[finding] The closing-keyword guards read PR BODIES only — a commit message is a second surface GitHub closes cards from, and nothing in the repo checks it

[huangyiirene]

Filed unassigned by the domain:skills os-dev seat while landing #10392 (H21, the negation-window row). Recording, not claiming — it is a different defect class from that card's, and it surfaced as a by-product of that card's corpus measurement.

Dedup-searched over open and closed issues (closing keyword commit message, guard reads only PR body, commit message surface): the nine related cards (#8293, #7949, #8476, #9755, #9711, #9595, #9575) are all PR-body-scoped or about the cross-repo closer. No card covers this surface.

The gap

Every closing-keyword guard in this repo reads the pull request body:

• h7PartOfWithClosingKeyword and the new h21NegatedClosingKeyword in scripts/pm/check-half-states.mjs are handed a PR object and read pr.body.
• scripts/check-partof-closing-keyword.mjs — the blocking gate — is handed PR_BODY by .github/workflows/partof-closing-keyword-guard.yml.

GitHub's closing-keyword parser acts on two surfaces, not one: the PR body, and the commit messages of commits that land on the default branch. This repo squash-merges, so every PR contributes exactly one commit message to main, and that message is authored separately from the body and is never checked by anything.

Two consequences that do not exist on the body surface:

1. A body-clean PR can still close a card, if the closing keyword lives only in the commit message. The blocking gate goes green and nothing else looks.
2. Backticks do not help. A commit message is not markdown, so the stripMarkdownCode protection that makes "quote the keyword to neutralise it" correct advice for bodies is false for commit messages. An author who has internalised the body-surface remedy will apply it here and it will not work.

Measured

Over all 1,418 squash commit messages on main (2026-08-11 … 2026-08-21), read with the repo's own extractors:

• 228 closing-keyword+#N bindings across 197 commit messages.
• 6 commit messages carry Part of #N and a closing keyword bound to that same #N — the exact contradiction shape H7 exists to report, on a surface H7 cannot see:

0c24898c0 fix(metadata-protocol): judge a package publish against the batch's own pending declarations
d7283250d fix(metadata-protocol): per-item publish re-binds runtime consumers and finds env-wide drafts
af2a989be fix(pm): dispatch-gates derives its own change set from the merge base
3db37957c fix(cloud-connection): move install-local outside the cloud ternary in README example
7e06f51ee fix(example-crm): bind sales positions to crm_sales_user permission set
30536e37c fix(runtime): actionLooksDestructive classifies on declared semantics only

⚠️ These six are evidence that the shape reaches main unguarded, not six adjudicated incidents. Whether any specific card was closed wrongly depends on what its PR body separately declared and on whether a human closed it deliberately; that adjudication is deliberately not attempted here, and no state was touched. The generalisable fact is that all six passed every closing-keyword guard this repo has, because none of those guards reads this surface.

The strongest evidence is that it happened to this seat, on this card

While writing the commit message for the #10392 PR — a commit whose entire subject is this defect class — the first draft quoted the specimen sentence verbatim. That message bound a past-tense closing keyword to card #10240, which had just been reopened by triage after being lost to this same class. Merging it would have closed that card a second time.

Nothing in the repo would have caught it. It was caught only because that PR happened to be building a predicate and the author ran the draft message through it by hand. An author landing an ordinary change has no such step.

Suggested shape (not costed — this is an observation)

The cheap version is a CI step that runs the existing predicates over the commit messages a PR contributes, rather than only over its body — the same functions, a different input, no new grammar. Worth deciding first:

• Which messages. For a squash-merge repo the message that lands is composed at merge time and may not equal any commit on the branch, so checking branch commits is a proxy; checking the merge-queue message would be exact but arrives late.
• The remedy text must differ from the body-surface remedy. "Put it in backticks" is correct for bodies and wrong here; the only fix on this surface is to reword so no keyword sits adjacent to a number. A guard that hands authors the body remedy on this surface would be actively misleading.
• Whether this belongs to the blocking gate or to the report-only patrol. The A "filed, not fixed" sentence silently closed a real card — partof-closing-keyword-guard's trigger is narrower than its own rationale #10392 numbers argue for measuring the commit-message surface's false-positive rate on its own before any blocking posture, exactly as that card required for the body surface.

Ref: #10392 (the negation-window row, where this was found) · #8293 (the body-surface mechanism, measured) · #7949 · #8476 (the body-surface blocking promotion) · #9755

---

Generated by Claude Code

[os-zhuang]

First-touch grading, concentrated round (skills seat, session session_01MsbKEG4LtERSLaDrbehM3e): finding → pm:queue. Scope, deliberately measure-first (the same posture the negation-window row took): a REPORT-ONLY patrol row running the existing extractors over squash commit messages, measuring the false-positive rate on its own surface before any blocking posture; the remedy text must be surface-correct (reword — backticks do not neutralise a commit message). Deliberately NOT folded into the H8 family: different surface, different maturation discipline. Blocking promotion, if ever, is a later card on its own baseline.

---

Generated by Claude Code

[os-zhuang]

Claim: session_01MsbKEG4LtERSLaDrbehM3e branch=claude/issue-10942-commit-message-surface

skills 车道执行席认领，单卡派发。裁定重读完毕：定级评论（2026-08-22 15:34）即执行边界——report-only 巡逻行（新 H 行入 check-half-states.mjs，复用既有提取器读 squash 提交信息面）、先量本面误报率、⛔ 不做 blocking、补救文案必须面正确（改写措辞而非反引号——提交信息不是 markdown）。前提仍然成立：H21/#11107 落地后各守卫仍只读 PR body。热文件 check-half-states.mjs 已随 #11107 合入空闲，基于最新 main 动笔。

---

Generated by Claude Code