AGENTS.md never tells a seat not to submit an approving review on a governed pull request — the rule the retired sha pin made weightier

[claude]

Found while rewriting the two stale sha-pin passages for #7605 (PR #7616). Filed rather than fixed there: that dispatch scoped the prose to the ruled predicate and its accepted cost, and this is a different sentence.

The gap

scripts/check-governed-queue-guard.mjs states a normative seat rule in its header, and its refusal text prints it too:

⛔ An agent seat never submits an approving review on a governed-surface pull request, under any account. Every seat in this repository writes under a shared GitHub identity, so GOVERNED_APPROVERS is a technical control that is only as good as that normative rule.

AGENTS.md § 受管面 does not carry it. What it carries (the paragraph beginning 「⛔ 绝不 gh pr ready」) is the adjacent set — never ready, never enqueue, never auto-merge, never self-merge, plus 「别去等一个不存在的 approval」. A seat that reads only AGENTS.md learns not to merge and not to wait for an approval; it is never told not to leave one.

Why it got weightier on 2026-09-04

The maintainer's ruling (executed as PR #7606, merged 0d8fd7ce) retired the sha pin, and the guard's header says out loud what that does to this rule:

⚠️ With the pin retired that rule carries MORE weight, not less: one approval now clears every later push on the same pull request.

Before the ruling, a seat-submitted approval decayed on the next push. It no longer does — one such approval clears the queue leg for every commit that follows on that pull request. GOVERNED_APPROVERS restricts which account may approve, and under a shared GitHub identity that is not the same thing as restricting who may approve.

Suggested shape

One sentence in AGENTS.md § 受管面, next to the existing ⛔ list, copied from the guard header rather than paraphrased — the same discipline #7605 applied to the predicate, so the three places cannot drift. AGENTS.md is a governed surface, so this is a draft pull request parked for a human merge; that is the regime, not an obstacle.

Filed by the objectui skills seat working #7605, session session_019RfFHiRCSs3JXLK4cwcfox, under a shared GitHub identity — the branch on that card's pull request is the identifying detail, not the author field. Left unassigned and unlabelled for PM triage.

---

Generated by Claude Code

[claude]

Triage — skills seat (session session_019RfFHiRCSs3JXLK4cwcfox, os-steve), 2026-09-04T07:46Z (the seat grades its own programme's follow-ups while the objectui triage seat is vacant, per the #7359 precedent). Graded priority:p3, domain:skills, documentation, state pm:queue. Applied and read back after this comment.

Scope as queued (XS, serial behind PR #7616 — same AGENTS.md § 受管面 paragraph, so it lands after that governed PR merges): add the guard header's normative seat rule to AGENTS.md § 受管面 beside the never-ready / never-enqueue / never-auto-merge / never-self-merge list — copied from scripts/check-governed-queue-guard.mjs (the sentence beginning "⛔ An agent seat never submits an approving review on a governed-surface pull request, under any account"), in the section's language and voice, with the reason the retired pin made it weightier (one approval now clears every later push). Governed (AGENTS.md) ⇒ draft PR, both approvers, human merge. ⛔ No other prose.

---

Generated by Claude Code

[claude]

Claim: PM loop round 5
Session: session_019RfFHiRCSs3JXLK4cwcfox
Branch: claude/issue-7618-agents-never-approve-rule
Worktree: objectui-issue-7618
Domain: domain:skills
File surface: AGENTS.md only — § ⛔ 受管面, beside the never-ready / never-enqueue / never-auto-merge / never-self-merge list (near L488 at 11edab8): add the guard header's normative seat rule, copied from scripts/check-governed-queue-guard.mjs (L93–96: "⛔ An agent seat never submits an approving review on a governed-surface pull request, under any account. Every seat in this repository writes under a shared GitHub identity, so GOVERNED_APPROVERS is a technical control that is only as good as that normative rule."), quoted verbatim inside the section's Chinese voice with the reason the retired pin made it weightier (one approval now clears every later push). ⛔ No other prose; ⛔ no change to the guard.
Container & model: XS, mode:subagent, model: opus (governance prose that must keep the rule's force)
Clause-②: no (prose only)
Serial constraints cleared: objectui PR #7616 (#7605) MERGED 11edab88 11:00:03Z — the same paragraph region now states the unpinned predicate; no open PR touches AGENTS.md. GOVERNED (AGENTS.md) ⇒ draft PR, in-seat review, both authorized approvers requested by the seat, human merge — ⛔ never ready / queue / auto-merge from the seat.

Decision re-read (11:27Z): no ruling owed — the sentence already stands as the guard's own normative rule (2026-08-27 approver ruling, 2026-09-04 unpinning); the card only mirrors it into the instruction surface, scoped by the 07:46Z triage.

---

Generated by Claude Code

[claude]

os-dev-report

{
"issue": 7618,
"status": "done",
"branch": "claude/issue-7618-agents-never-approve-rule",
"pr": "#7630",
"premise_still_valid": true,
"summary": "Added the guard header's normative seat rule to AGENTS.md § 受管面 as a fifth prohibition beside the existing never-ready / never-enqueue / never-auto-merge / never-self-merge line (L488 at 11edab8), which is untouched. The diff is +6 lines in AGENTS.md and nothing else: a short Chinese lead-in naming both places the rule already lives (the file header of scripts/check-governed-queue-guard.mjs and the cleared branch of its verdict text), the English sentence as a blockquote, and one Chinese clause carrying why the retired sha pin made it weightier (maintainer ruling 2026-09-04, executed as #7606, written into the paragraph below by #7616): one authorized APPROVED review now clears every later push, so the prohibition is the only thing between an agent-operated approver account and a governed landing it cleared for itself. STOP-CONDITION FIRED, and I took the instructed branch. The sentence quoted by the card and by the dispatch ends '...only as good as that normative rule.' with a terminal period; at origin/main the sentence does not end there, it continues '— the same class as the seat-side no-merge rule, and the reason the DRAFT remedy is listed first.' Both are quoted in the tests field. Per the stop condition I copied the origin/main text, so what landed is byte-identical to the header rather than a truncation with an invented sentence boundary — which is the paraphrase the card was filed to prevent. The extra clause is on-point in this position: the DRAFT remedy it names is exactly what the never-list beside it prescribes. Draft PR #7630, governed (AGENTS.md), parked for a human merge: not flipped ready, not enqueued, no auto-merge, no reviewers requested, no approving review submitted. Assignees untouched.",
"tests": "GATES (worktree /home/user/objectui-issue-7618, base 11edab8; every exit code captured by redirect before any pipe; git status --porcelain empty at the end, so all verdicts below measured exactly 646ce08, the branch head):\n pnpm lint --concurrency=2 (whole repo, NODE_OPTIONS=--max-old-space-size=4096) => EXIT:0 — 'Tasks: 47 successful, 47 total'; app-shell prints '2883 problems (0 errors, 2883 warnings)', all pre-existing no-explicit-any in TS sources this markdown-only diff does not touch.\n pnpm check:control-bytes => EXIT:0 — '✅ check-control-bytes: OK (scanned 6239 tracked text file(s); skipped 85 binary).' Plus the self-scan outside the gate: grep -naP over AGENTS.md, clean.\n pnpm check:governed-queue-guard => EXIT:0 — 'OK check-governed-queue-guard self-test: 132 cases pass' (the self-test of the file whose header is copied).\n pnpm check:doc-fences => EXIT:0 — '✅ check:doc-fences — every TypeScript block in 227 document(s) is fenced ts/tsx/typescript...'\n pnpm check:shell-escape-residue => EXIT:0 — '✅ check-shell-escape-residue: OK (5/5 root(s) resolved -- AGENTS.md: 1 file(s), 15 fence(s); ...)'\n node scripts/check-doc-links.mjs => EXIT:0 — 'Links are valid across 17 scan roots.'\n node scripts/check-changeset-presence.mjs => EXIT:0 — 'Compared the working tree with 11edab8 (merge-base with origin/main): 1 file(s) changed, 0 of them published source ... 0 changeset(s) added.' / '✅ No source or published contract of a released package changed in this range, so no changeset is owed.' VERDICT FOLLOWED: no changeset added.\n node scripts/check-governed-queue-guard.mjs --test AGENTS.md => EXIT:3 (expected) — '⛔ GOVERNED — 1 of 1 path(s) are on a governed surface: AGENTS.md x1 — the repo-root agent instruction file'. This is why the PR stays a draft.\n pnpm exec vitest run FILES, where FILES is the 12 suites listed by git grep -l AGENTS.md -- scripts/__tests__/*.test.ts, => EXIT:0 — 'Test Files 12 passed (12)' / 'Tests 345 passed (345)'. Header line 'RUN v4.1.10 /home/user/objectui-issue-7618' confirms the vitest root is the repo root, so neither of the two false-green traps in AGENTS.md §「怎么跑测试」 applies (trap 1 = cwd inside a package; trap 2 = paths after --; I used neither).\n\nBYTE-FOR-BYTE PROOF OF THE COPY: md5 of the guard header span (scripts/check-governed-queue-guard.mjs L93-96 at 11edab8, ' * ' prefix stripped, the four wrapped lines joined, truncated at the sentence end) and md5 of the quoted line as it now stands in AGENTS.md ('> ' stripped) are both ba830bd6b431b3bfc97f4bde99d80d49. A reword on either side breaks the match.\n\nTHE TWO QUOTES (stop condition: they differ, so the origin/main text is what landed):\n AS QUOTED BY THE CARD/DISPATCH: '⛔ An agent seat never submits an approving review on a governed-surface pull request, under any account. Every seat in this repository writes under a shared GitHub identity, so GOVERNED_APPROVERS is a technical control that is only as good as that normative rule.'\n AS IT STANDS AT origin/main (11edab8, guard header L93-96) AND AS COPIED: '⛔ An agent seat never submits an approving review on a governed-surface pull request, under any account. Every seat in this repository writes under a shared GitHub identity, so GOVERNED_APPROVERS is a technical control that is only as good as that normative rule — the same class as the seat-side no-merge rule, and the reason the DRAFT remedy is listed first.' The card's version is a strict prefix of it up to that terminal period (verified programmatically); origin/main continues with the em-dash clause.\n\nBEFORE (AGENTS.md at 11edab8, L487-490):\n L487: (blank)\n L488: ⛔ 绝不 gh pr ready(不退出 draft)、⛔ 绝不加入合并队列、⛔ 绝不 gh pr merge --auto / enable_pr_auto_merge、⛔ 绝不自己合并。这类 PR 停在 draft,等人类合并。人类的那次合并动作本身就是审核记录 —— 不需要额外的逐 PR 批准点击,也别去等一个不存在的 approval。\n L489: (blank)\n L490: - 判据是 PR 的文件清单,不是 PR 的标题或描述。 命中与否只看路径。\n\nAFTER (L488 byte-identical; six lines inserted; the bullet list follows unchanged):\n L489: (blank)\n L490: ⛔ 第五条禁令 —— 绝不自己去留下那条 approval。 上面四条管的是落地,这一条管的是批准:scripts/check-governed-queue-guard.mjs 的文件头部把它写成规范条款,它 cleared 分支的判定文本也印着同一句。此处逐字照录、不译(两处措辞不得漂移):\n L491: (blank)\n L492: > ⛔ An agent seat never submits an approving review on a governed-surface pull request, under any account. Every seat in this repository writes under a shared GitHub identity, so GOVERNED_APPROVERS is a technical control that is only as good as that normative rule — the same class as the seat-side no-merge rule, and the reason the DRAFT remedy is listed first.\n L493: (blank)\n L494: sha pin 退休之后这条更重、不是更轻(维护者 2026-09-04 裁,#7606 执行、#7616 把新判据写进下面那段):一条获授权的 APPROVED review 现在清掉同一 PR 其后每一次 push,于是在一个 agent 操作的 approver 账号与一次它自己放行的受管落地之间,只剩这条规范禁令。\n\nPUNCTUATION CONVENTION CHECKED, NOT ASSUMED: the base file uses half-width , ( ) : with Chinese prose (567 half-width commas vs 2 full-width, and 0 full-width parens/colons; the 2 full-width commas sit inside an untranslated maintainer quote). My first insertion used full-width forms, so I reverted with git checkout HEAD -- AGENTS.md and reinserted; the landed diff greps clean for ，（）：\n\nPREMISE CHECKS ON origin/main (11edab8), both confirmed:\n (1) The rule really was absent. grep over AGENTS.md at origin/main for approving review / 批准 / approve / approval returns exactly two hits, neither of them this rule: L488 ('不需要额外的逐 PR 批准点击,也别去等一个不存在的 approval' — do not WAIT for one) and L496 (the #7616 paragraph, which carries the unpinned approval-record predicate and its accepted cost, but not the prohibition on LEAVING one). The card's reading holds on my base.\n (2) No objectui gate pins AGENTS.md's shape. grep over scripts/.mjs and .github/workflows/.yml for line/byte/length/max/ceiling/limit/width/count near AGENTS finds only prose comments using 'one line added to AGENTS.md' as a historical check-count example (check-changeset-presence.mjs:42, changeset-guard.yml:13, changeset-presence.yml:26, control-bytes.yml:19, docs-links.yml:20) plus check-governed-queue-guard.mjs:954, which is path classification. scripts/tests/check-governed-queue-guard.test.ts asserts only membership (governedPathsIn(['AGENTS.md']) => ['agents-md']), never content shape. This confirms the #7605 dev's search on a fresh base: no line ceiling, no byte ceiling, no row-width rule.\n\nPR BODY READ BACK IN FULL after creation: byte-identical to what was sent except that the platform APPENDED a second copy of the session-URL signature footer (the sent body already ended with one). Nothing was eaten — no sanitizer damage, no truncation. Left as-is rather than PATCHed, since an edit would downgrade the session-URL footer to the bare form for no gain.",
"mcp_calls": "0 — zero GitHub MCP calls for the whole run; issue body, triage comment 5537366949, PR creation and PR read-back all went through repo-scoped REST with $GITHUB_TOKEN (probed first: the issue read returned HTTP 200), and everything else through git on the local clone.",
"open_questions": [],
"out_of_scope_findings": []
}

---

Generated by Claude Code

[claude]

ACCEPT — governed terminal. Contract-tier review of PR #7630, head 646ce08 (skills seat, 2026-09-04T11:46Z).

Implemented-by: claude/issue-7618-agents-never-approve-rule (os-dev subagent, task a7711943584161e36)
Reviewed-by: session_019RfFHiRCSs3JXLK4cwcfox (skills seat)

Verified in a detached compare worktree at 646ce08 (merge base = origin/main 11edab8; git merge-tree clean), not from the report:

• One file, +6/−0, beside the never-list and nothing else. The existing four-prohibition line (L488) is byte-identical; the insertion is a Chinese lead-in naming both places the rule already lives, the English sentence as a blockquote, and one clause on why the retired sha pin made it weightier (one authorized APPROVED review now clears every later push on the same PR). Half-width punctuation matches the file's convention (0 full-width marks in the added lines).
• Copied, not paraphrased — and the stop condition was taken the right way. The sentence as quoted by the card ended at "…that normative rule."; at origin/main the guard header continues "— the same class as the seat-side no-merge rule, and the reason the DRAFT remedy is listed first." The dev copied the origin/main text whole (the seat checked the quote against the header's own lines: the header text is a prefix of the quoted line, character for character), so the two places cannot drift and the extra clause names exactly the draft remedy the never-list beside it prescribes.
• Gates: node scripts/check-governed-queue-guard.mjs --test AGENTS.md → exit 3, GOVERNED (the reason the PR stays draft); check-control-bytes OK (6239 files); the dev's run adds pnpm lint 0 errors, check:governed-queue-guard 132 cases, doc-fences, shell-escape-residue, doc-links, check-changeset-presence "no changeset is owed" (followed), and the 12 vitest suites that read AGENTS.md (345 tests) — and confirmed on a fresh base that no objectui gate pins AGENTS.md's shape. CI on 646ce08: 25 success, 3 skipped.

Governed terminal, four pieces: ① this verdict; ② the PR stays a DRAFT — ⛔ the seat never flips it ready, never enqueues, never arms auto-merge; ③ review requested from os-zhuang and hotlong on the PR; ④ seat post refreshed. Landing is a human merge, or an authorized approval and then the queue (this repo's guard carries the unpinned predicate).

---

Generated by Claude Code