pm-dispatch: SKILL.md wave-2 family — two-tier triage inventory, decision re-read, subagent tier fuse, git-history guard index

[os-zhuang]

Fixes #10190
Fixes #10544
Fixes #10523
Fixes #10508

Maintainer-authorized family fold (recorded on the chain-head claim, live PM chat 2026-08-22): four graded cards, one commit per member, one shared branch. Governed surfaces throughout (.claude/skills/**, AGENTS.md) — draft PR, human merge only; never queue/arm/ready.

Members (one commit each, in order)

1. Two-tier triage inventory (c2d962af) — SKILL.md gains the principle in the 分诊座位职责 section: hourly rounds read since windows anchored to the previous round-close brief's timestamp from the seat post (the mutex guard's existing reading — no new state; label writes refresh updated_at; anchoring on the brief covers a skipped-fire backlog); one daily fire runs the four-repo full reconciliation and clusters the already-daily duties; the tier is picked by fire-time-of-day, never a counter; the brief states which tier ran. The since-equivalent readings, cost targets and daily-duty clustering land in references/dispatch-runbook.md (new section 分诊两级盘点细则 with the maintainer's 2026-08-20 verbatim quote).
2. Decision re-read (ad9c80fe) — the pre-brief protocol (候选与批次) gains the decision re-read, symmetrical to the claim protocol's race re-read: read the card's comments to the last page before writing a brief; quote any maintainer ruling verbatim into the brief's ruling partition as binding, or state none exists. Lead 1 of the card only; leads 2–3 (greppable marker, body-carried ruling) are noted adjacencies, not implemented.
3. Subagent tier fuse (9857ead3) — the needs:contract-review chain now states: a mode:subagent get_session reading measures the dispatching session (the two fields agreeing is one source quoted twice), so the downgrade fuse is valid only for seat sessions; clause-② subagent dispatches always keep needs:contract-review, and the dispatched-at-the-floor review skip is seat-session-only. Composes with the 2026-08-21 self-review relaxation (that path is a seat session reading its own row) — no conflict measured.
4. git-history guard index (9e58b7d8) — AGENTS.md (Multi-agent working discipline) gains the windowed-history rule using the card's drafted wording (answer or REFUSE; shallow clones answer windowed git log/rev-list at exit 0 with no warning; historyHorizon() is the read-only predicate), and the pm-dispatch 机械守卫索引 gains the scripts/pm/git-history.mjs row, ratchet-paid. Both placements landed (the SKILL.md row was affordable).

Ratchet ledger (all three files at their ceilings before AND after; verified at 9e58b7d8)

file	before	after	ceiling
.claude/skills/pm-dispatch/SKILL.md	682	682	682
.claude/skills/pm-dispatch/references/dispatch-runbook.md	243	243	243
AGENTS.md	961	961	961

One dispatch mechanism assumption was falsified and is reported here: the runbook is not unratcheted — it carries its own 243-line ceiling at zero headroom, so member 1's runbook section is paid in-file too. Additions are paid by content deletions, never bare re-wrap: trimmed redundancy/incident-narrative in the runbook (读数成本三则 intro, 落卡细则 measured-narrative per its own 行文纪律, 活性巡查 tail, roster/SendMessage duplication with the main file, 紧急卡 restatement of a rule the main file owns, subagent-checkout incident detail), in SKILL.md (objectos onboarding history, dispatch-gates clause duplicated by the runbook's 清单取数补遗, one stale-premise rationale clause), and in AGENTS.md (merged-language incident narrative, duplicated-branch incident sentence, stash/registry incident detail — case law lives in git history per the ratchet's own header). Per-commit ratchet stays ≤ ceiling at every step (682 → 680 → 681 → 682).

Gates (final tree 9e58b7d8, derived by node scripts/pm/dispatch-gates.mjs on the actual change set; exits captured before any pipe)

• check:pm-skill-ratchet — ✓ all 20 files within ceilings (verdict lines: "682 lines (ceiling 682; headroom 0)", "243 (243; 0)", "961 (961; 0)")
• check:pm-skill-id-lint — ✓ "17 file(s) clean (pattern /#[0-9]{3,}/g)"
• check:skill-frame-sync — ✓ "binding sentence present in all 4; 4 count mention(s) agree"; check:skill-frame-freshness — ✓ "frame itself is unchanged"
• check:pm-governed-prose — ✓ "2 instruction surface(s) name all 5 registered governed surfaces"
• check:pm-governed-merges — ✓ self-test 119 assertions
• check:doc-authoring — ✓ "389 files clean"; check:docs-audit-scope — ✓; check:required-contexts — ✓
• pnpm --filter @objectstack/lint run check:doc-formula-expressions — ✓ (after building the lint dependency closure; first run failed only on unbuilt @objectstack/formula dist)
• check:nul-bytes — ✓ "no raw ASCII control bytes" (6414 files)

Instruction-file-only diff, releases nothing → skip-changeset (label applied and read back after PR creation).

Generated by Claude Code

---

Generated by Claude Code

[github-actions]

⛔ merge queue 构建失败 — 先分诊，再决定要不要重排

队列构建 32586272156 红了。队列跑的是全量套件（PR 侧 CI 只跑 affected 子集），
所以失败的测试可能在本 PR 没碰过的包里 —— 那不是重排能修的。每次盲目重排都会让排在后面的所有 PR 重建一轮。

失败的 job（日志抽取，best effort）：

• Console Pin Gate — 失败步骤: Build the Console SPA at the pinned objectui SHA

  ✗ Build failed in 5.95s
  

↳ 失败原因 是判读的关键：超时（Test timed out in … / Hook timed out in …）多半是负载/时序，不是本 PR 的回归；
断言（AssertionError: …）才指向真实的行为改变。两者的 FAIL 行长得一模一样，只有这一行能区分。

跨 PR 相同签名（24h，按失败测试文件聚合）：

• ⚠️ 本次没有可用的聚合签名（日志里没有能解析出测试文件名的 FAIL 行）—— 这不是「没有同签名的其他 PR」，是这一轮没测到。跨 PR 聚合本次不可用，请手工比对其他 PR 的同类评论。
• ⚠️ 24h 评论账本没读完（超过 5 页仍未读到窗口尽头），所以上面的「不同 PR 数」是下界，不是全量。

历史信号：

• 本 PR 过去 24h 无队列失败记录（首次）。
• 过去 24h 队列共有 78 个失败构建（不含本次）。

分诊清单：

1. 失败测试在本 PR 改动的包里 → 真回归，修 PR。
2. 失败测试与本 PR 无关 → 看上面的「跨 PR 相同签名」；已有汇总 issue ⇒ flaky/环境问题实锤，去那张 issue 上谈，修好前重排只会再烧一轮全队列。
3. 两者都不是 → 可能与同组 PR 语义冲突；等前面的 PR 落地或失败出队后再重排一次即可，不要连续重排。

---

Generated by Claude Code · merge-queue-triage workflow (#4859)