[objectstack-fleet]

Requested by Steedos · project thread

Fixes #31551

维护者速读

• 改了什么:认领(评论+标签+指派)与验收落地(评论+转正式+开自动合并)各合成一次中继写入,不再分几笔先后发。
• 为什么改:分笔发送会留下“评论已发、标签未改”的半态,且多占写入节流的名额。
• 风险与回滚:只改两个 PM 写入脚本与两行流程文字,不碰中继本身;回滚即撤回本 PR,旧的分笔写法仍可用。
• 先看:scripts/pm/label-write.mjs 中的 foldStroke,以及 post-stamped.mjs 新导出的 judgeBody。
• 两分钟验证:运行 node scripts/pm/label-write.mjs --self-test,看到 196 条通过、14 组。
• 需要你做的:按治理层级审阅本 PR(含 .claude/** 改动,保持草稿)。

What changes

• scripts/pm/post-stamped.mjs: the body checks (render, signature first line, claim keyed lines, record lines, decision facets, vacant ledger) move into one exported judgeBody. main calls it; behaviour is unchanged.
• scripts/pm/label-write.mjs: new --comment-file FILE sends the stamped comment, the label diff and the assignee change as ONE fleet-write dispatch (the --release shape, via releaseActions). --land (pull requests only, needs --comment-file) appends pr_ready and automerge_enable to the same stroke. The comment is judged by judgeBody, so the signature first line and the one-clock {{NOW}} stamp hold. RELAY-ONLY: a direct route is refused (exit 3).
• Read-back, one printed line per part: ④ labels, ④ assignees, ④ comment (plus the stamp-drift and byte-class lines from post-stamped's verdict), and with --land ④ pull (draft judged; REST auto_merge printed, not judged, the executor's landed check judges the arming). Any mismatch exits 4 and is never repaired; an unreadable read exits 3.
• No new relay op and no change to scripts/pm/fleet-write/ops.mjs: the fold uses issue_comment, labels_*, assignees_*, pr_ready and automerge_enable as they are.
• Two pm-dispatch procedure lines (claim atomic pair, partial-landing release) and one landing line now name the one-stroke form. Line-neutral.

Why

A claim or a verdict went out as two or three separate writes, each paced and read back on its own. Between them the card sat half-written. Folding them into one dispatch keeps pacing (the executor still spaces each write), keeps every read-back, and keeps the signature check.

Verification

• node scripts/pm/label-write.mjs --self-test: 171 cases / 13 batteries before, 196 / 14 after. The new battery pins the claim fold, the verdict fold, a refusal for each part that fails read-back (labels, assignee, comment, pr_ready), unsigned / typed-stamp / bad-keyed-line refusals, --land on an issue, the direct-route refusal, no-run, failure and dry run.
• node scripts/pm/post-stamped.mjs --self-test: 753 / 26 before, 761 / 27 after (the one-judge battery).
• node scripts/pm/fleet-write/dispatch.mjs --self-test: 358 / 30, including the no-run conformance.
• node scripts/pm/dispatch-gates.mjs --commands: 40 derived, 40 run, all exit 0; --ran reconciles 0 NOT-MEASURED.

Clause-②: no
Metadata-protocol: none

Human Minimum

• 1. Schemas start as Zod; packages/spec holds no runtime logic. (not touched)
• 2. Every name I added uses the casing for its kind.
• 3. No formatter ran; eslint passes on both changed scripts (--no-inline-config, the pnpm lint flags).
• 4. One issue, one branch off an up-to-date main, one pull request.
• 5. A change that publishes carries a changeset, and both declaration lines are filled in here and in the changeset. (nothing publishes: skip-changeset)
• 6. packages/spec touched: its generated artifacts are regenerated and committed. (not touched)
• 7. No hand edits to content/docs/references/, packages/*/CHANGELOG.md or content/docs/releases/.
• 8. Title, body and comments are in English. (except the required 维护者速读 block)
• The Before Submitting list in CONTRIBUTING.md is done.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QsR2jQze2kJDxaxMsAW2kY

[objectstack-fleet]

domain:skills#2 · zhuangjianguo · session_01QsR2jT34CXs5fCn3BDirZw

Contract-review-hold: session_01QsR2jT34CXs5fCn3BDirZw · head 3e24b82 · director seat off duty since 16:01Z (clock-out 6723731915), and the maintainer's standing rule routes this project's CONTRACT_REVIEW_TIER reviews to the director seat only · until the maintainer re-sits the director or names another reviewer

Seat read at this head (2026-11-08T16:10Z): the diff matches order #31551 (one stroke for claim and verdict+landing, one judgeBody path, no new relay op, ops.mjs unchanged). Self-tests rerun by the seat: label-write 196/14, post-stamped 761/27, ratchet exit 0, both .claude files line-neutral with no line over 120 B. Merges clean with main and with PR #31571. The dev's out-of-scope finding (--release comment not signature-judged) is filed as #31585.

[objectstack-fleet]

domain:skills#2 · zhuangjianguo · session_01QsR2jT34CXs5fCn3BDirZw

Contract-review-claim: session_01QsR2jT34CXs5fCn3BDirZw · head 3e24b82 · 2026-11-08T16:14Z

The hold above is lifted: the maintainer ruled at 16:12Z that a lane with Fable runs its own CONTRACT_REVIEW_TIER review. The seat's isolated reviewer is on this head from the --brief-template brief; the record follows on this PR.

[objectstack-fleet]

domain:skills#2 · zhuangjianguo · session_01QsR2jT34CXs5fCn3BDirZw

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: 3e24b82230fa7ca2b0a7f885b589ba005ae1d0a6
Local-runs: none
Scope: .claude/skills/pm-dispatch/references/execution-duties.md, .claude/skills/pm-dispatch/references/landing-operations.md + scripts/pm/label-write.mjs, scripts/pm/post-stamped.mjs

Inputs: card #31551 (body, triage 6722994869, claim 6723407287, os-dev-report 6723796105), PR #31583 body and file list, git diff origin/main...pr31583 (merge-base 645ed3d, two commits: 158742d + merge 3e24b82), the 45 check-runs on the head, and the live tree at the head for every cited symbol. The dispatching seat's conclusions were not read as inputs. Remainder of the diff outside Scope: none (four files changed, all in Scope).

① Derived judgments

• Accept-set: nothing under packages/** changes; no schema, no .d.ts, no Zod. No accept-set widens or narrows — right.
• Public surface: none. scripts/pm/** are repo tooling, not a package export; post-stamped.mjs gains a module export judgeBody and label-write.mjs gains LAND_OPS/foldActions — internal tooling surface, not a published one — right.
• Governed path: .claude/skills/pm-dispatch/references/{execution-duties,landing-operations}.md are governed rule text ⇒ rules face ⇒ Tier S, CONTRACT_REVIEW_TIER; Governed Surface Queue Guard success on the head — right.
• Relay surface: scripts/pm/fleet-write/ops.mjs unchanged; the fold uses the existing ops comment, labels_add, labels_remove, assign, unassign, pr_ready, automerge_enable ({ op, pull: N } is the shape ops.mjs:2405/2423 declares) and stays under MAX_ACTIONS = 20 (ops.mjs:1974): a claim fold is 4 actions, a verdict fold 3 — right, and it answers the claim's "ops.mjs is out unless…" constraint with no new op.
• Claim's file surface held: the four files changed are exactly the surface the claim named — right.
• Check-runs on the head (read twice): every concluded run success or skipped, including Lint · PM dispatch-gates, Lint · ESLint, Lint · repo & tooling gates, Check Changeset, PR-body guards, Governed Surface Queue Guard; Test Core (1/8) was still in_progress at the second read — it does not touch this diff (no packages/** change), and the landing-time "全绿" check belongs to the landing seat, not this record.
• PR author objectstack-fleet[bot] — relay-written; no off-relay note owed.

② Semver level

• Nothing publishes: no packages/**, no changeset owed. The skip-changeset label on the PR and Check Changeset success agree with the diff — right.
• Clause-②: no on the PR body (line-initial) and on the claim 6723407287 — right: no accept-set widens, no public surface grows. Metadata-protocol: none — right: no metadata author surface changes.

③ Boundary flags

• Dev flag (1), landing-operations.md:51 drops its "拼写见 rest-channel.md" pointer to fit 120 bytes — NOT accepted as-is; see ④ F2 (should-fix on the prose face).
• Dev flag (2), found-not-changed: the --release body (composeReleaseBody → renderBody, label-write.mjs:716–731) is not judged for the signature line — pre-existing on main, outside the card's surface, and no write LOSES its signature check by this diff (Wanted 3 holds). Answered: accepted deviation here; the lane seat owes a follow-up card for --release to go through judgeBody too, filed before ACCEPT and named in it.
• needs_decision: none, open_questions: none in the report; none found in the diff.
• Card Wanted 1 (claim/release as one stroke, each part read back): met — foldStroke prints ④ labels, ④ assignees, ④ comment (+ post-stamped's drift/byte-class lines) and exits 4 on any ✗, never repairing.
• Card Wanted 2 (verdict + pr_ready + automerge_enable as one stroke): met via --land, PR-only, pr_ready before automerge_enable as rest-channel.md:58–59 orders.
• Card Wanted 3 (no write loses pacing, read-back or signature; one-clock stamp): met — the one dispatch goes through sendFleetWrite behind write-pace as every stroke; the executor spaces the writes inside it; judgeBody refuses an unsigned body and a typed stamp with post-stamped's own text; the stamp is substituted from the act's clock (c.clock().getTime()), exactly as post-stamped's main substitutes from Date.now().
• Card Acceptance (one day of Fleet Write read as 6722165466; relay self-tests green): the measurement is post-landing and is the lane seat's; the relay self-tests are not in this diff and Lint · PM dispatch-gates is green on the head.
• Two check-runs were in_progress at the first read; one (Lint · PM dispatch-gates) concluded success before this record was written, the other (Test Core (1/8)) is unrelated to the diff — escalated to the landing seat as the ordinary pre-landing "全绿" read, nothing more.

④ Design judgment

execution-duties.md (hits). Line 70 (partial-landing release as one stroke) and line 72 (the dispatch atomic pair as one label-write stroke) are judged against the CLI they now name: --assign LOGIN --add pm:dispatched --remove pm:queue --comment-file claim.md parses (parseOptions: --comment-file moved out of the release-only set; a lone --comment-file is refused; --land beside --release is refused) and runs steps ①② first — the kind rule (pm:dispatched onto a pull refused) and refuseTwoStates still fire before any send — then foldStroke. The claim comment ② is judged by judgeBody with mode: 'comment', so SKILL.md's template lines (Branch:, Clause-②:, Metadata-protocol: line-initial) are refused on the same claimKeyedLineRefusals path as post-stamped --comment (lines 74–75 keep their reader). Line 86 (限流不拆原子对) stays coherent: a write-pace refusal (exit 10) now holds the whole pair, and an executor failure on the comment — the first action — writes nothing behind it. Both replaced lines ≤ 120 bytes (114, 57 unchanged, 118), file 188 → 188 lines, so check-skill-line-ratchet holds (Lint gates green).

• F1 — should-fix (prose face, outside the diff but inside the governed skill): .claude/skills/pm-dispatch/references/dispatch-runbook.md:36–37 still teaches 「写序:标签翻转只在认领评论写成功之后,两笔不原子」 and the rollback-the-labels obligation for the member-card pointer comment + label flip — precisely the two-stroke sentence the card's "Reader who acts" names, and now contradicted by execution-duties.md:72. A seat that reads the runbook for a family dispatch is told two strokes; a seat that reads execution-duties.md is told one. The fold keeps the runbook's reason intact (the comment is the first action, so a failed run never leaves labels flipped without a comment — releaseActions/foldActions, label-write.mjs:729–740; executor stop-at-first-failure, execute.mjs:1771), so the fix is one prose line naming the fold, not a code change.
• F3 — nit (prose): execution-duties.md:73 「step ① 之后获得 assignee 的直接弃出本批」 keeps a between-①-and-② observation point that no longer exists once both ride one stroke; the only read is the post-stroke ④ assignees line, where a concurrent login is concurrentAdds — printed in the read-back list but by design NOT a mismatch (readBackVerdict/verdictIsClean, label-write.mjs:513–527), and unlike labels it gets no ④ note (label-write.mjs:1313–1316). The race is still settled by line 87 (③ 竞态复读, the claim timestamp as arbiter), so no correctness effect; a later prose touch can say 「回读见他人 assignee 即弃出」.

landing-operations.md (hits). Line 51 is judged against --land: PR-only (refused on an issue before the comment is read), needs --comment-file, appends pr_ready then automerge_enable (LAND_OPS, label-write.mjs:733) — the order rest-channel.md:58–59 mandates because auto-merge on a draft stores nothing; a landing stroke may also carry --remove pm:awaiting-merge, so SKILL.md:196's 「清标、ready、auto-merge」 fits one stroke. Read-back: ④ pull reads GET /pulls/{n} (classifyOp 'pull-read' added), judges draft === false, prints auto_merge and does not judge it — agreeing with platform-readings.md:13 (auto_merge unstable) and leaving the arming to the executor's PR_LANDED_STATE.automerge_enable (ops.mjs:2187). SKILL.md:188 「席位记录落 PR 先于 pr_ready」 holds inside the stroke: the comment is action 1, pr_ready action 2+. Line 118 bytes, file 102 → 102 lines.

• F2 — should-fix (prose face): line 51 now presents label-write --land as THE landing spelling and drops the rest-channel.md pointer, while the tool refuses a bare landing (--land without --comment-file, parseOptions label-write.mjs:~896: 「a bare landing is the relay actions file rest-channel.md spells」). Three lines below, line 54's 「undraft 撞 429 ⇒ … 到点一次重试」 now has no spelled route on this face: with the fold, a 429 on pr_ready fails the run AFTER the comment is stored (executor stops at the first failure; label-write exit 5 says so), and a literal retry of label-write --land --comment-file would write a second comment. The bare F = [pr_ready, automerge_enable] route survives only in rest-channel.md:58 and in the tool's refusal text. Fix on prose: have line 51 (or 54) name the retry spelling or restore the pointer within 120 bytes. No code change needed; allowing --land without a comment would be a design choice the card did not ask for.

scripts/pm/post-stamped.mjs (implementing). judgeBody is a pure lift of main's refusal chain in the same order — render + signature beside it (empty body told once), then claim keyed lines, record lines, decision facets (comment mode), vacant ledger (body mode) — and main now calls judgeBody(input, options) with the defaults Date.now() and 'post-stamped', then rebinds rendered; behaviour unchanged, the dry-run-refused-as-live ordering kept (judge before resolveRoute). The structural self-tests were re-pointed at the new shape rather than loosened, and the new battery pins main reaching every refusal through judgeBody alone and label-write importing it with no reader of its own. Agrees with the contract sentence by sentence; no finding.

scripts/pm/label-write.mjs (implementing). foldStroke: judge → route (route.error exit 3; non-dispatch transport exit 3 with the RELAY-ONLY text, zero writes) → packRequest → one c.send → read-back. Exit register matches --release's: refusal 2 (= post-stamped's EXIT_REFUSED 2), prerequisite 3, mismatch 4, run failure 5, no-run/timeout 6 via unconfirmedText (a replay would be a second comment — right, the comment is action 1). pickRelayComment + post-stamped's readBackVerdict locate and judge the stored comment the way runRelease does (same since = dispatchedAt − 60 s). --json carries mode: 'fold', refusal, comment, land. Header text agrees with the code. Self-test battery registered in SELF_TEST_BATTERIES with the floor raised 13 → 14 (a battery that stops registering names itself).

• Accepted, not a finding: a folded comment over the relay's 60,000-byte body cap is refused by packRequest (exit 2, nothing written) where post-stamped would route it direct — the RELAY-ONLY rule the header states, and a refusal rather than a silent direct write.
• Accepted (dev flag 2): --release's body is not judgeBody-judged — pre-existing; follow-up card owed by the lane seat (③).

Verdict basis: no must-fix. F1 and F2 are should-fix on prose faces of the same skill and earn a delta review of their fix by this reviewer; F3 is a nit.

Implemented-by: claude/project-thread-sxds3e
Reviewed-by: session_01QsR2jT34CXs5fCn3BDirZw

VERDICT: PASS

[objectstack-fleet]

domain:skills#2 · zhuangjianguo · session_01QsR2jT34CXs5fCn3BDirZw

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: 137e74ac405997b9eb57f249c3e44a112e2f6daf
Local-runs: none
Scope: .claude/skills/pm-dispatch/references/execution-duties.md, .claude/skills/pm-dispatch/references/landing-operations.md, .claude/skills/pm-dispatch/references/dispatch-runbook.md + scripts/pm/label-write.mjs, scripts/pm/post-stamped.mjs

Delta record, same reviewer as record 6723957070 (PASS on 3e24b822). The head moved by ONE prose commit, 137e74ac 「docs(pm): name the one-stroke claim and the bare-op landing retry」, +3/−3 over two .claude/** files; git diff 3e24b822 pr31583 touches nothing else, so the ①–③ judgments and the two implementing files' ④ carry over unchanged from 6723957070 and are restated only where the delta bears on them. Inputs: card #31551 and its thread, PR #31583 body and file list (now five files, all in Scope; remainder outside Scope: none), the net diff against origin/main, the delta diff, the check-runs on 137e74ac, and the live tree at the head for every cited symbol.

① Derived judgments

• Accept-set: unchanged, none — no packages/** in the diff. Public surface: none. Right.
• Governed path: three .claude/skills/pm-dispatch/references/*.md files now ⇒ rules face, Tier S, CONTRACT_REVIEW_TIER. Right.
• Relay surface: fleet-write/ops.mjs still unchanged; the delta's prose names only existing spellings — label-write --comment-file, label-write without it, and rest-channel.md:52/58's bare actions file F = [pr_ready, automerge_enable]. Right.
• Skill line budget: the replaced lines are 119, 120 and 118 bytes (≤ 120; the ratchet header itself counts 34 live lines at exactly 120), and every file keeps its line count (dispatch-runbook.md 245 → 245, landing-operations.md 102 → 102, execution-duties.md 188 → 188).
• Check-runs on 137e74ac: 38 runs at the second read for this record; every concluded run success (11) or skipped (16), 11 still in_progress (the Lint and Type Check families, incl. Lint · PM dispatch-gates, and Test Core shards 1–2) — a push of three prose lines re-runs them; their conclusions are the gate verdicts and the landing seat's pre-landing "全绿" read, not this record's. Nothing in the delta can move a derived gate other than check:pm-skill-ratchet, which the byte and line readings above satisfy.
• PR author objectstack-fleet[bot]; no off-relay note owed.

② Semver level

• Unchanged: nothing publishes; skip-changeset label stands; Clause-②: no (PR body, claim 6723407287) and Metadata-protocol: none remain right — a prose delta to governed rule text changes no accept-set and no metadata author surface.

③ Boundary flags

• Dev deviation on this head: landing-operations.md:54 drops 「不自定节奏」 to fit 120 bytes. Judged in ④ — no rule is lost.
• Carried from 6723957070: dev flag (2) (--release body not judgeBody-judged, label-write.mjs:716–731) stays an accepted pre-existing deviation; the lane seat still owes the follow-up card before ACCEPT and names it there.
• F3 (execution-duties.md:73, nit) left as is by the dev — accepted deviation, as a nit may be; the ③ race re-read (line 87) remains the arbiter.
• needs_decision/open_questions: none.
• Card Wanted 1–3 and the claim's surface constraint: met as before; the delta adds no write path and removes none.

④ Design judgment

F1 — CLOSED. dispatch-runbook.md:36–37 now read 「写序:认领一笔 label-write --comment-file,评论排首、标签随后;执行器首败即停,标签不先翻。」 and 「评论存而标签未翻 ⇒ 非零退出;补标签用不带 --comment-file 的 label-write,⛔ 不重发评论。」 Judged against the code: foldActions → releaseActions puts the comment first and the additive verbs after (label-write.mjs:729–740); the executor issues actions in order and stops at the first failure (execute.mjs header line 63, battery 「actions in order, stop at the first failure」), so labels can never flip before the comment is stored — the runbook's old reason (真半状态 = labels flipped, no comment) is now impossible rather than mitigated by rollback, and the rollback sentence is rightly gone. 「评论存而标签未翻 ⇒ 非零退出」 holds on both routes that can produce it: a run that fails after the comment is exit 5 with the text 「a failed run may have left the comment with nothing behind it」 (label-write.mjs:1287), and a success run whose read-back shows ④ labels ✗ is exit 4, 「Not repaired here — under the relay a re-send is a second write」 (:1342). The repair the line prescribes — label-write with the label plan and no --comment-file — is the ordinary relay stroke of the same tool, carries no comment, and is read back the same way; it is also exactly what the tool's own refusal texts ask the seat to do deliberately. The line sits in the member-card section where the pointer comment is a signed, stamped comment like any other, so judgeBody governs it too. Agrees with execution-duties.md:72; the contradiction 6723957070 named is gone and no other 「两笔」 sentence remains in the skill (grep on the head). No finding.

F2 — CLOSED. landing-operations.md:51 is unchanged (the landing is label-write --land); line 54 now reads 「undraft 429 ⇒ 卡记 resets at,到点一次重试裸 op(rest-channel.md),⛔ 不重发评论、不报阻塞。」 Judged against the code: a 429 on pr_ready fails the run after the comment (action 1) is stored, label-write exits 5 and never retries; the retry the line now prescribes is the bare op through rest-channel.md:52's dispatch.mjs --actions-file F with F from :58 (pr_ready then automerge_enable, :59's order), which carries no comment — so the duplicate-comment path 6723957070 described is closed by the text, and it matches the refusal label-write prints for --land without --comment-file (「a bare landing is the relay actions file rest-channel.md spells」). The pointer the first head dropped from line 51 is back where it is needed, on the retry line. No finding.

The 「不自定节奏」 deviation — no rule lost. The old line prescribed the pace positively (「该刻记卡,到点一次重试」: record the moment, retry ONCE, AT the reset time) and then forbade its negation (「不自定节奏」). The new line keeps the positive form whole — 「卡记 resets at,到点一次重试」 — which fixes both the moment (到点) and the count (一次), so no retry rhythm other than the prescribed one is left open; and the general discipline the dropped clause restated still stands two lines above (:53 「定点文本照定时器写法纪律」) and in the 「不报阻塞」 that survived. Dropping a restatement to keep the retry spelling under 120 bytes is the right trade; accepted deviation, not a finding.

F3 — left as is, nit stands. execution-duties.md:73 unchanged; no correctness effect (the post-stroke ④ assignees read plus line 87's timestamp arbiter settle the race). Accepted deviation on a nit; a later prose touch may reword it.

execution-duties.md:70, :72, landing-operations.md:51, scripts/pm/label-write.mjs, scripts/pm/post-stamped.mjs — byte-identical to 3e24b822; the ④ of 6723957070 stands for them: the one-stroke claim and release, --land PR-only with pr_ready before automerge_enable, judgeBody as the one judging path, per-part read-back with exit 4 never repaired, RELAY-ONLY exit 3, pacing through sendFleetWrite. No new finding.

Verdict basis: both should-fix findings of 6723957070 are closed on this head by prose that agrees with the code it cites; the one nit is an accepted deviation; no must-fix, no new should-fix. This ends the round.

Implemented-by: claude/project-thread-sxds3e
Reviewed-by: session_01QsR2jT34CXs5fCn3BDirZw

VERDICT: PASS

[objectstack-fleet]

domain:skills#2 · zhuangjianguo · session_01QsR2jT34CXs5fCn3BDirZw

维护者速读

• 改了什么：席位的认领、释放和落地，原来要分两次经中继写入（评论一次，标签和指派一次；或评论一次，ready 和 auto-merge 一次），现在一次写完。label-write 新增 --comment-file，可以把评论和标签、指派放进同一次写入；新增 --land，可以把验收评论和 pr_ready、automerge_enable 放进同一次写入。评论的检查直接调用 post-stamped 新导出的 judgeBody（时间戳、签名首行、认领键行），和 post-stamped 本身走同一条路径。每一项写完都会逐行回读核对。另外改了三处规则文字，说明这种一次写入的用法。
• 为什么改：ci(glue): Fleet Write starts about 4,600 runs a day and every CI completion starts a Merge Queue Triage and a Scheduled Full Run Card run — measure which senders can fold their writes and narrow the workflow_run triggers #31521 的读数显示，成对的两次写入让每次认领多一次中继运行、多等一轮；你说「同意，立卡」「高优先」。
• 风险与回滚：没有新的中继操作，ops.mjs 没改，旧的两次写法仍然可用。回读不一致时以 exit 4 停下，不自动重写。规则文件行数不变。回滚即 revert 本 PR。
• 先看：scripts/pm/label-write.mjs 的 foldStroke；execution-duties.md L70/L72，landing-operations.md L51/L54，dispatch-runbook.md L36–37。
• 两分钟验证：node scripts/pm/label-write.mjs --self-test（196 例，14 组）与 node scripts/pm/post-stamped.mjs --self-test（761 例，27 组）都通过。
• 需要你做的：无。这是 Tier S，契约复核通过后由席位自己落地。
• 席位意见：ACCEPT。车道自有的隔离复核器按你 16:12Z 的话，在 CONTRACT_REVIEW_TIER 档复核：首轮 PASS（记录 6723957070），带两条 should-fix，开发已修；增量复核 PASS（记录 6724244229，Head 137e74ac），没有新的发现。--release 的评论还没走签名检查，这是原有缺口，已另立 label-write --release sends its comment without the signature check #31585。

席位核对

Head-sha: 137e74ac405997b9eb57f249c3e44a112e2f6daf

1. Shape: draft, base main, body carries Fixes #31551, skip-changeset. Five files: label-write.mjs, post-stamped.mjs, and three .claude files, each line-neutral (execution-duties 188/188, landing-operations 102/102, dispatch-runbook 245/245), with no line over 120 B. Tier S (.claude/**).
2. Premise measured: on main, a claim needed post-stamped for the comment and then label-write for the labels and assignee, which is two dispatches. Landing needed a separate hand-written actions file. The seat re-ran both self-tests at 3e24b82: 196/14 and 761/27. The rework touched prose only.
3. Gates: contract review at CONTRACT_REVIEW_TIER, round 1 PASS 6723957070 on 3e24b82 (F1 and F2 should-fix, F3 nit). The delta is PASS 6724244229 on this head: F1 and F2 are closed. Dev gates: 40 derived, 40 exit 0, --ran 40/40 (report 6723796105, re-run at the rework). Check-runs on this head: no failure so far; auto-merge waits for the rest. K (first-parent commits from PRBASE to main) = 9, under 15.
4. Accepted deviations: dispatch-runbook.md joins the surface by amendment 6723965134 (review F1). L54 drops 「不自定节奏」 to fit 120 B; the reviewer judged that no rule is lost. F3 (execution-duties L73) stays as a nit.
5. Out of scope / Skills-docs: label-write --release sends its comment without the signature check #31585 covers --release going through judgeBody. Metadata-protocol: none, so no Skills-docs: line is owed.
6. Landing: Tier S. pr_ready + automerge_enable on this head; the queue guard reads record 6724244229.