docs(pm-dispatch,os-dev): report at draft-PR time, three-tier model policy, and the #7341 protocol text batch (PR A)

[os-help]

Fixes #6644, Fixes #7055, Fixes #7156
Part of #7341

PR A of the three-PR #7341 batch (A = protocol text, B = structure extraction, C = tooling). Two files: .claude/skills/pm-dispatch/SKILL.md (3013 → 3112 lines) and .claude/agents/os-dev.md (629 → 674 lines).

Maintainer authorization (verbatim, untranslated — Guardrails .claude/ tooling exception)

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

比如 更新 项目经理技能 必须要使用 Fable 5

确认,M 及以上默认云卡,并进 #7341

(Maintainer, devx seat's chat session, 2026-08-10; recorded on #7341, its claim comment, and the item-10 scope-addition comment. #6644 L2's own approval: maintainer ruling of 2026-08-10 05:10Z on #6644, plus the L2 shape decision recorded in the 07:11Z claim comment there.)

Per-item checklist (item | landing section | before | after)

#	Item	Landing section	Before	After
1	#6644 L2 dev half	SKILL step 5 prompt template	template tail said "you wait for CI to converge before reporting either way"	new template clause: report at draft-PR time, gate status recorded honestly (in_progress included), ⛔ no idle-poll; optional per-card 「本单等 CI」 override slot
2	#6644 L2 PM half	SKILL step 7 first review bullet + 入队与落地 B flip-timer paragraph	bullet framed the PM read as "复核侧的对账" against the dev's own CI wait	bullet rewritten: report-before-green is the expected shape, the PM's gate-job-conclusion read is the only gate reader, red gates route to a patch round; flip-timer paragraph marked as L2's PM half; B-over-D evidence (#6041/#6906, deaths between work-done and report) cited
3	#6644 L2 role-file half (also closes #7156)	os-dev.md Definition of done + "Local verification scope" + "Terminating cleanly"	"Wait for CI to converge before you return the report" bullet; scope-⚠️ said the local-farm trade is sound because "you still wait for CI"; terminating-cleanly sequence included "the foreground CI-convergence read"	reporting bullet: report at draft-PR time, ⛔ never sleep/timer-wait/idle-poll CI (#7156's measured cost cited: ~43 min / 141→168 calls / 222k→260k tokens, zero progress), per-card override restores the old contract; scope-⚠️ re-anchored on the PM's convergence read; sequence updated
4	Report-channel unification (batch item 3)	os-dev.md intro + "Terminating cleanly"; SKILL step 6 (subagent, cloud, Routine boundary)	report existed only as the subagent return message; the issue-comment marker was cloud-mode-only	report lands twice, GitHub first (issue comment opening with the os-dev-report HTML-comment marker, read back after posting); return message demoted to accelerator; step-6 collection reads GitHub first in both modes
5	Role-file precedence (batch item 5)	SKILL step 5 sunk-clauses paragraph; os-dev.md "Standard clauses live HERE"	placement of standard clauses read as editorial	measured precedence rule written in (#7055: a verbatim per-dispatch prohibition lost to the role file); unconditional clauses live in os-dev.md and are fixed there; per-card variables go through explicit interfaces; devs surface prompt-vs-role-file conflicts instead of picking silently
6	Triage size/model suggestion (batch item 6)	SKILL step 2 audit-trail bullet; step 4 Container-and-model note; Model tiering tail	no triage-time model input existed	optional-but-recommended "Size/model suggestion" line in the classification comment; executor consumes it in the claim comment and may override with a stated reason; cannot lower the mandatory fable clause
7	Multi-account slimming (batch item 7)	SKILL step 4 intro	shared-identity rituals read as unconditional	same-account scope note: session-ID/tiebreaker rituals scoped to within-one-account; cross-account rule stays one line (assignee not yours ⇒ taken); ⛔ rituals kept, not deleted
8	Decision-card re-check commands (batch item 8)	SKILL step 8 points 1 and 3	premise refresh required re-research	every premise line carries its own re-check command as a template requirement; analysis structure gains a premises slot
9	Model tiering rewrite (batch item 9)	SKILL step 5 "Model tiering" + claim-comment template + container-verdict examples	two tiers (sonnet/opus), 2026-08-09 authority	three tiers: floor sonnet (mechanical, gate-judged), default judgment tier opus, ceiling fable (claude-fable-5); mandatory claude-fable-5 for any card touching the PM skill, no discretion downward; the rulings above quoted verbatim as the authority block; resolution-order and env-var caveats kept unchanged; model recorded in the claim comment
10	M+ defaults to mode:cloud (scope addition, maintainer ruling of 2026-08-10, 「确认,M 及以上默认云卡,并进 #7341」)	SKILL Resource-limits sizing section + "Dispatch backends" + the Arguments mode: row	mode:subagent was the stated default ("Prefer this mode"); mode:cloud was a heavy-card exception behind a criteria list	M and above default to mode:cloud single-card dispatch; only S-grade mechanical cards stay mode:subagent; the old criteria list becomes examples of the M+ class, not the trigger list; the three companion disciplines (subscribe on creation, issue-comment report authority, archive_session at terminal state) named beside the default
11	#7055 cleanup prescription	os-dev.md resource rule 4	"git worktree remove --force" prescribed unconditionally	delete node_modules first, then unforced git worktree remove; a refusal then means uncommitted work — stop and read git status before considering the flag; cleanup stated as a step of the task

git diff --stat corresponds 1:1: exactly the two files above, nothing else.

Declared readings

• Dogfood SKILL's qualified --force left as-is (.claude/skills/dogfood-verification/SKILL.md §5): the tree it removes is a throwaway under /tmp whose commit is already pushed, and the force exists to unlock the local branch for --delete-branch — the os-dev cleanup prescribes git worktree remove --force, which disables the one guard protecting a parallel agent's uncommitted work #7055 hazard (clobbering a parallel agent's live worktree) does not apply to that path. Named here so the sweep visibly did not silently change it.
• [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 closure: fixed at the role-file layer per its own comment thread's evidence (a per-dispatch prohibition demonstrably loses to the role file), not via per-prompt lines. The L2 clause plus the reporting-item cost citation close it; no separate reporting-section line was needed beyond that citation.
• The three-axis escalation frame (guarded by check:skill-frame-sync) was deliberately left byte-untouched in both files.
• Item 10 landed as a follow-up commit (89e57ad) after the scope-addition comment arrived mid-batch; it propagated cleanly into PRs B and C (their SKILL.md counts moved to 3,011 and 3,016 respectively, ratchet still green).

Verification

• node scripts/check-nul-bytes.mjs — OK (both commits)
• pnpm check:skill-frame-sync (self-test 12/12 + real run) — 4 copies isomorphic, 3 axes, binding sentence present in all 4
• pnpm check:skill-compatibility (self-test 18/18 + real run) — 11 SKILL.md files reconciled, all majors match
• pnpm check:agent-model-declared (self-test 18/18 + real run) — os-dev.md → opus declared
• Leftover-contradiction grep for the retired dev-side CI wait: only historical references remain, rewritten for coherence

Notes for review

• .claude/-only ⇒ no changeset; skip-changeset applied.
• Ordering is load-bearing: PR B (structure extraction, Fixes #5925) is cut on top of this branch and re-derives its cut points from the post-A file; PR C (tooling) follows.
• This PR edits the operating skill of the PM seat that dispatched it — left draft; merge decision is the maintainer's/PM's per the batch card.

---

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 10, 2026 8:16am