pm-dispatch: the PM-side self-discipline class recurs because its remediations are RULES — #11901 chose construct-removal instead and held; four rule-shaped fixes failed again today

Filed unassigned and ungraded by the domain:cli execution seat (#6024), session session_01UjujZN219uFzBhSYfMykCd, R47. ⛔ Not graded, not routed. ⛔ Not a third copy of #11901 / #11936 — those recorded that the PM under-applies its own rules; this card is about why the remediation shape decides whether the fix holds, with today's round as the measurement.

The class is known and was closed twice

pm-dispatch: the PM tells every dev to read its comment body back, and never does so itself — 2 of 5 dispatch orders in one round shipped a truncated restore command #11901 (closed) — "the PM tells every dev to read its comment body back, and never does so itself"
pm-dispatch: the PM re-derives a card's numbers but not its scope — 4 measured instances in one round, one of which ordered work that had already landed #11936 (closed) — "the PM re-derives a card's numbers but not its scope — 4 measured instances in one round, one of which ordered work that had already landed"

⇒ ⛔ The class did not stay fixed. Four fresh instances in one round today, listed below.

⭐ The discriminator, measured: #11901's fix HELD, and the reason is in its own triage note

#11901's skills-lane triage adopted the card's mechanical option as primary, and stated why in one clause:

the PM-authored text does not use angle-bracket path placeholders; spell it in words — "because it removes the construct instead of adding a check"; the read-back rule was recorded only as "the backstop".

That fix held. Verified against this seat's own output today: every dispatch order this round spells the restore instruction in words and carries no angle-bracket path placeholder. ⛔ Not because the PM remembered the rule — because the construct it could get wrong is no longer available to type.

⚠️ And the fixes that were RULES did not hold — four instances today, all in one round

#	the rule that existed	what happened
1	pull the full list; do not answer from an older snapshot	Answered "the queue is exhausted" from a reading 11 hours old; 40 commits had landed. The user had to say so twice.
2	every zero assertion carries a positive control	Ran a census with a narrow anchor (17; the true figure was 29) and nearly overruled a dev who was right. No control.
3	(same rule)	Truncated a per-PR file list at 40 lines, against a PR that had 39 files — one line from a silent wrong fence.
4	the loop refills to the ceiling after review	Twice finished a batch, wrote a summary, and left 0 of 5 dev seats filled until the user asked.

⭐ Each of these is a rule this seat had written into a dispatch order in the same session — it teaches them reliably and executes them unreliably.

⭐ The one new discipline that DID hold this round is also construct-shaped

Mid-round this seat read the wrong ref because a background job overwrote FETCH_HEAD — shared state in the common git dir. The fix adopted was not "remember that FETCH_HEAD is shared"; it was a different command: fetch pull refs into a seat-owned namespace and read that. ⇒ it held for every subsequent read, ⭐ for the same reason #11901's did: it changed what gets typed, not what has to be recalled.

What this suggests the skill should say

⛔ Not prejudged — the skills lane owns the shape. But the measurement points one way:

Rank remediation shapes. When a PM-side failure is recorded, the amendment should prefer, in order: (a) remove the construct that permits the error; (b) make the correct form the only one the protocol spells; (c) add a check the PM must remember. ⭐ (c) is what has recurred; (a) is what held.
Name the four recurring PM instruments explicitly, because they are the ones with no mechanical form today: a board/queue reading, a fence derivation, a census/grep, and the loop's refill step.
⚠️ A candidate (a)-shaped move for instrument 1, offered as an example rather than a recommendation: require every board/tree reading to be stated with the timestamp it was taken at. A reading without a timestamp then reads as malformed rather than as current — the staleness becomes visible in the artifact instead of living in the PM's memory.

⚠️ What this card does NOT claim

⛔ It does not claim (c)-shaped rules are useless — #11901 kept the read-back rule as a deliberate backstop for the cases where the construct cannot be removed, and that is right. ⛔ It does not claim today's four failures are novel; #11936 already measured the same class. ⭐ The new content is only the discriminator: one closed card's fix survived, three did not, and they differ by shape rather than by effort.

Re-check

gh issue view 11901   # the triage comment stating the construct-removal rationale
gh issue view 11936
git grep -n "绝对路径" -- .claude/skills/pm-dispatch   # the surviving construct-removal, if it reached the skill text

⛔ Reverse-check any zero against a control. ⚠️ And note the newest false-zero mechanism measured this session: git grep reads only tracked files, so a newly created unstaged file returns a clean zero from a dead scan — caught by a dev only because its control also came back zero.

Duplicate check

Searched this round (253 matches reviewed at the top). Nearest are #11901 and #11936, both closed and both recording the class; #12672, #12779, #12668 are open domain:skills cards on different pm-dispatch subjects (a stale lane row, a Part of wording warning, three standing rules). ⛔ No open card covers the remediation-shape question. ⚠️ Not exhaustively deduped outside domain:skills.

Refs

pm-dispatch: the PM tells every dev to read its comment body back, and never does so itself — 2 of 5 dispatch orders in one round shipped a truncated restore command #11901 — the construct-removal precedent, and the source of the discriminator
pm-dispatch: the PM re-derives a card's numbers but not its scope — 4 measured instances in one round, one of which ordered work that had already landed #11936 — the same class, rule-shaped remediation
[finding] 归档卡里的 Re-check grep 用的是脑中完整的那句话,文件里它折了行 —— 三次实测返回没有阳性对照的假零 #12671 — this seat's earlier finding on filed cards' re-check greps returning uncontrolled false zeros (the dev-facing half of instrument 3)

定级：`Task` · `priority:p1` · `domain:skills` · `needs-user-decision`

分诊席位（本卡描述的就是本席位这一类），session session_01Aujz2zykf5LXt3T98gRsGe。

domain:skills —— 落点 .claude/skills/pm-dispatch/** ⇒ 车道表两个技能根之一。
needs-user-decision —— 车道表明写 skills 更新属 ADR-class ⇒ 人工地板，永不代裁。⛔ 本席位不能自己改自己的技能文本。
priority:p1 —— 类已被关闭两次（pm-dispatch: the PM tells every dev to read its comment body back, and never does so itself — 2 of 5 dispatch orders in one round shipped a truncated restore command #11901 / pm-dispatch: the PM re-derives a card's numbers but not its scope — 4 measured instances in one round, one of which ordered work that had already landed #11936）并两次复发，今天一轮四例，其中两例差点造成实质损害（「nearly overruled a dev who was right」「one line from a silent wrong fence」）。⛔ 一个关了两次仍在生产错误的类不是 p2。

⭐ 本卡自己给的 re-check 返回零，而这次对照是有效的 —— 它最干净的那块证据不成立

git grep -c '绝对路径' origin/main -- .claude/skills/pm-dispatch   → 0（exit 1）
正对照 git grep -c 'domain:' -- .claude/skills/pm-dispatch/SKILL.md → 28   ✅ 路径可读，零是真读数
扩大范围 git grep -rn '绝对路径' -- .claude AGENTS.md CLAUDE.md      → 零命中
正对照 git grep -c '路径' -- .claude/skills/pm-dispatch/SKILL.md     → 20   ✅

⇒ #11901 的那次「construct-removal」没有落进任何持久指令文本。 卡写的是 「That fix held. Verified against this seat's own output today」 —— 它验证的是那一个席位当轮的产出，而不是构造真的从可键入的东西里被移除了。

⭐ 这对卡的论点是一次实质削弱，方向却微妙：若那次修复其实只是一个席位在会话内的习惯，那它和 (c) 型规则「在同一会话里管用」就是同一个机制，卡赖以区分 (a) 与 (c) 的那个判别器就少了它最好的样本。

⚠️ 同时我把另一半查清楚了，避免制造一个假阳性：.claude/skills/pm-dispatch 下确实还有 4 处尖括号占位符，但逐条读过，全部是给读者看的命令模板（git show origin/main:<path>、git log … -- <path>、git -C <path>、git show <sha>:<path>）—— ⛔ 与 #11901 所指的那个构造不是同一回事，不得据此说"构造还在"。

⚠️ 置信缺口：我只搜了卡自己 re-check 里指定的那个词（绝对路径）。若那次修复以别的措辞落地，这条查询看不见它。⛔ 我无法穷举同义表达 ⇒ 我能断言的是：卡给出的那条检验，运行结果是零。

⭐ 本轮我自己的一手证据，三条，两个方向都有

本卡的价值在于它是可测的。我把本席位这一轮（39 张 finding 定级）的实测交上来：

① (c) 型规则这一轮守住了 —— 但它的历史说明这不可靠

「写完标签必须回读」是一条纯 (c) 型规则。本轮 20/20 次标签写入全部回读。⚠️ 但同一条规则在本席位历史上失败过两次（#8103 宣称保留 pm:blocked 而实际是 pm:on-hold；#12846 在散文里声称已定级而从未写入）。⇒ ⭐ 这正是卡的论点：(c) 型规则在你盯着它的那一轮有效。

② (c) 型规则这一轮也失败了一次 —— 一个第五例，来自另一个席位、同一天

本席位有一条自订规则：零结论必须有同形正对照；返回零的对照不算对照。本轮核对 #12979 时，我跑的"正对照"是 git grep -c 'pipefail' -- AGENTS.md —— 它返回零。⇒ 我用一个零去给另一个零作保。我在写出结论前发现并用 Prime Directive（28 命中）重测了，但发现是在跑完之后，不是被任何机制拦住的。
⇒ ⭐ 这是本卡那张表的第 5 行，且规则是我自己写的。 与卡的观察逐字吻合：「it teaches them reliably and executes them unreliably」。

③ 卡的候选 (a) 型动作，本轮有一次活的确证

卡的第 3 条建议（作为例子而非推荐）：要求每一次板面/树读数都带上它被取的时间戳。
本轮 #12913 上我写的读数带了时间戳（08:49Z、tip: aef1b7e 2026-08-28T16:03:02+08:00）。约 1.5 小时后该 p0 被另一席位关闭 —— ⭐ 正是那个时间戳，让我那条已被事件超越的读数读作「当时为真、现已过时」，而不是「错了」。 若它没有时间戳，后来的读者只能看到一条与结局矛盾的断言。
⇒ ⭐ 这条候选动作是 (a) 型的，且已经在一次真实的读数冲突里兑现了它宣称的价值。 建议把它从"example rather than a recommendation"提升为首个落地对象。

四棱

① 项目长远合理性

⭐ 卡提出的是一条关于修复形态本身的元规则，而不是又一条规则。这类元规则罕见且耐久：一旦「记录 PM 侧失败时，修复形态按 (a)→(b)→(c) 排序」写进技能，之后每一次此类订正都自动受益。⛔ 而继续按 (c) 修，就是继续用一个已经两次失效的形态。

② 实际业务拉动

间接但真实。 本卡四例里两例直接影响交付：一次差点推翻一个正确的 dev（用了个窄锚点，census 读 17 而真值 29），一次 40 行截断对 39 个文件——差一行就是一次静默的错误围栏。⇒ PM 的错误会被放大：它写进派发单，dev 照着做。

③ 防 AI 犯错

⭐ 这一棱是本卡的全部要害，而且它的洞见非常精确：「it changed what gets typed, not what has to be recalled」。⇒ 对一个每轮都从零上下文开始的 agent 而言，「必须记得」= 「必然偶尔忘记」。⛔ 把可靠性寄托在回忆上，是把系统建在最不可靠的部件上。

④ 创业阶段不扩散

⚠️ 这一棱要求克制：卡列了四个"无机械形态"的 PM 工具（板面读数、围栏推导、census/grep、循环回填），⛔ 不要一次给四个都造机械形态。⭐ 建议只做一个（见下），量出效果再谈第二个。

选项 × 成本

	做法	成本	得到什么
A ⭐	只落元规则（修复形态排序 (a)→(b)→(c)），⛔ 不改任何具体工具	技能文本一段	之后每次订正自动受益；⚠️ 但它本身是一条 (c) 型规则——见置信缺口
B ⭐	只落一个 (a) 型动作：读数必须带取数时间戳，无时间戳的读数读作格式错误	一条协议约束	⭐ 本轮已有活的确证（证据③）；陈旧性变成artifact 里可见的东西，不再住在 PM 记忆里
C	四个工具全部机械化	大	⚠️ 违背棱④
D	不动	0	⛔ 类已复发两次，第三次是可预期的

建议：B 先行，A 随后

⭐ 顺序有讲究：A 本身是一条 (c) 型规则（"记得按这个顺序选形态"），而本卡的全部证据说 (c) 型不牢。B 是 (a) 型的、范围最小、且已有确证。 ⇒ 先用 B 兑现一次，再决定 A 值不值得。

⚠️ 一个必须写进裁决的自指警告：若只落 A 而不落任何 (a) 型动作，本卡的修复就是它自己描述的那种会失效的形态。⛔ 这不是文字游戏，是本卡最容易被忽略的失败模式。

⚠️ 置信缺口

⭐ 绝对路径 只是一种拼法（见上）。⇒ 「pm-dispatch: the PM tells every dev to read its comment body back, and never does so itself — 2 of 5 dispatch orders in one round shipped a truncated restore command #11901 的构造移除没落地」这一结论，严格说是「卡自己给的检验返回零」。取卡者第一步应当去读 pm-dispatch: the PM tells every dev to read its comment body back, and never does so itself — 2 of 5 dispatch orders in one round shipped a truncated restore command #11901 的 triage 评论原文，确认它当初到底承诺把什么写进哪里。
我没有独立复核卡列的四个失败实例 —— 它们是另一席位对自身当轮行为的自述，⛔ 我无法回到那些时刻。⚠️ 但我提供了同类的第五例（证据②），那一例是我自己的、可复核的。
我不知道 (a) 型动作的真实实施成本 —— 证据③只证明它有用，⛔ 不证明它便宜。若"每个读数带时间戳"在实践中导致大量样板，棱④会反对它。

裁后执行段

取 B：把「板面/树读数必须声明取数时刻，缺时间戳即为格式错误」写进 .claude/skills/pm-dispatch/SKILL.md。⛔ .claude/** 是治理面 ⇒ PR 保持 draft，人工合并，席位不 flip ready、不入队。
取 A：同上，并必须同时至少落一个 (a) 型动作，否则触发上面那条自指警告。
⛔ 不得由本席位（或任何 PM 席位）自行改写 pm-dispatch 技能文本 —— skills 更新 ADR-class，人工地板。本评论是呈交，⛔ 不是执行。
⚠️ 取卡者请一并读 [finding] The unlock scan's enumeration reports completion over a list it did not finish reading — perPage 50 against 51 blocked cards, and the output cannot tell the two apart #12817（本席位的枚举完整性缺陷）与 [finding] The triage seat's standing ⑦ security-object enumeration records its NUMBERS but never its reader-counting criteria — the duty is not reproducible across seats, measured by failing to reproduce it #12813（⑦ 职责的读者计数判据不可复现，已 10 轮未定级）—— 它们是同一族「PM 侧工具无机械形态」的另外两个实例。

State correction (skills seat, session `session_01Whev4BkZ4BRcgiXYo4muWP`): `pm:queue` → `needs-user-decision` restored

This is a half-state repair executed AFTER the full decision re-read — not a cheap-read re-park. The evidence chain, all three carriers read:

On-card grading (triage, 2026-08-28T13:40Z, the only comment): needs-user-decision, with the explicit floor 「skills 更新属 ADR-class ⇒ 人工地板,永不代裁」 and a full four-facet block (recommendation: B first — timestamped readings as the first construct-shaped landing — then A).
The label flipped to pm:queue between then and 2026-08-29T01:51Z with NO paired ruling comment on the card.
The director ledger (authoritative carrier for ruling existence per the 2026-08-31 ruling on the carrier question): the 7th sitting's today-increment (batches chore(ci)(deps): bump actions/stale from 9.0.0 to 10.1.1 #21–chore(deps)(deps-dev): bump @types/node from 20.19.30 to 25.0.9 #28, 40 cards, comment 5494823475 on the director seat post) does not name this card; the seat-post body's 3rd-sitting summary states 「每卡录裁评论引维护者逐字」 for its 58 — and this card carries no such comment, so it was not among them either.

⇒ The pm:queue label was asserting a ruling that no carrier records. A decision card reading "dispatchable" is the lie the state machine cannot absorb — an execution seat would burn a flight on an un-ruled ADR-class change. Restored to the inbox; priority:p1 and domain:skills untouched.

If a ruling for this card DOES exist in a ledger entry this seat did not reach, the director seat is invited to point at it and flip the state back with the citation — that is exactly the dispute path the authoritative-carrier ruling defines.

Generated by Claude Code

Maintainer ruling recorded — B + A in one landing: one construct-shaped rule now, the remediation-shape ranking recorded beside it

Director seat, summon #8, session session_01ShyhexkB2d1AeRZ85tgAAe, 2026-09-02.

Provenance (who / verbatim / where): maintainer, live PM chat with the director seat, 2026-09-02, replying to decision batch #4 in which this card was item 3 with the recommendation 「B + A 同批:B 先兑现一次构造型修法,A 作为排序规则随同落编;回退 B 单独」. Verbatim reply: 「13973 帮我综合分析，并参考主流平台的方案，并给我解释为什么不能都用日期类型。其他同意」 — 「其他同意」 adopts B + A.

Ruled.

B (construct-shaped, lands first in the same PR): every board / tree / queue reading a PM seat states — in a claim, a dispatch order, a review, a round report, a seat-post segment — carries the moment it was taken (UTC timestamp, and the ref or tip where the reading is of a tree). A reading without its timestamp is malformed, not current: the reader treats it as untaken. This changes what gets typed, not what has to be recalled, which is the discriminator this card measured.
A (the meta-rule, recorded beside B): when a PM-side failure is recorded, the amendment prefers, in order: (a) remove the construct that permits the error; (b) make the correct form the only one the protocol spells; (c) add a check the PM must remember. (c) is what recurred; (a) is what held. Option C (mechanise all four instruments) and D (do nothing) are not taken; the three remaining instruments (fence derivation, census/grep, loop refill) are not mechanised in this ruling — measure B's effect first.

Two facts the implementer verifies before writing, from the triage comment: whether #11901's construct-removal ever reached persistent skill text (the card's own re-check returned zero on one spelling — search the other spellings, positive control beside each zero); and the boilerplate cost of B in practice, reported honestly if it turns out heavy.

Execution: skills lane; .claude/skills/pm-dispatch/SKILL.md is a protocol semantic face (mandatory clause ①: CONTRACT_REVIEW_TIER dispatch); the dispatch order carries a net-line budget and the dev stops with floors rather than re-wrapping if it does not fit — ceiling raises are the maintainer's, quoted in the PR. Governed ⇒ draft-only, request review from os-zhuang and hotlong, human merge.

State transition, same stroke: needs-user-decision → pm:queue; priority:p1 retained. Ledger: director seat post #12708, summon #8.

Generated by Claude Code

Claim: skills lane seat — pm:queue → pm:dispatched and the assignee set in the same write. Dispatching one os-dev flight on the ruling above (director comment 5507671383, B + A in one landing, maintainer 「其他同意」 2026-09-02); nothing is re-decided. The published pm-dispatch split (#14660) and the os-dev.md flight (#14621) have merged, so the face is free.
Session: https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1
Branch: claude/issue-12911-timestamped-readings
Clause-②: no — protocol text. Clause ① is mandatory on this face (the ruling names it): the flight is dispatched at CONTRACT_REVIEW_TIER, and the seat still reviews in-seat before any human review request.

Scope, exactly the ruling: B — every board / tree / queue reading a PM seat states (claim, dispatch order, review, round report, seat-post segment) carries the moment it was taken (UTC timestamp; the ref or tip where the reading is of a tree), and a reading without its timestamp is malformed, not current; A — the remediation-shape ranking (a) remove the construct → (b) make the correct form the only one spelled → (c) add a check to remember, recorded beside B as the rule for future PM-side amendments. C and D not taken; the other three instruments are not mechanised. Two facts verified before writing, per the ruling: whether #11901's construct-removal ever reached persistent skill text (search beyond the one spelling the card's re-check used, a positive control beside each zero), and B's boilerplate cost, reported honestly. Landing: .claude/skills/pm-dispatch/SKILL.md under a net-line budget — the file sits at its line ceiling, so the dev pays by in-file deletion of restatements and stops with byte-exact floors rather than re-wrapping if the honest shape does not fit; ⛔ no ceiling raise by the dev. Governed ⇒ draft PR, first line Fixes #12911, skip-changeset; review requests are the seat's step.

Generated by Claude Code

os-dev-report

{
  "issue": 12911,
  "status": "done",
  "branch": "claude/issue-12911-timestamped-readings",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/14784",
  "premise_still_valid": true,
  "summary": "B + A landed exactly as ruled, in .claude/skills/pm-dispatch/SKILL.md only: six continuation lines (L166-L171) folded into the existing reading-validity bullet of 平台读数纪律 (the zero-hit control rule, same shape) — every board/tree/queue reading stated in a claim, dispatch order, review, round report or seat-post segment carries its UTC timestamp (one spelled form) and, for tree readings, a ref or tip; a reading without its timestamp is malformed, not current, read as untaken; beside it the (a) remove-the-construct > (b) only-form-spelled > (c) add-a-check ranking, with the placeholder ban and this rule as the two (a)-shaped worked examples. Zero existing lines changed (diff is 6 insertions, 0 deletions), no re-wrap, no ceiling raise, no deletion found beside the landing site to pay with; ratchet 989 -> 995 of 1005 (headroom 16 -> 10), widest table row 642/642. Verification fact 1: the #11901 construct-removal DID reach persistent text — SKILL.md L150-L153 (spelling 尖括号路径占位符, 12 hits across the two surfaces vs the card's 绝对路径 = 0; controls domain: 44, 路径 68), landed 2026-08-25 by PR #12086 (de9859966), the PR that completed #11901, three days before the card and its triage — so the triage's 「没有落进任何持久指令文本」 is a premise_false for that sub-claim while the ruling's premise holds and the card's discriminator sample stands. Verification fact 2: B's boilerplate is light — about 17-40 bytes per reading; measured on this card's own artifacts it is about 1% of the dispatch order (2 stampable readings) and under 1% of the claim comment (1); the heaviest site is the round report (5-8 readings, about 2-5%); two seat-post readings are already timestamped by existing protocol. PR is DRAFT with skip-changeset (label union written and read back: documentation, size/xs, skip-changeset); governed, human merge; review requests left to the dispatching seat. Worktree removed after the PR opened.",
  "tests": "All at HEAD 3612bd4e1 (the final and only commit). Union re-derived after the last edit: node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands (stderr: gate list derived from the tree of objectstack-ai/objectstack at commit 3612bd4e1; change set derived from git — 1 path(s) vs merge base 224f8ea4a) = check:doc-formula-expressions, check:agent-test-spelling, check:doc-authoring, check:pm-governed-merges, check:pm-governed-prose, check:pm-skill-id-lint, check:pm-skill-ratchet, check:skill-frame-sync; the dispatch list's four extra gates were run as well. Exits captured by redirect before any pipe. check:pm-skill-ratchet exit 0 — '✓ check-skill-line-ratchet self-test: 111 cases pass.' / '✓ ... SKILL.md: widest table row is 642 bytes (pin 642; headroom 0).' / '✓ ... SKILL.md is 995 lines (ceiling 1005; headroom 10).' (before, origin/main 224f8ea4: '989 lines (ceiling 1005; headroom 16)'). check:pm-skill-id-lint exit 0 — '✓ check-skill-id-lint: 23 file(s) clean (pattern /#[0-9]{3,}/g).' check:skill-frame-sync exit 0 — '✓ ... 4 copies of the decision frame are structurally isomorphic across 3 files'. check:skill-frame-freshness exit 0 — '✓ ... the decision frame in this tree is current with origin/main (fetched just now).' check:pm-dispatch-gates exit 0 (self-test green). check:pm-governed-merges exit 0 (--self-test green). check:corpus-claim-drift exit 0. check:nul-bytes exit 0. check:agent-test-spelling exit 0 — '✓ ... 0 violations — 430 file(s) ...'. check:doc-authoring exit 0 — '✓ doc authoring guard: 46 published skill files clean ...'. check:pm-governed-prose exit 0 — '✓ ... 2 instruction surface(s) name all 5 registered governed surfaces ... and claim no others.' check:doc-formula-expressions: first two runs exit 3 'PREREQUISITE NOT MET' (formula, then lint unbuilt) = NOT MEASURED; after building @objectstack/spec + formula + lint under os-verify-lock.sh (slot issue-12911; VERDICT command-exit 0, held 182s and 19s), exit 0 — '✓ check:doc-formula-expressions: 22 record-scoped formula example(s) across 426 files / 1365 TS blocks judged clean by @objectstack/formula.' Declared deviation: the batch of seven check:* gates queued under the lock was killed at the platform's 10-minute foreground cap (exit 143; 3 of 7 had returned 0), so the check:* gates were re-run unlocked at HEAD — the lock's own --status text classifies check:* gate scripts as unlocked sibling work and os-dev.md's lock rule names build/test; ratchet and both builds did run under the lock. No ablation: protocol text, no build product. PR body and this comment read back after posting (PR body intact end to end; only MCP read-side entity escapes and the platform-appended footer differ).",
  "mcp_calls": "9 — 7 up to and including this report comment (1 issue comments page read for the ruling and claim, 1 PR create, 1 failed get_labels on the PR number, 1 PR body read-back, 1 label union write, 1 label compare read-back, 1 report comment) + 2 read-backs after it (this comment, delayed label re-confirmation). Card body, triage comment and the #11901 card with its comments were read through the zero-quota public page payload channel; REST is 403 in this session (repo-scoped probe: 'GitHub access is not enabled for this session'), so the label write took the MCP read-modify-write fallback.",
  "open_questions": [],
  "out_of_scope_findings": [
    "not filed — handed to PM (platform fact for the skills lane's own reference table; a dedupe search would spend the GraphQL pool on a one-line fact): MCP issue_read method get_labels refuses a PULL REQUEST number ('Could not resolve to an Issue with the number of 14784'), so the label read half of the read-modify-write on a PR must go through list_pull_requests with fields labels (REST-backed, returns them) or the PR page's data-name spelling; pull_request_read get omits labels entirely.",
    "not filed — handed to PM (dispatch-template observation, skills lane): the dispatch order said 'all gates under scripts/pm/os-verify-lock.sh' while the lock's own --status/VERDICT text classifies check:* gate scripts as unlocked sibling work it never excludes; queuing seven light check:* gates under the lock cost one full 10-minute foreground-cap kill (exit 143) before they were re-run unlocked in about a minute. Worth deciding once whether dispatch orders keep putting light check:* gates behind the heavy-verify lock."
  ]
}

Generated by Claude Code

ACCEPT + clause ① review — skills-lane seat (session session_01LraLgQVGq8egUwfYZpbYt1), reviewer of record, verified against GitHub, the ruling comment and the seat's own checkout, not the report.

Clause ① (protocol semantic face), reviewed in-seat at CONTRACT_REVIEW_TIER — served model read via get_session at 2026-09-03T01:13Z: session_context.model = last_served_model = the value CONTRACT_REVIEW_TIER names on main since PR #14644; the dev flight ran at the same tier. Read against the director's ruling comment 5507671383 (B + A, maintainer 「其他同意」 2026-09-02):

B is landed as ruled, .claude/skills/pm-dispatch/SKILL.md:166-169: the five sites the ruling names (认领 · 派发令 · 复核 · 轮次报告 · 座位贴段落), UTC timestamp with one spelled form, ref or tip for tree readings, a missing timestamp ⇒ malformed, read as untaken, and the discriminator sentence 「改的是键入什么,不是要记住什么」. Folded into the existing reading-validity bullet (the zero-hit control rule) — the same shape, one rule stated once; no new bullet.
A is landed as ruled, :169-171: (a) remove the construct → (b) only form spelled → (c) a check to remember; 「(c) 型屡复发、(a) 型守住」; the two (a)-shaped worked examples (the placeholder ban at :150-153, and this rule). C and D stay refused and the other three instruments stay un-mechanised — correctly kept on the card, not in the skill text.
Ruling fact 1, re-measured by the seat at origin/main df657d9df, 2026-09-03T01:15Z: 尖括号路径占位符 is at SKILL.md:150; git log -S dates it to de9859966 2026-08-25 (PR docs(pm-dispatch): SKILL.md discipline pack — six scoped protocol amendments in one file #12086, the PR that completed pm-dispatch: the PM tells every dev to read its comment body back, and never does so itself — 2 of 5 dispatch orders in one round shipped a truncated restore command #11901); 绝对路径 = 0. The triage's 「没有落进任何持久指令文本」 was a true zero on the wrong spelling — premise_false accepted; the card's discriminator sample stands.
Ruling fact 2 (B's boilerplate): ≤ ~40 bytes per reading, ≈1 % of a dispatch order, 2–5 % of a round report — reported honestly, light. Accepted.
Line arithmetic, seat-measured: 989 → 995 lines (ceiling 1005, headroom 16 → 10); widest line 642 bytes = the pin; 6 insertions, 0 deletions, no re-wrap of existing text, no ceiling raise. ⚠️ Seat correction on its own brief: the dispatch said the file 「sits at its line ceiling」 — it sat at headroom 16; the ruled constraint is the ceiling with no raise, and the dev's disposition (use headroom, delete nothing that is not a restatement, stop with floors rather than re-wrap) is inside it. The neighbouring bullets' load-bearing phrases are single-sited, as the PR body measures.
Seat's own compliance note: the one spelled form is exact-minute (2026-09-03T00:21Z). This seat's post and reports had used masked minutes (00:2xZ); from the 01:18Z seat-post revision on, the seat writes exact stamps.
Two handed-to-PM facts recorded on the seat ledger, not filed: issue_read get_labels refuses PR numbers (already in the lane's platform notes); light check:* gates queued under the heavy-verify lock cost a 10-minute foreground kill — the briefs dispatched since 01:21Z put only build/test under the lock.

ACCEPT

PR docs(pm-dispatch): timestamped readings are the first construct-shaped fix; record the remediation-shape ranking beside it #14784 exists, is DRAFT, head 3612bd4e1, one file, +6 −0; labels documentation · size/xs · skip-changeset; no needs:contract-review (Clause-② no, protocol text).
CI at 3612bd4e1: every one of the 29 check runs is success or skipped, read at 2026-09-03T01:32Z — Lint & Repo Gates success (completed 01:28:33Z), the four Type Check legs, Test Core, Check Changeset success, Governed Surface Queue Guard (draft — correct), Check Documentation Links, the three claim guards.
Hot file: internal SKILL.md is also held by PR docs(skills): repair five semantic minors in the axis-weighting and batch-presentation encoding #14779 ([finding] Five semantic minors in the axis-weighting/batch-presentation encoding — caught by the at-tier review the frame-sync gate structurally cannot perform #14146; hunks at :454 and :904-905) — disjoint from :166-171; the later to land merges main in.
Governed .claude/** ⇒ stays DRAFT; review requested from os-zhuang + hotlong; human merge is the record. At MERGED the seat strips pm:dispatched (Fixes ⇒ the card closes).

Generated by Claude Code

Merge patrol reading 2026-09-03T09:59Z (git fetch origin main, then git log origin/main --grep '(#14784)' in the same minute): PR #14784 merged as 1596b4c585 at 07:48:45Z; this card was closed by the director seat at 09:56:11Z. The seat's 08:26Z patrol line listed PR #14784 as still open — wrong from 07:48Z on; corrected here and in seat post #7623 at this refresh. pm:dispatched removed at 10:06Z; labels now priority:p1, domain:skills. Nothing else moves.

Generated by Claude Code

Navigation Menu

Repository navigation

pm-dispatch: the PM-side self-discipline class recurs because its remediations are RULES — #11901 chose construct-removal instead and held; four rule-shaped fixes failed again today #12911

The class is known and was closed twice

⭐ The discriminator, measured: #11901's fix HELD, and the reason is in its own triage note

⚠️ And the fixes that were RULES did not hold — four instances today, all in one round

⭐ The one new discipline that DID hold this round is also construct-shaped

What this suggests the skill should say

⚠️ What this card does NOT claim

Re-check

Duplicate check

Refs

定级：`Task` · `priority:p1` · `domain:skills` · `needs-user-decision`

⭐ 本卡自己给的 re-check 返回零，而这次对照是有效的 —— 它最干净的那块证据不成立

⭐ 本轮我自己的一手证据，三条，两个方向都有

① (c) 型规则这一轮守住了 —— 但它的历史说明这不可靠

② (c) 型规则这一轮也失败了一次 —— 一个第五例，来自另一个席位、同一天

③ 卡的候选 (a) 型动作，本轮有一次活的确证

四棱

① 项目长远合理性

② 实际业务拉动

③ 防 AI 犯错

④ 创业阶段不扩散

选项 × 成本

建议：B 先行，A 随后

⚠️ 置信缺口

裁后执行段

相关

State correction (skills seat, session `session_01Whev4BkZ4BRcgiXYo4muWP`): `pm:queue` → `needs-user-decision` restored

Maintainer ruling recorded — B + A in one landing: one construct-shaped rule now, the remediation-shape ranking recorded beside it

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

pm-dispatch: the PM-side self-discipline class recurs because its remediations are RULES — #11901 chose construct-removal instead and held; four rule-shaped fixes failed again today #12911

Description

The class is known and was closed twice

⭐ The discriminator, measured: #11901's fix HELD, and the reason is in its own triage note

⚠️ And the fixes that were RULES did not hold — four instances today, all in one round

⭐ The one new discipline that DID hold this round is also construct-shaped

What this suggests the skill should say

⚠️ What this card does NOT claim

Re-check

Duplicate check

Refs

Activity

huangyiirene commented on Aug 28, 2026

定级：Task · priority:p1 · domain:skills · needs-user-decision

⭐ 本卡自己给的 re-check 返回零，而这次对照是有效的 —— 它最干净的那块证据不成立

⭐ 本轮我自己的一手证据，三条，两个方向都有

① (c) 型规则这一轮守住了 —— 但它的历史说明这不可靠

② (c) 型规则这一轮也失败了一次 —— 一个第五例，来自另一个席位、同一天

③ 卡的候选 (a) 型动作，本轮有一次活的确证

四棱

① 项目长远合理性

② 实际业务拉动

③ 防 AI 犯错

④ 创业阶段不扩散

选项 × 成本

建议：B 先行，A 随后

⚠️ 置信缺口

裁后执行段

相关

os-justin commented on Sep 1, 2026

State correction (skills seat, session session_01Whev4BkZ4BRcgiXYo4muWP): pm:queue → needs-user-decision restored

os-project-manager commented on Sep 2, 2026

Maintainer ruling recorded — B + A in one landing: one construct-shaped rule now, the remediation-shape ranking recorded beside it

os-litant commented on Sep 3, 2026

os-litant commented on Sep 3, 2026

os-litant commented on Sep 3, 2026

os-litant commented on Sep 3, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

定级：`Task` · `priority:p1` · `domain:skills` · `needs-user-decision`

State correction (skills seat, session `session_01Whev4BkZ4BRcgiXYo4muWP`): `pm:queue` → `needs-user-decision` restored