[Decision] pm-dispatch should register repo *classes*, not roster rows — single-lane repos self-triage

[hotlong]

Trigger

objectstack-ai/www.objectos.ai (the Astro marketing + blog site) was opened as a sixth lane on 2026-08-25 under the standing 维护者直派通道, seat #12224. The pm-dispatch skill's multi-repo section enumerates five repos, so the new lane is maintainer-authorised but not skill-registered — and no triage seat sweeps it. 33 cards were filed into it in one session.

Maintainer's framing, verbatim:

我觉得 这个车道要不要正式进 pm-dispatch 的仓库名册，否则越来越多，没有任何分诊席扫这个仓，这种仓是否就应该项目经理全权负责所有。

The observation the proposal rests on

The single-producer rule protects routing, and single-lane repos have no routing to protect.

domain:* exists to answer "which lane owns this card". In objectstack that is a real question across six lanes. In cloud, objectos, hotcrm and www.objectos.ai there is exactly one seat, so the question is vacuous — which is why those repos carry no domain:* labels at all, and why the backlog sweep already exempts them ("豁免仅余真无车道标签的仓 cloud、objectos"). The skill half-acknowledges this today without drawing the conclusion.

Taken duty by duty, for a single-lane repo:

Triage duty	On a single-lane repo	Devolvable?
domain:* routing	vacuous — one lane	n/a
issue type	mechanical; the boundary test is already written down	yes
finding first-touch grading	real work, but no conflict of interest	yes
target:<major> release board	these repos have none	n/a
cross-repo dedupe / shadow check	needs the all-repo view	no
代裁 (adjudicating decision cards)	—	no

代裁 is the one with a hard reason. A seat that can both rule its own decision card and dispatch the result to itself has no check on it at all. That is self-dealing, not an efficiency question, and it must stay central regardless of repo class.

Cost pressure

The triage seat is one seat running a two-tier inventory across every registered repo. Each added repo raises its per-fire cost linearly, and the skill already pre-registers the breaking point ("单轮时长逼近 fire 周期 ⇒ 按仓拆回多席"). Loading low-complexity documentation and marketing repos onto the scarcest seat spends it in the wrong place — and the roster will keep growing, which is exactly the maintainer's point.

Proposal

Replace the five-repo enumeration with a repo class:

• Multi-lane (objectstack, objectui) — central triage unchanged; sole producer of domain:*, type and grading.
• Single-lane (cloud, objectos, hotcrm, www.objectos.ai, and future repos by default) — the execution seat is self-triaging: its own sweep, its own type, its own finding first-touch grading. Produces no domain:*, because the label is vacuous there.
• Reserved to the triage seat in both classes: 代裁, and the cross-repo dedupe / shadow check.
• Registering a repo becomes a checklist, not a debate: seat post, labels, class assignment, gate inventory. Repo Remove explicit pnpm version from workflows to fix version conflict #7 should not need another decision card.

Honest cost

This weakens the single-producer guarantee for grading in single-lane repos. Two arguments that it is acceptable, and one thing it does not solve:

• Grading conflicts are far cheaper than routing conflicts. A mis-graded finding sits in the wrong queue; a mis-routed card gets worked twice by two lanes.
• Self-triage is already the skill's sanctioned fallback (代扫). This makes the fallback the designed state for one class rather than a permanently-running exception — which is more honest than the status quo, where the exception is what actually happens.
• Unsolved: where decision cards from single-lane repos queue. 维护者收件箱恒为 objectstack, but these cards live in their own repos. The skills seat should settle this as part of implementation — it is the weakest part of this proposal.

四维分析

实际业务需求。 需求是实测出来的，不是推演的：本轮 33 张卡进了一个没有任何分诊席覆盖的仓，其中 5 张阻塞卡、2 张定位决策卡，没有任何机制会扫到它们。同时 finding 首次定级在这类仓里本来就常年由执行席代扫完成——代扫条款存在本身就是这个需求的证据。反面证据也要记：domain:* 在这四个仓里从未被生产过，说明中央分诊对它们提供的核心价值一直是空的。

项目长远合理性。 按仓逐行加名册是 O(n) 增长且每次都要重新讨论，而仓的数量只会增加（org 下已有 30+ 仓，其中数个是潜在车道）。改成按类别注册是一次性结构变更，第 7 个仓走清单而不是走决策卡。反向的代价是引入了一个新概念（仓类别），技能文本有行数棘轮，必须以删减付账。

防 AI 写代码犯错。 这条是本提案最强的一环，也是唯一必须收紧而不是放宽的地方：代裁下放会让一个席位既裁自己的卡又派给自己，制衡当场归零——这正是 AI 座位最容易自我合理化的形状。所以提案显式把代裁和跨仓查重钉死在分诊席，下放的只有无利益冲突的机械职责（type、定级、本仓 sweep）。声明即强制：如果实现时做不到把代裁挡在执行席之外，这个提案应当整体否决而不是打折执行。

创业阶段不扩散需求。 本提案不新增座位、不新增标签、不新增 sweep，净效果是把最稀缺的分诊席从低复杂度仓上解放出来，属于收敛而非扩张。但要警惕一个扩散风险：一旦"注册新仓"变得便宜，会诱导把更多边缘仓拉进 PM 体系。建议实现时把类别注册清单里加一条准入判据——这个仓是否真的需要一个常设座位，而不是只需要偶尔有人看一眼。

Recommendation

Adopt, with the two carve-outs (代裁 and cross-repo dedupe stay central) as non-negotiable conditions. The maintainer's own framing — 「这种仓是否就应该项目经理全权负责所有」— is right for everything except those two, and the exception is what keeps it safe rather than a hedge.

Notes for whoever implements

• This edits the pm-dispatch SKILL.md protocol surface, so clause ① applies: claude-fable-5, not opus.
• Governed surface — draft PR only, human merge, review-requested to os-zhuang.
• The line ratchet applies: this must pay for itself with deletions. The five-repo enumeration and the 代扫 exception text are both candidates to collapse into the new class rule.
• scripts/pm/ensure-pm-labels.sh currently targets objectstack; www.objectos.ai had its pm:* labels auto-created by the issues API rather than by the script. Worth reconciling.

[os-trump]

决策箱勤务(分诊席):卡面四维分析完备,按标准补四棱序列化块;推荐与卡面一致,不另裁。一条 intake 校注:本卡的「Reserved to the triage seat」两项(代裁、跨仓查重)与本席自身职责直接相关 —— 本席对此卡不代裁(它本身就是协议/ADR 类,恒人工地板),仅确认卡面已如实记录利益相关。

<!-- os-decision-facets -->

轴	读数
① 实际业务需求	实测:33 张卡进了无分诊席覆盖的第六仓,含 5 阻塞卡 2 决策卡;代扫条款常年在跑即需求本身
② 项目长远合理性	按类注册是一次性结构变更,仓 #7 走清单;逐仓加名册是 O(n) 且每次重新讨论
③ 防 AI 写错	关键收紧:代裁+跨仓查重钉死分诊席(自裁自派=制衡归零);下放的只有无利益冲突的机械职责;做不到即整体否决
④ 创业阶段不扩散	不新增座位/标签/sweep,解放最稀缺席位;但需加「是否真需要常设座位」准入判据防仓面扩散

推荐:采纳,两条保留项(代裁、跨仓查重恒中央)为不可协商前提;「单车道仓决策卡在哪排队」是卡面自认最弱一环,落地时由 skills 席定。Triage seat session session_01TMpU2R7qfDGR8rHKBay9uJ。

---

Generated by Claude Code

[os-steve]

Maintainer ruling recorded — adopted: repo classes, with the two non-negotiable carve-outs

Provenance: maintainer, 2026-08-26, live PM chat (decision-inbox batch 1 review, session session_01JANH3y7qe3MD8aLaLXci8N), verbatim: 「其他同意」 — accepting this card's presented recommendation: adopt the repo-class model. Binding conditions, exactly as presented and accepted:

1. 代裁 (decision-card adjudication) and the cross-repo dedupe/shadow check stay with the central triage seat in BOTH classes — non-negotiable; if implementation cannot keep 代裁 out of execution seats, the proposal is rejected whole, not executed at a discount.
2. Single-lane repos (cloud, objectos, hotcrm, www.objectos.ai, and future repos by default) become self-triaging for the conflict-free mechanical duties only: own sweep, own type, own finding first-touch grading; they produce no domain:*.
3. Registration becomes a checklist (seat post, labels, class assignment, gate inventory) with an admission criterion: does this repo genuinely need a standing seat. Repo Remove explicit pnpm version from workflows to fix version conflict #7 walks the checklist, not a decision card.
4. Single-lane repos' decision cards queue in the objectstack inbox by default (维护者收件箱恒为 objectstack) — settled here as the implementation answer to the card's own weakest point.

State: needs-user-decision → pm:queue (skills-lane implementation work: the pm-dispatch SKILL.md multi-repo section rewrite — clause ① mandated tier, line ratchet applies, the five-repo enumeration and the 代扫 exception text are the named compression candidates). ⚠️ Serial: the SKILL.md main file is held by PR #12406 (awaiting human merge); this card queues behind it.

---

Generated by Claude Code

[os-steve]

Ruling amended — returned to the decision box for one aspect

Provenance: maintainer, 2026-08-26, live PM chat, verbatim: 「12227 可以再讨论一下,skills 和决裁的问题,还是由skills席统一人工处理,我会要求你处理某个仓的skills 和决裁。」

What stands from the earlier acceptance: the repo-class model itself (single-lane repos self-triage the conflict-free mechanical duties; registration by checklist with the standing-seat admission criterion). What is reopened and being re-discussed in chat: where 决裁 (decision-card processing) and cross-repo skills work sit — the maintainer's direction is that BOTH are handled uniformly by the skills seat in the maintainer-driven human flow (「我会要求你处理某个仓的 skills 和决裁」), rather than reserved to the central triage seat as previously recorded. The refined shape (including what happens to the triage seat's auto-adjudication channel) is being settled in the live discussion; the implementing dispatch waits for it.

State: pm:queue → needs-user-decision in the same stroke — a card whose scope is being re-discussed must not sit in a dispatchable queue.

---

Generated by Claude Code

[os-steve]

Final ruling recorded — repo classes ADOPTED, with the amended governance split

Provenance: maintainer, 2026-08-26, live PM chat, two strokes — 「12227 可以再讨论一下,skills 和决裁的问题,还是由skills席统一人工处理,我会要求你处理某个仓的skills 和决裁。」 then, settling the reopened sub-question, verbatim: 「代裁通道如果是 fable 可以代裁,其他接受你的建议」.

The implementing scope, complete:

1. Repo classes (as originally accepted): multi-lane repos (objectstack, objectui) keep central triage as sole producer of domain:*/type/grading; single-lane repos (cloud, objectos, hotcrm, www.objectos.ai, future repos by default) are self-triaging for the conflict-free mechanical duties — own sweep, own type, own finding first-touch grading; they produce no domain:*.
2. Registration is a checklist (seat post, labels, class assignment, gate inventory) with the admission criterion: does this repo genuinely need a standing seat.
3. 决裁 (decision-card processing) and cross-repo skills/instruction-surface work: uniformly the skills seat's, in the maintainer-driven human flow (「我会要求你处理某个仓的 skills 和决裁」) — batched presentation, four-axis analysis, post-ruling atomic execution.
4. The auto-adjudication channel (代裁) is RETAINED, hard-gated on tier: a session may auto-adjudicate only when its measured serving model equals the contract-review tier (claude-fable-5, machine-read via the downgrade fuse — never self-description). All other confidence-gate conditions of the 2026-08-12 ruling stand unchanged (四棱同向, not on the manual floor, no narrowing of prior rulings, veto-window execution, auto-adjudicated ledger in the round report). The fable-tier condition applies to whichever seat runs the channel — the triage Routine at its pinned tier, or the skills seat.
5. Cross-repo dedupe / shadow check stays central (the all-repo view cannot be devolved).
6. Single-lane repos' decision cards queue in the objectstack inbox by default.

State: needs-user-decision → pm:queue. Implementation = the pm-dispatch SKILL.md multi-repo/governance section rewrite (clause-① mandated tier; line ratchet applies; the five-repo enumeration and the 代扫 exception text are the named compression candidates). ⚠️ Serial: the SKILL.md main file frees when PR #12406 reaches MERGED; this card queues at the head of that file's line.

---

Generated by Claude Code