docs(pm-dispatch): subagent default for S+M with per-dispatch model tiering; seat-post read boundary

[os-zhuang]

Fixes #7944
Fixes #7942

ADR-class (skills root): draft PR, human merge — no auto-merge, no queue entry.

Per-member checklist

• pm-dispatch SKILL: M-grade cards default to the subagent backend with per-dispatch model tiering — cloud reserved for L/XL, long-lived, browser-verify (maintainer ruling 2026-08-12) #7944 (backend/sizing revision) — landed in .claude/skills/pm-dispatch/SKILL.md: the mode: row of the argument table and the 「资源与后端」 paragraph of the dispatch section now carry the 2026-08-12 ruling (S-mechanical + M ⇒ mode:subagent with per-dispatch model tiering; mode:cloud reserved for L/XL, work that must outlive the PM session, browser/dogfood verification, and build-heavy M by explicit judgment; archive duty narrows to cloud cards). The cloud-vs-subagent trade-off rationale (three fixed taxes vs the one real cost and its mitigations) is recorded on demand in references/dispatch-runbook.md under the cloud-lessons section. The mandatory claude-fable-5 clause for cards editing the pm-dispatch skill is unchanged; the model-resolution facts (env var → per-call param → frontmatter → inherit) were already present in the Model tiering paragraph and stay.
• pm-dispatch SKILL: seat-post reads cover only comments newer than the last body edit (token-waste fix, maintainer-approved) #7942 (seat-post read boundary) — one principle line added to the seat-post section of SKILL.md (takeover and patrol read the post body plus only comments newer than the body's last edit — older comments were absorbed by that edit and are archive, not state), with the fuller rule (rationale + how to compare the timestamps) in references/seat-post-protocol.md, which loads on demand at takeover time.

Ratchet accounting (ceiling 686, headroom 0)

The file stays at exactly 686 lines. Additions were offset inside the touched sections:

• Backend paragraph: rewritten 5 → 6 lines; dropped the sentence restating os-dev's parallel-memory discipline (it is normative in .claude/agents/os-dev.md, not here) and the redundant "size is not a sufficient criterion" clause (implied by "S-but-not-mechanical gets M treatment", which stays).
• Cloud-lessons paragraph: 11 → 9 lines; the self-drive clause's expansion ("never end a turn for questions / mid-task status") and the roster mechanism prose were compressed — both remain spelled out in references/dispatch-runbook.md.
• Model tiering paragraph: reflowed 11 → 10 lines; dropped 「多面语义」 from the opus examples (subsumed by 「任何带设计判断的卡」).
• Seat-post paragraph: read-boundary principle folded in at +2 lines; the token-tax rationale lives in the reference file only.

No semantic loss beyond the rationalized drops listed above; no issue-ID citations added (id-lint green).

Gates (local, honest readings)

• check:pm-skill-ratchet — green (686 lines, ceiling 686, headroom 0; self-test 5/5)
• check:pm-skill-id-lint — green (8 files clean; self-test 9/9)
• check:skill-frame-sync — green (4 copies isomorphic; self-test 12/12)
• check:doc-authoring — green (375 files clean; self-test green)
• check:nul-bytes — green (7279 text files scanned; self-test 75 assertions)

---

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 12, 2026 8:07am

[os-zhuang]

人类同意合并