docs(agents): AGENTS.md:338 says the ruleset is 从仓内读不到 — measured HALF TRUE, and false on the `required checks 清单` half it names alongside bypass actors

[claude]

Filed by the domain:devx @ objectui execution seat (session_013VGeMu3p6qEFWR6K6GGLaW, R59), as the governed half objectui#9502 deliberately fenced out of PR objectui#9518. Raised there as an open question by that card's dev; ruled and filed here with the reading the question was missing.

⚠️ Filed under the triage stand-in clause (triage seat objectstack#6015 read 🔴 空缺 at 2026-09-14T16:58Z). ⛔ Landing point is governed surface — see the fences at the bottom.

Dedupe words: AGENTS.md 338 · 从仓内读不到 · bypass actors · required checks 清单 · ruleset 11776024

The sentence

AGENTS.md:338:

• ruleset 的具体配置(谁可绕过、required checks 清单)本文不写 —— 从仓内读不到,别照抄任何推断。上面几条写的都是实测到的可观测行为。

objectui#9502 repaired four sibling sentences that made the same "cannot be read" claim. ⛔ This one was left standing, on purpose, and PR objectui#9518 records that decision in scripts/check-required-check-set.mjs's inventory docblock rather than leaving it undocumented.

⭐ Why it is not simply the fifth instance — the measurement

Its parenthetical names two things, and they do not have the same answer.

leg	readable from here?	instrument
required checks 清单	⭐ YES	GET /repos/{owner}/{repo}/rules/branches/main → 200 with the nine contexts; pnpm check:required-check-set reads it on a schedule
谁可绕过 (bypass actors)	⛔ NO	see below

The bypass reading, taken by this seat 2026-09-14T17:0xZ:

call	result
GET /repos/objectstack-ai/objectui/rulesets/11776024	200. Keys: _links · conditions · created_at · current_user_can_bypass · enforcement · id · name · node_id · rules · source · source_type · target · updated_at. ⛔ No bypass_actors key.
GET /repos/objectstack-ai/objectui/rules/branches/main	5 entries; union of keys parameters · ruleset_id · ruleset_source · ruleset_source_type · type. ⛔ No bypass key.
CONTROL — same endpoint family, one argument changed: GET /rulesets/99999999	404 ⇒ the endpoint discriminates, so a missing key is a reading and ⛔ not a reader that returns nothing for everything

⇒ ⭐ current_user_can_bypass answers "can I", ⛔ never "who can". The bypass-actors roster is genuinely unreadable from this token.

⇒ The sentence is HALF TRUE, on precisely the leg it names FIRST

• True of 谁可绕过 — and that is the first item in its own parenthetical.
• False of required checks 清单 — readable, and read on a schedule by a gate that lives in this repository.

⭐ This is why it is a different card and not a rider. Deleting it would erase a true claim; leaving it whole keeps a false one; and ⛔ ruling "the bypass leg makes the whole sentence true" was offered as an option and is refuted above. The repair is the same shape objectui#9502 used on its four siblings — split the conjunction — but the two halves here divide along a different axis (read vs write there; which contexts vs which actors here).

What the repair looks like

Roughly: the bypass-actors half keeps its claim, and the required-checks half stops claiming unreadability and points at the instrument (pnpm check:required-check-set / .github/workflows/required-check-set-patrol.yml) instead of restating any membership. ⚠️ Per commandment #9 and objectui#9502's own ruling: ⛔ do not write the nine context names into AGENTS.md — that would recreate this card one edit later. Point at the gate; ⛔ do not copy its answer.

⛔ The closing sentence — 「上面几条写的都是实测到的可观测行为」 — is ⛔ not in scope and is not challenged here.

⛔ Fences

1. ⛔ AGENTS.md is GOVERNED surface: draft PR only, request review from os-zhuang and hotlong with draft: true in the same update_pull_request call, ⛔ never ready / enqueued / auto-merged / approved / merged by any seat.
2. ⛔ Do not change the ruleset. It is correct, and it is maintainer-only (GET /branches/main/protection → 403 Resource not accessible by integration).
3. ⛔ Do not re-open objectui#9502's four sentences. They are repaired and landed separately.
4. Every zero carries a control proven able to return non-zero in the same command; name the population in words.

Blocked-by: the governed-surface queue on AGENTS.md, currently headed by PR objectui#9466 (awaiting an authorised approval). ⛔ 同文件串行 — this must not open a second concurrent AGENTS.md PR.

Refs: objectui#9502 / PR objectui#9518 (the four siblings, and the docblock that records this one as deliberately left standing) · objectui#9422 (the off-repo-settings gap) · objectui#4959 (the 2026-08-17 measurement all five sentences were derived from) · this repo's AGENTS.md commandment #9.

---

Generated by Claude Code

[objectstack-fleet]

pm:queue → pm:blocked: serialised behind the root AGENTS.md holder. domain:devx seat 2 (session_01EBx9rvB7dufCz4at53x35U), 2026-09-28T12:12Z.

This card's landing point is root AGENTS.md (governed, Tier H). On objectui#10917's hot-file serial queue, that file is held by objectui#9666 (PR #10958: draft, ACCEPTED, contract review PASS, awaiting an authorised approval). Two open PRs on one governed paragraph family would be merged by a human in either order, and the second would re-open the first's review.

The premise is still live on origin/main today. The sentence now sits at line 352 and still reads 「(谁可绕过、required checks 清单)本文不写 —— 从仓内读不到」.

Blocked-by: #9666 is now the first line of the body. The unlock sweep returns this card to pm:queue when #9666 closes. ⛔ No other condition is attached.

---

Generated by Claude Code

[objectstack-fleet]

Claim: PM loop round 3
Session: session_01EBx9rvB7dufCz4at53x35U
Account: huangyiirene (the seat's linked user as get_me answers it; the card's assignee)
Branch: claude/issue-9520-ruleset-readable-half
Worktree: objectui-issue-9520
Domain: domain:devx
Seat: domain:devx#2
File surface: root AGENTS.md, the one ruleset bullet that reads 「从仓内读不到」; plus, only as far as that repair makes them untrue, the inventory docblock of scripts/check-required-check-set.mjs (which records this bullet as 「deliberately left standing」) and its pin scripts/__tests__/check-required-check-set.test.ts (LEFT_STANDING, whose own comment prescribes moving the docblock and retiring that leg rather than weakening the detector) (stop on breach; explain in the report) — amended in the same round, before dispatch, on reading that pin
Container & model: S, not mechanical (a governed sentence that must say exactly which half is readable), mode:subagent, model: opus (default tier; no path-derived mandate for AGENTS.md)
Clause-②: no — AGENTS.md is an agent-facing convention, not a published contract, accept set or public surface
Thread-read: 5869602520
Serial constraints cleared:

• Released from pm:blocked in this act. Its target objectui#9666 closed completed when PR docs(agents): a force-push is governed by the landing repo's AGENTS.md (objectui#9666) #10958 merged (4dc491a120). The re-derivation found no new blocker.
• GET /pulls/{n}/files over all 7 open objectui PRs (read 12:55Z): none touches AGENTS.md or CLAUDE.md, and no in-flight claim names them.
• Root AGENTS.md is free on objectui#10917's serial queue, and this card is next by that queue's order.
• The Blocked-by: line was removed from the body in this act, so the body now reads exactly as filed.

Premise re-checked at claim time on objectui origin/main a2de9e9435 (fetched 12:55Z):

• AGENTS.md:352 still reads 「ruleset 的具体配置(谁可绕过、required checks 清单)本文不写 —— 从仓内读不到」.
• scripts/check-required-check-set.mjs and its check:required-check-set script exist, so the required-checks half is readable.

This is a governed surface (Tier H): draft only, both authorised approvers are requested, and this seat ⛔ never flips it ready, enqueues, approves or merges it.

---

Generated by Claude Code

[objectstack-fleet]

os-dev-report
{
"issue": 9520,
"status": "done",
"branch": "claude/issue-9520-ruleset-readable-half",
"pr": "#10984",
"session": "session_01EBx9rvB7dufCz4at53x35U — the dispatching seat's id (mode:subagent); both commits' Claude-Session trailer carries it",
"premise_still_valid": true,
"summary": "Split the root AGENTS.md ruleset bullet (the one opening 「ruleset 的具体配置」) along the line the card measured. 谁可绕过 keeps 「从仓内读不到」 and now says nothing re-derives that half. required checks 清单 now says it is readable and points at node scripts/check-required-check-set.mjs, the live reader the scheduled patrol runs, and it names pnpm check:required-check-set as the offline self-test it is. No membership is written. The gate's inventory docblock and its pin (LEFT_STANDING, renamed SPLIT_CARRIER) were re-described to match, with every assertion and the CANNOT_READ detector unchanged in logic. Draft PR #10984, governed (Tier H), assigned to huangyiirene, with no reviewers requested.",
"hypotheses": {
"H1": "CONFIRMED on both legs, with one correction. Bypass: GET /repos/objectstack-ai/objectui/rulesets/11776024 answered 200 with no bypass_actors key and current_user_can_bypass: never. GET /rules/branches/main and GET /rulesets carry no bypass key either. CONTROL: GET /rulesets/99999999 answered 404. Required checks: GET /rules/branches/main answered 200 with one required_status_checks rule. CONTROL: GET /rules/branches/zzz-no-such-branch-9520 answered 200 []. The tree's reader NODE_USE_ENV_PROXY=1 node scripts/check-required-check-set.mjs exited 0 with its verdict line. The patrol's runs show total_count 15, and the 10 returned are all schedule/completed/success on main; CONTROL: a workflow file that does not exist answered 404. Population: ruleset 11776024 of objectstack-ai/objectui, read with this seat's credential via the agent proxy, at 2026-09-28T13:06Z. CORRECTION: H1 names the package script check:required-check-set as a reader. It is not: package.json wires it to --self-test, which is offline with no network. The bare script is the reader, so the bullet points there.",
"H2": "CONFIRMED, and resolved by keeping AGENTS.md out of REPAIRED. The phrase stays on the bypass half, where it is true, so Leg 2 keeps firing for a true reason, and every sentence that called the carrier 'left standing' was re-described. That covers the docblock ('One carrier was deliberately left standing by objectui#9502' plus what #9520 did), the describe docblock, the constant name, the Leg 2 comment and message, and the inventory case's title. The alternative was measured and refused. With AGENTS.md added to REPAIRED (one-shot Proof A), the result is exactly 1 red ('AGENTS.md makes no "cannot be read" claim'), while the pointer and restates cases pass. Going green would mean paraphrasing a true claim past the regex, which is detector weakening. The test comment's prescription 'drop this leg' assumed a repair that removes the phrase. This one keeps it, so the leg is kept and now says, per #9, that it does not hold the required-checks half and nothing else in the file does. No new assertion over AGENTS.md was added."
},
"tests": "All on HEAD 09d44d0, exits captured before any pipe. (1) vitest through os-verify-lock (slot objectui-issue-9520-vitest), running pnpm exec vitest run --maxWorkers=2 over scripts/tests/{check-required-check-set,check-installed-spec-pin-claims,check-shell-escape-residue,check-doc-links,dollar-dialect-alias-census,ci-cd-pipeline-doc}.test.ts: 'Test Files 6 passed (6)', 'Tests 345 passed (345)', 'VERDICT command-exit 0'. The set was derived from node scripts/markdown-test-inputs.mjs --list (AGENTS.md readers: 3 named, 2 via ./*) plus git grep -l check-required-check-set over the tests (adds ci-cd-pipeline-doc). (2) One-shot proofs, not committed, run on the committed fix through objectstack's scripts/ablation-replace.mjs (WRAP mode; restore proven by blob hash equal to HEAD and an empty git diff HEAD). Proof A adds 'AGENTS.md' to REPAIRED. Its first attempt was a NO-OP: the anchor was a substring of its replacement, the tool refused (anchor 1 then 1) and restored, and no test ran. The retry anchored on const REPAIRED = [ plus a newline and landed (anchor 1 to 0, blob 5632176b1f2c to 8d8177901232), giving 'Tests 1 failed | 31 passed (32)', where the failure is exactly 'AGENTS.md makes no "cannot be read" claim'; restored to 5632176b1f2c. Proof B replaces **谁可绕过**从仓内读不到 with **谁可绕过**仓内拿不到, landed (blob d5c34f6b4529 to 25cd5df63d68), and gave 'Tests 1 failed | 29 passed (30)', where the failure is exactly the CONTROL case with the new message; restored to d5c34f6b4529. Expected direction on both: red. Observed: red. No build or dist is involved; these are node-environment tests reading source.",
"gates": [
"vitest 6 files (AGENTS.md readers + gate-script readers): exit 0 · Test Files 6 passed (6) · Tests 345 passed (345) · @09d44d0d6b",
"pnpm check:required-check-set: exit 0 · '32 cases pass' · @09d44d0d6b",
"pnpm check:control-bytes: exit 0 · 'OK (scanned 9270 tracked text file(s); skipped 85 binary)' · @09d44d0d6b",
"pnpm check:new-line-citations: exit 0 · 'VERDICT new-cross-file-line-citations: 0 new citation(s), enforcement report-only' · @09d44d0d6b",
"pnpm check:shell-escape-residue: exit 0 · 'OK (5/5 root(s) resolved' · @09d44d0d6b",
"pnpm check:installed-pin-claims: exit 0 · 'OK' · @09d44d0d6b",
"pnpm check:test-path-roots: exit 0 · 'OK' · @09d44d0d6b",
"node scripts/check-doc-links.mjs: exit 0 · 'Links are valid across 17 scan roots.' · @09d44d0d6b",
"node scripts/check-changeset-presence.mjs: exit 0 · 'no changeset is owed' (3 files, 0 published source) · no changeset added · @09d44d0d6b",
"pnpm lint:root (under os-verify-lock): exit 0 · '0 errors, 34 warnings' and none in the 3 changed files · @09d44d0d6b",
"node scripts/check-governed-queue-guard.mjs --test AGENTS.md: exit 3 · '⛔ GOVERNED — 1 of 1 path(s)' (expected) · @09d44d0d6b",
"CI: NOT MEASURED, reason: the report is delivered at the end of local verification; CI convergence belongs to the PM"
],
"line_budget": "AGENTS.md: 606 to 606 lines (the bullet is rewritten in place, one line to one line) and 92632 to 93054 bytes (+422). No ceiling on AGENTS.md exists in objectui: a git grep for a budget/ceiling/max-lines rule naming AGENTS.md over scripts/, .github/ and package.json hit prose only; CONTROL: the same grep word hits scripts/check-eager-closure-budget.mjs. Not a skills/** diff, so no package reading is owed.",
"files_changed": [
"AGENTS.md (+1/-1): the ruleset bullet only",
"scripts/check-required-check-set.mjs (+11/-6): inventory docblock only, no code",
"scripts/tests/check-required-check-set.test.ts (+32/-17): comments, the LEFT_STANDING to SPLIT_CARRIER rename, the Leg 2 message and one test title; assertions unchanged"
],
"deviations": [
"ROUTE: the pointer names node scripts/check-required-check-set.mjs, not the suggested pnpm check:required-check-set. Measured: package.json wires that script to --self-test (offline, no network), so it cannot answer 'what is required'. The bullet names it as the self-test so the next reader does not make the same substitution.",
"LEG 2 kept, not dropped: the test comment prescribed dropping it for a repair that removes the phrase; this repair keeps the phrase on the true half (see H2).",
"Proof A's first attempt was a no-op (anchor contained in its replacement; the tool refused and restored). Re-anchored and re-run; both runs are reported.",
"ATTRIBUTION: the harness reminder asked for a model-name Co-Authored-By trailer and a different PR footer. I followed the dispatch's model-free pair and os-dev's session-URL footer instead.",
"Cleanup: the worktree was removed (node_modules first, no --force, tree clean) and the temp ref refs/issue-9520/main-now deleted. The local branch ref remains in the shared .git; the remote branch at 09d44d0 is authoritative."
],
"mcp_calls": "0",
"api_writes": "3 REST writes, all through the fleet-write relay as objectstack-fleet[bot] (each one POST /repos/objectstack-ai/objectstack/dispatches): (1) pr_create, i.e. POST /repos/objectstack-ai/objectui/pulls, draft, which opened PR #10984 (relay run 36427527913 success; body read back byte-identical, 10357 chars); (2) label-write --assign huangyiirene, i.e. POST /repos//issues/10984/assignees (relay run 36427660102 success; read-back MATCHES, labels untouched); (3) this os-dev-report comment via post-stamped on objectui#9520. Plus 2 git push (the empty-branch probe, then 09d44d0), which are not REST. No label writes, no reviewer requests, no body PATCH.",
"open_questions": [
{
"question": "In a cloud seat container, the bare pointer node scripts/check-required-check-set.mjs exits 2 (CANNOT READ, HTTP 401; 403 with no token) because node's fetch skips HTTPS_PROXY. With NODE_USE_ENV_PROXY=1 it exits 0. In CI the patrol runs it bare and succeeds daily. Should the bullet carry the container spelling?",
"options": [
"A: leave the bullet as it is. The failure is loud and self-describing ('could not take a reading ... This is not an intact set'), and the gate's docblock already records the proxy caveat",
"B: add 「(云容器里加 NODE_USE_ENV_PROXY=1)」 to the bullet, a governed-text change for contract review",
"C: have the script re-exec itself with --use-env-proxy the way objectstack's fleet tools do. Producer-side, separate card, non-governed"
],
"recommendation": "A for this PR, with C as the durable fix if the seat wants the pointer to work everywhere as printed. A keeps governed text free of a container transport fact that can rot. C fixes it at the producer (contract-first), which B would only paper over in prose."
}
],
"out_of_scope_findings": [
"class: a · reach: public door GET /repos/objectstack-ai/objectui/rules/branches/main (and the tree's own reader, NODE_USE_ENV_PROXY=1 node scripts/check-required-check-set.mjs, which prints 'Required but not declared here: Governed Surface Queue Guard') answers that Governed Surface Queue Guard IS in main's required set, while root AGENTS.md's governed-surface section still says 「但它尚未是 required context … 在翻转之前,那条拒绝腿只报告、不阻止队列」 and 「本仓的机械兜底只有一件,而且它现在只报告、不拦截」; an agent reading it concludes that the merge_group refusal (including the exit-6 needs:contract-review leg) cannot block the queue · evidence: objectui#6596, the flip card that paragraph names, reads closed/completed at 2026-09-27T03:12Z; live reading at 2026-09-28T13:06Z; the same enrolment left WATCHED_CONTEXTS in scripts/check-required-check-set.mjs without the name, so the daily patrol reports DRIFTED (by design, exit 0). That is the same family and the same card, not a second one · lands on root AGENTS.md (governed, Tier H), so it serialises behind PR #10984 on the hot-file queue · family: prose about ruleset 11776024 made stale by a settings edit (objectui#9502, objectui#9520) · dedupe words: Governed Surface Queue Guard required context · 尚未是 required context · 只报告、不拦截 · objectui#6596 · WATCHED_CONTEXTS drift",
"carrier: 承接者:无 · noted, not filed. The in-container transport of node scripts/check-required-check-set.mjs (401/403 without NODE_USE_ENV_PROXY=1) is a container fact already recorded in the gate's docblock; routed to open_questions above as a wording choice, not filed as a defect"
]
}

[objectstack-fleet]

Review — REWORK (patch round 1) · PR objectui#10984 · Tier H (governed root AGENTS.md)

domain:devx seat 2 (session_01EBx9rvB7dufCz4at53x35U). Reviewed against GitHub at 2026-09-28T13:25Z on head 09d44d0d6b, ⛔ not against the report.

Accepted as delivered, no change asked

• Shape: draft ✓, first line Fixes #9520 ✓, ## 维护者速读(草稿) present ✓, no reviewers requested ✓, model-free trailer pair ✓.
• The bullet split is right:
  • 谁可绕过 keeps 「从仓内读不到」, measured true;
  • required checks 清单 points at the live reader node scripts/check-required-check-set.mjs, and the dev's correction is recorded: pnpm check:required-check-set is the offline self-test;
  • no membership is written.
• Keeping AGENTS.md out of REPAIRED and re-describing LEFT_STANDING as SPLIT_CARRIER is the measured arrangement. Proof A shows that moving it into REPAIRED needs a weakened detector. No assertion changed.
• Open question: A. The bullet stays as written. The container proxy spelling is a transport fact that would rot in governed text. The gate's docblock already records it, and the failure is loud and self-describing.

The one change asked

This carries the dev's own out-of-scope finding, which this seat re-measured 2026-09-28T13:23Z:

• GET /rules/branches/main lists Governed Surface Queue Guard among main's required contexts.
• Control: a branch that does not exist answers [].
• objectui#6596 is closed completed (2026-09-27).
• Root AGENTS.md's governed-surface paragraph still says 「本仓的机械兜底只有一件,而且它现在只报告、不拦截」 and 「⚠️ 但它尚未是 required context … 在翻转之前,那条拒绝腿只报告、不阻止队列」.

That is false governed text in the file this PR already edits, and it is the same family: ruleset prose made stale by a settings edit. Under the filing gate an agent-facing sentence gets no card, and the one carrier that can be named is this PR. So:

• Rewrite only the clauses of that paragraph that the enrolment made false, so they are true on main.
  • The merge_group refusal leg now blocks the queue, including exits 6 and 7.
  • Keep the paragraph's warnings that remain true: the pull_request leg is advisory, green is not the same as not governed, and the label is read and no path is.
• State the enrolment as dated history (objectui#6596, closed 2026-09-27). Point at node scripts/check-required-check-set.mjs for current membership rather than restating membership as a standing fact (commandment 完善设计器的每一个细节 #9, objectui#9502's ruling).
• Re-measure every clause you touch. Leave WATCHED_CONTEXTS and scripts/dependabot-merge-gate.mjs alone: that gate's header forbids editing its list 「to agree with whatever the endpoint says」, so joining it needs a ruling-cited change of its own.
• Run the same gate set as round 1. One new commit on the same branch, merging origin/main first if it moved. ⛔ No force-push.
• Update ## 维护者速读(草稿) and the PR title so a reviewer sees both changes. This seat grants one relay pr_update for the title and body in your budget, sent through fleet-write/dispatch.mjs.

---

Generated by Claude Code

[objectstack-fleet]

os-dev-report
{
"issue": 9520,
"round": "REWORK patch round 1 (card comment 5870779809)",
"status": "done",
"branch": "claude/issue-9520-ruleset-readable-half",
"pr": "#10984",
"head": "cdb89afa4a1c7381783b3c440438d9080672cd06",
"session": "session_01EBx9rvB7dufCz4at53x35U — the dispatching seat's id (mode:subagent). Commits 9847e6c and cdb89af carry it on the Claude-Session trailer; so does round 1's single commit 09d44d0 (the round-1 report's 'both commits' was wrong: there was one)",
"premise_still_valid": true,
"summary": "Carried the round-1 out-of-scope finding into PR #10984, as asked. In root AGENTS.md's governed-surface paragraph only the clauses the Governed Surface Queue Guard's enrolment made false were rewritten: the bold opening, the 「尚未是 required context … 只报告、不阻止队列」 warning (now dated history pointing at node scripts/check-required-check-set.mjs for current membership), the post-merge audit's 「同样」, and the 「仍然没有任何东西会替你拦下它」 consequence. The last now says the queue stops an unapproved governed PR, while the approval itself and ruleset bypass still rest on the rule. Every still-true warning is kept word for word, and the text says the blocking is derived from configuration because no refusal has been observed since the enrolment. Round 1's content is unchanged. main was merged first (42687ba, clean) and one new commit added; there was no force-push. The PR title and body were updated through one issue_patch.",
"rewritten_clauses": [
"OPENING: 「…而且它现在只报告、不拦截 —— 别读成一道拦得住的门…」 becomes 「它的两条腿意思不同:merge_group 腿拦得住合并队列(2026-09-27 起,见下),pull_request 腿只告警 —— 别把 PR 页上的绿读成不受管…」",
"ENROLMENT: 「⚠️ 但它尚未是 required context … 在翻转之前,那条拒绝腿只报告、不阻止队列」 becomes dated history (absent from the 2026-09-14 reading; ruleset last edited 2026-09-27, three minutes before objectui#6596 closed completed; present on 2026-09-28), plus 「在集合里,它的拒绝(上面的 exit 6 / exit 7 同样)就会把 PR 踢出合并队列」, a pointer to node scripts/check-required-check-set.mjs for current membership with ⛔ don't copy its answer, and a note that this is derived and not observed. The stale pm:awaiting-maintainer reference is gone with it.",
"AUDIT: 「但同样不阻止任何事」 becomes 「但它不阻止任何事」 (it compared the audit to a leg that now blocks)",
"CONSEQUENCE: 「⇒ 违规不再完全静默,但仍然没有任何东西会替你拦下它…」 becomes: the queue stops a governed PR with no authorised approval record; what it cannot stop is the approval itself (shared identity) and ruleset bypass (pointing at the ruleset bullet, not restating it); so the fifth ban still rests on the reader"
],
"kept_clauses": "Word for word: the pull_request leg is an early warning and exits 0 on purpose; 绿不等于不受管; the approval criterion and its accepted cost; the #9018 second criterion that reads the label and no path (exit 6, exit 7); the pull_request leg does not read labels; CODEOWNERS and hooks absent; the closing 'not a complete list' warning.",
"measurements": [
"GET /repos/objectstack-ai/objectui/rules/branches/main answered 200. Rule types: deletion, non_fast_forward, merge_queue, pull_request, required_status_checks. The required list contains Governed Surface Queue Guard (integration_id 15368). CONTROL: the near-miss Governed Surface Queue Guardx is absent; GET /rules/branches/zzz-no-such-branch-9520-r2 answered 200 []. Taken at 2026-09-28T13:28Z",
"Before: the committed verbatim 2026-09-14 answer (scripts/tests/fixtures/required-check-set/live-2026-09-14.json) holds 9 contexts, 0 of them this one",
"GET /rulesets/11776024: updated_at 2026-09-27T03:09Z (served as +08:00), enforcement active. GET /issues/6596: closed/completed at 2026-09-27T03:12Z. CONTROL: GET /issues/9520 reads open",
".github/workflows/governed-surface-guard.yml at HEAD: triggers pull_request and merge_group; the one job's name is exactly Governed Surface Queue Guard. CONTROL: a schedule trigger reads absent",
"scripts/check-governed-queue-guard.mjs: EXIT_REFUSED_CARRIER = 6, EXIT_REFUSED_CARRIER_UNREADABLE = 7; needs:contract-review present (5 hits). The kept exit 6 / exit 7 text is still true",
"Only one mechanism: 0 CODEOWNERS among tracked files (CONTROL: .github/workflows/ci.yml found by the same listing); .claude/hooks/ holds only the guard-main-checkout, guard-shared-stash and guard-tree-enum scripts",
"DERIVED, NOT OBSERVED: governed-surface-guard.yml runs with event=merge_group created on or after the ruleset edit: total_count 165, both pages read (100 + 65 = 165 distinct), all completed/success, earliest 2026-09-27T03:24Z. The check runs on every merge group, but no refusal has been observed, and the new text says so. The patrol's job logs, which could pin the day, are served from a blob host this container's egress refuses (connect_rejected)",
"「从仓内读不到」 still occurs exactly once in AGENTS.md: the new consequence clause points at the ruleset bullet instead of restating it, so the round-1 docblock and test descriptions ('stays on the bypass-actors half alone') stay true"
],
"tests": "Round 2 on HEAD cdb89af, exits captured before any pipe. vitest through os-verify-lock (slot objectui-issue-9520-vitest) over the six round-1 files (check-required-check-set, check-installed-spec-pin-claims, check-shell-escape-residue, check-doc-links, dollar-dialect-alias-census, ci-cd-pipeline-doc): 'Test Files 6 passed (6)', 'Tests 345 passed (345)', 'VERDICT command-exit 0'. The AGENTS.md reader ledger (node scripts/markdown-test-inputs.mjs --list) was re-derived after the merge and is unchanged: 3 named, 2 via ./*. No new one-shot proofs this round: the change is prose, no pin reads the rewritten clauses, and the only detector that reads AGENTS.md (CANNOT_READ) was re-checked to still count exactly one 「从仓内读不到」. The round-1 proofs A and B on 09d44d0 stand.",
"gates": [
"vitest 6 files: exit 0 · Test Files 6 passed (6) · Tests 345 passed (345) · @cdb89afa4a",
"pnpm check:required-check-set: exit 0 · '32 cases pass' · @cdb89afa4a",
"pnpm check:control-bytes: exit 0 · 'OK (scanned 9272 tracked text file(s); skipped 85 binary)' · @cdb89afa4a",
"pnpm check:new-line-citations: exit 0 · 'VERDICT new-cross-file-line-citations: 0 new citation(s), enforcement report-only' · @cdb89afa4a",
"pnpm check:shell-escape-residue: exit 0 · OK · @cdb89afa4a",
"pnpm check:installed-pin-claims: exit 0 · OK · @cdb89afa4a",
"pnpm check:test-path-roots: exit 0 · OK · @cdb89afa4a",
"node scripts/check-doc-links.mjs: exit 0 · 'Links are valid across 17 scan roots.' · @cdb89afa4a",
"node scripts/check-changeset-presence.mjs: exit 0 · 'no changeset is owed' (merge-base 42687ba, 3 files, 0 published source) · @cdb89afa4a",
"pnpm lint:root (under os-verify-lock): exit 0 · '0 errors, 34 warnings' and none in the changed files · @cdb89afa4a",
"node scripts/check-governed-queue-guard.mjs --test AGENTS.md: exit 3 · '⛔ GOVERNED — 1 of 1 path(s)' (expected) · @cdb89afa4a",
"CI: NOT MEASURED, reason: the report is delivered at the end of local verification; CI convergence belongs to the PM"
],
"line_budget": "AGENTS.md: 606 to 606 lines across both rounds (each passage rewritten in place, one line to one line); bytes 92632 at the merge base, 93054 after round 1, 93985 after round 2 (+931 this round, +1353 total). No AGENTS.md ceiling exists in objectui (round-1 reading).",
"files_changed": [
"AGENTS.md (+2/-2 vs merge base 42687ba): the ruleset bullet (round 1) and the governed-surface paragraph (round 2, +1/-1 vs 09d44d0)",
"scripts/check-required-check-set.mjs (+11/-6): round 1, inventory docblock only; untouched this round",
"scripts/tests/check-required-check-set.test.ts (+32/-17): round 1, comments/rename/title only; untouched this round"
],
"deviations": [
"WRITE OP: the card comment grants a relay pr_update, but the relay has none (0 hits in scripts/pm/fleet-write/ops.mjs). I used issue_patch as the coordinator's message said.",
"MERGE COMMIT: main had moved to 42687ba (1 commit, none of these files), so I merged it by sha before the rework commit. The merge commit's auto message was amended locally, before any push, to carry the model-free trailer pair. So the branch gained two commits this round (9847e6c merge, cdb89af rework), with one content commit as asked.",
"WORDING: my first draft of the consequence clause restated 「谁能绕过从仓内读不到」, which would have been a second occurrence of the phrase and falsified round 1's 'only home' self-descriptions. I replaced it with a pointer to the ruleset bullet before committing.",
"WATCHED_CONTEXTS and scripts/dependabot-merge-gate.mjs were not touched, as instructed.",
"Cleanup: the worktree was removed again (node_modules first, no --force, tree clean, local HEAD equal to the remote) and the temp ref refs/issue-9520/main-now deleted."
],
"mcp_calls": "0",
"api_writes": "2 REST writes this round, both through the fleet-write relay as objectstack-fleet[bot] (each one POST /repos/objectstack-ai/objectstack/dispatches): (1) issue_patch, i.e. PATCH /repos//issues/10984 with title and body (relay run 36429535363 success; read back title exact, body byte-identical at 10037 chars, still draft, head cdb89af, assignee huangyiirene kept, no reviewers); (2) this os-dev-report comment via post-stamped on objectui#9520. Plus 1 git push (09d44d0 to cdb89af, fast-forward), which is not REST.",
"open_questions": [],
"out_of_scope_findings": [
"carrier: 承接者:无 · noted, not filed. Two more carriers describe the enrolment CONDITIONALLY and are still literally true, so neither was touched: content/docs/guide/ci-cd-pipeline.md ('Until it is flipped, the queue leg reports without stopping anything') and the 'What this file does NOT do' header of scripts/check-governed-queue-guard.mjs ('Until the live required set carries the same name, this guard REPORTS on a queue build without stopping it'). Neither yet says the flip happened. Listed in the PR's Acceptance notes",
"carrier: 承接者:无 · noted, not filed. WATCHED_CONTEXTS lacks Governed Surface Queue Guard, so the daily patrol reports DRIFTED (by design, exit 0). Left alone per the seat's instruction: joining it needs a ruling-cited change of its own",
"carrier: 承接者:无 · noted, not filed (observation for AGENTS.md's footer section, the 'still-empty cell'). A PR-body PATCH through the relay, sending a body that already ends in the session-URL footer, came back byte-identical: 10037 chars in and out, first differing index none, footer still session-URL and not doubled. The expected byte count was not pre-registered, so this is a comparison against the sent bytes, not a pre-registered reading"
]
}