installed in this harness; the walk ran without it.
Checklist walk — coding-standards.md §6, the foundation-strength invariants:
| Item | Status | Reasoning |
|---|---|---|
| 1 Boundary enforcement red-green | ✅ as superseded (AD #31) | Eight arch specs on main — FeatureSlicesSpec, LayeringSpec, NoSchemaQualifierSpec, NullMarkedSpec, CollectionBindingSpec, FixtureReachabilitySpec, RoleComparisonSpec, ProductionClasses; FS-14 AC-5 red-greened six probes; FS-10 AC-12 |
2 application.yml shape + fail-fast + .example twins; ≤ 30-min seeded onramp |
✅ first half · ⏳ the onramp half is FS-8’s, not expected yet | EtProperties carries 12 validation annotations; one committed .example twin per repo; fail-fast red-greened at FS-1 AC-9, FS-2 AC-8/AC-19a, FS-6 AC-1, FS-11 AC-8 |
3 Error envelope + correlation id + @RestControllerAdvice on the reference endpoint |
✅ | FS-10 AC-8…AC-11 on the wire; FS-5 AC-3 closed the traceId hand-off — the real tracer populates the envelope, and a 500’s id equals the ERROR line’s |
| 4 One scheduled job, one queued job, one DLQ alarm end-to-end in dev, with an idempotency-key replay test | ✅ in-process halves · ☁ the alarm and “in dev” are cloud-gated | FS-7 AC-2…AC-4 and AC-7 (consume, replay no-op, poison → DLQ, trace across the queue) against LocalStack; FS-9 AC-1/AC-2/AC-5 by hand-enqueue; the DLQ alarm is the FS-5/FS-7 cloud half; no development environment exists |
5 The four provider seams — AIProvider, MailboxProvider, ErpHandoffService, AuctionPriceProvider — each with a local stub; PriceModel first-class |
❌ → R-FND-001 | git grep on main: none of the four exists under src/main; PriceModel exists (FS-2 FR-C2). No slice owned them and no spec recorded the deferral. FS-11’s FileScanProvider / RecordVisibilityCheck behind SeamGuard (fail-closed in cloud shapes) is the pattern each should follow |
| 6 Tokens render one screen in Empire’s theme and a neutral theme | ✅ | FS-3 AC-3; re-proven at FS-14 AC-12 (one import swapped; --et-radius-sm 9px → 3px; 0 fonts loaded under neutral) |
7 data-testid + role-query conventions demonstrated in the first Playwright smoke test |
◐ role queries yes; a data-testid from Playwright not yet |
Both a11y.spec.ts files use getByRole only — the shell has no element that needs an id; the check:testids gate exists in both frontends (FS-3). Closes with the first feature verify-WI that needs one |
8 CI wall-clock < 10 min; every job declares timeout-minutes |
✅ with a trend → R-FND-004 | All six jobs declare timeout-minutes: 10. Backend Slow lane measured 3m26s (FS-2), 4m03s (FS-4), 4m49s (FS-10), 5m37s (FS-6) — about 17% growth on FS-6 alone |
| 9 Every §7 tool wired and proven red-green once | ✅ | FS-4 NFR-7 (every gate red-greened; Knip advisory by design); the surviving guards re-proven at FS-14 AC-5/AC-7/AC-11/AC-15 |
| 10 §0.1 honest failure: a reachable fixture fails the build; a fixture string in an artifact fails CI; the first real list screen renders all six §3.2 states | ✅ ✅ ❌ → R-FND-002 | FS-1 AC-7, FS-3 AC-8/AC-9, FS-14 AC-11 (ArchUnit + dependency-cruiser, two hops); FS-4 AC-16 (jar and bundle, nested jars included). No list screen exists — FS-3 Q-5 deferred the Table and app frame to the first screen-shipping feature / F32 — so the third clause is unprovable at foundation close |
| 11 Identity allocation per §2 from the first entity | ✅ on the substrate; the first production series is F26’s | FS-2 AC-10 (GENERATED ALWAYS AS IDENTITY rejects a supplied key; OVERRIDING SYSTEM VALUE is the deliberate escape), AC-11 (the allocator under two concurrent issuers with a gap accepted — on a test-fixture sequence; no CREATE SEQUENCE ships yet), AC-12 (typed duplicate → FS-10’s 409 DUPLICATE_VALUE), AC-13 (classified grep baseline; main shows no derivation) |
Summary: 7 pass · 1 pass with a cloud-gated half (item 4) · 1 pass with a trend accepted as a risk (item 8) · 1 partial (item 7) · 2 fail, both accepted as risks (item 5; item 10’s third clause) · item 2’s onramp half is FS-8’s by construction.
Hand-off ledger — every “owed to X” across the eleven plans and tracker notes, reconciled:
| Hand-off | Finding | Disposition (lead, 2026-09-15) |
|---|---|---|
FS-10: the OpenAPI listPrice nullability fix “owed before the first FE generator run” |
Already discharged — FS-6 plan Decision 25 shipped platform.web.NullableRecordComponentConverter, on main; the FS-10 tracker note still said owed |
FS-10 row note corrected |
FS-7 ↔ FS-6: the SYSTEM audit principal for job-originated rows |
Unowned; no SYSTEM principal in the auth package |
Stays on the FS-7 row note; owner still to name |
| FS-1 FR-F9 → FS-5: the frontend error reporter | Unowned; forge-et-fe/src/app/telemetry.ts says “FS-5 supplies the implementation” while FS-5 OQ-1 declined it — a dangling pointer in code |
FS-5 row note; the code comment re-points when the cloud half is specced |
FS-9: check-no-scheduled.py and @EnableScheduling |
Carried in the script’s header; FS-6’s Spring Session shape is now on main |
Already on the FS-9 row note |
FS-10 F-006: the huge-page 500 and the body TYPE_MISMATCH mapping |
Two spec Revs owed, both ASK FIRST | Already on the FS-10 row note |
FS-10 subtask 6b: the compose.yaml Postgres bind mount on Docker Desktop VirtioFS |
Unowned; a fresh clone could not start the local database as shipped | Onramp material → the FS-8 spec (draft body amended at this review — FR-R6) |
FS-11: the ET_FILES_STORAGE_BUCKET CI env key; OQ-P1 / OQ-P2 |
The key is absent from the CI env block; both confirmations are owed by the lead | FS-11 row note; the lead’s call |
FS-3: the pnpm check:tokens gate; the icon-registry convention in CLAUDE.md |
Open lead calls from the FS-3 plan’s subtasks 12/13 | FS-3 row note |
| FS-3 F-014: no repo home for slice-level visual evidence | Unowned harness call — a lessons entry or a rule | FS-3 row note; the lessons-vs-rule choice stays open |
FS-2 PriceModelSpec’s “double wins” half |
Decorative since FS-12’s cancellation — it asserts against its own stub | FS-2 row note |
Owned and unchanged, for the record: first-administrator provisioning (F5 + the FS-1 cloud half, D2); the Node pin (FS-8, D4); the FR-T6 opt-in build check (the first shipper of a temporal table — F26); F4’s OVERRIDING SYSTEM VALUE + setval; the first own/team ownership context (F1/F2); the Unassigned queue (F11); the real RecordVisibilityCheck and the attachment type allow-list (the first attaching features); Modal overflow and toast-over-modal (the first real modal screen); the Table and app frame (F32 / the first screen-shipping feature); the §2.14 alerts and market.is_active (F26); the coverage + Trivy gate flip (Phase 1→2); OQ-53(b), FS-6 Q-3, OQ-56 (client).
Advisory:
- The FS-8 draft contradicted the record on operating systems. It put Windows support out of scope “because the team is on macOS”; FS-1 spec Rev 2 says two of the three developers work on Windows, and the FS-3 F-001/F-002 and FS-6 subtask-1 records are Windows sessions. Corrected in the draft body at this review as FR-R6 and AC-14 (the spec is
Status: draft, so no## Revisionsentry — the peer read sees the amended text). - FS-14 bookkeeping: plan subtasks 13 and 14 are unticked though the cascade and Reflect landed; the spec’s Appendix B annotation for the two
et.migration-conventionsrules retired without replacement (dotted sub-versions,R__underdb/migration) was absent; Appendix A still showedmodel/undercatalog/. Fixed in the review’s follow-up harness PR: both boxes ticked, both appendices annotated. - Substrate load: 94 spec files, 35
@SpringBootTestannotations, 18@DynamicPropertySourcecontexts in the backend test tree; the test Postgres already raised tomax_connections=300(L-030). An input to Gate 3 work-item sizing (R-FND-004). - Process: the
/forge-test-verifygate was never runnable for a foundation slice and four slices bypassed it by precedent (FS-11 Reflect Q3); the skill change is owed its own PR. - Trivy:
.trivyignore.yamlis empty in all three repos; the known findings are report-only and pre-existing (Tomcat, excluded above;fast-uri/js-yamlin the commitlint tooling lockfile). - The frontend README still says “Node 22+” against CI’s 24 — already in the FS-8 draft (FR-R4).
Risks accepted:
| ID | Owner | Reasoning |
|---|---|---|
| R-FND-001 | Tharinda | §6 item 5 — the four provider seams do not exist at foundation close. Accepted under AD #29’s first-consumer rule: AIProvider is built by the first AI feature (F11/F12), MailboxProvider by F20, ErpHandoffService by F19, AuctionPriceProvider by F26 — each with a local/test stub and a fail-closed cloud guard on the FS-11 SeamGuard pattern. Gate 3 must make that assignment explicit in each feature’s spec so no feature assumes the seam is already there. |
| R-FND-002 | Tharinda | §6 item 10, third clause — the six §3.2 list-screen states are unproven because no list screen exists (FS-3 Q-5). Owed by F32 or the first screen-shipping feature, whichever lands first; its spec must carry the six-state acceptance criterion (empty vs empty-filtered; refresh-failed keeping its data with an as-of label). |
| R-FND-003 | Tharinda | The cloud-gated halves remain unwritten specs — the FS-1 Terraform bootstrap, the FS-5/FS-7/FS-9 alarms and real queues/schedules, FS-11’s GuardDuty enablement, FS-6’s cloud login — all gated on the AWS account Altrium now runs (AD-O-10 resolved). §6 item 4’s “end-to-end in dev” and the DLQ alarm cannot be proven until then. |
| R-FND-004 | Tharinda | Slow-lane headroom. The backend Slow lane grew from 3m26s to 5m37s over the foundation against a 10-minute cap that is a required-check contract; each @DynamicPropertySource IT adds a cached context. Gate 3 sizes work items with this in view; the lever when it runs out is fewer container-backed specs, never a raised cap. |
Consequences recorded elsewhere: tracker.yaml setup.foundation.review_completed / review_audit_ref; the slice row notes named above; the FS-8 draft body (FR-R6, AC-14, Input Sources); coding-standards.md §6 items 5 and 10 annotated with the accepted-risk reading; .claude/CLAUDE.md § Active Context. Gate 3 (/forge-decompose) may now run — this note is its foundation precondition; FS-8’s approval runs beside it (D3).
Addendum — 2026-09-16: one slice added after this review. PRD Rev 27 (2026-09-15) removed the products catalog, and the backend’s substrate stand-in entity — product (FS-2 FR-C1), FS-10’s reference endpoint, FS-6’s audit-proof fixture and product.* keys — contradicts it. FS-15 (retire the product catalog) was added to the backlog on 2026-09-16 with a drafted spec and plan (.forge/specs/foundation/015-retire-product-catalog-spec.md). It is not covered by this review’s verdict: it takes its own lead approval (Tharinda — the data-model.md §2 half is his under AD #31), and its acceptance criteria re-prove the FS-10 and FS-6 substrate criteria it re-homes. Gate 3 may run before FS-15 ships — the slice changes no substrate guarantee, only which entity carries the proofs — but F26’s spec must be written against the revised data-model.md §2, so PR A of FS-15 precedes the F26 spec.
Addendum — 2026-09-16: foundation sealed. Every slice reached an exit state on 2026-09-16 — FS-15 (forge-et-be#29 + harness#102/#103, reflected via harness#105) and the closing FS-8 (forge-et-fe#16, forge-et-pwa#17, forge-et-be#28, harness#98/#106, reflected via harness#109/#110) closed the list; FS-12/FS-13 stay dropped (AD #29). Tharinda (lead) confirmed the seal the same day: tracker.yaml setup.foundation.status: done, setup.decomposition unblocked. The accepted risks R-FND-001…004 above stand unchanged; the cloud-gated halves remain owed to the AWS account. Gate 3 (/forge-decompose) may run.
Gate 2 Run 1 — 2026-07-13 (Dinil)
Mode: full Outcome: pass-with-risks-and-spikes Trigger: Gate 1 passed-with-risks (2026-07-07); architecture & feasibility probe before decomposition (Gate 3)
Sync-check: architecture.md and CLAUDE.md § Architecture Decisions were both blank templates prior to this run and were authored together in this session — no contradictions to reconcile.
| Section | Item | Status | Reasoning |
|---|---|---|---|
| Tech stack viability | Stack chosen and justified against PRD constraints | ✅ | Java 21/Spring Boot, TypeScript/React/Next.js, PostgreSQL (RDS), AWS — each tied to a PRD constraint (custom build, Claude-direct, relational-preferred, web-first). |
| Tech stack viability | Mandatory tech constraints from PRD honored | ✅ | Claude called directly from backend (AIProvider), relational DB, responsive web-first, Google/Microsoft email-calendar-contact connectivity all architected. |
| Tech stack viability | Stack capability gaps named | ✅ | Exchange-EWS/raw-IMAP named as materially harder than Gmail API/Microsoft Graph — scoped as SP-ARCH-002. |
| Foundational decisions | Data model approach decided | ✅ | Relational, single-tenant; admin custom fields (F34) as JSONB, not dynamic DDL. |
| Foundational decisions | Auth model decided (provider, session/token, RBAC) | ✅ | Dual-path OAuth2 (Google/Microsoft) + local email/password onto one User; Spring Session JDBC (not JWT); RBAC on the PRD’s 4 roles. |
| Foundational decisions | Tenancy model decided | ✅ | Single-tenant, per PRD §Multi-Tenancy — carried forward, not re-litigated. |
| Foundational decisions | Sync vs. async boundaries identified | ✅ | Interactive request/response path vs. SQS-backed worker path (mail/calendar sync, batch AI triage, migration, scheduled alerts) — architecture.md § Data Flow. |
| Foundational decisions | Integration patterns for external systems decided | ✅ | Push+poll for mailboxes, synchronous request/response for Claude, manual-document seam (ErpHandoffService) for the ERP fork. |
| Build feasibility | Highest-risk PRD requirements identified | ✅ | 6-row table in architecture.md § Build Feasibility & High-Risk Requirements. |
| Build feasibility | Paper sketch per high-risk item | ✅ | Every row carries a sketch (email-sync matching engine, AIProvider seam, ErpHandoffService seam, DocRule table, stable stage_type, JSONB custom fields). |
| Build feasibility | Spike work scoped where a sketch isn’t enough | ✅ pass-with-spike | SP-ARCH-001 (self-hosted AI) and SP-ARCH-002 (EWS/IMAP) — see below. |
| Resource and timeline reality | Team capacity vs. PRD scope sanity-checked | ❌ | Not established this session — no prior data available to this architect walk. |
| Resource and timeline reality | Skills gap identified | ❌ | Spring Boot confirmed as the team’s existing pattern; React/Next.js and AWS familiarity unconfirmed. |
| Resource and timeline reality | Critical-path estimate vs. delivery window | ❌ | No delivery window supplied yet. |
| In-house-first audit | External dependencies enumerated | ✅ | 5-row table: Claude API, Google/Microsoft OAuth+APIs, AWS managed services, OWASP Dependency-Check, discarded no-code/marketplace platforms. |
| In-house-first audit | Each dependency justified vs. in-house alternative | ✅ | |
| In-house-first audit | Rationale recorded for each external choice | ✅ | |
| Security & supply chain | Dependency-vulnerability CI gate planned | ✅ | OWASP Dependency-Check, CVSS ≥ 7.0 fails build, per-repo suppression file with justification+expiry (Constitution Decision #9). |
| Honesty | Architectural unknowns/TBDs explicitly listed | ✅ | architecture.md § Resource & Timeline Reality states the gaps rather than fabricating numbers. |
| Honesty | Reversible vs. irreversible decisions distinguished | ✅ | CLAUDE.md’s “DO NOT REVERSE” table is the irreversible set; everything else in architecture.md (e.g. ECS-vs-alternative FE hosting, CI platform) is explicitly draft/TBD. |
| Honesty | Decisions deferred to later phases noted with trigger | ✅ pass-with-spike | Same two spikes, each with an explicit trigger condition. |
Summary: 18 pass (2 of which pass-with-spike) · 3 fail (all in Resource & Timeline Reality, accepted as R-ARCH-001) · 0 N/A.
Advisory:
<backend-repo>and<frontend-repo>do not exist yet — this is expected at Gate 2; their creation is Foundation slice 1 inarchitecture.md § Foundation Backlog.- No new architectural open questions were surfaced beyond what’s already tracked — the EWS/IMAP contingency (SP-ARCH-002’s trigger) is already covered by the existing OQ-24 in
project-prd-signals.md; no duplicate OQ added.
Risks accepted:
| ID | Owner | Reasoning |
|---|---|---|
| R-ARCH-001 | Dinil | Team capacity/composition, skills-fit beyond the confirmed Spring Boot pattern, and delivery window were not established in this session (checklist: Resource and timeline reality — all 3 sub-items). Accepted to avoid blocking Gate 2 on data unavailable to this architect walk; revisit before Gate 3 decomposition sequences wave/timeline planning. |
Spikes scoped:
| ID | Owner | Description | Trigger |
|---|---|---|---|
| SP-ARCH-001 | Dinil | Self-hosted AI model integration path (model choice, compute sizing, latency/cost budget) — architecture assumes the Claude API behind an AIProvider seam today. |
Revisit if OQ-15 resolves to “no external AI permitted.” |
| SP-ARCH-002 | Dinil | Exchange-EWS / raw-IMAP mailbox sync mechanics — architecture designs first-class support for Gmail API + Microsoft Graph API only. | Revisit when OQ-24 confirms Exchange-on-prem or generic IMAP accounts are actually in use (vs. Gmail/Microsoft 365 covering all reps). |
Gate 2 Run 2 — 2026-08-06 (Dinil)
Outcome: ⚠️ passed-with-risks-and-spikes. Verdict given by Dinil.
Why this run happened. architecture.md was found ~3 weeks stale against project-prd.md (last updated 2026-08-03). Two client decisions and a whole confirmed subsystem had landed after Run 1 and were unreflected. The document was revised the same day (Rev 1 / 1a / 1b) and re-walked. A dry-run walk preceded this full run; findings were reconciled before the verdict.
What changed in the artifact since Run 1
| Area | Change |
|---|---|
| Mailbox scope | Narrowed to Microsoft 365 via Microsoft Graph only. Confirmed against project-proposal.md — already issued — which names Microsoft 365 exclusively in six places. Gmail/Exchange/IMAP were never quoted; out of scope, additive only via change request behind a new MailboxProvider seam. Drove PRD Rev 7 (five locations). |
| AI provider | OQ-15 accepted 2026-07-28 — cloud Claude approved, pending contract-stage formal approval. AIProvider seam retained as safeguard. |
| ERP | Current ERP named as IFS; automated integration is a separately-priced follow-on. |
| Weekly Bulk Pricing | New section. Confirmed in the PRD 2026-08-03, after Run 1, and previously unmodelled: effective-dated prices/costs, (standard × market × week) key, draft→published workflow, shared price/quote expiry clock, AuctionPriceProvider seam. |
| Scheduling | EventBridge Scheduler → SQS for all timed work; in-process @Scheduled explicitly rejected (fires duplicates across ECS tasks). |
| Idempotency | Idempotency keys required on every queued job — SQS standard is at-least-once. |
| AuthZ | Six roles (was four) plus a (role, resource, permission-level) model per OQ-47. |
| Added treatment | CSRF/session hardening · shared-mailbox modelling · Graph subscription renewal + silent-failure alarms · AI resilience and cost control · search · API conventions · virus scanning · backup/recovery · data retention. |
| Foundation Backlog | 8 slices → 11 (scheduler substrate, API conventions, file pipeline); slices 2/5/6/7 gained scope. |
| Constitution | AD rows 12–16 promoted to CLAUDE.md § Architecture Decisions. |
| New | ## Open Decisions — 8 architectural choices (AD-O-1…8), each with a settle-by point. |
Checklist walk: 19 pass (2 carrying flagged gaps) · 2 fail · 0 N/A.
The 2 failures are unchanged from Run 1 and both sit in Resource & Timeline Reality — team capacity vs. scope, and critical-path estimate vs. delivery window. Both remain blocked on business information unavailable to an architecture walk, and both are already covered by R-ARCH-001, carried forward. Run 1’s third failure in this category (skills gap) now passes — Spring Boot is confirmed as the team’s pattern, with React/Next.js and AWS familiarity explicitly flagged as to-be-confirmed rather than assumed.
Sync-check contradictions found, and their resolution — all reconciled before this verdict:
| # | Contradiction | Resolution |
|---|---|---|
| A | CLAUDE.md § Project Overview listed 4 user roles; PRD confirms 6 |
Fixed in CLAUDE.md (commit f90a2a2) |
| B | CLAUDE.md § Stack omitted EventBridge Scheduler (a locked requirement) |
Added; SES correctly left marked “proposed” (AD-O-3) |
| C | CLAUDE.md AD table had 11 rows; architecture proposed 12–16 |
Rows 12–16 promoted |
| D | CLAUDE.md § Active Context stale — both spikes shown active, three client requirements shown “pending capture” |
Rewritten; all three are in the PRD body |
| E | tracker.yaml spikes still showed SP-ARCH-001/002 open |
Resolved by this run’s writes |
| F | PRD §Users and Access confirms 6 roles but the Per-Role Capability Matrix has only 4 columns | The gap was already documented in a matrix note, but its completion trigger read “at spec time” — too late, since Foundation slice 6 builds the authorization model and runs before any feature spec. Trigger moved to “before Foundation slice 6.” Completing the two columns remains an open task. |
Advisory:
.forge/checklists/architecture-checklist.mdis still the uncalibrated template (# [Project Name], 21 generic items). Run 1 also walked it in this state. Calibrating it for this engagement is recommended before Gate 3.<backend-repo>/<frontend-repo>still do not exist — expected pre-Foundation.<backend-repo>/docs/data-model.mdis now a hard prerequisite for the F26 pricing spec (see SP-ARCH-003).- No new architectural OQs surfaced. OQ-9 (product-held certification storage) remains open in signals and gates F9.
.forge/design/archive/core-reuse-feasibility.mdwas authored alongside this run as a standalone, unapplied analysis of isolating a reusable CRM core from the tea domain. It is deliberately not part of this gate — nothing in it is adopted, and it is blocked on an IP/contract carve-out question, R-ARCH-001, and a funding decision.
Risks accepted:
| ID | Owner | Reasoning |
|---|---|---|
| R-ARCH-001 | Dinil | Carried unchanged from Run 1. Team capacity/composition and delivery window still not established (checklist: Resource and timeline reality — 2 of 3 sub-items; the skills-gap sub-item now passes). Accepted again to avoid blocking on data unavailable to an architecture walk. Must resolve before Gate 3 — it gates wave sequencing and any date-setting via /forge-set-dates. |
Addendum — same day, after the verdict (2026-08-06). Both inputs behind R-ARCH-001 were established within hours of this run closing, so the risk has been restated rather than closed (see
tracker.yaml setup.architecture.accepted_risks). The verdict above stands as given; this note records what changed:
- Team confirmed: 2 developers.
- The delivery window was never missing.
project-proposal.md § 7— already issued to the client — carries a full indicative timeline (~11–13 weeks feature-complete, ~3.5–4 months signed-off) plus a day-for-day slippage clause, and § 6 defines the delivery phases. Both this audit andarchitecture.mdhad recorded it as “not supplied.” Nobody had read the proposal. This is the third finding of that class in one day (the others: mailbox scope, and three “open decisions” thatA.5/A.6.3/A.2already specified)./forge-estimaterun: ~58–74 wall-clock cycle-days for 11 Foundation slices + 30 features vs. a ~55–65 working-day budget. Build fits on paper; headroom thin because the estimate excludes queue time, and the queue is already enumerated (~14 OQs on the Pipedrive audit, OQ-9). Basis: org baselinesv0.34.0, zero project samples.- Follow-on artifact changes:
architecture.md § Resource & Timeline Realityrewritten; AD-O-1, AD-O-5, AD-O-6 closed as already-answered by the proposal (frontend on Fargate; backup RPO/RTO and retention fully specified); AD-O-3 strengthened; the Multi-AZ-prod claim corrected to the single-AZ costed baseline with Multi-AZ as a priced add-on; AI cost controls replaced with the committed$50/mo, alert at $150figures; committed security baseline and AI data-handling boundary sections added fromA.4/A.6; and the Phase-1b provider comparison recorded as a delivery commitment, which means theAIProviderseam will genuinely be exercised with a second implementation rather than merely held in reserve.- New signal: OQ-48 — the trade-show capture PWA is in the issued proposal but absent from the PRD and architecture, and the proposal contradicts itself on whether it is in V1. Parked by the developer; recorded so it is not lost.
Process lesson for the next gate:
project-proposal.mdis a binding client artifact and was not in this gate’s input list. It should be. Three separate “unknowns” were answered inside it.
Second addendum — verification dry-run (2026-08-06). After the first addendum’s fix landed (
project-proposal.mdadded to/forge-arch-probe’s step-3 inputs), a/forge-arch-probe dry-runwas run to verify it. No verdict was taken and nothing was written by the run itself — this note records it because its findings were subsequently applied toarchitecture.md(Rev 3) and the PRD, and the audit trail should explain why those documents changed.Fix verified. Step 3 read the proposal in full; step 4 surfaced genuine proposal-vs-internal contradictions. The pre-fix walk, hours earlier, had completed clean and reported none. This is the verification step specified in
.forge/lessons.mdL-001.Checklist delta vs. the Run 2 walk: the two Run 2 failures — team capacity vs. scope and critical-path estimate vs. delivery window — both now read as pass, since both inputs are known (2 developers; the window was in proposal §7 all along). Honesty — architectural unknowns listed read as incomplete, because two unknowns were absent from
## Open Decisionsand one of them blocks the immediate next task. That item is now resolved by the Rev 3 additions below.Four findings, all applied (commit
b04d499):
# Finding Applied as 1 Environment count. architecture.md, its CI/CD line, Foundation slice 1 and the PRD hosting NFR all assumed three environments. ProposalA.1funds two (development + production, development doubling as UAT and scheduled off outside hours);A.2.3prices separate staging as an unpurchased +$45/mo add-on;A.2.4frames three environments as the rejected ~$520/mo option.Corrected in all four places; CI/CD now promotes dev → prod directly. Same failure shape as the Multi-AZ correction in the first addendum — infrastructure scoped that the client had not bought. Two instances in one day. 2 Phase assignments. F23 (import/export) sat in P0 and F19 (ERP handoff) in P1a; proposal § 6 places both in Phase 1c, and §§ 6/17 tie milestone acceptance and payment to those boundaries. Both moved in the PRD phasing table, with a reconciliation note recording that the import engine is still built in P0 for F4, and that deals can reach Won before F19 exists. 3 AD-O-7too narrow. Proposal § 8 commits a parallel run with Pipedrive as system of record, then a freeze-window final migration — so F4 is an initial load plus a delta reconciliation pass, not one import.AD-O-7widened to cover delta reconciliation, the held-open rollback position, and the verifiable migration report.4 Two untracked unknowns. AWS region (10–25% cost variance, A.1; tied to EU data-protection, § 12) and who holds and pays the AWS account (§ 12/§ 17, “to be confirmed”). Neither appeared in## Open Decisions.Added as AD-O-9 and AD-O-10. AD-O-10 blocks Foundation slice 1 outright — the AWS accounts cannot be created without it — and slice 1, CLAUDE.mdand the tracker now all say so.New signal: OQ-49 — the proposal contradicts itself on environment count (§ 11 Deliverables promises three at handover; five places in Appendix A fund two). Recorded in
project-prd-signals.md. This one is a client conversation, not a document edit;architecture.mdfollows the costed appendix until it is settled.Open decisions after Rev 3: 7 active (AD-O-2, 3, 4, 7, 8, 9, 10), 3 closed (AD-O-1, 5, 6). (An earlier count of “8 → 4” was arithmetic error — closing 3 of 8 leaves 5.)
The Gate 2 Run 2 verdict above is unchanged. No Run 3 was opened. If the developer wants these findings to carry a formal verdict of their own, that is a full-mode run and a separate decision.
Spikes scoped:
| ID | Owner | Description | Trigger |
|---|---|---|---|
| SP-ARCH-003 | Dinil | Bulk-pricing data model. Reconcile .forge/discovery/research/bulk-tea-pricing.md into <backend-repo>/docs/data-model.md and settle the effective-dating scheme, the (standard × market × week) price key, and the publish/expiry state transitions. The largest unmodelled area in the architecture and the client’s stated priority. |
Before the F26 spec is written, and before Foundation slice 2 fixes the temporal-row conventions. |
Spikes stood down:
| ID | Why |
|---|---|
| SP-ARCH-001 | OQ-15 accepted 2026-07-28 — no on-prem/self-hosted LLM; AI runs on cloud Claude, subject to formal client approval at contract stage. The trigger condition (“if OQ-15 resolves to no external AI permitted”) did not occur. The AIProvider seam is retained as a safeguard should contract-stage approval fail — if it does, re-scope this spike. |
| SP-ARCH-002 | OQ-24 resolved — Microsoft 365 is the sole provider in use, and the issued proposal quoted only Microsoft Graph. The trigger condition (“when OQ-24 confirms Exchange-on-prem or generic IMAP in use”) did not occur. Re-scope only if a Gmail/EWS/IMAP change request is accepted. |
Spikes resolved:
Appended 2026-08-21. The “Spikes scoped” table above is the Run 2 record and is left as written; this block records what happened to a spike afterwards. A resolved spike is distinct from a stood-down one — here the trigger fired and the question was answered.
| ID | Resolved | By | Output | Outcome |
|---|---|---|---|---|
| SP-ARCH-003 | 2026-08-21 | Dinil | forge-et-be/docs/data-model.md §2 + docs/proofs/pricing/ (forge-et-be#6) |
Answered. Price grain (buying_standard × market × validity); cost at its own grain (buying_standard × validity) — it has no market axis; validity anchored to a numbered tea.auction_sale row rather than a calendar week, because auction sale days shift; publication as a draft → published transition enforced by a partial exclusion constraint. PostgreSQL 17 pinned. DDL verified on 17.11 — an assertion suite plus the AD #17 core-alone acceptance test, gated by a bidirectional doc-vs-scripts drift check — rather than asserted on paper. Review tightened three doc-vs-schema gaps before merge (unwired standard_no allocator; an unguarded denormalisation on the quote snapshot; a nullable cost_at_quote that let a line reach the OQ-40 margin gate with nothing to evaluate), each now enforced in the DDL with its own assertion. |
What the spike found that the paper design had not. Four reconciliations against
bulk-tea-pricing.md (2026-07-16, predating the 2026-08-03 pricing answers), whose §7 DDL is superseded:
the priced unit is a Buying Standard not a Product (OQ-42); currency and Incoterm leave the price key
(OQ-43/44); cost moves off the market-keyed price row; and — the load-bearing one — the quote-line price
provenance FK would have been a core -> tea foreign key, which AD #17 forbids and CI rejects. That last
one was latent and would have surfaced as a red CI check partway through F26 implementation, with data
already in the schema. It makes LineItem a fifth typed 1:1 extension entity (architecture Rev 22).
(FS-14 note, 2026-09-10: AD #31 later removed the core → tea FK check and collapsed the two schemas into one public schema, so LineItem’s provenance FK is now an ordinary intra-schema FK. This spike record stands as what was true in August; the typed-extension entity shape survives as data-model design.)
Consequences recorded elsewhere: architecture Rev 22; tracker.yaml setup.architecture.resolved_spikes
(a new list — a resolved spike had no home in the schema, since stood_down_spikes means the trigger never
fired); FS-2 unblocked with its substrate scope enumerated in the data model §2.13. No Gate 2 re-run is
implied — resolving a scoped spike is the spike doing its job, not a change to the Run 2 verdict.
Gate 1 Run 1 — 2026-07-07 (Dinil)
Mode: full Outcome: pass-with-risks Trigger: End of discovery — initial PRD readiness check before architecture (Gate 2)
| Section | Item | Status | Reasoning |
|---|---|---|---|
| Scope and boundaries | In-scope features explicitly listed | ✅ | §Scope > In Scope (V1) enumerates F1–F26 with P0/P1/P2 priorities, grouped by theme. |
| Scope and boundaries | Out-of-scope + negative-space, each with won’t-build reason | ✅ | §Scope > Out of Scope lists 9 exclusions (ERP, estate ops, auction/grading AI, native mobile, generic automation builder, integration marketplace, marketing suite, SLTB/ASYCUDA submission, autonomous AI), each with a Won’t-build reason distinct from a deferral. |
| Scope and boundaries | Deferred separated, tagged with phase/trigger | ✅ | §Scope > Deferred is a distinct table with a “Target phase / trigger” column (ERP API sync, RAG, F27–F35, mobile, etc.). |
| Scope and boundaries | Phasing / sequencing with ordered priorities + dependencies | ✅ | §Phasing table (P0 → P1a/b/c → P2) plus per-feature priority and dependency rationale (“clean core data before AI”). |
| Scope and boundaries | Reference-system classification complete | ✅ | §Reference-System Classification assigns Replicate/Redesign/Defer/Discard to every Pipedrive capability, P-01…P-34. |
| Domain model | Key entities named | ✅ | §Domain Model > Key Entities names User, Account, Contact, Lead, Deal, Pipeline, Stage, Product Brief, Product, Sample, Quote, Activity, Email, Template, Sequence, etc. |
| Domain model | Entity relationships (1:1, 1:N, N:N) | ✅ | Key Relationships column states cardinalities (Account 1:N Contacts; Deal 1:1 ProductBrief; Deal/Quote N:N Products; Account/Product N:N Certifications). |
| Domain model | Lifecycle states named where applicable | ✅ | §Lifecycle States covers lead, both deal pipelines, 7-state sample, quote, activity, sequence enrollment, rotting, and ERP handoff. |
| Domain model | Glossary for client-specific terminology | ✅ | §Glossary defines 24+ tea-domain and CRM-parity terms (BOPF, CIF/FOB, Lion Logo, CUSDEC/ASYCUDA, rotting, stage probability, sequence, merge field). |
| Users and access | User roles enumerated | ✅ | §Users > Roles: Sales Rep, Sales Manager, R&D/Sample Owner, Administrator (candidate roles flagged for OQ-3). |
| Users and access | Per-role × per-resource capability breakdown | ✅ | §Per-Role Capability Matrix maps ~22 resources × 4 roles with explicit C/R/U/D/A/Cfg cells — not narrative. |
| Users and access | Multi-tenancy / org-hierarchy model | ✅ | §Multi-Tenancy: single-tenant, access by role + ownership + region tag; formal office hierarchy deferred (OQ-3). |
| Functional surface | Each in-scope feature ≥ one-paragraph description | ✅ | §Functional Surface has a paragraph for F1–F34. |
| Functional surface | User journeys for headline flows (end-to-end) | ✅ | 5 end-to-end journeys: A lead triage→opportunity, B private-label deal→Won→ERP, C quote readiness/approval, D Pipedrive migration, E email+follow-up. |
| Functional surface | Integration points with external systems named | ✅ | §Integration Points table: ERP, Pipedrive, Google/MS (email/calendar/contacts), Claude, inbound email, courier, LinkedIn, enrichment, doc storage, SLTB. |
| Non-functional | Performance expectations (order-of-magnitude) | ✅ | §NFR > Performance: 10–20 concurrent users, thousands of records, <2s pages, <10s sync AI, provider rate limits, batch migration. |
| Non-functional | Security / compliance requirements named | ✅ | §NFR > Security: RBAC + pricing-approval restriction, data classification, external-AI-privacy decision, HITL governance, mailbox/SPF-DKIM, marketing consent, file-upload validation. |
| Non-functional | Accessibility commitments named | ✅ | §NFR > Accessibility: WCAG 2.1 AA target, keyboard nav, no-color-only status. |
| Non-functional | Observability / audit requirements named | ✅ | §NFR > Observability: audit trail (stage moves, pricing/AI approvals), AI provenance + confidence, email/AI monitoring + token-cost tracking. |
| Constraints | Tech stack (mandatory vs preferred vs open) | ✅ | §Constraints > Tech Stack: custom-build + Claude-direct + email connectivity mandatory; Postgres preferred; frameworks/hosting/auth explicitly open (Gate 2). |
| Constraints | Regulatory constraints | ✅ | §Constraints > Regulatory: data protection for EU-buyer PII; SLTB/CUSDEC referenced-not-enforced; anti-spam if F28 promoted. |
| Constraints | Deployment / hosting constraints | ✅ | §Constraints > Deployment: cloud web app reaching Claude + email providers; region open pending AI-privacy; CI/CD + env separation. |
| Honesty | Open questions / TBDs explicitly listed | ✅ | project-prd-signals.md lists 30 OQs, each anchored to a PRD section with owner and Blocks features. |
| Honesty | Known risks / unknowns logged | ✅ | §Risks R-01…R-16, each with owner + status. |
| Honesty | Success criteria stated | ✅ | §Success Criteria: functional acceptance (Pipedrive parity + domain + AI + parity gaps) + business outcomes + adoption. |
| Honesty | Input sources / provenance recorded | ✅ | §Input Sources table + inline [prototype] / [research:…] / [Pipedrive doc research] tags on requirements throughout. |
Advisory:
- Derivative scaffolding not yet filled:
.claude/CLAUDE.md(§Project Overview, §Stack) and.forge/design/architecture.mdremain template stubs whileproject-prd.mdandtracker.yamlare populated. Expected pre-Gate-2 (filled by/forge-arch-probe); noted so it isn’t mistaken for drift.
Risks accepted:
| ID | Owner | Reasoning |
|---|---|---|
| R-PRD-001 | Dinil | ERP identity/API unconfirmed (OQ-2); F19 ships the manual handoff path, automated CRM↔ERP sync deferred until the ERP API is confirmed at Gate 2. |
| R-PRD-002 | Dinil | External-AI data-privacy decision pending (OQ-15); if the client forbids sending buyer content to the Claude API, the AI layer needs a self-hosted model — resolve before AI-layer build. |
/forge-deliver state reconciliation on 2026-09-18
Ticket: F5-a
Healed: features.F5-a.waves[1].status pr-open → merged (+ merged_at: null → 2026-09-18)
Evidence: both Wave-1 wave PRs merged — gh pr view 21 --repo altrium-io/forge-et-fe → state MERGED (mergedAt 2026-09-18T08:34:09Z); gh pr view 32 --repo altrium-io/forge-et-be → state MERGED (mergedAt 2026-09-18T08:34:05Z). All F5-a app-repo branches (wave + per-WI) already deleted on remote; no local worktrees — Stage 14 branch cleanup a no-op.
Operator: Tharinda Rajapaksha
/forge-deliver state reconciliation on 2026-09-19
Ticket: F5-a
Healed: features.F5-a.waves[2].status pr-open → merged (+ merged_at: null → 2026-09-18)
Evidence: Wave-2 (final, T-E2E) wave PR merged — gh pr view 25 --repo altrium-io/forge-et-fe → state MERGED (mergedAt 2026-09-18T19:56:56Z, base main). Per-WI PRs terminal (#24 e2e MERGED; #18/#19 sub, #31 be sub CLOSED as superseded; #20 verify MERGED). All F5-a app-repo branches already deleted on remote. Local cleanup still pending in forge-et-fe (main checkout on feature/F5-a-wave-2; local feature/F5-a-wave-2 + feature/F5-a-WI-2.1-e2e-signin-signout) — offered at Stage 14.
Note: wave PR #25 carries no recorded /forge-review-pr comment and no review/approval object — mandatory review gate to be confirmed before Reflect.
Operator: Tharinda Rajapaksha
/forge-deliver state reconciliation on 2026-09-22
Ticket: F20-a (invoked --wave 2)
Healed: nothing. Two divergences found and both left deliberately unhealed — see below.
Evidence:
-
features.F20-a.waves[1].workitems[F20-a-WI-1.1].impl_statusreadspr-open, butgh pr list --repo altrium-io/forge-et-be --head feature/F20-a-WI-1.1-mailbox-connection-core --state all→ PR #37 MERGED, its remote branch deleted on merge. Same shape forF20-a-WI-1.2: forge-et-fe PR #37 MERGED, branch deleted. Both merged directly tomainrather than through a wave PR, so nofeature/F20-a-wave-1was ever created and the actor that writes thepr-open → wave-closedterminal (forge-pr-open --wave) never ran. Stage 0’s rule table has no row for this shape. Not healed, on purpose: wave 1 is not complete —F20-a-WI-1.3is mid-flight — so flippingwaves[1].status → mergedwould be false, and the Wave-1 rows are being written right now by the parallel WI-1.3 session. Writing them from here would produce exactly thetracker.yamlconflict this ticket’s two live sessions are trying to avoid. Left to the WI-1.3 session to close. -
features.F20-a.waves[1].workitems[F20-a-WI-1.3].impl_statusreadsdispatched, with no remote branch and no PR (git ls-remote --heads origin feature/F20-a-WI-1.3-verify-connect-disconnect→ empty;gh pr list --head …→ empty). Stage 0’s table classes this as “heal → re-dispatchable: resetdispatched → pending”. Not healed, and the rule would have been wrong here: WI-1.3’s agent is alive in a parallel session (local worktreeworktrees/F20-a/F20-a-WI-1.3/forge-et-feat5261fd0, onfeature/F20-a-WI-1.3-verify-connect-disconnect). The no-heartbeat note states a dispatched agent is indistinguishable from a dead one, and decides on what reached the remote — but a live agent that has not pushed yet presents identically to a dead one that never pushed, and the auto-reset would have re-dispatched a second implementer onto a worktree already being written. The out-of-band knowledge that the session is running is the only thing that distinguished them. Framework finding, owed at F20-a’s Reflect: theno branch + no PR → auto-reset to pendingrow needs a local-worktree existence check before it heals, or the heal must be an ASK rather than an auto-write.
Operator: Dinil
/forge-deliver --wave override on 2026-09-22
Ticket: F20-a
Wave: 2
Deps unmet: wave-1:in-progress
Operator: Dinil
Decision: proceeded — Wave 2’s code dependency is satisfied even though the wave is not. WI-2.1’s depends_on is F20-a-WI-1.1, whose PR (forge-et-be #37) is merged to main, and Wave 2 branches from main, so every substrate it consumes (the two seams, mailbox_connection, MailboxConnectionService, MailboxScopes, MailboxProperties, CloudSeamFakes) is present at its base. What keeps wave 1 in-progress is F20-a-WI-1.3, a verify-WI that writes only forge-et-fe/e2e/**; Wave 2 is backend-only, so the two are file-disjoint — verified by diffing both Wave-2 plans’ ## Files to Modify against WI-1.3’s (zero intersection; WI-2.1 states “Nothing in forge-et-fe is touched”). Rebase risk accepted on that basis.
/forge-deliver state reconciliation on 2026-09-24
Ticket: F20-b
Healed: waves[1].status pr-open → merged; waves[1].merged_at null → 2026-09-24
Evidence: gh pr list shows forge-et-be#51 MERGED 2026-09-24T07:20:19Z, forge-et-fe#49 MERGED 2026-09-24T07:20:38Z, forge-et-harness#192 MERGED 2026-09-24T07:20:44Z; origin/main in both app repos is the wave-1 merge commit (be 994a696, fe ac8ce84).
Operator: Dinil
/forge-deliver Stage 7 Step 3 review skip on 2026-09-24
Ticket: F20-b
Wave: 2
Skipped: the per-WI forge-plan-review pass in Stage 7 Step 3 for WI-2.1, WI-2.2, WI-2.3 and WI-2.4.
Reason: all four plans were already reviewed and approved (> Status: approved), and merged to main in forge-et-harness#183; the Wave-2 contract is Status: frozen. The developer instructed that they not be re-reviewed.
Carried into dispatch without a plan re-review (recorded in each affected plan’s ## Notes): WI-2.1 takes the four ingest-under-load findings deferred from the be#51 review (cloud-gated-validation.md § B F20-b, “Deferred review findings”) and applies spec Rev 1 (out-of-scope ingest leaves no claim; reconnect re-runs the back-fill) to the delta poll. WI-2.4 keeps its planned fresh sales_rep principal (plan D3), confirmed by the developer.
Operator: Dinil
/forge-deliver state reconciliation on 2026-09-25
Ticket: F20-b
Healed: waves[2].status pr-open → merged; waves[2].merged_at null → 2026-09-25
Evidence: gh pr view shows forge-et-be#57 MERGED 2026-09-25T04:27:58Z, forge-et-fe#55 MERGED 2026-09-25T04:28:17Z, forge-et-harness#197 MERGED 2026-09-25T04:28:23Z; origin/main in both app repos is the wave-2 merge commit (be 41acdef, fe 17a2d4c).
Operator: Dinil
/forge-deliver state reconciliation on 2026-09-25
Ticket: F1-a
Healed: waves[3].status pr-open → merged; waves[3].merged_at null → 2026-09-25
Evidence: gh pr view shows forge-et-be#59 MERGED 2026-09-25T07:26:33Z (merge commit 03b641e) and forge-et-harness#201 MERGED 2026-09-25T07:26:48Z (88ab4ec); both wave branches were auto-deleted on merge.
Operator: Biseni Perera
/forge-deliver state reconciliation on 2026-09-28
Ticket: F5-i
Healed: waves[1].status pr-open → merged; waves[1].merged_at null → 2026-09-28
Evidence: forge-et-fe’s wave PR #69 is MERGED (2026-09-28T04:37:12Z, merge commit d1bfc11). forge-et-be and forge-et-harness, also in the wave’s touched_repos union, have no wave PR on feature/F5-i-wave-1, which Stage 14’s rule treats as a halt. Healed anyway on complete evidence: the lead merged Wave 1’s backend and harness changes directly as per-WI PRs, recorded as pr_role: ship — forge-et-be#86 (MERGED) plus its follow-up #87 (MERGED), and forge-et-harness#241 (MERGED). forge-et-be’s wave branch was identical to main and was deleted unopened. Nothing in Wave 1 is unmerged.
Operator: Tharinda Rajapaksha
/forge-deliver state reconciliation on 2026-09-28 — wave-model deviation
Ticket: F5-k
Wave: 1
Deviation: the two sub WIs’ per-WI review PRs were merged straight to main instead of being integrated into feature/F5-k-wave-1 and shipped by the wave PR. forge-et-fe#61 (WI-1.2, frontend) merged 2026-09-27T12:49:09Z (b058e2b); forge-et-be#79 (WI-1.1, backend) merged 2026-09-27T19:12:50Z (462a845). Only WI-1.3’s tests (forge-et-fe#64, merged into the wave branch) remain for the wave PR, forge-et-fe#67.
Consequences: (1) for about 6h23m main was split — the frontend gated navigation on the new keys and signed e2e in as the all-permissions user and crm01@empire.lk, while the backend still returned the old keys and created neither user; closed when #79 merged. (2) fe#61 merged without /forge-review-pr (mandatory on code PRs); a post-merge review is owed. (3) the backend wave branch carried nothing beyond main and was deleted, so wave 1 has one wave PR (frontend) instead of two.
Healed: waves[1].workitems[F5-k-WI-1.1, F5-k-WI-1.2, F5-k-WI-1.3].impl_status pr-open → wave-closed (terminal; the per-WI PRs are superseded — two by merging to main directly, one by the open wave PR).
Evidence: gh pr view on forge-et-fe#61, forge-et-be#79 and forge-et-fe#64 (all MERGED) and forge-et-fe#67 (OPEN, base main, head feature/F5-k-wave-1).
Operator: Tharinda Rajapaksha
Post-merge /forge-review-pr audit on 2026-09-28
Scope: every app-repo feature/* and fix/* PR merged to main since 2026-09-20 (the mandatory review rule has applied since 2026-09-18), checked for a recorded /forge-review-pr in its PR body or comments, bugs.md, the plans and the tracker notes. Per-WI PRs that merged into a wave branch are covered by that wave PR’s review.
Found without a recorded review (16): forge-et-be #43, #44, #48, #52, #86, #87; forge-et-fe #28, #30, #36, #40, #41, #43, #44, #45, #61, #65; forge-et-pwa none (pwa#25 was reviewed).
Method: post-merge review by backend-reviewer / frontend-reviewer, each finding judged against origin/main today, not only the PR’s diff; findings already fixed by later work listed as resolved.
Verdicts: be #52, #86 (code), #87 (code) clean; be #43, #44, #48 has-issues; fe #36, #40, #41, #43, #44, #61, #65 clean or nits only; fe #28, #30, #45 has-issues. No Blocker anywhere.
Fixed (each fix PR had its own /forge-review-pr, and its safe findings applied):
- forge-et-be#92 —
AccountControllerIT,ContactControllerITandLogHygieneITstop deleting whole shared tables and clean only their own rows (from #48; order-dependence on the shared test database); theContributedTokenRefreshGrantcomment corrected (from #44). - forge-et-fe#70 — tablet drawer focus management and Esc closing only the innermost layer (from #28; F5-b NFR-2); a nav choice closes the drawer (consistent with design-system §9 per the review); one no-horizontal-scroll helper that also measures
<main>and inner scrollers (from #30, #45 — the old checks could not fail for overflow inside<main>); an 800px drawer-keyboard browser row; weak e2e assertions tightened (from #28, #30, #61, #65). Live run: seam 33 passed / 5 skipped, chromium 132 passed. Deferred to F40 (Accessibility and responsive checks at every supported size): the 320px rows for the app frame and/email/mailbox(design-system Rev 28). Left for the lead (not applied — behaviour, design or CI decisions): - be #43 — the Graph subscription delete/create run inside the transaction holding the mailbox row lock; move them after commit, or record the risk in BUG-018.
- be #44 — classify a malformed Entra certificate PEM as a configuration fault instead of
TransientFailure. - be #86 / #87 — merged while the required Slow lane was red or pending (admin bypass);
main’s Slow lane last completed at69175b8. The Slow-lane runtime is being cut in forge-et-be#88 (another session). - fe #28 — the gold active/hover nav glyph and edge (2.2–2.75:1, below the 3:1 non-text minimum) and the “Sales CRM” eyebrow (10.5px gold text) — darken to
--et-accent-gold-deepor record the eyebrow as part of the logo lockup; tooltips on the icon-only rail (design-system §5 asks for them); the neutral count-badge token (4.45:1). - fe #65 — a CI check of the built bundle for stub strings; a dev-server guard when
VITE_ENTRA_LAPTOPandVITE_API_BASE_URLare both set (or document it inforge-et-be/docs/auth/laptop-sign-in.md). - fe#70 review — where focus lands after a nav choice (the toggle, as built, or the new page’s heading); the toggle does nothing visible at laptop width.
Noted, already planned:
DefaultPermissionCheckerletsteamreach unowned records; F5-f narrows unowned records toall(single record) and F11-b does the lead list. Operator: Tharinda Rajapaksha
Lead decisions on the items above, 2026-09-28 (Tharinda). Re-checked against main first: be#90, be#92 and fe#70 had merged, and nothing on main had changed any of the items.
- be #43 — keep the Graph calls inside the transaction and record the risk: moving them after commit would reopen the duplicate-subscription race BUG-018’s fix closes, and
GraphMailboxProvider’s 5 s / 15 s timeouts bound the lock hold. Recorded inbugs.mdBUG-018. - be #44 — fixed as a configuration fault:
signingKey()throwsEntraKeyMaterialExceptionand the refresh grant rethrows it instead of reportingTransientFailure(forge-et-be#95). The laptop-mode startup check (EntraLaptopGuard) catches a malformed pair at boot only inentra-laptop; the cloud profiles have no startup parse, so this path is the one a cloud environment reaches. - be #86 / #87 — bypass merges allowed only when the Slow lane was cancelled by its timeout after every step passed, recorded in the PR body, until forge-et-be#88 merges. Recorded in
.claude/rules/git-conventions.md § Branch Protection. - fe #28 — fix all three contrast items (nav glyph and edge ≥ 3:1, the eyebrow and the neutral badge ≥ 4.5:1, both themes) and add the rail tooltips, in one fe fix PR.
- fe #65 — build both: the bundle stub-string check as a new step in the existing CI job (a CI change the lead approved), and the dev-server guard for
VITE_ENTRA_LAPTOPwithVITE_API_BASE_URL. - fe#70 review — after a nav choice, focus moves to the new page’s main heading (Esc and the close button still return it to the toggle); the toggle is hidden at laptop width. Same fe fix PR as fe #28.
- F5-k wave 2 ships as one wave PR per repo, backend first — no per-WI PRs merged to
maindirectly. - F5-c’s decomposition plan, Wave-1 contract and Wave-1 plans are revised now, in parallel with F5-k wave 2: they still name the retired
user.view/user.createkeys and predate AD #37.
Follow-up decisions from the fix PRs’ own reviews, 2026-09-28 (Tharinda).
- The fix PRs forge-et-be#95 and forge-et-fe#71 each had
/forge-review-pr(pass with notes, and pass with nits), with their safe findings applied, and both merged. forge-et-be#95’s review also added the refusal of a mismatched key pair. - forge-et-be#95 merged while its Slow lane had been cancelled inside
Build, check and test. That does not meet the timeout-bypass rule, which needed every step to have passed first. The local./gradlew checkon the merged head was green (2622 tests, 0 failures). forge-et-be#88 merged the same day, which ended the timeout exception, so.claude/rules/git-conventions.md § Branch Protectionnow says never to bypass a required check to merge a delivery PR. - The focused page heading after a drawer choice shows no outline: a heading is not interactive.
- The Empire theme’s neutral badge text moves from
#7A6650to#77634D, taking the generic Badge on its tinted fill from 4.46:1 to 4.66:1. - The app-wide focus ring moves to
accent/gold-deep, after a check of every surface with focusable controls. - An expired or not-yet-valid Entra certificate is refused on every signing (forge-et-be#99). The heading, badge and focus-ring changes are forge-et-fe#73, recorded as design-system Rev 31.
Decisions on the reviews of forge-et-be#99 and forge-et-fe#73, 2026-09-28 (Tharinda).
- be#99, sign-in with an unusable key pair: redirect to the SPA with its own marker,
?login=unavailable, and log the fault once at ERROR, never a raw 500 onapi.*. The backend half is in forge-et-be#99, the message is forge-et-fe#75, and the rule is F5-a spec Rev 5 (FR-8, AC-10b). - fe#73, disabled-but-focusable controls: draw the full focus ring while focused. Their 0.55 opacity had faded it to about 2.1–2.4:1.
- fe#73, neutral Badge text: darkened again, to
#685641, so the Badge clears 4.5:1 on every surface it could sit on, not only a white card. - fe#73, drawer scrim: taken out of the Tab order. It was a keyboard-reachable button whose ring fell outside the viewport. This is forge-et-fe#74, separate because the defect predates fe#73. Without the scrim as the last Tab stop, the open drawer gains a Tab wrap, so that focus cannot leave a modal overlay.
- fe#73, the capture app: forge-et-pwa#28 ports the ring and neutral-text values. The capture app has no toast and no disabled-but-focusable control. Its theme file’s header claims its
:rootblock is byte-identical to forge-et-fe’s, which has not been true for some time (several tokens are missing). That is left for the lead. - fe#73, left as is: the form fields’ gold focus border inside the gold-deep ring.
Decisions on the reviews of forge-et-fe#73’s fixes, #74, #75 and forge-et-pwa#28, 2026-09-28 (Tharinda). All four passed with nits, and none had a Blocker.
- fe#74, closing the drawer: add a visible close button to the drawer’s header. With the scrim hidden from assistive tech, a touch screen-reader user on a tablet had no named control that closed the drawer.
- fe#73, the disabled look: it is the same at rest and on focus, and it uses no element opacity. Before, a focused disabled secondary button looked almost like an enabled one, and a disabled primary button changed colour when it was focused.
- forge-et-pwa, shared files: re-copy
tokens.cssandbreakpoints.tsfrom forge-et-femain(comment-only drift, whichcheck:sharedmisses because it compares recorded hashes), and correct the theme file’s header, in their own PR (forge-et-pwa#29). - The review nits: applied on each PR. The F38-e alarm is recorded in
cloud-gated-validation.md§B FS-6.
Post-merge /forge-review-pr backfill — 2026-09-30
Scope and method. An audit of every app-repo PR merged since the mandatory /forge-review-pr rule started (2026-09-18, harness commit 13c1fb31), Dependabot PRs excluded, checked for a review record in the PR body and comments, the plans, bugs.md, the tracker notes and unmerged branches. Each PR still without a pre-merge review was then reviewed by backend-reviewer / frontend-reviewer, and every finding was checked against current origin/main; only findings still present are actioned.
| Group | Count | PRs |
|---|---|---|
| Code PRs audited | 91 | — |
| Reviewed before merge | 64 | — |
| Ambiguous | 1 | fe#33 — a frontend-reviewer pass recorded in bugs.md, not a /forge-review-pr |
| Reviewed only after merge (mostly by the 2026-09-28 audit) | 16 | — |
| Skip written in the PR body, but no tracker note | 2 | be#85, pwa#20 |
| No review evidence at all | 8 | be#35, be#40, be#75, be#76, be#88; fe#21, fe#25, fe#87 |
App-repo docs / ci / chore PRs, listed as “exempt?” for the lead to decide |
24 | — |
Patterns:
- Per-WI PRs merged straight to
main, so no wave PR existed to review: be#37, fe#37, be#86, fe#61, fe#65. - Follow-up fix PRs that were not reviewed themselves: be#40, be#44, be#75, be#76, fe#45; and forge-et-be#124’s three late fix commits on 2026-09-30.
- Review records are spread across PR comments, PR bodies, plan notes,
bugs.mdand unmerged branches, and the PR checkboxes were often left unticked. - The 2026-09-28 audit started at 2026-09-20, so it missed fe#21 and fe#25; it also missed be#40, be#75 and be#76; and it listed be#48 and fe#44 as unrecorded, though both had records made before merge.
Results per PR.
| PR | Findings | Action |
|---|---|---|
| be#35 | Nit, superseded: V28 deleted the superadmin role that V12 renamed | None |
| be#40 | Important: the GCM associated-data binding in TokenCipher.java:105,148 has no test that fails if it is removed, because both test constructions use LocalTokenKeyProvider, which already refuses a foreign context. Nits: a circular “stable across re-consents” row in TokenEncryptionContextSpec.groovy:77; a stale AESWRAP-1 reference at KmsTokenKeyProvider.java:67 (from the later KMS change). Known and already registered, not new: an unreadable key answers 500 and leaves the connection ACTIVE with no reconnect prompt (cloud-gated-validation.md §B F20-a) — needs a lead decision |
Fixed in forge-et-be#131 (fix/post-merge-review-backfill); the 500 item left for the lead |
| be#75 | Nit: LeadResponse.java:46-69 keeps the default record toString, which prints the contact’s email, phone and name and the inquiry text — the NFR-3 exposure be#75 closed on four other records |
Same forge-et-be branch |
| be#76 | Nit: the comment claiming V27’s constraint “matches Java’s isBlank for ASCII” misses U+001C–U+001F | Comment corrected in LeadFieldRules.java on the same branch; the constraint is unchanged (not worth a migration) |
| be#88 | Its Slow-lane scope gating was superseded by 69cb4f8 / 6767007, and no test lost coverage. Nit: ScanResultQueueIT.groovy:497 parses the poll timeout in whole seconds only |
Same forge-et-be branch |
| fe#21 | Three Importants: BUG-052 (a failed sign-out shows the user as signed out), BUG-053 (no error screen on /login or the sign-in gate), BUG-054 (the return route is lost across the real redirect). Nits: the return path drops search and hash and is not validated; an empty 200 is not parsed as no body; the permission resolver is not memoised; absence-only unit tests with no positive check first |
Bugs filed; fixes and nits in forge-et-fe#101 (fix/BUG-052-sign-in-post-merge-fixes) |
| fe#25 | Nits: the AC-9 browser test never asserts that the token rotated (session-lifecycle.spec.ts:137); an adjacent F5-b absence check before the heading (:77-79) |
Same forge-et-fe branch |
| fe#33 | Nits: absence-only unit tests (LoginRoute.test.tsx:320-333); the cross-tab marker |
Tests fixed on the same forge-et-fe branch; the marker left as is |
| fe#87 | Clean | None |
| forge-et-be#124 (F5-k wave 2) | Merged 2026-09-30 before its three late fix commits were reviewed. Post-merge review: one Important — the lock is taken after the service’s first read, so a later transaction making several changes could deadlock — and nits | Fixed in forge-et-be#128, reviewed while the PR is open; that review’s one further Important, on the javadoc’s claim, was fixed in 3bddcf1. Deferred to F5-c / F5-d: the stronger options — a self-check that throws when a transaction has already written before taking the lock, locking in every write method, and a lock timeout |
For fe#21, findings already fixed by later work: the sign-in busy state (fe#57, BUG-027), the CSRF token before the first POST (fe#22), the CSRF cookie name (fe#59, F5-j) and the 320px login reflow (fe#62).
Lead decisions, 2026-09-30 (Tharinda).
- A failed sign-out keeps the user signed in and shows “Sign-out didn’t complete. Try again.”; a 401 from the logout request counts as signed out (BUG-052).
- The return route is carried across the redirect in
sessionStorage(BUG-054). - The
/loginerror screen offers a “Reload the page” action (BUG-053). - A service-unavailable notice shows when the first
/mefails with anything other than a 401. - Left as is: a sign-out in one tab shows “your session expired” in another tab.
Systemic fix. The review gate design, .forge/design/harness/forge-review-gate.md (forge-et-harness#328, in review), makes the review record a required check on app-repo PRs, so a merge without one is refused rather than found by a later audit.
Second pass, 2026-10-01.
- Scope: every app-repo PR merged since 2026-09-18 again (65 backend, 54 frontend, 19 capture app), checked for a review record in the PR, the harness
mainand every harness branch. Docs-only PRs (only*.md) are exempt, by the lead’s rule of 2026-10-01. - Merged since the first pass without a record:
- forge-et-be#127 (F4-a trade-fair rule);
- forge-et-pwa#44 (BUG-043; see the correction below);
- forge-et-be#130, a Dependabot bump with a person’s reformat commit added.
- The first pass counted as reviewed, but no record exists: be#34, be#39, be#61, fe#31, fe#35, fe#62, fe#81, pwa#25.
ci:/chore:PRs that changed code without a review: be#83, be#96, be#101, fe#60, fe#76, pwa#30. The others in that group changed only docs, comments or config.- Reviewed after merge by
backend-reviewerandfrontend-reviewer, each finding checked againstmain:- Clean or nits only: be#127, be#83, be#130 (the reformat is whitespace-only, proved by a whitespace-stripped diff), be#34, be#61 (since superseded by F5-l), be#96, fe#31, fe#35, fe#62, fe#81, pwa#25, fe#76, pwa#30, fe#60.
- be#96’s build cache: it cannot reuse a stale test pass for any repo file, because every file the tests read is a declared input. Environment variables, moving container tags and the clock are not tracked (Nit).
- be#101, I-1:
local-build-evidence.pychecked the tree and HEAD only before the build. Fixed in forge-et-be#133. - be#101, I-2: nothing in CI checks that a ready backend PR carries a build record. A skipped Slow lane counts as passing. This is a CI-policy decision for the lead.
- be#39, I-3: the BUG-016 token-cache invalidation clears one ECS task only. Filed as BUG-060, renumbered BUG-067 when
main’s BUG-058 to BUG-066 merged first. Lead decision, 2026-10-02: store the issued scopes in the cached token and refuse a cache hit that does not cover the scopes needed. - pwa#44: the fix is correct, but the error screens had no
<main>landmark. Fixed in forge-et-pwa#47. Correction, 2026-10-02: pwa#44 did have a pre-merge/forge-review-pr(frontend-reviewer, pass-with-nits). It was recorded only inbugs.mdon the BUG-043 fix’s harness branch, which this pass had not seen, and that review had already filed the missing landmark as BUG-060. This pass filed it again as BUG-058, which the merge withmainfolded into BUG-060.
- Reviewed PRs whose findings have no lasting home: almost every deferred Important was fixed later or is tracked. Still open:
- BUG-050 and BUG-051, open in
bugs.md. - The F20-b decisions owed since 2026-09-25 (tracker F20-b notes): a
backfill_generationcolumn with a CHECK constraint;$filteron the first delta call;SchedulerConfigGuardrefusingauto-startup=false; the lapse notice not announced in the/emaillive region; the lapse sentence usingtype/meta. - The F5-l WI-1.1 hand-off note: F5-k WI-3.1 removes the
allkey before a view key through the stub. - The be#128 deferrals and BUG-055/BUG-056, which this PR records.
- BUG-050 and BUG-051, open in
Lead decisions, 2026-10-02 (Tharinda).
- The review gate becomes a hard gate:
Forge reviewis made a required check in the three app repos. PRs that change only*.mdfiles pass it automatically, and a review record must name the reviewer agent for the repo (backend-reviewer,frontend-reviewer). Delivered through the review gate design’s rollout (forge-et-harness#328). - CI checks the backend build record (be#101 I-2): a cheap step fails a ready
forge-et-bePR whose local-build record is missing or names an older commit. - BUG-067’s fix: store the token’s issued scopes in the cache and refuse a cached token that does not cover the scopes needed.
- The F20-b items owed since 2026-09-25 are filed as BUG-068 to BUG-072, to be decided one by one when each is picked up.
Operator: Tharinda Rajapaksha
Third pass, 2026-10-02 — up to the required check.
- Why:
Forge reviewbecame a required check in the three app repos on 2026-10-02 (.forge/design/harness/forge-review-gate.md, rollout step 4). Everything merged before that could merge without a review record, so the window since the second pass was audited once more. - Scope: every PR merged into
mainin forge-et-be, forge-et-fe and forge-et-pwa from 2026-10-01 to the last merge before the check (be#143, 2026-10-02 08:57 UTC): 30 PRs — 25 code, 3 Markdown-only (be#132, be#143, pwa#42; exempt) and 1 Dependabot PR whose only commit is Dependabot’s own (pwa#43; exempt). pwa#44 was handled by the second pass. - Reviewed before merge, with a record on the merged head: be#128, be#131, be#133, be#141, fe#95, fe#98, fe#101, fe#106, fe#108, fe#109, pwa#41, pwa#46, pwa#47, pwa#48, pwa#49. Every one has a
Forge reviewsuccesson the merged head, noOpenfinding, and (backend) a build record for that head. - A review ran, but no record was written: be#122 (the third pass,
approve, covered the merged head8463d58; nothing further needed), fe#99 (the last review was onf456a11;0921665, applying its seven Nits, was not re-reviewed) and fe#102 (8f8bfa2, which says it applies review fixes, was not re-reviewed). - No review record, or fix commits never re-reviewed: be#136 (
58a01f3Dockerfile digest pins,28ca04fthe Dependabotdockerentry), fe#103 and pwa#45 (BUG-046), fe#104 (24c969f, the fix for its review finding) and fe#107 (BUG-058). - No PR merged with an open Blocker or Important, or with a
failurestatus on its head. - Post-merge reviews, 2026-10-02 (
backend-reviewerfor be#136,frontend-reviewerfor the rest; each finding checked againstmain):- be#136: the digests are valid pins of the intended tags, and the
dockerDependabot entry is correct. Important: a Dependabot PR that moves a base-image digest would get an automaticForge reviewsuccessfromforge-review-dependabot.yml, although no CI job builds or scans the image. Open forge-et-be#144 removes the digest pins and thedockerentry, which removes the case. If #144 is abandoned, adependabot/docker/*PR must be excluded from the automatic status (or CI must build and scan the image for it). Nit, left as is:# syntax=docker/dockerfile:1is a floating tag. - fe#107, fe#103 / pwa#45, fe#104, fe#99
0921665: pass-with-nits, no Blocker or Important. fe#1028f8bfa2: clean.
- be#136: the digests are valid pins of the intended tags, and the
- Lead decisions, 2026-10-02 (Tharinda):
- The Nits are fixed in one follow-up PR per frontend (
fix/post-merge-review-nits), reviewed and recorded under the required check:- fe#107: the auth setup’s cleanup no longer hides the first failure, and the subscription message names the received value;
- fe#104: a guard ties the engine to the size project’s name;
- fe#99: the expired re-read after a back-forward restore gets its test row;
- forge-et-pwa:
rules.mjs,ci-check.mjs,browserGateCiCheck.test.tsandbrowserGateRules.test.tsjoinscripts/shared-with-fe.json, sopnpm check:sharedand the gate’s live comparison catch a one-sided edit.
- Follow-up (done 2026-10-02): the rest of the browser gate is also byte-identical in the two frontends but unlisted:
scripts/browser-gate.mjs,scripts/browser-gate/run.mjs,lockfile.mjs, the.d.mtsfiles,gate-inputs.jsonand the otherbrowserGate*.test.tsfiles (onlyrepo.mjsandexpected-skips.jsondiffer by design). Listing them needs aSOURCE OF TRUTHheader on each copy in forge-et-fe, a re-copy,check-shared-drift.mjs --update, and the list insharedDrift.test.ts. Raised byfrontend-revieweron forge-et-pwa#55 (I-1); left out of that PR because it changes none of those files. Done in forge-et-fe#116 (the headers, andCLAUDE.md’s list) and forge-et-pwa#56 (the re-copy,shared-with-fe.json, the manifest,sharedDrift.test.ts,CLAUDE.md), each with anapproverecord; they merge after fe#114 and pwa#55, on which they are stacked. The list also took the gate suites’ helpergitFixture.tsand thefixtures/dependency-change/files, which the sharedbrowserGateDependencyChange.test.tsreads; the JSON and lockfile fixtures carry no header. - Follow-ups from forge-et-pwa#56’s review (open, Nits): the comment on the shared-file step in forge-et-pwa’s
.github/workflows/pr-validation.ymlstill says “the four design files” (a workflow edit, so it waits for the lead’s approval); ande2e/browser/support/page-check.spec.tsandscripts/check-testid-coverage.mjsare byte-identical in the two frontends but unlisted (both outside the browser gate; listing them needs the forge-et-fe header first). - BUG-046’s open follow-up: a skip reason with no letter or digit is refused, in both frontends and the harness hook together (
bugs.md§ BUG-046).
- The Nits are fixed in one follow-up PR per frontend (
- The check itself: probe forge-et-pwa#51 (closed) showed a PR with green
Fast laneand no recordBLOCKED, thenCLEANafter a record. The design’s implementation note has the detail.
Operator: Tharinda Rajapaksha
/forge-deliver state reconciliation on 2026-09-28
Ticket: F38-c
Healed: waves[2].status pr-open → merged; waves[2].merged_at null → 2026-09-28
Evidence: gh pr list --head feature/F38-c-wave-2 --state all shows forge-et-be#94 MERGED 2026-09-28T06:21:02Z; the remote wave branch was auto-deleted on merge.
Operator: Tharinda Rajapaksha
/forge-deliver --wave override on 2026-09-28
Ticket: F26-a
Wave: 2
Deps unmet: wave-1:planned (on main; Wave 1 was dispatched on 2026-09-28 in another session, whose record is on WI-1.1’s unpushed harness branch)
Gates unmet: none
Operator: Biseni Perera
Decision: proceeded — operator asserts wave is file-disjoint from unmet deps and accepts rebase risk if those deps mutate shared files later; any unmet gate is overridden by the operator’s explicit choice. Run with --no-implement: Wave 2’s plans are authored and reviewed only, nothing is dispatched, and the plans cite WI-1.1’s ### Locked contract for everything Wave 1 builds.
/forge-deliver state reconciliation on 2026-09-28 (Wave 1 merged)
Ticket: F26-a
Healed: waves[1].status pr-open → merged; waves[1].merged_at null → 2026-09-28
Evidence: gh pr view 100 --repo altrium-io/forge-et-be → MERGED 2026-09-28T12:41:27Z, merge commit c5876e5 (head feature/F26-a-wave-1); gh pr view 268 --repo altrium-io/forge-et-harness → MERGED 2026-09-28T12:42:13Z, merge commit 5abf81d; per-WI PRs forge-et-be#97 and #98 CLOSED; git ls-remote --heads origin "feature/F26-a-*" prints nothing in either repo. Every local F26-a Wave 1 branch passed git merge-base --is-ancestor <branch> origin/main and was deleted with git branch -d; the two Wave 1 worktrees were removed.
Operator: Biseni Perera
/forge-deliver state reconciliation on 2026-09-29 (Wave 2 merged)
Ticket: F26-a
Healed: waves[2].status pr-open → merged; waves[2].merged_at null → 2026-09-29; phase dev → review → ship (Stage 15, every wave merged)
Evidence: gh pr view 110 --repo altrium-io/forge-et-be → MERGED, merge commit fc21b20 (head c04572a, the commit its local build record names); gh pr view 279 --repo altrium-io/forge-et-harness → MERGED, merge commit 4009a2d; per-WI PRs forge-et-be#106, #107 and #109 CLOSED. The three per-WI remote branches and every local F26-a branch passed git merge-base --is-ancestor <branch> origin/main before deletion (git push --delete, git branch -d); the Wave 2 worktrees were removed.
Operator: Biseni Perera
/forge-deliver state reconciliation on 2026-10-01 (Wave 2 merged)
Ticket: F5-i
Healed: waves[2].status pr-open → merged; waves[2].merged_at null → 2026-09-30; phase dev → review → ship (Stage 15, every wave merged)
Evidence: gh pr view 322 --repo altrium-io/forge-et-harness → MERGED 2026-09-30T12:29:35Z, merge commit 84d4dc6 (head e588f58, feature/F5-i-wave-2, the manual wave’s record); git ls-remote --heads origin "feature/F5-i-*" prints nothing, and no local feature/F5-i-* branch or worktree remains in forge-et-harness, forge-et-be or forge-et-fe, so there was nothing to clean up.
Operator: Tharinda Rajapaksha
/forge-deliver state reconciliation on 2026-10-02 (Wave 1 merged)
Ticket: F5-l
Healed: waves[1].status pr-open → merged; waves[1].merged_at null → 2026-09-30; phase dev → review → ship (Stage 15, every wave merged)
Evidence: gh pr list --head feature/F5-l-wave-1 --state all → forge-et-be#126 MERGED 2026-09-30T18:57:52Z (head 1a34521, the commit its local build record names), forge-et-fe#97 MERGED 2026-09-30T18:57:56Z (head 277d5fb), forge-et-harness#319 MERGED 2026-09-30T18:58:00Z. No WI has pr_role: ship, so no per-WI ship-PR check applied. No feature/F5-l-* branch remains, remote or local, in forge-et-harness, forge-et-be or forge-et-fe. The two app wave PRs carry no Forge review status because they merged before .claude/lib/forge-review-record.sh existed; their /forge-review-pr reviews are in F5-l-WI-1.1-plan.md and F5-l-WI-1.3-plan.md.
Operator: Tharinda Rajapaksha
/forge-deliver state reconciliation on 2026-10-02 (F5-k Wave 2 merged)
Ticket: F5-k
Healed: features.F5-k.waves[2].status pr-open → merged (+ merged_at: null → 2026-10-02)
Evidence: gh pr list --head feature/F5-k-wave-2 --state all → forge-et-be#124 MERGED 2026-09-30T14:39:31Z, forge-et-harness#316 MERGED 2026-10-01T05:07:27Z, forge-et-fe#95 MERGED 2026-10-02T06:48:52Z. Every wave-2 WI is pr_role: review, so no per-WI ship-PR check applied. No feature/F5-k-* branch remains, remote or local, in forge-et-be or forge-et-fe; branch cleanup was a no-op. The heal is committed on feature/F5-k-wave-3, cut from origin/main, because F5-k’s spec, plan and wave-2 branches have all merged.
Operator: Tharinda Rajapaksha
/forge-deliver state reconciliation on 2026-10-03 (F5-k Wave 3 merged)
Ticket: F5-k
Healed: features.F5-k.waves[3].status pr-open → merged (+ merged_at: null → 2026-10-03); phase dev → review → ship (Stage 15, every wave merged)
Evidence: gh pr list --head feature/F5-k-wave-3 --state all → forge-et-harness#394 MERGED 2026-10-03T10:07:14Z (head c78178c), forge-et-fe#121 MERGED 2026-10-03T10:16:56Z (head fc01500, Forge review success, Fast lane success). WI-3.1 is pr_role: integration; the wave-1 ship PRs forge-et-be#79 and forge-et-fe#61 were re-checked and every required check other than Forge review is green on their merged heads. No feature/F5-k-* branch remains on any remote; the three wave-3 worktrees and the local feature/F5-k-wave-3 branches in forge-et-fe and the harness were removed after confirmation. The heal is committed on chore/F5-k-reflect, cut from origin/main, because F5-k’s spec, plan and wave branches have all merged.
Operator: Tharinda Rajapaksha
/forge-deliver state reconciliation on 2026-10-02 (F5-f Wave 1 merged)
Ticket: F5-f
Healed: waves[1].status pr-open → merged; waves[1].merged_at null → 2026-10-02
Evidence: gh pr view 153 --repo altrium-io/forge-et-be → MERGED 2026-10-02T16:24:30Z, merge commit 68a2a8d (head eae896c, feature/F5-f-wave-1, which carries the approve review record and the local build record). Wave 1 touched forge-et-be only. No WI has pr_role: ship, so no per-WI ship-PR check applied. The per-WI remote branches are already gone; the local wave branch and its worktree remain.
Operator: Tharinda Rajapaksha
/forge-deliver state reconciliation on 2026-10-04
Ticket: F38-j
Healed: workitems[F38-j-WI-1.2].impl_status dispatched → pending
Evidence: git ls-remote --heads origin feature/F38-j-WI-1.2-scan-off-path in forge-et-be prints nothing, and gh pr list --head feature/F38-j-WI-1.2-scan-off-path --state all returns []. The first agent stopped on an API session limit before pushing; its eight local commits (e67ebcb…7c4da04, subtasks 1–9) remain in the WI-1.2 worktree, and the re-dispatch resumes at subtask 10 on them.
Operator: Tharinda Rajapaksha
/forge-deliver state reconciliation on 2026-10-04 (F38-j Wave 1 merged)
Ticket: F38-j
Healed: features.F38-j.waves[1].status pr-open → merged (+ merged_at: null → 2026-10-04)
Evidence: gh pr list --repo altrium-io/forge-et-be --head feature/F38-j-wave-1 --state all → forge-et-be#171 MERGED 2026-10-04T06:52:54Z (head 75876d0, Forge review success). Both wave-1 WIs are pr_role: review and their per-WI PRs (#165, #170) are CLOSED as superseded, so no per-WI ship-PR check applied.
Operator: Tharinda Rajapaksha
/forge-deliver cross-WI gate on 2026-10-04 (F38-j Wave 2)
Ticket: F38-j
Wave: 2 (WI-2.1, WI-2.2, WI-2.3 — drafted in parallel by forge-plan-author)
Contract alignment: consistent. WI-2.1 D10 (the sweep does not advance version) and WI-2.2 D4 (the event records the status the handler read) together make a racing scan result win, as FR-10 requires; WI-2.3’s R-4 sentence holds. Recorded in both plans’ Notes.
Gap closed: forge-et-be/docs/api/README.md § 8 (“stays PENDING permanently”) was owned by no wave-2 plan; moved to WI-2.3 by its R-8 fallback (plan revised before review).
File collision: forge-et-be/docs/audit/README.md in WI-2.1 and WI-2.2 (adjacent rows, as the approved Decomposition Plan assigns). Decision (lead): accept; both WIs stay parallel from main, WI-2.1 merges first at Stage 11 Pass 1 and the conflict is resolved by keeping both edits. The spec’s AC checkboxes are shared harness ticks, not a collision.
Operator: Tharinda Rajapaksha
/forge-deliver dispatch stop on 2026-10-04 (F38-j Wave 2)
Ticket: F38-j
Changed: workitems[F38-j-WI-2.2].impl_status dispatched → pending; workitems[F38-j-WI-2.3].impl_status dispatched → pending
Reason: the lead asked to work sequentially with few agents because the token budget was low; the orchestrator stopped both background implementers. WI-2.1 kept running.
State left: WI-2.2 subtasks 1-5 committed locally (a8b2243…ad74fb2), waiting on the shared lock at subtask 6; WI-2.3 subtasks 1-7 done (forge-et-be 8bd4ed4…0f5de71; harness edits committed on feature/F38-j-plan). Neither branch is pushed; no PR. Both worktrees clean. Resume points are in the feature’s tracker notes.
Whole-project audit — 2026-10-04
Scope and method. The harness, the three app repos and the review trail, at the lead’s request: review records on every app-repo PR merged since Forge review became required (none missing), the harness machinery and instruction text, the backend and both frontends against coding-standards.md and design-system.md, a stale-text sweep, and whether the reviewer and implementer agents carry the standards. Every finding ends fixed (PR), guarded (check or agent line) or deferred (owner).
Report: .forge/audits/2026-10-04-project-audit.md; working record and decisions: .forge/audits/2026-10-04-project-audit-progress.md.
Operator: Tharinda Rajapaksha
/forge-deliver state reconciliation on 2026-10-04
Ticket: F38-j
Healed: waves[2].status pr-open → merged; merged_at null → 2026-10-04
Evidence: gh pr list --repo altrium-io/forge-et-be --head feature/F38-j-wave-2 --state all returns forge-et-be#177 MERGED at 2026-10-04T09:01:46Z, head ccf5a6c, which carries a success Forge review status. Plan PR forge-et-harness#414 has also merged, so the close-out writes go on chore/F38-j-reflect, cut from origin/main at 215dfddc.
Operator: Tharinda Rajapaksha
/forge-deliver state reconciliation on 2026-10-04
Ticket: F5-c
Healed: waves[1].status pr-open → merged; waves[1].merged_at null → 2026-10-04
Evidence: gh pr list --head feature/F5-c-wave-1 --state all reduces to MERGED in both touched repos — forge-et-be#172 (merged 2026-10-04T10:40:29Z) and forge-et-fe#130 (merged 2026-10-04T10:53:10Z); the per-WI PRs forge-et-be#162 and forge-et-fe#126 are CLOSED (superseded) and forge-et-fe#129 MERGED into the wave branch, consistent with wave-closed. Written on feature/F5-c-wave-2, cut from origin/main, because the plan PR #407 has merged.
Operator: Tharinda Rajapaksha
/forge-deliver wave-merge exception on 2026-10-06
Ticket: F5-c
Wave: 3
Healed: waves[3].status pr-open → merged; waves[3].merged_at null → 2026-10-06
Evidence: forge-et-be#201 MERGED 2026-10-05T20:19:33Z (merge commit 1a17343) and forge-et-fe#155 MERGED 2026-10-05T20:19:36Z (merge commit abb5105). The harness wave PR forge-et-harness#468 (feature/F5-c-wave-3, head 51eeb59) was OPEN when this was written: its content is final and reviewed, and it waited only on its required Repo hygiene check, which stayed queued because GitHub Actions hosted runners were not taking jobs during a GitHub incident.
Exception: the lead’s decision for this session — count wave 3 as merged although #468 is open, because it is a records-only harness PR. Its tracker row was therefore read from origin/feature/F5-c-wave-3, not origin/main, and the wave-4 harness branch feature/F5-c-wave-4 was cut from origin/feature/F5-c-wave-3. The wave-4 harness PR merges after #468. The harness branches feature/F5-c-wave-3, feature/F5-c-WI-3.2-activation-controls and feature/F5-c-WI-3.3-verify-activation are left in place until #468 merges.
Operator: Tharinda Rajapaksha
/forge-deliver state reconciliation on 2026-10-15
Ticket: F38-a
Healed: waves[1].status pr-open → merged; waves[1].merged_at null → 2026-10-15
Evidence: gh pr view 705 --repo altrium-io/forge-et-harness shows MERGED at 2026-10-15T09:45:31Z (merge commit 99640e28); wave 1 touched only the harness, and no per-WI PR was opened (the wave branch was a fast-forward of the WI-1.1 branch). Written on chore/F38-a-wave-1-merged, cut from origin/main, because the plan PR #702 has merged and no wave is being dispatched.
Operator: Tharinda Rajapaksha
/forge-deliver state reconciliation on 2026-10-16
Ticket: F20-h
Healed: waves[2].status pr-open → merged; waves[2].merged_at null → 2026-10-16
Evidence: gh pr view 732 --repo altrium-io/forge-et-harness shows MERGED at 2026-10-16T08:02:34Z (merge commit 2606d643); wave 2 is a manual wave that touched only the harness, with no per-WI PR. Written on chore/F20-h-reflect, cut from origin/main, because the plan PR #657 has merged; Stage 15 follows on the same branch.
Operator: Dinil Gamage
/forge-deliver state reconciliation on 2026-10-19
Ticket: F38-d
Healed: waves[1].status pr-open → merged; waves[1].merged_at null → 2026-10-19
Evidence: gh pr view 772 --repo altrium-io/forge-et-harness shows MERGED at 2026-10-19T04:40:53Z (merge commit a2598c92); wave 1 touched only the harness, and its per-WI PR #771 was closed as superseded (pr_role: review, so no ship-PR check applies). Written on feature/F38-d-wave-2, cut from origin/main, because the plan PR #770 has merged and wave 2 is being planned.
Operator: Tharinda Rajapaksha
/forge-deliver state reconciliation on 2026-10-20
Ticket: F38-g
Healed: waves[1].status pr-open → merged; waves[1].merged_at null → 2026-10-20
Evidence: gh pr list --head feature/F38-g-wave-1 --state all shows MERGED in all three touched repos: forge-et-fe#311 at 2026-10-20T10:03:38Z (merge commit b576aaf2), forge-et-pwa#84 at 10:04:14Z (174553bb), forge-et-harness#798 at 10:08:30Z (36a503f5). Both per-WI PRs (forge-et-fe#306, forge-et-pwa#83) were closed as superseded (pr_role: review, so no ship-PR check applies). Written on feature/F38-g-wave-2, cut from origin/main, because the plan PR #790 has merged and wave 2 is being planned.
Operator: Tharinda Rajapaksha
/forge-deliver state reconciliation on 2026-10-21
Ticket: F1-c
Healed: waves[1].status pr-open → merged (merged_at 2026-10-21)
Evidence: gh pr list --head feature/F1-c-wave-1 --state all — forge-et-be [“MERGED”] (#284, merge aa987d73), forge-et-fe [“MERGED”] (#312, merge b2dc3ee7); forge-et-harness has no wave PR: the work items’ harness branches merged into feature/F1-c-plan (#775, #778, #780, #800), which ships to main as the plan PR #774. No WI is at pr_role: ship, so no ship-PR check applies.
Operator: Biseni Perera
/forge-deliver state reconciliation on 2026-10-22
Ticket: F38-g
Healed: waves[1].status pr-open → merged; waves[1].merged_at null → 2026-10-20
Evidence: gh pr view shows MERGED for all three wave-1 PRs: forge-et-fe#311 at 2026-10-20T10:03:38Z (merge commit b576aaf2), forge-et-pwa#84 at 10:04:14Z (174553bb), forge-et-harness#798 at 10:08:30Z (36a503f5). The same heal is on feature/F38-g-wave-2 (15bc0cae, not yet merged), with the same two tracker lines and the feature’s wave_merged metrics event; it is written again here on feature/F38-g-wave-3, cut from origin/main for wave 3, which runs in parallel with wave 2. No second wave_merged event is added, so the merge is counted once.
Operator: Tharinda Rajapaksha
/forge-deliver state reconciliation on 2026-10-22
Ticket: F38-g
Healed: waves[3].status pr-open → merged; waves[3].merged_at null → 2026-10-22
Evidence: gh pr list --head feature/F38-g-wave-3 --state all shows forge-et-harness#858 MERGED at 2026-10-22T06:11:37Z (merge commit c218234b); wave 3 touched only the harness, and WI-3.1 is at wave-closed (no pr_role: ship, so no ship-PR check applies). Written on feature/F38-g-wave-2, the live harness wave branch, after origin/main was merged into it at Stage 7 M2.
Operator: Tharinda Rajapaksha
/forge-deliver state reconciliation on 2026-10-26
Ticket: F38-b
Healed: waves[3].status pr-open → merged; waves[3].merged_at null → 2026-10-26
Evidence: gh pr list --head feature/F38-b-wave-3 --state all shows forge-et-harness#940 MERGED at 2026-10-26T07:13:47Z (merge commit cc707069); wave 3 is a manual wave that touched only the harness, and WI-3.1 is at wave-closed (no pr_role: ship). Every other F38-b wave PR reads MERGED. Written on chore/F38-b-reflect, cut from origin/main for Stage 15, with the dev → review → ship flips.
Operator: Tharinda Rajapaksha
/forge-deliver state reconciliation on 2026-10-22
Ticket: F38-g
Healed: waves[2].status pr-open → merged; waves[2].merged_at null → 2026-10-22
Evidence: gh pr view 862 shows forge-et-harness#862 (head feature/F38-g-wave-2) MERGED at 2026-10-22T06:54:50Z (merge commit ccf61eb2); wave 2 is a manual wave touching only the harness, and WI-2.1 is at wave-closed (no pr_role: ship, so no ship-PR check applies). Written on feature/F38-g-wave-4, cut from origin/main for wave 4, because the plan PR #790 has merged.
Operator: Tharinda Rajapaksha
/forge-deliver --wave override on 2026-10-26
Ticket: F38-h
Wave: 3
Deps unmet: wave-1:pr-open (forge-et-iaac#11 open, not yet approved or merged)
Gates unmet: none
Operator: Tharinda Rajapaksha
Decision: proceeded — wave 3 is not file-disjoint from wave 1: WI-3.1’s overlay sets the variables WI-1.1 adds (a compile coupling in forge-et-iaac). WI-3.1’s plan is authored and reviewed now; its build is stacked on wave 1 — its forge-et-iaac branch is cut from feature/F38-h-wave-1, and its PR to dev opens only once forge-et-iaac#11 has merged (or is rebased onto dev then). A change to wave 1 during its review is merged into the wave 3 branch before its PR opens.
/forge-deliver --wave override on 2026-10-26
Ticket: F38-h
Wave: 4
Deps unmet: wave-3:in-progress (WI-3.1 built and pushed on forge-et-iaac feature/F38-h-wave-3 at e78b12e, no PR; it waits for wave 1’s forge-et-iaac#11). Wave 2 is merged.
Gates unmet: none (wave_merged F38-d wave 7: forge-et-harness#960 MERGED)
Operator: Tharinda Rajapaksha
Decision: proceeded for plan writing and review only — WI-4.1’s and WI-4.2’s plans are written against wave 3’s branch for the et-uat-* names (the Decomposition Plan’s Shared Contracts). Wave 4 is file-disjoint from wave 3 (different repositories), but its workflows name what wave 3 creates, so no wave 4 work item is dispatched until wave 3 has merged to forge-et-iaac dev; a change to wave 3’s names before then is carried into both plans by a revision. Stage 0’s expected halt on WI-3.1 (dispatched, branch pushed, no PR — the F38-h tracker note) is left as it is: it concerns wave 3’s dispatch, not this plan writing.
/forge-deliver state reconciliation on 2026-10-27
Ticket: F3-a Healed: waves[1].status pr-open → merged; merged_at null → 2026-10-27 Evidence: every wave 1 PR is MERGED — forge-et-be#273 (2026-10-27T03:24:00Z), forge-et-fe#296 (03:24:17Z), forge-et-harness#767 (03:24:38Z); the per-WI PRs forge-et-be#272 and forge-et-fe#294 are CLOSED as superseded and forge-et-fe#295 MERGED into feature/F3-a-wave-1. No F3-a branch or worktree remains locally or on the remote. Operator: Nanndish
/forge-deliver state reconciliation on 2026-11-02
Ticket: F38-i
Healed: waves[2].status pr-open → merged; waves[2].merged_at null → 2026-10-29
Evidence: gh pr view 1035 shows forge-et-harness#1035 (head feature/F38-i-wave-2) MERGED at 2026-10-29T18:03:35Z; wave 2 touched only the harness, and WI-2.1 is at wave-closed (no pr_role: ship, so no ship-PR check applies). The live harness wave branch feature/F38-i-wave-3 was cut before #1035 merged, so its row still shows wave 2 as planned; it is left untouched (another session’s wave). Written on feature/F38-i-wave-10, cut from origin/main for wave 10 (--wave 10: no depends_on, no gates).
Operator: Tharinda Rajapaksha
/forge-deliver state reconciliation on 2026-11-03
Ticket: F43
Healed: waves[2].status pr-open → merged; waves[2].merged_at null → 2026-11-03
Evidence: gh pr list --head feature/F43-wave-2 --state all shows forge-et-be#351 MERGED at 2026-11-03T06:19:18Z, forge-et-fe#384 MERGED at 2026-11-03T06:19:36Z (after the backend, as the Decomposition Plan asks) and forge-et-harness#1086 MERGED at 2026-11-03T04:29:17Z. Every wave 2 WI is at pr_role: review / wave-closed, so no ship-PR check applies. Written on feature/F43-plan (plan PR #1072 still open).
Operator: Tharinda Rajapaksha
/forge-deliver --wave override on 2026-11-03
Ticket: F43
Wave: 4
Deps unmet: wave-3:pr-open (forge-et-iaac#16 waits for Shyamin Ayesh’s approval; forge-et-harness#1091 waits for #16’s apply to be recorded). Waves 1 and 2 are merged.
Gates unmet: none
Operator: Tharinda Rajapaksha
Decision: proceeded — wave 4 is file-disjoint from wave 3 (forge-et-be and the harness; wave 3 is forge-et-iaac plus three harness records WI-4.2 does not edit, and WI-4.2’s harness PR merges after #1091). The Decomposition Plan § Notes check still holds: wave 4’s forge-et-be PR merges only after #16’s dev.yaml apply succeeded and aws iam get-role-policy --role-name et-dev-backend-task --policy-name bedrock lists all six profiles, because that merge deploys dev’s Global background keys. Started early so Shyamin Ayesh’s review is off the critical path (the developer’s “Start wave 4 now”, 2026-11-03).
/forge-deliver state reconciliation on 2026-11-03
Ticket: F3-a
Healed: waves[2].status pr-open → merged; merged_at null → 2026-11-03
Evidence: every wave 2 PR is MERGED — forge-et-be#327 (2026-11-03T07:46:19Z), forge-et-fe#353 (07:46:33Z), forge-et-harness#1005 (07:46:38Z). No WI is at pr_role: ship, so no ship-PR check applies. No F3-a branch or worktree remains locally or on the remote in forge-et-be, forge-et-fe or the harness. Written on feature/F3-a-wave-3, cut from origin/main for wave 3.
Operator: Nanndish
/forge-deliver state reconciliation on 2026-11-07
Ticket: F3-a
Healed: waves[3].status pr-open → merged; merged_at null → 2026-11-06
Evidence: every wave 3 PR is MERGED — forge-et-fe#431 (2026-11-06T13:30:05Z, head 4ccb40cc), forge-et-harness#1183 (2026-11-06T13:51:17Z). No WI is at pr_role: ship, so no ship-PR check applies. Written on feature/F3-a-wave-4, cut from origin/main for wave 4.
Operator: Nanndish (run overnight by Claude under Nanndish’s instruction to finish F3-a up to ready PRs)
/forge-deliver state reconciliation on 2026-11-08
Ticket: F3-a
Healed: waves[4].status pr-open → merged; merged_at null → 2026-11-08
Evidence: every wave 4 PR is MERGED — forge-et-fe#445 (2026-11-08T04:43:32Z, head 8750f015), forge-et-harness#1192 (2026-11-08T04:44:51Z, head 52f3a97e). WI-4.1 is at pr_role: integration, not ship, so no ship-PR check applies. GitHub deleted every F3-a remote branch on merge. Written on chore/F3-a-reflect, cut from origin/main (c6ddc19b) for Stage 15.
Operator: Nanndish
/forge-deliver state reconciliation on 2026-11-11
Ticket: F34-c
Healed: waves[2].status pr-open → merged; merged_at null → 2026-11-11
Evidence: every wave 2 PR is MERGED — forge-et-be#405 (2026-11-11T15:58:01Z, head b10e654b, merge ff0ed4d5), forge-et-fe#497 (2026-11-11T15:58:17Z, head b13b1629, merge 63123ad1), merged by Tharinda’s decision in the session. Every WI is at pr_role: review, so no ship-PR check applies. The harness records reached main through the checkpoint PR forge-et-harness#1288. Written on chore/F34-c-reflect, cut from origin/main (e085d2be) for Stage 15.
Operator: Tharinda
/forge-deliver state reconciliation on 2026-11-13
Ticket: F36-a
Healed: waves[1].status in-progress → pr-open; F36-a-WI-1.1, F36-a-WI-1.2 and F36-a-WI-1.4 impl_status pr-open → wave-closed; F36-a-WI-1.3 impl_status dispatched → wave-closed
Evidence: Stage 12’s forge-pr-open --wave 1 writes never reached a pushed branch. Wave 1’s PRs: forge-et-be#423 and forge-et-fe#531 MERGED (2026-11-13T10:00Z, heads 0f36d6b3 and a676001c), forge-et-harness#1368 MERGED, forge-et-pwa#128 OPEN (draft, head c09d1b90); the per-WI drafts forge-et-be#421, forge-et-fe#530 and forge-et-pwa#127 CLOSED as superseded. WI-1.3 (dispatched, branch pushed, no PR of its own) is not the ambiguous Stage 0 case: its branch head 61c26aa is an ancestor of forge-et-pwa origin/feature/F36-a-wave-1 (Stage 11 Pass 2, c543b88), and its plan’s ## Progress records both phases green on that head; its test-verify (Pass) and pre-PR review (Nits only) were run in this session. The live tracker row was on feature/F36-a-wave-3-plan (a harness branch Step 0a does not recognise as a wave branch); written there after merging origin/main into it.
Operator: Nanndish
/forge-deliver state reconciliation on 2026-11-14
Ticket: F36-a
Healed: waves[1].status pr-open → merged, merged_at null → 2026-11-14; waves[2].status planned → in-progress; F36-a-WI-2.1 and F36-a-WI-2.2 impl_status pending → dispatched, with run_by: main-session, their branches and touched_repos
Evidence: every wave 1 PR is MERGED — forge-et-be#423 and forge-et-fe#531 (2026-11-13T10:00Z), forge-et-harness#1368, and forge-et-pwa#128 (2026-11-14T07:23:21Z, merge e6bcc1321, head 09783ce with success Forge review and Test record), merged by Nanndish; every wave 1 WI is at pr_role: review, so no ship-PR check applies. WI-2.1 (forge-et-be feature/F36-a-WI-2.1-captured-leads) and WI-2.2 (forge-et-pwa feature/F36-a-WI-2.2-capture-screen) were built from 2026-11-13 on top of feature/F36-a-wave-1, by Nanndish’s approval, before the tracker recorded a dispatch; no agent runs them now, so they are marked run_by: main-session and continued in the main session. Written on the harness wave branch feature/F36-a-wave-2, cut from origin/main (c6f7f7438).
Operator: Nanndish
Documentation review — 2026-10-26
Scope and method. At the lead’s request, every harness document that states current facts, and the Markdown and instruction files of forge-et-be, forge-et-fe, forge-et-pwa and forge-et-iaac, checked for anything outdated, inaccurate or inconsistent. Reviewers checked each claim against its source, independent verifiers tried to refute each finding, and only confirmed findings were fixed: in the harness by forge-et-harness#936, and in the app repositories by Markdown-only PRs.
Report: .forge/audits/2026-10-26-doc-review.md — the findings, where each one ended, and the status at hand-off. Entry added on 2026-11-12: the report had no inbound link once the .forge/owed.md rows that pointed at it were closed.
Operator: Tharinda
/forge-deliver --wave override on 2026-11-14
Ticket: F36-a
Wave: 3
Deps unmet: wave-2:in-progress (its work items are integrated on feature/F36-a-wave-2 in forge-et-be and forge-et-pwa; WI-2.4, the verify, is running; no wave 2 PR is open yet)
Gates unmet: none
Operator: Claude, under Nanndish’s overnight delegation of 2026-11-14 (“complete as much as possible for F36-a … don’t stop to ask”), for his review in the morning
Decision: proceeded — wave 3 is not file-disjoint from wave 2: WI-3.1 changes CapturedLeadService and CapturedLeadResponse (the capture stores the campaign id), which WI-2.1 creates, and WI-3.3 sends the active campaign through WI-2.2’s submit module. So wave 3’s work items are built stacked on wave 2, as wave 2 was on wave 1 by Nanndish’s approval of 2026-11-13: each first work item’s branch is cut from feature/F36-a-wave-2 in its repo, and feature/F36-a-wave-3 is cut from feature/F36-a-wave-2, not main. Once wave 2’s PRs merge, origin/main is merged into each wave 3 branch, and wave 3’s PRs open against main only then (or open now as drafts whose diff includes wave 2, retargeted after the merge). A change made to wave 2 in its review is merged into the wave 3 branches the same way.
/forge-deliver --wave override on 2026-11-17
Ticket: F36-c
Wave: 2
Deps unmet: none (wave 1 merged)
Gates unmet: F36-b wave 3 merged in forge-et-be and forge-et-pwa — F36-b wave 3 is being built on F36-a’s unmerged wave 4 (its own override, recorded on harness feature/F36-b-wave-3); no F36-b wave 3 PR is open yet
Operator: Dinil
Decision: proceeded, by Dinil’s decision, to save time, for the backend work items only. F36-c WI-2.1 is cut from forge-et-be origin/feature/F36-b-WI-3.1-review-reads-convert (F36-b WI-3.1, code done at 0c2ee4cb, itself on F36-a’s feature/F36-a-wave-4); WI-2.2 stacks on WI-2.1 and WI-2.3 on WI-2.2, as planned. The capture-app work items WI-2.4 and WI-2.5 are not started until F36-b WI-3.4’s branch (feature/F36-b-WI-3.4-read-again, forge-et-pwa) is pushed, because they change the Leads screens F36-b WI-3.2 to WI-3.4 change; WI-2.4 is then cut from it, and WI-2.6 follows the plan. Every F36-c wave 2 PR stays a draft until F36-a wave 4 and F36-b wave 3 are on main, merged in that order; then origin/main is merged into each F36-c wave 2 branch, the local record runs again and the review records the merge as a delta. A change made lower down in review is merged up once, before the record.
/forge-deliver state reconciliation on 2026-11-16
Ticket: F36-a
Healed: waves[3].status pr-open → merged; merged_at null → 2026-11-16
Evidence: gh pr list --head feature/F36-a-wave-3 --state all reads MERGED in every repo the wave touched — forge-et-be#435 (2026-11-16 05:55 UTC), forge-et-pwa#136 (06:58 UTC) — and harness #1414 (07:14 UTC); no work item of wave 3 is at pr_role: ship
Operator: Nanndish
/forge-deliver --wave override on 2026-11-17
Ticket: F36-b
Wave: 3
Deps unmet: none (wave 2 merged 2026-11-17)
Gates unmet: F36-a’s review (wave 4) on main in forge-et-be and forge-et-pwa — forge-et-be#457 (ready, its build record running after a conflict merge) and forge-et-pwa#147 (draft) are open; both feature/F36-a-wave-4 branches were 0 commits behind main when checked
Operator: Dinil
Decision: proceeded, by Dinil’s decision (“start wave 3 on his branches”), to save time. Wave 3’s work items are built stacked on F36-a’s wave 4: WI-3.1’s branch is cut from origin/feature/F36-a-wave-4 in forge-et-be (07e40690) and WI-3.2’s from origin/feature/F36-a-wave-4 in forge-et-pwa (ee48727); WI-3.3 and WI-3.4 stack on WI-3.2 as planned. The harness wave branch feature/F36-b-wave-3 is cut from origin/main, because harness feature/F36-a-wave-4 is far behind main; WI-3.2’s pwa-current drawing subtask, which builds on F36-a’s wave-4 drawings, waits until forge-et-harness#1448 merges. F36-b’s per-WI and wave PRs stay drafts until both F36-a wave-4 PRs are on main; then origin/main is merged into each wave 3 branch, the local record runs again and the review records the merge as a delta. Nanndish was told on both PRs (not to force-push or rewrite the branches). A change F36-a makes to wave 4 in its review is merged into the wave 3 branches once, before the record.
/forge-deliver --wave override on 2026-11-18
Ticket: F36-c
Wave: 3 (new, from PRD Rev 101: the campaign focus and the capture ranking carried into the CRM; not yet in the Decomposition Plan or the tracker)
Deps unmet: wave-2:in-progress (its PRs are drafts; WI-2.6’s verify runs are in progress in another session, on harness feature/F36-c-wave-2)
Gates unmet: none defined yet — wave 3’s gates are set when it is added to the Decomposition Plan
Operator: Dinil
Decision: proceeded, by Dinil’s decision, to save time; wave 3 is not file-disjoint from wave 2 (it changes the captured-lead conversion of WI-2.1 and the ranking job of WI-2.2), so it is built stacked on wave 2, as wave 2 was on F36-b. The harness branch feature/F36-c-wave-3 is cut from origin/feature/F36-c-wave-2 (b289204c), and carries the F36-c spec Rev 6, the F11-g spec revision, the Decomposition Plan’s wave 3 and its plans. Each wave 3 work item in forge-et-be and forge-et-pwa is cut from that repo’s feature/F36-c-wave-2 branch (or from a wave 3 sibling’s branch, when the plan stacks it); a forge-et-fe work item, which wave 2 does not touch, is cut from main. Every F36-c wave 3 PR stays a draft until wave 2’s PRs are on main; then origin/main is merged into each wave 3 branch, the local record runs again and the review records the merge as a delta. A change made to wave 2 in its review (recorded in .forge/plans/F36-c/F36-c-wave-2-handoff.md) is merged into the wave 3 branches the same way. origin/main is not merged into the harness wave 3 branch now: it conflicts on the tracker, owed.md and the pwa-current prototype, which wave 2’s harness PR resolves first.
Follow-up (2026-11-18, wave 3 plan review pass 2): one wave 3 PR is the exception to “stays a draft until wave 2’s PRs are on main”: WI-3.8’s forge-et-fe PR, cut from main with base main and containing no wave 2 code. It is marked ready on its own and merges to main before the forge-et-be wave 3 PR and before wave 3’s Stage 11 Pass 1 and Stage 12; merging it stays Dinil’s. WI-3.4 is cut from WI-3.8’s branch rather than from main. The Decomposition Plan’s § Notes and § Wave Ship Plan C1 record the merge order.
/forge-deliver state reconciliation on 2026-11-20
Ticket: F20-i
Healed: waves[5].status pr-open → merged; merged_at null → 2026-11-20
Evidence: gh pr list --head feature/F20-i-wave-5 --state all reads MERGED — harness #1565 (2026-11-19 18:52 UTC, 2026-11-20 00:22 +0530); wave 5 is a manual wave touching only the harness, with no work item at pr_role: ship
Operator: Tharinda Rajapaksha
/forge-deliver state reconciliation on 2026-11-21
Ticket: F45
Healed: waves[2].status pr-open → merged; merged_at null → 2026-11-21
Evidence: gh pr list --head feature/F45-wave-2 --state all reads MERGED — harness #1632 (merge commit d2442c5c2, 2026-11-21 15:22 UTC); wave 2 is a manual wave touching only the harness, with no work item at pr_role: ship
Operator: Tharinda Rajapaksha