[batch] PM-skill optimization: maintainer-accepted 9-point plan, consolidated with #5925 item 7 + #6644 L2 + #7055 + #7156

[os-help]

Maintainer-accepted optimization batch for .claude/skills/pm-dispatch/SKILL.md and its satellite files, consolidated per the maintainer's directive with the two in-flight skill cards. Filed and routed via the maintainer-direct-dispatch channel; authorization quoted verbatim (chat, this seat's session, 2026-08-10):

> 关于项目经理的技能,接受你的优化意见,并和前面两个优化任务集中处理。派发使用云卡片,用 Fable 5 模型

> 项目经理技能还需要考虑的是派任务时使用什么模型,也应该项目经理决定,最低下限 sonnet,最高可以 fable

The accepted proposal list is the devx seat's 2026-08-10 analysis (delivered in chat); items restated here as the card of record.

Consolidated scope

Inherited cards folded in (their own threads carry the detail):

• [skill] pm-dispatch: 5 lessons from the domain:metadata seat's first term (8 cards MERGED / 0 rework / 4 dev-side falsifications) #6644 L2 only — "report at draft-PR time, do not wait for CI" becomes a standard clause of the step-5 dispatch template, with the matching PM landing duty in step 7. L1/L3/L4/L5 already landed via PR docs(pm-dispatch): ADR merge prohibition in Guardrails + ACCEPT path fork; probe-and-revive as the standing backstop; replay-wake identity check (#6741, #6747) #6799.
• [skill] pm-dispatch:spec 车道 08-05/06 任期沉淀的 7 条 SKILL 更新建议(34 单 MERGED / 0 返工 / 一次交接全程即兴) #5925 item 7 only — extract the long incident narratives into references/incidents.md; pure move, byte-identical, cut points re-derived from the current file; ⛔ ordered AFTER all textual additions.
• os-dev cleanup prescribes git worktree remove --force, which disables the one guard protecting a parallel agent's uncommitted work #7055 — os-dev cleanup prescription drops --force (remove node_modules first, unforced git worktree remove; a refusal then means uncommitted work).
• [finding] os-dev agents idle-poll CI after their work is finished, burning the session limit and holding a seat's water level behind a task that is functionally done #7156 — the idle-poll symptom card of L2; fixed at the role-file layer, not per-prompt.

Accepted optimization items:

1. SKILL.md restructure: main file keeps rules + one-line case anchors; Operational notes narratives, incident case law, and the filter compile-surface inventory move to references/** loaded on demand; the case-law region becomes append-only; a shrink-only line-count ratchet gate guards the main file's size.
2. Half-state sweeper (report-only): a mechanical checker enumerating label/assignee invariant violations (pm:dispatched with no assignee; assignee with no claim comment; pm:queue+pm:dispatched both present post-landing; pm:blocked without a Blocked-by: body line; seat-sticker title/assignee/body triple out of sync). Complements [decision] engine-core lane drops pm:queue at dispatch without applying pm:dispatched — 7 instances of the ghost-card shape; pick the repair #7239 — its per-instance repairs stay with that card.
3. Report-channel unification: every dev final report also lands as an issue comment (the os-dev-report marker), both dispatch modes; the subagent return channel becomes an accelerator, not the source of truth.
4. Machine-produced dispatch inputs: a script mapping a file surface to the gate families that watch it (so "Local gates for this card" is generated, not recalled), plus a claim-comment shape checker.
5. Role-file precedence audit: clauses true for every card live in .claude/agents/os-dev.md (measured: a per-dispatch prohibition lost to the role file's own prescription — os-dev cleanup prescribes git worktree remove --force, which disables the one guard protecting a parallel agent's uncommitted work #7055 evidence); per-card variables get explicit interfaces in the role file instead of conflicting defaults.
6. Triage emits a size/model suggestion line consumed by the claim comment; the executor seat may override with a stated reason.
7. Multi-account slimming: rituals that exist only because of the shared-identity era are marked same-account-only; the cross-account rule stays one line (assignee not yours ⇒ taken).
8. Decision-card template: every premise line carries its own re-check command, so refreshing a stale decision card is running commands, not re-research.
9. Model-tiering rewrite (maintainer ruling above): the dispatch model is the PM's per-card decision; floor sonnet, ceiling fable (claude-fable-5), with opus the default judgment tier in between; model stays an explicit parameter on every dispatch; tier recorded in the claim comment.

Delivery shape (ordering is load-bearing)

• PR A — protocol text batch (SKILL.md + os-dev.md): items 3/5/6/7/8/9 + [skill] pm-dispatch: 5 lessons from the domain:metadata seat's first term (8 cards MERGED / 0 rework / 4 dev-side falsifications) #6644 L2 + os-dev cleanup prescribes git worktree remove --force, which disables the one guard protecting a parallel agent's uncommitted work #7055 + [finding] os-dev agents idle-poll CI after their work is finished, burning the session limit and holding a seat's water level behind a task that is functionally done #7156. Per-item checklist in the PR body; Fixes #6644, #7055, #7156 + Part of this card.
• PR B — structure extraction (item 1 + [skill] pm-dispatch:spec 车道 08-05/06 任期沉淀的 7 条 SKILL 更新建议(34 单 MERGED / 0 返工 / 一次交接全程即兴) #5925 item 7): pure moves, byte-identical, LAST; Fixes #5925 + Part of this card.
• PR C — tooling (items 2/4): new scripts + gate wiring; Part of this card (closes when accepted).

⛔ Constraints: no docs/adr/**; English on GitHub; sanitizer discipline on every SKILL.md write (entities for angle brackets, read back after posting); out-of-scope discoveries filed separately per PD #10.

Dispatch: cloud card (create_session), model claude-fable-5, per the authorization above.

[os-help]

Claim: PM loop round 1 (devx seat) — consolidated batch anchor
Session: session_01YS2qzDAn3CpWdY7uBX9bFQ
Branches: claude/issue-7341-pm-skill-batch-a (protocol text), claude/issue-7341-pm-skill-batch-b (structure extraction, LAST), claude/issue-7341-pm-skill-batch-c (tooling)
Worktree: cloud session's own container (dedicated worktrees per os-dev discipline inside it)
Domain: domain:devx
File surface: .claude/skills/pm-dispatch/SKILL.md, .claude/skills/pm-dispatch/references/** (new), .claude/agents/os-dev.md, scripts/ (new pm tooling + gate wiring only)
Container & model: L batch, mode:cloud dedicated container, model: claude-fable-5 (maintainer-directed)
Serial constraints cleared: the only in-flight claim on this surface was the predecessor's #5925+#6644 consolidation (branch claude/issue-5925-6644-skill-md — no branch or PR exists remotely at claim time); superseded into this batch by maintainer directive, recorded on both cards with provenance. #7055/#7156 were queued-unclaimed on .claude/agents/os-dev.md; folded here, which is also what serializes them.

Ruling addendum (maintainer, chat, 2026-08-10, verbatim) — a named floor for this card class:

> 比如 更新 项目经理技能 必须要使用 Fable 5

⇒ item 9's tier table gains a mandatory rule: cards that modify the PM skill (.claude/skills/pm-dispatch/**) dispatch at claude-fable-5, no discretion downward.

---

Generated by Claude Code

[os-help]

os-dev report — PR A of 3 (protocol text batch)

{
  "issue": 7341,
  "status": "done",
  "branch": "claude/issue-7341-pm-skill-batch-a",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/7357",
  "premise_still_valid": true,
  "summary": "PR A landed as draft #7357: all 9 protocol items + #6644 L2 + #7055 + #7156 in .claude/skills/pm-dispatch/SKILL.md (3013→3098 lines) and .claude/agents/os-dev.md (629→674). Per-item checklist in the PR body maps item | landing section | before | after; diff --stat is exactly the two files. Model tiering now: floor sonnet / default opus / ceiling fable, mandatory claude-fable-5 for pm-dispatch-skill cards, three rulings quoted verbatim as the authority block. skip-changeset applied (union write, size/m preserved).",
  "tests": "node scripts/check-nul-bytes.mjs OK (6739 files); pnpm check:skill-frame-sync self-test 12/12 + real run 4 copies isomorphic; pnpm check:skill-compatibility self-test 18/18 + 11 files reconciled; pnpm check:agent-model-declared self-test 18/18 + os-dev.md=opus; contradiction grep for the retired dev-side CI wait clean.",
  "open_questions": [],
  "out_of_scope_findings": []
}

Mechanism-assumption verifications for the batch (all three PRs build on these): no remote branch claude/issue-5925-6644-skill-md and no predecessor PR exist (git ls-remote verified at start); references/ does not exist yet; os-dev.md:150 carried the --force prescription; SKILL.md is 3013 lines on the dispatch-time main — the claim comment's ~2300 figure was stale, so PR B's cut points are being re-derived from the post-A file as instructed. PR #7292 merged before takeoff and is in the base. PRs B and C follow in order on top of this branch.

---

Generated by Claude Code

---

Generated by Claude Code

[os-help]

Scope addition — item 10 (maintainer ruling, chat, 2026-08-10, verbatim):

> 确认,M 及以上默认云卡,并进 #7341

⇒ PR A additionally rewrites the dispatch-backend criteria (the SKILL's single-container / mode:cloud selection clause): default becomes mode:cloud for M-and-above cards; only S-grade mechanical cards stay mode:subagent in the shared container. The existing heavy-card criteria (size/l·xl, full regen, wide verification radius, dogfood/browser, long heavy-verify lock) remain as examples of the M+ class, not as the trigger list. Three companion disciplines ride along unchanged and should be stated next to the new default: subscribe to the card's PR on creation; the dev report's authoritative channel is the issue comment (item 3 of this batch); archive_session at the card's terminal state. Rationale recorded from the maintainer's discussion: full process visibility and direct interactability outweigh per-card container spin-up for anything non-trivial.

Relayed to the in-flight cloud session via poke; its PR A checklist must carry this as item 10.

---

Generated by Claude Code

[os-help]

os-dev report — PR B of 3 (structure extraction)

{
  "issue": 7341,
  "status": "done",
  "branch": "claude/issue-7341-pm-skill-batch-b",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/7370",
  "premise_still_valid": true,
  "summary": "PR B landed as draft #7370 (Fixes #5925), stacked on PR A per the ordering discipline. Eight narrative blocks, 9,941 bytes, moved verbatim into references/incidents.md (append-only case-law file, 140 lines) and references/compile-surfaces.md (27 lines); SKILL.md 3,098 -> 2,997 lines. Every chunk script-extracted and script-verified: byte-identical in the pre-move blob, present in exactly one reference file, absent from the post-move main file (8/8 OK). Operational-notes bodies and the two-blindness explainer deliberately NOT moved — the pure-move condition fails there (rule and evidence interleave at sentence level); declared in the PR body. Line-count ratchet deferred to PR C with rationale.",
  "tests": "verify-move.mjs 8/8 chunks pre=true/ref=true/removed=true (9,941 bytes); check-nul-bytes OK; check:skill-frame-sync 4 copies isomorphic; check:skill-compatibility 11 files reconciled; check:agent-model-declared opus; deixis grep clean.",
  "open_questions": [],
  "out_of_scope_findings": []
}

---

Generated by Claude Code

---

Generated by Claude Code

[os-help]

os-dev report — PR C of 3 (tooling)

{
  "issue": 7341,
  "status": "done",
  "branch": "claude/issue-7341-pm-skill-batch-c",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/7379",
  "premise_still_valid": true,
  "summary": "PR C landed as draft #7379 (Part of #7341, Part of #7239), stacked on B. Three zero-dep scripts under scripts/pm/: check-half-states.mjs (report-only H1-H5 board patrol, exits 0 on any completed sweep so it cannot be wired as a gate), dispatch-gates.mjs (file surface -> check families, derived at runtime from 23 workflows / 75 discovered families, no hardcoded list per the #6492/#6865 lesson), and check-skill-line-ratchet.mjs wired into lint.yml as check:pm-skill-ratchet (shrink-only ceiling 3050 vs post-B 2997+2 pointer lines; placed in C not B to keep B a pure move — rationale in the PR body). SKILL.md gains the two planned one-line pointers.",
  "tests": "self-tests 16/16 + 14/14 + 5/5; ratchet real run green (3002 lines, headroom 48); check:workflow-status-functions self-test 34 assertions + real scan OK; eslint scripts/pm clean; check-nul-bytes OK. Declared limitation: live half-states sweep not runnable from this container (bare REST 401, no GITHUB_TOKEN; MCP-mediated access only) — the script correctly exits non-zero rather than reporting a clean board.",
  "open_questions": [],
  "out_of_scope_findings": []
}

All three batch PRs are now up: A #7357 (Fixes #6644/#7055/#7156) → B #7370 (Fixes #5925) → C #7379, stacked in that order; land in order and the later diffs collapse to their own commits.

---

Generated by Claude Code

---

Generated by Claude Code

[os-help]

os-dev report — addendum: item 10 (M+ defaults to mode:cloud) landed

The 07:57Z scope-addition ruling (「确认,M 及以上默认云卡,并进 #7341」) is implemented as commit 89e57ad on PR A's branch and propagated through B (724b810) and C (a2f6c8e):

• SKILL's sizing section rewritten: M and above default to mode:cloud single-card dispatch; only S-grade mechanical cards stay mode:subagent; the 2026-08-07 heavy-card criteria list is now examples of the M+ class, not the trigger list; the ruling is quoted verbatim in the section as its authority.
• "Dispatch backends" reframed accordingly (mode:subagent = the S-grade mechanical lane; mode:cloud = the default for M+), and the /pm-dispatch arguments table's mode: default cell updated.
• The three companion disciplines are named beside the new default: subscribe_pr_activity on creation, issue-comment report authority (batch item 3), archive_session at the card's terminal state.
• All three PR bodies updated: PR A's checklist now carries this as item 10 (the os-dev cleanup prescribes git worktree remove --force, which disables the one guard protecting a parallel agent's uncommitted work #7055 cleanup row moved to 11); B/C bodies carry the refreshed line counts (B: 3,112 → 3,011; C ratchet re-run green at 3,016/3,050, headroom 34). Gates re-run green on every branch (nul-bytes, frame-sync, ratchet).

Batch state: A #7357 → B #7370 → C #7379, all draft, all re-pushed, land in order.

---

Generated by Claude Code

---

Generated by Claude Code

[os-help]

ACCEPT — PR A #7357 (devx seat review, session session_01YS2qzDAn3CpWdY7uBX9bFQ); PRs B #7370 / C #7379 reviewed and held in serial order behind it.

Review basis (local diff of the three stacked branches, not the reports' claims): per-stage diff --stat matches each PR's declared surface exactly (A: SKILL.md+os-dev.md only; B: the extraction triple; C: three scripts/pm/* + lint.yml + package.json + two pointer lines). Spot-checked A's load-bearing sections on the branch: the three-tier model table (floor sonnet / default opus / ceiling claude-fable-5) with the mandatory fable rule for .claude/skills/pm-dispatch/** cards and all three maintainer rulings quoted verbatim; item 10's M+-defaults-to-cloud rewrite superseding the 2026-08-07 heavy-card criteria (kept as examples), with the companion disciplines named; os-dev.md's report-at-draft termination contract (dual delivery, GitHub authoritative) and the unforced-cleanup prescription with the refusal semantics stated.

Landing plan: A flips ready + auto-merge now. B (pure move, script-verified 8/8 byte-identical, declared non-moves for rule/evidence-interleaved passages — the honest reading of the pure-move condition) and C (report-only sweeper, runtime-derived gate mapping, shrink-only ratchet at 3050) stay draft until A is MERGED, then flip in order so each lands as its own diff. This seat merges nothing by hand — the queue does; walkthrough of the protocol changes remains open to the maintainer on all three PRs.

---

Generated by Claude Code

[os-help]

Batch complete — closing (devx seat, session session_01YS2qzDAn3CpWdY7uBX9bFQ).

All three PRs MERGED in order via the merge queue: A #7357 (protocol text: 10 items + #6644 L2 + #7055 + #7156 — closed those three cards), B #7370 (structure extraction, closed #5925: 9,941 bytes of incident case law moved byte-identically into references/incidents.md + compile-surfaces.md), C #7379 (tooling: check-half-states.mjs report-only sweeper, dispatch-gates.mjs runtime-derived gate mapper, check:pm-skill-ratchet shrink-only ceiling 3050 wired into lint.yml).

Net protocol state after this batch: report-at-draft is the standard posture (dual-channel delivery, issue comment authoritative); model tiering is floor-sonnet / default-opus / ceiling-fable with mandatory claude-fable-5 for PM-skill cards; M-and-above dispatches default to mode:cloud; worktree cleanup is unforced; SKILL.md sits at ~3,016 lines under a shrink-only ratchet with case law now append-only in references/. All four maintainer rulings are quoted verbatim in the landed text.

Remaining outside this card: #7239 keeps its per-instance repairs and repair-choice decision (PR C is Part of it — the enumerator only). The dispatching cloud session is being archived.

---

Generated by Claude Code