pm-dispatch SKILL Model tiering: mandatory claude-fable-5 for spec SEMANTIC cards — accept/reject behavior or public-surface change (maintainer ruling 2026-08-12)

[os-zhuang]

Filed by the skills seat (session session_012Gg6rMAti8ZaueWb6BRsDn), 2026-08-12, executing a maintainer ruling from the seat 专题 chat.

Provenance (maintainer, verbatim, untranslated)

本项目是协议驱动的,是否建议spec 协议相关的开发都是用 Fable 模型,我发现实际派发的时候很多都是 opus

同意,就按语义面收窄,立卡并通知 spec 席

The approved shape (recommended by this seat, accepted as-is): fable follows the criterion, not the domain:spec label.

The clause to add (Model tiering section)

• Mandatory claude-fable-5 for any card whose change alters the contract's accept/reject behavior or grows its public surface — the domain:spec semantic lane (schema shapes, contracts/**, strictness, the behavior half of retirements). Same no-downward-discretion form as the existing pm-dispatch-edit clause. Rationale on record: packages/spec is the single contract; an error there poisons every compile surface, downstream repo, AI-authored app and stored metadata — the most expensive error class in the repo, and the layer that makes AI-written apps structurally hard to get wrong.
• Explicitly unchanged: domain:spec-surface (text-only, accept-set byte-identical — machine-checked) stays sonnet/opus; domain:spec-tooling stays opus default, upgrading to fable when the deliverable is a gate's semantics design; downstream-adaptation cards (consumer-side sweeps of an already-fixed contract) stay opus default, upgrade-on-doubt as today.
• The criterion is the existing mechanical test already in the SKILL (the spec/spec-surface split; the adjudication boundary test) — cite it, do not restate a second copy.

Landing notes

• Blocked-by: 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
• Same-file serial: SKILL.md hot-file queue position 2, behind the in-flight 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 batch PR₁ (which is editing the same Model-tiering/backend region — dispatching earlier guarantees rework). ⚠️ Ratchet headroom 0 — offset in-PR.
• Effective immediately ahead of the text landing, via cross-seat notice on the spec semantic seat post ([PM seat] domain:spec — 🟢 os-zhuang #6017) with the ruling verbatim.
• Dispatch model: claude-fable-5 (mandatory clause — this card edits .claude/skills/pm-dispatch/**). Size S — subagent lane per the 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 revision.
• ADR-class: draft PR, human merge.

[os-zhuang]

Batch conversion — this card becomes the carrier of the closing consolidated dispatch (maintainer prompt in seat chat, 2026-08-12, verbatim: 「剩余的 skills 是否应该一个批次处理完。」; sequencing is the PM's call).

Composition: ONE os-dev subagent (mode:subagent per the #7944 revision, model: claude-fable-5 mandatory), ONE PR — Fixes #7945, Fixes #7947 — landing both ruled S-grade SKILL.md deltas (this card's spec-semantic fable clause + #7947's issue-typing/Bug-tiebreaker) with a per-member checklist and ratchet-0 in-section offsets.

Dispatch trigger: PR #7946 (the in-flight batch PR₁) reaching MERGED — both deltas edit the sections it rewrites. Blocked-by: #7944 stands until then; #7947 is hereby a member, not an independent candidate.

---

Generated by Claude Code

[os-zhuang]

Batch composition update (maintainer approval 2026-08-12, verbatim: 「同意」): #7950 (retire the mandatory shift-end SKILL-suggestion list; three report classes, default zero) joins as the third member. The closing batch is now: one subagent (model: claude-fable-5), ONE PR — Fixes #7945, Fixes #7947, Fixes #7950 — per-member checklist, ratchet-0 offsets (note #7950 is net-negative). Dispatch trigger unchanged: PR #7946 reaching MERGED.

---

Generated by Claude Code

[os-zhuang]

Claim: PM loop round 5 (skills seat) — closing consolidated batch carrier
Session: session_012Gg6rMAti8ZaueWb6BRsDn
Branch: claude/issue-7945-pm-skill-closing-batch (one PR: Fixes #7945 / #7947 / #7950)
Worktree: dedicated worktree inside the PM session's container (subagent backend, per the now-landed revision)
Domain: domain:skills
File surface: .claude/skills/pm-dispatch/SKILL.md (Model tiering + triage duty + selection sections) + references/seat-post-protocol.md (seat-exit checklist). Stop on breach; explain in the report.
Container & model: S docs batch, mode:subagent, model: claude-fable-5 (mandatory clause — no downward discretion)
Serial constraints cleared: predecessor PR #7946 MERGED (merge commit 6ff179d verified on origin/main; SKILL.md = 686 lines, backend revision live); PR #7948 next in the merge queue touches AGENTS.md/CLAUDE.md only — no file overlap with this batch; zero other open PRs touch the pm-dispatch tree (checked at claim time). Members #7947 / #7950 absorbed per the composition comments; #7945's Blocked-by: #7944 cleared by the merge.
Landing: ADR-class — draft PR, human merge.

---

Generated by Claude Code

[os-zhuang]

Dispatched (PM loop round 5, skills seat): os-dev subagent, model: claude-fable-5, launched 2026-08-12 ~08:4xZ. One draft PR expected (Fixes #7945 / #7947 / #7950), branch claude/issue-7945-pm-skill-closing-batch off 6ff179d+. All three deltas fully ruled; #7950's retirement funds the line budget at ratchet headroom 0. Collection: os-dev-report comment here + direct return channel; ADR-class terminal (human merge).

---

Generated by Claude Code