pm 席位 GraphQL 限流:高频读改走 REST(curl+代理凭证),GraphQL 只留给 mutation-only 操作

[os-zhuang]

维护者 2026-08-23 指示:「项目经理经常被 GraphQL 限流……rest api 余额很多,排查哪些请求可用 REST,能用 REST 的尽量用 REST。」本卡由 skills 席(session 757f0dc0-5dae-51d7-9053-c15c57ebfe3b)带实测立卡。

实测(2026-08-23 01:1x UTC,共享身份,GET /rate_limit 免费端点)

桶	本小时 used	limit
core(REST)	~50	15000
graphql	~4965	5000
search	0	30/min(独立桶)

两个结构性证据(不受并发席位噪声影响):

1. MCP list_issues 返回 pageInfo.startCursor/endCursor = GraphQL Connection 游标——issue 列表走 GraphQL;
2. 夹逼测量:MCP 调用间 graphql used +7/+4,core 几乎不动。

即:所有席位共用一个身份,GraphQL 桶是公共牧场且已在天花板上;REST 桶几乎全空。

路由表(PM 循环高频操作)

操作	现走	改走 REST
拉 backlog(按标签列 issue)	list_issues(GraphQL 已证)	GET /repos/{o}/{r}/issues?labels=…&state=open&per_page=…
决策重读(评论翻页到最后)	issue_read get_comments	GET /issues/{n}/comments?per_page=100&page=N
卡正文/状态读	issue_read get	GET /issues/{n}
标签读-改-写	get_labels + POST	GET /issues/{n}/labels + POST(POST 两条已在 #11112 进 allowlist)
PR 读(正文/文件/commits)	pull_request_read	GET /pulls/{n}、/files、/commits
CI 判定	get_check_runs	GET /commits/{sha}/check-runs(本席今日已用)
PR 正文改写	update_pull_request	PATCH /pulls/{n}(本席今日已用,含 sanitizer 尾读回)

保留 MCP/GraphQL(无 REST 等价或语义缺失):enable_pr_auto_merge(mutation)、resolve_review_thread / unresolve_review_thread、pull_request_read get_review_comments(线程 isResolved 态 REST 不给)。search_* 独立桶,不动。

交付物

1. .claude/settings.json permissions.allow 增补 curl GET/PATCH 条目(比照 Make the additive label POST the prescribed write and permit it for dev agents #11112 的两条 POST 形状,objectstack+objectui 双仓),使席位免弹窗走 REST。⚠️ .claude/** 属 governed:draft PR + 人工合并。
2. 路由指引落在哪:platform-readings.md 134/134 零余量、SKILL.md 682/682 零余量——允许以真删付账(优先挖重复),付不起就只落 allowlist、prose 部分显式移交 Whole-text restructuring round of the pm-dispatch protocol face — maintainer-ordered, fires when the skills-lane backlog drains #11086 全文重构轮(在 PR 正文声明)。
3. 曾有席位汇报过被限流的现象记录可引用本卡实测,不必重测。

验证

• allowlist 生效:worktree 内以 seat 权限跑一条表内 curl GET 不弹窗;
• 不回归:.claude/hooks/* selftest 全绿;若动 prose,line-ratchet 门绿并在 PR 正文给付账明细。

[os-zhuang]

CLAIM — session 757f0dc0-5dae-51d7-9053-c15c57ebfe3b(skills 席),branch claude/issue-11181-pm-rest-routing。维护者 2026-08-23 当面指示立即执行,立卡即派发。

---

Generated by Claude Code

[claude]

{
 "issue": 11181,
 "status": "done",
 "branch": "claude/issue-11181-pm-rest-routing",
 "pr": "https://github.com/objectstack-ai/objectstack/pull/11186",
 "premise_still_valid": true,
 "summary": "Deliverable 1 landed: `.claude/settings.json` `permissions.allow` gains 19 narrow curl entries — nine REST path shapes per repo as objectstack/objectui twins (issue list, issue get, issue comments, issue labels read, PR get, PR files, PR commits, commit check-runs, PATCH PR body) plus the free `GET /rate_limit`. Shape mirrors the #11112 additive-label POST exactly: method + host + path shape pinned, URL immediately after the `-X` flag, trailing ` *` absorbing the Bearer header and payload; no bare `curl *`. Deliverable 2 (routing prose) is DEFERRED to the #11086 whole-text round and declared as such in the PR body — `platform-readings.md` is 134/134 and SKILL.md 682/682, and no honest deletion was available (detail in open_questions). No ceiling was raised. PR is a DRAFT on a governed surface; no seat flips it ready, enqueues it, or arms auto-merge — read back as draft:true, auto_merge:null, base:main. `skip-changeset` applied via the additive POST and read back after 25s: labels = size/s, skip-changeset. Incidental confirmation of the card's premise: the PR body PATCH and every read-back in this task were done with real REST curl from this seat (HTTP 200), so the seat's direct-REST channel is live.",
 "tests": "Gate families derived with `node scripts/pm/dispatch-gates.mjs` with NO hand-fed paths — it read the change set itself (`1 path(s) vs merge base be7262e72`, working tree 1) and named 4 families. Union re-run on the FINAL commit `99f2db40e2` with a clean `git status`; every exit code captured before any pipe (redirect-then-capture), every verdict quoted from the gate's own line. ratchet --self-test EXIT=0 '19 cases pass'; ratchet EXIT=0, 20 green ceilings incl. 'platform-readings.md is 134 lines (ceiling 134; headroom 0)'; check:doc-authoring EXIT=0 '389 files clean — no bare metadata literals'; check:skill-frame-sync EXIT=0 '4 copies of the decision frame are structurally isomorphic across 3 files'; check:pm-governed-merges EXIT=0 '119 assertions'; check:doc-formula-expressions EXIT=0 '22 record-scoped formula example(s) across 416 files / 1447 TS blocks judged clean' (needed `pnpm install --filter @objectstack/lint...` and `pnpm --filter '@objectstack/lint^...' build` first, both under `scripts/pm/os-verify-lock.sh -c`, VERDICT command-exit 0, held 2s and 198s); check:nul-bytes EXIT=0 'scanned 6485 text file(s), no raw ASCII control bytes'. Hook selftests (not touched by this diff, run because the card asks): guard-main-checkout-bash EXIT=0, guard-shared-stash EXIT=0. governed predicate `check-governed-merges.mjs --test .claude/settings.json` EXIT=3, verdict line 'GOVERNED — a human merge is the review record for this PR'. JSON validity: parses, 47 allow entries / 21 curl. Repo-wide ESLint NOT run locally — declared narrowing with all three proofs in the PR body: (1) population read from `eslint.config.mjs` itself — `**/*.{ts,tsx,mts,cts,js,jsx,mjs,cjs}` plus five .ts-only blocks, no JSON plugin, no `files:` naming .json; (2) file count of the intersection with this diff = 0, the diff being exactly one .json path; (3) invariance — a permission allowlist defines no rule, plugin, parser or files pattern, so it cannot move any untouched file's verdict. No ablation was performed (nothing in this card asserts a gate can fail).",
 "open_questions": [
 {
 "question": "The card's 2026-08-23 measurement and an existing dated line in `platform-readings.md` § API 配额 contradict each other on seat identity, and a second line there is overturned by this task's own evidence. Both are measured readings; resolving them is a supersession edit, not a line-budget payment, so I did not land it inside an allowlist PR. Who resolves, and where?",
 "options": [
 "A — resolve in the #11086 whole-text restructuring round, where both lines are already in scope and a reviewer is reading the prose: replace the per-seat-accounts claim with the newer measurement (dated, both readings' provenance kept) and replace the 'a pure-MCP session has no list channel until reset' clause, whose premise this task disproved — the seat ran GET /rate_limit, GET issue/PR/check-runs and PATCH /pulls/11186 at HTTP 200 today. Costs nothing now; the stale lines keep misinforming seats until that round lands.",
 "B — file a dedicated follow-up card now for just these two lines, funded by deleting the superseded text (a genuine 真删, roughly the headroom the routing row needs). Fixes it sooner and self-funds the prose row, but spends a dispatch slot and duplicates #11086's reading pass over the same section.",
 "C — land the supersession as a rider on this PR. Rejected and not done: it would resolve a conflict between two dated measurements inside a permissions diff where no reviewer is looking for it, and it needs a ceiling payment I cannot honestly fund by rewrap."
 ],
 "recommendation": "A, because the conflict is between two dated measurements rather than a defect, and #11086 already has the whole-text reading pass scheduled over exactly this section — that is the round where a reviewer is positioned to judge which reading survives and to carry the routing row in with it. B only wins if seats are actively being misled before #11086 lands; the practical harm is bounded (the stale lines make seats MORE conservative about REST, not less), so it does not justify the extra slot. Note for triage either way: the allowlist landed here is independent of the outcome — it removes prompts, it instructs nobody."
 }
 ],
 "out_of_scope_findings": []
}

---

Generated by Claude Code