[finding] `pm:blocked` 的封锁被写成散文而非 `Blocked-by: #N`，解锁扫描读不到 —— 实测三张卡的上游早已关闭却仍挂着，其中一张是 p1 + security

[os-project-manager]

在一次全车道状态稽核中测出（49 张 open domain:cli 卡）。⛔ 不是推断，三个实例都已逐条核过并当场修正。

缺陷

SKILL.md:109 把 pm:blocked 定义成一对：

pm:blocked + 正文行 Blocked-by: #N；等上游：选择期跳过，#N 关闭时由解锁扫描放回

解锁靠的是那一行 Blocked-by: #N。而实际发生的是：席位把封锁写成散文，写得清清楚楚、人一看就懂，⛔ 但机器读不到。于是 #N 关闭时没有任何东西把卡放回来，它就一直挂在 pm:blocked 里。

三个实例，全部实测，全部已修正

卡	封锁写成	上游现状	挂了多久
#15942 ⚠️ p1 + security	「State: needs-user-decision → pm:blocked (blocker: #16293) … Unblocks to pm:queue when #16293 merges」	#16293 已 closed（completed）	自 09-06
#16674 p2	「解锁条件：PR #16628 合并（或其分支释放）」	PR #16628 已 closed	自 09-08
#16545 p2	「解除条件（唯一）：PR #16380 合并或关闭」	PR #16380 已 closed	自 09-07

三张的解锁条件都早已满足。已由本席逐张一笔 replace 放回 pm:queue（⛔ 定级、kind、lane 一字未动）。

⚠️ #15942 是这里最该被看见的一张：p1 + security，主题是 MCP run_action 从不强制 ai.requiresConfirmation —— 一个「承诺了一个不存在的停顿」的安全闸门。它的裁定（决策批次 #54，「同意」采纳 option A）在 09-06 就下了，spec 半边 #16293 也已完成，而执行半边就这样静默停了四天。

⭐ 最能说明问题的一条：有人已经踩过，并且写下了正确的教训，却仍然只写散文

#16545 自己的评论里逐字写着：

「一条只活在评论里的持有关系，对下一个按标签取卡的席位是不存在的。」

那位席位当时正是因为这个原因补上了 pm:blocked 标签。⇒ 教训是对的、动作是对的，但仍然没有写 Blocked-by: 行 —— 所以卡从"标签缺失"变成了"标签有了但解锁不了"，⛔ 问题只是换了个形状。

⇒ 这不是某个席位粗心。散文是人类最自然的表达，而机器可读的那行需要被刻意记起 —— 只靠纪律不会稳定成立。

⛔ 本卡不主张的

• ⛔ 不主张这三张是仅有的实例。本次只稽核了 domain:cli 的 49 张，其中 16 张 pm:blocked；其它车道未查。
• ⛔ 不主张 check-half-states.mjs 有缺陷 —— 恰恰相反，它的 H4（Blocked-by: 在 body 与 comment 两个通道皆无）与 H19（Blocked-by: 目标已 CLOSED）正是为这个类写的，且 :118 明写该行「may be DECORATED in either channel」。模型是对的。
• ⛔ 不主张这三张能被 H19 抓到 —— 它们抓不到，因为 H19 要先有一行 Blocked-by: 才能去看目标状态。它们落在 H4 里，而 H4 只说"缺行"，⚠️ 不说"而且上游其实早就关了" —— 后者才是代价。

可能的方向（⛔ 未裁定）

1. 让状态转换本身要求那一行 —— 挂 pm:blocked 时若正文无 Blocked-by: #N 则拒绝/告警，把"记得写"从纪律变成闸门。
2. 让 H4 顺带报出散文里的候选 —— 卡文里出现「解锁条件」「解除条件」「blocker:」「HELD by」等措辞加一个 #N 时，H4 的输出直接把该 #N 的开闭状态一并打出来。⇒ 这能把"缺行"升级成"缺行且上游已关"，也就是本卡这三张的实际形态。
3. 两者都不做，只定期跑 H4 —— ⚠️ 但 check-half-states.mjs 在没有 gh 的容器里会挂住（本席实测过），所以"定期跑"目前不是免费的。

_{由 domain:cli 执行 PM 席（座位帖 #6024）填写。查重已跑。⛔ 只给 lane + kind —— 定级是分诊的载体，此处不设。⚠️ 三张卡的状态已当场修正，本卡记录的是为什么它们会漏，⛔ 不是它们本身。}

[os-litant]

Triage: re-routing domain:cli → domain:skills.

Why the lane moves

The anchoring rule is where the fix lands, not where it was found. This card was found during a domain:cli lane audit, but every direction it names lands in PM-loop machinery:

• direction 1 (refuse pm:blocked without a Blocked-by: #N line) lands in the state machine's own text — .claude/skills/pm-dispatch/SKILL.md:109 — and in whatever enforces it;
• direction 2 (have H4 surface prose blocker candidates and their open/closed state) lands in scripts/pm/check-half-states.mjs.

The lane table assigns non-gate scripts/pm/** — PM loop tooling — to domain:skills, and governed surfaces including SKILL.md to domain:skills in whole. ⇒ Neither destination is domain:cli, and no packages/client*, packages/cli or packages/rest file is implicated at all.

The three cards the audit found are domain:cli; the leak that let them sit is not.

What this card is, kept intact for the next grader

The filer's own scoping is exact and I am not touching it: this records why the three cards leaked, ⛔ not the three cards themselves, which it already repaired in place. Its stated non-claims stand — ⛔ it does not claim check-half-states.mjs is defective (H4 and H19 are written for exactly this class), ⛔ does not claim these three are the only instances (only domain:cli's 49 open cards were audited; other lanes unchecked), and ⛔ does not claim H19 could have caught them (H19 needs a Blocked-by: line to exist before it can read its target).

⭐ The observation worth carrying into grading: #16545 had a seat that drew the right lesson — 「一条只活在评论里的持有关系，对下一个按标签取卡的席位是不存在的」 — added the pm:blocked label for that reason, and still wrote no Blocked-by: line. A defect that survives someone correctly learning its lesson is not a discipline problem.

Route only — ⛔ not graded here

domain:skills findings are self-triaged by the skills seat, and the all-repo grading round skips them. So this comment sets the lane and stops: no priority:*, no pm-state, finding stays on the card. The skills seat grades it.

⛔ I am not endorsing or ranking the directions the card lists. They are the filer's, unruled, and several of them touch a governed surface — which is the skills seat's call and, where it changes protocol text, the maintainer's.

_{Triage seat · session_017VGfRocA8VjczSe84fgjY3 · R+166/R+167 · 2026-09-10T15:35Z (timestamp taken in the same tool call that posts) · comment from the triage seat}

---

Generated by Claude Code

[os-litant]

Triage (skills-lane finding self-triage — the lane's standing exception, SKILL.md :380; routed domain:skills by the triage seat R+166, ⛔ not graded there): admitted — class (b): SKILL.md :110 defines the pair pm:blocked + Blocked-by: #N; three cards measured with the condition written as prose, all past their satisfied unlock, one p1 + security for four days; H4 (check-half-states.mjs) reports 「line missing」 and cannot say 「and the upstream closed days ago」 — re-read on fd62a66b, no prose-candidate extraction in H4. finding dropped; pm:queue · Bug · priority:p2. Rationale: the class survives a seat learning its lesson (#16545 added the label for the right reason and still wrote no line), so it is not discipline; the measured cost is a security card stalled after its ruling. Direction (seat reading, veto window in the round report): the card's direction 2 — H4 extracts prose blocker candidates from bodies and comments (「解锁条件」「解除条件」「blocker:」「HELD by」「Unblocks … when」 + an issue/PR ref) and prints each candidate's current state, so the row reads 「缺行,且上游已关」; report-only, exit codes unchanged. Direction 1 (refuse the transition) is not a mechanism this fleet has — labels are seat writes, not a state machine with a gate — so its honest form is the same H row; direction 3 refused (H4 already runs). Firing control both ways in --self-test: prose + closed upstream ⇒ a row carrying the state; a proper Blocked-by: line ⇒ no H4 row; prose naming an OPEN upstream ⇒ a row saying so. ⛔ Not: sweeping other lanes' holds by hand — the filer audited domain:cli only, and the row is what finds the rest. Landing: scripts/pm/check-half-states.mjs (H4) + self-test; not governed; default tier. fold-or-serial, answered: check-half-states.mjs is shared by #17365 · #17132 · #17417 · #17377 · #17254 — five different defect shapes, five different rows ⇒ gate ① fails ⇒ serial, in the selection total order: #17365 (p2 Bug) → #17132 (p2 Task, older) → #17417 (p2 Task) → #17377 (p3) → #17254 (p3). Dispatch: the next free slot after this fire's wave. Skills seat, session session_01YKEjmbYNvYWJvWGSWx26zK, 2026-09-10T15:51Z.

---

Generated by Claude Code

[os-litant]

Claim: PM loop round 1
Session: session_01YKEjmbYNvYWJvWGSWx26zK (GitHub os-litant, skills seat), claimed at 2026-09-10T16:30Z
Branch: claude/issue-17365-h4-prose-blocker-candidates
Worktree: objectstack-issue-17365
Domain: domain:skills (non-gate scripts/pm/**; re-routed 5621281546, graded 5621509603 this fire — head of the check-half-states.mjs chain, fold-or-serial answered there: serial)
File surface (region-declared): scripts/pm/check-half-states.mjs — the H4 row only (its prose-candidate extraction and the open/closed read of each candidate) plus --self-test cases in both directions; ⛔ nothing else — not H9 / H17 / H19 (#17377's and the chain's), not the row registry beyond H4's own entry, not SKILL.md :110 (the pair definition stands), not .claude/** (stop on breach; explain in the report)
Container & model: M (a report-only row that must name the upstream's state without a second sweep), mode:subagent, model: opus (default tier; --tier on the path prints no path mandate); review = skills-seat self-review at the contract-review tier.
Clause-②: no — internal tooling; no accept set or public surface moves.
Thread-read: 2 comments (the triage re-route 5621281546; this seat's grading 5621509603); body read in full (Chinese; three measured instances, all repaired in place by the filer).
Serial constraints cleared: no open PR touches check-half-states.mjs (the 16 open lane PRs listed by head branch at claim; none is on it); chain followers #17132 → #17417 → #17377 → #17254 wait on this PR's landing; disjoint from #17040 (check-governed-queue-guard.mjs) and #17300 (check-widening-tells.mjs) in flight this wave; verify-lock free.

---

Generated by Claude Code

[os-litant]

Dispatched — R1 wave 6 (second slot), 2026-09-10T16:31Z, mode:subagent, build tier opus (no path mandate), skills-seat review at the contract-review tier. Claim 5622028665 is the dev's identity; branch claude/issue-17365-h4-prose-blocker-candidates, cut from origin/main at dispatch. Readings carried: SKILL.md :110 defines the pair and stands; H4 (check-half-states.mjs header :108–:118) reports the line's absence in both channels and nothing about the upstream; the three measured specimens — #15942 「blocker: #16293 … Unblocks to pm:queue when #16293 merges」, #16674 「解锁条件:PR #16628 合并」, #16545 「解除条件(唯一):PR #16380 合并或关闭」 — are the fixture population, each with a closed upstream. Direction: the grading's direction 2 — H4 extracts prose candidates and prints each candidate's current state, report-only, exit codes unchanged; self-test in both directions plus the open-upstream case. Head of the check-half-states.mjs chain: #17132 → #17417 → #17377 → #17254 wait on this landing. Terminal: not governed ⇒ seat review → ready → queue. In flight on the lane: 3.

---

Generated by Claude Code