[kyle-sexton]

Refs: melodic-software/claude-code-plugins#9307

Summary

Two pr-review-cli publish fixes that block claude-code-plugins#9307 from moving its security lane onto pr-review-cli:

• The claude-lane-review-complete notice had no annotation title. pr-review.yml and pr-review-security.yml emit ::notice title=claude-lane-review-complete::, and babysit's evidence reader keeps only annotations with that title. Without it, every pr-review-cli review read as no-record.
• When one bad anchor demotes the whole inline batch to the summary, the model's critical-findings: line still left those findings out. A demoted CRITICAL finding then counted as non-blocking.

Fix

• publish.cjs: the three completion notices (whole or incremental review, skip=no-change, skip=docs-only) pass { title: "claude-lane-review-complete" }.
• countedSummary: on a demoted batch, it adds the moved CRITICAL findings to the model's one critical-findings: line, changing that line in place. A finding counts as CRITICAL unless it has exactly one IMPORTANT or SUGGESTION severity: line, which is how the gate reads a thread. With no line, it adds one that counts every summary finding as CRITICAL. With two or more lines it leaves the text alone, because a gate already counts every finding then.

Not changed: the skipped jobs' names. When a pr-review-cli job is skipped, its ${{ inputs.lane ... }} name renders as the raw expression. Literal per-lane job names would emit a skipped wrong-lane check (pr-review-security / claude-review-status) on every run. babysit's AI_REVIEW_CHECKS matches claude-review-status on the last segment, so that check would hold every PR. Each reader treats a skipped AI review check the same as a missing one (babysit_merge.py ai_review_holds, merge.md lane()), so the raw name changes no gate.

Verification

• node --test .github/scripts/publish-review.test.cjs: 17 pass. Three of the new tests fail on main: the title on all completion notices, a demoted CRITICAL count raised in place, and a line added when missing.
• node --test .github/scripts/claude-lane-*.test.cjs .github/scripts/publish-review.test.cjs: 201 pass.

Related

• Follow-ups, not in these lines: the inline comment footer is duplicated when the model writes its own, and the review-key leg skips a human-pushed head on a Dependabot-authored PR.

🤖 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 2026-10-23T04:53:10.186476Z	1f41a44	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.

[claude]

Claude finished @kyle-sexton's task in 2m 25s —— View job

---

Security review

Reviewed the three changed files against origin/main...HEAD (1f41a44).

publish.cjs

• RECORD = { title: "claude-lane-review-complete" } passed as the core.notice annotation-properties argument (static string, no injection surface).
• SEVERITY / CRITICAL_LINE regexes are anchored, linear (no nested quantifiers), and bounded (\d{1,6}, MAX_BODY=8000, MAX_SUMMARY=60000) — no ReDoS concern, and matchAll's spec-mandated regex clone means the shared g-flag regex objects don't leak lastIndex state across calls (.replace also resets lastIndex before use).
• countedSummary's fallback paths (no existing critical-findings: line, or an anchor-problem's extra unitemized findings) are conservative — they add to the critical count rather than suppress it, so a parsing edge case fails toward holding the merge gate, not bypassing it.
• comment.body / review.summary are always coerced through stripHtml's String(text ?? ""), so a malformed/non-string model field can't throw before the existing secret-scan gate (scan(findSecret, …)) runs over the final summary text — that gate still covers the new code path unchanged.
• The second core.notice argument is new but additive; it doesn't change what's already scanned or posted.

pr-review-cli.yml: diff is comment-only (documents the new notice title and the demoted-count behavior); no job, trigger, permission, or run: changes.

publish-review.test.cjs: new tests only, no production logic.

No .claude-plugin manifest changes, and no deletion/narrowing of a standing instruction (CLAUDE.md/AGENTS.md/rules file) in this diff, so the instruction-surface-deletion and manifest-flag checks don't apply.

No security vulnerabilities found (logic flaws, auth/access-control gaps, injection, or credential handling). This is an internal CI-bookkeeping fix with no new trust-boundary or data-flow surface; GitHub Actions hardening concerns (triggers, permissions, pins) are unaffected and out of scope for this lane regardless.

critical-findings: 0

claude-lane-summary: lane=claude-security-review sha=1f41a44879912310dd405eca27aea18c6c575d6a run=39902588980

[claude]

Claude finished @kyle-sexton's task in 4m 9s —— View job

---

Code review: pr-review-cli publish fixes

• Gather context (diff, PR context, rules)
• Resolve standards index / paved paths (no paved-path index row found; no plugin manifest touched)
• Review publish.cjs changes
• Review publish-review.test.cjs changes
• Review pr-review-cli.yml changes
• Check closing issue criteria (no PR-context file / closing issue to check)
• Check path-scoped rules (.claude/rules/pr-body-contract.md — PR body opens with Refs: and has ## Summary/## Fix/## Verification/## Related: compliant)
• Post findings

Summary

No blocking or high-signal findings.

Traced both fixes against the test suite by hand (the sandbox denies node --test execution in this lane, so I could not run it, only read it):

• Notice title (publish.cjs lines 30-31, 355, 357, 473-475): all three claude-lane-review-complete notice sites (whole/incremental review, skip=no-change, skip=docs-only) now pass RECORD = { title: "claude-lane-review-complete" }. Grepped the file for every core.notice call carrying that record prefix — none were missed, and the unrelated shadow ... notices correctly stay untitled. This matches the title README.md already documented at L1312-1313, so the doc and code are now consistent.
• countedSummary/isCritical (publish.cjs lines 166-190): placeComments is all-or-nothing (every comment anchors, or the whole batch is demoted), so unanchored.length is always either 0 or review.comments.length — the rest = findings - comments.length computation in the "line missing" branch is sound against that invariant. Manually recomputed all four new/changed test cases (publish-review.test.cjs lines 489-572) against the implementation and they check out: the in-place raise, the added line, and the two-lines-left-alone case all match what the regexes (SEVERITY, CRITICAL_LINE) produce.
• pr-review-cli.yml's header comment update (lines 92-100) accurately describes both behavior changes; no functional workflow change in this file, consistent with the PR's own "not changed" note about job names.

Nothing in .github/actions/publish-review/publish.cjs or the workflow diff touches security-sensitive surfaces (secrets, permissions, trigger gating), so nothing here is withheld for the security lane either — there's simply nothing of that kind in this diff.

Author-claimed and unverified (I did not independently run these): the PR's own "17 pass" / "201 pass" node --test results.

claude-lane-summary: lane=claude-review sha=1f41a44879912310dd405eca27aea18c6c575d6a run=39902588980

[github-actions]

claude-security-review has reviewed this pull request through 1f41a44; a later push is reviewed from there.

[github-actions]

claude-review has reviewed this pull request through 1f41a44; a later push is reviewed from there.