skills(pm-dispatch): clause-② cards build at the default tier and are gated by the contract review at CONTRACT_REVIEW_TIER (#16905) - #16915

维护者速读(补丁轮后,head e1edc4196)

• 改了什么:.claude/skills/pm-dispatch/ 下 4 个文件、11 行改 11 行,每个文件行数不变(SKILL.md 811、contract-review.md 60、core-rules.md 150、lanes/cli.md 29),每行 ≤120 字节。强制条款②从「凡放宽接受集或扩大公开面的卡一律契约复审档」改为「默认判断档施工、契约复审档复核」;references/** 里五处仍写「条款②卡派/锁契约复审档施工」的句子(contract-review.md :10 :13 :15、core-rules.md :113、lanes/cli.md :18)同轮改成只定复核档;删掉两句因此失效的特例(额度耗尽豁免及于条款②;入队闸门里的「派发档位低于契约复审档」谓词);〈轮次报告与节奏〉加一行计数器。条款①、needs:contract-review 双载体纪律、降档保险丝、lanes/spec.md 分派判据、四轴框架块一字未动。
• 为什么改:你的原话「开skill卡,在裁决落地之前,你直接按新规则试行」;架构本已把契约复审定为低档派发的补偿(references/contract-review.md 降档保险丝「豁免对象是派发,复核正为补偿低档派发而存在」)和 2026-08-20 的「fable 当审计师用,不当施工队用」;条款②是唯一还把贵档压在施工这条大项上的地方。施工与复审 token 量级实测 33 万 vs 21 万,档位只改单价。复审席指出 references/** 五处与 SKILL.md 互相矛盾,总监席把文件面扩到那三份,本轮一并改齐。
• 风险与代价(含回滚):① 主风险 = 默认档施工把公开面判错的概率是否高于达档施工,这一点没有测量,卡片自己的基线是 1 FAIL / 1(样本一,不是基线)。计数器因此是改动的另一半:轮报席在复审清单记默认档施工的条款②复审 FAIL 率,与改制前达档施工的历史值比;超出 ⇒ 立决策卡交你定回退,席位不自裁。② :639 入队闸门被重写后的后果要你知情:凡路径肢命中(diff 触 packages/spec/src/**)的 PR,不论申报 Clause-②: yes 还是 no、不论施工档位,入队唯一出口都是席内契约复审档的 PASS 在案 —— 比 skills(pm-dispatch): the clause-② enqueue gate at SKILL.md:640 has two incompatible readings — one of them freezes the whole domain:spec lane whenever fable is out, and makes check-clause2-carriers --pair exit 0 unreachable #16912 记录的 spec 车道现行做法(--pair 退出 0 即入队)更严。本 PR 不裁 skills(pm-dispatch): the clause-② enqueue gate at SKILL.md:640 has two incompatible readings — one of them freezes the whole domain:spec lane whenever fable is out, and makes check-clause2-carriers --pair exit 0 unreachable #16912。③ references/contract-review.md 同时被排队卡 pm-dispatch: a contract-review verdict without the label handoff is invisible to every seat — name the FAIL end-state in one sentence, and add H51 for "verdict posted on this head, carrier still on" #16836(未认领、无 PR)申报,其认领者需 re-merge。回滚 = revert 本 PR 两个 commit(4 文件、11 行)。
• 席位意见:席位推荐落地。四轴:业务需求 — 派发席实测在付的钱在施工端;长远 — 把特例并进已有的「降档施工的补偿控制 = 复核席跑契约复审档」规则,且五份拷贝与主文一致;防 AI 犯错 — 闸门只收紧不放松,清标仍只由达档裁决完成,错误由复审拦而不是靠施工档位;创业聚焦 — 不加脚本、不加标签、不加状态,只加一行计数。
• 你要做的:读 PR 正文里 11 行的 before/after 与「每条删行的内容现在住哪」;确认上面 ② 的闸门后果是你要的(不是 ⇒ :639 需插「判定 yes」,回补丁轮);同意则 approve + 人工直合;不同意直接关 PR,文本无副作用。此 PR 是 强制条款① 面(SKILL.md 主文件),已按契约复审档施工,改动不豁免自身落地。

Generated by Claude Code

Contract review (claude-fable-5-1, isolated seat) — PR #16915 @ bd26a0839

Verdict: CHANGES REQUIRED

Isolated seat: ref fetched to refs/review/16915, diffed against the merge-base 0939ba008 (= origin/main at dispatch; origin/main is now 70f7d6d73, 4 commits ahead, merge-tree onto it is clean). Nothing in the PR body was trusted; every number below is re-measured on an extracted copy of bd26a0839.

Ruling implemented: yes, exactly the ruling zone — and no more. Card #16905 records the maintainer's instruction verbatim: 「开skill卡，在裁决落地之前,你直接按新规则试行。」 — the authorisation for a trial, not a ruling on the text (the claim 5588071161 says the same: 「⛔ Not pre-approved … The PR is the place the ruling is made, by merge」). The claim's ruling zone — governed four-piece treatment · narrowing only · 条款① untouched · carrier discipline untouched · tier fuse untouched · tripwire mandatory · ratchet paid by deletion · skip-changeset · file surface SKILL.md (+ contract-review.md / lanes/skills.md only conditionally) — is implemented: 条款① (:509) byte-identical; :646-647 (carrier on card + PR) byte-identical; references/contract-review.md not in the diff, so :60 (降档保险丝) untouched; tripwire at :675; one deletion (:522) pays for the one insertion (:675); skip-changeset hung. The conditional surfaces were not needed and were not touched.

Governed paths touched: YES — maintainer-only merge. One file, .claude/skills/pm-dispatch/SKILL.md (.claude/skills/**, Prime Directive #14; the file is itself a 强制条款① surface). 5 insertions / 5 deletions, three hunks (@511, @637, @673).

Per-line reading (merge-base 0939ba008 → head bd26a0839)

Gate readings (extracted tree of bd26a0839, dependency-free scripts run by this seat)

• node scripts/pm/check-skill-line-ratchet.mjs → exit 0, SKILL.md is 811 lines (ceiling 811; headroom 0). Re-measured: 811 lines; changed lines 514/517/524/639/675 are 108 / 97 / 79 / 118 / 120 bytes (:675 is exactly at the 120-byte budget).
• node scripts/check-skill-frame-sync.mjs → exit 0, 2 copies … structurally isomorphic.
• node scripts/pm/check-skill-id-lint.mjs → exit 0, 26 file(s) clean.
• node scripts/check-skill-frame-freshness.mjs → exit 0 but proves nothing in an extracted tree (freshness could NOT be established: there is no origin/main). Substituted: frame block :734-752 md5 22f2339f0acb64cdb50c7adc9db681c3 on merge-base, on head, and on current origin/main 70f7d6d73 — identical all three.
• Diff vs merge-base: exactly one file; origin/main since the merge-base does not touch .claude/skills/**, .claude/agents/os-dev.md or AGENTS.md.
• Commit bd26a0839 trailers: Co-Authored-By + Claude-Session only — no closing keyword / Part-of / Refs bound to a card (RULE 2 of scripts/check-partof-closing-keyword.mjs); CI check Part-of PR must not also close its card green.
• Four governed pieces: draft: true ✓ · labels needs-user-decision + skip-changeset (+ documentation, size/s) ✓ · 速读 comment 5588349195 ✓ · reviewers hotlong, os-zhuang ✓. mergeable_state: blocked (draft + required review — expected).
• CI (get_check_runs, 32 runs): polled once at 16:28Z — 30 of 32 completed (16 success, 14 skipped, 0 failures); 2 still in_progress at 16:28Z: Lint & Repo Gates (job 102142928896) and Type Check · workspace (job 102142928428), both started 16:19Z. Not a green reading; the director seat re-reads CI before recording. This diff is markdown-only, outside eslint's files population, so neither job's verdict can vary with it.

Findings

1. [major] The protocol still contradicts itself on the clause-② build tier inside references/**, and none of it is in pm(dispatch-gates): clause-② prose in dispatch-gates.mjs (docblocks + every --tier printout) and core-rules.md:124 still describes a BUILD-tier mandate once the 强制条款② narrowing lands #16914. After this PR, SKILL.md:514 says default-tier build, while on the same head:

   • references/contract-review.md:10 「Clause-②: yes | no 按设计临时:只定派发档位的保守方向」 — the declaration's stated purpose is to set the dispatch tier, which it no longer does;
   • references/contract-review.md:13 「新导出符号或已发布载荷上的新键恒 yes,直接锁契约复审档」 — a claim-time lock of the ceiling for the build;
   • references/contract-review.md:15 「claim 拿不准 ⇒ 按 yes 派契约复审档」 — an explicit instruction to dispatch at CONTRACT_REVIEW_TIER;
   • references/core-rules.md:113 「派发调用 … 协议语义面与契约变更卡恒走契约复审档」 — an explicit dispatch mandate for contract-change cards;
   • references/lanes/cli.md:18 「条款②从卡内容判且优先于它」 — clause ② overriding the path-derived tier floor only makes sense if clause ② sets a build tier.

   pm(dispatch-gates): clause-② prose in dispatch-gates.mjs (docblocks + every --tier printout) and core-rules.md:124 still describes a BUILD-tier mandate once the 强制条款② narrowing lands #16914 exists and covers exactly what its body lists: scripts/pm/dispatch-gates.mjs prose (three sites) and core-rules.md:124. It does not cover the five lines above, and the PR body's "Untouched" paragraph presents references/contract-review.md as consistent with the change (「the tier fuse … untouched」) without disclosing that :13/:15 of the same file mandate the opposite build tier. The claim fenced contract-review.md out of the file surface unless the counter lived there, so the dev could not edit it — but the claim also says 「stop on breach; explain in the report」, and the report explains neither. A PM reading contract-review.md:15 after this lands is told to dispatch at the ceiling; one reading SKILL.md:514 is told not to.
   Expectation: either (a) the director widens the surface and this PR rewrites those five lines in place — all five can be reworded line-neutrally (e.g. :15 「claim 拿不准 ⇒ 按 yes 挂标走契约复审档复核」, :13 「…恒 yes,直接锁契约复审档复核」, core-rules :113 「协议语义面卡恒走契约复审档,契约变更卡恒过契约复审档复核」), so the ratchet (contract-review.md 60/60, core-rules.md 150/150, cli.md 29/29) is not disturbed; or (b) the maintainer accepts the interim contradiction explicitly, in which case pm(dispatch-gates): clause-② prose in dispatch-gates.mjs (docblocks + every --tier printout) and core-rules.md:124 still describes a BUILD-tier mandate once the 强制条款② narrowing lands #16914 must be widened to name these five sites and the PR body's "Untouched" paragraph must disclose them. (a) is the smaller total change; the choice is the maintainer's.

2. [minor] :639 — direction of the gate change is not "outcome unchanged" as the body says; it is the stricter of skills(pm-dispatch): the clause-② enqueue gate at SKILL.md:640 has two incompatible readings — one of them freezes the whole domain:spec lane whenever fable is out, and makes check-clause2-carriers --pair exit 0 unreachable #16912's two readings and stricter than skills(pm-dispatch): the clause-② enqueue gate at SKILL.md:640 has two incompatible readings — one of them freezes the whole domain:spec lane whenever fable is out, and makes check-clause2-carriers --pair exit 0 unreachable #16912's own proposal. Old text, literal (Reading A of skills(pm-dispatch): the clause-② enqueue gate at SKILL.md:640 has two incompatible readings — one of them freezes the whole domain:spec lane whenever fable is out, and makes check-clause2-carriers --pair exit 0 unreachable #16912): any path-limb hit + below-ceiling dispatch ⇒ hard bar, no exit. Old text as the instrument reads it (Reading B): the diff is adjudicated, any tier can pass. New text: any limb hit ⇒ an in-seat CONTRACT_REVIEW_TIER PASS on record is the only exit, regardless of tier. Consequences: (i) for an at-tier build the gate line itself now demands a PASS, where before only :643 / 落地前检① did — narrowing; (ii) for a below-tier build the absolute bar becomes a PASS requirement — the only reading compatible with :514, and it removes the lane freeze skills(pm-dispatch): the clause-② enqueue gate at SKILL.md:640 has two incompatible readings — one of them freezes the whole domain:spec lane whenever fable is out, and makes check-clause2-carriers --pair exit 0 unreachable #16912 reports; (iii) for a Clause-②: no PR whose diff merely touches packages/spec/src/** (path limb), a PASS is now required — skills(pm-dispatch): the clause-② enqueue gate at SKILL.md:640 has two incompatible readings — one of them freezes the whole domain:spec lane whenever fable is out, and makes check-clause2-carriers --pair exit 0 unreachable #16912 records that fix(spec): isValueDomainMember refuses an off-vocabulary domain instead of failing open #16899 / docs(spec): say on the title grade which INPUT it describes #16901 were enqueued on --pair exit 0 without a PASS, and skills(pm-dispatch): the clause-② enqueue gate at SKILL.md:640 has two incompatible readings — one of them freezes the whole domain:spec lane whenever fable is out, and makes check-clause2-carriers --pair exit 0 unreachable #16912's suggested wording would only bar a yes verdict below tier. lanes/spec.md:33 「条款②路径肢照命中,入队前仍过契约复核」 already states (iii) for the pre-approved ledger class, so the new line is consistent with the lane's own rule and with 「按更严的一条行动」 — defensible; but the 速读 tells the maintainer 「闸门不变」, which is only true against Reading A + :643. So skills(pm-dispatch): the clause-② enqueue gate at SKILL.md:640 has two incompatible readings — one of them freezes the whole domain:spec lane whenever fable is out, and makes check-clause2-carriers --pair exit 0 unreachable #16912 describes exactly this ambiguity and this line resolves it. Not deciding skills(pm-dispatch): the clause-② enqueue gate at SKILL.md:640 has two incompatible readings — one of them freezes the whole domain:spec lane whenever fable is out, and makes check-clause2-carriers --pair exit 0 unreachable #16912 here.
   Expectation: the 速读 / PR body says in one sentence that every path-limb hit now needs an at-tier PASS to enqueue, including Clause-②: no spec-lane PRs, so the maintainer rules on that knowingly; if that is not intended, :639 needs 「判定 yes」 inserted before the PASS clause (fits the 120-byte budget only with compression).

3. [minor] :675 — the recording half is executable, the revert half is not. Who/where/when: the round-report writer, in the 复审清单 section that contract-review.md:40 does define (「轮次报告设复审清单专节,形状与代裁清单同为强制审计」), every round — fine. But 「实质高于达档施工 ⇒ 条款②回退」 names no actor, no threshold, and its comparison arm 「达档施工」 is produced by nothing in the rule once every clause-② card builds at the default tier: the at-tier sample stays at the historical 1/1 unless a PM keeps exercising :508 「按卡取用」 discretion. 「基线待测」 admits this but does not say who will measure it. The revert itself is a governed-text change, so the seat's only executable action is to escalate — the line does not say so.
   Expectation: non-blocking. Acceptable as-is if the maintainer treats the counter as a report line that feeds a human decision; otherwise the line should name the escalation (决策箱) as the trigger's output.

4. [info, non-blocking] PR body carries two attribution footers (a bare footer, then --- + footer, appended by the platform on create). AGENTS.md:432 「⛔ never re-send a body that already carries an appended footer」 — correctly left alone.

5. [info] Process. The freshness gate cannot be run meaningfully outside a git checkout; the md5 substitute above is the reading. The dev's claim recorded CONTRACT_REVIEW_TIER for the build, as 条款① requires for this file; this seat did not verify the transcript (contract-review.md:56) — that reading belongs to the dispatching seat.

Maintainer-only merge: yes

Generated by Claude Code

Contract review (claude-fable-5-1, isolated seat) — PR #16915 @ e1edc4196

Verdict: CHANGES REQUIRED (record-only — the head's text needs no further edit; the one owed item is the PR body, see finding 1)

Isolated seat, re-review of the moved head. Ref fetched to refs/review/16915 = e1edc4196 (as expected). Merge-base with origin/main is 70f7d6d73 = current origin/main (the branch merged main in b5be3f049); merge-tree --write-tree onto origin/main is clean. Nothing in the PR body or the 速读 was trusted; every reading below is re-measured on git show/git diff of the ref and on an extracted copy of the head tree.

Prior findings (review 5588464964 @ bd26a0839)

F1 per-line reading (origin/main → e1edc4196)

Measurements (extracted tree of e1edc4196)

• Line counts head / origin/main: SKILL.md 811/811 · contract-review.md 60/60 · core-rules.md 150/150 · lanes/cli.md 29/29.
• Diff vs origin/main: exactly 4 files, 11 insertions / 11 deletions (SKILL.md 5/5, contract-review.md 3/3, core-rules.md 2/2, cli.md 1/1). Patch commit e1edc4196 alone (vs its parent b5be3f049): the same 4 files, 7/7. The merge commit b5be3f049 adds nothing of its own (diff bd26a0839..b5be3f049 -- .claude AGENTS.md CLAUDE.md scripts/pm is empty; diff origin/main..b5be3f049 is only bd26a0839's SKILL.md 5/5). The raw bd26a0839..head --stat shows 20 files because it spans the main merge — all 16 extra files are origin/main's, none is the PR's. scripts/pm/**, .claude/agents/os-dev.md, AGENTS.md, CLAUDE.md, content/docs/releases/** untouched.
• Every added line ≤ 120 B: max added line is 119 B (:675); changed lines measured individually: 108 / 97 / 79 / 118 / 119 / 112 / 116 / 116 / 119 / 115 / 109 B.
• Frame block SKILL.md:734-752 md5 22f2339f0acb64cdb50c7adc9db681c3 on head and on origin/main 70f7d6d73 — unchanged.
• node scripts/pm/check-skill-line-ratchet.mjs → exit 0; the four touched files read 811 (ceiling 811), 60 (ceiling 60), 150 (ceiling 150), 29 (ceiling 29), headroom 0 each; total ceilings down 9 lines.
• node scripts/check-skill-frame-sync.mjs → exit 0, 2 copies … structurally isomorphic across 2 files.
• node scripts/pm/check-skill-id-lint.mjs → exit 0, 26 file(s) clean.
• Governed pieces: draft: true ✓ · labels needs-user-decision, skip-changeset, documentation, size/s ✓ · 速读 = one comment (5588349195), edited in place 16:40:34Z, now titled 「补丁轮后,head e1edc4196」 ✓ · reviewers hotlong, os-zhuang ✓ · mergeable_state: blocked (draft + required review, expected).
• Commits (3): bd26a0839 and e1edc4196 carry Co-Authored-By + Claude-Session only; b5be3f049 is a plain Merge remote-tracking branch 'origin/main' (no trailers). No closing keyword / Part-of / Refs in any commit — RULE 2 of check-partof-closing-keyword.mjs clean; CI Part-of PR must not also close its card success.
• CI on e1edc4196 (get_check_runs, read once at 16:44Z): 24 runs — 11 success, 9 skipped, 0 failures, 4 still in_progress (Lint & Repo Gates 102149754082, Type Check · debt ledger 102149754452, Type Check · consumer gates 102149754376, Type Check · source gates 102149754362; all started 16:39:36Z). Not a green reading; the director seat re-reads before recording. The diff is markdown-only under .claude/**, outside eslint's files population, so none of the four can vary with it.
• 速读 wording 「闸门只收紧不放松」, checked line by line for what a seat may do without a review: nothing is loosened. :514 / cli:18 / core-rules:113 lower the build tier (the change itself) while every one of them keeps the at-tier review as the mandate; :639 and core-rules:124 convert a tier predicate into a PASS requirement — a narrowing for at-tier builds (the gate line itself now demands the PASS) and, for below-tier builds, the absolute bar of the literal old text becomes PASS-gated, which is exactly the replacing guarantee; the deleted :522 half's guarantees live at SKILL.md:526, :646-647 and contract-review.md:60 (降档保险丝, byte-identical). The claim holds.

Findings

1. [minor — the one owed item] The PR body still describes bd26a0839 and now contradicts the head. On e1edc4196 the body says 「One file, .claude/skills/pm-dispatch/SKILL.md, 5 insertions / 5 deletions」, its table is headed 「after (bd26a0839)」, its :675 row quotes the superseded sentence, the 「Adjacent cards」 section says pm-dispatch: a contract-review verdict without the label handoff is invisible to every seat — name the FAIL end-state in one sentence, and add H51 for "verdict posted on this head, carrier still on" #16836's references/contract-review.md is not touched (it is: 3 lines), and neither the five references/** rewrites nor core-rules.md:124 nor the :639 consequence appear anywhere in it. The 速读 tells the maintainer 「读 PR 正文里 11 行的 before/after 与「每条删行的内容现在住哪」」 — the body holds 6 rows for 1 file. The director's item 1 asked for the pm-dispatch: a contract-review verdict without the label handoff is invisible to every seat — name the FAIL end-state in one sentence, and add H51 for "verdict posted on this head, carrier still on" #16836 overlap 「in the body」; F2's expectation named the body too. A governed PR is ruled on by the maintainer reading the body; a body that affirmatively states the wrong file surface is a record defect even though every line of the head is right.
   Expectation: rewrite the PR body for e1edc4196 — 4 files / 11 lines with the six references/** rows added to the before/after table, 「Where each deleted piece lives now」 kept, the :639 consequence sentence (as in 速读 ②), pm-dispatch: a contract-review verdict without the label handoff is invisible to every seat — name the FAIL end-state in one sentence, and add H51 for "verdict posted on this head, carrier still on" #16836 disclosed as touched, its claimant re-merges, pm(dispatch-gates): clause-② prose in dispatch-gates.mjs (docblocks + every --tier printout) and core-rules.md:124 still describes a BUILD-tier mandate once the 强制条款② narrowing lands #16914 as re-scoped. Do it the way AGENTS.md:432 prescribes: send the footer you want stored (strip the appended block first), pass draft: true explicitly (platform-readings.md:41 — update_pull_request sends the draft bit whether or not you pass it), read the body back. No push, no head move, no CI round; the director seat can verify the body against this table without a second tier review.

2. [info, non-blocking] Merge commit b5be3f049 carries no trailers (author Claude <noreply@anthropic.com>). AGENTS.md:517/542 sanction git merge origin/main on a PR branch and AGENTS.md:578 exempts the merge commit itself; no closing keyword anywhere, so nothing for RULE 2. Recorded so the squash-merge author does not read the three-commit shape as two.

3. [info, non-blocking] The frozen comparison arm of :675 is a sample of one. 「改制前达档史值」 is, per the card and the 速读, 1 FAIL / 1 at-tier-built clause-② card (objectui#8583 / PR sys_position's shipped en bundle still asserts installation-wide uniqueness that #8556 corrected at source #8601). The 速读 discloses this (「样本一,不是基线」); the line itself no longer says 「基线待测」. Acceptable because the output is a decision card to the maintainer, who judges the n — but the report seat should carry the n alongside the rate so the card is not read as a threshold breach on 1 vs 1.

4. [info] 速读 says 「回滚 = revert 本 PR 两个 commit」. The PR has three commits; the merge commit has no changes of its own, so reverting the two content commits is the correct recipe. Neutral.

Maintainer-only merge: yes (governed .claude/skills/**)

Generated by Claude Code