Detect stale magic-close keywords in PR commit bodies

Context

PR #11183 empirically proved that stale magic-close keywords inside branch commit bodies can still close a GitHub issue even when the PR body has been corrected to Refs #N.

Observed sequence:

• PR #11183 originally had first commit deb022d0c with Resolves #11182 and stale AC wording.
• Later PR body and second commit changed the live framing to Refs #11182; gh pr view 11183 --json closingIssuesReferences returned [] before merge.
• PR #11183 was squash-merged as dev commit 5c7c5a2f4.
• GitHub auto-closed #11182 anyway.
• #11182 had to be manually reopened. Claude's reopen comment ties the auto-close to the stale Resolves #11182 in the squash commit body.

This validates the PR #11183 Cycle 2 branch-history RA as load-bearing, not cosmetic.

The Problem

The current review discipline treats PR body close-target state and GitHub's pre-merge closingIssuesReferences surface as sufficient evidence that a PR will not auto-close an issue. That is false under the repository's squash-merge path.

GitHub's squash commit body can concatenate branch commit bodies. If any branch commit body contains Closes #N, Fixes #N, or Resolves #N, the resulting dev-tip commit can still trigger auto-close, even if:

• the PR body says Refs #N,
• the latest commit body says Refs #N,
• closingIssuesReferences is empty before merge,
• reviewers have already downgraded the issue to non-blocking branch hygiene.

This creates a repeatable false-negative in the close-target audit.

The Architectural Reality

Relevant workflow surfaces:

• .agents/skills/pr-review/references/pr-review-guide.md §5.2: reviewer-side close-target audit already says to parse PR body and commit messages, but it does not explicitly capture squash-merge commit-body preservation as a post-merge auto-close mechanism.
• .agents/skills/pull-request/references/pull-request-workflow.md: author-side PR workflow controls commit and PR body hygiene before handoff.
• PR #11183 / merge commit 5c7c5a2f4: empirical anchor showing stale commit-body Resolves #11182 survived squash merge and auto-closed #11182.
• #11182 reopen comment: confirms actual issue state damage and manual repair.

The Fix

Update the workflow documentation / skill substrate so both authors and reviewers treat branch commit bodies as merge-time close-target surfaces.

Recommended implementation shape:

1. In pr-review-guide.md §5.2, add an explicit squash-merge note: commit-body magic close keywords are load-bearing even when the PR body is corrected and closingIssuesReferences is empty.
2. In the reviewer close-target audit, require scanning git log origin/dev..HEAD --format='%h%x09%s%n%b' or equivalent for magic close keywords whenever a PR is partial / multi-phase / leaves the referenced issue open.
3. Define the resolution ladder:
   • clean branch / Drop+Supersede when no operator auth exists for rewrite,
   • operator-explicit auth for amend/rebase/force-push when preserving the same PR is preferred,
   • do not approve with stale magic-close commit bodies on partial-resolution PRs.
4. In pull-request-workflow.md, add the author-side pre-handoff check: if the PR body uses Refs #N because the ticket stays open, no commit body in origin/dev..HEAD may contain Closes/Fixes/Resolves #N.
5. Keep wording concise. This is a high-severity tripwire, not a new long-form theory section.

Contract Ledger Matrix

Acceptance Criteria

• pr-review-guide.md §5.2 explicitly states that squash-merge can preserve branch commit-body magic-close keywords and trigger post-merge issue closure.
• Reviewer audit instructions include a branch commit-body scan command or equivalent.
• Reviewer resolution guidance blocks approval for partial-resolution PRs with stale magic-close commit bodies unless the branch is cleaned or a clean superseding PR is used.
• pull-request-workflow.md includes the symmetric author pre-handoff check for Refs #N partial PRs.
• The docs cite PR #11183 / commit 5c7c5a2f4 / #11182 reopen as the empirical anchor.
• Added text is compact and placed in existing close-target sections rather than creating a large new top-level rule.

Out of Scope

• Changing GitHub repository merge settings.
• Adding automation that rewrites commits.
• Reworking all close-target discipline across AGENTS.md.
• Retrospectively editing already-merged commit history.

Avoided Traps

• Relying on closingIssuesReferences alone: empirically insufficient before squash merge.
• Treating stale commit bodies as cosmetic: disproven by #11182 auto-close.
• Forcing amend/rebase/force-push without operator authorization: conflicts with workflow safety; clean replacement PR is the compliant path.
• Adding a large new rule block: this should be a surgical close-target audit tripwire.

Related

• PR #11183
• PR #11184
• Issue #11182
• Merge commit 5c7c5a2f4

Origin Session ID

22713fa8-23d2-4b31-918b-6e2f48d69c06

Handoff Retrieval Hints

• query_raw_memories(query="squash merge stale Resolves commit body auto close")
• query_raw_memories(query="PR 11183 branch-history RA closingIssuesReferences")
• gh api repos/neomjs/neo/commits/5c7c5a2f4 --jq '.commit.message'