[kyle-sexton]

Refs: #967

Summary

Two callers on one pull request share the reviewed-head marker lane=claude-review. A docs-only skip by the legacy pr-review.yml recorded a head that pr-review-cli.yml then skipped as already reviewed, so a workflow change merged without code review on pr-pipeline-sandbox#273.

This PR delivers the reusable-side fix. A marker now counts as a prior review only for the caller that wrote it.

Fix

• publish.cjs always writes the second marker line, now with the caller workflow path from GITHUB_WORKFLOW_REF: <!-- claude-lane-reviewed-base ref=main merge-base=<sha|none> caller=.github/workflows/review.yml -->.
• The scope step in pr-review-cli.yml (all three legs) reads only a marker naming its own caller.
• The scope steps in pr-review.yml and pr-review-security.yml ignore a marker naming a caller, so a CLI skip cannot narrow a legacy review.

Why the caller is on line 2:

• Line 1 stays unchanged because other repositories match it with anchored patterns: clear-escalations.mjs:517, address-feedback.mjs:75, prior-findings.mjs:68, babysit_review_evidence.py:264.
• The only line-2 consumer outside this repo, prior-findings.mjs:93, matches by prefix.
• The issue's other option, a format check on line 2 without a caller, does not work: pr-review.yml on main also writes that line when it has a merge base.

Transition: an open pull request's existing CLI marker names no caller, so its next push gets one whole review and a new lane comment. That cost is one review per open PR.

Verification

• New tests in claude-lane-scope.test.cjs cover the issue's fixture (CLI marker at A, newer legacy marker at the head, reviews incrementally from A), a legacy marker alone, another caller's marker, the caller's own marker still skipping, and both legacy lanes ignoring a marker naming a caller.
• Negative control: with the three workflows restored from origin/main, all 6 new tests fail.
• node --test passes on the scope, publish, CLI, API-call, completion and script-body suites.
• GitHub documents that in a reusable workflow "the github context is always associated with the caller workflow" (docs). That the GITHUB_WORKFLOW_REF variable follows the same rule is checked on this PR's own review lanes only after merge, through the sandbox step below.

Left out: the live sandbox re-run, because the sandbox pins a released reusable (see Remaining).

Remaining

• Re-run the sandbox caller after this merges and the sandbox moves its pin: on a pull request with a code push after the first review, pr-review-hosted reviews the new head while pr-review-no-bots runs. fix(pr-review-cli): a legacy pr-review.yml marker for lane claude-review skips the code review #967 closes on that result.

Related

• Refs #8994: this PR is the vehicle for that issue's live ready probe.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VQb9Q1VgLEiwGgRVshK3b7

---

^{Open in [code]smith · Need help on this PR? Tag @codesmith-bot with what you need.}

[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 2026-11-03T04:04:40.066197Z	e50d559	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-security-review reviewed e50d5598ce82095071b46019300eb631826e7394: 0 finding(s), 0 as inline comments.

No security findings in PR #972. The skip gate does not apply, because the change touches review-scoping and skip logic, so I read the whole diff.

What changed. The claude-lane-reviewed-base marker line now also records a caller= value, taken from GITHUB_WORKFLOW_REF with whitespace replaced. pr-review-cli.yml reads only markers that name its own caller. pr-review.yml and pr-review-security.yml ignore any marker that carries a caller=. The tests and README are updated to match.

What I checked.

• Who can write the marker. The comment.user?.login === "github-actions[bot]" filter is unchanged in all five scope steps. A PR author cannot forge a marker that makes the lane skip or narrow a review.
• Where caller comes from. It comes from the workflow ref of the calling workflow, which is not PR-body text. The \s replacement and the (\S+) regex keep it to a single token. A PR that renames the caller workflow only makes the lane fail safe, since it finds no matching marker and reviews the whole pull request.
• Legacy markers. In pr-review-cli.yml, a marker with no caller, an unparseable second line, or another caller's name no longer counts as a record. The result is a whole review, not a skip. The test "a newer legacy marker at the head neither skips nor narrows the review" covers this.
• Model-written summary text. The marker body also holds the model's summary, which comes from untrusted PR content. The /^<!-- claude-lane-reviewed-base .* caller=/mu test in the legacy workflows scans the whole body, so a summary line shaped like that header could make a legacy lane discard its own marker. That causes a whole re-review, not a skipped one. In pr-review-cli.yml only line 2 is matched, and publish.cjs writes that line from fixed values.
• base dereference. BASE_LINE.exec(lines[1]) is called after the filter has already matched that line, so base cannot be null.
• Trust controls. I read nothing that widens a trusted-actor list, token scope, kill switch, merge hold or isolation. publish-review/publish.cjs and the workflow permissions: and triggers are otherwise unchanged. The publish.cjs edit only writes the extra marker field.

I did not read docs/contracts/instruction-exception-register/README.md, since no standing instruction is removed or softened. The workflow scripts' marker handling is covered by the zizmor lane, and nothing here needed deeper reasoning.

manifest fields: none hit (no plugin manifests changed).

critical-findings: 0

[github-actions]

claude-review reviewed e50d5598ce82095071b46019300eb631826e7394: 0 finding(s), 0 as inline comments.

Code review: no findings

I read the full diff (9 files) and the PR context. I did not run the test suites. The PR body says they pass, but that is the author's claim and I haven't verified it.

What I checked

• Caller value on both sides. publish.cjs and the three scope steps in pr-review-cli.yml derive the caller from GITHUB_WORKFLOW_REF the same way (drop @ref, drop owner/repo/, replace whitespace with _). The scope steps read process.env directly, and the publish step runs as a composite action, so the variable is present in both places.
• Missing variable. If GITHUB_WORKFLOW_REF is unset, the marker is written as caller=. The scope regex needs caller=(\S+), so it never matches and the lane reviews the whole pull request. That is a safe failure.
• merge-base=none. The scope step maps it back to an empty merge base, the same as the old no-base-line case. The last-summary slice moves from "skip 2 lines, then 1" to slice(3), which is the same offset.
• marker-id. It is now set only for the caller's own marker, so a different caller's marker is never updated in place. A new comment is created instead, as the PR's transition note says.
• Legacy lanes. pr-review.yml and pr-review-security.yml still write the old second line without a caller and still use their old BASE_LINE. Their added filter ignores any marker whose base line names a caller. Neither legacy workflow uses the publish-review action, so they never write a caller marker themselves.
• Tests. The new scope tests cover the fixture from the issue, a legacy marker alone, another caller's marker, the caller's own marker still skipping, and both legacy lanes ignoring a caller marker. The existing fixtures that now need a caller on line 2 are updated.
• Closing issues. The PR uses Refs: #967, so no closing issue is read and no acceptance criteria apply.
• Path-scoped rules. The only rule file, pr-body-contract.md, covers the PR body. The body opens with Refs: and has the required sections.

Unverified: that GITHUB_WORKFLOW_REF carries the caller's path inside a reusable workflow. The PR acknowledges this and defers the live check to the sandbox re-run after merge. If the variable named the reusable instead of the caller, two callers would again share one marker value, so this check matters.

Not flagged (cosmetic): the new README sentence leaves one line well past the surrounding wrap width.