docs(pm-dispatch): dispatch token economics — sonnet tier for S-class cards, scoped local gates, slimmed prompts (#6863)

[os-project-manager]

Fixes #6863

Three coupled changes to the dispatch machinery, all in .claude/. Nothing ships to
users; this PR releases nothing and writes no changeset.

Maintainer authorization

Per the Guardrails exception for .claude/ internal tooling PRs (#5925-6, precedents
#5597 / #5872 / #6650), the authorization is quoted here verbatim and untranslated.
Maintainer, in-session, 2026-08-09:

全部同意,先开任务

Approving the domain:spec seat's token-economics assessment (session
session_011sGk4SKHqGRgmmqUok1P8M). Review stays with another seat or a maintainer
walkthrough — this PR is a draft, auto-merge is not armed, and the seat that
implemented it will not merge it.

Item 1 — model tiering, and the precedence finding the card asked for

The finding first, because the card flagged it as the one thing that could make item 1
inert: it does not. The per-dispatch argument wins.

Claude Code resolves a subagent's model in four steps, verified 2026-08-09 against the
subagent documentation rather than inferred from the repo's own prose:

1. the CLAUDE_CODE_SUBAGENT_MODEL environment variable, when set
2. the per-invocation model parameter on the Agent call
3. the agent definition's model: frontmatter
4. the main conversation's model

So .claude/agents/os-dev.md's frontmatter model: opus (#6688, guarded by #6836's
check:agent-model-declared) does not defeat a dispatch that passes
model: "sonnet" — step 2 outranks step 3. The pin governs only the case where the
dispatcher passes nothing, which is exactly the #6686 incident it was written for. Item 1
takes effect as written.

Two adjacent traps are recorded next to it, because both are silent:

• CLAUDE_CODE_SUBAGENT_MODEL outranks both the argument and the frontmatter. It is
  not set in the dispatch container (checked 2026-08-09), but if it ever is, it
  overrides every tier decision the PM makes and nothing in this repo would show it.
• A tier blocked by the org's availableModels allowlist does not fall back to the
  frontmatter pin — it falls back to the inherited model, i.e. straight back into the
  silent-inheritance failure the pin exists to stop.

On the contradiction the card named. The old rule existed in two places: step 5's
「Model split (maintainer policy): pass model: "opus" on every dev dispatch」 and the
08-08 handover's 「所有派发至少 opus」. The second is a handover note, not a file — grep
finds no copy of it in the tree — so only the first was editable. Rather than quietly
swapping the sentence, the new §5 names the old wording, quotes it, and declares it
废止, and instructs the reader to override any 「一律/至少 opus」 they meet elsewhere
(old handover notes, seat stickers, second-hand retellings). A reader cannot come away
holding both.

Tiering itself: S-class mechanical cards go to sonnet, where the criterion is "the
gate farm decides correctness mechanically", not "the diff is small"; M/L, ruling
implementations, multi-face semantic cards and anything with design judgment stay on
opus; ties round up. model is passed explicitly on every dispatch rather than
leaning on the pin, so the tier is a property of the dispatch and not of a file's current
state. The claim comment's container-assessment line now carries the tier beside the
size, since both come from one reading — with an explicit warning that "S-class" does not
imply sonnet on its own.

Item 2 — local gate scoping, with the safety half kept mandatory

os-dev.md gains a Local verification scope section: build closure, the affected
packages' suites, and the gate families touching the card's surface — instead of
enumerating all 55+ check:* steps out of lint.yml. The dispatch template now names
the relevant families per card.

The card accepts an occasional extra push-fix lap as the cost, so the rule states that
cost and then pins the half that makes it safe. Written into os-dev.md verbatim:

This is not licence to report before CI converges — it is the opposite: the local farm
was the thing that could be dropped precisely because the CI wait cannot be.

And reconciled on the PM side at step 7: the CI-convergence check moves from "the second
of two safety nets" to the only one, so a non-obvious gate reddening in CI is now the
expected shape rather than an anomaly, the extra lap is a price already paid rather
than grounds for REWORK, and "local green" in a report covers a smaller surface than it
used to. Kept intact per the card: build-first, prefix-direction consumer sweep, reverse
verification, and CI convergence before reporting.

Item 3 — slimmed prompts, boilerplate sunk to the producer

• The template no longer pastes the issue body. It instructs the dev to read the full
  body and all comments on GitHub, which premise-first already forced. It still
  pastes what GitHub reading cannot supply: ruling quotes verbatim (Chinese
  untranslated), the 裁决 / PM-机制假设 partition, card-specific clauses and same-day
  churn.
• Mitigation for the notes-12 truncation risk is written as the price of the risk
  transfer, not a courtesy: pasting made truncation the PM's risk, self-reading makes
  it the dev's, so the prompt requires a completeness self-check and a double-read on
  suspicion. Precedent [移交自 objectui] ComponentPropsMap 未声明 objectui 渲染器实读的 5 个顶层 prop(page:header ×3 / page:tabs.tabStyle / page:accordion.variant)—— #5775 的清点未完 #6776 is cited.
• The unconditional non-negotiables move into .claude/agents/os-dev.md once
  (producer-side fix, the PD#12 instinct the SKILL itself cites): build-first ordering,
  filter direction, cross-package reverse verification, the authorable-surface anchor
  rules with the gen:schema MERGE-state ban, and the Fixes vs Part of judgment.
  Worktree discipline, ADR-0112 rejection assertions, foreground posture, English policy
  and the report contract were already there and were not duplicated.
• Conditional standard clauses — the ones carrying 适用判据 (multi-face shared
  conformance coverage, rejection-case code+status, the filter compile-face
  inventory) — deliberately stay in the SKILL and are still pasted per card. They are not
  universal, so sinking them would attach them to cards they do not apply to.
• os-dev.md says plainly that a prompt's silence about a standard clause is the
  expected shape and never permission.

Overlap with PR #6799

#6799 is merged — cd704cc43, 2026-08-09T01:50:17Z — so the card's premise that it
is open and being held for a human merge had already expired by dispatch time. I branched
from origin/main with #6799 in it rather than composing around it.

Region overlap: none. Hunk ranges do not intersect at any point:

	regions touched
#6799	1820-1855, 1846-1875, 1927-1957, 2090-2099, 2101-2131, 2539-2568
this PR	1302-1310, 1368-1394, 1395-1441, 1453-1459, 1685-1693, 1999-2005

Verified by content, not by line arithmetic: every line #6799 added is still present
verbatim in the file, checked by replaying its added lines through grep -qF
(zero missing). All four of its additions are untouched — the ADR merge prohibition in
Guardrails, the step-7 ACCEPT path fork, probe-and-revive as the standing backstop, and
the replay-wake identity check. My nearest edit (the step-7 CI-convergence bullet, in the
review checklist) sits above the ACCEPT verdict section where #6799's fork lives.

Deliberately NOT done

• ⛔ No docs/adr/**, no content/docs/releases/, no code, no AGENTS.md.
• The card's "Already in effect (seat practice)" list is explicitly marked
  "no file change needed" — batched seat-post refreshes, terse ACCEPT comments,
  thread-tail race re-reads, sweep-first S-card batching. Not written.
• Kept untouched per the card, each carrying incident receipts: claim protocol, race
  checks, premise-first verification, write-after-read, serial landing relay, and the
  45-minute patrol cadence. The claim comment gained one line and lost nothing; the
  「最后一行」 back-reference that the [skill] pm-dispatch:services 车道 08-05/06 班次交接沉淀的 10 条 SKILL 更新建议(28 PR / 两次 CI 红 / 三次前提证伪) #5885 serial-constraints sentence depends on was
  preserved by inserting the new line above it, not below.
• scripts/check-agent-model-declared.mjs's header describes the resolution as "three
  steps" and omits CLAUDE_CODE_SUBAGENT_MODEL. Accurate in effect today (the variable
  is unset) and outside this card's declared file surface, so it was not edited — filed
  separately as an observation.

Verification

No package code changed, so pnpm test / pnpm typecheck are not meaningful here —
saying so beats manufacturing template-shaped evidence. What was run, serialized under
the shared verification lock with the capped heap, is every gate whose scan surface
includes .claude/ (found by grepping scripts/*.mjs for the path rather than from
memory), plus the repo-wide prose and changeset gates. 8 gates, all green, each with
EXIT=0:

• check:agent-model-declared — 18 self-test cases pass; os-dev.md → opus, 0 justified inherits
• check:skill-frame-sync — 12 self-test cases; 4 copies isomorphic across 3 files, 3 axes, 42 files scanned
• check:skill-frame-freshness — 12 self-test cases; frame current with origin/main
• check:skill-compatibility — 18 self-test cases; 11 SKILL.md reconciled against 77 packages
• check:nul-bytes — 56 assertions; 6403 tracked text files, no raw ASCII control bytes
• check:doc-authoring — 373 files clean
• check:role-word — 44 baselined files, no new occurrences
• check:empty-changeset — no empty-frontmatter changeset introduced

Byte discipline: a manual control-byte self-scan beyond the gate's surface
(grep -naP over the two edited files, the full 0x00-0x1f plus 0x7f range) returned
no matches. This PR is dogfooding its own item 2 — the gates above are the families
touching its surface, not the whole farm.

---

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 2:06am