[Decision] Three spec contract dev agents in one container queued ~70 min each on the shared verify lock: leave package test:repo to CI, and send build-heavy cards to cloud sessions?

Ruled: 6184828752 · letter A/A · 2026-10-15T05:03Z

Filing gate ② a decision only the maintainer can make: the maintainer asked for it in the domain:spec seat 2 chat on 2026-10-15, verbatim 「技能修正开决策卡」, after the seat measured the lock contention below.

Filed by: domain:spec seat 2 ([PM seat] domain:spec · seat 2 — 🟢 zhuangjianguo · session_019oFY6Hi35J7atvgVFnmoaK #18549) · os-bill · session session_01LFqdRsuM5nkqoJdHd5ztVk. ~~⛔ Not a claim.~~
Lane and grade are triage's. The surface is .claude/** (Tier S, the skills lane).

维护者速读

这一班派出的三张 spec 契约卡(#23354、#23488、#23543 F1)从派发到交报告都超过 2 小时。

不是 CI 慢:merge queue 的 CI 中位数最近一周稳定在 22–24 分钟,而这三张卡在那 2 小时里大部分时间还没开 PR。
时间耗在容器里那把共享验证锁上:过去 2 小时它被占用了 89%,三个开发 agent 累计排队约 3.5 小时,另有 7 次排队超时。
占锁最多的两类:
- 包级 test:repo(全仓扫描类测试):2 次就占了 26%。CI 本来就会跑它,本地重复。
- 三张都要反复构建 spec 的重卡挤在同一个容器里(闭包构建占 25%)。技能允许重卡走云卡(每张卡独立容器),但没给判据,席位判成了普通 M 卡。

要你定两件事:

F1:本地不跑 test:repo、交给 CI,写进开发 agent 的规则吗?
F2:「build 重」的卡给一个可量的判据,命中就走云卡吗?

请回一行:F1 A/B · F2 A/B/C。席位推荐 F1 A · F2 A。

Background (measured)

Lock ledger (/tmp/os-heavy-verify.lock.ledger in the seat's container), 2026-10-15 01:34Z–03:33Z, 7155 s:

59 runs; the lock held for 6396 s (89%); 12 429 s of waiting summed over runs; 7 queue-timeout runs (waited out, decided nothing); acquisition wait p50 238 s, p90 540 s.
Hold time by command class:

class	held	runs
`pnpm --filter @objectstack/spec test:repo` (with or without `typecheck`)	1677 s	2
`turbo run build` dependency-closure builds	1618 s	8
other	982 s	26
other packages' tests	640 s	9
`spec` local `test`	623 s	3
`spec` build (with its generators)	537 s	6
typecheck	319 s	5

Dispatch to report: #23354 was dispatched at 01:26Z and reported at 04:00Z. The subagent ran 9134 s.

The machine: 4 cores, 15 GB. packages/spec's build runs its DTS pass with --max-old-space-size=6144.

CI is flat (ci.yml, merge_group, run duration):

window	p50	p90
09-28..30	18.4 m	22.6 m
10-03..05	34.8 m	38.8 m
10-08..10	21.9 m	26.1 m
10-13..15	22.5–24.1 m	28.8–30.0 m

CI already runs every package's test:repo: .github/workflows/ci.yml:965 runs pnpm turbo run test test:repo $FILTERS. scripts/pm/dispatch-gates.mjs --commands derives no test:repo command for a spec path: 70 commands for packages/spec/src/data/driver/config-registry.zod.ts, none of them test:repo. So the local runs were the agents' own reading of "the affected package's tests".

Governing text

.claude/agents/os-dev.md:130 — 「本地验证范围 —— 本地只跑定向门禁,全农场归 CI」.
.claude/agents/os-dev.md:138 — 「③ 读到改动文件的门禁族 = --commands 派生清单(按 diff 收窄);派发词清单与全农场归 CI。」
.claude/agents/os-dev.md:158 — 「仓级扫描(以 pnpm lint = eslint . --no-inline-config 为首)是 CI 拥有的运行,永不是你欠的。」
.claude/skills/pm-dispatch/references/execution-duties.md:142 — 「mode:cloud 只保留给 L/XL、活过 PM 会话的工作、浏览器/dogfood 验证;build 重的 M 卡逐卡判。」
None of them says whether a package's test:repo project counts as "the affected package's tests" or as a repo-wide scan. None gives "build-heavy" a measurable test.

Protocol declaration: no product protocol changes. Every option edits agent and PM rules under .claude/** (Tier S).

Premises, each with its re-check

CI runs test:repo on a PR. Re-check: git grep -n "turbo run test test:repo" origin/main -- .github/workflows/ci.yml. Expect :965. As a control in the same file, git grep -n "merge_group" origin/main -- .github/workflows/ci.yml must hit.
The derived gate list does not send agents to test:repo. Re-check: node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack packages/spec/src/data/driver/config-registry.zod.ts | grep -c test:repo. Expect 0; the same command's total line count is the control (70 on cdc42eb49).
The lock is one slot per container. Re-check: bash scripts/pm/os-verify-lock.sh --status prints one state: holder line.

F1 — local `test:repo`

一句话问题: 开发 agent 在本地跑包级全仓扫描测试,每次占锁 13–15 分钟,而 CI 在 PR 上本来就会跑同一套。

option	what	what the fleet sees	workload
A	`os-dev.md`'s local-verification section names a package's `test:repo` project as CI's run: run it locally only when the diff edits a `repo`-project test, else declare the narrowing in `deviations`	about a quarter of the lock's hold time returns to the queue (measured 26% in the window above); a `repo`-project failure shows on the PR's CI instead of before the push	skills lane · 1 Tier S PR, a few lines in `.claude/agents/os-dev.md` · precedent PR #16271 (+11/−11, 2 files, a rule line in the same family)
B	keep the text; seats keep sending a per-dispatch lock note, as this seat did at 03:40Z	depends on each seat remembering	none

业务含义: A 像规定「全楼消防演练由物业统一做,各户不必自己再跑一遍」;B 是靠每次口头提醒。
四轴:
- ① 长远:A 把「本地定向、全量归 CI」这条既有原则补到它漏掉的那一类测试上,不新增机制。
- ② 拉动:本班实测 2 次就占锁 26%;三张卡同时被拖慢。
- ③ 防 AI 犯错:A 让 agent 不必猜「包自己的测试」包不包括全仓扫描;代价是这类失败在 CI 上才出现,而不是推送前就被发现。
- ④ 不扩散:只改一句规则,不加门禁。
推荐 A,回退 B。置信缺口:只量了本容器 2 小时;其它席位容器的占比没测。

F2 — routing build-heavy cards

一句话问题: 技能允许「build 重」的卡走云卡,但没有判据;席位把三张要反复构建 spec 的契约卡都判成普通卡,塞进同一容器同一把锁。

option	what	what the fleet sees	workload
A	`execution-duties.md:142` gets a test: a card whose declared surface writes `packages/spec/src/**` (a published schema or contract) and at least one other package is build-heavy and goes `mode:cloud`; a container runs at most one build-heavy subagent at a time	each heavy card gets its own container and lock; more cloud sessions to run and archive; each pays its own install and closure build	skills lane · 1 Tier S PR, a few lines in `references/execution-duties.md` · precedent PR #16271 (+11/−11)
B	keep 「逐卡判」, and the seat's own judgment	the next seat can repeat this round	none
C	stay `mode:subagent`, and raise the container size plus the lock's slot count	fewer sessions; needs a measured memory ceiling (spec's DTS pass alone asks for 6 GB) and a lock-script change	the environment's size (a maintainer setting) + 1 script PR + a measurement round · no precedent

⛔ C's lock change is not "a second lock class": [finding] The dispatch cap counts CARDS, but the binding resource is one container-wide heavy-verify lock — measured convoy: holder 284s, queue depth 3, waits of 442s/376s/312s #14944's Q2 ruled the class split out. C widens the one lock's capacity.
业务含义: A 像重活派到各自独立的工位,轻活共用一个工位;C 是把共用工位换大、多开一个口。
四轴:
- ① 长远:A 用已有的云卡通道、给已有的「逐卡判」一个可量判据,不新增机制;C 改锁与机器规格,要先量内存上限。
- ② 拉动:本班三张重卡实测平均每张排队约 70 分钟;spec(conversions): eight ADR-0087 D2 conversions stay on the 18 load path after #23487 — five families each need a ruling before they can retire #23543 还有五族重卡要逐个派。
- ③ 防 AI 犯错:A 把一个靠席位记忆的判断变成看文件面就能判的规则;B 下一个席位照样会判错。
- ④ 不扩散:A 只改一句;C 要动脚本和环境。
推荐 A,回退 C。置信缺口:云卡每张要重装依赖、重建闭包的固定开销没实测;判据线(spec 加一个以上其它包)是按本班三张卡定的。

自检:只看①选 F1 A、F2 A;②③④ 是否翻转:否。

os-decision-facets

① 项目长远合理性:两项 A 都是把既有原则(本地定向、全量归 CI;重卡可上云)补上可判的那一句,不新增机制;C 才扩锁与机器规格。
② 实际业务拉动:本班实测锁占用 89%、三个开发 agent 累计排队约 3.5 小时、test:repo 两次占 26%;spec(conversions): eight ADR-0087 D2 conversions stay on the 18 load path after #23487 — five families each need a ruling before they can retire #23543 还有五族重卡待派。
③ 防 AI 犯错:A 让 agent 和席位不必猜「包自己的测试」与「build 重」;代价是 test:repo 的失败晚到 CI 才出现。
④ 创业阶段不扩散:两项 A 各改一句规则,零新增门禁;C 需改脚本与环境。
Prior rulings read: test:repo,verify-lock,mode:cloud,subagent,build-heavy,local verification → 0 ADR hits; [finding] The dispatch cap counts CARDS, but the binding resource is one container-wide heavy-verify lock — measured convoy: holder 284s, queue depth 3, waits of 442s/376s/312s #14944 ruling 5548762263 (Q1 B the depth hold, Q2 no lock-class split, Q3 A the wait discipline) and finding(skills): full-farm verification runs vs the single shared heavy-verify lock — dependency-bump cards serialize every sibling agent for ~30 minutes #8716 (directions 1 and 2 not taken for want of measured need, re-openable on a future measurement); thread: none; repo: objectstack-ai/objectstack

After the ruling

F1 A and F2 A: one Tier S PR in the skills lane, edits to .claude/agents/os-dev.md (F1) and .claude/skills/pm-dispatch/references/execution-duties.md (F2), at CONTRACT_REVIEW_TIER because os-dev.md is the protocol-semantic surface. Until it lands, domain:spec seat 2 applies both by hand: lock notes to its dev agents, and heavy spec cards (spec(conversions): eight ADR-0087 D2 conversions stay on the 18 load path after #23487 — five families each need a ruling before they can retire #23543's next families) dispatched as cloud sessions.
F2 C: the maintainer sets the environment size first, then a measurement card for the lock's slot count.
Any B: this card closes with the ruling quoted; nothing changes.

Dedupe

Two semantic search_issues queries on objectstack-ai/objectstack, all states:

"os-dev local verification test:repo shared verify lock contention cloud card build-heavy subagent mode:cloud": 0 hits.
The control, "os-verify-lock heavy verify lock queue wait parallel agents container": 15 hits, all closed. The nearest are finding(skills): full-farm verification runs vs the single shared heavy-verify lock — dependency-bump cards serialize every sibling agent for ~30 minutes #8716, [finding] The dispatch cap counts CARDS, but the binding resource is one container-wide heavy-verify lock — measured convoy: holder 284s, queue depth 3, waits of 442s/376s/312s #14944, [finding] Decide with a measurement whether CPU-heavy check:* runs should be routed through os-verify-lock — the throughput/contention trade is currently unmeasured #12795, [finding] check:pm-dispatch-gates repeatedly holds the shared verify lock 500-900s+, starving the queue #14521 and Dev contract: the verify-lock waiting discipline (acquire once, stay blocked, always name OS_VERIFY_LOCK_SLOT) — the #14944 ruling's Q3 A half #16262. None is about test:repo or about routing build-heavy cards to cloud.

Generated by Claude Code

席位意见（增量）· director seat (summon #37 continuation 2, `session_01LqNgdPZB2Jnr5HpGbaqLeB`), read as batch #348 item 5, 2026-10-15T04:50Z

Premises re-run on origin/main 8d05456cb: .github/workflows/ci.yml:965 runs pnpm turbo run test test:repo on the merge group (P1 holds); dispatch-gates.mjs --commands for packages/spec/src/data/driver/config-registry.zod.ts derives 71 commands, 0 of them test:repo (P2 holds; the script read at this session's HEAD 6a40a09ed); the governing lines stand (os-dev.md:130, :138, :158; execution-duties.md:142); the lock ledger (P3) is container-local and is not re-read from here. Prior rulings (check-prior-rulings --card 23607): 440 candidates, none on local verification scope; #14944 (Q2: no lock-class split) read as the card cites it. Both files F1 and F2 edit sit at or beside their ceilings (os-dev.md 400/402, execution-duties.md 183/183): under #23606 B (same batch) this ruling authorizes the lines.

Facets: ① agree (local verification is targeted gates; a repo-wide scan is CI's; a heavy card gets its own container through the channel the skill already has) · ② agree (the lock held 89 % of two hours, three agents queued about 3.5 hours summed, test:repo alone 26 %) · ③ agree, with the cost named: a repo-project failure then shows on the PR's CI instead of before the push, one CI cycle (about 23 minutes at p50) against 13–15 minutes of serial lock per local run · ④ agree.

Recommend F1 A · F2 A, fallback F2 C, with 口径 for the executor:

F1: a package's test:repo project is CI's run unless the diff edits a repo-project test file or the derived --commands list names it; the dev's deviations line says which; the seat's ACCEPT reads the PR's CI for that project.
F2: the test is "the declared file surface writes packages/spec/src/** and at least one other package" ⇒ mode:cloud, one container per heavy card; a container runs at most one build-heavy subagent at a time. The cloud session pulls the turbo remote cache (the framework's closure reads HIT REMOTE on today's forced run), so its fixed cost is install plus cache pull; the first such dispatch measures and records it on its card.
⛔ No lock-script change and no environment-size change (C) without a measured memory ceiling; C returns only if the measured cloud overhead exceeds the queue it removes. 「多花钱的我不要」: A adds container-hours per heavy card on the existing plan and no new SKU; C changes the environment's specification.

Workload of the letters (the card's column): domain:skills · 1 Tier S PR at CONTRACT_REVIEW_TIER (os-dev.md is the protocol-semantic surface) · precedent PR #16271 (+11/−11). F2 C = the environment size (the maintainer) + 1 script PR + a measurement round · no precedent.

自检：只看①选 F1 A、F2 A；②③④ 是否翻转：否。

Decision-box entry: passes · `priority:p2` · `domain:skills` · `area:devpath` added · triage seat · `session_01AavokzJ5DndAwitDXvKy4U` · 2026-10-15T05:00Z

Entry check on the card body. These are all present:

the maintainer's words, quoted;
the speed read with a one-line ask (F1 A/B · F2 A/B/C);
the measured lock ledger;
the governing text (os-dev.md:130/138/158, execution-duties.md:142);
premises with re-checks;
per family: options × consequence × workload;
the four facets, with the token;
the recommendation (F1 A, F2 A), the fallback (F2 C) and the confidence gaps.

It passes.

Labels added: priority:p2 + domain:skills + area:devpath.

It is a measured fleet-velocity cost: the lock was held 89% of two hours, and three cards each queued about 70 minutes.
F1 A extends an existing rule ("local is targeted, the full farm is CI's") to the one class it left ambiguous. F2 A gives an existing per-card judgment a measurable test. No new gate either way.

Ruling: batch #348 item 5 · letter A/A · maintainer 「同意」 2026-10-15T05:02Z

Director seat, summon #37 (continuation 2), session_01LqNgdPZB2Jnr5HpGbaqLeB, on the card's analysis (F1 A · F2 A; fallback F2 C; filed on the maintainer's 「技能修正开决策卡」) and the seat's delta 6184678084; the maintainer's 「同意」 takes the recommendation on both questions: F1 A · F2 A.

What the ruling binds:

F1 A. A package's test:repo project is CI's run (ci.yml runs test test:repo on the merge group). A dev runs it locally only when the diff edits a repo-project test file or the derived --commands list names it; otherwise the dev's deviations line says so, and the seat's ACCEPT reads the PR's CI for that project. os-dev.md's local-verification section carries the sentence.
F2 A. execution-duties.md:142 gets the test: a card whose declared file surface writes packages/spec/src/** and at least one other package is build-heavy and is dispatched mode:cloud, one container per such card; a container runs at most one build-heavy subagent at a time. The cloud session pulls the turbo remote cache for the closure; the first such dispatch measures its fixed cost (install plus cache pull) and records it on its card.
⛔ No change to the verify-lock script and no change to the environment's size (F2 C) without a measured memory ceiling; C returns only with a measurement showing the cloud overhead exceeds the queue it removes.
Until the PR lands, domain:spec seat 2 applies both by hand (lock notes to its devs; spec(conversions): eight ADR-0087 D2 conversions stay on the 18 load path after #23487 — five families each need a ruling before they can retire #23543's remaining families dispatched as cloud sessions).

Workload of the letters: domain:skills · 1 Tier S PR at CONTRACT_REVIEW_TIER (.claude/agents/os-dev.md is the protocol-semantic surface; references/execution-duties.md sits at its ceiling and under #23606 ruling B this ruling authorizes the raise) · precedent PR #16271 (+11/−11).

Execution: this card is the execution card: needs-user-decision removed, pm:queue, domain:skills and priority:p2 added in this act; the body's first line carries Ruled:. The card closes on the PR's landing.

Claim: PM loop round 1
Session: session_01VqGhCg9wdsBXPz54uAEWg7
Account: os-elon-musk (the seat's linked user as GET /user answers it; always the card's assignee)
Branch: claude/issue-23607-test-repo-ci-heavy-cloud
Worktree: objectstack-issue-23607
Domain: domain:skills
Seat: domain:skills#2
File surface: .claude/agents/os-dev.md (〈本地验证范围〉, the F1 sentence: a package's test:repo project is CI's run) · .claude/skills/pm-dispatch/references/execution-duties.md (L144, the mode:cloud line: the F2 build-heavy test) · scripts/pm/check-skill-line-ratchet.mjs (only the two CEILINGS rows for those files, each moved by its measured delta under ruling record 6184799323 on #23606; ⛔ the header docblock and the refusal text are #23606's, in flight with seat 1) (stop on breach; explain in the report)
Container & model: S, mode:subagent, model: TIER_DEFAULT (reason: the measured quota exemption — --tier at 57994c311 reads MANDATORY CONTRACT_REVIEW_TIER for '.claude/agents/os-dev.md' (clause ①), and that tier refused this session's at-tier subagent with HTTP 429 (usage limit reached), reported to the seat by 2026-10-15T12:03Z; never lower; the at-tier contract review stays owed before landing)
Clause-②: no
Responsibility: n/a — not a defect card
Ruling-ref: 6184828752
Thread-read: 6184828752
Serial constraints cleared: #23579 merged (PR #23660, 881eeff31), the execution-duties.md chain head this card waited on; #23606 (seat 1, claim 6190726655, in flight) shares scripts/pm/check-skill-line-ratchet.mjs, in the header and refusal region against this card's CEILINGS rows, so the two are disjoint and each PR merges origin/main before opening; #23662 (this seat) touches only references/decision-analysis.md. Every open PR's file list and every open pm:dispatched claim were read 2026-10-15T12:09Z.

Generated by Claude Code

Dispatch order — posted 2026-10-15T12:12Z by the claiming seat (session_01VqGhCg9wdsBXPz54uAEWg7, os-elon-musk, domain:skills#2); the same text is the dev's prompt, mode:subagent, model tier TIER_DEFAULT under the measured quota exemption (the path mandate is CONTRACT_REVIEW_TIER; see the claim's Container & model line).

Dispatch — issue #23607 (objectstack-ai/objectstack) · `domain:skills` · PM `session_01VqGhCg9wdsBXPz54uAEWg7` (os-elon-musk)

You are the os-dev developer for exactly one card: objectstack-ai/objectstack issue #23607. Read the card's body and EVERY comment yourself (the director's delta, triage's entry, the ruling, the seat's claim, this order) and check that the body is complete before you start. The PM's Claim: comment on the card (id 6191711722) is your identity: its Branch: is claude/issue-23607-test-repo-ci-heavy-cloud, its worktree objectstack-issue-23607. Before any edit, verify that the newest Claim: on the card names that branch; on a mismatch, stop and report. ⛔ Leave no second claim, ⛔ never write the card's assignee; the PR's assignee is os-elon-musk (pr_create with assignees: ["os-elon-musk"]). Your standard duties (worktree-first, push the empty branch first, the shared verify lock, the report shape, the sanitizer rules) live in .claude/agents/os-dev.md and AGENTS.md; this order carries only the increments. The repository checkout is /home/user/objectstack (shared, ⛔ never edit it; your worktree goes beside it). Run everything in the foreground; a round ends only with the report.

1. Ruling — not re-decidable (`Ruling-ref: 6184828752`, letters F1 A · F2 A, the maintainer's 「同意」 2026-10-15)

Quoted from the ruling of record on the card (comment 6184828752):

What the ruling binds:

F1 A. A package's test:repo project is CI's run (ci.yml runs test test:repo on the merge group). A dev runs it locally only when the diff edits a repo-project test file or the derived --commands list names it; otherwise the dev's deviations line says so, and the seat's ACCEPT reads the PR's CI for that project. os-dev.md's local-verification section carries the sentence.

F2 A. execution-duties.md:142 gets the test: a card whose declared file surface writes packages/spec/src/** and at least one other package is build-heavy and is dispatched mode:cloud, one container per such card; a container runs at most one build-heavy subagent at a time. The cloud session pulls the turbo remote cache for the closure; the first such dispatch measures its fixed cost (install plus cache pull) and records it on its card.

⛔ No change to the verify-lock script and no change to the environment's size (F2 C) without a measured memory ceiling; C returns only with a measurement showing the cloud overhead exceeds the queue it removes.

Until the PR lands, domain:spec seat 2 applies both by hand (lock notes to its devs; spec(conversions): eight ADR-0087 D2 conversions stay on the 18 load path after #23487 — five families each need a ruling before they can retire #23543's remaining families dispatched as cloud sessions).
The raise authorization, quoted from ruling record 6184799323 on [Decision] The line ratchet (check:pm-skill-ratchet): 32 of 37 files at zero headroom, so a sentence already ruled still needs a second ruling to land — keep it, let the ordering ruling carry its lines, add slack, or narrow it #23606 (letter B, the maintainer's 「同意」):

A ruling of record that orders text into a ratcheted file is the authorization to raise that file's ceiling by the measured lines the text needs: the raising PR quotes the ruling comment's id and the measured line delta in its body, and moves the CEILINGS row in the same PR by at most that delta. No second word from the maintainer is needed.

In force from this ruling for every ruling of record, this batch's included ([Decision] A card triaged to one lane carries an item that lands in another (#23513: an engine card with a skills/** item) — who splits it, and when #23579's line in execution-duties.md, [Decision] Three spec contract dev agents in one container queued ~70 min each on the shared verify lock: leave package test:repo to CI, and send build-heavy cards to cloud sessions? #23607's lines in os-dev.md and execution-duties.md, [Decision] contract review of a mixed diff: does one governed path send the whole PR to CONTRACT_REVIEW_TIER, or only the governed hunks? #23578's line in contract-review.md): each such PR quotes its ruling and its delta.

Deliverable: ONE draft PR to main, first body line Fixes #23607, editing exactly:

.claude/agents/os-dev.md, the 〈本地验证范围〉 section: the F1 sentence (item 1).
.claude/skills/pm-dispatch/references/execution-duties.md: the F2 test (item 2), at or beside the mode:cloud line. The ruling names it :142; it is L144 on 57994c311, because PR pm-dispatch execution-duties: a split-off off-lane card does not inherit the parent's domain; ceiling 183 → 185 by ruling #23660 inserted two lines above it.
scripts/pm/check-skill-line-ratchet.mjs: only the CEILINGS rows for those two files, each moved by exactly its measured delta, with a comment block above each moved row in the file's own convention. Read the execution-duties.md 183 → 185 block that PR pm-dispatch execution-duties: a split-off off-lane card does not inherit the parent's domain; ceiling 183 → 185 by ruling #23660 added for the shape.

⛔ Nothing else. Item 3 forbids touching scripts/pm/os-verify-lock.sh or the environment. Item 4 is another seat's interim duty. No line in review-checklist.md or any other reference.

2. PM mechanism assumptions — measure before you build; falsify and report where wrong

(a) PM readings at origin/main 57994c311:
- os-dev.md is 400 lines, ceiling 402 (headroom 2). The F1 text may fit without a raise; if it needs more than 2 lines, the row moves by the measured excess.
- execution-duties.md is 185 lines, ceiling 185 (headroom 0). Any F2 line moves its row by the measured delta.
- Every line in these files is held to the ratchet's 120-byte per-line cap (see the ratchet header's "Why 120 bytes"). execution-duties.md's widest line is 120 B today.
(b) CI runs test:repo: git grep -n "turbo run test test:repo" origin/main -- .github/workflows/ci.yml hits L965. Control in the same file: merge_group must hit. Re-run both.
(c) The F1 sentence belongs among the numbered local-verification items (① build closure / ② the affected package's pnpm test / pnpm typecheck / ③ derived families). The open question the ruling closes is whether item ②'s "affected package's tests" includes a package's test:repo project. Read os-dev.md L130–160 and place the sentence where item ② is read; ⛔ do not reword ①–③.
(d) Concurrent sibling: PR feat(pm): line ratchet — a raise is authorized by the ordering ruling of record (id + measured delta in the PR body) #23668 ([Decision] The line ratchet (check:pm-skill-ratchet): 32 of 37 files at zero headroom, so a sentence already ruled still needs a second ruling to land — keep it, let the ordering ruling carry its lines, add slack, or narrow it #23606, seat 1, Tier S, already out of draft) edits the ratchet's header docblock, its verdict and refusal text, its self-tests and one SKILL.md row. It does not touch the CEILINGS rows. Merge origin/main before you open the PR; if feat(pm): line ratchet — a raise is authorized by the ordering ruling of record (id + measured delta in the PR body) #23668 has landed by then, the ratchet's own text names the ruled-raise form, and your PR body follows it (ruling id + measured delta). ⛔ Do not edit the header or the refusal text yourself.
(e) Same-day churn: execution-duties.md and scripts/pm/check-skill-line-ratchet.mjs merged changes today (881eeff31, PR pm-dispatch execution-duties: a split-off off-lane card does not inherit the parent's domain; ceiling 183 → 185 by ruling #23660; 8d05456cb, PR docs(agents): a test file that declares a measured budget with its class is held to that budget, not the 10 s line #23596). Work from the merged code; the snippet the card cites may have moved, so check current main first.

3. PM suggested route — optional; measurement wins

F1, one line in the file's register, e.g. 「② 不含包的 test:repo 项目:它归 CI(ci.yml 在 merge group 跑);仅 diff 改该项目测试或派生清单点名时本地跑,否则 deviations 写明。」 Split it if it exceeds 120 B, and the headroom of 2 may then cover it.
F2, e.g. 「声明文件面写 packages/spec/src/** 且至少另一包 ⇒ build 重,mode:cloud 一卡一容器;一容器同时至多一个 build 重子代理;首张实测固定开销记卡。」 Give it two lines if the bytes need it, and move the row by what you measure.
The PR body quotes, verbatim: ruling id 6184828752 with letters F1 A · F2 A; each file's measured delta; and the authorization id 6184799323, letter B, item 1. The seat record and the queue guard's reader check that quote.

4. Verification you owe (the derived gate families below are a clue, not the spec)

Derived at 57994c311 by node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack .claude/agents/os-dev.md .claude/skills/pm-dispatch/references/execution-duties.md scripts/pm/check-skill-line-ratchet.mjs (39 commands). After your FINAL commit, re-derive with no paths. Run the local subset, capture each exit code before any pipe, and reconcile with --ran. Quote git rev-parse --short HEAD of that run in the report and the PR body. CI runs the full farm; cite the PR's check-runs (name + conclusion) for every family CI runs.

node scripts/check-ci-filter-parity.mjs
node scripts/check-closing-keyword-parity.mjs
node scripts/check-closing-keyword-parity.mjs --self-test
node scripts/check-comment-mask-corpus.mjs
node scripts/check-declaration-mirrors.mjs
node scripts/check-declaration-mirrors.mjs --self-test
node scripts/check-scripts-symbol-anchors.mjs
node scripts/check-scripts-symbol-anchors.mjs --self-test
node scripts/check-self-test-wired.mjs
node scripts/check-self-test-wired.mjs --self-test
node scripts/check-self-test-workflow-commands.mjs
node scripts/check-self-test-workflow-commands.mjs --self-test
node scripts/check-whole-set-label-write.mjs
node scripts/check-whole-set-label-write.mjs --self-test
node scripts/pm/bare-root-worklist.mjs --self-test
node scripts/pm/check-governed-queue-guard.mjs --self-test
node scripts/pm/check-harness-current.mjs --self-test
pnpm --filter @objectstack/lint run check:doc-formula-expressions
pnpm check:agent-model-declared
pnpm check:agent-test-spelling
pnpm check:bash32-floor
pnpm check:cli-command-ids
pnpm check:commit-card-trailers
pnpm check:cross-package-test-inputs
pnpm check:doc-authoring
pnpm check:driver-memory-census
pnpm check:entry-guard
pnpm check:gitlink-declared
pnpm check:nul-bytes
pnpm check:parse-guard
pnpm check:pm-dispatch-gates
pnpm check:pm-governed-merges
pnpm check:pm-skill-id-lint
pnpm check:pm-skill-ratchet
pnpm check:pnpm-filter-targets
pnpm check:ratchet-remedy-authority
pnpm check:refd-timer-probe
pnpm check:skill-frame-sync
pnpm check:watch-hint-literal

Plus the edited script's own suite: node scripts/pm/check-skill-line-ratchet.mjs --self-test and the plain run (its verdict lines for both files are readings the PR body quotes). pnpm check:pm-dispatch-gates is a long battery: run it with nohup … > LOG 2>&1 &, then tail --pid, as os-dev.md says.

5. PR body and labels

First line Fixes #23607; then the line Clause-②: no at line start (copied from the claim; no arm).
A ## 维护者速读(草稿) section (Chinese, business angle; five fixed parts: 改了什么 / 为什么改 / 风险与代价(含回滚) / 席位意见(留空) / 你要做的).
The diff publishes nothing (.claude/**, scripts/pm/**). Write no changeset, and add the skip-changeset label through scripts/pm/label-write.mjs (labels only).
This PR is governed, Tier S (.claude/** plus the non-governed scripts/pm/**). Push, open the draft PR, deliver the report and STOP. The seat lands it after the at-tier contract review PASS. ⛔ Never flip it out of draft, never arm auto-merge.
Footer in the session-URL form; no model identifier anywhere in the PR, commits or comments.

6. Writes

Writes: REST proxy only, through scripts/pm/ tools (git push is not a REST write). Budget:

git push (as needed);
pr_create once (draft, assignees: ["os-elon-musk"]);
label-write once (skip-changeset);
post-stamped --comment=23607 once (the os-dev-report comment).

Report api_writes (count + endpoints) and mcp_calls (count + tool names; a write tool in the list refuses the report).

7. Report

Deliver the JSON report twice, as .claude/agents/os-dev.md specifies: first as a comment on #23607 whose first line is the bare text os-dev-report, then as your final message (the JSON alone, no surrounding prose). Machine-checkable fields only: gates / tests / deviations / files_changed, with each file's line delta and the ratchet verdict lines.

8. The four-axis decision frame — use it only if a `needs_decision` arises; ⛔ keep no copy in the repo

Verbatim copy of .claude/skills/pm-dispatch/SKILL.md lines 241–261 at 7b3d5585c, md5 abff5f852c8fd09b79ec0623439aa4f0:

每个方案必须沿四条固定评估轴分析,这是决策分析的核心原则,不是可选项:

实际业务需求 — 它服务的是真实存在的业务场景,
还是投机性能力面?判据要求实测(谁在写这个键、谁在读、
示例应用与真实部署的用法),「读起来像有用」不作数。这条轴会改变结论,不是陪衬。
项目长远合理性 — 哪个方案符合北极星方向与可持续架构(no workarounds、
contract-first),临时补丁式选项要明说长期代价。
防 AI 写代码犯错,尤其是防 AI 写元数据 app 犯错 — 哪个方案让 AI 在结构上更难写错:
契约收紧(严格 schema、publish 时响亮拒绝)优于消费端宽容(?? 回退、静默容错)——
宽容恰是 AI 批量犯错被掩盖的温床;声明即强制,绝不让 AI 声明一个运行时不兑现的能力。
创业阶段不扩散需求 — 创业阶段聚焦原则:创业项目先专注核心能力,能力扩张默认
从紧,无拉动的声明面按 implementation-first 处置,已发布零消费能力不因沉没成本豁免。
过渡也从紧 —— 创业阶段不渐进:废弃别名/拼写与能力退役默认立即退休,无分阶段
窗口、无双拼写宽限;staged 仅凭具名外部用户证据方可荐。新增门禁默认否:门禁是只减
不增的零件,例外只有维护者点名;荐新增门禁须引其点名原话,否则荐「不加」。

基本裁决原则:spec 声明 > 实现 > 文档面;声明而未兑现是实现缺口,补实现或退役,
⛔ 不在消费端收窄;未声明键随实现,文档随之。「声明即强制」只及 spec 与元数据契约、
运行时能力,⛔ 不及 AGENTS.md 这类给 agent 读的约定 —— 那些的读者是 agent,不是门禁。
推荐意见必须基于这四条轴给出理由;四轴冲突时如实呈现权衡,
交维护者拍板。

Generated by Claude Code

os-dev-report
{
"issue": 23607,
"status": "done",
"branch": "claude/issue-23607-test-repo-ci-heavy-cloud",
"pr": "#23676",
"session": "session_01VqGhCg9wdsBXPz54uAEWg7",
"premise_still_valid": true,
"summary": "F1 A: two lines in os-dev.md's local-verification section, right after item 2's continuation lines. Item 2 excludes a package's test:repo (the vitest repo project), which CI's Test Core runs on both PR and merge group; a dev runs it locally only when the diff edits that project's tests or the derived list names it, otherwise says so in deviations, and the seat's ACCEPT reads its CI. F2 A: three lines beside the mode:cloud line in execution-duties.md. A declared file surface that writes packages/spec/src/** plus at least one other package is build-heavy and goes mode:cloud, one card per container; a container runs at most one build-heavy subagent at a time; the cloud session pulls the turbo remote cache for the closure; the first such dispatch measures its fixed cost and records it on its card. The mode:cloud line itself is unchanged. Ratchet: the execution-duties.md CEILINGS row moves 185 -> 188 (+3, the measured delta) under ruling 6184828752 and authorization 6184799323 letter B item 1. os-dev.md uses its standing headroom of 2, so its row stays at 402. Draft PR #23676 is Tier S, assignee os-elon-musk, label skip-changeset; it waits for the seat's at-tier contract review.",
"files_changed": [
".claude/agents/os-dev.md: +2 / -0, 400 -> 402 lines, ceiling 402 unchanged (row moved by 0); new lines 101 B and 119 B",
".claude/skills/pm-dispatch/references/execution-duties.md: +3 / -0, 185 -> 188 lines, ceiling 185 -> 188; new lines 112 B, 95 B, 79 B",
"scripts/pm/check-skill-line-ratchet.mjs: +21 / -1, only the execution-duties.md CEILINGS row and its comment block"
],
"line_budget": "execution-duties.md: tightest one-line spelling of the four ruled parts is 245 B against the 120 B cap; neither two-line split fits (test plus per-container limit 137 B; limit plus cache plus fixed-cost 141 B), so +3. os-dev.md: tightest one-line spelling 131 B, so 2 lines, which fit headroom 2.",
"tests": "At head 4428e17: node scripts/pm/check-skill-line-ratchet.mjs --self-test exit 0, '✓ check-skill-line-ratchet self-test: 164 cases pass.' Plain run exit 0 with 0 red lines, printing '✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/references/execution-duties.md is 188 lines (ceiling 188; headroom 0).' and '✓ check-skill-line-ratchet: .claude/agents/os-dev.md is 402 lines (ceiling 402; headroom 0).' The per-line byte rule prints only on failure. No *.test.ts names the script (git grep exit 1). No ablation: the diff is agent text plus one ceiling number, and no tested behaviour changes.",
"gates": "Derived at head 4428e17 with dispatch-gates --commands --repo objectstack-ai/objectstack and no paths: 39 commands, 3 paths against merge base 5477020, a list identical to the derivation at 0f02a71. Local: all 39 exited 0, each captured before any pipe. pnpm check:pm-dispatch-gates ran with nohup and tail --pid, printed '1884 cases pass; 6 slow section(s) deferred', 268 s, exit 0. --ran printed '✓ dispatch-gates --ran: 39 derived famil(ies) accounted for — 39 run, 0 NOT-MEASURED (a DERIVED zero — all 39 recorded an exit code and none of them is 3).' The doc-formula-expressions gate first exited 3 (PREREQUISITE NOT MET: formula and lint dist absent); both were built under the verify lock (VERDICT command-exit 0) and the gate re-ran with exit 0. Extra ratchet readers, not derived: check-published-list-mirrors exit 0, check-skills-token-ratchet exit 0. PR check-runs on 4428e17 at report time: 12 success, 11 skipped, 17 in_progress. Success: Governed Surface Queue Guard, Type Check · workspace, Check Changeset, Check PR Size, Auto Label, Check Documentation Links, Dogfood Regression Gate, filter, the three claim and single-writer guards, Part-of guard. In progress: the eight Lint lanes, Test Core 1-6/6, Type Check · consumer gates, debt ledger and source gates. The seat reads the final CI.",
"deviations": [
"The F1 line says Test Core runs test:repo on both PR and merge group, where the ruling's parenthetical says 'on the merge group'. Measured: Test Core's if: (ci.yml:501) has no event exclusion, unlike the four full-battery jobs. The PR body states this.",
"test:repo was not run locally (the F1 rule applied to itself): the diff edits no repo-project test and the derived list does not name it. No package is touched, so there is no item-1 closure build and no item-2 package test run.",
"origin/main was merged twice before opening the PR. The second merge brought #23668 (the ratchet header's ruled-raise form); the merge was clean, and the full local set was re-run on the merged head.",
"The harness attribution reminder asked for a model-named Co-Authored-By trailer and a different PR footer. The commit instead carries AGENTS.md's model-free trailer pair, and the PR carries the session-URL footer.",
"Ran at TIER_DEFAULT under the claim's quota exemption; the at-tier contract review is still owed."
],
"mcp_calls": "0",
"api_writes": "3 relay strokes, each a POST /repos/objectstack-ai/objectstack/dispatches executed as objectstack-fleet[bot]. (1) pr_create: POST /repos/objectstack-ai/objectstack/pulls (draft) plus POST /repos//issues/23676/assignees; read-back 9799 B sent and stored, identical. (2) label-write labels_add: POST /repos//issues/23676/labels with skip-changeset; read-back matches. (3) post-stamped --comment=23607: POST /repos//issues/23607/comments, this report. git push is not counted.",
"open_questions": [],
"out_of_scope_findings": [
"carrier: none (承接者:无) · noted, not filed (PR Acceptance notes) · os-dev.md's report JSON template carries no deviations key, while the ruled F1 line, two earlier os-dev.md lines and dispatch orders all name a deviations field · agent-text consistency, no class a/b/c reach · dedupe words: os-dev report template deviations field, report JSON keys, files_changed line_budget"
]
}

Generated by Claude Code

ACCEPT → PR #23676 comment 6192430896 · seat domain:skills#2, session_01VqGhCg9wdsBXPz54uAEWg7 · 2026-10-15T12:49Z

Tier S. The PR stays draft with needs:contract-review until the at-tier review passes on its head. The build ran at TIER_DEFAULT under the quota exemption, so that review is the compensating control and is not waived. This card stays pm:dispatched until the PR merges.

Generated by Claude Code

Landing refused by the session classifier · seat domain:skills#2, session_01VqGhCg9wdsBXPz54uAEWg7 · 2026-10-15T21:27Z

The at-tier review passed. The PASS record is comment 6202616611 on PR pm: a package's test:repo project is CI's run; a spec-plus-one-package card is build-heavy and goes mode:cloud #23676, at head 4428e1746ad86d37069ae7b3dd838de4bc8cadd5, with Served-tier: CONTRACT_REVIEW_TIER. The reviewer transcript shows 57 at-tier stamps and 0 at any other tier.
The pre-landing checks hold. All 47 check-runs are success or skipped. check-governed-merges --pr 23676 reads Tier S. mergeable_state reads clean.
Refused command, step 1 (landing): node scripts/pm/fleet-write/dispatch.mjs --repo objectstack-ai/objectstack --actions-file F, with F = [{"op":"pr_ready","pull":23676},{"op":"automerge_enable","pull":23676}]. This session's auto-mode classifier refused it before anything ran, reason Merge Without Review.
Refused command, step 2 (label): node scripts/pm/label-write.mjs --repo objectstack-ai/objectstack --issue 23676 --remove needs:contract-review. Refused the same way, reason CI Bypass.
What the seat does now: stops. Per landing-operations.md B, it does not rephrase either command or try another route, and reports to the maintainer.
Current state: PR pm: a package's test:repo project is CI's run; a spec-plus-one-package card is build-heavy and goes mode:cloud #23676 is unchanged. It is still draft, with no auto-merge and needs:contract-review still on it. Landing it now takes the maintainer's hand (ready plus auto-merge), or a permission rule in this session's settings.

Generated by Claude Code

Landed · seat domain:skills#2, session_01VqGhCg9wdsBXPz54uAEWg7 · 2026-10-15T23:08Z

probe	reading
PR #23676	`merged` by the merge queue at 2026-10-15T23:06Z, merge commit `1b9e526d1`; it was enqueued at 2026-10-15T22:50Z on head `4428e1746a`, the head that the at-tier PASS 6202616611 names
`origin/main` (tip `3b46855fb`)	`1b9e526d1` is an ancestor. The three files are byte-identical to the reviewed head. The F1 lines are `.claude/agents/os-dev.md` L138–139 and the F2 lines are `references/execution-duties.md` L145–147. `check-skill-line-ratchet` exits 0: `os-dev.md` 402/402 and `execution-duties.md` 188/188
merge queue	no `gh-readonly-queue/main/*` ref remains

State: the PR's Fixes closed this card completed. pm:dispatched is removed in this act, and only this card was closed. The landing stroke (pr_ready + automerge_enable, relay run 38819636530) followed the refusal recorded at 6202646278 and ran on the maintainer's word in the seat's session, 「同意你复核落地」. The serial queue on os-dev.md is released: #23665 is next.

Generated by Claude Code

Navigation Menu

Repository navigation

[Decision] Three spec contract dev agents in one container queued ~70 min each on the shared verify lock: leave package `test:repo` to CI, and send build-heavy cards to cloud sessions? #23607

维护者速读

Background (measured)

Governing text

Premises, each with its re-check

F1 — local `test:repo`

F2 — routing build-heavy cards

After the ruling

Dedupe

席位意见（增量）· director seat (summon #37 continuation 2, `session_01LqNgdPZB2Jnr5HpGbaqLeB`), read as batch #348 item 5, 2026-10-15T04:50Z

Decision-box entry: passes · `priority:p2` · `domain:skills` · `area:devpath` added · triage seat · `session_01AavokzJ5DndAwitDXvKy4U` · 2026-10-15T05:00Z

Dispatch — issue #23607 (objectstack-ai/objectstack) · `domain:skills` · PM `session_01VqGhCg9wdsBXPz54uAEWg7` (os-elon-musk)

1. Ruling — not re-decidable (`Ruling-ref: 6184828752`, letters F1 A · F2 A, the maintainer's 「同意」 2026-10-15)

2. PM mechanism assumptions — measure before you build; falsify and report where wrong

3. PM suggested route — optional; measurement wins

4. Verification you owe (the derived gate families below are a clue, not the spec)

5. PR body and labels

6. Writes

7. Report

8. The four-axis decision frame — use it only if a `needs_decision` arises; ⛔ keep no copy in the repo

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

[Decision] Three spec contract dev agents in one container queued ~70 min each on the shared verify lock: leave package test:repo to CI, and send build-heavy cards to cloud sessions? #23607

Description

维护者速读

Background (measured)

Governing text

Premises, each with its re-check

F1 — local test:repo

F2 — routing build-heavy cards

After the ruling

Dedupe

Activity

objectstack-fleet commented on Oct 15, 2026

席位意见（增量）· director seat (summon #37 continuation 2, session_01LqNgdPZB2Jnr5HpGbaqLeB), read as batch #348 item 5, 2026-10-15T04:50Z

objectstack-fleet commented on Oct 15, 2026

Decision-box entry: passes · priority:p2 · domain:skills · area:devpath added · triage seat · session_01AavokzJ5DndAwitDXvKy4U · 2026-10-15T05:00Z

objectstack-fleet commented on Oct 15, 2026

objectstack-fleet commented on Oct 15, 2026

objectstack-fleet commented on Oct 15, 2026

Dispatch — issue #23607 (objectstack-ai/objectstack) · domain:skills · PM session_01VqGhCg9wdsBXPz54uAEWg7 (os-elon-musk)

1. Ruling — not re-decidable (Ruling-ref: 6184828752, letters F1 A · F2 A, the maintainer's 「同意」 2026-10-15)

2. PM mechanism assumptions — measure before you build; falsify and report where wrong

3. PM suggested route — optional; measurement wins

4. Verification you owe (the derived gate families below are a clue, not the spec)

5. PR body and labels

6. Writes

7. Report

8. The four-axis decision frame — use it only if a needs_decision arises; ⛔ keep no copy in the repo

objectstack-fleet commented on Oct 15, 2026

objectstack-fleet commented on Oct 15, 2026

objectstack-fleet commented on Oct 15, 2026

objectstack-fleet commented on Oct 15, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

[Decision] Three spec contract dev agents in one container queued ~70 min each on the shared verify lock: leave package `test:repo` to CI, and send build-heavy cards to cloud sessions? #23607

F1 — local `test:repo`

席位意见（增量）· director seat (summon #37 continuation 2, `session_01LqNgdPZB2Jnr5HpGbaqLeB`), read as batch #348 item 5, 2026-10-15T04:50Z

Decision-box entry: passes · `priority:p2` · `domain:skills` · `area:devpath` added · triage seat · `session_01AavokzJ5DndAwitDXvKy4U` · 2026-10-15T05:00Z

Dispatch — issue #23607 (objectstack-ai/objectstack) · `domain:skills` · PM `session_01VqGhCg9wdsBXPz54uAEWg7` (os-elon-musk)

1. Ruling — not re-decidable (`Ruling-ref: 6184828752`, letters F1 A · F2 A, the maintainer's 「同意」 2026-10-15)

8. The four-axis decision frame — use it only if a `needs_decision` arises; ⛔ keep no copy in the repo