docs(pm-dispatch): ADR merge prohibition in Guardrails + ACCEPT path fork; probe-and-revive as the standing backstop; replay-wake identity check (#6741, #6747)

[os-project-manager]

Fixes #6747
Part of #6741
Part of #5925
Part of #6644

Sweep of four cards, all landing in .claude/skills/pm-dispatch/SKILL.md — this repo's hottest serialisation surface, so one claim, one PR, a per-item checklist (per the packaged-promotion rules and the shape PR #6720 used). Every anchor was re-derived against the post-#6720 file by content; no line numbers or counts were inherited from the card bodies.

Per-item checklist

#6741 (p1) — ADR merge prohibition — Part of, and why

• Guardrails: new clause directly after the version-release guardrail it mirrors. The maintainer ruling (2026-08-08, ⛔ Discipline: ADRs are confirmed and merged by the maintainer only — no AI seat may merge, queue, or auto-merge a docs/adr/** PR #6741) is quoted verbatim and untranslated: 「adr 只能由维护者自己确认,人工合并,ai 不得擅自合并。」 The clause names the prohibited actions explicitly (merge / add to merge queue / enable_pr_auto_merge), carries the "reviewed + approved + fully green does not override" sentence, cites the worked example docs(adr): ADR-0123 — layered master data, template/override linkage and the resolution organization (#4585) #6668 (a fully-green ADR draft closed on demand grounds no gate could evaluate) and the inverse failure ADR-0005 的租户可定制白名单表说 flow/workflow/approval 一律 ❌ per-org override,注册表却写着 allowOrgOverride: true——反转没有走 ADR(Prime Directive #13) #6191/ADR-0005:57 白名单表与注册表的分歧不止 flow：page/app/action/permission/tool/skill 六类 ADR 写 ❌ 而 allowOrgOverride:true，另有 dataset/book/position 三类表里没有却默认成了 true #6483, and anchors on AGENTS.md PD [WIP] Add Chinese version of the documentation #13 for why an ADR merge is the adoption of a governance position.
• Step 7 ACCEPT path — the load-bearing placement. The ACCEPT paragraph's "once every check is green" sentence now routes through a path-fork before ready/auto-merge/queue: read the PR's changed paths via get_files (never the report's self-description); any hit on docs/adr/** switches ACCEPT to a different terminal state — (1) review recorded on the issue as usual, (2) PR left for the maintainer (no merge, no queue, no auto-merge — visibly pending, not silently parked), (3) the round report lists it as awaiting a human merge. The passage states the incident's mechanism plainly: the general ACCEPT rule was not wrong — it executed correctly on PR docs(adr),chore(devx): reconstruct ADR-0079 + make every cited ADR number resolve #6732 and produced the prohibited act, which is why the carve-out must live where the hands are, not only in Guardrails.
• Mixed diffs: one docs/adr/** hit is enough — no proportionality, no "mostly other files" pass; split the ADR part into its own PR instead.
• Eviction mechanism recorded in both places: converting the PR to draft is the only reliable way to evict an already-queued PR; disable_pr_auto_merge alone does not remove merge-queue membership (measured on docs(adr),chore(devx): reconstruct ADR-0079 + make every cited ADR number resolve #6732's reversal).
• AGENTS.md (requirement 1 of the card) — NOT in this PR. Verified absent on current origin/main (no clause matching the ruling anywhere in AGENTS.md). My dispatch fixes the file surface to SKILL.md only and instructs to report rather than widen. The Guardrails clause honestly names AGENTS.md as the rule's repo-wide authoritative home per the card, pending that half. Hence Part of #6741, not Fixes — merging this with Fixes would close a card whose requirement 1 is unimplemented, which is precisely the failure mode the Fixes-vs-Part of rule in this same file exists to prevent.

#6747 — PM half of the dev-termination contract — Fixes

• (a) Probe-and-revive as the standing backstop. New passage in step 6 after the five probe rules. The old trigger face ("dispatched and no remote branch/PR yet") is falsified by the 2026-08-08 measurement — 7 of 7 dispatches failed to hand back cleanly after opening a correct PR, and 3 of the 4 dispatches carrying the termination clause verbatim died anyway — so the trigger bullet itself was re-worded to "dispatched and report not yet arrived" (producer of the contradiction fixed, not patched around), and the passage carries the consequence that makes the backstop cheap: every death so far was fully recoverable from transcript with zero work lost, so the cost is latency, not correctness — which is why the response is a probe and never a redispatch into a possibly-live worktree.
• (b) A stale wake is recognisable without re-reading the deliverable. New passage after the stall-correction block: identity before content. publish/lint 面对字面量空组合子响亮拒收 —— #5322 拍板单位元后的可选编写期信号(PD #12 形状) #5330 emitted six notifications for one card, five of them byte-identical replays of the same full JSON — each indistinguishable at arrival from a real completion, so each was re-adjudicated. The PM-side procedure: compute the identity tuple (issue, branch, PR head sha) plus the monitor's self-declared subject (the dev-side half PR docs(agents): make the os-dev termination contract explicit and honest about its measured failure rate (#6586) #6740 landed); on a match with an already-reviewed report, log one replay line and stop — no re-review, no re-read of the diff, no duplicate ACCEPT comment. Neither the arrival (a monitor fires on its own deadline) nor the non-arrival of notifications is evidence of liveness; the dev-side self-identification makes this cheaper but is explicitly not a precondition (3-of-4 clause-carrying deaths — the same cause defeats both clauses).

#5925 — seven items — Part of (verification only, zero new text)

All seven items were checked against the current file; every one is declined, each with a reason:

• Items 1–6 are already in the file, landed by PR docs(pm-dispatch): 座位贴协议(一座位一贴)+ 两班次沉淀的 16 条 SKILL 更新 #6029 (2026-08-06), verified by content grep at the merged ref: handover checklist (交接收尾清单), whole-body-replace trap (structurally obsoleted by the one-sticker-per-seat architecture, 座位贴协议), the three API traps (Operational notes 13 for list_issues OR-filter and labels whole-set replace; the REST hourly-reset quota note), priority:p0 (step 3), standby posture (待命姿态 under Stop conditions), and the PM-tooling-PR exception (Guardrails, tagged #5925-6). Writing them again would be the redundant-second-statement defect the dispatch names. This matches the card's own re-scoping comment of 2026-08-08 ("items 1–6 are historical record, not work").
• Item 7 (extract long narratives into references/incidents.md) is excluded by the card itself (「⛔ 不与前 6 条同 PR」, maintainer-approved as a separate pure-move PR) and stays open — which is why this line is Part of.

#6644 — remainder — Part of (nothing left that is writable)

Determined from PR #6720's merged diff: L1 (Fixes/Part-of), L3 (hot-file serial queue section), L4 (PM narrowing judgment as falsifiable assumption), L5 (sweep products grouped in ACCEPT) are all in the file — no restatement written. L2 (report-at-draft-PR-time vs wait-for-CI) is ⛔ excluded from this sweep: it is escalated to the maintainer with four live options and a dev-side inversion argument (PR #6740), and this PR leaves every passage it touches on that question untouched. The card stays open on L2, hence Part of.

Scope

One file: .claude/skills/pm-dispatch/SKILL.md (+100/−4 before the routine merge of origin/main). No docs/adr/**, no content/docs/releases/**, no code.

Verification

No package code changed; pnpm test / pnpm typecheck are not meaningful for this diff (saying so beats manufacturing template-shaped evidence). What was run, serialized under the shared verification lock with the capped heap, is the full check:* gate list enumerated from .github/workflows/lint.yml — 41 gates, all green, including the two that read this file family (check:skill-compatibility: OK; check:skill-frame-sync: OK), check:nul-bytes (OK, 6269 files), and a manual control-byte scan of the edited file beyond the gate's surface (no matches). Honest sequence: 37/41 green on the first pass; the 4 reds (check:app-nav-i18n, check:i18n, check:i18n-coverage, check:type-check-debt) were the fresh-worktree missing-build-artifact class — each gate's own classifier named its prerequisite (built CLI / built packages closure), and after turbo run build --filter=@objectstack/cli plus the packages closure build, all four re-ran green with echo EXIT=$? discipline. None of the four reads .claude/.

Labels

skip-changeset: this PR touches only .claude/ internal agent tooling — releases nothing, writes no changeset. Applied as a union with the labels already present, not as a bare set.

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
objectstack	Ignored		Aug 8, 2026 6:04pm