fix(check-partof-closing-keyword): the already-pushed red says what discharges it — a merge whose squash message is the PR body — beside the rewrite ban

[hotlong]

Part of #16516

This PR delivers the gate half of the card — item 2 of triage's re-scope (comment 5579821742): the BRANCH ALREADY PUSHED paragraph of scripts/check-partof-closing-keyword.mjs now states, positively, what discharges the red. The other half — the queue-entry rule's third case in .claude/skills/pm-dispatch/SKILL.md and its references/core-rules.md twin — is not in this PR: it stopped under the dispatch's 回翻条款 because it cannot be expressed in place or paid for by a fold in either ratcheted file (measurements below). It needs a ceiling decision that is the maintainer's, not a seat's; the full reading and the ready-to-land text are in the dev report on the card.

What changed — scripts/check-partof-closing-keyword.mjs only (+43 / −0)

Output. In judge(), the already-pushed paragraph — after …and it does not decide whether the pull request merges. — gains one paragraph:

  What DISCHARGES it is the merge, and only a merge whose squash message is the PR BODY: then the
  trailer above never reaches the default branch. Whether that happens by itself is the repository
  setting `squash_merge_commit_message`. At `COMMIT_MESSAGES` (the state fact 2 measured) it takes
  the lander replacing the assembled commit list with the PR body BY HAND at the merge button — a
  queue merge edits nothing, so there the residue lands. At `PR_BODY` every squash does it, queue
  included. Until that merge the red stays on this branch; ⛔ discharging it never rewrites history.

Header. The same statement, as the authority on detail, in the section RULE 2 — what the output may ask for, and when it may ask for nothing, directly under the measured-squash paragraph it conditions.

Self-test. Three pins in the RULE 2 battery (its ledger count is a floor, so no ledger edit): the discharge sentence prints on the already-pushed branch; the setting is named with BOTH of its values and the by-hand clause; and the sentence sits BESIDE ⛔ Do NOT amend, rebase or force-push, never in place of it.

Why this wording, and the setting-D probe

The card set the wording's timing on setting D (squash_merge_commit_message moving from COMMIT_MESSAGES to PR_BODY, the maintainer's own change ruled on #16502). Probed at claim time on origin/main 7862fb711 (re-read on 7f96e1417): the three newest multi-commit squashes — 68fd85a41 (#16812), 30b099078 (#16788), 094b8fd9c (#16775) — each begin their body with a * bullet followed by a commit subject, the COMMIT_MESSAGES shape; a PR_BODY squash would begin with the PR body's first line. D is not in effect. So the sentence takes the before-D form — the merge needs the PR body as its squash message, and today that is a by-hand act — but it is written as a CONDITION on the setting's value rather than as today's value, so it stays true once D is applied and no history is rewritten either way. The existing fact 2 (assembled from the COMMIT messages … 0a61db1f5, pinned) is untouched; the new paragraph names it as the measurement behind the COMMIT_MESSAGES value.

What is deliberately untouched

• RULE 2's detection — commitRelations(), commitTrailerFindings(), the four counterfactual cases — byte-identical (the diff is 43 insertions, 0 deletions).
• The three loci triage fenced off (on d4401f75b: the finding sentence at 594, the repair paragraph at 746, the guidance pins at 1017–1018) — each present exactly once, byte-identical, checked by content rather than line number.
• The six existing guidance pins.

Verification

• node scripts/check-partof-closing-keyword.mjs --self-test → exit 0, ✓ check-partof-closing-keyword self-test: 95 cases pass. (92 before, 3 new).
• Ablation (from the committed state b93320823, blob fb822185aacb5f3887c3f1099eec8faf9fb3722f): the output paragraph deleted on disk (anchor What DISCHARGES it is the merge, and only a merge counted 1 → 0; git diff --stat HEAD = 7 deletions) → self-test exit 1, 3 of 95 case(s) failed, exactly the three new pins. Restored with git checkout HEAD -- scripts/check-partof-closing-keyword.mjs; git hash-object on the restored file = the HEAD blob above; git diff HEAD empty, porcelain empty; self-test 95/95 again.
• Governed predicate: node scripts/pm/check-governed-merges.mjs --test scripts/check-partof-closing-keyword.mjs → exit 0, NOT governed (the diff no longer touches .claude/**; the seat routes the landing).
• Mergeability, driverless: git merge-tree --write-tree from a bare --shared probe clone against origin/main 7862fb711 and again against 7f96e1417 (fix(pm): dispatch-gates' reconciliation line enumerates every OUTSIDE block, not three of five #16802, scripts/pm/dispatch-gates.mjs only) → exit 0, clean. origin/main 7862fb711 is merged into the branch; 7f96e1417 is not (it is disjoint from this file — §10 scoping).

Gate reconciliation

Derived in the worktree on the merged head 27bea86dc with node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack (no paths; the change set is read from the merge base): 31 commands, identical before and after the merge. The seat's list (39 on d4401f75b, derived over three paths) carried eight more — check-governed-queue-guard --self-test, check:doc-formula-expressions, check:doc-authoring, check:pm-governed-merges, check:pm-governed-prose, check:pm-skill-id-lint, check:pm-skill-ratchet, check:skill-frame-sync — every one of them placed by the two .claude/skills/** paths this PR does not touch; they belong to the stopped half.

Every command was run from a script that captures the exit by redirect before any pipe (bash -c "$cmd" > log 2>&1; ex=$?), NODE_OPTIONS=--max-old-space-size=4096, no verify lock (no build or test — the diff touches no package):

• 30 of 31: exit 0, each gate's own verdict line read from its log — among them ✓ check-partof-closing-keyword self-test: 95 cases pass., check-nul-bytes: OK (scanned 8302 text file(s) … no raw ASCII control bytes), ✓ check-self-test-wired: every one of the 191 script(s) …, check-closing-keyword-parity: OK (3 parsers agree …), ✓ comment-mask corpus sweep … 6356 files, 0 disagree, ✓ check-whole-set-label-write --self-test: all cases pass, check-ratchet-remedy-authority: 227 scripts swept ….
• pnpm check:pm-dispatch-gates (exceeds the container's foreground cap; run detached with output to a file and awaited with tail --pid, restarted once on the merged tree, 09:03Z → 09:15Z): its own verdict line ✓ dispatch-gates self-test: 1552 cases pass., no signal branch printed.
• node scripts/pm/dispatch-gates.mjs --ran over the run record: exit 0, ✓ dispatch-gates --ran: 31 derived famil(ies) accounted for — 31 run, 0 NOT-MEASURED.

Not in this PR's derived total and named as such by the tool: the always-runs tail, 10 whole-root families, 1 CI-measured family (this very script, read from the workflow event payload — its self-test is the local half and is green above), 3 workflow-valued families, and the artifact-roster families (silent on every card; not clearances).

Mirror reading

skills/objectstack-pm-dispatch/SKILL.md (the published mirror) carries no queue-entry rule — 0 hits for 入队 / 队列, and its queue mentions are the pm:queue backlog label; check:skill-frame-sync compares only the decision frame (COPIES: the internal and the published SKILL.md); check:pm-skill-id-lint scans .claude/skills/pm-dispatch/**, os-dev.md and AGENTS.md. Two references restate the all-green rule as pointers — references/review-checklist.md:43 and references/true-green.md:3 — no gate holds them to SKILL.md's wording; they become the follow-up's concern when the third case lands.

验收备注

• noted, not filed: review-checklist.md:43 and true-green.md:3 restate 「入队资格 … 全 check 绿」 without the third case; a pointer, not a gated mirror — 承接者: the follow-up PR that lands the two skill loci of this card.
• noted, not filed: dispatch-gates.mjs --commands printed STALE TREE … 1 file(s) it derives from CHANGED on 27bea86dc immediately after merging 7862fb711; the file was scripts/pm/dispatch-gates.mjs itself, moved by fix(pm): dispatch-gates' reconciliation line enumerates every OUTSIDE block, not three of five #16802 one commit later — correct behaviour, recorded so the seat does not read it as a derivation defect. 承接者: 无.
• noted, not filed: the gate's fact 2 (assembled from the COMMIT messages, not from the PR body, pinned with 0a61db1f5) reads as a live fact today and as a dated measurement once D lands; the new paragraph conditions it on the setting's value, so no edit is owed now. 承接者: whoever applies setting D reads the paragraph, not the fact.

维护者速读(草稿)

改了什么:check-partof-closing-keyword 门禁在「分支已推送」那段输出里补了一句正面陈述:这个红由合并化解,且只有 squash 提交信息取 PR 正文的合并才化解;仓库设置 squash_merge_commit_message 在 COMMIT_MESSAGES 时要人在合并按钮手动换成 PR 正文(队列合并不改信息,残留会落地),在 PR_BODY 时每次 squash 自动完成;两种情况都不改写历史。文件头同步一段,自测加三条钉子。

为什么改:此前输出只说「什么都清不掉、⛔ 不许改写历史」,没说它到底怎么被清掉;读到红的 agent 不知道该做什么、也不知道可以什么都不做。措辞按设置值写成条件句,D 生效前后都成立。

风险与代价(含回滚):只加不删(+43/−0),检测逻辑与四条反事实逐字未动;三处已落地成果逐字未动。回滚 = revert 这一个提交。⚠️ 本 PR 只交付卡的门禁那一半;入队规则第三种情形那一半(SKILL.md + core-rules.md)实测无法在既有行内表达、两文件均无可折并的相邻行(601 对与 121 对实测 0 对 ≤120 字节),需要天花板裁决,已在卡上报告。

席位意见:(留空)

你要做的:一个动作 —— 对卡上报告里的天花板问题作裁决(两文件各 +1 行,还是别的路);设置 D 仍未生效,顺手改了这一半就自动化解。

🤖 Generated with Claude Code

Generated by Claude Code

---

Generated by Claude Code