[objectstack-fleet]

Requested by Irene · project thread

Fixes #28921

Clause-②: no
Metadata-protocol: none

维护者速读

• 改了什么:分诊不再出进度表,只在七种情况下推送你;清单写进分诊席章程 lanes/triage.md。
• 为什么改:你说「有问题的时候要找我，没问题我也不需要看」,并「同意推送清单」;另补「pr 超过 6小时没更新」这一条。
• 风险与回滚:只改席位规则文本和行数上限,不加门禁不加标签;回滚即 revert。分诊章程上限 8 → 15,分诊职责 102 → 99,合计 +4 行。
• 先看:① lanes/triage.md 第 8–15 行(清单本身);② triage-duties.md 三处合并行;③ check-skill-line-ratchet.mjs 两条上限及注释。
• 两分钟验证:node scripts/pm/check-skill-line-ratchet.mjs,预期 lanes/triage.md 15 行(上限 15)、triage-duties.md 99 行(上限 99),退出 0。
• 需要你做的:无,席位复核后落地;若不同意把分诊章程上限抬到 15 行,在卡上回「否」。

What changes

lanes/triage.md now carries the push list (lines 8–15): no progress table and no chat round report; the channel follows SKILL.md's push clause; everything else goes only into the seat post's round record. At-once cases ① security finding, ② p0 still unclaimed one round after the name escalation or a self-dispatched stop-bleed, ③ main or the merge queue red over 2 h with no fix PR in flight. End-of-round cases, merged into at most one push per round: ④ human-floor decision (names the decision card and the one letter), ⑤ release (new blocker, RC retest regression, Version PR waiting only on the maintainer's merge), ⑥ a lane that cannot move (p1 waiting over 4 h on a seatless lane; name escalation unanswered over 6 h where the reclaim needs the maintainer's word; an open PR with no update for over 6 h after the seat knock goes unanswered past the next round, with the addendum's exclusions), ⑦ a permission block, never done in the seat's place. Not pushed: grading, healing half-states, unlocks, contract reviews (PASS/FAIL go to the lane), doorbell answers, a green health reading.

Ruling of record (card body, maintainer verbatim): 「由分诊每天出一次中文进度表，频率太低，而且有问题的时候要找我，没问题我也不需要看」 · 「需要我协助的时候你应该推送消息给我 os-zhuang，但是需要和我沟通哪些情况下」 · 「同意推送清单，立卡？」; addendum comment 6568980556: 「pr 超过 6小时没更新，如果不是带决策或者是被阻塞，应该都属于有问题吧」. The raise follows the ratchet header's ruling-of-record path (ruling record 6184799323) by the measured delta only, on the seat's reading that these words order the text into lanes/triage.md.

Line budget (measured on a418225)

File	Before	After	Ceiling
lanes/triage.md	8	15	8 → 15 (+8 landed, −1 paid in place)
triage-duties.md	102	99	102 → 99 (lowered)
both	110	114	net +4

New lines in lanes/triage.md, bytes against the 120-byte line budget: 117, 113, 106, 118, 118, 118, 120, 111.

Deletions and the surviving home of each

Deleted text	Surviving home	Reader
lanes/triage.md scope line (objectstack incl. ui/, first touch, cross-lane dedupe, no app-repo triage seat)	cross-repo-coordination.md lines 4, 9, 11, 31; business rationale in filing-gate.md 24–26	triage seat, before every routing and dedupe (SKILL.md index)
triage-duties.md 10–11: the owed-review reading set, same-round claim-then-take, isolated subagent same-shape record, ⛔ no tier downgrade	contract-review.md lines 21, 23–27, 32, 45–46	triage seat, when it takes an owed review
triage-duties.md 94: the trigger "two reds on untouched packages in an hour, any seat flips to report-only"	AGENTS.md Build & Test ("any seat flips a gate that reds untouched packages twice in an hour to report-only before any ruling")	every seat
triage-duties.md 101: report only at milestones, real blockers, when asked	superseded by the push list (lanes/triage.md line 8); its tail (intermediate state and rate-limit detail go only to the seat post and trigger) kept on the new pointer line	triage seat
triage-duties.md 102: human-floor items still need explicit words, batched not one by one	explicit words: SKILL.md 216 and the needs-user-decision row (human only) plus the human-floor list in triage-duties.md; batching: the push list's one merged end-of-round push	triage seat

Two wordings were tightened in the merged lines without changing the rule: 「不同意即翻回」 → 「否即翻回」, 「同笔送待决策汇总入下批,附两红 run」 → 「同笔附两红 run 入下批待决」.

Tests

No code path changed; check-skill-line-ratchet.mjs changed only in two CEILINGS rows and their comments. node scripts/pm/check-skill-line-ratchet.mjs --self-test :: exit 0.

Acceptance notes

• seat-lifecycle.md still tells every seat to send a Chinese round-report table by chat each round; the new rule overrides it for triage only. Left alone on the seat's order (out-of-scope finding, carrier: seat).
• PR pm-dispatch(execution): an idle lane seat does four pre-triage chores on its own cards, never grades, and rings the triage mailbox instead of waiting #28935 also edits triage-duties.md (adds one line, deletes the round-report line); no overlap with the lines here. Whichever lands second re-measures the ceiling.
• No new gate and no new label.

Verification

dispatch-gates derivation and reconciliation (head a418225)

node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack :: exit 0 — 39 commands (the ratchet script in the diff widens the derivation). All 39 run, each exit 0; pnpm --filter @objectstack/lint run check:doc-formula-expressions first answered exit 3 (prerequisite: unbuilt formula/lint), re-run after turbo run build --filter=@objectstack/formula --filter=@objectstack/lint (lock VERDICT command-exit 0) :: exit 0. pnpm check:pm-dispatch-gates :: exit 0 (fast tier, 1937 cases). check:pm-skill-ratchet :: exit 0.

node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --ran FILE :: exit 0 — "39 derived famil(ies) accounted for — 39 run, 0 NOT-MEASURED (a DERIVED zero — all 39 recorded an exit code and none of them is 3)".

---

Generated by Claude Code

[objectstack-fleet]

domain:skills#1 · huangyiirene · session_01J8VVAnMUWA9rufXzFaEh4J

Contract-review-claim: session_01J8VVAnMUWA9rufXzFaEh4J · head a418225 · 2026-11-02T05:17Z

The seat review runs at CONTRACT_REVIEW_TIER in an isolated subagent (the seat session is not at that tier).

[objectstack-fleet]

domain:skills#1 · huangyiirene · session_01J8VVAnMUWA9rufXzFaEh4J

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: a418225588d7e813612c43dd837158c9ad2e10f9
Local-runs: none
Scope: .claude/skills/pm-dispatch/references/lanes/triage.md, .claude/skills/pm-dispatch/references/triage-duties.md + scripts/pm/check-skill-line-ratchet.mjs (the gate that prices the two rule files, two CEILINGS rows and their comments), and the cited files read to judge the deletions and conflicts: .claude/skills/pm-dispatch/SKILL.md (lines 86, 216, 269, 273–274), references/seat-lifecycle.md (65–66), references/cross-repo-coordination.md (4, 7, 9, 11, 31), references/filing-gate.md (24–26), references/contract-review.md (11–12, 21–36, 45–46), references/platform-readings.md (385), AGENTS.md (140–141, Prime Directive 15), the triage seat post #6015 body (standing ruling of 2026-11-02), and PR #28935 head c53bdb0.

Reviewer: isolated contract reviewer for the domain:skills#1 seat, read-only on GitHub; inputs were card #28921's body and all five comments (addendum 6568980556, claim 6569310846, dev reports 6569436821 and 6569660808 as input not conclusions, seat decision 6569462916), the PR body, the net diff of a418225 against origin/main 135054e (merge-base ff88105, one commit), the head blobs read with git show for line and byte counts, and the head's check-runs; the #28935 overlap was computed with git merge-tree --write-tree on the two fetched heads, a git read with no checkout and nothing of the project run.

① Derived judgments

Acceptance set and public face: none. The diff touches no packages/spec/src/**, no published package, no protocol key, no gate logic. check-skill-line-ratchet.mjs changes only in two data rows of CEILINGS and their comments; --self-test and the verdict functions are untouched. Governed rule text (.claude/skills/**) is the face, hence this tier.

Judgment 1 — the seven cases, read against the card body, the addendum 6568980556 and the seat post #6015 standing ruling, each against the new lanes/triage.md lines 8–15 (head blob):

• No daily table / no chat round report: line 8 「推维护者仅限需其出手,通道照 SKILL.md;⛔ 无进度表与聊天轮报,余只进座位贴轮次记录。」 carries 「没问题我也不需要看」 and the card's "written in the round record on the seat post and pushed to no one". Correct. The card's channel format (one line, starting with what to do) is not restated; it survives verbatim-equivalent at SKILL.md 274 「一行 ≤200 字,先写动作与编号;每项恰一次,例行进展、刚问刚答与在场即见的不推。」, which line 8's 「通道照 SKILL.md」 points at, and SKILL.md is read in full by every seat session. Not a drop.
• Case 1 security finding: line 9 「①安全发现(机制只在会话给)」 — correct, mechanism stays in session.
• Case 2 unclaimed p0: 「②p0 点名升级后一轮仍无认领,或分诊已自派止血」 — both arms present. Correct.
• Case 3 long red: line 10 「③main 或合并队列红逾 2h 且无修复 PR 在飞」 — correct.
• Split and cap: 「即推:」 heads 1–3; line 10 「余下轮末合为一推,每轮至多一推:」 heads 4–7. Correct, one merged push per round.
• Case 4 human-floor decision: line 11 「④人工地板待裁,点名决策卡与所需一字」. The card's enumeration of what the human floor covers (closing on merits, downgrade, new gate, protocol key, security boundary, rule change) is not restated; it lives at triage-duties.md 17–19 (the human-floor and workload criteria, same file, same fire) and SKILL.md 216. "the one letter, or the yes/no" is tightened to 「所需一字」, which covers both. Tighter, not dropped. "A human-floor question still also gets its decision card" survives at SKILL.md 269 「卡先于弹窗:需裁决先落 needs-user-decision 卡」.
• Case 5 release: 「⑤新发版阻塞、RC 复测回退、Version PR 只待其合」 — three arms present. The card's sentence "Triage never merges it, never dispatches release workflows and never tags" is not restated; AGENTS.md Prime Directive 15 states it for every seat and says of itself that a rule binding every seat lives there, not only in a lane file. Tighter, no conflict.
• Case 6 lane that cannot move: lines 12–14 「⑥车道不动:p1 卡在无席车道逾 4h;点名升级逾 6h 无答且回收需其一句;open PR 逾 6h 未更新,先敲其席(无信箱按名)过下轮无答且回收需其一句;待裁或被具名 Blocked-by:、status:parked 卡、所骑栈阻者与信箱、Version PR 不算。」 Both card arms present (6 h name escalation with reclaim needing the maintainer's word; p1 over 4 h on a seatless lane). The addendum's stale-PR case is complete: the 6 h threshold, the knock on the PR's seat by name when it has no mailbox, the push only when the knock is unanswered past the next round and the reclaim needs their word, and all five exclusions (waiting on a decision; named Blocked-by:; status:parked; the stack it rides on; mailbox PRs and the Version PR). The addendum places it under case 6; so does the head. Correct.
• Case 7 permission block: line 14 「⑦席位或子代理被拒、需人做的动作,⛔ 不代做。」 — correct, no laundering.
• Not-pushed list: line 15 「不推:定级、治半状态、解锁、契约复核(PASS/FAIL 归车道)、门铃答复、健康绿读数。」 — all six items of the card (grading, healing half-states, unlocks, contract reviews with PASS and FAIL to the lane, doorbell answers, green health reading). Correct. It also matches [PM seat] triage (objectstack-wide) — 🟢 Routine · session_01AavokzJ5DndAwitDXvKy4U · trig_01NcnCtMS2tH46nUg1a5TaMJ (hourly) · state = body + the round records newer than it #6015's 「其余(定级、补状态、放行、复审、敲门应答、健康绿)只记轮次记录,不推」.
• No case dropped, none changed in substance. No gate, label, check or workflow is added; Governed Surface Queue Guard and PR-body guards are green on this head.

Judgment 2 — ratchet raise legality against the header discipline ("RAISING one is authorized by the RULING OF RECORD that orders text into the file … quotes that ruling comment's id and the measured line delta … moves the CEILINGS row in the same PR by at most that delta"):

• Counts read from the head blobs: lanes/triage.md 15 lines (main 8), triage-duties.md 99 lines (main 102). Rows: lanes/triage.md 8 → 15 (+7), triage-duties.md 102 → 99 (−3). The raise equals the measured net delta (+7) and is below the eight ordered lines (eight landed, one paid in place by the scope line). Lowering to 99 is the file's landed count. Correct.
• Byte cap: MAX_LINE_BYTES = 120, the check is bytes ≤ 120, a line on the cap passes. New lanes/triage.md lines measure 117, 113, 106, 118, 118, 118, 120, 111 bytes (line 14 sits exactly on the cap, which passes); the merged triage-duties.md lines 10, 93 and 99 measure 120, 120, 119. The PR body's eight numbers match the blobs. Lint & Repo Gates (which runs the ratchet) is green on this head.
• The ruling-of-record citation is present in the lanes/triage.md row comment: the three maintainer quotes from the card body 「由分诊每天出一次中文进度表，频率太低，而且有问题的时候要找我，没问题我也不需要看」, 「需要我协助的时候你应该推送消息给我 os-zhuang，但是需要和我沟通哪些情况下」, 「同意推送清单，立卡？」, plus the addendum comment id 6568980556 with its quote 「pr 超过 6小时没更新，如果不是带决策或者是被阻塞，应该都属于有问题吧」, the card number, and the delta ("Eight lines landed, one paid in place … measured delta is +7"). The PR body repeats the same quotes, the comment id and the delta table. This is the same convention as the 64 → 68 raise on filing-gate.md (maintainer 「立卡」 on a proposal, quoted on the card). Correct in substance.
• One inaccuracy in that comment, should-fix: "… 「同意推送清单，立卡？」, which name this file" — the maintainer's quoted words do not name lanes/triage.md. The file is named by the card body (the triage seat's filing: "Reader: domain:skills. The file is lanes/triage.md") and by the seat post [PM seat] triage (objectstack-wide) — 🟢 Routine · session_01AavokzJ5DndAwitDXvKy4U · trig_01NcnCtMS2tH46nUg1a5TaMJ (hourly) · state = body + the round records newer than it #6015's convention that standing rulings are written into lanes/triage.md (「收班时仍未写进 lanes/triage.md 的,随简报交维护者定去留」, and line 59 「规则卡 pm-dispatch(triage): triage pushes the maintainer only on the seven escalation cases (security, unclaimed p0, long red, decision cards, release, unreachable lanes, permission blocks) — no daily progress table #28921 待写进 lanes/triage.md」). The PR body says this correctly ("on the seat's reading that these words order the text into lanes/triage.md"); the ratchet comment should say the same, e.g. "and 「同意推送清单，立卡？」 on the push list whose card names this file". The 64 → 68 precedent is worded that way ("「立卡」 on the proposal that names the field").
• Cross-file arithmetic: lanes/triage.md is not a CROSS_FILE_MOVES destination, so no ruledRaises record applies, as the comment states. triage-duties.md is a declared source of the director.md move with was: 120; lowering it to 99 only widens that move's net source decrease (18 → 21) and pushes crossFileMoveTotalVerdict's net further negative. lanes/triage.md is not a participant, so its raise is outside condition (b) by construction. Nothing trips.

Judgment 3 — the five deleted or merged restatements, surviving home, reader, same fire:

• lanes/triage.md scope line 「范围 = objectstack(ui/ 在内)每张卡首触(业务理由与定级)与跨车道查重;应用仓无分诊席。」 — survives at cross-repo-coordination.md 4 「多车道仓 objectstack:中央分诊唯一生产 domain:*/type/定级」, 11 (ui/ lives in this repo, UI cards stay here), 31 「多车道仓:分诊唯一生产路由/type/查重/shadow/duplicate_of」, 9 「应用仓无分诊席」; the business-rationale reading at filing-gate.md 24–26 and 29; first-touch grading at triage-duties.md 23. Reader: the triage seat 「分诊席路由与查重前」 (SKILL.md index line 85), i.e. the same fire, before any routing. Not lost. Minor: the earlier 7 → 8 ratchet comment anchors line 5 「之后补」 to the now-deleted 「应用仓无分诊席」; the ordered text itself (line 5) is intact.
• triage-duties.md 10–11 「复审:每轮先跑 check-owed-reviews.mjs(open PR 含 draft,读交接/认领/hold 行),欠审同轮认领后接。」 and 「隔离子代理达档跑同形记录、署本席;本席也被配额拒 ⇒ PR 写一次留标,⛔ 永不降档审。」 → head line 10 「复审每轮先跑 check-owed-reviews.mjs,照 contract-review.md 接;本席也被配额拒 ⇒ PR 写一次留标。」. Surviving home contract-review.md: 23 (draft counts, no record and no live claim), 24–25 ((a)(b)(c) same-round take; check-owed-reviews.mjs read-only), 26–31 (claim and hold lines), 32 (same-shape record, seat and subagent alike), 45–46 (independence pair; ④ run by an agent other than the drafter). Reader: the triage seat when it takes an owed review, which the new line sends there by name, same fire. The one clause whose explicit triage spelling is gone is 「⛔ 永不降档审」: contract-review.md 21 states 「⛔ 不降档审」 for the dispatch seat's quota refusal and hands the PR to triage; 34–35 make a review at any tier other than the face's tier 「不成裁决」; platform-readings.md 385 「强制档不因临缺授权降档 … 不达档起隔离达档子代理或等档」 binds every seat but is not in the triage fire's index; the seat post [PM seat] triage (objectstack-wide) — 🟢 Routine · session_01AavokzJ5DndAwitDXvKy4U · trig_01NcnCtMS2tH46nUg1a5TaMJ (hourly) · state = body + the round records newer than it #6015 carries it as the standing ruling of 2026-11-01 (「永不降档、永不以 TIER_DEFAULT 自审加宽面」), which is not inherited across shifts. The rule survives by the tier-equals-face construction and the dispatch-seat parallel, so it is not lost; the explicit prohibition for this seat is weaker than on main. Should-fix: the merged line sits at 120 bytes, but a 116-byte spelling keeps the ⛔: 「- 复审每轮先跑 check-owed-reviews.mjs,照 contract-review.md 接;配额拒 ⇒ PR 留标,⛔ 不降档审。」 (measured 116 bytes; 「本席也被配额拒 ⇒ 留标,⛔ 不降档审」 without 「每轮」 measures 119).
• triage-duties.md 94–95 「例外二:一小时内两次红未触包的门禁,任一席位可先翻 report-only,三条硬条件缺一不可:」 / 「①只翻 report-only,不删门不跳测;②同笔送待决策汇总入下批,附两红 run;③不同意即翻回。」 → head line 93 「例外二:AGENTS.md 报告态翻转,须①不删门不跳测;②同笔附两红 run 入下批待决;③否即翻回。」. The deleted trigger and actor survive at AGENTS.md 140–141 "any seat flips a gate that reds untouched packages twice in an hour to report-only before any ruling". Reader: every seat, every session (root instruction file). The three hard conditions are all still in the triage file; the two tightened wordings (「否即翻回」, 「同笔附两红 run 入下批待决」) keep the rule. Not lost.
• triage-duties.md 101 「汇报只在里程碑、真阻塞、被问到时;中间状态与限流/重试类细节只进座位贴和触发器。」 → the first clause is superseded by the push list (lanes/triage.md 8–15, the card's own purpose); the tail survives on head line 99 「推送维护者只照 lanes/triage.md 推送清单;中间状态与限流/重试细节只进座位贴和触发器。」. Reader: the triage seat at fire start (SKILL.md 86), same fire. Not lost.
• triage-duties.md 102 「人工地板项(发版、天花板、契约扩大)仍需明确字句,但攒批不逐项弹。」 — explicit words survive at SKILL.md 216 「人工地板与风险/工作量判据(triage-duties.md)任一命中即呈维护者,⛔ 不入 PM 裁量」, 269 (card before pop-up) and the needs-user-decision row 「只由人工裁」, plus lanes/triage.md 11 「点名决策卡与所需一字」; batching survives as 「轮末合为一推,每轮至多一推」 on line 10. Reader: the triage seat (SKILL.md every session; lanes/triage.md every session). Not lost.

Judgment 4 — conflicts:

• SKILL.md 273–274 PushNotification clause: the triage list is a specialisation of 「受管 PR 待批合、决策卡待字、发版、仅其可解之阻 ⇒ 另发 PushNotification」 (case 5 third arm is the governed PR awaiting merge; case 4 the decision card; case 5 release; cases 1–3, 6, 7 are blocks only the maintainer can lift), and lanes/triage.md 8 defers the channel to SKILL.md. 「每项恰一次」 is compatible with one merged end-of-round push. No conflict.
• seat-lifecycle.md 65 「每轮向维护者打中文轮次报告(chat 通道):issue → 判决 → PR 链接 → 备注的表。」 is read by every seat every fire and now contradicts lanes/triage.md 8 「⛔ 无进度表与聊天轮报」 for the triage seat. The lane charter is the more specific text and the maintainer's words rule for triage; line 66 already carves the project-seat exception, so a per-seat carve-out is the established pattern. Not a must-fix for this card (the acceptance names lanes/triage.md and the pointer only; the seat ruled it out of scope), but see ③: it should be filed, not left as "noted, not filed".
• AGENTS.md Prime Directive 15: the head keeps 「Version PR 只待其合」 as a push trigger and never as an action; no release act is authorised. No conflict.
• Seat post [PM seat] triage (objectstack-wide) — 🟢 Routine · session_01AavokzJ5DndAwitDXvKy4U · trig_01NcnCtMS2tH46nUg1a5TaMJ (hourly) · state = body + the round records newer than it #6015 standing ruling (line 59): every case, the split, the cap and the not-pushed list agree with the head; [PM seat] triage (objectstack-wide) — 🟢 Routine · session_01AavokzJ5DndAwitDXvKy4U · trig_01NcnCtMS2tH46nUg1a5TaMJ (hourly) · state = body + the round records newer than it #6015 lacks the stale-PR arm because the addendum came later, and the head carries it. No conflict.
• PR pm-dispatch(execution): an idle lane seat does four pre-triage chores on its own cards, never grades, and rings the triage mailbox instead of waiting #28935 (head c53bdb0): it edits triage-duties.md by adding one line after main 34 and deleting main 98 「轮次报告设呈裁清单专节」; its hunks only share context with this PR's changes. git merge-tree --write-tree of the two heads produced tree 9731a722 with no conflict; in that tree triage-duties.md is 99 lines against this PR's ceiling 99, no line over 120 bytes, and the ratchet row from this PR survives. So the second to land needs no re-measure and no ceiling edit in either order (if pm-dispatch(execution): an idle lane seat does four pre-triage chores on its own cards, never grades, and rings the triage mailbox instead of waiting #28935 lands first, this PR's 102 → 99 remains the landed count; if this PR lands first, pm-dispatch(execution): an idle lane seat does four pre-triage chores on its own cards, never grades, and rings the triage mailbox instead of waiting #28935's +1/−1 nets 0 at 99/99). Both PR bodies promise a re-measure anyway; it will read 99.

② Semver level

none. No published package is touched (.claude/skills/** and scripts/pm/** only); Clause-②: no and Metadata-protocol: none on the PR body are correct; no changeset is owed and skip-changeset is the right carrier for this repo (Check Changeset skipped as designed). Fixes #28921 is correct: the card's whole deliverable (seven cases, split, cap, not-pushed list, 6 h addendum, pointer line, no gate, no label) lands in this PR, so merge should close it.

③ Boundary flags

• Dev flag (6569436821 open question, A/B/C): answered by the seat's decision 6569462916 for option A; the head implements A exactly (list in lanes/triage.md, 8 → 15 by the measured +7, triage-duties.md 102 → 99, no gate, no label). Closed.
• Dev out-of-scope finding, carrier seat, "noted, not filed": seat-lifecycle.md 65 still orders a chat round-report table from every seat. This record agrees it is out of this card's scope and that the lane charter overrides it for triage, but a live contradiction between two files the same seat reads in one fire should not end as an unfiled note: should-fix, file the one-line follow-up card (reader domain:skills; a carve-out on seat-lifecycle.md 66, which has 10 bytes of slack, or on 65 with 15) and name it on this PR.
• Dev out-of-scope finding, pm-dispatch(execution): an idle lane seat does four pre-triage chores on its own cards, never grades, and rings the triage mailbox instead of waiting #28935 overlap: measured here (Judgment 4); no action owed.
• Dev deviations: the empty-branch push and the 39-command derivation are process notes, not boundary crossings; skipping assign and reviewer per the order is correct.
• Ratchet comment accuracy ("which name this file"): should-fix, wording only (Judgment 2).
• Explicit 「⛔ 永不降档审」 for the triage seat dropped from triage-duties.md 10: should-fix, 116-byte respelling given in Judgment 3.
• Author: PR user.login is objectstack-fleet[bot], so no login to name.

④ Design judgment

• lanes/triage.md against the card's scenario: a triage seat reading its charter at session start now knows, without any other file, when it may push the maintainer and when it may not; the at-once cases are the three that lose value by waiting an hour, the end-of-round cases are the four that batch. The 6 h stale-PR arm is placed where the addendum placed it, with its knock-first step, so a vacant seat (the measured instance, PR fix(metadata-protocol): the by-id OCC comparison answers only for a row the caller reads #28731) is pushed only after the knock fails. The text reads in the maintainer's register (「余只进座位贴轮次记录」) and is tighter than the card, not looser. Holds.
• triage-duties.md against the files it now defers to: each merged line points at a file the same seat already reads in the same fire (contract-review.md when taking a review, AGENTS.md always, lanes/triage.md always); no rule is reachable only by inference except the explicit downgrade ⛔ discussed above, which survives structurally. Holds, with the should-fix.
• check-skill-line-ratchet.mjs against its own header: the raise follows the ruling-of-record path exactly as written (ruling quoted, comment id cited, delta stated, row moved by at most the delta, deletions paying the rest), the lowering is the landed count, and the cross-file-move arithmetic is unaffected. The one defect is a relative clause that overstates what the maintainer's words name. Holds, with the should-fix.
• No gate, label, protocol key or acceptance set changes; rollback is a plain revert.

Check-runs at record time (2026-11-02T05:26Z, head a418225): 45 runs, 44 completed, 0 failed, 1 in progress (Test Core 1/8). Green: Lint & Repo Gates, TypeScript Type Check, Lint · PM dispatch-gates, Lint · repo & tooling gates, Lint · docs & code guards, Lint · CI release & engine gates, Lint · ESLint, the three scoped lint jobs, Type Check · workspace / source gates / consumer gates / debt ledger, Governed Surface Queue Guard, PR-body guards, Check PR Size, Auto Label, Dogfood Regression Gate, Test UI, Test Core 2/8, Plan Test Core shards, filter. Skipped by design: Check Changeset (skip-changeset), UI/E2E/docs/conformance/opt-in jobs. Clause ② of contract-review.md (all checks green) is read at queue time, not here.
Implemented-by: claude/project-thread-vzhnx0
Reviewed-by: session_01J8VVAnMUWA9rufXzFaEh4J

VERDICT: PASS

---

Generated by Claude Code

[objectstack-fleet]

domain:skills#1 · huangyiirene · session_01J8VVAnMUWA9rufXzFaEh4J

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: 510240a7913df0e8fe980847ecb5a67cb29116dc
Local-runs: none
Scope: .claude/skills/pm-dispatch/references/lanes/triage.md, .claude/skills/pm-dispatch/references/triage-duties.md + scripts/pm/check-skill-line-ratchet.mjs (the gate pricing the two rule files; the lanes/triage.md CEILINGS row comment is the delta's second hunk), and the cited files re-read for the delta: references/contract-review.md (21–22, 26, 34–35), the needs:contract-review section of the same file, and current origin/main 00d0668 (which now carries PR #28935).

Reviewer: isolated contract reviewer for the domain:skills#1 seat, read-only on GitHub. Delta review over record 6569884536 (PASS on a418225): the new head is one commit 510240a on top of a418225 (ancestor confirmed, no force-push), the dev's delta report 6570014885 was read as input, and the delta diff a418225..510240a7 touches exactly two hunks: one line of triage-duties.md and three comment lines of the lanes/triage.md CEILINGS row. The commit lands only findings from the prior record, so this is the same reviewer's delta review; the two should-fixes it answers are re-judged below and nothing else in the diff against main changed (git diff --stat origin/main...510240a7: the same three files, 30 insertions, 9 deletions, the one extra insertion being the comment re-wrap).

① Derived judgments

Acceptance set and public face: unchanged, none. The delta touches no packages/spec/src/**, no published package, no protocol key and no gate logic; check-skill-line-ratchet.mjs changes only inside one CEILINGS row comment, the row value stays 15.

Judgment 1 — should-fix 1 (ratchet comment attribution): the lanes/triage.md row comment now reads 「… and 「立卡」 on the push list (「同意推送清单，立卡？」) whose card names this file, plus the addendum comment 6568980556 「pr 超过 6小时没更新，如果不是带决策或者是被阻塞，应该都属于有问题吧」」. The naming of the file is attributed to the card, not to the maintainer's words, in the 64 → 68 precedent's form; all three maintainer quotes are kept verbatim, the addendum id 6568980556 and its quote are kept, and the delta sentence 「Eight lines landed, one paid in place … so the measured delta is +7; triage-duties.md fell 102 → 99 in the same PR, net +4 over both」 is unchanged. Resolved.

Judgment 2 — should-fix 2 (the dropped downgrade prohibition): triage-duties.md line 10 on the head reads 「- 复审每轮先跑 check-owed-reviews.mjs,照 contract-review.md 接;配额拒 ⇒ PR 留标,⛔ 不降档审。」 at 116 bytes. Against main's 「隔离子代理达档跑同形记录、署本席;本席也被配额拒 ⇒ PR 写一次留标,⛔ 永不降档审。」:

• 「⛔ 永不降档审」 → 「⛔ 不降档审」: the same prohibition; a ⛔ clause is absolute without 「永」, and it is the spelling contract-review.md 21 already uses for the dispatch seat, so the two seats now read one spelling. Same meaning.
• 「本席也被配额拒」 → 「配额拒」: the subject is this seat by position (the line is the triage seat's review duty) and the parallel to the dispatch seat's refusal is carried by 「照 contract-review.md 接」 and line 21 there. Nothing lost.
• 「写一次留标」 → 「留标」: the dropped 「一次」 said the pending-review mark is written once. The mark is the needs:contract-review label, whose carrier is the PR and which a second application cannot duplicate (contract-review.md 等复核标记 section: 「单载体 = PR」, hung once at ACCEPT, lifted once on PASS), and the handoff line it travels with is per head (contract-review.md 22 names the head; 26 「一 head 一行」 for the claim discipline the same reader follows). So the once-ness is carried by the carrier and the per-head line, and the triage seat reads both in the same fire when it takes a review. Nothing lost by the fit to 120 bytes.
• The isolated-subagent clause 「隔离子代理达档跑同形记录、署本席」 was already judged in 6569884536 as surviving at contract-review.md 32 and 45–46; unchanged here.
  Resolved.

Judgment 3 — counts, bytes and the #28935 merge: head blobs measure lanes/triage.md 15 lines (ceiling 15) and triage-duties.md 99 lines (ceiling 99), no line over 120 bytes in either file, line 10 at 116. PR #28935 merged to main at 05:33Z (merge commit 00d0668); main's triage-duties.md is 102 lines at ceiling 102 (its +1/−1 netted zero). git merge-tree --write-tree origin/main 510240a7 merges clean (tree ee73cac1) with no conflict; in that tree triage-duties.md is 99 lines against this PR's ceiling 99, lanes/triage.md 15 against 15, no line over 120 bytes, and both CEILINGS rows carry this PR's values. The ratchet will read 15/15 and 99/99 after landing; no re-measure is owed, and the PR bodies' promise to re-measure is already answered by this reading. Lint & Repo Gates (which runs the ratchet) is green on 510240a.

Judgment 4 — no other change: the delta diff is the two hunks above; the seven cases, the split, the cap, the not-pushed list, the 6 h addendum (lanes/triage.md 8–15) and the other four deletions' surviving homes are byte-identical to a418225 and stand as judged in 6569884536. No gate, label, check or workflow added.

② Semver level

none, unchanged. Clause-②: no, Metadata-protocol: none, skip-changeset remain correct; no published package is touched; Fixes #28921 remains correct.

③ Boundary flags

• Prior should-fix 1 (ratchet comment attribution): closed by 510240a.
• Prior should-fix 2 (explicit 「⛔ 不降档审」 on the triage review line): closed by 510240a in the 116-byte spelling.
• Prior should-fix 3 (seat-lifecycle.md 65 chat round-report table vs lanes/triage.md 8, file a follow-up card): not addressed by this commit and not owed by it; still owed by the seat as a one-line follow-up card (reader domain:skills). Carried, not a FAIL.
• Dev deviation in 6570014885: the PR body's Verification block still names head a418225 while the readings are identical on 510240a. Body prose, not a record field; the seat corrects it in its ACCEPT, as the coordinator noted. No boundary crossed.
• Author: PR user.login is objectstack-fleet[bot]; no login to name.
• No new flags or open questions in the delta report.

④ Design judgment

• triage-duties.md line 10 against the files it defers to: the triage seat now reads, on its own duty line, that a quota refusal means leave the mark and never review below the face's tier, and 「照 contract-review.md 接」 sends it to the claim, hold and handoff discipline it needs in the same fire. The line is now at least as strong as main's for the reader this card serves, and shorter. Holds.
• check-skill-line-ratchet.mjs against its header: the raise still follows the ruling-of-record path (quotes, comment id, measured delta, row moved by at most the delta, deletions paying the rest), and the comment no longer overstates what the maintainer's words name. Holds.
• lanes/triage.md: unchanged from the judged head. Holds.

Check-runs at record time (2026-11-02T05:43Z, head 510240a): 41 runs, 41 completed, 0 failed, 0 in progress. Green: Lint & Repo Gates, TypeScript Type Check, Lint · PM dispatch-gates, Lint · repo & tooling gates, Lint · docs & code guards, Lint · CI release & engine gates, Lint · ESLint, the three scoped lint jobs, Type Check · workspace / source gates / consumer gates / debt ledger, Governed Surface Queue Guard, PR-body guards, Check PR Size, Auto Label, Dogfood Regression Gate, Test UI, Test Core (both shards and the aggregate), Plan Test Core shards, filter. Skipped by design: Check Changeset (skip-changeset) and the UI/E2E/docs/conformance/opt-in jobs. All checks green, so contract-review.md clause ② is met on this head as well.
Implemented-by: claude/project-thread-vzhnx0
Reviewed-by: session_01J8VVAnMUWA9rufXzFaEh4J

VERDICT: PASS

---

Generated by Claude Code

[objectstack-fleet]

domain:skills#1 · huangyiirene · session_01J8VVAnMUWA9rufXzFaEh4J

维护者速读

• 改了什么: lanes/triage.md 写进维护者的推送清单:不发进度表、不发聊天轮报;只有需要维护者出手时才推,并分「即推」三项(安全发现、p0 无人认领或已自派止血、main / 合并队列红逾 2h 无修复 PR)与「轮末合一推、每轮至多一推」四项(人工地板待裁、发版、车道叫不动含 PR 逾 6h 未更新、权限被拒);另列不推的事项。triage-duties.md 合并或删去五行已有家的重述,加一行指针。
• 为什么改: 落实你的裁决「同意推送清单，立卡？」与补充「pr 超过 6小时没更新…应该都属于有问题吧」。
• 风险与回滚: 纯内部指令文本;lanes/triage.md 上限 8 → 15(实测 +7,依卡上的裁决记录),triage-duties.md 102 → 99,合计 +4 行;不加门禁或标签。回滚即 revert。
• 先看: lanes/triage.md 第 8–15 行;triage-duties.md 第 10 行与新指针行;check-skill-line-ratchet.mjs 两行上限的注释。
• 两分钟验证: node scripts/pm/check-skill-line-ratchet.mjs 绿;grep -n "无进度表" .claude/skills/pm-dispatch/references/lanes/triage.md 命中第 8 行。
• 需要你做的: 无。Tier S,本席落地。
• 席位意见: ACCEPT。七类情形与 6h 补充逐条都在,被删五行各有同席同 fire 读到的家;补丁轮修正了上限注释的出处措辞,并补回「⛔ 不降档审」。

席位核对 — ACCEPT · domain:skills#1 · head 510240a7913df0e8fe980847ecb5a67cb29116dc

1. 形态: draft → main,Fixes #28921,无他卡关键词;作者 objectstack-fleet[bot](中继)。
2. 前提实测: main 上 lanes/triage.md(8 行)与 triage-duties.md 无推送清单(复核记录 6569884536 读树确认,delta 记录 6570082751 沿用)。
3. 门禁 check-runs(钉 head 510240a7): success 25 / skipped 16 / 无红无在跑。check-governed-merges --pr 28972:Tier S(.claude/** ×2;scripts/pm/check-skill-line-ratchet.mjs 只动 CEILINGS 两行数据与注释),+30/−9。与已合入的 pm-dispatch(execution): an idle lane seat does four pre-triage chores on its own cards, never grades, and rings the triage mailbox instead of waiting #28935 合并无冲突,合并后 99/99、15/15(delta 记录 ③)。
4. 接受的偏差: 抬上限依「立卡即裁决记录」惯例(卡正文引维护者原话 + 补充评论 6568980556,同 filing-gate.md 64 → 68 先例),幅度等于实测增量;合并行去掉「一次」以守 120 字节,delta 复核判其不丢义。PR 正文 Verification 一节所引 head 应为 510240a7(读数与 a4182255 相同)。
5. 范围外去向: seat-lifecycle.md 第 65 行「每轮向维护者打中文轮次报告」与本 PR 的「⛔ 无进度表与聊天轮报」对分诊冲突;改 pm 规则须维护者点名,已呈维护者是否立卡(⛔ 席位自提不立)。Skills-docs: 不适用(Metadata-protocol: none)。
6. 落地条款: Tier S,现 head 同形达档 PASS 在案(6570082751)+ 全绿 ⇒ 本席经中继 pr_ready + automerge_enable。

---

Generated by Claude Code