os-dev quota shaping: PM-supplied dedup reading rides the dispatch prompt; dev search-before-file goes through the REST channel — one dev measured ~5,658 GraphQL points in ~15 minutes

[claude]

Authorized: maintainer 2026-08-23, live PM chat — options ①②③ approved verbatim as 「1+2+3」. Filed by the skills seat (session 5213b871-5164-5bc3-8874-28b336bbcd40), self-triaged under the skills-lane exception. This card is option ②.

Measured facts

• The biggest GraphQL consumer in the fleet is the dispatched dev subagent, not PM patrol: one dev burned ~5,658 points in ~15 minutes (a whole hourly seat bucket), dominated by dedup/search reads. PM inventory reads are noise by comparison.
• Every dev in a batch independently re-scans the same dedup surface: N devs = N copies of the same expensive read.
• The dev-side curl REST channel is proven writable AND readable from dev seats (the skip-changeset additive-endpoint clause in os-dev.md is the existing precedent spelling).

Change (three halves, smallest edits that close the class)

1. Dispatch-word standard clause (references/dispatch-runbook.md 派发词构造细则): the PM takes ONE dedup/board reading at dispatch time and hands it to every dev in the batch; the dev re-verifies only its own increment (new cards since the reading's timestamp), never re-runs the full scan.
2. os-dev.md 立单纪律 (rule 3 search-before-file): the search goes through the curl REST channel (list endpoint + keyword grep), not MCP list/search — same reason the skip-changeset write already goes REST. os-dev.md is a protocol semantic face ⇒ clause ① fable-mandatory construction.
3. review-checklist one row: spot-check git-first compliance (branch existence, content reads via git — zero quota) when a report's evidence shows API reads where git would answer.

Ceilings: dispatch-runbook 242 headroom 0 after the restructuring round; review-checklist 82 headroom 0 — payment by in-place compression per the ratchet header.

Serial constraint

Queues BEHIND the #11086 restructuring PRs (same hot-file chain). Related: option ① card (REST-first reads) and option ③ card (right-sized reads + pacing), filed in the same stroke.

---

Generated by Claude Code

[claude]

Trio index (one ruling, three cards): #11364 = option ① REST-default reads · #11365 = option ② dev usage shaping · #11366 = option ③ right-sized reads + mutation pacing. All three queue behind the #11086 restructuring PRs on the pm-dispatch hot-file chain; ①+③ share the platform-readings quota section — fold-or-serial must be answered at claim time (five-gate test likely passes for folding ①+③).

---

Generated by Claude Code

[os-zhuang]

Claim: skills seat dispatch (session session_01RMTpSRF5CjMmQBFfPtPCwJ). Serialization released: the #11086 hot-file chain and the trio's ①+③ fold (PR #11749) are both MERGED to main; premise re-verified against the landed rest-channel.md / platform-readings.md this stroke.

• Branch: claude/issue-11365-osdev-quota-shaping
• Worktree: objectstack-11365-quota-shaping
• Ruling being implemented: maintainer 2026-08-23, verbatim 「1+2+3」 — this card is option ② (dev usage shaping): the PM-supplied dedup reading rides the dispatch prompt, and the dev's search-before-file goes through the REST channel.
• File surface: .claude/agents/os-dev.md (the dev-side channel rule) and .claude/skills/pm-dispatch/references/dispatch-runbook.md (the PM-side prompt-shaping row). ⚠️ rest-channel.md sits at ceiling 46/46 headroom 0 and platform-readings.md at 130/130 — compose with them BY POINTER; any addition there needs a ceiling raise argued in the PR body, and preference is to need none. Stop on breach.
• Governed face (.claude/**): draft PR + human merge (PD feat: Comprehensive CRM example demonstrating all ObjectStack protocol features #14). Clause-②: no.
• Disjointness note: [finding] Census of all 47 pm:blocked cards: 32 park Blocked-by: outside the body (invisible to the unlock scan), 16 wait on already-closed upstreams, 7 name no blocker at all #11747's fold PR edits state-machine.md and patrol scripts, not these two files; if implementation pressure pushes either dispatch into the other's surface, stop and report rather than widen.

---

Generated by Claude Code

[claude]

{
 "issue": 11365,
 "status": "done",
 "branch": "claude/issue-11365-osdev-quota-shaping",
 "pr": "https://github.com/objectstack-ai/objectstack/pull/11870",
 "premise_still_valid": true,
 "summary": "Option 2 of the 2026-08-23 「1+2+3」 quota ruling landed as two shaping rules in the two files the claim comment declared. PM side (dispatch-runbook.md 派发词构造细则): the gate-list row and the new dedup/board-reading row merged into ONE row under one shared judgement — a reading the PM already holds is handed down once (query + timestamp, whole batch shares it; dispatch word says treat as given, re-verify only the increment, never re-run the scan), N devs never re-run it. Dev side (os-dev.md rule 3 立单纪律): search-before-file and list reads go through the REST channel (list endpoint + local grep), stated ONCE and composed BY POINTER with references/rest-channel.md, carrying the dev half of the PM rule in the same clause. Both files land exactly at their ceilings (os-dev.md 403->405 using its 2 lines of headroom; dispatch-runbook.md 242->242 paid in place). ZERO edits to rest-channel.md (46/46) and platform-readings.md (130/130); state-machine.md, the patrol scripts and the ratchet script itself untouched; no ceiling raised.",
 "tests": "All at final commit 7b246a30 (worktree clean, tree identical to the commit); every exit code captured with EXIT=$? on a redirected command, never through a pipe. Gate set derived with `node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack` (no path args; header confirmed 'gate list derived from the tree of objectstack-ai/objectstack at commit c4db3116' and '--repo ... it holds'; change set taken from git = the 2 files). 9/9 matched families green — verdict lines each gate printed: check:pm-skill-ratchet '✓ check-skill-line-ratchet: .claude/agents/os-dev.md is 405 lines (ceiling 405; headroom 0).' + '✓ ... dispatch-runbook.md is 242 lines (ceiling 242; headroom 0).'; check:agent-model-declared '✓ ...: 1 agent definition(s) under .claude/agents/ all declare a model'; check:skill-frame-sync '✓ ...: 4 copies of the decision frame are structurally isomorphic across 3 files'; check:pm-skill-id-lint '✓ check-skill-id-lint: 22 file(s) clean (pattern /#[0-9]{3,}/g).'; check:doc-authoring '✓ doc authoring guard: 389 files clean — no bare metadata literals.'; check:nul-bytes 'check-nul-bytes: OK (scanned 6601 text file(s) ... no raw ASCII control bytes).'; check:agent-test-spelling '✓ check-agent-test-spelling: 0 violations — 357 file(s)'; check:pm-governed-merges '✓ check-governed-merges --self-test: 129 assertions'; check:doc-formula-expressions '✓ check:doc-formula-expressions: 22 record-scoped formula example(s) across 421 files / 1448 TS blocks judged clean by @objectstack/formula.'. NOTE: doc-formula-expressions first exited 1 twice with PREREQUISITE NOT MET (@objectstack/formula, then @objectstack/lint not built — 'Nothing was measured'); both built via the shared verify lock (VERDICT command-exit 0), gate then green. Repo-wide `pnpm lint` NARROWED to the changed files, declared and measured on three counts: (1) population read from eslint's own config — both files return 'File ignored because no matching configuration was supplied', i.e. no .md file is judged by this config; (2) --format json reports 2 files judged, 0 errors, 2 'ignored' warnings; (3) invariance — the diff contains no TS/JS and no eslint config change, so no untouched file's verdict can move. Reflow identity proof (the runbook's payment): whitespace-normalized old lines 219-221 vs new 220-221 compare EQUAL, 121 chars each side — the 文件面两句原文 bullet lost a line without losing a byte of text. No ablation was run: this diff adds no gate and mutates no runtime path, so there is nothing whose failure mode a mutation could demonstrate.",
 "gates": "9 named families, all green at 7b246a30: check:agent-model-declared, check:agent-test-spelling, check:doc-authoring, check:doc-formula-expressions (after building @objectstack/formula + @objectstack/lint), check:nul-bytes, check:pm-governed-merges, check:pm-skill-id-lint, check:pm-skill-ratchet, check:skill-frame-sync. Re-derivation against the ACTUAL changed paths added nothing (same 2 paths, same 9 families). CI convergence is the PM's read per the 2026-08-10 ruling; no changeset — .claude/**-only, declared with the skip-changeset label on PR #11870 (additive POST .../issues/11870/labels, HTTP 200, read back).",
 "deviations": [
 "Card half ③ (a review-checklist.md row spot-checking git-first compliance) was NOT implemented: the claim comment's declared file surface is os-dev.md + dispatch-runbook.md only, review-checklist.md sits at 82/82 headroom 0, and the dispatch prompt named two shaping rules. Left for PM triage — see open_questions.",
 "The runbook's line payment includes a PURE REFLOW of an unrelated bullet (文件面两句原文, 3 lines -> 2, byte-identical text). It widens the diff by one bullet that changed no content; declared here and in the PR body with the mechanical equality proof rather than left to look like an edit.",
 "The measured story (~5,658 GraphQL points / ~15 min) is NOT copied into either instruction file — the ratchet header rules that stories live on cards, not in operational text. os-dev.md carries the one-clause reason ('GraphQL 池是舰队最紧的桶'); the number lives on this card and in the PR body.",
 "dispatch-gates printed 'Model tier — MANDATORY: claude-fable-5' for .claude/agents/os-dev.md (clause ①, protocol-semantics face). This dev ran on opus and the claim comment carries no `Container & model` exit line. Flagged, not acted on — the seat/model choice is the PM's."
 ],
 "open_questions": [
 {
 "question": "Half ③ of the card (review-checklist row: spot-check git-first compliance when a report's evidence shows API reads where git would answer) is unimplemented and review-checklist.md is at ceiling 82/82. Where should it land?",
 "options": [
 "A: file it as a follow-up card for the review-checklist seat, paid by in-place compression there (this PR stays exactly the surface the claim declared)",
 "B: add it to THIS PR — needs the review-checklist file surface, which the claim did not declare and a sibling wave may hold",
 "C: drop it — the PM already reads the dev's `tests` evidence, so a checklist row may be redundant"
 ],
 "recommendation": "A. The claim comment is the serialization record other agents read; widening it mid-flight is exactly the collision the claim protocol exists to prevent, and the row is cheap to land on its own once someone can pay 82/82's compression. C is wrong: the row's value is that it makes a silent regression (dev reverting to MCP reads) visible at ACCEPT time, which is not something the reading itself surfaces."
 }
 ],
 "out_of_scope_findings": []
}

---

Generated by Claude Code

[os-zhuang]

ACCEPT — skills seat (session session_01RMTpSRF5CjMmQBFfPtPCwJ), PR #11870 at head 7b246a30, verified against the actual diff, not the report's prose.

• The two shaping rules land exactly as the 「1+2+3」 ruling's option ② prescribes, one home each: os-dev.md rule 3 carries the dev-side channel rule (REST list endpoint + local grep, ⛔ no MCP list_issues/search_issues, composed BY POINTER with rest-channel.md) plus the treat-as-given clause for the PM-supplied dedup reading; dispatch-runbook.md merges the gate-list and dedup rows under the one shared judgement (a reading the PM holds is handed down once, N devs never re-run it), preserving the pnpm rider and the 文件面两句原文 bullet byte-identically (reflow equality proven in the report).
• Ceiling discipline held the hard way: both files land AT ceiling (405/405, 242/242), zero edits to the two 0-headroom files, no raise anywhere. 9/9 derived gate families green by their own verdict lines; skip-changeset applied and read back.
• Tier note: dispatch-gates flags os-dev.md as a clause-① protocol-semantics face (mandatory tier claude-fable-5); the implementing dev ran below it and flagged that honestly. This ACCEPT is the tier-qualified review — reading this seat's own last_served_model = claude-fable-5 (= CONTRACT_REVIEW_TIER at origin/main dispatch-gates.mjs:3070) and the full diff above.
• Open question ruled: A — the review-checklist spot-check row is filed as its own follow-up card (paid by in-place compression at 82/82), keeping this PR exactly the claimed surface. Option C rejected for the report's own reason: the row's value is making a silent dev regression visible at ACCEPT time.

PR stays draft — governed face, human merge; it joins the maintainer's merge list.

---

Generated by Claude Code