[finding] the seat-vacancy threshold is 一天 in SKILL.md:286 and 3 小时 by maintainer ruling, and a multi-seat lane has no seat-SELECTION procedure at all — the four 开轮互斥 readings all presume the seat number is already known #19646

⛔ Filed unassigned, ⛔ no priority:*, ⛔ no domain:*, ⛔ no type — routing and grading are triage's. Routing suggestion: domain:skills. ⛔ Not a claim. ⛔ Not a ruling — it records one and asks what text should carry it.

Filed by the domain:spec execution seat 1 (session_013RDBh5DqXd2xnLwvHLgLFr, seat post #6017), 2026-09-22T03:3xZ. Companion to #19645, which is the instrument half; this is the text half. ⛔ Deliberately not folded: different file, different remedy, and #19645 needs no maintainer decision while this one does.

① A maintainer ruling now contradicts the written threshold

.claude/skills/pm-dispatch/SKILL.md:286, verbatim:

无心跳,活性惰性判定:超过一天无自有产出即可降级回收。

The maintainer, in session, 2026-09-22 — quoted verbatim, ⛔ not translated, ⛔ not paraphrased:

座位帖更新不及时，你要看一下实际哪个座位已经3个小时没更新了，就是没人了。

⇒ 一天 (written) vs 3 小时 (ruled) — a ~8× divergence on the one number that decides whether a seat may be taken.

Under SKILL.md's own 「两条细则冲突 ⇒ 按更严的一条行动并立卡;⛔ 不当场改文本了结」 this seat acted on the ruling (维护者裁决 outranks all written detail) and filed here rather than editing the line on the spot. ⚠️ But the text is what the next seat reads: a seat that reads :286 alone will sit idle for 24h beside a lane the maintainer considers vacant after 3h. ⇒ the divergence has to be resolved in the text, ⛔ not left as a ruling only this session's transcript remembers.

⚠️ The two numbers may not even be the same quantity — :286 reads as a reclamation threshold (when may a successor take over), the ruling as a vacancy threshold (when is nobody there). They could legitimately coexist as two numbers with two names. ⛔ This seat does not choose; that is the question.

② A multi-seat lane has no seat-SELECTION procedure

:210-211 establish that a lane may run several seats (· seat N, one post each) and that mutual exclusion is per seat number (「同席认领互斥,他席认领不触发自退」). The four 开轮互斥 readings at :77-84 are all filtered 「同席号」.

⇒ every one of them presumes the seat number is already known. Nothing defines how an incoming session summoned as /pm-dispatch <domain>@<repo> — with no seat number in the arguments — picks one out of five.

Measured cost of the gap, this session: this seat surveyed five domain:spec posts, judged seat 4 vacant at 10h17m from its post timestamp, and was ~one write from seating into a seat whose occupant was mid-R21 and had written 4 minutes earlier. The recovery was ad-hoc (see #19645), ⛔ not a procedure. Had it completed, two sessions would have been dispatching under one seat number, and the per-seat-number exclusion that :211 relies on would have been silently void for both.

What a remedy would have to buy, ⛔ and what it costs

A seat-selection step would need roughly: survey every pm:seat post in the lane → read each one's actual last write by session id (#19645's instrument) → pick the stalest past threshold → and ⭐ prefer a seat that left a 收班简报, since a brief is an explicit release and needs no inference at all. ⚠️ In this session the one seat with a brief (seat 1) was also unambiguously safe to take, while the seat that looked safest by timestamp was occupied — so "brief first" may be most of the remedy.

⚠️ This card carries a real cost and cannot be waved through. SKILL.md is under scripts/pm/check-skill-line-ratchet.mjs — 行数只降不升 — so any added line must be paid for with deletions, or by a maintainer ruling raising the cap quoted in the PR body. And SKILL.md is 协议语义面 ⇒ 强制条款① puts any such PR at the contract-review tier. ⇒ ⛔ the cheapest honest outcome may be no new text at all: ① could be a one-word edit to :286, and ② might be discharged entirely by #19645's tool plus the existing 「简报即最新事件 ⇒ 立即坐席」 at :86.

Dedup — queries run by this filer, with hit counts

⛔ search/issues ⛔ not used — 403 for session-bound tokens here. Both listings are enumerations read to exhaustion (totalCount 31 and 22, matching returned counts).

Generated by Claude Code