[Decision] objectos is a fourth repository with no lane, no seat and no pm labels — which shape should it get?

[os-zhuang]

The gap

pm-dispatch models three repositories — "产品横跨三仓: objectstack → objectui and cloud". objectstack-ai/objectos is a fourth, and the protocol has no answer for it:

• No domain:* lane covers it and no repo:objectos seat exists, so the seat/lane bijection ("域 X 谁管" / "PM Y 管什么" each have exactly one answer) has no answer for it.
• The triage seat is defined as "三仓唯一" and produces classification for three repositories. Nobody triages objectos.
• ensure-pm-labels.sh has never run there: no pm:queue, no pm:dispatched, no domain:*. The only pm-vocabulary label present is finding.

This is not hypothetical — work is already flowing through that repository outside the protocol. Open cards objectstack-ai/objectos#61 and objectstack-ai/objectos#64 carry finding with no triage, no domain, no owner. And a change landed there written in-session by a PM seat with no issue filed first, precisely because no lane existed to route it through. That deviation is recorded on objectstack-ai/objectos#66; this card is about the structural cause rather than that one instance.

Update 2026-08-18: dispatch authorized, structure still open

The maintainer has authorized this session to dispatch work in objectos (原话: 「现在允许你派 objectos 的活」), under the standing 维护者直派通道. That unblocks the immediate queue and does not answer this card: an ad-hoc authorization to one session is not a seat, and the next session starts from the same ambiguity that produced the deviation above. The options below stand.

Options

A — full sibling-repo treatment. Extend the triage seat to four repositories; mint a repo:objectos execution seat; run ensure-pm-labels.sh there. This is exactly the shape objectui and cloud already have, so both invariants survive unchanged.
Cost: one more seat Routine and its fire cycles, for a repository with a handful of open issues.

B — triage extended, execution folded into domain:devx (recommended). Triage seat covers four repositories (one extra lane-inventory query per round, no new seat); objectos execution routes to the existing devx lane, which already owns content/docs/** and apps/docs in this repository — objectos is entirely docs and site code. Mint repo:objectos as a seat only when volume justifies it, which is the protocol's own registered scale-out trigger.
Cost: one lane briefly owns docs in two repositories, which is a wrinkle in the bijection until it is split out.

C — declare objectos out of scope. It stays outside the PM loop; issues there are handled ad hoc by the maintainer.
Cost: the gap becomes permanent and deliberate. Every future agent touching objectos hits the same ambiguity and resolves it the same improvised way.

Four facets

• Platform long-term coherence — B and A both close the gap; A is more coherent long-run (repo → seat, as with the other two siblings), B is more coherent today (one lane already owns docs, and objectos is only docs). C leaves a documented hole in an otherwise total protocol, and the protocol's value comes from being total. Slight edge to B, converging on A as volume grows.
• Measured business pull — thin, and worth saying plainly: a handful of open issues, ~15 merged PRs since the repository was repositioned in July. This is real but low-volume, which argues against minting a seat now and for the cheapest change that removes the ambiguity. Points at B.
• AI-agent error-resistance — this is the decisive facet. The failure this card exists because of is an agent finding no lane and improvising one. A and B both make the answer lookup-able; C guarantees the improvisation recurs, and each recurrence will look locally reasonable. Rules out C.
• Startup scope discipline — a seat is a permanent obligation (a Routine, a seat post, a fire budget). B adds one query per triage round and no new obligation; A adds an obligation ahead of demand. Points at B.

Four facets converge on B, so this would meet the confidence gate for auto-adjudication on facets alone — but it sits squarely on the human floor: it changes .claude/skills/pm-dispatch/**, which is ADR-class and governed. Filed for you rather than adjudicated.

Implementation notes for whoever takes it

• Any PR touching .claude/skills/pm-dispatch/** is model: claude-fable-5 (mandatory clause ①) and merges by hand — governed face, no auto-merge, no queue.
• The lane table and the multi-repo section both need editing in the same PR ("新增或退役一个 domain:* 必须同批改本表"); under B the edit is to the repo-coordination section rather than a new domain:* row.
• The line ratchet applies: check-skill-line-ratchet.mjs only lets the file shrink, so this has to be paid for with deletions elsewhere or an explicit ratchet-raise decision quoted in the PR body.
• bash scripts/pm/ensure-pm-labels.sh needs a repo argument or a run against objectos under either A or B.
• objectos has no changeset flow, no packages/, and a different CI shape from the three modelled repositories. Whichever option lands should say so, or dispatched agents will look for machinery that is not there.

[os-support-ai]

Adjudication record (maintainer, 2026-08-18, PM chat with the skills seat, session session_01WswfK2yNYT9hNnMH6TzAwL). Ruling verbatim, untranslated: 「就按 B,带你补的两条,落地吧」

Decided: Option B — triage extends to four repositories; objectos execution folds into domain:devx; no repo:objectos seat is minted now — plus the two seat-proposed additions the ruling adopts:

1. The A-upgrade trigger is written down, not felt: the implementation states in the multi-repo section that objectos escalates to a full repo:objectos seat when either registered scale-out trigger fires (a round's duration approaching the fire cycle, or the objectos queue persistently starving under the selection total order). B is a recorded transition, not a second hole.
2. objectos agent-instruction files are governed surface: per the maintainer's standing governed-surface extension ruling (2026-08-18, 「任何对 agents.md 等文件的修改…包括 objectui cloud仓库」→「同意」), the implementation names objectos in that clause — execution routes to devx, but instruction-file PRs there stay draft/human-merge.

Label transition needs-user-decision → claimed in the same stroke (this card lands in .claude/skills/pm-dispatch/** ⇒ skills lane). Option C is rejected; option A deferred behind the written trigger. The single-session objectos dispatch authorization recorded in the body retires once this lands — the lane answer becomes lookup-able.

---

Generated by Claude Code

[os-support-ai]

Claim: PM loop round 2 (skills seat)
Session: session_01WswfK2yNYT9hNnMH6TzAwL
Branch: claude/issue-9551-objectos-lane-shape-b
Worktree: objectstack-issue-9551
Domain: domain:skills
File surface: .claude/skills/pm-dispatch/SKILL.md (multi-repo coordination section + domain:devx table row + governed-surface clause) — possibly scripts/pm/ensure-pm-labels.sh if the repo-argument route is taken (stop on breach; explain in the report)
Container & model: M, mode:subagent, model: claude-fable-5 — clause ① path mandate (.claude/skills/pm-dispatch/**), per dispatch-gates --tier derivation quoted on #9528 this shift
Clause-②: no — internal protocol text; changes no contract accept/reject behaviour, widens no public surface
Serial constraints cleared: NOT yet — hard same-file serial behind PR #9536 (also edits SKILL.md, including the same domain:devx row). #9536 is ACCEPT'd and awaiting the maintainer's merge; dispatch holds until #9536 is MERGED (this session holds a PR subscription on it and dispatches on the merge event). No other open PR touches .claude/skills/**.

---

Generated by Claude Code

[os-zhuang]

One more constraint, measured while running the first dispatched batch in objectos rather than reasoned about: objectos has no merge queue.

The protocol's landing path for a reviewed dev PR is ready → queue, with "队列是唯一被认可的落地路径, ⛔ 永不队列外合并". In objectos that path does not exist — objectstack-ai/objectos#65 landed by a human pressing merge, and there is no queue workflow in the repository. So a PM seat working objectos can review and flip to ready, and then has nowhere protocol-sanctioned to put the PR. That is not a hypothetical: objectstack-ai/objectos#69 is sitting in exactly that state now, reviewed and green, waiting on a human because the seat has no legitimate way to land it.

Whichever option is chosen needs to say what landing looks like there. Three shapes, no recommendation attached — this is narrower than the seat question and follows from it:

• objectos adopts a merge queue, and the existing rule applies unchanged;
• the protocol grants a named exception for repositories without a queue (direct merge after review, which weakens a rule that exists because queues serialise a shared resource);
• objectos PRs are always merged by a human, which is the de-facto status quo and costs the maintainer one action per PR.

Related to the existing implementation note that objectos has no changeset flow: the gap is not one missing label vocabulary, it is that the repository's CI and landing shape differ from the three modelled ones in more than one place.

---

Generated by Claude Code

[os-support-ai]

{
 "issue": 9551,
 "status": "done",
 "branch": "claude/issue-9551-objectos-lane-shape-b",
 "pr": "https://github.com/objectstack-ai/objectstack/pull/9558",
 "premise_still_valid": true,
 "summary": "Landed Option B per the adjudication: SKILL.md multi-repo section now states the objectos fold-in (triage sweeps four repos, execution folds into domain:devx, no repo:objectos seat), the written A-upgrade trigger (either registered scale-out trigger from rule 4), and the no-machinery notice (no changeset flow, no packages/, no merge queue = human merge, different CI). The governed-face unified definition (ACCEPT path fork) now names the cross-repo extension for objectui/cloud/objectos with the extension ruling quoted verbatim. ensure-pm-labels.sh adds objectos to the state-machine label loop with commented deliberate exclusions (domain lanes, target boards); running it against objectos is the PM's landing step. Ratchet held at 682/682 — +7 added lines paid by deleting Guardrails text that survives verbatim in the seat invariants, the ACCEPT path fork, and AGENTS.md (payment table in the PR body; no re-wrap used).",
 "tests": "All gates at final head 6a8b11b (clean tree, re-run after the only commit): pnpm check:pm-skill-ratchet exit 0 (SKILL.md 682 lines, ceiling 682); pnpm check:pm-skill-id-lint exit 0 (9 files clean); pnpm check:skill-frame-sync exit 0 (4 copies isomorphic, 3 axes); pnpm check:pm-governed-merges exit 0 (self-test 25 assertions); pnpm check:doc-authoring exit 0 (377 files clean); pnpm --filter @objectstack/lint run check:doc-formula-expressions exit 0 (upstream closure built first; note: '--concurrency=2' placed after --filter was forwarded to tsup and failed the first build — the documented trap, retried with --workspace-concurrency before the filter); node scripts/check-nul-bytes.mjs exit 0 (6157 files). Gate list re-derived with node scripts/pm/dispatch-gates.mjs (no paths, script-derived changed set) — 6 gates matched, identical to the dispatch prompt's list, no additional families for this diff; tier line confirms clause-1 fable floor. bash -n on ensure-pm-labels.sh clean. files_changed: .claude/skills/pm-dispatch/SKILL.md, scripts/pm/ensure-pm-labels.sh (+27/-20). skip-changeset applied via read-modify-write, read back: [size/s, skip-changeset].",
 "open_questions": [
 {
 "question": "Falsified PM mechanism assumption: no cross-repo governed-extension clause exists anywhere in SKILL.md or references/ on main at da808c6 (grep for 包括 objectui / 任何对 agents / governed across the skill dir and AGENTS.md: zero hits), so there was no existing clause to name objectos in. Where should the extension live?",
 "options": ["A - anchor it at the governed-face unified definition in the ACCEPT path fork, which the Guardrails bullet already points to (done in the PR)", "B - a new standalone clause/section (forbidden by the dispatch instruction: do not invent a new section)"],
 "recommendation": "A, because the unified definition is the single source every governed check dereferences; the deviation is recorded here rather than silently absorbed."
 },
 {
 "question": "Unadjudicated sub-question raised on this card after the claim: objectos has no merge queue — which landing shape does the protocol adopt (adopt a queue / named no-queue exception / always human merge)?",
 "options": ["A - objectos adopts a merge queue, existing rule applies unchanged", "B - named protocol exception for queueless repos (direct merge after review)", "C - human always merges (de-facto status quo)"],
 "recommendation": "None issued - this is a maintainer decision outside the adjudicated scope. The PR records only the factual status quo (落地 = 人工合并) as the dispatch-expectations notice; the sub-question stays open on this card."
 }
 ],
 "out_of_scope_findings": []
}

---

Generated by Claude Code

---

Generated by Claude Code

[os-support-ai]

ACCEPT — PR #9558 (reviewer of record: skills seat, session session_01WswfK2yNYT9hNnMH6TzAwL). Governed surface ⇒ terminal three-step: conclusion here, the PR stays draft awaiting the maintainer's merge, round report lists it under "awaiting a human merge".

Checklist conclusions, verified against GitHub / local runs at head 6a8b11b — not the report's self-description:

• Shape: draft, base main, Fixes #9551; changed files exactly .claude/skills/pm-dispatch/SKILL.md + scripts/pm/ensure-pm-labels.sh (+27/−20), within the claimed surface (the script was declared conditional); skip-changeset present, read back.
• Content vs the ruling: Option B landed with both additions — fold-in + written A-upgrade trigger + no-machinery notice in the multi-repo section; governed cross-repo extension names objectui/cloud/objectos at the unified definition with the ruling quoted verbatim; triage sections went four-repo with the per-site judgment the dispatch demanded (dependency-direction intro and release-board metric legitimately stay three-repo; the triage section explicitly skips a nonexistent objectos release board — consistent).
• Ratchet payment audited cut-by-cut, not taken on faith: all three Guardrails cuts survive verbatim at their pointed homes — the PM-no-files exception in the seat invariants, the governed prohibitions + draft-only dequeue exit in the ACCEPT fork, and the release-act enumeration in AGENTS.md PD#15 (which is richer than the cut bullet: same acts plus the incident case law). The removed AGENTS.md-wins Guardrails line survives in the invariants section. 682/682 held, no re-wrap.
• Local gate re-run (my own): ratchet · id-lint · frame-sync · governed-merges self-test · bash -n — all clean at 6a8b11b, matching the report's head.
• Deviation, publicly owned by the PM: my dispatch asserted an existing cross-repo governed-extension clause to name objectos in — that assumption was false (the dev's grep evidence stands; no such clause existed). The dev's resolution — anchoring the extension at the governed-face unified definition, the single source every check dereferences — is the right one and is hereby approved. Good falsification, exactly what the mechanism-assumption tier is for.

Open on this card after the PR merges: ① the queueless-landing sub-question (comment 5326477159) — with the seat's recommendation going to the maintainer in chat: adopt C (human always merges) as the standing answer now (volume is a handful of PRs; a named queue exception would weaken the flagship rule for negligible gain), revisit at the written A-upgrade trigger, at which point queue adoption rides the seat upgrade; the PR's factual notice (落地 = 人工合并) then doubles as the rule with zero further diff. ② The PM landing step: run the label vocabulary against objectos.

---

Generated by Claude Code