[objectstack-fleet]

domain:skills#1 · huangyiirene · session_01J8VVAnMUWA9rufXzFaEh4J

Requested by Irene · project thread

Fixes #29467

Clause-②: no
Metadata-protocol: none

维护者速读

• 改了什么:总监每日摘要加一行「外部需求行」三个数;取卡全序同级内非车队卡排在车队卡之前(车队 = App 作者或正文首行是席位签名);H67 读序同步。
• 为什么改:维护者对「车队在给自己造活」的评估回了「同意」;目前没有任何读数说明一周的工作有多少回应车队之外的需求。
• 风险与回滚:只改规则文本与一个巡检的读序,不升级、不越 p0;回滚即 revert 本 PR。
• 先看:lanes/director.md 职责四新增三行;execution-duties.md 同级平手规则;check-half-states.mjs 的 h67ReadOrder。
• 两分钟验证:node scripts/pm/check-half-states.mjs --self-test,预期末行 H67 组 158/158 全过。
• 需要你做的:无,席位落地(Tier S)。

What changes

• Daily summary (lanes/director.md, duty four): the summary lists an outside-demand line, defined in three new lines. The three numbers are read when the summary is written: (1) cards opened in the window by non-fleet authors; (2) the share of the window's closed cards that non-fleet authors opened; (3) the open non-fleet cards with no pm:* state, each named with its age. The query is spelled out and posted with the line. For 1 and 2 it takes the all-state since listing from the window start, which is a superset, and counts a card only when its created_at (for 1) or closed_at (for 2) falls in the window. For 3 it takes the open listing. Pull requests are excluded from all three.
• Fleet card, per the seat's ruling as amended in its contract review: a card whose author is an App (user.type Bot, every [bot] login, github-actions[bot] included), OR whose body's first non-empty line is a seat signature. Seat-operated human accounts file cards too, and the card's own measurement excluded seat posts. Every other card is non-fleet.
• Candidate order (execution-duties.md): within one level of the order, a non-fleet card sorts first, then Bug, then card age. The rule never raises a card's level and never jumps priority:p0.
• h67ReadOrder (scripts/pm/check-half-states.mjs) is the only scripts/pm reader that sorts candidates by priority. It now orders by rank, then fleet status, then age. It reuses two readers from the same file: artefactAuthor, and readSignatureLine, the signature reader record-recognisers.mjs imports from this file. No new regex is added. An unreadable author counts as the fleet's, so it is never promoted. No new patrol row, battery or gate.

Ratchet payment (both files stay at their ceilings, no ceiling row moved)

file	before / after / ceiling	added	paid by
lanes/director.md	86 / 86 / 86	3 summary lines plus 1 moved line (the fleet definition and the window test were later edited in place)	removed the one-bullet section ## 与 skills 席的分界: its heading, its blank line and the blank line above it. Its bullet moved unchanged into 授权与形态 next to the scope line.
execution-duties.md	188 / 188 / 188	1 tie-break line (the seat-signature clause was later edited in place)	deleted the separate ruling-implementation re-check of the action surface. Its content moved into the work-item surface line, which now names the ruling's items next to the card's and the triage comment's.

Every line in both files is at most 120 bytes; the longest new line is 119.

Tests

• node scripts/pm/check-half-states.mjs --self-test on cbc61eb: 5219 cases pass, and the battery H67 queued merged-delivery reading reads 158/158. The pin moved from 142 to the count, so none of the six new cases can be removed without turning the battery red.
• New cases:
  • A human-opened p2 card is reached before an older fleet-opened p2 card.
  • Two fleet cards of one priority stay oldest-first.
  • The fleet leg never lifts a card out of its priority or past p0.
  • github-actions[bot] counts as fleet, and an unreadable author is not promoted.
  • A human-account card whose body opens with a seat signature stays behind an older fleet card.
  • The same human card with the signature moved off its first line is reached first again.
• Reverse verification, two ablations, each on a committed head and each restored from HEAD and green afterwards:
  • On 1d7e6fb, with fleet(a) - fleet(b) replaced by 0, the run printed 1 of 5189 case(s) failed: ⭐ H67 order: within ONE priority a human-opened card is reached before an OLDER fleet-opened one (got "21,22", want "22,21").
  • On cbc61eb, with the signature test replaced by false, the run printed 1 of 5219 case(s) failed: ⭐ H67 order: a HUMAN account whose body opens with a seat signature is the fleet's — it stays behind the older fleet card (got "22,21", want "21,22").

Acceptance notes

• The director's payment removes a section heading and two blank lines, not prose; the bullet moved byte-identically. The seat's contract review ruled this a real deletion.
• The seat's should-fixes S1 (seat-signed cards are the fleet's) and S2 (the summary query names its local window test) are in cbc61eb, paid in place in both texts. The delta review's should-fix is in f64ec72: the query line names the all-state listing again, because number 2 needs closed cards and the default listing is open-only. It was edited in place, and director.md stays at 86 lines.
• PR feat(pm): a ruled decision card still in the box is a patrol row, and the director's ruling lands with its knock #29515 landed first and touched lanes/director.md and check-half-states.mjs. main was merged in with no conflict, and both files were re-measured on the merged tree.

Verification

Local gates on cbc61eb: 54 derived, 54 run, all exit 0 (`dispatch-gates --ran`)

• On head cbc61eb, node scripts/pm/dispatch-gates.mjs --commands derived 54 commands. All 54 ran, each with a 900 s timeout, with the exit codes recorded before any pipe. All 54 exited 0.
• The --ran reconciliation printed: 54 derived famil(ies) accounted for — 54 run, 0 NOT-MEASURED.
• Both ratcheted files read headroom 0 at their ceilings (86 and 188), and every line is at most 120 bytes.
• On f64ec72, a one-line text edit to director.md, these gates were re-run: check:pm-skill-ratchet, check:pm-skill-id-lint, check:doc-authoring, check:nul-bytes. All exited 0, and director.md read 86 lines (ceiling 86, headroom 0).
• Earlier, on 7d105ed: pnpm check:pm-skill-ratchet had exited 1 on a 125-byte line, which was trimmed to 119 bytes. The full run on cbc61eb above supersedes those readings.
• Prerequisites: pnpm install --frozen-lockfile, and a turbo build of @objectstack/formula and @objectstack/lint for check:doc-formula-expressions.
• Families CI runs beyond the derived list are read from this PR's check runs, not re-run locally.

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; pnpm lint passes. (no formatter ran; pnpm lint was not run locally on this diff, so the CI lint job decides it)
• 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. (publishes nothing: .claude/** and scripts/pm/**; 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 维护者速读 block, which the seat contract writes in Chinese)
• The Before Submitting list in CONTRIBUTING.md is done.

🤖 Generated with Claude Code

https://claude.ai/code/session_01J8VVAjYZRbZ8TG8TRzG5zd

---

Generated by Claude Code

[objectstack-fleet]

domain:skills#1 · huangyiirene · session_01J8VVAnMUWA9rufXzFaEh4J

Contract-review-claim: session_01J8VVAnMUWA9rufXzFaEh4J · head 7d105ed · 2026-11-02T19:00Z

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: 7d105ed7642a058e5e6170170ba854cba03c8d12
Local-runs: probe — on a detached worktree of the head (merge-base with origin/main 1b2043e; none of the three files moved on main since; origin/main at 2789ba7 when fetched; no node_modules in the worktree): node scripts/pm/check-half-states.mjs --self-test (5217 cases, exit 0, battery H67 queued merged-delivery reading 156/156), node scripts/pm/check-skill-line-ratchet.mjs (default, exit 0, director.md 86/86 and execution-duties.md 188/188, and --self-test, 164 cases), node scripts/pm/check-skill-id-lint.mjs (108 files clean), node scripts/pm/dispatch-gates.mjs --commands (54 commands derived for the three paths; all 54 run in their node form, 42 exit 0, 10 PREREQUISITE NOT MET on a missing yaml/typescript/@typescript-eslint/parser, 2 exit 1 on a missing @changesets/cli and the pnpm --filter build, all twelve about this container and covered by the head's green lint jobs), check-governed-merges.mjs --pr 29572 (Tier S, exit 3 = governed), five reviewer-authored ablations applied to the committed file, measured, and restored with git checkout HEAD -- to the head blob (git hash-object 07966427ee equal to the PR's blob after every one, git status --porcelain empty), and the three daily-summary queries run read-only against the live board by page number (the proxy here refuses the repositories/{id} URLs --paginate follows, so pages were walked by page=N); the worktree was removed after the runs and the main checkout was not touched
Scope: .claude/skills/pm-dispatch/references/lanes/director.md, .claude/skills/pm-dispatch/references/execution-duties.md, scripts/pm/check-half-states.mjs (the PR's three files; h67ReadOrder, its battery pin and four cases live in the third) + read, not reviewed: scripts/pm/check-skill-line-ratchet.mjs (header, the two ceiling rows and the ledger precedents at its lines 1172 and 1213), .claude/skills/pm-dispatch/SKILL.md line 291 and .claude/skills/pm-dispatch/references/lanes/skills.md line 20 (ratchet currency), .claude/agents/os-dev.md lines 296 to 297 (the 折行合并 definition of re-wrap), artefactAuthor in check-half-states.mjs (line 16165), H67_SELECTION_ORDER_LABELS (line 17593), PR #29515's director.md diff (commit 728c73e), card #29467 body with comments 6581730754 (triage), 6585217502 (claim) and 6585720791 (dev report, read as input), PR #29572 body and its one comment 6585772094 (the seat's review claim on this head)

Reviewer: an isolated contract reviewer at CONTRACT_REVIEW_TIER, read-only on GitHub (REST reads only; nothing written, pushed or commented), fed the card body and comments, the PR body, the net diff against the merge-base, the head's check-runs, contract-review.md and review-checklist.md. The dev report was read as input, not as conclusions; every count below is this review's own measurement. No environment variable value is printed anywhere in this record; the one token-bearing run passed it by substitution and never echoed it.

① Derived judgments

The card's three Change items, its two done-when lines, the claim's surface and the PR's claims, one at a time:

1. File surface. Four commits (1d7e6fb feature, 1bf1b21 merge of origin/main, 526a039 and 7d105ed payment and width fixes; commit author Claude, PR author objectstack-fleet[bot], head repo = base repo, draft, base main); git diff 1b2043e7 7d105ed7 --stat, the REST file listing (+42/−18, 3 files) and the PR body agree on exactly lanes/director.md (+5/−5, blob f50d7a150c), execution-duties.md (+3/−3, blob 807da2dbd4) and scripts/pm/check-half-states.mjs (+34/−10, blob 07966427ee). The claim (6585217502) named these three files with these contents (daily-summary line paid in place at 86, tie-break paid in place at 188, h67ReadOrder plus a same-priority pair and the H67 floor) and said stop on breach; nothing else is touched. No .github/**, no check-governed-merges.mjs, no new H row (grep of the diff for a new sweepInto/findings.push/HALF_STATE_FAMILY_BAND entry finds none), no filing-gate text. Correct, and on surface.
2. director.md, the daily-summary line (lines 74 to 77). Line 74 adds 「外部需求行」 to the summary's roster; 75 defines the row (「一行三数,摘要时现查并贴查询;非车队 = 作者 user.type 非 Bot 的人工账号」); 76 names the three numbers ① 窗内非车队开卡数 ② 窗内关卡中非车队所开占比 ③ open 无 pm:* 非车队卡逐张带龄; 77 the query (①② a since all-state listing, ③ the full open listing, PRs excluded, computed locally on user.type and labels). Against the card's Change 1: the three numbers are the card's three, term for term; 「摘要时现查」 is "read by query at summary time"; 「贴查询」 is the done-when's "and their queries". Correct. The query is runnable: run here against the live board for the window 2026-10-26T00:00Z → now, ① 19 cards opened by User authors (1,707 by Bot), ② 56 of 1,689 cards closed in the window were opened by User authors (3.3%), ③ 16 open User-authored cards with no pm:* label (39 open User-authored in all; 254 open issues). One precision gap: since filters on updated_at, so the ①② listing is a superset and the window test on created_at (①) and closed_at (②) is made locally; line 77 says 「本地按 user.type 与标签求」 and does not name the date test, though 「窗内」 on line 76 defines it. S2. One wording change rode in: line 74's 「等复审逾 24h 的 PR」 became 「复审逾 24h 的 PR」 to fit 120 bytes (103 → 118 B); same referent. Cosmetic.
3. director.md, the payment. Section 「## 与 skills 席的分界」 (heading, the blank line above it and the blank line after it, three lines) is removed and its one bullet moves byte-identically (verified by diff of merge-base line 86 against head line 22) into 授权与形态, directly under 「不占 domain:* 车道 … 永不写码」, which is where a seat-boundary rule belongs. 86 → 86 lines, ceiling row 86 unmoved, ratchet green. Ruling on the dev's question in ③. No obligation lost: the bullet is the heading's entire content and survives word for word with a home; PR feat(pm): a ruled decision card still in the box is a patrol row, and the director's ruling lands with its knock #29515's two edits (line 13 the Notify: ruling knock, the merged 职责一 line now at line 40) are both present unchanged in the head. Correct.
4. execution-duties.md, the tie-break (line 55): 「同级先非车队作者(user.type 非 Bot)卡,再 Bug,再卡龄;⛔ 不升级、不越 priority:p0。」 Against Change 2: within one level a non-fleet card sorts first; the rule neither raises a level nor jumps p0, in as many words; the pre-existing 「先 Bug 再卡龄」 is kept behind the new leg. Paid by deleting line 15 「裁决实施卡的动作面再核被点名的动作在 origin/main 上还没被做掉。」 and folding its referent into the work-item line 17 「工作项面:卡片、裁决与分诊评论列的工作项逐条对树核验是否仍未完成,⛔ 不对卡核验。」 The obligation (a ruling's named actions are re-checked against the tree before dispatch, so a done action is not re-dispatched) survives as the ruling's items on the work-item surface, under line 13's 「裁决同罪,三面都查」; what is dropped is the word origin/main, which the three-surface check already reads against (line 14's git log runs on the seat's origin/main checkout). Not a re-wrap: one line deleted, one word added to a sibling, no reflow. 188 → 188, row unmoved, line 17 at 119 B. Correct.
5. The [bot] definition, consistently applied. Claim: fleet = every login ending in [bot]; non-fleet = a human account. Text: user.type 非 Bot (director.md 75, execution-duties.md 55). Code: artefactAuthor(issue) returns isApp: type === 'Bot', and the H67 header comment spells 「an App (user.type Bot, every [bot] login)」. The two spellings are one predicate on GitHub (an App installation's user is type: Bot with a [bot] login), and the live read confirms it: over 1,919 issue rows of the window listing, zero logins ending in [bot] with another type and zero Bot rows without the suffix; the Bot logins seen are objectstack-fleet[bot] 1,790, github-actions[bot] 39, claude[bot] 7. Consistent in claim, both texts and code. What the definition does not separate is a human account a seat operates (see ③, S1).
6. h67ReadOrder, the leg. Sort is rank → fleet → age → number; rank is unchanged (H67_SELECTION_ORDER_LABELS: p0, pm:blocking, p1, p2, p3, then unlabelled), so the author leg can only reorder within one rank and can never place a card above a p0 of another rank; fleet is 0 for a readable non-App author and 1 otherwise, so an unreadable user (null from artefactAuthor) reads as the fleet's and is never promoted, as the comment says. h67ReadOrder is the only reader in scripts/pm that sorts candidates by priority (grep for SELECTION_ORDER/取卡全序 finds only this file), so the card's "the selection code (if any reads author) follows it" is met by the one site. The row's population, cap, request count and output text are untouched; H67 remains report-only. Correct.
7. Self-test, measured here. Head: 5217 cases pass, exit 0, H67 queued merged-delivery reading 156/156 (the dev's 5189 was measured on 1d7e6fb before main's H73 battery, 28 cases, was merged in; 5189 + 28 = 5217). Five ablations on the committed file, each restored to the head blob: A1 the fleet leg removed → exit 1, 2 red (the ⭐ human-before-older-fleet pair, the dev's own control, among them); A2 an unreadable author read as human (author?.isApp ? 1 : 0) → 2 red (the github-actions[bot]/undefined-author case); A3 the author leg placed before rank → 2 red (the never-past-p0 case); A4 pin back to 142 and the four new cases deleted → exit 0 (the pre-PR state, as expected); A5 the four new cases deleted with the pin left at 156 → exit 1, 1 red, the roster floor. So the pair the claim asked for is pinned, each of the three properties the card names (human first, within one priority, never past p0) has a control that reds when its mechanism is reverted, and the floor at the count holds the cases in.
8. Gates. dispatch-gates --commands derives 54 commands for the three paths (measured 54, the PR's number). All 54 were run here in their node form with exit codes captured before any pipe: 42 exit 0, among them check-half-states (check:pm-half-states), skill-ratchet (both forms), skill-id-lint, dispatch-gates (both forms), declaration-mirrors (both), issue-citations (both; 0 citations added), self-test-wired (both), whole-set-label-write (both), platform-checklist, commit-card-trailers, closing-target-claim, partof-closing-keyword, post-stamped, nul-bytes, entry-guard, parse-guard, skill-frame-sync, agent-test-spelling, scripts-symbol-anchors, ratchet-remedy-authority, single-claim-paths, bash32-floor, watch-hint-literal, gitlink-declared, pnpm-flag-spellings, pnpm-filter-targets, refd-timer-probe, cross-package-test-inputs, release-section-coverage (both), sweep-closed-cards, board-snapshot, bare-root-worklist, check-harness-current, plan-test-core-shards, scheduled-full-run, render-release-coverage-anchor. Ten answered PREREQUISITE NOT MET (exit 3: closing-keyword-parity ×2, comment-mask-corpus, platform-checklist-watchdog ×2, ci-filter-parity, cli-command-ids, doc-authoring, driver-memory-census, governed-merges self-test — each naming the uninstalled yaml, typescript or @typescript-eslint/parser), and two exited 1 for the same reason (check-self-test-workflow-commands on release-pending-publish.mjs missing @changesets/cli; check:doc-formula-expressions with no node_modules), findings about this container, not the diff. CI covers those: Lint & Repo Gates, Lint · repo & tooling gates, Lint · docs & code guards, Lint · CI, release & engine gates, Lint · PM dispatch-gates, Lint · ESLint and Lint · scoped · tooling self-tests are all success on the head.
9. Ratchet, read directly. check-skill-line-ratchet.mjs default run exit 0: both rows at their ceilings (director.md 86/86, execution-duties.md 188/188), headroom 0, no CEILINGS or CROSS_FILE_MOVES row changed in the diff; widest line in each file 120 B, the three new director lines 117/119/117 B, the new execution-duties line 120 B and the edited line 17 119 B.
10. check-governed-merges.mjs --pr 29572: 2 of 3 paths hit the register, both .claude/**, Tier S; scripts/pm/check-half-states.mjs off the register; size 60 changed lines, under the human-merge threshold. Landing is the owning seat's through the queue once this record is on the thread.
11. PR body against the head. 86/86/86 and 188/188/188 (measured); 54 derived (measured); 156/156 (measured); longest new line 119 B is right for director.md and the edited execution-duties line, while the new tie-break line is 120 B (still inside the budget; cosmetic); Clause-②: no, Metadata-protocol: none, skip-changeset right (.claude/** and scripts/pm/** publish nothing; Check Changeset success); Fixes #29467 right because both done-when lines are met once the next summary is written (the roster line and the three-number definition are in the charter; the tie-break text and its code site match, with the pair pinned). One stale figure: "5189 cases pass" is the pre-merge count; the head reads 5217. S3.
12. Nothing new on any surface: no gate, no label, no workflow, no CODEOWNERS, no register row, no package, no new patrol row, no change to the filing gate (card Change 3).

② Semver level

None. No published package, no packages/spec/src/**, no TypeScript; governed rules text and scripts/pm tooling only. skip-changeset is the right declaration and Check Changeset is success on the head.

③ Boundary flags

• Dev open_questions 1, the director.md payment — ruled here: a real deletion, counted. The ratchet's rules define the forbidden currency narrowly: os-dev.md 296 「⛔ 不拿 re-wrap(折行合并)当筹行」 and 297 「分界只问折行有没有为新增内容买行」, SKILL.md 291 「⛔ re-wrap 不得用作筹行,新增以删减付账;不买内容的密度修复允许」, skills.md 20 「⛔ re-wrap 筹行;砍点逐条核幸存,被砍内容若在别处无家,砍掉就是丢规则」, and the ratchet header "deletion at the source pays, restatement does not". A re-wrap is 折行合并: prose reflowed so the same words occupy fewer lines. Here no prose line was merged or reflowed; three non-rule lines (a heading and two blanks) left the file and the one rule under the heading kept every byte and a home. The ratchet's own ledger prices headings and blank lines as lines in both directions (rest-channel row: 「that restatement and its adjacent blank are retired (−2)」; landing-operations row: 「29 rule lines, its heading and blank lines — moved」), so counting them as payment is the ledger's existing convention, not a new reading. The skills.md survival test passes (nothing cut lacks a home). Two cautions the seat should hear: the payment is thin against what the ratchet prices (three short structural lines bought ~350 bytes of new rule text), and it is exhaustible — one more one-bullet section remains (基本原则, line 28 to 30). Had a prose payment been required, the line to delete without losing an obligation is line 38 「出处:维护者 2026-10-18 令,原话载卡 24241 正文。」: a provenance line that the rules-only rewrite ruling (recorded in the ratchet ledger: provenance dates and ruling citations leave the corpus) already sent out of the skill files, and whose rule (line 35 to 37) stands without it.
• S1, definition gap, not a blocker: the claim ruled fleet = App accounts, and the PR applies that consistently. The card's own measurement, though, counted 「open issues by non-fleet authors (excluding seat posts)」, and several seats file under human accounts (os-zhuang, os-support-ai, os-warren, os-justin, os-steve sign the card and its triage). Under the PR's definition they are non-fleet: of the 16 cards in ③ today, 8 are theirs (Permission Model v2 — named follow-up ADRs (post-ADR-0090 parking lot) #2776, 架构:把审批 UI 重建在标准 SDUI 渲染器上(ListView + RecordDetailView + ApiDataSource + 多态记录预览组件) #25950, Response bodies are never checked against the schemas that declare them — staged plan, not a repo-wide sweep #3877, service-storage: IStorageService.list(prefix) means two different things on the two shipped adapters (local: one level, directories as files; S3: recursive, silently capped at 1000) #5266, [裁决] driver-memory / driver-mongodb 投入冻结 —— 维护者 2026-08-05 口径（跨单锚点） #5499 os-zhuang; [tracking] Envelope-position convergence line for packages/rest's flat dialect — the live ratchet owner #9559 os-support-ai; [Half-state patrol] check-half-states live sweep — generated view (please pin) #9857 os-warren; [program] CLAUDE.md is a pointer to AGENTS.md in every repo — no second instruction set to maintain (maintainer 2026-09-09) #17161 os-justin), and 4 of the 19 ① openings (os-warren 2, os-zhuang 1, os-steve 1). The summary posts its query beside the numbers, so the reader can see what was counted, but "outside demand" will read high by roughly half on ③ until the seat or the maintainer says whether a seat-operated human account is the fleet's. Follow-up card for the skills seat; a ruling, not a dev fix.
• S2, query precision (director.md 77): the since listing is an updated_at superset; the local test on created_at for ① and closed_at for ② is implied by 「窗内」 but not named. One-line edit paid in place, same follow-up card.
• S3, PR body: "5189 cases pass" → 5217 on the head; "longest new line is 119" while the tie-break line is 120 B. Cosmetic.
• Dev deviations: the PR carries no assignee or reviewer; os-dev.md's rule, no flag.
• Independence: PR author objectstack-fleet[bot], commit author Claude; no off-relay author note is owed. The PR thread carries the seat's Contract-review-claim: for this head (6585772094) and no other record; no reviews on the PR; auto-merge not armed; labels documentation, size/s, skip-changeset.

④ Design judgment

Per hits file, against the card's scenario (the maintainer wants one reading of how much of the week answers demand from outside the fleet, and wants a human-opened card reached before a fleet-opened one of the same priority, with no new mechanism):

• lanes/director.md: a three-number line read by query at summary time is the right instrument for a reading nobody has, because it costs two listings the summary's author already makes and invents no label, state or row; posting the query with the numbers makes the definition auditable on the day it is read, which is what makes S1 a flag and not a defect. Defining non-fleet by user.type, the one field the listing serves, is the cheapest correct predicate; its blind spot (seat-operated human accounts) is a policy question for the seat, not a text error. The payment keeps the file at its ceiling without a ruled raise and loses no rule. Holds.
• execution-duties.md: putting the author leg inside the level, ahead of Bug and age, is the one placement that honours the card's two prohibitions by construction: the level is decided first, so no promotion and no jump past p0 is possible however the leg is read. Folding the ruling's items into the work-item surface tightens line 13's 「三面都查」 without a second home for the same check. Holds.
• scripts/pm/check-half-states.mjs: taking the leg in h67ReadOrder is the right scope, since it is the only code that models 取卡全序 and the card says the code follows the text wherever the code reads author; reusing artefactAuthor keeps one author predicate in the file; reading an unreadable author as the fleet's is the conservative direction (a card is deferred, never promoted, on missing data); pinning the floor at the count makes the four cases load-bearing, and the controls above show each property reds on its own reversal. Holds.

Must-fix: none.

Should-fix:

• S1 (follow-up card, skills seat; a ruling for the seat or the maintainer, no new head on this PR): whether a seat-operated human account (os-* and the seat accounts that sign cards) counts as the fleet for the outside-demand line; today it halves ③ (16 → 8) and trims ① (19 → 15).
• S2 (same card, one line paid in place): name the local created_at/closed_at window test on director.md line 77.
• S3 (PR body, cosmetic): 5189 → 5217; "longest new line is 119" → 120.

Check-runs at record time (2026-11-02T19:18Z): 51 check-runs on 7d105ed, all completed, 31 success, 20 skipped, 0 failure, 0 in progress (a first read a few minutes earlier had 3 in progress: Live half-state sweep, Test Core (1/8), Type Check · consumer gates; all three are success now). Success: Lint & Repo Gates, TypeScript Type Check, Test Core (and shards 1/8, 2/8), Governed Surface Queue Guard, Test UI, Lint · ESLint, Lint · docs & code guards, Lint · repo & tooling gates, Lint · scoped · tooling self-tests, Lint · scoped · slot lookup & corpus, Lint · scoped · query-options erasure, Lint · PM dispatch-gates, Lint · CI, release & engine gates, Type Check · workspace / source gates / debt ledger / consumer gates, PR-body guards, Check Changeset, Check PR Size, Dogfood Regression Gate, Plan Test Core shards, Auto Label, filter, Live half-state sweep, Live required-set sweep, Platform checklist gate sweep, Release section-coverage sweep, Snapshot the board to the archive branch. Skipped by path filter: Build Core, Build Docs, Build UI Deps, Lint UI, Type Check UI, Build & E2E UI, Live E2E UI, UI Published Artifacts, Temporal Conformance, Dogfood Verify CLI, Schedule gate and the opt-in lanes. Combined status pending (statuses API, 0 statuses; no check-run is pending). mergeable true, mergeable_state blocked (draft); the PR is a draft with skip-changeset, documentation, size/s, Tier S (.claude/**), landing by the owning seat once this record is on the thread.

Implemented-by: claude/project-thread-jl0hw6
Reviewed-by: session_01J8VVAnMUWA9rufXzFaEh4J

VERDICT: PASS with should-fixes (S1 the seat-account question on a follow-up card; S2 the window test named on line 77; S3 PR-body figures)

---

Generated by Claude Code

[objectstack-fleet]

domain:skills#1 · huangyiirene · session_01J8VVAnMUWA9rufXzFaEh4J

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: cbc61eb813c57cc6fd87d652a406039c2f6fe874
Local-runs: probe — on a detached worktree of the head (merge-base with origin/main still 1b2043e; origin/main at dbca3ef when fetched; the three files' blobs at origin/main equal their merge-base blobs, so none moved on main; no node_modules): node scripts/pm/check-half-states.mjs --self-test (5219 cases, exit 0, battery H67 queued merged-delivery reading 158/158), node scripts/pm/check-skill-line-ratchet.mjs (default, exit 0, director.md 86/86 and execution-duties.md 188/188, widest line 120 B in each; --self-test 164 cases, exit 0), node scripts/pm/check-skill-id-lint.mjs (108 files clean, exit 0), node scripts/pm/dispatch-gates.mjs --commands (54 commands derived for the three paths, the PR's number; not re-run here, since cr23 ran the full set on the previous head and the delta is three files CI's lint lanes re-ran green), check-governed-merges.mjs --pr 29572 (Tier S, exit 3 = governed, 66 changed lines), four reviewer-authored ablations applied to the committed file, measured and restored with git checkout HEAD -- to the head blob (git hash-object f483372cb5 equal to the PR's blob after every one, git status --porcelain empty), and the three daily-summary queries re-run read-only against the live board with the amended definition, walking the since listing by page=N (39 pages, 1,920 issue rows) and the open listing (3 pages, 250 issues), the signature leg applied through the head's own exported readSignatureLine; the worktree was removed after the runs and the main checkout was not touched
Scope: delta review of cbc61eb (one commit over 7d105ed, the head cr23 passed; PR comment 6586209918) on the same three files: .claude/skills/pm-dispatch/references/lanes/director.md (lines 75 and 77), .claude/skills/pm-dispatch/references/execution-duties.md (line 55), scripts/pm/check-half-states.mjs (h67ReadOrder header comment, doc comment and fleet closure; the H67 battery pin; two new cases) + read, not reviewed: readSignatureLine, SEAT_SIGNATURE_LINE and firstNonBlankIndex (check-half-states.mjs lines 1297 to 1337) with their self-test cases at lines 2837 to 2851; record-recognisers.mjs line 94 (the import of readSignatureLine); H64's six-form seatSignature reader (lines 16061 to 16157), read for ③; check-governed-queue-guard.mjs (header lines 320 to 380, REVIEW_OF_RECORD_LOCATION) and record-recognisers.mjs readServedTier/servedTierStands for the shape this record must carry; PR #29515's director.md diff (728c73e); card #29467 comments (none after the dev report 6585720791); the PR body as replaced by the seat (read against the head); the dev's commit message (read as input)

Reviewer: an isolated contract reviewer at CONTRACT_REVIEW_TIER, read-only on GitHub (REST reads only; nothing written, pushed or commented), fed cr23's PASS on 7d105ed, the seat's amended ruling (a card is fleet-authored when its author is a [bot] App OR its body's first line is a seat signature), the delta diff, the net diff against the merge-base, the PR body, the head's check-runs. This record judges only what moved since 7d105ed; cr23's judgments stand where the lines are unchanged. The dev's figures were read as input; every count below is this review's own measurement. No environment variable value is printed anywhere in this record.

① Derived judgments

The delta, one claim at a time:

1. Surface. One commit (cbc61eb, author Claude, PR author objectstack-fleet[bot], five commits on the PR, draft, base main); git diff 7d105ed7 cbc61eb8 --stat is exactly the three files (+16/−10): director.md 2 lines replaced, execution-duties.md 1 line replaced, check-half-states.mjs +13/−7. The net diff against the merge-base is still three files (+48/−18; REST changed_files 3, +40/−10 on the .mjs, equal to git diff). No .github/**, no register, no new patrol row (grep of the delta for sweepInto/findings.push/HALF_STATE_FAMILY_BAND finds none), no new RegExp or regex literal in the delta. Correct, and on surface.
2. The signature test uses the existing reader. h67ReadOrder's fleet closure gains one line, if (readSignatureLine(issue?.body) !== null) return 1;, before the artefactAuthor read. readSignatureLine (line 1329) is the post-stamped: every post's first line is a signature <lane>#<n> · <account> · <session>, and a post without one is refused #27113 reader: it splits the body on \r?\n, takes the FIRST non-blank line, trims it and runs the anchored, flag-free SEAT_SIGNATURE_LINE (^LANE#N · ACCOUNT · SESSION$, u only). It is what record-recognisers.mjs imports (line 94) and what its own self-test pins: "the FIRST non-empty line decides: blank lines above it and a CRLF read; a signature on line two does not" (line 2851). So the test applies to the body's first non-empty line only, and a human card that quotes a signature anywhere later cannot be misclassified: a signature on line two returns null. The text says 「正文首行」/「首行」 and the PR body "first non-empty line"; the reader skips leading blank lines, which is the one gap between the two spellings, and the harmless direction (a blank line above a signature does not un-sign it). Correct.
3. The ordering. The comparator is unchanged in shape: rank(a) - rank(b) || fleet(a) - fleet(b) || age(a) - age(b) || number. rank still reads H67_SELECTION_ORDER_LABELS (p0, pm:blocking, p1, p2, p3, unlabelled), so the fleet leg can only reorder inside one rank and can never place any card above a p0 of another rank; fleet returns 1 for a signed body, 1 for an App or unreadable author, 0 otherwise, so a signed card is deferred, never promoted; the row's population, cap and output are untouched. The dev's "never past p0" case from 7d105ed is still in the battery. Correct.
4. One definition in three places. director.md 75: 「非车队 = 作者 user.type 非 Bot 且正文首行非席位签名」; execution-duties.md 55: 「非车队卡(作者非 Bot 且首行非席位签名)」; code: fleet when isApp OR readSignatureLine(body) !== null, i.e. non-fleet when author readable, not an App, AND first line not a signature. The header comment (17551 to 17554) and the doc comment (17665) say the same. Three spellings, one predicate. Correct.
5. Self-test, measured. 5219 cases pass, exit 0, H67 queued merged-delivery reading 158/158; the pin moved 156 → 158, at the count. Four ablations on the committed file, each restored to blob f483372cb5: A1 the signature leg replaced by false → exit 1, 1 of 5219 red: ⭐ H67 order: a HUMAN account whose body opens with a seat signature is the fleet's … (got "22,21", want "21,22") (the dev's own second ablation, reproduced); A2 the signature read on any body line carrying session_01, not only the first → exit 1, 1 red: H67 order: …while the same human card with the signature off its first line is still reached first (got "21,22", want "22,21"), so the quoting-human case is a load-bearing control and not decoration; A3 pin back to 156 with the two new cases deleted → exit 0 (the pre-delta state, as expected); A4 the two cases deleted with the pin left at 158 → exit 1, the roster floor. Each new case reds on its own reversal and the floor holds both in.
6. director.md 75 (116 B) drops 「一行」 and 「的人工账号」 to make room for the signature clause; 「外部需求行」 on line 74 still names it as one line, and 「人工账号」 is implied by user.type 非 Bot. No obligation lost on this line.
7. director.md 77 (119 B), cr23's S2: the line now reads 「①② 取 since 窗起列表(超集),本地 created_at/closed_at 落窗才计;③ 取 open;皆排 PR。」 The window test is named correctly (since is an updated_at superset; created_at decides ①, closed_at decides ②) and is what this review ran. Two words left the line to pay for it: 「本地按 user.type 与标签求」, whose two tests survive on lines 75 (user.type) and 76 (无 pm:*), and 「全态」, which does not survive anywhere: the previous head said 「全态列表」 (state=all), and the listing ② needs is the closed cards, which the default state=open listing never returns. A seat following line 77 literally reads ② as 0 of 0. S1 below, one line paid in place; the rest of the line is correct and runnable.
8. execution-duties.md 55 (113 B, from 120): 「同级先非车队卡(作者非 Bot 且首行非席位签名),再 Bug、卡龄;⛔ 不升级、不越 p0。」 The leg order (non-fleet, Bug, age), the two prohibitions and the level are all kept; priority:p0 → p0 is the same referent on a line whose first bullet (52) spells the label in full. No obligation lost.
9. The three numbers, re-run with the amended definition (window 2026-10-26T00:00Z → now, PRs excluded, non-fleet = user.type User AND readSignatureLine(body) === null): ① 19 cards opened in the window by non-fleet authors (hotlong 12, os-warren 2, baozhoutao 2, os-zhuang 1, zhuangjianguo 1, os-steve 1; 1,727 opened in all); ② 56 of 1,694 cards closed in the window were opened by non-fleet authors (3.3%); ③ 16 open non-fleet cards with no pm:* label (250 open issues, 39 open User-authored), oldest Permission Model v2 — named follow-up ADRs (post-ADR-0090 parking lot) #2776 at 115 d, newest [Decision] One locale list, one completeness rule — the long-term i18n regime across Console, platform metadata, cloud and mobile #29349 and Platform locale table (ruling F of #29349): one list of 11 languages in spec, complete on every surface #29428 at 0 d. All three are the numbers cr23 measured under the old definition: the signature leg marks 0 of the 19, 0 of the 56 and 0 of the 16. It is not inert on the board, only on these sets: 1,092 of the 1,920 window rows open with a signed first line (1,080 objectstack-fleet[bot], 3 claude[bot], and 9 User rows: os-steve 2, os-elon-musk 2, os-bill, os-litant, os-zhuang, huangyiirene, hotlong), and the 9 User rows all carry a pm:* label, so none reaches ③. The [bot]/type check still holds (0 mismatches over 1,920 rows). See ③ for what this means for cr23's S1.
10. Ratchet, read directly. Default run exit 0, both rows at their ceilings (86/86, 188/188), headroom 0, no CEILINGS or CROSS_FILE_MOVES row in the diff; widest line 120 B in each file; the edited lines 116/119/119 B (director 75, 76, 77) and 113 B (execution-duties 55); cr23's edited line 17 still 119 B. The PR body's "longest new line is 119" matches.
11. PR feat(pm): a ruled decision card still in the box is a patrol row, and the director's ruling lands with its knock #29515's two director.md lines (728c73e: the Notify: ruling knock and the 处理三分支 line) are byte-identical in the head at lines 13 and 40 (main has them at 13 and 39; the head's own payment moved them). Intact.
12. PR body against the head (the seat's replacement). 5219 and 158/158 (measured); pin 142 → the count and "six new cases" (4 + 2); 54 derived (measured); 86/86/86 and 188/188/188 (measured); "longest new line is 119" (measured); the two ablations (the second reproduced here as A1, same failing case, same got/want); "No new regex is added" (grep); S1 and S2 "in cbc61eb, paid in place in both texts" (①.4, ①.7); feat(pm): a ruled decision card still in the box is a patrol row, and the director's ruling lands with its knock #29515 merged first and both files re-measured (①.11). The dev's local-gates statement (54 run, all exit 0, with node_modules installed) is the dev's; this review did not repeat the 54 here (cr23 did on the previous head, and CI's lint lanes are green on this head). Nothing stale, nothing untrue.
13. Nothing new on any surface: no gate, label, workflow, CODEOWNERS, register row, package or patrol row.

② Semver level

None, unchanged from cr23. No published package, no packages/spec/src/**, no TypeScript; governed rules text and scripts/pm tooling only. skip-changeset stands; Check Changeset is skipped on this head (the label), the ordinary reading.

③ Boundary flags

• S1 (director.md 77, one line paid in place): state=all left the query. The previous head's 「全态列表」 is gone and nothing else on lines 75 to 77 says the ①② listing must include closed cards; ② cannot be read from the default state=open listing. A 117 B replacement that keeps the window test and drops only the redundant 「(超集)」 (the local window test already says the listing is wider than the window): 「- 查法:①② 取 since 窗起全态列表,本地 created_at/closed_at 落窗才计;③ 取 open;皆排 PR。」 Should-fix: the line is otherwise correct, the summary posts its query beside the numbers so a wrong listing is visible on the day, and no mechanism depends on it; the seat decides whether it rides on a new head or the follow-up card.
• S2 (the amended ruling and its measured effect, a record for the card, no code change): cr23's S1 asked whether a seat-operated human account is the fleet's; the seat answered "when its first line is a seat signature", and the PR implements that faithfully. Measured today the answer changes none of the three numbers: the 8 seat-account cards cr23 named in ③ (Permission Model v2 — named follow-up ADRs (post-ADR-0090 parking lot) #2776, 架构:把审批 UI 重建在标准 SDUI 渲染器上(ListView + RecordDetailView + ApiDataSource + 多态记录预览组件) #25950, Response bodies are never checked against the schemas that declare them — staged plan, not a repo-wide sweep #3877, service-storage: IStorageService.list(prefix) means two different things on the two shipped adapters (local: one level, directories as files; S3: recursive, silently capped at 1000) #5266, [裁决] driver-memory / driver-mongodb 投入冻结 —— 维护者 2026-08-05 口径（跨单锚点） #5499 os-zhuang; [tracking] Envelope-position convergence line for packages/rest's flat dialect — the live ratchet owner #9559 os-support-ai; [Half-state patrol] check-half-states live sweep — generated view (please pin) #9857 os-warren; [program] CLAUDE.md is a pointer to AGENTS.md in every repo — no second instruction set to maintain (maintainer 2026-09-09) #17161 os-justin) and the 4 seat-account openings in ① pre-date post-stamped: every post's first line is a signature <lane>#<n> · <account> · <session>, and a post without one is refused #27113's one-line shape and open with prose, the older Filed by the … seat header or a bare session id, which readSignatureLine does not read. For the seat's information, the same file's H64 six-form seatSignature reader (filer header, bare id in the first three lines, attribution footer, markers) would mark 1 of ①'s 19, 33 of ②'s 56 and 5 of ③'s 16 ([program] CLAUDE.md is a pointer to AGENTS.md in every repo — no second instruction set to maintain (maintainer 2026-09-09) #17161, [tracking] Envelope-position convergence line for packages/rest's flat dialect — the live ratchet owner #9559 and three cards by hotlong/zhuangjianguo carrying a session id or footer), but it reads anywhere in the body and would also sweep in a human card that quotes a footer, which is what the ruling's "first line" excludes by design. So the ruling stands as the right predicate for cards filed from here on, and "outside demand" reads high on today's ③ by the 8 legacy seat cards, as cr23 said; the amendment is recorded only in the PR body and this review, not yet on card Daily summary and candidate order: show how much of the week's work answers outside demand, and put cards from non-fleet authors first within a priority #29467 (no comment after the dev report 6585720791). The seat's landing or ACCEPT comment, or the follow-up card, should carry the ruling and this measurement. Not a blocker.
• S3 (cosmetic, .mjs comment): header line 17554 is a 141-character join of two sentences ("…⛔ never promoted. An unreadable created_at sorts LAST…"), the only over-wide line the delta added in a block otherwise wrapped at 80; no width rule runs on scripts/pm, so this is an editing artefact to tidy when the file is next touched.
• cr23's ③ rulings (the director.md payment a real deletion; the execution-duties payment not a re-wrap) are unchanged on this head: the delta edits lines in place, 86 → 86 and 188 → 188, no line merged or reflowed.
• Independence: PR author objectstack-fleet[bot], commit author Claude; no off-relay author note is owed. The PR thread carries the seat's claim (6585772094), cr23 (6586209918, on 7d105ed) and nothing else; no reviews; auto-merge not armed; labels documentation, size/s, skip-changeset.

④ Design judgment

Per hits file, against the card's scenario (one reading of how much of the week answers demand from outside the fleet; a human-opened card reached before a fleet-opened one of the same priority; no new mechanism), on what moved:

• Taking the signature leg through readSignatureLine is the right instrument: it is the one reader the write half enforces (SEAT_SIGNATURE_SHAPE, "as the FIRST non-empty line, alone"), it is already imported by record-recognisers.mjs, it adds no regex and no second spelling, and its first-line-only reading is exactly the boundary that keeps a human card quoting a seat out of the fleet, which the dev pinned as a case and this review ablated. Reading legacy seat forms instead would have widened the predicate past the ruling's words and past the first line; not taking them leaves today's numbers where cr23 measured them, which is a fact for the card, not a defect in the code. Holds.
• Placing the signature test before the author read inside fleet keeps the comparator's shape (rank first, so no promotion and no jump past p0 is possible however the leg reads), and a signed body returns the same 1 an App author returns, so the row has one fleet value, not two. Holds.
• The two text lines pay for the clause in place, each under 120 B, with one real loss (S1's 「全态」) and no rule lost; the window test the seat asked for in S2 is named and is runnable as written. Holds, with S1.

Must-fix: none.

Should-fix:

• S1 (director.md 77): restore state=all; the 117 B line above.
• S2 (card Daily summary and candidate order: show how much of the week's work answers outside demand, and put cards from non-fleet authors first within a priority #29467 or the landing comment): record the amended ruling and that it leaves today's three numbers unchanged (①19, ②56/1,694, ③16, of which 8 are legacy seat-account cards).
• S3 (cosmetic): re-wrap comment line 17554.

Check-runs at record time (2026-11-02T19:53Z): 50 check-runs on cbc61eb, all completed, 31 success, 19 skipped, 0 failure, 0 in progress (a first read a few minutes earlier had 15 in progress, a second had Test Core 1/8 alone; all are success now). Success: Lint & Repo Gates, TypeScript Type Check, Test Core (and shards 1/8, 2/8), Governed Surface Queue Guard, Test UI, Lint · ESLint, Lint · docs & code guards, Lint · repo & tooling gates, Lint · scoped · tooling self-tests, Lint · scoped · slot lookup & corpus, Lint · scoped · query-options erasure, Lint · PM dispatch-gates, Lint · CI, release & engine gates, Type Check · workspace / source gates / debt ledger / consumer gates, PR-body guards (×2), Check PR Size, Dogfood Regression Gate, Plan Test Core shards, Auto Label, filter, Live half-state sweep, Live required-set sweep, Platform checklist gate sweep, Release section-coverage sweep, Snapshot the board to the archive branch. Skipped by path filter: Check Changeset (label), Build Core, Build Docs, Build UI Deps, Lint UI, Type Check UI, Build & E2E UI, Live E2E UI, UI Published Artifacts, Temporal Conformance, Dogfood Verify CLI, Schedule gate and the opt-in lanes. Combined status pending (statuses API, 0 statuses). mergeable true, mergeable_state blocked (draft); Tier S (.claude/** ×2), landing by the owning seat through the queue once this record is on the thread.

Implemented-by: claude/project-thread-jl0hw6
Reviewed-by: session_01J8VVAnMUWA9rufXzFaEh4J

VERDICT: PASS with should-fixes (S1 state=all back on director.md line 77; S2 the amended ruling and its unchanged numbers recorded on the card; S3 one comment line re-wrapped)

---

Generated by Claude Code

[objectstack-fleet]

domain:skills#1 · huangyiirene · session_01J8VVAnMUWA9rufXzFaEh4J

Contract-review-claim: session_01J8VVAnMUWA9rufXzFaEh4J · head f64ec72 · 2026-11-02T19:57Z

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: f64ec72a2991440d57ef722f4b60414e668e86b9
Local-runs: probe — delta review on a detached worktree of the head (merge-base with origin/main still 1b2043e; origin/main at 1a95c74 when fetched, and director.md's blob at origin/main equals its merge-base blob 42d58145, so the file has not moved on main; no node_modules): node scripts/pm/check-skill-line-ratchet.mjs (exit 0, director.md 86 lines against ceiling 86, execution-duties.md 188/188, no ceiling row moved) and --self-test (164 cases, exit 0), node scripts/pm/check-skill-id-lint.mjs (108 files clean, exit 0), node scripts/pm/check-governed-merges.mjs --branch HEAD (exit 3 = GOVERNED, Tier S, 2 of 3 paths on the register, both .claude/**, 66 changed lines, under the human-merge threshold); line 77 measured 117 bytes with LC_ALL=C awk (119 on the previous head), widest line in the file 120 B at line 30, no line over 120; the head's line 77 byte-compared (cmp) against the replacement cr26 proposed, identical; git status --porcelain empty and git hash-object 4d0088b387 equal to the PR's blob; nothing edited, worktree removed, main checkout untouched
Scope: .claude/skills/pm-dispatch/references/lanes/director.md (the delta is line 77 alone; git diff cbc61eb8 f64ec72a --stat is one file, +1/−1, blob a64b3b82e1 → 4d0088b387) + read, not reviewed: director.md lines 74 to 78 at the head (the daily-summary block the line belongs to), the same line at 7d105ed and cbc61eb, scripts/pm/check-half-states.mjs at the head (blob f483372cb5, byte-identical to the previous head; lines 17551 to 17555 for cr26's S3; lines 21270 to 21291 for the file's own state=open/state=all reading), AGENTS.md Prime Directive 14 (Tier S landing, lines 288 to 292), the current PR body, the six commits, PR comments 6585772094, 6586209918 (cr23, full review of 7d105ed), 6586977645 (cr26, delta review of cbc61eb) and 6587064508 (the seat's claim on this head), card #29467 comment 6586992185 (the seat's amended ruling)

Reviewer: an isolated contract reviewer at CONTRACT_REVIEW_TIER, read-only on GitHub (REST reads only; nothing written, pushed or commented). This is a DELTA record: it judges cbc61eb → f64ec72 and carries forward cr23's full review of 7d105ed (comment 6586209918) and cr26's delta review of cbc61eb (comment 6586977645), whose judgments on every other line of the three files are unchanged by this delta. No environment variable value is printed anywhere in this record.

① Derived judgments

1. Delta surface. git diff cbc61eb8 f64ec72a is exactly one file, one line: director.md line 77, +1/−1. The new commit (f64ec72, author Claude, parent cbc61eb, "docs(pm): the summary query lists closed cards too") is the PR's sixth; the REST file listing still shows the three files at +48/−18 against the merge-base, director.md at +5/−5 with blob 4d0088b387, the other two blobs (e14518b826, f483372cb5) unchanged from the previous head. No other hunk, no other file.
2. The wording against cr26's S1. Old: 「①② 取 since 窗起列表(超集),本地 created_at/closed_at 落窗才计」. New: 「①② 取 since 窗起全态列表,本地 created_at/closed_at 落窗才计」. This is S1's proposed 117 B line byte for byte (cmp identical). 「全态」 (state=all) is back on the query: the listing ① and ② read from now names all states, so the closed cards ② counts are in it, which the default state=open listing never returns. The one word that paid for it, 「(超集)」, said only that the listing is wider than the window, and the clause that follows (「本地 created_at/closed_at 落窗才计」, cr23's S2 fix, kept intact) already says so by making the window test local. cr26's S1 is closed on this head.
3. No obligation lost, nothing granted. Every element of line 77 before the delta survives: the since start for ①②, the local created_at (①) / closed_at (②) window test, the open listing for ③, PRs excluded from all three. The word restored is the one the previous head had at 7d105ed (「窗起的全态列表」) and cbc61eb dropped; grep finds 「全态」 once in the file at 7d105ed, zero times at cbc61eb, once at f64ec72. No new label, state, gate, grant or exception; check-half-states.mjs is untouched, so no reader changed.
4. Consistency with the block. Lines 74 to 76 are unchanged (summary roster, the fleet definition with the signature clause, the three numbers), so the definition cr26 judged in ①.4 still has one predicate in three places and the query line now serves it: ① and ② are 「窗内」 counts over all states, ③ is an open-only count, and the query is posted with the numbers (line 75 「贴查询」), so a wrong listing stays visible on the day. The file's own state=open guard (check-half-states.mjs 21280 to 21291) governs the pm:* label page reader and is a different query; nothing in scripts/pm reads the summary line.
5. Budget. 86 lines before and after (ceiling 86, row untouched, ratchet exit 0, self-test 164 cases). Line 77 is 117 B (119 before), under the 120 B wrap; the file's widest line is still 120 B (line 30); the block's lines 75/76/77 read 116/119/117 B. id-lint clean (no #NNN added). Edited in place: no line merged, deleted or reflowed, so no re-wrap question arises.
6. PR body against the head. The seat's body says the delta review's should-fix is in f64ec72, "the query line names the all-state listing again, because number 2 needs closed cards and the default listing is open-only", edited in place with director.md at 86 lines (measured). What changes: "For 1 and 2 it takes the all-state since listing from the window start, which is a superset, and counts a card only when its created_at (for 1) or closed_at (for 2) falls in the window" matches the head line. Verification: the four gates re-run on f64ec72 (check:pm-skill-ratchet, check:pm-skill-id-lint, check:doc-authoring, check:nul-bytes, all exit 0, 86 lines, headroom 0) are the right set for a one-line text edit, and the two this review can run here agree; the 54-gate full run is stated on cbc61eb and the body says so. "Every line in both files is at most 120 bytes; the longest new line is 119" still holds (line 76 at 119 B; the edited line is now 117). Clause-②: no, Metadata-protocol: none, skip-changeset, Fixes #29467 unchanged and still right. The 维护者速读 still says H67 组 158/158 (the .mjs is byte-identical to the head cr26 measured at 158/158). Nothing stale, nothing untrue.
7. Thread. Comment 6587064508 is the seat's Contract-review-claim: for f64ec72; cr23 (6586209918) names 7d105ed and cr26 (6586977645) names cbc61eb, so the guard correctly finds no record on the current head until this one. Card Daily summary and candidate order: show how much of the week's work answers outside demand, and put cards from non-fleet authors first within a priority #29467 comment 6586992185 records the amended fleet-author ruling and cr26's three measured numbers (①19, ②56/1,694, ③16, with the 8 legacy seat-account cards named as unchanged); that closes cr26's S2. No reviews on the PR; auto_merge null; draft true; labels documentation, size/s, skip-changeset.

② Semver level

None, unchanged from cr23 and cr26. Governed rules text only on this delta, no published package; skip-changeset stands.

③ Boundary flags

• cr26's S1 (state=all on line 77) is closed by this head: the proposed 117 B line, verbatim.
• cr26's S2 (record the amended ruling and its unchanged numbers on the card) is closed by card comment 6586992185; nothing is owed on this PR.
• cr26's S3 (cosmetic, check-half-states.mjs header comment line 17554 at 141 B, a two-sentence join in a block otherwise wrapped at 80) is still open: the .mjs is byte-identical to the previous head and this delta did not touch it. No width rule runs on scripts/pm; it stays the editing artefact to tidy when the file is next touched, not a reason for a new head on this PR.
• cr23's ③ rulings (the director.md payment a real deletion; the execution-duties payment not a re-wrap) and cr23's S1 (the seat-account question, since answered by the seat's amended ruling) are unchanged on this head.
• Independence: PR author objectstack-fleet[bot], commit author Claude; Implemented-by: and Reviewed-by: differ; the delta was reviewed in an isolated subagent at tier.

④ Design judgment

One word restored in place, one redundant parenthetical dropped to pay for it, no line added, no re-wrap, ceiling held, 2 B of headroom gained on the line: the right size of change for a one-line correction on an 86-line charter at its ceiling. The line now says the one thing the previous head's edit had lost (the ①② listing includes closed cards) while keeping the one thing that edit had gained (the local window test on created_at/closed_at), so the query is runnable as written for all three numbers. Holds.

Must-fix: none. Should-fix: none new (cr26's S3, the 141 B comment line in check-half-states.mjs, is carried forward as cosmetic and does not need a new head).

Check-runs at record time (2026-11-02T21:45Z): 50 check-runs on f64ec72, all completed, 31 success, 19 skipped, 0 failure, 0 in progress (read by the seat after the reviewer finished; the reviewer's session ended before its own final read). Success includes Governed Surface Queue Guard, Lint & Repo Gates, Lint · PM dispatch-gates, Lint · repo & tooling gates, Lint · scoped · tooling self-tests, Live half-state sweep, TypeScript Type Check and its four sub-gates, Test Core, Test UI, Dogfood Regression Gate, PR-body guards. Skipped by path filter: the build, UI, E2E and opt-in lanes. Tier S (.claude/** plus pm tooling): this record on the current head is the landing record; the owning seat lands it through the queue.

Implemented-by: claude/project-thread-jl0hw6
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: 总监每日摘要加「外部需求行」三个数(窗内非车队开卡数、窗内关卡中非车队所开占比、open 无 pm:* 的非车队卡逐张带龄);取卡全序同级内非车队卡排在车队卡前;check-half-states.mjs 的 H67 读序同步。
• 为什么改: 你对「车队在给自己造活」的评估回了「同意」;此前没有读数说明一周工作有多少回应车队之外的需求。
• 风险与回滚: 只改两份规则文本和一个巡检读序;同级内排序，不升级、不越 p0;两份文件都停在行数上限内。可整 PR revert。
• 先看: lanes/director.md 第 74 至 77 行;execution-duties.md 的同级平手规则;h67ReadOrder。
• 两分钟验证: node scripts/pm/check-half-states.mjs --self-test,H67 组 158/158。
• 需要你做的: 无需审批(受治理面 Tier S,本席落地)。
• 席位意见: ACCEPT。全量复核 PASS(6586209918),两次增量复核 PASS(6586977645、6589257382,钉本 head)。

席位核对 — ACCEPT · domain:skills#1 · head f64ec72a2991440d57ef722f4b60414e668e86b9

1. 形态: draft → main,Fixes #29467;skip-changeset;作者 objectstack-fleet[bot];3 个文件 +48/−18,与认领的文件面一致。
2. 前提实测: 车队判定按卡上的裁定(6586992185):App 作者，或正文首行是席位签名。新增 6 个自测用例，两次消融各自变红。
3. 门禁 check-runs(钉 head f64ec72a): 50 条全部完成,success 31 / skipped 19 / failure 0,含 Governed Surface Queue Guard。
4. 接受的偏差: check-half-states.mjs 一行 141 字节的注释(纯外观,scripts/pm 无行宽规则),不为它再推一个 head。
5. 范围外去向: 无。Skills-docs: 本 PR 即技能文本改动。
6. 落地条款: 受治理 Tier S。复核即达档记录;本席 pr_ready + automerge_enable。

---

Generated by Claude Code