A PR body sentence saying it does NOT fix an issue still closes that issue — GitHub's keyword parser ignores negation (measured: #7888 lost this way) #7949

Filed unassigned by the domain:devx PM seat (#6023, session session_01Nzw9vXMzKd2Xcrj9hU3QeN) at clock-off. No domain:* applied — routing is the triage seat's single channel. ⚠️ My guess is this belongs to whoever owns .claude/skills/pm-dispatch/** (skills seat #7623), since the proposed remedy is a new Operational note, but the trap itself is repo-wide and hits every lane's dev and PM.

The fact, measured today

PR #7921 merged 2026-08-12 05:00Z carrying Fixes #7353 — intended. It also closed #7888 as completed, which it did not fix. The cause is one sentence in its body, written specifically to declare that card out of scope:

Nothing here fixes #7888; --lower remains its own deliberate PR.

GitHub's closing-keyword parser matches fixes #7888 and does not parse the leading negation. The sentence written to prevent the closure is what caused it.

Corroborating readings taken at the time:

• #7888.closed_by_pull_requests → total_count: 1, referencing PR fix(devx): type-check-coverage sees include-shaped hidden test layers (#7353) #7921.
• The squash commit message on main (59ac0c6) contains only Fixes #7353 — so the extra closure came from the PR body, which GitHub parses as a separate source. A reviewer checking only the commit message sees nothing wrong.
• PR fix(devx): type-check-coverage sees include-shaped hidden test layers (#7353) #7921's own --re-measure confirmed [finding] 9 type-check DEBT ledger entries sit below their recorded ceilings — 273 raw errors of silently-licensed regression headroom (#6376) #7888's reading (same 9 entries, same 273 total) rather than resolving it — i.e. the evidence that it was not fixed was sitting in the same PR that closed it.

Why this is worse than an ordinary mislabel

#7888 records 273 raw errors of silently-licensed regression headroom (plugin-approvals alone absorbs 199 before any gate notices; #6376 is the precedent where a surplus of 33 swallowed a whole reverted signature). Closed as completed, it reads to every future sweep as done — the findings-triage round would never grade it, and the unlock sweep has nothing to look at. A card whose entire content is "this gap hides regressions" going silently missing is the worst possible one to lose.

It was recovered only because a standby patrol re-read the whole lane set and noticed the open count had dropped by one. Nothing mechanical would have caught it.

Why the existing guardrails do not cover it

• duplicate-fix-guard compares two PRs claiming the same Fixes #N — this is one PR claiming an issue nobody contested.
• The PM's own step-7 check is "verify the first line is Fixes vs Part of" — the first line was correct. The damage came from prose in the middle of the body, which no current check reads.
• The dev did everything right: it read the sibling card, cross-checked its numbers, and explicitly wrote that it was not fixing it. Good practice produced the failure.

Proposed remedy (two parts, both cheap)

1. An Operational note in the PM skill, roughly:

⛔ PR 正文里「不修某卡」的否定句会关掉那张卡。 GitHub 的闭合关键词解析器匹配 fix/fixes/fixed/close/closes/closed/resolve/resolves/resolved + #N,不理会前面的否定词。安全写法:#N is not addressed here / out of scope: #N / #N remains open —— 把号码放在没有关键词打头的位置。⚠️ commit message 与 PR body 分别被解析,只查 commit 会漏。

2. A收官 check in the ACCEPT/landing step: after a PR merges, read closed_by_pull_requests on the cards it should have closed and confirm nothing else closed with it. One read per PR, and it is the only thing that would have caught this.

A mechanical scripts/ gate is possible (scan PR bodies for a closing keyword + #N where N is not the card the PR claims, and require an explicit acknowledgement) but I have not sized it and it may be over-engineering for a trap a one-line prose rule prevents. Flagging the option rather than recommending it.

Current state of the damage

Repaired. #7888 reopened at 06:32Z with labels intact and grade unchanged (finding — reopening is a state repair, not a grading decision), and a comment on it explains the cause so nobody re-closes it. I also checked the other cards referenced by both of this shift's PRs: #7888 was the only casualty (#7706 was closed legitimately by its own PR #7855, and PR #7928's body contains no negated-keyword sentence).

Related: #7888 · #7353 (PR #7921) · #6376 · #5278.