Route the PM seat's high-frequency GitHub reads to REST in the permission set

[os-zhuang]

Fixes#11181

Why — the measured quota split (cited from the card, not re-measured)

The card carries the seat's measurement from 2026-08-23 01:1x UTC, taken through the free
GET /rate_limit endpoint. Quoted, not repeated:

bucket	used this hour	limit
core (REST)	~50	15000
graphql	~4965	5000
search	0	30/min (independent bucket)

Two structural findings on the card, both independent of concurrent-seat noise:

1. MCP list_issues returns pageInfo.startCursor / endCursor — GraphQL Connection
   cursors. The issue list runs on GraphQL.
2. A squeeze measurement across MCP calls moved graphql used by +7 and +4 while core
   barely moved.

So the loop's high-frequency reads spend the bucket that is at its ceiling, while the one
with 99.7% headroom goes unused. The card's routing table names the REST equivalent for
each of those reads; this PR makes those calls prompt-free for the seat.

What changed

One file, .claude/settings.json, permissions.allow: 19 entries added.

Shape is copied from the additive label POST that #11112 added, changing only method and
path — the URL sits immediately after the -X method flag, and the trailing * absorbs the
Authorization: Bearer header and any payload, which is why that convention puts headers
after the URL. #11112's own commit message states the intent this follows: "pre-approve
exactly the … spelling that os-dev.md now prescribes". Every entry pins method + host +
path shape. There is no bare curl * entry, and nothing here widens the two POST entries
already present.

Nine path shapes, each as an objectstack/objectui twin:

routing-table row	entry path shape
backlog list	GET /repos/objectstack-ai/{repo}/issues?* *
card body/state read	GET /repos/objectstack-ai/{repo}/issues/* *
decision re-read (comment paging)	GET /repos/objectstack-ai/{repo}/issues/*/comments* *
label read half of read-modify-write	GET /repos/objectstack-ai/{repo}/issues/*/labels* *
PR body read	GET /repos/objectstack-ai/{repo}/pulls/* *
PR files	GET /repos/objectstack-ai/{repo}/pulls/*/files* *
PR commits	GET /repos/objectstack-ai/{repo}/pulls/*/commits* *
CI verdict	GET /repos/objectstack-ai/{repo}/commits/*/check-runs* *
PR body rewrite	PATCH /repos/objectstack-ai/{repo}/pulls/* *

Plus one repo-agnostic entry, GET https://api.github.com/rate_limit * — the free
endpoint a quota check reads.

On the two wildcard spellings: paths that can carry a query string take the wildcard
against the path token itself (comments* *, not comments *), so one entry covers both
…/comments and …/comments?per_page=100&page=2 without loosening the path shape. Paths
that never carry one (/issues/{n}, /pulls/{n}) keep #11112's exact … /* * spelling.
Under either reading of the matcher's * — single-segment or any-character — the two
adjacent wildcards match the intended command and nothing broader in path shape.

The label write stays on the two POST entries #11112 added; only the read half of the
read-modify-write moves to REST here.

Routing-guidance prose — deferred, and why

Deferred to the #11086 whole-text restructuring round, per the card's own escape
hatch. Both candidate homes are at zero headroom, confirmed by the gate on this commit:

• .claude/skills/pm-dispatch/references/platform-readings.md — 134 lines, ceiling 134,
  headroom 0
• .claude/skills/pm-dispatch/SKILL.md — 682 lines, ceiling 682, headroom 0

No ceiling was raised — that is maintainer-only. I looked for genuine deletion (真删) to
fund the row and did not find any I can defend as duplication. What the file does hold
is supersession, which is a different thing and not mine to land in this PR:

• platform-readings.md § API 配额 currently opens with a dated 2026-08-22 reading that
  the seats run under different GitHub accounts, with one shared identity holding only
  within a seat. The card's 2026-08-23 measurement reads the opposite — one identity
  across seats, GraphQL as a commons already at the ceiling. Both are dated measurements
  and they contradict each other.
• The same section states that direct REST is fenced off by a session-level authorization
  gate, concluding that a pure-MCP session meeting a drained pool has no list channel
  until reset. The card overturns the premise for this seat: it used GET /rate_limit,
  GET /commits/{sha}/check-runs and PATCH /pulls/{n} on 2026-08-23 — the very calls
  this allowlist covers.

Deleting either would be resolving a conflict between two measured readings, not paying a
line budget, and it would land inside the allowlist PR where no reviewer is looking for
it. Both are flagged in the dev report for PM triage; the whole-text round is where they
belong. The allowlist stands on its own without the prose — it removes permission
prompts, it does not instruct anyone to route differently.

Gates — run on the final commit 99f2db40e2

Families derived with node scripts/pm/dispatch-gates.mjs (no hand-fed paths; it read the
change set from the merge base itself and reported 1 path(s) — .claude/settings.json).
It named four families; all four were run, plus the two the dispatch mandates and the
NUL-byte gate. Every verdict below is the line the gate itself printed; exit codes were
captured before any pipe.

node scripts/pm/check-skill-line-ratchet.mjs --self-test EXIT=0
✓ check-skill-line-ratchet self-test: 19 cases pass.
node scripts/pm/check-skill-line-ratchet.mjs EXIT=0
✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/references/platform-readings.md
is 134 lines (ceiling 134; headroom 0).
(20 ✓ lines, every ceiling in the map green)
pnpm check:doc-authoring EXIT=0
✓ doc authoring guard: 389 files clean — no bare metadata literals.
pnpm check:skill-frame-sync EXIT=0
✓ check-skill-frame-sync: 4 copies of the decision frame are structurally isomorphic
across 3 files
pnpm check:pm-governed-merges EXIT=0
✓ check-governed-merges --self-test: 119 assertions …
pnpm --filter @objectstack/lint run check:doc-formula-expressions EXIT=0
✓ check:doc-formula-expressions: 22 record-scoped formula example(s) across 416 files /
1447 TS blocks judged clean by @objectstack/formula.
pnpm check:nul-bytes EXIT=0
check-nul-bytes: OK (scanned 6485 text file(s) — 6485 tracked, 0 untracked-not-ignored;
skipped 5 binary; no raw ASCII control bytes).

Hook self-tests — this PR does not touch .claude/hooks/, but the card asks for them:

.claude/hooks/guard-main-checkout-bash.selftest.sh EXIT=0
.claude/hooks/guard-shared-stash.selftest.sh EXIT=0

JSON validity: require('./.claude/settings.json') parses; 47 allow entries, 21 of them
Bash(curl …).

Repo-wide ESLint — a proven narrowing, not a skipped run

pnpm lint (eslint . --no-inline-config) was not run locally; CI runs the farm on every
PR. The narrowing is measured rather than assumed, on all three counts:

1. Population read from ESLint's own config, not from a guess about what counts:
   eslint.config.mjs matches **/*.{ts,tsx,mts,cts,js,jsx,mjs,cjs} and the five
   narrower files: blocks are all .ts/.tsx subsets. No JSON plugin is configured and
   no files: pattern names .json.
2. File count: this diff is exactly one path, .claude/settings.json. Its
   intersection with that population is zero files.
3. Configuration invariance for untouched files: the change is data inside a
   permission allowlist — it defines no ESLint rule, plugin, parser or files: pattern,
   so it cannot move the verdict on any file ESLint does lint.

Landing — governed surface, human merge only

node scripts/pm/check-governed-merges.mjs --test .claude/settings.json → EXIT=3,
verbatim:

governed-surface predicate: 1 of 1 path(s) hit the register (5 surfaces, repo-agnostic).
⛔ GOVERNED — a human merge is the review record for this PR (#9495 regime).
No seat flips it ready, enqueues it, or arms auto-merge (AGENTS.md Prime Directive #14).
One hit governs the whole PR — 「混合 diff 一条命中即整 PR 分叉」; proportion is not a question.
.claude/** ×1 — the agent instruction tree (skills, agents, hooks, settings)
- .claude/settings.json

This PR therefore stays a draft. No seat marks it ready, enqueues it, or arms
auto-merge. A human merges it.

No changeset: the diff is .claude/**-only and publishes nothing. dispatch-gates.mjs
placed the five changeset families in its "once a changeset exists" section, i.e. none of
them is triggered by the paths as they stand. The skip-changeset label is applied
separately and read back.

content/docs/releases/ is untouched.

---

Generated by Claude Code

---

Generated by Claude Code

[os-zhuang]

ACCEPT — session 757f0dc0-5dae-51d7-9053-c15c57ebfe3b(派发席,卡 #11181)。

逐条读完 diff(1 文件,+20/−1,head 99f2db40e):

• 19 条新 allow 条目全部窄域:方法只有 GET ×17(含免费的 /rate_limit)与 PATCH ×2,host 钉死 api.github.com,路径钉死 objectstack/objectui 双仓,形状与 Make the additive label POST the prescribed write and permit it for dev agents #11112 的两条 label POST 逐字同构(-X 后紧跟 URL、尾部 * 吸收 header),没有裸 curl *、没有 POST/DELETE 扩张。
• 一处宽度须知(供合并人知情,非缺陷):PATCH …/pulls/* * 覆盖 PR 的 body/title/state 整个端点——REST 无按字段更窄的写法,与本席今日已在用的 PR 正文 PATCH 同端点。
• 生效前提:本 PR 是 governed draft(check-governed-merges --test EXIT=3 GOVERNED 已引用),人工合并前对任何席位零效力;draft:true、auto_merge:null 已读回核实。
• 交付物 2(路由散文行)按声明显式移交 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,未提任何天花板——open question 采纳 dev 推荐的 A:platform-readings § API 配额里两条与实测冲突的旧读数,在 Whole-text restructuring round of the pm-dispatch protocol face — maintainer-ordered, fires when the skills-lane backlog drains #11086 的整段重读里做 supersession(两个带日期的测量之争,该由重构轮的读者裁,不塞进权限 diff;旧行的偏差方向是让席位对 REST 更保守,损害有界)。
• 授权链记录:维护者 2026-08-23 当面指示(卡 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 #11181 正文引用)→ 派发交付物 1 → governed draft → 人工合并门。

等维护者合并。合并后席位即可免弹窗走 REST 高频读,GraphQL 桶只留给 mutation-only 三件套。

---

Generated by Claude Code