Distillation

Date: Sep 10, 2026

• 🔴 (00:52) User stated the CI poller always re-adds the ci-ready label in .github/workflows/ci-poller.yml:356-359; waiting-for-CI removes stale ci-ready in .github/workflows/publish.yml:113-129.
• 🔴 (00:52) User directive: “Always run trusted code.” For the CI poller checkout, .github/workflows/ci-poller.yml:48-53 must use ref: ${{ github.event.repository.default_branch }} with persist-credentials: false.
• 🟡 (00:52) Independent review issued DO-NOT-MERGE after finding 2 MUST-FIX blockers and 3 security regression tests that passed for the wrong reason.
• 🟡 (00:52) MUST-FIX blocker 1: .github/workflows/auto-approve.yml:18-21 sparse-checks out only auto-approve-repos.txt, but .github/workflows/auto-approve.yml:28-36 executes src/publish/record-auto-approval-attestation.js. That source and its dependencies are absent, so automated approval fails with MODULE_NOT_FOUND. src/publish/__tests__/auto-approval-workflow.js:7-13 checks only textual ordering and does not verify checkout contents.
• 🟡 (00:52) src/publish/record-auto-approval-attestation.js depends on ../modules/approval-attestation.js and ./validate-approval-attestation.js; recordAutoApprovalAttestation() fetches the issue and rejects unless issue.state === "open", issue.title === issueTitle, and issue.user.login case-insensitively equals autoApprover, then calls createAutoApprovalAttestation({ autoApprover, issue }).
• 🟡 (00:52) MUST-FIX blocker 2: .github/workflows/ci-poller.yml:36-46 validates ATTEMPT with ^(0|[1-5][0-9])$, which accepts only 0 and 10–59 despite the stated range 0–59. The first run dispatches attempt 1 at .github/workflows/ci-poller.yml:453-463, so the next run rejects it and self-dispatch stops after the first retry. src/publish/__tests__/ci-poller-workflow.js:40-51 explicitly preserves the faulty regex.
• 🟡 (00:52) Security-test concern 1: the self-approval fixture in src/publish/__tests__/authorize-approval.js:96-107 omits state: "open", causing src/publish/authorize-approval.js:202-210 to reject at the open-state guard before exercising the intended self-approval guard; removing the self-approval check would therefore not fail this regression.
• 🟡 (00:52) Security-test concerns 2–3: CI-ready mutation fixtures in src/publish/__tests__/record-ci-ready-attestation.js:48-63 and :66-91 omit state: "open", so both short-circuit at src/publish/record-ci-ready-attestation.js:46-55 rather than testing title mutation or accepted-event replacement. Removing the intended title or event checks would not fail these tests.
• 🟡 (00:52) Review PASS: human approval requires write, maintain, or admin; automated requesters and direct Internal App authorization are denied by the generic path in src/modules/approval-authorizer.js:1-49, and src/publish/authorize-approval.js:186-219 denies case-insensitive self-approval after checking the live requester.
• 🟡 (00:52) Review PASS: Internal App auto-approval requires an allowlisted parsed path, approved bot requester, current accepted event, and request-bound comment authored by github-actions[bot], implemented in src/publish/authorize-approval.js:135-169.
• 🟡 (00:52) Review PASS: approval and CI-ready validation bind the live open issue, title, body/dry-run digest, current labels, accepted event, trusted comment author, and CI-ready actor in src/modules/approval-attestation.js:14-55, src/modules/approval-attestation.js:176-240, and src/publish/validate-approval-attestation.js:63-100.
• 🟡 (00:52) Review PASS: CI promotion captures the accepted event before polling, revalidates it after CI while recording proof, and validates again immediately before adding ci-ready in .github/workflows/ci-poller.yml:121-153 and :272-359.
• 🟡 (00:52) Review PASS: publish validates once at entry and again immediately before Craft in .github/workflows/publish.yml:197-220 and :328-354.
• 🟡 (00:52) Review PASS: failure cleanup removes accepted even when publish inputs are absent or malformed, via .github/workflows/publish.yml:409-414, src/publish/update-issue.js:8-18, and src/modules/update-issue.js:62-79.
• 🟡 (00:52) Review PASS: manual recovery uses the permissionless, secret-free workflow_dispatch relay behind production in .github/workflows/ci-poller-dispatch.yml:3-16. The privileged poller accepts only a successful default-branch relay run and checks out the default branch before token creation in .github/workflows/ci-poller.yml:8-31 and :48-74. The accepted organization-secret exposure is documented in README.md:97-99, and the diff does not widen its organization scope.
• 🟡 (00:52) Review PASS: changed workflows pin GitHub Actions to full 40-character SHAs and Craft to a full image digest in .github/workflows/publish.yml:191-223, :266-274, and :353-354; src/publish/__tests__/workflow-action-pinning.js:30-75 enforces this.
• 🟡 (00:52) Review verification results: yarn test --printConsoleTrace passed all 84 tests across 16 files; yarn lint passed in 1.54s; git diff --check passed with no output.
• 🟡 (00:52) Exact-worktree review inspected all 21 modified files and all 20 untracked files. The tree remained byte-for-byte unchanged: HEAD 83d210b29553ea6f4d97508821724a834c43856e; unstaged diff SHA-256 8ca4aa7afcb6a1c171f695d488a06b8cf4d08436a414a3223a18b33a08db4fc6; staged diff empty; every untracked-file hash matched the initial snapshot.
• 🟡 (00:52) Assistant planned the repair sequence: first make tests fail for their intended reasons; then fix sparse checkout and retry parsing; rerun each regression without its corresponding guard; complete formatting, full tests, lint, whitespace checks; and obtain a final independent exact-worktree review.
• 🟡 (00:52) Work queue, in exact order: 1. Add fail-first regression for auto-approval source checkout (in_progress, high); 2. Add fail-first regression for retry attempts 1–9 (pending, high); 3. Repair self-approval and CI-ready mutation fixtures (pending, high); 4. Fix auto-approval checkout and retry validation (pending, high); 5. Verify regressions fail without each guard (pending, high); 6. Run formatting, full tests, lint, and whitespace checks (pending, high); 7. Obtain final independent exact-worktree review (pending, high).
• 🟡 [requested-tests] (00:53) Assistant specified targeted regressions: require checkout of every file needed by node src/publish/record-auto-approval-attestation.js; test every canonical retry attempt from 0 through 59; and add valid open-issue snapshots to existing self-approval and CI-ready guard tests so they reach the intended branches.
• 🟡 (00:53) recordCiReadyAttestation() in src/publish/record-ci-ready-attestation.js:30-59 concurrently fetches the issue, events, and authenticated login; obtains currentAcceptedEvent(events); and throws "The approval changed before CI could be marked ready" if the issue is not open, title differs, accepted is absent, no accepted event exists, or the event actor/ID differs from expectedAcceptedEvent. On success it calls createCiReadyAttestation({ acceptedEvent, ciReadyActor, issue }).