skills(pm-dispatch): pm:awaiting-maintainer owes a Maintainer-action: entry line, read by H55; 具名不升级类 gains the four bookkeeping classes

[os-justin]

Fixes #17017

Maintainer, 2026-09-09, verbatim and untranslated — the two instructions that opened the card. The first is the authority for Half B, the second for Half A:

「5986 这种任务不影响我们的产品质量，不应该让人类决策，项目经理自己当场决定就可以，而且这个应该开技能卡纠正」

「这个不应该是决策卡吗？哪个流程出了问题？」

Governed surface (.claude/**) ⇒ draft, human merge; skip-changeset (nothing here publishes: .claude/** and scripts/pm/** ship in no package files[]).

Half A — pm:awaiting-maintainer gains a machine carrier

The state had a semantic contract (决定已做, one action a seat cannot perform) and no machine carrier: pm:blocked owes Blocked-by:, pm:on-hold owes Restart-when:, and this state owed nothing, so a card waiting for a JUDGEMENT was indistinguishable from one waiting for a HAND.

• references/state-machine.md §pm:awaiting-maintainer 与 pm:retriage now specifies the entry line (the two placeholders are spelled as words in this PR body because the platform mutates angle-bracket fragments; the file carries them literally, in the same shape as Restart-when: closed OWNER/REPO#N):
  • awaiting 入态带 Maintainer-action: ACTION — done when EVIDENCE,双通道行锚定。 — ACTION = 席位做不了的动作, EVIDENCE = 可查证据; body OR comment; line-anchored and case-sensitive like the two sibling lines.
  • 写不出即决定待做;本态无机器出口:维护者完成后由总监席核验摘标 + 证据评论同笔。 — the classification test the card asked for (「裁完之后」 is the definition of needs-user-decision), folded onto the existing exit line.
• SKILL.md 状态模型 row (pointer form, table row extended in place, 234 B ≤ the 342 B table pin): …;入态恒带行首 Maintainer-action: 行(细则见 references/state-machine.md),无行即半态.
• references/core-rules.md twin (in place): pm:awaiting-maintainer 只剩站外人工动作,恒带 Maintainer-action: 行;pm:blocking ⛔ 不手工挂。
• scripts/pm/check-half-states.mjs — H55, report-only, band state: an open card carrying the label whose body and thread carry no complete Maintainer-action: line. The line is read by the shared decorated-directive reader (directiveValues), so it is line-anchored, case-sensitive, decoration-tolerant, and a mid-sentence mention is prose. Both halves are required: a value without its done when half fires as "names an act but never its completion". The thread is bought only for a judged, body-clean card (needsMaintainerActionComments, H9's gate one field over) on the shared comment cache; three input states for the thread (unconsulted / unreadable / read), the unreadable one fires and says so. Remedy text: write the line if the act exists; if the line cannot be written because the card is asking for a RULING, the state is needs-user-decision (决定待做) and the entry was misfiled; the 台账 reader is the director seat's 职责三 (维护者动作台账). ⛔ Not a machine exit — the state still has none by construction; the row judges the ENTRY only.
  • Reconciliations, pinned on the source (H54's discipline): the H11 important-parked note now reads "has no machine exit BY CONSTRUCTION … the state now has an ENTRY predicate … and STILL no machine exit", and its awaiting remedy tells the re-ask to read the card's Maintainer-action: line first (still never prescribing a Restart-when: re-check — the existing pin at :19486 holds). H25's deferral note ("any requirement that the card NAME the awaited action … a protocol face the state-model row has not been written for yet") now says that face exists and H55 reads it; H25 stays the EXCLUSIVITY half, H55 is the ENTRY half, neither widened into the other.

Population scoping (ruling 2) — the landing day must not be red on the healthy case

Entries are split at a pinned instant, MAINTAINER_ACTION_LINE_SINCE = '2026-09-10T00:00:00Z' (the UTC day after the rule was written, in the file's own pinned-date idiom): a card last touched BEFORE it is a legacy entry and renders as ONE summary census clause — Awaiting-maintainer entries (H55): N open pm:awaiting-maintainer card(s) last touched before … — COUNTED here and filed as no row …, oldest #N (last touched YYYY-MM-DD); … — never as rows; a card touched ON/AFTER it is judged and can be a row.

The split reads updated_at, not the labeled event — the trade, stated. The entry instant is in no payload the sweep already holds (assumption C, measured: the loop holds the label pages' issue payloads, and the only event reader, H35, consumes the repo-wide stream filtered to gate-semantic labels — pm:awaiting-maintainer is not one), and this file's standing design makes no per-card timeline fetch. updated_at is the cheapest honest proxy and its error runs one way: updated_at is never earlier than the labelling, so a card last touched before the pin certainly entered before it (the census can never hide a post-pin entry); a legacy entry TOUCHED after the pin lands among the judged and can render as a row — an over-read costing one cheap idempotent act (write the line, or reclassify) on a card someone is already handling. An unreadable updated_at is judged, never counted as legacy.

Measurement — the live board at authoring (2026-09-09T17:4xZ; REST list endpoint + per-card /events; 0 MCP calls)

repo	open pm:awaiting-maintainer	cards (entry = last labeled event → last touch)
objectstack	5	#16737 (09-08T14:01 → 09-09T15:31, +25 h), #15233 (09-09T10:29 → same), #17129 (09-09T11:42 → same), #16804 (09-09T11:44 → 13:55, +2 h; also carries needs-user-decision — H25 already fires), #17151 (09-09T13:21 → same)
objectui	2	#5828 (08-29T06:45 → 08-29T06:46), #6596 (09-09T05:34 → same)
hotcrm	0	—
total	7	5 of 7 have updated_at within 3 s of the labeled event; 2 were touched after entering

Premise shift, reported rather than pressed: the card measured 55 (37 / 16 / 2) on 2026-09-09 and 55 → 55 across ~14 h. By dispatch the population was 7: the repo-wide events stream (H35's channel, read 30 pages per repo back to 2026-09-09T03:38Z / 2026-09-08T17:07Z) shows a reclassification pass that morning — objectui 15 unlabeled pm:awaiting-maintainer at 02:56–03:04Z, objectstack's named specimens at 02:29–05:12Z (#13636, #16500, #12675, #12884, #15406 unlabeled and closed; #12034 back to pm:queue). The three defects the card names (no carrier, H8 blind to the chosen state, the sweep reading shape and time only) are unchanged by the count, so the design stands; what changes is the landing-day shape — with the pin above, all 7 render as the census clause on the first sweep, not as rows. The search endpoint answers 403 from this container, so the control reading is the events stream plus a list-endpoint control on a populous label (pm:dispatched = 26 open).

Half B — the floor of what may reach the maintainer

具名不升级类 (SKILL.md 〈升级与决策〉) gains the four bookkeeping classes, each bounded by 「无产品可见行为变化」 so none reads as widening; the 人工地板 lines (SKILL.md :390–:392 after the shift; core-rules :95–:96) are byte-untouched and remain the exclusion set.

• SKILL.md, after 具名不升级类(立即行动):…:
  • - 四类记账事同属不升级类,恒以无产品可见行为变化为界:去重与卡片合并;台账与记账整理; (120 B)
  • - 纯文档措辞更正(无契约声明改动);门禁内部参数与扫描器盲区修复(加强,非削弱)。 (110 B)
• core-rules.md twin, after 具名不升级类:…: - 记账事不升级(无产品可见行为变化):去重并卡、台账整理、纯文档措辞、门禁盲区加强。 (119 B; the compressed spelling the 人读摘要 file uses for every twin).

Authority: the maintainer's first sentence above — cited here, ⛔ not inside the skill text (the skill states rules; check:pm-skill-id-lint keeps citations out).

Line budget — every file at its ratchet ceiling, headroom 0, line-neutral

pnpm check:pm-skill-ratchet on af85c612: SKILL.md is 812 lines (ceiling 812; headroom 0) · widest table row is 342 bytes (pin 342) · state-machine.md is 42 lines (ceiling 42) · core-rules.md is 151 lines (ceiling 151). No re-wrap, no ceiling raise, check-skill-line-ratchet.mjs untouched.

file	ceiling	before → after	paid by (deleted line, verbatim, and why it is not a loss)
SKILL.md	812	812 → 812	(+0) row :111 extended in place. (+2) the two Half B lines. (−1) - 机会主义重启条件必须点名触发文件。 (:131) — the rule is stated in references/state-machine.md :18–:19 (Restart-touch: mechanics) and SKILL.md :124 already points there. (−1) - 先过升级门槛:明显的问题直接修,大多数感觉像决定的不是决定。 (:711) — restated by the three lines under it (:712–:713 the two thresholds and 其余归 PM 裁量; :714 具名不升级类(立即行动)); introduced by the rules-only rewrite, not a ruled clause.
references/state-machine.md	42	42 → 42	(+1) the entry-line spec. (0) :37 rewritten in place to fold the 写不出即决定待做 test onto the exit line (「维护者完成该操作后,由项目总监席核验」 compressed to 「维护者完成后由总监席核验」, meaning intact). (−1) - 标签须五仓存在,首次应用时创建。 (:42) — stated in references/lanes/hotcrm.md :46 (「pm 状态机词表五仓统一,幂等创建在 scripts/pm/ensure-pm-labels.sh」) and mechanised by that script (it creates pm:awaiting-maintainer idempotently, :212).
references/core-rules.md	151	151 → 151	(+1) the Half B twin. (0) :36 rewritten in place — the carrier joins the definition; pm:blocking keeps its ⛔ and loses only the WHY (「是反向索引推导的缓存」), which SKILL.md :112 keeps. (−1) - `pm:dispatched` 与摘 `pm:queue` 恒在同一次标签写入里成对落地。 (:30) — an intra-file duplicate of :109 (「认领原子对:一次标签写入完成认领与状态对调…」) and :38 (「转换恒一笔 replace」); SKILL.md :107 and :466 keep the rule, so the twin obligation (SKILL.md :44) is untouched.
references/lanes/director.md	72	72 → 72	untouched (assumption B, below).

Mechanism assumptions (派发词 — tested before acting)

• A (pinned lines). check-governed-prose pins the SKILL.md region between **ACCEPT 之后的路径分叉 and 本段只适用本循环派发的 dev PR; check-half-states pins RELEASE_RULE_LINE (SKILL.md's 释放是显式动作 line) and the retriage answer line; check-skill-frame-sync pins the four-axis frame; check-clause2-carriers / check-widening-tells pin references/contract-review.md and the claim spelling. None of the deleted or edited lines is in any pinned set — all six gates are green below. Held.
• B (director.md). :52 「pm:awaiting-maintainer 跨仓标签是待人工动作的唯一机器载体」 was read as the label being the unique carrier of the STATE (against seat-post prose, :55 「标签即状态」), while the new line is the label's machine half specified in state-machine.md, which :56 already points at (「出口条款见 ../state-machine.md」); :54 「台账与出口归本席」 is what the H55 remedy names. No sentence contradicts the line ⇒ left untouched, per the ruling. Noted in Acceptance notes as the one reading a reviewer may want to overturn.
• C (entry date from what the sweep holds). Falsified — measured above; ruling 2's stated fallback (updated_at) taken, trade stated in the header, the row text and the summary clause.
• D (line numbers). The card's SKILL.md:110/:126 are :111/:127 on a9550bf8; check-half-states.mjs:2298 is :2302 (the note) / :2307 (the remedy); 具名不升级类 :714 and 人工地板 :391–:393 as cited. Worked from the tree.

Gates (local, on af85c612; the list derived by node scripts/pm/dispatch-gates.mjs on a9550bf8 — 45 families — and re-derived from the real changeset on the final tree: identical 45; --ran reconciliation below)

Every exit captured before any pipe (cmd > log 2>&1; EXIT=$?).

family	verdict
node scripts/pm/check-half-states.mjs --self-test (check:pm-half-states)	✓ check-half-states self-test: 3103 cases pass. (baseline on a9550bf8: 3019; +84 H55 cases)
pnpm check:pm-skill-ratchet	exit 0 — the four ceiling lines quoted above
pnpm check:pm-skill-id-lint	✓ check-skill-id-lint: 26 file(s) clean (pattern /#[0-9]{3,}/g).
pnpm check:skill-frame-sync / check:skill-frame-freshness	2 copies of the decision frame are structurally isomorphic across 2 files / the frame itself is unchanged
pnpm check:pm-governed-prose	2 instruction surface(s) name all 5 registered governed surfaces … and claim no others.
pnpm check:pm-clause2-carriers / check:pm-widening-tells (self-tests)	exit 0 / exit 0
pnpm check:nul-bytes + the control-byte grep over the four files	no raw ASCII control bytes / empty
pnpm --filter @objectstack/lint run check:doc-formula-expressions	first run exit 3 = PREREQUISITE NOT MET (unbuilt @objectstack/formula) — NOT a measurement; rebuilt formula + lint under os-verify-lock.sh (VERDICT command-exit 0 · held the lock 2s) and rerun: exit 0
the other 16 pnpm check:* families (agent-test-spelling, bash32-floor, changeset-gate-self-tests, cli-command-ids, cross-package-test-inputs, doc-authoring, driver-memory-census, entry-guard, parse-guard, partof-closing-keyword, pm-governed-merges, pnpm-filter-targets, ratchet-remedy-authority, refd-timer-probe, watch-hint-literal, pm-half-states)	all exit 0
the 22 direct node families (reference-carrier-shape ×2, changeset-no-major ×2, ci-filter-parity, closing-keyword-parity ×2, comment-mask-corpus, declaration-mirrors ×2, scripts-symbol-anchors ×2 — 3094 anchors across 238 scripts resolve, self-test-wired ×2, self-test-workflow-commands ×2, whole-set-label-write ×2, bare-root-worklist --self-test — OK … none stale, none missing, none contradicted, governed-queue-guard --self-test, sweep-closed-cards --self-test, report-test-timings --self-test)	all exit 0
pnpm check:pm-dispatch-gates (> 600 s foreground cap; run detached, tail --pid waited)	run detached on the same tree (setsid nohup, waited with tail --pid): ✓ dispatch-gates self-test: 1624 cases pass., exit 0. node scripts/pm/dispatch-gates.mjs --ran RAN_FILE on the final tree: 45 derived famil(ies) accounted for — 45 run, 0 NOT-MEASURED
CI-measured only (not runnable here, per the derivation): check-governed-queue-guard.mjs (workflow payload), check-partof-closing-keyword.mjs (env), shard attestation, test completeness, the --format=markdown --provenance sweep run	NOT MEASURED locally, by construction; CI

No package touched ⇒ no build closure owed; no changeset (skip-changeset: .claude/** and scripts/pm/** publish nothing — no gate demanded one; check-changeset-no-major exit 0).

Reverse verification (one-time ablation, committed state, restore proven by bytes)

Committed first (af85c612), then mutated under a trap restore EXIT INT TERM with absolute paths: the first statement of h55AwaitingWithoutMaintainerAction was made return null (marker ABLATION-17017-MARKER). Landing proven on disk before reading any result: marker count 0 → 1 on the anchor line, blob f015273c… ≠ HEAD blob 0ec90893…. Expected direction: RED (the positive control and every row-text case), the census/registry/summary cases unaffected — observed exactly that: ✗ check-half-states self-test: 20 of 3103 case(s) failed., the 20 being the six H55 fires: cases, the three thread-state cases, the undated case and the ten H55 row: sentence cases; every neighbour (H9, H11, H25, H54, summary clauses, band registry) stayed green. Restore leg: git checkout HEAD -- ABSOLUTE_PATH (never a bare checkout — the index is the mutation), proven by git diff HEAD --stat empty, git status --porcelain empty, git hash-object = 0ec90893d36c84e19c8721b755252781b8b8b805 = git rev-parse HEAD:scripts/pm/check-half-states.mjs, marker count back to 0; the restored run reads ✓ check-half-states self-test: 3103 cases pass. again. No dist/ is involved (the sweep runs from source), so no dist preflight applies; no permanent test file was left behind.

Acceptance notes

• The bounded reclassification worklist is a LATER, separate act — named here, ⛔ not filed (ruling 2): the 7 live carriers above, each read per-thread and either given its Maintainer-action: line or moved to needs-user-decision (决定待做) / closed. objectstack dev has no https mode, so the advertised OAuth path for interactive MCP clients cannot be tried against a local dev server #16804 already carries both needs-user-decision and pm:awaiting-maintainer (H25 fires on it today) — first in that worklist. No live card was touched by this PR.
• director.md :52 reading (assumption B): left untouched on the reading above. If the reviewer reads 「唯一机器载体」 as contradicted by a second machine-read carrier, the in-place edit is …唯一机器载体,其 Maintainer-action: 行即台账行。 — it measures 122 B, 2 over the cap, so it would cost a fold elsewhere in the 72/72 file; not done here.
• done when required — H55 reads "both parts" literally: a Maintainer-action: value without a done when half fires as half a line (its own sentence). The dash spelling is not the contract (—, -, -- all carry it); the KEY is byte-exact.
• The EVIDENCE placeholder reads 可查证据, not 机器可查证据, in the state-machine line: 6 bytes were needed to fit the 120-byte cap; "evidence a reader can check" is stated in the H55 row text.
• noted, not filed — 承接者:无: check-half-states.mjs's directiveValues JSDoc named only two keys; widened to three (comment only). noted, not filed — 承接者: the reclassification worklist above: the awaiting label's ensure-pm-labels.sh description (:212) still says "no machine exit" — true, and now incomplete (no mention of the entry line); a description string, no reader.
• GET /search/issues answers 403 from this container (so did MCP-free dedup rely on the events stream + list endpoint); Maintainer-action had 0 hits on origin/main at dispatch and this PR is its first spelling.

维护者速读(草稿)

改了什么

给 pm:awaiting-maintainer(等您亲手做一件事的状态)加了一条必填的机器可读行:Maintainer-action: 要做的那件事 — done when 怎么验证做完了。写不出这一行的卡,说明它等的不是您的手而是您的裁决,应当放进决策箱(needs-user-decision)。巡查脚本新增 H55 报"进了这个状态却没写这行"的卡(只报告不改标签);今天之前已在状态里的卡只汇总成一行计数,不逐张报红。另外,四类纯记账事(去重合卡、台账整理、纯文档措辞更正、门禁内部参数与盲区加强)明确列为 PM 当场自决、不上报您。

为什么改

您 09-09 的两句话:5986 那种任务不该让人决策、该开技能卡纠正;以及"这不应该是决策卡吗?哪个流程出了问题?"。溯源结果是:这个状态只有定义没有载体,谁都能挂、没人能证伪,H8 只查有没有转状态、不查转对没有,巡查只看形状和时间不看含义。卡上当天量到 55 张里约 49 张挂错;派发时板上已剩 7 张(上午有一轮改判),缺陷本身没变。

风险与代价(含回滚)

三个技能文件都在行数上限,每加一行删一行(删的都是别处已有的重复表述,PR 正文逐行列出)。H55 判"新旧"用的是卡片 updated_at 而不是打标签时刻(不做逐卡 timeline 请求),代价是:老卡在 09-10 之后被碰过就会被当新卡报一次,补一行即消;绝不会漏报新卡。回滚 = revert 本 PR,无数据、无标签、无 workflow 变化。

席位意见

(留空)

你要做的

读完后人工合并本 PR(受管面,席位不合)。合并后如同意,由席位另开一张有界的改判工作单处理现存 7 张卡;本 PR 不动任何在板卡片。

---

Generated by Claude Code

[os-justin]

维护者速读

改了什么 — 给「等您亲手做一件事」的状态(pm:awaiting-maintainer)加一条必填的机器可读行:Maintainer-action: 要做的那件事 — done when 怎么验证做完了。写不出这一行的卡,等的是您的裁决而不是您的手,应进决策箱。巡查脚本新增 H55,只报告不改标签:今天之前已在该状态的卡只汇总成一行计数(现存 7 张),不逐张报红;之后进入的卡没写这行才报。另外,四类纯记账事(去重合卡、台账整理、纯文档措辞更正、门禁内部参数与盲区加强)明确列为 PM 当场自决、不上报您,以「无产品可见行为变化」为界,人工地板清单一字未动。

为什么改 — 您 09-09 的两句话:「5986 这种任务不影响我们的产品质量,不应该让人类决策,项目经理自己当场决定就可以,而且这个应该开技能卡纠正」和「这个不应该是决策卡吗?哪个流程出了问题?」。溯源结果:这个状态只有定义没有载体,谁都能挂、没人能证伪;H8 只查有没有转状态、不查转对没有;巡查只看形状和时间不看含义。卡上当天量到 55 张里约 49 张挂错;派发时板上已剩 7 张(上午有一轮改判),缺陷本身没变。

风险与代价(含回滚) — 三个技能文件都在行数上限,每加一行删一行,删的四行在别处都已成文(本席逐条核对;其中 SKILL.md「先过升级门槛:明显的问题直接修,大多数感觉像决定的不是决定」这一行是判断而非纯重复,请您过目)。H55 判新旧用卡片 updated_at 而不是打标签时刻,代价是老卡在 09-10 之后被碰过会被报一次,补一行即消,绝不漏报新卡。纯内部 PM 工具与技能文本,不发布、不进产品、无 changeset。回滚 = revert 本 PR,无数据、无标签、无 workflow 变化。

席位意见 — 本席以契约复审档复核通过:七条裁决逐条对分支核实,自测 3103 例全过,CI 31 项全绿(含线上巡查实跑),合并无冲突。建议合并;合并后本席另开一张有界的改判工作单处理现存 7 张卡。

你要做的 — 受管面走人工合并:看一眼新增那行的写法与被删的那四行,认可即批准并合并本 PR。是否合并?

---

Generated by Claude Code