docs(pm-dispatch): give the objectui local backlog its own target:v17 producer (#6903)

[os-project-manager]

Fixes #6903

Maintainer authorization (quoted verbatim, per the card's guardrail note)

This is a .claude/ internal-tooling PR. Under the standing exception it may be implemented by a PM seat only with the maintainer's authorization quoted here:

• 2026-08-09, on the three proposed actions: 「1 补上生产者」
• 2026-08-09, follow-up: 「三个动作都创建issue」
• 2026-08-09, dispatch instruction via the PM session: 「派发这两个任务」

Review guardrail: this PR requires review by another seat or a maintainer walkthrough. No self-review-self-merge. It is opened as a draft and stays that way until that review happens.

Premise verification against origin/main (the issue body is a lead, not a spec)

Both halves of the card's premise still hold at origin/main (19d8948):

• The 发版板 section reads exactly as described. Before this PR it said 「生产者唯一 = 分诊座位」 and, on the release-day line, 「姊妹仓同标签:objectui 随 console bundle 入板」 — one producer, and a sibling-repo clause with consumers but no standing producer.
• The measured gap re-measured today: objectui's target:v17 label has been used 7 times, all closed (the 2026-08-06/07 audit). objectui's local backlog is 119 open / 52 pm:queue (the card recorded 117 / 48 when filed a few hours earlier) with 0 open board items. The numbers moved slightly; the gap is unchanged.

Both readings are pinned in the amended text (the filed reading and this PR's re-measure), rather than silently overwriting one with the other.

What changed

Exactly the three numbered points of the deliverable, edited into the 发版板 section of .claude/skills/pm-dispatch/SKILL.md. The section's voice and bullet structure are kept; this is a surgical amendment, not a rewrite.

1. Producer scope becomes "exactly one producer per backlog." The bullet formerly titled 「生产者唯一 = 分诊座位」 becomes 「每个 backlog 恰好一个生产者」, with two named producers: the objectstack triage seat for the main backlog, and the objectui whole-repo seat (/pm-dispatch repo:objectstack-ai/objectui) for objectui's own local backlog, applying the same binary criterion and the same four blocking classes during its local backlog sweep. Other execution seats still never add or remove the label; mislabels go through escalation. The single-producer discipline is not relaxed — what is corrected is the wrong premise that one seat could sweep every backlog, which is why the step-0 sweep's repo-scoping (the 「析取 2 必须 repo-scoped」 note) is cross-referenced rather than contradicted. The measured gap is recorded inline as the reason the clause exists.
2. Stock-clearing posture, now its own bullet because it covers two backlogs: objectstack was cleared once on 2026-08-06 (309 to 46); objectui's stock gets one equivalent pass (One-time target:v17 clearing pass over the objectui local backlog (48 queued + defect-class findings) #6904); after each side's single pass, both are incremental-only and never full-rescan.
3. The sibling-repo line points at the new producer: objectui now boards items in its own repo, produced by the objectui whole-repo seat, and fixes reach the release through the console bundle at pin-bump time. It also records that release-day inventory is therefore two queries (objectstack and objectui, each label:target:v17 is:open), both cleared to empty. The consumer bullet above still says one query — that wording update, and the pre-release console-bump linkage, belong to Release-day inventory spans two queries (objectstack + objectui target:v17) — update board consumer docs and verify the pre-release console bump linkage #6906 and are deliberately not touched here.

Zero riders

git diff --stat is one file, 24 insertions / 6 deletions, all inside the 发版板 section. Three other queued cards target the same file (#6410 compiler-face checklist, #6490 fan-out ordering, #6741 ADR guardrail) — none of them is implemented here, and no other section of the SKILL is touched. They serialize behind this PR.

Gates

Enumerated from .github/workflows/lint.yml (not from memory), then filtered to what a .claude/**-only change can reach. The applicable set was derived mechanically: grep -rln '\.claude' scripts/ plus a read of each doc-scoped gate's scan roots. All run locally on the pushed commit, all green:

Gate	Result
pnpm check:nul-bytes	OK — 6415 tracked text files, no raw ASCII control bytes (self-test 56 assertions)
pnpm check:doc-authoring	OK — 374 files clean; scan root is .claude (not .claude/skills), so this file is in scope
pnpm check:skill-frame-sync	OK — 4 copies of the decision frame structurally isomorphic across 3 files, incl. this SKILL.md; 42 markdown files scanned for undeclared copies
pnpm check:skill-compatibility	OK — 11 SKILL.md files reconciled against 77 workspace packages. Note its scan root is the published skills/ dir, so it does not read .claude/skills/; run and reported because the dispatch named it
pnpm check:agent-model-declared	OK — 1 agent definition under .claude/agents/, declares opus
pnpm check:docs-audit-scope	OK — reads .claude/workflows/docs-accuracy-audit.js; 179 hand-written docs in sync
pnpm check:role-word	OK — 44 baselined files, no new occurrences
pnpm check:quick-reference-counts	OK — 13 sections, every heading matches its table
pnpm check:adr-anchors	OK — 40 anchored files, 20066 citations across 3290 files resolve
pnpm lint (ESLint)	OK, exit 0 — eslint.config.mjs carries only ts/tsx/mts/cts/js/jsx/mjs/cjs globs, so a markdown-only change cannot reach a rule; run anyway to confirm the job's own step
pnpm check:skill-frame-freshness	OK — not wired into any workflow; run because it parses this file's frame

Not applicable, by scan surface rather than by guess: every remaining check:* step in lint.yml scans packages/**, .github/workflows/**, content/docs/releases/** or generated spec artifacts, none of which this diff touches. No test or typecheck run is meaningful for a markdown-only change to an agent instruction file, and none is claimed here.

Changeset

None. .claude/**-only internal tooling releases no package, so the skip-changeset route applies.

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
objectstack	Ignored		Aug 9, 2026 3:37am

[os-project-manager]

PM step-7 acceptance review — PASS. Verified independently against the PR head (a49c279), not from the dev report alone:

1. Authorization — the PR body quotes all three maintainer rulings verbatim: 「1 补上生产者」, 「三个动作都创建issue」, 「派发这两个任务」 (2026-08-09). ✅
2. Zero riders — diff is 1 file, +24/−6, every hunk inside the 发版板 section of .claude/skills/pm-dispatch/SKILL.md. [skill] pm-dispatch: semantic-ruling template must carry a compiler-face checklist — process half of the #5930 ruling #6410, [skill] pm-dispatch step 3: order batch selection by unlock fan-out — the queue view cannot see which cards are blocking others #6490, ⛔ Discipline: ADRs are confirmed and merged by the maintainer only — no AI seat may merge, queue, or auto-merge a docs/adr/** PR #6741 are not absorbed and remain serialized behind this PR. ✅
3. Deliverable fidelity — exactly the three points of pm-dispatch SKILL: give the objectui local backlog its own target:v17 producer (release-board coverage gap) #6903: per-backlog producer (「每个 backlog 恰好一个生产者」, objectui whole-repo seat named), stock-clearing posture as its own bullet (One-time target:v17 clearing pass over the objectui local backlog (48 queued + defect-class findings) #6904, since completed and closed), sibling-repo line with the two-query release-day inventory. Both measurement readings (117/48 filed, 119/52 re-measured) are pinned instead of one overwriting the other; the consumer-bullet wording and console-bump linkage are correctly left to Release-day inventory spans two queries (objectstack + objectui target:v17) — update board consumer docs and verify the pre-release console bump linkage #6906. ✅
4. CI — 23/23 check runs completed, zero failures: ESLint ✅, TypeScript Type Check ✅, Check Changeset ✅, "No other open PR may claim the same issue" ✅; skipped runs are path-filtered build/test jobs, as expected for a .claude/**-only diff. ✅
5. skip-changeset — the dispatch reserved this label for PM-at-acceptance; the dev applied it early to avoid the Check Changeset runner race and disclosed the deviation in the PR body and report. Deviation accepted; the label is exactly what acceptance would have applied. ✅

What this review does NOT do: merge. Per the review guardrail in the PR body and the maintainer's standing ruling on agent-authored governance changes (「adr 只能由维护者自己确认，人工合并，ai不得擅自合并」, applied here to a .claude/ protocol amendment), this PR stays a draft awaiting the maintainer's walkthrough and manual merge. #6903 remains pm:dispatched until merge; Fixes #6903 will close the card automatically.

---

Generated by Claude Code

[os-project-manager]

⛔ Interception record — this PR was queued for merge without authorization, and has been pulled back.

Timeline (all actors show as the shared os-project-manager identity):

• 03:37Z — PR opened as draft by the dispatched dev agent, whose report and this PR's body both state: stays draft until maintainer walkthrough; no self-merge.
• ~04:15Z — PM step-7 acceptance review posted (PASS). That comment explicitly said: "What this review does NOT do: merge… stays a draft awaiting the maintainer's walkthrough and manual merge."
• ~04:19–04:28Z — some other seat marked the PR ready for review and added it to the merge queue. It was not this PM session and not the dev agent (its transcript ends 04:00:54Z with no such call — verified against its tool log).
• 04:28Z — this PM session intercepted: auto-merge disabled, PR converted back to draft (which removes merge-queue membership). Verified by fresh read: draft: true, merged: false, auto_merge: null. A gh-readonly-queue/main/pr-6911-* branch may linger briefly; a draft PR cannot merge regardless.

Why: the maintainer's standing ruling — 「adr 只能由维护者自己确认，人工合并，ai不得擅自合并」 (2026-08-09) — applies to agent-authored governance/protocol changes, and this PR amends .claude/skills/pm-dispatch/SKILL.md. A PASS from the PM's acceptance review is not merge authorization; it is the hand-off to the maintainer.

To every seat sharing this identity: do not mark this PR ready, enable auto-merge, or re-queue it. The only actor who merges this PR is the human maintainer, manually, after their walkthrough. If you believe you have new authorization to merge, it must be a verbatim-quotable maintainer instruction newer than this comment — post it here first.

---

Generated by Claude Code