finding(pm): pm:queue 半态「已判未转」——两条可机判的 H 规则,加 #11086 的一条输入

[os-zhuang]

Finding,只记录不认领(维护者 2026-08-23 口头采纳立卡建议;记录席位 session 757f0dc0-5dae-51d7-9053-c15c57ebfe3b)。

事故

另一 PM 席位把 #10534 / #10746 / #10556 挂在 pm:queue 长期不处理。维护者点破后,该席位自查结论:技能原文清楚,是执行没跟上,至少四处违规——

技能原文	实际做的
「挂 needs-user-decision、退出活动队列」	贴了分析,留在 pm:queue
「状态转换同笔……永不留挂」	正是留挂
前提刷新:「零命中/变形的就地改写或撤卡」	只加注解,没改也没撤
pm:queue 定义「无可问之事」	#10556 带两个未决设计问题入队

外加 #10534 咬到的鲜度门:「录前重读晚于正文最后编辑的评论,有修正的先调和正文」没执行。

判读

四条违规同一签名:判断做了(有分析产出),状态写入没做(纯开销的那半)。长上下文席位注意力衰减最先丢的就是两段式义务的后半段——这不是"技能写得不清楚",是散文规则在长会话里的可预测衰减。本仓已有的结论一贯如此:活下来的规则都是有门的规则(worktree/stash hooks、check-half-states.mjs、各 ratchet)。

提案:进 scripts/pm/check-half-states.mjs 的两条新 H 规则

1. 矛盾态:卡同时挂 pm:queue 与 needs-user-decision(或等价未决标记)= 红。pm:queue 的定义就是「无可问之事」,共现即违反定义,机器可判。
2. 腐烂告警:pm:queue 卡零活动超过 N 天 = 点名,不判对错,只逼出一次显式转换(派发 / 转 needs-user-decision / 撤卡 / 改写)。

第三条(鲜度门)机判不了意图但可判事实:结论型评论时间戳早于卡正文最后编辑时间戳 → 提示需重新调和;可作 warning 级,不作 red。

不做的事

不往 .claude/skills/pm-dispatch/SKILL.md 加散文(682/682 零余量,且散文正是刚失效的手段)。「散文规则 → 机器门」的整体审视喂进已裁决待触发的 #11086 全文重构轮,本卡只落上面可机判的规则。

[os-zhuang]

Promoted to pm:queue + scope expanded with today's measured instances (skills seat, session 5213b871-5164-5bc3-8874-28b336bbcd40, self-triaged under the skills-lane exception; authorization: maintainer 2026-08-24, live PM chat — the half-state batch report plus verbatim 「以上问题要排查skills原因」).

The maintainer's sweep surfaced four half-state CLASSES in one afternoon. Each has a protocol root cause, and this card is the vehicle for the machine-checkable rules that close them:

1. Dual-hang pm:queue+pm:blocked (Retire the nine dead members of the plugin-manifest contributes block (events/menus/themes/translations/actions/drivers/fieldTypes/functions/commands) — tombstone route, kinds survives #10724, sat 3 days) — the unlock/park transitions are two independent label writes with no exclusivity invariant. Rule: the pm-state labels (pm:queue/pm:dispatched/pm:blocked/pm:on-hold) are MUTUALLY EXCLUSIVE; a patrol row flags any card carrying two, and every transition in the skill text is worded as replace-not-add.
2. Assigned-but-pm:queue (7 instances: [Epic] Migrate SCIM from @better-auth/scim 1.7.0-rc.1 to stable 1.7.x — seven-model provision, sys_scim_provider disposition, token reissue coordination, net-new Setup surface #11632, pm-dispatch: unlock scan must distinguish "upstream merged" from "consumer can install" — release-gated cross-repo unblocks go pm:on-hold with a Restart-when, not back to queue #11498-post-close, objectui bug(platform-objects): 超预算后台 seed 时,空库自证仍可能签在尾部数据落定之前 —— 首启一次启动两套契约 #4795/fix(sharing)!: give the share-management surface the authorization layer it never had (ADR-0111 P0, #3902) #3965/fix(plugin-sharing): 记录删除后 sys_share_link 令牌立即失效 —— resolve 存在性校验 + 删除级联 (#5190) #5214/fix(plugin-auth): 每号码 OTP 发送预算改用惰性解析的共享计数存储 —— 多节点下不再按节点数倍增 (#4790) #4806/✅ spec 双源清账主单:基线 52 → 0(2026-08-03 收官)—— #4446 gate 落地后的偿还 worklist #4535) — three distinct sub-causes, measured today: (a) an epic owner decomposing without flipping ([Epic] Migrate SCIM from @better-auth/scim 1.7.0-rc.1 to stable 1.7.x — seven-model provision, sys_scim_provider disposition, token reissue coordination, net-new Setup surface #11632); (b) assignee residue from the 2026-08-24 identity incident and from filer self-assign, with the claim protocol saying "comments are the arbiter" but NOTHING instructing residue cleanup; (c) decision-shaped cards mislabeled pm:queue when their own bodies demand a ruling (bug(platform-objects): 超预算后台 seed 时,空库自证仍可能签在尾部数据落定之前 —— 首启一次启动两套契约 #4795 「需裁决」, fix(plugin-auth): 每号码 OTP 发送预算改用惰性解析的共享计数存储 —— 多节点下不再按节点数倍增 (#4790) #4806 「本条不带 pm:queue 的原因」 — the card text and the label directly contradicting). Rules: assignee⊕pm:queue is a flagged pair unless a fresh claim comment exists; first-touch grading must check label-vs-body-disposition agreement.
3. needs:contract-review lingering after PASS ([Decision] Does react-blocks.ts keep objectName / viewType as sanctioned React-tier props, or do they converge on the metadata-tier spelling? — the carrier objectui#2890's last phase has never had #11284 card-side, this seat's own miss today: verdict posted, PR cleared, card write skipped) — the dual-carrier clear is two writes with read-back required per carrier, but the checklist doesn't enumerate the carriers as a PAIRED postcondition. Rule for references/contract-review.md: a PASS is complete only when BOTH carriers read back clear in the same stroke; a patrol row flags a card whose latest review comment says PASS while the label survives on either carrier. Companion rule: needs:contract-review on a card with NO open PR carrier (Should a definition-level input-schema refusal be non-retryable? #9889's parity floor makes a flow burn its whole retry budget on a verdict that cannot change #10025, standing noise for days) is premature — the gate label hangs at PR-open time, not at ruling-request time.
4. Stale pm:blocking derivations (A failing turbo run test task can land in Test Core with ZERO captured output, and the completeness guard's green does not cover a package that printed nothing #10032 pointing at a 404'd dependent; Migrate the release toolchain to @changesets/cli v3 — one atomic PR carrying the bump, the pre-mode restructure, and the gates that model v2's semantics #9465 with all dependents closed) — H14 derives additively but never retires. Rule: the derivation is a two-way sync — recompute the dependent set each patrol and STRIP when it is empty or every named dependent is closed/404.

The original card's two H-rules plus these four make one coherent deliverable: a scripts/pm/ patrol extension (report-only first, per the H21 precedent) + the matching one-line transition wordings in the skill text. Governed surface ⇒ draft PR, human merge; the patrol script half is ungoverned.

---

Generated by Claude Code