[kyle-sexton]

Refs: melodic-software/claude-code-account-rotation#259

Summary

Moves the live code review lane onto ci-workflows pr-review-cli.yml v0.63.1 (83b739f), lane code, funded by the reviews role: REVIEWS_CLAUDE_API_KEY first, REVIEWS_CLAUDE_CODE_OAUTH_TOKEN only after the key hands off (or when the repository variable REVIEWS_KEY_EXHAUSTED is true). This mirrors the security lane's switch in #9307. Both review lanes now run on pr-review-cli.yml.

Fix

• .github/workflows/pr-review-hosted.yml: calls pr-review-cli.yml@83b739f with lane: code, publish: comments, the same model-args, allowed-bot-sender-ids and allowed-author-ids, and docs-only-paths: '' passed explicitly (pr-review.yml's default at 2fe38c0 was empty, so the code lane never skipped documentation; the three security globs are not copied). Job id stays pr-review, so the check stays pr-review / claude-review-status. Permissions are contents: read and pull-requests: write (no id-token); runner is dropped (the new reusable has no such input). cancel-in-progress: false, since a cancelled run reads as red to the guardrails merge guard. The bare CLAUDE_CODE_OAUTH_TOKEN is no longer passed.
• review-venue.sh: a pr-review-cli.yml caller whose path does not contain security is now the code-review lane (before, no pr-review-cli.yml run counted for code-review, so it would have read local absent after this switch). The reference table says the same.
• babysit_review_evidence.py, address-feedback.mjs and feedback-dispatch.mjs already accept any review lane's github-actions[bot] footer and an account=fallback record since feat(pr-pipeline): run the security review on pr-review-cli with key-first funding #9307, so they need no logic change. They gain code-lane tests (claude-lane: claude-review footer, account=fallback record for claude-review), and comments that named only pr-review.yml or called claude[bot] the code lane's reviewer are corrected.
• .github/lane-funding.json: the pr-review lane is run-claude-cli reading both REVIEWS_* secrets.
• docs/conventions/pr-pipeline/README.md, pr-run-activity.md, and changelog fragments for pr-pipeline and source-control.

Verification

Live proof on head b1a70599799668bc4621981158e438b00d65bc3f, run 39910029055 (REVIEWS_KEY_EXHAUSTED not set):

• pr-review / claude-review-key: run-claude-cli: pr-review ended class=auth on REVIEWS_CLAUDE_API_KEY api_error_status=401, "handoff":true. The key itself did not review: the API refuses it.
• pr-review / claude-review-subscription: Claude review completed: credential=REVIEWS_CLAUDE_CODE_OAUTH_TOKEN fallback=auth turns=14 max-turns=75 models=claude-sonnet-5-5.
• pr-review / claude-review-status: notice titled claude-lane-review-complete, claude-lane-review-complete lane=claude-review sha=b1a70599799668bc4621981158e438b00d65bc3f findings=0 inline=0 scope=whole account=fallback; the summary marker <!-- claude-lane-reviewed-head lane=claude-review sha=b1a70599… --> was posted by github-actions[bot] (type Bot). Every job green.
• babysit's evidence reader (babysit_merge._lane_evidence, as the gate calls it, with changed_paths=[] because this PR edits .github/ and the lane-source gate would unbind it): claude-review bound, 0 open findings; claude-security-review bound, 0 open, 0 blocking.

Local, one suite at a time under nice:

• Node: callers.test.mjs 131 pass; address-feedback.test.mjs 89; pr-unblock/feedback-dispatch.test.mjs 77; resolve-config/feedback-dispatch.test.mjs 6; check-trusted-trigger.test.mjs 332; 0 fail.
• Python: test_babysit_review_evidence 137 OK. review-venue.test.sh 57 pass.
• The new review-venue test fails on the base script (code-review read local absent) and passes after. The new evidence-reader, address-feedback and feedback-dispatch code-lane tests pass on base too, since feat(pr-pipeline): run the security review on pr-review-cli with key-first funding #9307 made those readers lane-generic; each was shown to fail when its footer rule is narrowed to the security lane, then restored.
• runner-policy.mjs --root ., actionlint, zizmor (offline), pinned ruff, markdownlint and both changelog fragment gates pass; CI lint-repo passed on this head. ci-status is red only for the do-not-merge label.

Related

• feat(pr-pipeline): run the security review on pr-review-cli with key-first funding #9307 (the security lane's switch)
• feat(runner-policy): register the ci-workflows v0.63.1 pr-review-cli contract standards#849 and chore: sync standards components #9335 (runner-policy contract for v0.63.1)

🤖 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 2026-10-23T06:13:28.810760Z	b1a7059	PR opened

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-security-review needs no new review of 611e6360cc2e90ea9ad7d3fdf1da44e8cc3947ef: no file of this pull request changed since the review of b1a7059.

Security review of PR #9348: no security findings.

critical-findings: 0

I read the full diff (14 files) and the PR context. The PR context contained no embedded instructions. The skip gate did not apply, because the diff changes a workflow caller and review-trust logic.

What I checked:

• Workflow caller (pr-review-hosted.yml):
  • The reusable workflow is pinned to a full commit SHA (83b739f…, v0.63.1).
  • permissions: is narrower than before. id-token: write is dropped, and the job keeps only contents: read and pull-requests: write.
  • Secrets are passed by name (CLAUDE_API_KEY and CLAUDE_CODE_OAUTH_TOKEN, from the REVIEWS_* secrets). There is no secrets: inherit, and the bare CLAUDE_CODE_OAUTH_TOKEN is no longer forwarded.
  • The trigger set is unchanged, and allowed-bot-sender-ids and allowed-author-ids are carried over unchanged, so no trusted-actor widening.
  • cancel-in-progress: false and docs-only-paths: '' cause no security regression. The reusable still skips fork and Dependabot runs, and the lane reviews documentation changes as it did before.
• Review-bot trust (feedback-dispatch.mjs, address-feedback.test.mjs, babysit_review_evidence.py and tests):
  • Only comments and tests changed. The logic that accepts github-actions[bot] comments (footer on the last line, reviewed-head marker from the same bot id) already existed from feat(pr-pipeline): run the security review on pr-review-cli with key-first funding #9307.
  • The new tests pin the existing footer-last, wrong-lane and no-footer rejections for the claude-review lane. Trust did not widen.
• review-venue.sh:
  • The lane is now chosen with (.path // "" | test("security")) == $sec. It reads workflow-run metadata and decides only whether CI covers a lane or a local review is needed. It runs no input and has no injection surface.
  • $sec is passed through --argjson from a fixed array of false/true, so no caller-controlled data reaches jq code.
  • A caller whose path contains "security" but serves the code lane would be misread as the security lane. That affects venue health reporting only and gates nothing.
• Plugin manifests: none changed, so there are no manifest <field> lines.
• lane-funding.json: the change is metadata that lists the lane's secrets and launcher. It grants nothing.
• Instruction-surface deletions: the removed workflow-header text is rationale comments. No guardrail was removed.
• Trust-gate paths: the diff touches no trusted-actors, check-trusted-trigger, CODEOWNERS or .gitmodules file, so no hold applies.

The protected-class register was not consulted. Nothing in the diff removes a standing rule, so it was not needed.

[github-actions]

claude-review needs no new review of 611e6360cc2e90ea9ad7d3fdf1da44e8cc3947ef: no file of this pull request changed since the review of b1a7059.

No findings. I read the full diff (14 files) and the PR context. I did not run any tests or scripts, so everything below comes from reading the code.

• pr-review-hosted.yml: The caller now matches the already-merged security caller (pr-review-security-hosted.yml) on the pinned pr-review-cli.yml@83b739f (v0.63.1). That covers publish: comments, the pull-requests: write permission with no id-token, cancel-in-progress: false, and the two REVIEWS_* secrets mapped by name. docs-only-paths: '' is set explicitly, so the code lane keeps reviewing documentation.
• review-venue.sh: The jq now treats a pr-review-cli.yml caller as code-review when its path has no security, and as security-review when it does. I traced both lanes' selection logic. The existing test, where a code-lane caller of pr-review-cli.yml is listed beside the pr-review.yml run, still resolves to the same caller. The new test covers the case where the security caller is listed first.
• Closing issue: The PR says Refs: an issue in another repository and closes none, so there were no acceptance criteria to check.
• Manifests: No plugin.json or marketplace.json changed, so no manifest lines apply. Both changelog fragments use bump: minor, which fits the behavior change.
• PR context text: I found no instructions embedded in the PR text.
• Not reviewed: I did not read the changed .github/lane-funding.json against its schema or against the secret setup in repo settings. I also did not check that the pr-review / claude-review-status check name carries over, which depends on the reusable at the pinned SHA.
• Cost note, not a finding: The code lane now spends the reviews role's API key first. There is no hold field in this output schema, so none is set.