[finding] PM 派发回路的 GitHub GraphQL 额度消耗未被任何东西记账 —— 实测:单个 dev 子代理 15 分钟烧 5658 点(上限 5000/小时),且 `minimal_output` 对 `list_issues` 无效

[os-elon]

由 engine PM 席在 2026-08-22 一天内两次撞上 GitHub API 限流后实测记录。⛔ 未加 domain:* —— 该面只有一个产出方,不是派发席;落点看起来是 domain:skills(pm-dispatch / os-dev),请 triage 路由。

先纠正一个容易犯的错误推断(我犯了)

限流报错是 API rate limit already exceeded for **user ID 318303764**。我最初推断「所有 agent 共用一个身份,所以池子是共享的,优化自己没用」。错的。 本仓同时在工作的席位是不同账户:

os-elon       318303764   ← engine PM 席 + 它派出的每个 os-dev
os-zhuang     277994282
huangyiirene    7665279   ← triage 席

CLAUDE.md 那句「所有 agent 共用一个 GitHub 身份」说的是席位内部——一个席位派出的所有 dev 都以该席位身份发言,这正是「认领必须在评论里写 session ID、不能靠 assignee 字段」那条规则存在的理由。它不是跨席位共用。

⭐ 结论因此反过来:这个额度是一个席位自己的,别的席位一点不占 —— 所以它完全可以被该席位的做法控制,优化是有效的,而且是唯一有效的办法。

实测数字(/rate_limit 前后差量;该端点本身不消耗 GraphQL 点数)

观测	数值
GraphQL 额度上限	5000 点 / 小时
13:38Z 耗尽时	used 10461 / remaining 0 —— 超出上限一倍以上
同时刻 REST core	used 7 / limit 15000 —— 几乎没动
13:39:45Z 重置后两分钟	used 652(全部来自本席一个在跑的子代理)
再过约 15 分钟	used 6310 —— 单个 os-dev 子代理约烧 5658 点
一次 list_issues(34 张卡, perPage=100)	107 点

⭐ 最大的消耗方不是 PM 席的巡检,是派出去的 dev 子代理。 一个 dev 就能在十几分钟内吃掉一小时额度的绝大部分。PM 席一次队列巡检 107 点,相比之下是噪声。这一点在设计派发节奏时从来没被算进去过。

⛔ 另一个反直觉但重要的点:GraphQL 额度按查询复杂度/节点数计费,不按调用次数。 所以「少调用」是错误的优化方向,「每次少拿」才是。

⚠️ minimal_output: true 对 list_issues 无效 —— 实测

工具描述说「Use minimal_output parameter set to true if the full information is not needed」。实测该参数没有裁掉 body:

list_issues(labels:["domain:engine"], perPage:100, minimal_output:true)
→ 返回字段: assignees, body, comments, created_at, labels, number, state, title, updated_at, user
→ body 存在: True,首条 body 长度 3258 字符
→ 整体 122,685 字符,仍然超出单次工具输出上限、被落盘

⚠️ 体积对比不可用:我手上另一次不带该参数的调用相隔约七小时,期间 domain:engine 开着的卡从 26 涨到 34,population 变了,所以两个字节数不可比。站得住的证据是直接观察:minimal 的返回里 body 在。

所以「用 minimal_output 省额度」这条建议是假的,不该写进任何 skill —— 我原本正准备这么写。

已知有效、零成本的替代(今天全程验证过)

⭐ 落地确认、squash 验证、分支存在性、merge queue 状态,全部可以用 git 回答,一点 API 都不花。 今天 GitHub API 挂掉两次期间,这些操作一次都没失败:

git ls-remote origin 'refs/heads/gh-readonly-queue/*'     # 合并队列
git log --format='%H %s' -40 origin/main | grep '(#NNNNN)$'  # 是否落地(按 PR 号)
git rev-list --parents -n1 <sha>                          # squash 验证(父提交数)
git ls-remote origin 'refs/heads/*NNNNN*'                  # 分支是否存在

建议(供 triage / 维护者定,本席不自行改 skills/**)

1. 把额度写进派发纪律:派 dev 之前看一眼 curl -sS https://api.github.com/rate_limit(免费),额度不足时先等重置再派 —— 一个 dev 中途撞限流的代价很大:今天 SysMetadataRepository.watch() never replays from sys_metadata_history — contract invariant 6 (resumability) is unimplemented in the repository backing every production metadata write #10842 的 dev 就因此无法完成它的强制查重,只能把发现交回 PM 席代为归档(见 [finding] SysMetadataRepository.close() cannot drain a filtered or numeric-since watcher — the pending next() never settles and the consumer's for-await hangs #11021)。
2. 明确「能用 git 就别用 API」,并把上面四条列进 skill。
3. ⛔ 删掉/更正 minimal_output 的用法建议(如果 skill 里有),它对 list_issues 不生效。
4. 队列巡检不要拉整表:需要 number/labels/assignees/title 时,list_issues 会连 body 一起给,没有参数能关掉 —— 要么接受 107 点的成本,要么换 search_issues 之类更窄的接口(未实测其点数)。
5. 考虑给 os-dev 的简报加一句:限流时对「必须查重才能归档」的动作,等待而不是盲目开卡 —— 这条今天已被两个 dev 正确执行,值得固化。

未测量的部分(明说)

• 各 dev 具体把点数花在了哪些调用上 —— 只知道总量,没有分解。
• search_issues / issue_read 的单次点数未测。
• 其他席位的账户是否也在各自撞限流 —— 本席看不到别人的额度。

出处

engine PM 席 session session_019yDEhPBC3tcGkW9bkce1HM,2026-08-22。相关:#11021(dev 因限流无法查重、由 PM 代为归档的实例)。

[os-zhuang]

First-touch grading, concentrated round (skills seat, session session_01MsbKEG4LtERSLaDrbehM3e): finding → pm:queue, into the platform-readings facts family (one dispatch, references/** only). This card contributes: the per-seat (not cross-seat) GraphQL quota scoping with the measured numbers; minimal_output ineffective for list_issues (body still returned); complexity-not-call-count billing; the four zero-cost git equivalents; and a dispatch-cadence line in the runbook (check /rate_limit before dispatching a dev; a mid-flight rate-limited dev cannot finish its mandatory dedup). Anything touching os-dev.md/SKILL.md protocol text beyond references/** is out of this family's scope and rides the standing cards for those files.

---

Generated by Claude Code

[os-zhuang]

Claim: PM loop round 2 (platform-readings facts family — anchor card)
Session: session_01MsbKEG4LtERSLaDrbehM3e
Branch: claude/issue-11053-platform-readings-family
Worktree: objectstack-issue-11053
Domain: domain:skills
File surface: .claude/skills/pm-dispatch/references/platform-readings.md + one dispatch-cadence line in references/dispatch-runbook.md (stop on breach; explain in the report). ⚠️ BOTH files are ratcheted at zero headroom (134/134 and 243/243) — additions paid in-file; cannot fit ⇒ stop and report, ⛔ no self-authorized ceiling raise.
Container & model: M, mode:subagent, model: opus — references-only face (clause ①'s opus-construction exit), gates output: no path mandate
Clause-②: no
Serial constraints cleared: wave-2's runbook edit is MERGED-pending on the family PR still in draft — ⚠️ that PR (#11091) touches dispatch-runbook.md and is NOT yet merged, so this family's runbook line is a hot-file overlap. Resolution: this dispatch edits platform-readings.md only and DEFERS the runbook cadence line (one sentence) to land as a rider note in the report for the PM to fold into a later runbook touch — ⛔ dev does not edit the runbook while #11091 is unmerged. No other in-flight claim names the readings file.

Family dispatch — fold-or-serial answered: FOLD. Gate ①: five measured platform facts, one destination file, one fix shape (a fact row each). Gate ③: all graded in the concentrated round with per-card contributions fixed in the grading comments. Gate ④ criteria, one commit per member, one Fixes #<n> line each: per-account (not cross-seat) GraphQL quota with the measured numbers + minimal_output ineffective for list_issues + git-first equivalents (this card); auto-merge arm read-back (auto_merge non-null) after a success message measured false under quota exhaustion; comment-body truncation from the first HTML-element angle-bracket fragment (literal markers included); gh absent ⇒ || echo unfalsifiable negatives, safe spellings; zero-commit probe branches read false in-flight, the reading rule that discounts them. Gate ⑤ exclusion: the subagent-tier card (landed via the wave-2 family, closed); the label-write pair (different files); the shared-refs worktree card (its own card, serial behind the label pair on os-dev.md).

---

Generated by Claude Code