# Agent run status | Agent | Status | Last module processed | Last updated | Notes | | --------------- | ------------------- | ----------------------------- | ------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | baseline | complete | all (18 FE + 7 BE modules) | 2026-08-26 | 0 layering violations. 11 observations BL-001..BL-011. Thresholds for Phase 1 in §9 — agent 01's `[N] lines` = 40 (TS fn) / 60 (C# method) / 400 (file). Read BL-001, BL-002 and BL-004 before filing complexity, duplication or coverage findings. | | readability | skipped | n/a | 2026-08-27 | **skipped** — deliberate. BL-001: 23 of the 25 TS functions over CC 10 are reducers / `parse*` boundaries / `validate*`, all mandated house idioms; TS fn-length p99 is 34 with only 2 functions over 75 lines. Little left for this agent to find that is not a false positive. Revisit if the CC>10 population grows outside those three shapes. | | testability | complete | all 24 modules | 2026-08-26 | 9 findings TE-001..TE-009; 15 modules explicit "no findings". Corrected BL-004 — `libs/shared/domain` and `libs/beheer/contracts` are pure type declarations, 0% is unimprovable (amendment note in 00-baseline.md §10). TE-009 is also a latent correctness bug (dead `ActiveOn` branches). Acquitted `createDraftSync`, `httpClientFetch`, `Contracts/Mappers.cs` in writing. | | ddd-hexagonal | skipped | n/a | 2026-08-27 | **skipped** — deliberate. FE layering is clean (baseline §6: 0 violations, healthy instability gradient, `kernel` I=5% vs contexts I>=83%); backend `Domain/` is verified EF/ASP-free. The agent may only _extend_ existing hexagonal structure, and the one real target (`Program.cs`) has no `Features/` folder to extend — agent 04 already filed that as out-of-mandate OOM-A. | | cqrs-light | complete | all (16 FE + 6 BE modules) | 2026-08-26 | 7 findings CQ-001..CQ-007; 12 modules clean. Corrected BL-007 (see the amendment note in 00-baseline.md §10) and found 3 mutations the baseline missed. `Program.cs` → `Features/`+handlers filed as out-of-mandate OOM-A, not a ticket; CQ-006 is its prerequisite. `bhp/behandeling` named the reference implementation — do not "clean it up". | | bdd | skipped | n/a | 2026-08-27 | **skipped** — deliberate. No BDD tooling present, and the prompt forbids proposing any; it self-reduces to a single structural note. `gen:behaviour-spec` already extracts behaviours from spec names into `libs/shared/docs/behaviour-spec.mdx`, which covers the intent. | | adr-conformance | complete | all 6 ADRs | 2026-08-26 | 11 findings: 6 code, **5 ADR-fix (architect approval required)**. Sharpened BL-002 — `Principal` was never built, so ADR-0002 was untested not falsified (amendment note in 00-baseline.md §10). ADR-0005 fully conformed. Gates: ADR-C-005→ADR-C-004; ADR-C-003 gates contracts/ cleanup. 3 ADR-fixes need a matching CLAUDE.md correction in the same diff. | | bio2-compliance | complete | all modules + 7 control areas | 2026-08-27 | 20 findings BIO-001..BIO-020 (12 **defect now**, 8 **production gate**). High: BIO-008 BSN concatenated into the authz audit `Resource` (`Program.cs:674`, verified); BIO-004 `GET /uploads/{documentId}/content` has no authz at all (verified). Answered agent 06's handoff as BIO-002 — a production behandelportal build resolves to the seeded **zorgverlener**, failing closed on backoffice caps but open on citizen-scoped ones incl. `CanRevealBigNummer`. Carries the mandatory **"compliance review required"** list: 16 rows over agents 02/04/06. Also a pre-production checklist (~25 boxes). | | consolidation | complete (approved) | all inputs | 2026-08-27 | **HALTED for human approval** (per spec) — `99-backlog.md` written, nothing implemented. 33 open tickets RB-01..RB-33 + 5 ADR-fixes + 5 already-shipped, from 47 findings. RB-01 (no authz on upload content) and RB-02 (BSN in the audit `Resource`) sort above all structural work. Gate relaxed to the 4 agents that ran; a "Coverage of this backlog" note records what the 3 skips leave unowned. Caught two orchestrator errors: **CQ-002 is NOT fixed** (verified — `ApplicationsStore.cancel`/`AdminCasesStore.delete` still swallow errors → RB-20), and **CQ-004 shipped with half its compliance criterion unmet** (no audit row on `PUT /admin/flags/{key}`, verified → RB-07, which blocks signing ADR-C-009). OOM-D: re-run the baseline before using it to verify any ticket — ADR-C-006 and BL-008 moved it. **Approved 2026-08-27; HALT lifted.** | ## Phase 3 — implementation **All six batches complete, 2026-08-28. All 33 code tickets (RB-01..RB-33) and the four gated ADR-fixes (ADR-C-001, ADR-C-003, ADR-C-007, ADR-C-009) merged to `refactor/adr-c-006-shared-route-guards`, one commit per ticket. `npm run ci` green after every merge, verified independently before trusting any agent's own report.** Every batch carrying a **SIGN-OFF** ticket shipped only after the architect approval recorded 2026-08-27 (HALT lifted). Three of the four ADR-fixes required a matching CLAUDE.md correction (§2 once, §4 twice); all three landed in the same diff as their ADR amendment, per CLAUDE.md's own precedence rule. Across the six batches, several tickets turned out to be factually wrong, incomplete, or overstated relative to the actual code, and every one was reported rather than silently patched over — among them: two ADR-fixes (ADR-C-001's stale out-of-scope bullet, ADR-C-007 catching only half of its own finding), RB-18 (real scope one endpoint, not the several the stale line numbers implied), RB-23 (an unmentioned second `GetOrCreate` call site forced onto the same contract change), RB-25 (overstated which methods the missing token actually blocked), RB-27 (correctly declined to extract abort-vs-error into a signature that cannot express it, and declined an optional move outside its stated scope), RB-28 (TE-006's "cannot test the success case" claim was already false when written), and RB-31 (found a genuinely unreachable state — cursor 2 with no diploma chosen — that the old hand-rolled fixture had been asserting). None of these weakened a ticket; each was implemented correctly once the discrepancy was named. | CD batch | Tickets | Status | Notes | | -------- | ------------------------------------------ | ------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | 1 | RB-01, RB-02, RB-03, RB-04, RB-05, RB-06 | **complete** | Six commits on `refactor/adr-c-006-shared-route-guards`, one per ticket, each with `implementation/rb-0N.md`. `npm run ci` green. Every ticket left a test that was verified red without its fix. Carryover: RB-01's residual belongs to **RB-09** (the content endpoint is reached by a plain browser navigation with no identity header — BIO-002); `Pii.MaskTail` now lives in `Domain/People/Pii.cs`, **use it in RB-11** rather than hand-rolling a second masker; RB-06 additionally deleted `SubmissionRules.RejectRegistratie` (judgement call, recorded). | | 2 | RB-07, RB-08, RB-09, RB-10, RB-11 | **complete** | All five merged; `npm run ci` green **on the fixed gate** (see below). **RB-07** unblocks signing **ADR-C-009** and closes CQ-004's outstanding half. **RB-09** makes Production throw at startup when no real `IIdentityProvider` exists — note an environment that is neither Development nor Production (e.g. `Staging`) still fails fast, but at `GetRequiredService`, with a worse message. **RB-10** landed `parseStoredSession` twice, once per app, deliberately — TE-001/BL-002 say extract-to-shared contradicts ADR-0002; **RB-13** differentiates them. **RB-11** corrected a factual error in BIO-012 (the proefbrief error mapping was inlined, not already a separate function) and left the step-up as a literal moved one layer up to its only caller — BIO-006(c) stays a production gate. **RB-01's residual is still open** and is not solved by RB-09: the upload-content link is a plain browser navigation with no credential. | | 3 | RB-12, RB-13, RB-14, RB-15, RB-16, RB-17 | **complete** | All six merged; `npm run ci` green (14 steps — RB-14 added one — backend 260/260). **RB-12 rejected the ticket's binary framing:** of 47 routes only 16 use one of the five admin wrappers; of the remaining 31 only 10 are genuinely public, the other 21 are ownership-scoped inline (`ctx.Zorgverlener()`/`ctx.Caller()`) or use another mechanism. The allow-list therefore carries **a reason per route**, not a blanket "public" label. Known limitation: detection is `.Gate("XAdmin")` metadata declared at mapping time — **a declaration, not a derivation**, so it cannot catch a route that declares a gate it does not have. **This is RB-19's safety net; read `rb-12.md` before starting RB-19.** **RB-13** measured `ssp/auth` ↔ `bhp/auth` duplication at **32 lines each side, down from 168** (backlog expected <40); each app holds only its own `Principal` variant, which is ADR-C-004's own proposed resolution, and ADR-0002's "Known debt" section became an amendment. **RB-14** could not be built as written — `dotnet list package --vulnerable` exits 0 on a High advisory (verified), so a bare `- run:` would have been a gate that enforces nothing; `scripts/dotnet-audit.sh` matches the output instead and is shared by `ci.yml` and `ci-local.sh`. **RB-15** used a third environment name (`Staging`) in its test, since RB-09 makes Production fail to boot at all. | | | 4 | RB-18..RB-23 | **complete** | All six merged, one commit per ticket, each on its own merge. `npm run ci` green on the combined tree after every merge (14 steps, exit 0). Ran as three waves, because three of the six touch `Program.cs`: **A** = RB-18/20/21/22 in parallel (no file overlap), **B** = RB-23 after RB-22 (expand/contract), **C** = RB-19 alone and last, so it reordered final content. **Two tickets were incomplete, both reported rather than worked around.** RB-23 found `BriefStore.GetOrCreate` had a **second, unmentioned call site** — `GET /brief/preview` — so the split forced that endpoint to change too or the file would not compile; it got the same `Get` + 404 treatment. RB-18's real scope is **one** endpoint, not the nine BIO-018's stale line numbers implied: `Submit` has exactly one call site (`POST /change-requests`). **RB-22 deliberately left the `runResult` idiom** for `BriefAdapter.load()`: it hand-rolls try/catch to read the HTTP status, because `runResult` folds the error to a string and structurally cannot carry a 404. It still reuses the shared `problemDetail` mapper and models the outcome as the `BriefLoadFailure` union, not a sentinel string. Accepted — reviewed the diff before merging. Its once-only bound is stronger than the ticket asked: `recoverFromMissingBrief` never re-enters `load()`, so CQ-007's retry loop is absent, not merely capped. **RB-22 mispredicted one thing harmlessly:** it expected the regenerated client to parse a `ProblemDetails` 404, but `Results.NotFound()` declares no body so it throws a plain `SwaggerException` (matching the 17 other bare-404 endpoints). `isHttpNotFound` reads only `.status`, so it tolerated both — the pair held because the FE half was written defensively. **RB-19 verification, recorded because RB-12's test cannot do it:** RB-12 proves a `.Gate(...)` marker is present, not that it matches the wrapper the handler calls (its own stated declaration-vs-derivation limit). Checked centrally instead — the sorted list of all 47 route strings is identical before and after, **and so is every (route, `.Gate` marker, wrapper actually called in the handler) triple**, with zero gate/handler mismatches. `gen:api` produced an ordering-only diff in `swagger.json` + `api-client.ts` (only the two moved _and documented_ endpoints changed position; the other three moves are `.ExcludeFromDescription()`), committed rather than left to fail the drift job. | | 5 | RB-24..RB-30 | **complete** | All seven merged, one commit per ticket. `npm run ci` green on the combined tree after every merge. Ran as three waves, not the two the backlog implied: RB-24 rewrites imports in `brief.store.ts` and `org-template.store.ts`, which are two of RB-28's three targets — a dependency the backlog's "25/26/27 depend on 24" note never mentioned. **A** = RB-24 alone (the move), then **A2** = RB-29 + RB-30 in parallel (backend, no file overlap with the move or each other), **B** = RB-25 + RB-26 + RB-28 in parallel once RB-24 landed, **C** = RB-27 alone last, since it depends on RB-25's transport token. **RB-24 expanded its own scope, correctly.** Deleting the dependency-cruiser carve-out — the ticket's own acceptance criterion — exposed a second, real `ui-not-infrastructure` violation the old path had hidden: three UI components injected `UploadAdapter` for nothing but a one-line wrapper over its own exported pure function. The dispatch prompt said to report a second violation, not fix it; the agent judged this one was on the critical path (`dep:check` cannot pass with the carve-out gone otherwise) and fixed it minimally, reusing the existing pure function. Reviewed before merging — sound. **Two more findings were shown to be stale or overstated, on top of the two ADR-fixes found wrong and RB-18/RB-23's incompleteness from batch 4 — nine total now.** RB-25 found TE-003 overstated its own blocker: of the four methods named, only `upload()` and `cancel()` were actually unfakeable through the missing token — `delete()`/`pollReturning()` already went through the exported `UploadAdapter`. RB-28 found TE-006 already false at the time it was written: `brief.store.spec.ts` already had a `previewLetter` success test via jsdom's spyable `URL`/`window` stubs, contradicting the finding's "cannot test the success case" claim — the overall three-site diagnosis still held and was shipped as instructed. **RB-26 made one real design call**, reviewed before merging: `planFileSelection` must return `UploadMsg[]` per its literal signature, but an accepted file's real `localId` needs `crypto.randomUUID()`, which the ticket itself keeps in the controller. It ships a placeholder `localId: ''` discriminated by `.type` alone and never dispatched — verified the index alignment holds for both the multiple-rejection short-circuit and the per-file path. **RB-27 left one thing unextracted, correctly**: TE-005 lumped abort-vs-error disambiguation into the same extraction as `uploadOutcome`, but abort fires on a different event with no `status`/`responseText` at all — it structurally cannot fit the proposed signature. Left in place as a one-line ternary. The optional `currentScenario()` move into `KeepaliveTransport` was also correctly declined — it would have crossed into `upload-shell.service.ts`, outside this ticket's stated single-file scope. **End state of `libs/shared/upload`** (now split across proper layers): every layer that can hold pure logic has one and is spec'd — `upload.machine.ts` (domain, `planFileSelection`), `upload-shell.service.ts` (application, the `UPLOAD_TRANSPORT` seam), `upload-controller.ts` (application), `upload.adapter.ts` (infrastructure, `uploadOutcome`). Only the XHR/DOM boundary itself stays untested by design — TE-005 was explicit that abstracting `XMLHttpRequest` away is not wanted, since the file documents why XHR (not `fetch`) is required. | | 6 | RB-31, RB-32, RB-33 | **complete** | All three merged, one commit per ticket. `npm run ci` green on the combined tree after every merge. Dispatched as one wave — no file overlap at all (four machine specs in three apps, one docs file, one testing helper plus its one call site). **RB-31 found a real ADR-0006 violation, not a false alarm.** Two `registratie-wizard` tests asserted a state — cursor 2, no diploma chosen — that the real reducer cannot produce, since advancing past `beroep` (cursor 1→2) requires `KiesDiploma`/`KiesHandmatig` to have already run. This is exactly what forcing fixtures through message replay is for: a hand-rolled literal let an impossible state sit in the suite undetected. Fixed by replaying to cursor 1 and applying the diploma choice there instead; `submit()` validates the whole draft regardless of cursor, so the assertions are byte-identical to before — reviewed the diff before merging to confirm the old and new test bodies check the same thing. `intake.machine.spec.ts` also had an existing, correct `intake.testing.ts` sitting unused in its own folder, imported only by the acceptance spec — now wired to both. **RB-32 added the missing `language-switcher` row and took the ticket's explicitly-optional second step**: a ~14-line drift guard in `check-tokens.sh` that diffs every `CIBG-GAP EXTENSION` marker's component directory against the register's rows and fails naming what's missing. Verified the regex before trusting it — two existing rows carry parenthetical suffixes (`wizard-shell (error summary only)`) and the extraction correctly captures only the backtick-quoted name. This closes ADR-0003's own predicted failure mode ("if markers and this table drift, trust the code and fix the table") permanently rather than fixing it once more. **RB-33 made the real adopt-or-delete call the finding asked for, and chose delete.** `unwrapOk` had zero consumers anywhere in the repo since it shipped; manufacturing a first caller purely to satisfy the ticket would have removed no actual duplication, since there was only one occurrence to begin with. Deleted the helper and its doc mention; left the one candidate call site's inline guard alone, since it already satisfies ADR-0006 §3's real requirement (never a cast). | | ADR-fix | ADR-C-001, ADR-C-003, ADR-C-007, ADR-C-009 | **complete** | All four signed and landed by the architect on 2026-08-27, in one commit; doc-only, no code touched. Three carried the mandatory matching `CLAUDE.md` edit in the same diff (§4 twice, §2 once). **ADR-C-009's RB-07 gate was satisfied first** — all four clauses of its new test were verified against both `OrgTemplateStore` and `FeatureFlagStore` before signing, so the ADR does not ratify a control the code lacks. **Two findings were wrong and are corrected in the notes:** ADR-C-001 told us to keep an out-of-scope bullet reading "`SessionStore` is in-memory", which RB-10/RB-13 made false (the session now persists to `localStorage`; only multi-tab sync is still open), and ADR-C-007 flagged only the `.alert` half of ADR-0003's point 4 — its "header/side-nav use `.nav` + a local blue bar" clause is equally false (`site-header` composes the vendored `.titlebar`/`.logo__*`). ADR-C-007 also over-listed one path: `public/cibg-huisstijl/` never moved. ADR-C-003's open question was decided explicitly — **the 4 hand-written `contracts/*.dto.ts` stay**, because NSwag emits every property optional and flattens `RegistrationStatusDto` into five optional strings, which would make an illegal state representable (CLAUDE.md §3). Gates released: ADR-C-003 (contracts cleanup) and ADR-C-009 (a third runtime-editable surface). Still pending, untouched: **ADR-C-008 → RB-32** — 9 `CIBG-GAP` markers vs 8 register rows, missing row is `language-switcher`. | **Standing caveat for every batch:** `dotnet test` reports one failure, `OpenZaakIntegrationTests.Admin_cases_returns_the_seeded_zaak_mapped_through_real_HTTP_and_JWT`, which needs a live OpenZaak container. It fails identically on a stashed tree — it is not caused by any of these tickets. `npm run ci` does not run it. ## Gate integrity — read before trusting any "ci green" in this file `scripts/ci-local.sh` chained six of its steps as `cmd1 && cmd2` under `set -e`. Bash exempts every command of an AND-OR list **except the last** from `errexit`, so a crash in `cmd1` was swallowed: the paired check never ran and the script still printed "local CI passed". Verified directly — `bash -c 'set -e; false && true; echo hi'` prints `hi` and exits 0. This hid a **real** `gen:api` crash introduced by RB-09 (`dotnet swagger`'s design-time host defaults to Production, which RB-09 made throw). `.github/workflows/ci.yml` runs each step as its own `- run:` and would have caught it, so the local gate was strictly **weaker** than the remote one — the opposite of its purpose. The worst instance was `ng build ssp --localize && ng build behandelportal --localize`: a missing English translation in ssp could not fail the run. Fixed in `build: stop ci-local.sh swallowing the first half of every paired step`. **Every "ci green" recorded for batch 1 and for RB-07/RB-10/RB-11 predates that fix** and is therefore weaker than it reads; the batch-2 completion run above is the first one made on the honest gate (13/13 steps, exit 0). Nothing has since been found wrong with batch 1, but it has not been re-verified under the fixed gate either. ## Dispatching implementation agents — what actually goes wrong Batches 2 and 3 ran tickets as parallel agents in git worktrees. Six of seven agent-runs hit at least one of these. Put all of it in the prompt. 1. **The worktree base is not reliable.** **Four of the six** agents were handed a worktree branched from a stale ancestor — batch 3 was **three for three**, all landing on `ae7781e`, an unrelated lineage missing every RB ticket _and_ this backlog directory. Make step zero: `git log --oneline -8`, confirm a **named expected commit**, `git merge` the target branch if absent, and report which it was. The one agent that was not told to do this found out by luck. 2. **Agents park on background tasks.** Two agents in batch 2 ran `npm run ci` in the background, then ended their turn waiting for a notification that never usefully arrived; one finished its work twice and never committed it. Ban `run_in_background` and Monitor, **and say explicitly to pass the Bash tool's own `timeout: 600000`** — a batch-3 agent still auto-backgrounded because "run it in the foreground" alone does not defeat the 120s default. 3. **`git checkout ` destroys the work.** Agents verify a test is red by undoing the fix; restoring it with `git checkout` reverts the whole file. Tell them to undo and redo with edits. 4. **`behaviour-spec.mdx` conflicts on nearly every merge.** It is generated and every agent regenerates it. Resolve centrally with `npm run gen:behaviour-spec && git add`; tell agents to expect it and never hand-edit. 5. **Concurrent `dotnet test` runs flake** — a burst of `SQLite Error 1: 'no such table: Documents'`. Partly host pressure, but RB-12 found a real cause: a bare `new WebApplicationFactory()` races on the static `Db.ConnectionString`. Use the house `TestWebApplicationFactory` + `IClassFixture` idiom. Do not trust a backend green while another agent is testing. 6. **Agent worktrees live inside the repo**, so `prettier --check .` walks into them — fixed by ignoring `.claude/worktrees/` in both `.prettierignore` and `.gitignore`. 7. **The stale base is now the rule, not the exception.** Batch 4 ran six agents; **five were handed a bad base**, three of them the same unrelated `ae7781e` lineage. Across batches 2-4 that is **11 of 13 agent-runs**. Every one self-corrected at step zero. Keep the named-expected-commit check as the first instruction in every prompt — it is the highest-value line in there. Give the _current_ HEAD, not the batch's starting commit, when tickets merge sequentially. 8. **A spend limit can kill every agent mid-flight; resume, do not restart.** All four wave-A agents died on an org monthly-spend 429, three of them at the final CI re-run. Their worktrees kept the uncommitted work intact. Sending each agent a message resumed it from its own transcript and it finished from exactly where it stopped — nothing was redone. Check `git -C status` before assuming work is lost. 9. **The `99-backlog.md` conflict is mechanical and has a fixed recipe.** It fired on three of five merges. Cause: the central prettier pass reflows the table's column widths, so the whole table conflicts even though the two sides differ in only one or two status cells. Recipe: take HEAD's table, flip the incoming ticket's cell, re-run prettier. Verify by parsing both sides cell-by-cell and printing only the differing cells before discarding either side — do not eyeball a 33-row table. A ticket's cell reads `**done**` once merged (matching RB-01..RB-17), not `implemented`. 10. **For a zero-semantic-change commit, ask for evidence CI cannot give.** RB-19's diff is 181 lines in `Program.cs` and no test can prove it changed nothing. The sorted-route-list diff, plus the route/gate/handler-triple comparison, is what actually made it reviewable. Ask for it in the prompt and re-run it centrally before merging. 11. **`gen:api` is not always a no-op on a pure reorder.** OpenAPI operation order follows mapping order, so a reorder legitimately changes `swagger.json` and `api-client.ts`. Tell the agent to prove the diff is ordering-only (sort every line of both versions, diff, expect empty) and to commit the regenerated pair, or CI's drift job fails on a correct change. 12. **The backlog's own "depends on" column is not exhaustive — check actual imports before parallelizing a wave.** Batch 5's table said only "25/26/27 depend on 24"; it never mentioned that RB-24 rewrites imports in two of RB-28's three target files. `grep -rln` for the moved module's import path against every other open ticket's target files, before deciding what runs in parallel — not after a conflict. **Telling agents to report a ticket as wrong pays off.** Three did: BIO-012 was factually wrong about the proefbrief error mapping (RB-11), RB-12's wrapper/public binary did not fit the code, and RB-14 as worded would have shipped a non-gate. None of the three would have been caught by a review of the diff alone.