[Decision] A claimed card follows through to MERGED; cross-lane surfaces import the review instead of exporting the card; in-scope derived work and unclaimed blockers stay with the claiming seat (maintainer question after #17718, 2026-09-13)

[claude]

Filed by the domain:skills seat (session session_01DAcomhvR9kKizeYgg89Vo8, GitHub os-project-manager) at 2026-09-13T07:17Z, as the carrier for a ruling the maintainer asked for live in this seat's chat (2026-09-13T06:5xZ–07:1xZ). ⛔ Not dispatched: the charter change lands in this lane once ruled (SKILL.md 〈多仓协调〉/〈认领〉/〈状态模型〉, references/core-rules.md, references/contract-review.md, references/dispatch-runbook.md, references/state-machine.md; governed, maintainer hand-merge).

The maintainer's words (verbatim, untranslated)

各车道目前划分是不是太严格,17718 开发了一半换车道,压了好几天。
是不是已经立了卡片开始处理的,就应该跟进到底?或者有没有更好的方案?
已经立了卡片开始处理的,就应该跟进到底的话,那衍生出来的相关子任务,或者被 blocked 的任务是否也应该作为子任务合并处理,尤其是对于优先级高的卡片。

Per-card precedent already given on #17178 (2026-09-13T06:5xZ, quoted on comment 5651809638): 「17718 你直接跟进到合并。如果需要契约复审你可以派 fable」.

Governing text

• SKILL.md 〈多仓协调〉: 「凡触 packages/spec 一律转 domain:spec 座位,不论谁需要它。」 · 「跨座位转移:工作跨座位线,PM 永不跨(唯一豁免:简单阻塞项直接接手)。」
• SKILL.md 〈状态模型〉 pm:retriage row: 「向分诊提问(改判、跨域 PR 指定车道、改路由、拆卡、裁 dev 报告留下的分叉)」; 〈域车道〉: 「跨域例外路径:拆不动的跨域单 PR 由分诊席指定一个车道 PM 认领,认领评论申报文件面。」
• references/contract-review.md 〈复核归属与资格〉: 「spec 席审低档实现者的契约增量,非自身产物;余席条款②复核 = 默认判断档自审加门禁。」 · 「独立性件,spec 席:契约判断真分叉或 dev 挂旗 ⇒ 起上下文隔离的达档复核子代理。」
• references/dispatch-runbook.md 〈跨车道简单阻塞项直接接手〉: 「只覆盖等一周与做十分钟的形状。」 · 「packages/spec 恒归 spec 座位,本条不豁免:唯一所有者规则更硬。」
• references/state-machine.md: 「释放者 = 让卡离手的席位:行首 Release:(会话/因/去向)与清 assignee 同笔;下一任重新认领。」
• SKILL.md 〈Epic 子树车道〉: 「衍生问题三分,判据:不修它 epic 验收过不过得去;in-scope ⇒ 挂父单 sub-issue 下轮自动入队。」 (the subtree model already exists for epics).

Prior rulings read: 「凡触 packages/spec」 · 「跨座位转移」 · 「简单阻塞项」 · 「Release:」 → the five texts above; no ADR governs PM lane ownership; none in docs/adr/** for pm:retriage (control: ADR-0087 hits 1 file there).

Measured instance — #17178 / PR #17718 (all stamps from the card and PR)

UTC	event	wait
2026-09-11T18:08Z	dev delivers PR #17718, green (32 success / 6 skipped / 0 failures)	—
2026-09-11T18:11Z	cli seat: 4 of 8 files in packages/spec ⇒ pm:retriage
2026-09-13T01:57Z	triage answers: domain:cli → domain:spec, assignee left (「分诊不重指派」)	~32 h (single triage seat; vacant ~20 h; ~23 retriage backlog)
2026-09-13T06:32Z	cli seat releases (pm:queue, assignee cleared)	4 h 35 m structurally invisible (#17932)
2026-09-13T06:58Z	maintainer direct dispatch: cli seat drives it to merge, contract review by a fable agent	total ~37 h idle for a green PR

Same family this week: #17849 (spec half of #17323) has carried no domain:* since 2026-09-12T12:41Z (three triage windows); #17905 measured a 9-hour knock loss on a cross-lane review request. In this lane today, three cards (#17374 second half, #17569 rider, #17798) wait on devx PR #17803, itself awaiting the maintainer's hand merge.

Question 1 — who finishes a card that was dispatched into a surface another lane owns

• A (recommended) — 认领即跟到底: the claiming seat owns the card to MERGED. A cross-lane surface discovered after dispatch (packages/spec included) does NOT move the card: the seat imports the review — an isolated contract-review-tier reviewer subagent (the feat(spec): export SEED_WRITE_EXECUTION_CONTEXT and bind all three seeders to it #17718 precedent), record carrying Implemented-by: / Reviewed-by: so a self-review is machine-visible; governed paths still park at the human terminal. pm:retriage re-routing applies to undispatched cards only. The spec seat remains the sole intake for NEW packages/spec work (triage routing unchanged). Charter cost: two or three lines across SKILL.md, core-rules.md, contract-review.md, dispatch-runbook.md, paid by density.
• B — keep the transfer, fix only the mechanics: the patrol row from [finding] A re-routed card keeps the old lane's assignee, so it is structurally invisible to the lane it was routed to — and the patrol has no rule that sees it #17932 (assignee lane ≠ domain:*) and 「分诊不重指派」 narrowed to 「分诊不指派给新人」 so a re-route clears the assignee. Saves the 4.5 h; does not save the 32 h.

Question 2 — derived work and blockers of a card being followed through

class	test	disposition (recommended)
in-scope derived work	the card's acceptance cannot pass without it (三分判据 first class)	folds in: a sub-issue under the parent; the claiming seat owns the subtree automatically (no return to the pool); priority inherits from the parent; driven to MERGED by the same seat — the epic model generalised to any dispatched card
blocker (Blocked-by: target)	unclaimed vs already in flight	unclaimed ⇒ the blocked seat claims and does it, any size (the S-level cap of 「简单阻塞项」 lifted), contract surfaces by A's imported review; in flight elsewhere ⇒ wait, pm:blocking ranks it first in that lane, the waiter is written on the blocker card; p0/p1 parents propagate priority down the chain
incidental finding	unrelated to acceptance	does not fold: an independent card back to the pool (scope creep otherwise)

Two boundaries either way: one card = one PR (queue and review granularity kept; a trunk branch per AGENTS.md is the multi-card form); the spec seat keeps the intake of new packages/spec work.

维护者速读

改了什么 —— 什么都还没改;这是请你定一条 PM 规则:一张已经派出去、开始做的卡,谁负责把它做完。

为什么改 —— #17718 交付后因为「碰到了 spec 面就得换席」,一张全绿的 PR 空转了 37 小时:32 小时等分诊席答复改路由(单人席位、空缺近 20 小时),再 4.5 小时因为 assignee 没清而对接收席不可见,最后靠你当面直派才动。同一周还有 #17849(18 小时无人路由)、#17905(9 小时敲门丢失)是同一类交接损耗。车道买的两件事——并行写、契约复审档位——在交付之后都可以在席内满足,不需要把卡交出去。

风险与代价(含回滚) —— A 只改章程两三行,不加标签不加状态;复审仍达档、队列守卫与双载体照旧,去掉的只是「静默失败的那一次交接」。风险在 spec 席的独占感:它仍独占新工作的入口,不再收回已交付的活。回滚 = 把那几行改回来。B 只修机械面,便宜但省不下大头。

席位意见 —— 荐 A,并按上面的三类表处理衍生与阻塞项(in-scope 并入子树、无人认领的阻塞项直接做、顺带发现不并)。

你要做的(一个动作) —— 回一个字母:A / B;若同意三类表,回「同意」即可(不同意的类点名)。

os-decision-facets

• ① 项目长远合理性:A 让「归属在派发时定、不在交付后重议」成为规则,减少交接点(每个交接点都是最慢席位的串行点);B 保留交接、只修可见性。A 收缩特例,不增生契约。
• ② 实际业务拉动:实测三例(feat(spec): export SEED_WRITE_EXECUTION_CONTEXT and bind all three seeders to it #17718 37 h、Drop the M9.1 / M9.2 phase language from packages/spec and packages/formula (the spec-seat half of #17323, ruled B) #17849 18 h+ 无路由、[finding] A seat-post body refresh silently expires every unread knock — the read window is 「comments newer than the last body edit」, and a director-ruled cross-lane review took 9 hours to arrive because of it #17905 9 h),全部是交接损耗而非工程状态;本车道另有三张卡等同一张受管 PR 手合。零拉动不成立。
• ③ 防 AI 犯错:A 的复审仍在契约复审档位,记录带 Implemented-by / Reviewed-by,自审机读为 SELF-REVIEW,受管面照旧留 draft;失败方向从「静默不可见」变为「响亮的复核记录缺失」。B 的巡查行也响亮,但只盖住 4.5 小时那一段。
• ④ 创业阶段不扩散:A 不加标签、不加状态、不加工具,改两三行章程;子树归属推广的是既有 epic 模型,不是新机制。
• 推荐:A;子树三类表按推荐。回退:B。置信缺口:本分析未量化「交付后换席」在全舰队的频率,只有本周三例;spec 席对 A 的异议未征询(它今天没有在任会话可问)。

---

Generated by Claude Code

[claude]

Ruling recorded — question 3 (raised in the same discussion): execution seats self-route, and triage performs no after-the-fact audit of those routings

Maintainer, verbatim (live chat with this seat, reply received between 2026-09-13T07:30Z and this write), to the seat's proposal 「落点文件无争议的卡,执行席可以按车道表自己路由(留审计评论),分诊席改做事后审计」:

②同意,执行席自路由,分诊也无需事后审计

What is ruled. An execution seat may route a card by the lane table (the domain of the package the fix lands in) without waiting for the triage seat, and the triage seat does not audit those routings afterwards. This is a ruling on SKILL.md 〈域车道〉 「domain:* 只由分诊席产出;打标签不等于认领」 and 〈多仓协调〉 rule 4 「分诊座位唯一:定级/路由/type 由它一座生产」, and on references/core-rules.md's mirrors of both.

What the seat proposed in the same breath and is awaiting the maintainer's word on (⛔ not ruled yet, written here so the charter edit is one PR): (1) self-routing into one's own lane carries self-grading (priority + type) in the same write, with an audit comment; (2) routing a bare card into ANOTHER lane by the table is allowed, the receiving seat may re-route once on the same mechanical predicate, a second bounce goes to the decision box; (3) the triage seat keeps four duties — the sweep for bare cards nobody touched, cross-repo dedupe / shadow checks, the release board, decision-box hygiene — and cards whose body says the landing site is undecided stay with triage or go to the decision box.

Questions 1 (A / B) and 2 (the three-class table) on this card remain open. State: needs-user-decision kept until they are answered; the charter change lands as one PR in this lane.

Freshness gate: no comment newer than the card body at this write.

---

Generated by Claude Code

[os-tesla]

Ruling recorded — questions 1 and 2: A — a claimed card is followed through to MERGED by the claiming seat; a cross-lane surface discovered after dispatch imports the review instead of exporting the card; in-scope derived work and unclaimed blockers stay with that seat (director seat, decision batch #131 item 1, 2026-09-13)

Maintainer, verbatim (live PM chat, reply received 2026-09-13T07:3xZ), to decision batch #131 presented as 1A(+三类表)·2(1)·3A·4B: 「同意」. Question 3 (execution seats self-route; no after-the-fact triage audit) was ruled by the maintainer directly in the skills seat's chat and is recorded at 5651955479; this comment does not restate it. The card carries the maintainer's own questions verbatim and the per-card precedent 「17718 你直接跟进到合并。如果需要契约复审你可以派 fable」; this ruling generalises that precedent.

Derived first from the long-term axis: every hand-off is a serial point on the slowest seat, and the two things lanes buy — parallel writing and a contract-review tier — are both satisfiable inside the claiming seat after delivery; single ownership to delivery is the ordinary engineering norm. Measured this week: #17718 idle 37 h as a green PR, #17849 18 h+ unrouted, #17905 9 h. ②③④ do not overturn ①.

What is ruled

Q1 — ownership. The seat that claims a card owns it to MERGED. A cross-lane surface found after dispatch (packages/spec included) does ⛔ not move the card; the seat imports the review — an isolated contract-review-tier reviewer sub-round (the #17718 precedent), its record carrying Implemented-by: / Reviewed-by: so a self-review is machine-visible; governed paths still park at the human terminal. pm:retriage re-routing applies to undispatched cards only. The domain:spec seat keeps the sole intake of NEW packages/spec work. B's mechanics land too, because they are correct regardless: the #17932 patrol row (assignee lane ≠ domain:*) and 「分诊不重指派」 narrowed to 「分诊不指派给新人」 so a re-route clears the assignee.

Q2 — derived work and blockers, three classes.

class	test	disposition
in-scope derived work	the card's acceptance cannot pass without it	folds in as a sub-issue under the parent; the claiming seat owns the subtree automatically; priority inherits; driven to MERGED by the same seat
blocker (Blocked-by: target)	unclaimed ⇒ the blocked seat claims and does it, any size (the S-level cap of 「简单阻塞项」 is lifted; contract surfaces by the imported review); in flight elsewhere ⇒ wait, pm:blocking ranks it first in that lane, the waiter is written on the blocker card; p0/p1 parents propagate priority down the chain
incidental finding	unrelated to acceptance	⛔ does not fold — an independent card to the pool

Two boundaries: one card = one PR (a trunk branch per AGENTS.md is the multi-card form); the spec seat's intake of new spec work is unchanged.

State

needs-user-decision → pm:queue; domain:skills / priority:p2 kept. Landing as one charter PR in this lane together with Q3 (SKILL.md 〈多仓协调〉/〈认领〉/〈状态模型〉/〈域车道〉, references/core-rules.md, contract-review.md, dispatch-runbook.md, state-machine.md; governed, maintainer hand-merge).

---

Generated by Claude Code

[claude]

Ruling recorded — questions 1 and 2, and the two corollaries of question 3: the seat's recommendations accepted as written

Maintainer, verbatim (live chat with this seat, between 2026-09-13T07:52Z and this write, replying to the seat's five-item explanation): 「以上接受你的建议。」

The accepted text, verbatim from that explanation:

A(荐):认领席跟到 MERGED。具体动作:标签不换、assignee 不清;认领席起一个隔离的 fable 复核子代理审契约(就是你对 17718 那句授权的通用化),复核记录写明 Implemented-by: 分支、Reviewed-by: 席位会话;受管面照旧留 draft 给你。spec 席仍是新 spec 卡的唯一入口(分诊路由不变),只是不再收回已交付的活。

验收必需的衍生工作 → 立成父单的 sub-issue,认领席自动拥有,同席做到 MERGED,优先级继承父单。… 阻塞项(Blocked-by: 指向的卡):无人认领 → 被阻塞的席位直接认领做掉;已被别席在飞 → 等,但那张卡在对方车道排到最前(pm:blocking 机制已有)。… 顺带发现(与验收无关)→ 独立卡回池子,不并,防范围蔓延。边界不变:一卡一 PR;spec 新工作入口归 spec 席。

(a) 自路由是否含自定级?席位把卡路由进自己车道时,同笔打 priority 与 type。… 荐:含。 (b) 能否把裸卡路由到别的车道?… 荐:能;接收席不同意可按车道表改一次,第二次分歧进决策箱。… 分诊席剩四件:扫无人碰的裸卡、跨仓查重、发版板、决策箱勤务。

What is ruled. Q1 = A (a claimed card stays with the claiming seat to MERGED; a cross-lane surface imports an isolated contract-review-tier review with Implemented-by: / Reviewed-by: on the record; pm:retriage re-routing is pre-dispatch only; the spec seat keeps the intake of NEW packages/spec work). Q2 = the three-class table (in-scope derived work folds in as sub-issues owned by the claiming seat with inherited priority; an UNCLAIMED blocker is claimed and done by the blocked seat — the size cap was offered as an option and not taken; an in-flight blocker is waited on with pm:blocking ranking it first; incidental findings stay separate cards). Q3 corollaries: (a) self-routing carries self-grading (priority + type) in the same write; (b) a seat may route a bare card to another lane by the lane table, the receiver may re-route once on the same predicate, a second bounce goes to the decision box; the triage seat keeps the bare-card sweep, cross-repo dedupe / shadow checks, the release board and decision-box hygiene.

State: needs-user-decision → pm:queue in this stroke; priority:p2 → priority:p1 (maintainer-directed charter revision; it also releases #17742 / #17933 / #17932's hold). Execution shape (fold-or-serial answered on this chain head at dispatch): one charter-revision PR in this lane carrying this card with #17950 (tiering, ruled 「我点头」), #17933 (the per-lane applicability table), #17934 (③ exception, ruled 1), #17742 (the count + the disjunct decision), #17951 (director batch duty) and #17624 (ruled B) — every member ruled, one region (.claude/skills/pm-dispatch/**, AGENTS.md PD #14, scripts/pm/check-governed-*.mjs), one hand merge; the five gates and the exclusion list are written on the chain-head claim. Blocked-by: liveness: none on this card. Freshness gate: comments newer than the body at this write = 5651955479 (this seat's Q3 record) only.

---

Generated by Claude Code

[claude]

Record precedence, for the next reader: the ruling of record on questions 1 and 2 is the director seat's 5651976160 (2026-09-13T07:36Z); this seat's 5652079343 carries the verbatim chat text of the acceptance and is a supplement, not a second ruling. The one state change this seat made in that stroke: priority:p2 → priority:p1. Skills seat, session_01DAcomhvR9kKizeYgg89Vo8, 2026-09-13T07:58Z.

---

Generated by Claude Code

[claude]

Ruling: direct · letter ③ · maintainer 「同意」 (live chat with the skills seat, received immediately before this record, 2026-09-13T09:51Z)

Revision to this card's Q3 and Q2, recorded by the domain:skills seat (session session_01DAcomhvR9kKizeYgg89Vo8) in the fixed header form of #17748 item 6. The maintainer asked, in a discussion on token and API cost, whether everything should still pass through triage (「那如果还是要求经过分诊呢？你综合评判哪个划算」); the seat compared three shapes (① all through triage; ② self-routing + self-grading, this card's Q3 as accepted at 2026-09-13T07:36Z; ③ all through triage with one exception) and recommended ③; asked 「分诊最终的结论是什么？」 the seat stated it as five points; the maintainer answered 「同意」. The five points, verbatim as presented and accepted:

1. 一律经过分诊。 定级、路由、查重、跨仓 shadow 检查、duplicate_of 关单,全部收回分诊席一个人做;今早 Q3「执行席自路由自分级、分诊不审计」收回。理由是省的:分诊每轮本来就拉一次列表,查重零边际;分散到各席位则每个立卡者各拉一遍,API 更贵、token 打平、质量参差。
2. 唯一例外:在飞卡的衍生子卡(今早 Q2 第一类)。拥有它的席位带父卡的域和优先级直接立卡并派发,不等分诊;分诊的小时 sweep 会事后扫到它,查到重复按 duplicate 关。Q2 其余三类不变。
3. 立卡者不查重,只在卡面附 3–5 个查重词(包名、报错串、符号名);分诊拿这些词往自己的列表上 grep,「零命中要有对照词」的纪律随之搬到分诊。
4. 查重缓存住在分诊席自己的容器里:新会话开班全量拉一次,之后 since 增量;不建分支、不共享、不依赖备份。
5. 备份是另一条线,已立卡 Issue backup to Cloudflare R2 — a per-repo Actions snapshot of every issue and comment, so a GitHub account ban cannot take the audit chain with it (maintainer, 2026-09-13) #17989:Actions 定时把每张 issue 和评论快照推到你的 Cloudflare R2,与分诊缓存互不依赖。

Effect on this card's rulings of record: Q1 = A (follow-through to MERGED; cross-lane surfaces import the review) — unchanged. Q2 three-class table — unchanged, with class 1 (in-scope derived sub-issue, owned by the claiming seat at inherited priority) now the ONLY route that bypasses triage; classes 2–4 unchanged. Q3 (5651976160 / 5652079343: 「②同意,执行席自路由,分诊也无需事后审计」) is WITHDRAWN together with its corollaries (self-grading; routing bare cards to other lanes by the lane table): domain:*, priority:*, type and dedupe are again the triage seat's sole production; execution seats never grade or route a bare card. The triage seat's standing duties gain: dedupe on the keywords the filer attaches, with the zero-hit control; a session-local cache (full pull on a fresh session, since increments after). The filer's duty becomes: attach 3–5 dedupe keywords, do not pull the board.

Charter text (governed — rides the charter-revision PR, chain head this card): SKILL.md 立单前查重 line → 「分诊查重,立卡者附查重词」; the self-routing lines drafted for Q3 are dropped; the Q2 class-1 exception stated once; core-rules.md mirror in the same PR; lanes/triage.md gains the cache line. Ruling-ref for claims: this comment.

---

Generated by Claude Code