[kyle-sexton]

Closes #950
Refs: melodic-software/claude-code-plugins#7777

Summary

claude-code-plugins#12011 added plugins/review/lib/prior-findings.mjs and taught /review:code-review and /review:security-review to act on a PRIOR FINDINGS: prompt line. Nothing produced that line. This runs the script on every model leg of pr-review-cli.yml, for both lanes, and passes its note to the compose prompt the same way the settled-threads note is passed.

Fix

• New composite .github/actions/collect-prior-findings (one bash step, no model, no github-script, so no context shadowing as in fix(report-lane-outcome): stop shadowing github-script's context parameter #952). It:
  • removes any head-shipped .claude-lane/prior-findings.md;
  • reads the round cap file with gh api -X GET repos/$REPOSITORY/contents/docs/conventions/pr-pipeline.yaml -f ref="$DEFAULT_BRANCH" (raw), where DEFAULT_BRANCH is github.event.repository.default_branch, never the head or base SHA. Any failed read (404 included) leaves CAP_CONFIG unset, so the script counts no cap;
  • runs node "$SOURCE/plugins/review/lib/prior-findings.mjs" from the claude-code-plugins default-branch clone the plugin step already staged, with the job token (contents: read, pull-requests: read, unchanged);
  • on a script failure removes the output file and fails the step; the script writes note= only after a successful write, so the compose prompt gets no line and the lane reviews as before.
• pr-review-cli.yml: in review-key, review-subscription and review-fallback, a prior step after settled with the settled step's if:, continue-on-error: true and timeout-minutes: 2; lane maps code/security to claude-review/claude-security-review, the footer names the script matches. Compose gets PRIOR_NOTE: ${{ steps.prior.outputs.note }} and env.PRIOR_NOTE right after env.SETTLED_NOTE.
• Each leg job's limit rises from 93/101 to 95/103 minutes (code/security), so the job still exceeds its steps' sum by 5 as pr-review-cli.test.cjs requires.
• SECURITY MODEL header and README list the new file and action.

Verification

• node --test .github/scripts/*.test.cjs: 793 pass, 0 fail (4 of them new in collect-prior-findings.test.cjs: cap read from the default branch with the exact gh arguments; unread cap or empty default branch leaves CAP_CONFIG unset and still writes the file; a failed script leaves no file and no note, including a head-planted one; every leg wires the step before compose with read-only permissions).
• The script's GraphQL query, sent live with gh api graphql against claude-code-plugins#12011, returns no errors and github-actions with databaseId 41898282 on reviews and comments (the fix(collect-settled-threads): read resolvedBy as the User it is #953 failure class).
• zizmor on the action and workflow: no findings. shellcheck on the composite's run block: clean. Local actionlint 1.7.12 reports only the classes it already reports on main ($/ local action refs, cache-mode); CI's pinned actionlint is the authority.
• After merge: the next lane run's log should show prior-findings: N prior post(s), R round(s), cap 3 on a claude-code-plugins PR.

Related

• fix(report-lane-outcome): stop shadowing github-script's context parameter #952 and fix(collect-settled-threads): read resolvedBy as the User it is #953 (the two lane fixes this step avoids repeating).
• claude-code-plugins#12011 (the script and the skills' "Prior findings" rule).

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-01T18:42:36.029972Z	7495c6c	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[github-actions]

claude-review reviewed 7495c6c3adea7ad9354177fd424f5baadadda3db: 0 finding(s), 0 as inline comments.

No findings.

I read the whole diff (4 files), the PR context, and the existing collect-settled-threads step and the matching settled steps in pr-review-cli.yml for comparison.

• Wiring: The prior step is added to all three legs (review-key, review-subscription, review-fallback) with the same if:, continue-on-error and timeout-minutes: 2 as the settled step. Each leg passes PRIOR_NOTE into compose and adds env.PRIOR_NOTE right after env.SETTLED_NOTE. The job limits rise from 93/101 to 95/103 minutes, which matches the 2-minute step.
• Round cap: action.yml reads the cap from DEFAULT_BRANCH, never the head. Any failed read leaves CAP_CONFIG unset, so no cap applies.
• Planted file: The step removes a head-planted prior-findings.md before running. It removes the file again, and fails the step, when the script fails.
• Tests: The new tests cover those three behaviours and the wiring on every leg.
• Local action ref: uses: $/.github/actions/... is the same form the existing settled step uses, so it is not a new problem.
• Not verified: I could not read plugins/review/lib/prior-findings.mjs, because it is not in this checkout. I could not confirm that the env names the action passes (OUTPUT_PATH, LANE, CAP_CONFIG, GITHUB_TOKEN) are the ones the script reads. The settled action uses OUTPUT_FILE, a different script. The PR says the script was run live, but that is the author's claim and I did not confirm it. I also did not run the tests; the "793 pass" count is likewise the author's claim.
• Closing issue feat(pr-review-cli): run prior-findings.mjs and pass its note to the review lanes #950: Its body was omitted from the context because the author is not on the allowed list. I treated it as unread, so I could not check the change against any acceptance criteria.

[github-actions]

claude-security-review reviewed 7495c6c3adea7ad9354177fd424f5baadadda3db: 0 finding(s), 0 as inline comments.

No security findings. critical-findings: 0

What the PR does: it adds a collect-prior-findings composite action and wires it into all three model legs of pr-review-cli.yml (review-key, review-subscription, review-fallback). It also adds a test, a README entry and a 2-minute timeout bump. The action runs prior-findings.mjs from the plugin clone and adds a PRIOR FINDINGS: note to the compose prompt.

Trust-control and trust-boundary checks:

• Round cap: pr-pipeline.yaml is read through the contents API with ref=$DEFAULT_BRANCH, never the head. A failed or empty read leaves CAP_CONFIG unset, which means no cap and the review still runs. This does not weaken a merge hold: it only feeds the review prompt and no gate script consumes it.
• Script source: SOURCE is ${{ runner.temp }}/claude-code-plugins, the clone of the default-branch tip that the plugin step already stages. The step is gated on steps.plugin.outcome == 'success'. The head cannot supply the script.
• Planted output file: rm -f -- "$OUTPUT_PATH" runs before the script and again after a failure. A file the head ships at .claude-lane/prior-findings.md never reaches the model, and OUTPUT_PATH is a fixed literal from the workflow.
• Token scope: the action gets github.token as GH_TOKEN and GITHUB_TOKEN. The jobs' permissions are contents: read and pull-requests: read, which the new test asserts, so no scope grows and no secret or App token is added.
• Injection: the run: block interpolates nothing from the github context directly. Every value goes through env:, so there is no expression injection.
• Prompt text: the new PRIOR_NOTE path is a fixed file name. The script writes the note only on success, and the file is labelled DATA in the skill. The posts it lists are limited, per the workflow header, to bot-authored posts carrying the lane footer. I could not read prior-findings.mjs here (it lives in claude-code-plugins), so that filter rests on the header comment and the skill text, not on code I read.
• Failure behaviour: continue-on-error: true fails open, but only to the previous behaviour (a review with no prior findings).
• Timeouts and charges: the timeouts go from 93/101 to 95/103 minutes, and the step timeout is 2 minutes. This starts no billed resource.
• Manifests, .github/actions/* trust controls and instruction surfaces: the diff touches no plugin manifest, trusted-actor list, kill switch, token broker, merge-hold script or standing instruction.

Not reported: uses: $/.github/actions/collect-prior-findings matches the existing collect-settled-threads step's form, so I treated it as the repository's convention. Action pinning and workflow hardening are zizmor's lane.