docs(agents,pm-dispatch): a draft conversion drops only an unqueued PR's auto-merge; a queued governed pull-back goes to the maintainer

[objectstack-fleet]

Fixes #20847

Clause-②: no

What changes

Four sentences stop saying that a draft conversion, or draft plus an auto-merge disable, removes a PR from the merge queue. Each keeps the half that holds: on an unqueued PR, the conversion drops auto-merge. Prime Directive 14 now names the one removal act that is measured, the maintainer's manual removal, and says why a seat cannot do it.

site	before	after
AGENTS.md Prime Directive 14, "Already armed or queued" (was :285-:287, now :285-:288)	Convert it back to draft AND disable auto-merge — draft is what removes queue membership, disabling alone drops only the arming — then confirm from the remote that it is in neither the queue nor origin/main.	Disable auto-merge AND convert it back to draft, which disarms an unqueued PR; no seat act, those two included, is measured to dequeue one (platform-readings.md, queue membership), so ask the maintainer at once to remove a queued PR by hand — then confirm from the remote that it is in neither the queue nor origin/main.
AGENTS.md "State on your PR" (:428-:429 → :429-:430)	the draft flag, which flipped back destroys auto-merge and queue membership at once	the draft flag, which flipped back destroys an unqueued PR's auto-merge at once
AGENTS.md multi-agent §7 (:532-:533 → :533-:534)	flipping back to draft drops auto-merge and queue membership at once, and neither returns by itself	flipping back to draft drops an unqueued PR's auto-merge at once, and it does not return by itself
.claude/skills/pm-dispatch/SKILL.md:202	已入队才读到本条 ⇒ 转 draft 与 disable 都做;出队以阳性探针答,ref 缺席只旁证。	已入队 ⇒ 转 draft 并 disable,即请维护者手动出队(席位无实测手段);出队以阳性探针答。

Every clause that still holds is kept: "Draft is no barrier by itself — the barrier is this directive", "confirm from the remote that it is in neither the queue nor origin/main", "ready first, arm second", and the positive-probe reading of a dequeue. The rule text cites platform-readings.md and restates no reading. The readings themselves are in PR #20845 (issue #20764).

The grep. AGENTS.md and every file under .claude/** were searched for other copies of the claim: queue membership, draft next to queue or auto-merge, 转 draft, 出队, 踢队, dequeue, evict. The four sites above are all of them. references/platform-readings.md:41-:43 is PR #20845's, and this PR leaves it alone. lanes/cli.md:17 only points at that file.

Is there a seat act that removes a queued PR? (the claim's measurement)

Every step was read-only. ⛔ No mutation was called and no PR was dequeued.

1. Live GraphQL introspection was refused. The request was POST https://api.github.com/graphql with query { __type(name: "Mutation") { fields { name description } } }. It returned 403 from this session's egress, with the message "GitHub GraphQL is not available from Claude Code sessions". The same message lists the pull-request routes the egress does serve: ccr/auto_merge (PUT/DELETE), ccr/ready_for_review and ccr/convert_to_draft. None of them dequeues.
2. GitHub's published GraphQL schema does declare the mutation. This is a document reading, not a live introspection. The source is octokit/graphql-schema schema.graphql on main, fetched 2026-09-30T12:18Z (1,223,842 bytes, sha256 prefix 3c62d0526d133cee). It declares dequeuePullRequest(input: DequeuePullRequestInput!) with the description "Remove a pull request from the merge queue." Its input is id: ID!, the pull request id.
3. REST has no dequeue endpoint. In github/rest-api-description, api.github.com.json (info version 1.1.4, main), no path contains queue. The word "dequeue" appears only in the pull_request.dequeued webhook schemas.
4. GitHub's own docs list the ways a PR leaves the queue. The source is github/docs on main (data/reusables/pull_requests/merge-queue-removal-reasons.md). The list has four entries: CI failure, timeout, "User requesting a removal via the API or merge queue interface", and a branch-protection failure. Neither a draft conversion nor a close is on it. The how-to page (merging-a-pull-request-with-a-merge-queue.md) removes a PR through the web page's Remove from queue button and says the CLI cannot do it.
5. The relay cannot carry it. scripts/pm/fleet-write/ops.mjs at 05a7547c9 has 15 ops: comment, comment_edit, labels_add, labels_remove, assign, unassign, issue_patch, issue_create, pr_create, pr_request_reviewers, pr_ready, pr_draft, automerge_enable, automerge_disable and transfer. ALLOWED_MUTATIONS has five entries, and dequeuePullRequest is not one of them.

⇒ No dequeue act is measured to be available to a seat. The platform documents one (the GraphQL mutation), but neither this session's egress nor the relay can reach it. So the rule sends a queued pull-back to the maintainer and says why. The triage direction on the card reserves the next step for the triage seat: whether to file a tooling child for a relay dequeue op. That would be a new relay capability.

Coordination with PR #20845

The new rule cites the queue-membership section of platform-readings.md. At origin/main (05a7547c9), line :42 of that section still carries the old reading, and PR #20845 rewrites :41-:43. Until #20845 lands, the rule and line :42 disagree. #20845 should land first, or with this PR. The citation names a file section, not a PR, so it is correct once both have landed, whatever the order.

Line budget

• AGENTS.md: 1107 → 1108 lines, against a ceiling of 1116 (headroom 9 → 8). The +1 is all in Prime Directive 14. The new sentence adds the maintainer route and the reason for it, and it needs four lines where the old one took three. The other two AGENTS.md sentences are net 0.
• pm-dispatch/SKILL.md: 319 → 319 lines (ceiling 319, headroom 0). Line :202 is edited in place and goes from 107 to 116 bytes.
• Every added line is at most 120 bytes: 118 / 110 / 120 / 118 (Prime Directive 14), 83 / 87 ("State on your PR"), 88 / 88 (§7) and 116 (SKILL.md). check:pm-skill-ratchet reports no over-length line and leaves the table-row pins unchanged (768 and 342).

Deviations from the dispatch

1. SKILL.md:202 drops two parts. They are the clause 「ref 缺席只旁证」 and the 「才读到本条」 of the trigger. The 120-byte cap left 13 bytes of room on the line, and the maintainer route with its reason needs about 50. The dropped clause is kept in platform-readings.md:44 (「ref 缺席只作旁证,⛔ 永不承载结论」), and the positive-probe half of the line stays.
2. AGENTS.md is net +1, not ≤ 0. None of the three sentences frees a whole line by deleting content. Freeing one would need a re-wrap of text this PR does not otherwise change. The pin is not exceeded.
3. The introspection is a schema-document reading. The live endpoint refused this session (item 1 above).

Acceptance notes

• Same family, outside this PR's scope (the dispatch rules out any script edit). scripts/pm/check-governed-queue-guard.mjs makes the same claim in both of its refusal texts, at :1451-:1452 and :1911-:1912 (at 05a7547c9): "take the pull request out of the queue: convert it back to DRAFT (disarming auto-merge alone does NOT dequeue it)". Parking the PR in draft after the guard's red is still useful: an armed PR re-enters the queue, and a draft cannot be armed. The part in parentheses is the falsified premise. This is reported to the seat so it can join the card's family. No card is filed.
• An unmeasured candidate, for triage. Closing the PR is an act a seat has, through the relay's issue_patch with state: closed. A code comment in .github/workflows/cross-repo-issue-closer.yml:399-:401 says that closing drops queue membership. GitHub's list of removal reasons does not name it, and I know of no reading of it in this repo. It was not attempted.
• Changeset. None. .claude/** and AGENTS.md are not in any package's files[], so this PR publishes nothing. skip-changeset is for the seat to apply.
• Landing. This is a Tier H governed surface (AGENTS.md). The PR stays a draft until an authorized approval.

Verification (head b095ba02e)

• node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands, with no paths, derived 22 commands for this change set (2 paths against merge base 05a7547c9). All 22 exited 0:
  • check-closing-keyword-parity, plus its --self-test;
  • check-comment-mask-corpus;
  • check-governed-queue-guard --self-test and check-harness-current --self-test;
  • @objectstack/lint check:doc-formula-expressions, run after building @objectstack/lint and its dependency closure under os-verify-lock.sh (VERDICT command-exit 0);
  • the check: family: agent-test-spelling, cross-package-test-inputs, doc-authoring, docs-audit-scope, driver-memory-census, gitlink-declared, nul-bytes, pm-expected-skips, pm-governed-merges, pm-governed-prose, pm-skill-id-lint, pm-skill-ratchet, refd-timer-probe, required-contexts, skill-frame-sync and watch-hint-literal.
• pnpm check:pm-settings-deny-roster was also run, because its roster lives under .claude. It exited 0.
• The --ran reconciliation read "22 derived, 22 run, 0 NOT-MEASURED, 0 UNRUN". All 22 commands recorded an exit code.
• Readings from the gates:
  • check:pm-skill-ratchet: "AGENTS.md is 1108 lines (ceiling 1116; headroom 8)" and "pm-dispatch/SKILL.md is 319 lines (ceiling 319; headroom 0)".
  • check:pm-skill-id-lint: "34 file(s) clean".
  • check:pm-governed-prose: "2 instruction surface(s) name all 6 registered governed surfaces".
• NOT MEASURED:
  • check-required-contexts --verify-required-set, which reads the workflow event and so runs only in CI;
  • the four CI type-check lanes. This diff touches no TypeScript.

维护者速读(草稿)

• 改了什么:AGENTS.md 三句、pm-dispatch/SKILL.md 一行。旧文说「转 draft 会把 PR 踢出合并队列」,现在都不再这么说。改后:转 draft 只在未入队时掉 auto-merge。已入队的受管 PR 要撤回时,两手照做,并立即请你手动出队。
• 为什么改:实测里 4 个已入队后转 draft 的 PR 都没有因此出队,其中 2 个带着 draft 被队列直接合入,读数在 PR docs(pm-dispatch): the draft-conversion reading splits by queue state — no queued PR left the queue on a draft conversion #20845。「转 draft 能出队」的说法写在第 14 条指令里,那是防止坏合并的撤回手段,而这个手段实际不起作用。出队手段问题:目前席位没有经实测可用的出队手段。GitHub 公开 schema 里有 GraphQL dequeuePullRequest,但本会话出口拒绝 GraphQL(403),中继也没有这个 op。所以规则暂时把已入队的撤回交给你手动出队。要不要给中继加出队 op,属于新能力,由分诊决定是否立子卡。子卡立了,你会看到裁决请求。
• 风险与代价(含回滚):只改规则文字,不改代码。代价是已入队的撤回要等你点「Remove from queue」。队列落地约需 15–30 分钟,所以需要你及时响应。PR docs(pm-dispatch): the draft-conversion reading splits by queue state — no queued PR left the queue on a draft conversion #20845 应先合或同合,否则被引用的读数第 42 行暂时和本规则说法相反。回滚就是 revert 这一个 PR。
• 席位意见:
• 你要做的:审阅后批准本 PR(或直接合并)。

---

Generated by Claude Code

[objectstack-fleet]

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: b095ba02eee411e2db30b8336aa3cb8a7d7c9575
Local-runs: none

Reviewed read-only: card #20847 (body + comments 5910732998 / 5911100852 / 5911443452), PR #20854 body, file list and net diff, PR #20845 diff, the #20695 / #6732 / #20797 timelines, the head's check-runs, and the repository at 05a7547c9 (merge base), b095ba02e (head), origin/main (72f8c3820) and 02ea176a1 (#20845's tip) via git show / git grep / git diff only. The record template was printed with record-recognisers.mjs --template as the brief directs; no gate, build or test was run. Not read: the dispatch order or the dispatching seat's conclusions.

Net diff at the head. origin/main moved from the PR's snapshot base to 72f8c3820 after the branch forked at 05a7547c9; none of AGENTS.md, pm-dispatch/SKILL.md, platform-readings.md, fleet-write/ops.mjs or check-governed-queue-guard.mjs changed on main in that span, so the three-dot diff IS the net diff: 2 files, +9 / -8 (AGENTS.md +8 / -7 across three hunks; SKILL.md +1 / -1 at :202).

Check-runs on the head (filter=latest, 30 runs): success 18, skipped 12, no other conclusion; every run is completed. Every non-success run is a skipped: Auto Label (dup name), Build Core, Build Docs, Check Changeset (dup name), Check PR Size (dup name), Console Pin Gate, Dogfood Regression Gate (matrix), Dogfood Verify CLI, Packed-tarball smoke ×2, Temporal Conformance, Test Core (matrix). Of note among the greens: Governed Surface Queue Guard, Lint & Repo Gates, Check Changeset, the four Type Check · lanes, TypeScript Type Check, Test Core, Dogfood Regression Gate, and the three card/branch parity checks. PR state read: draft, auto-merge off, mergeable_state: clean, labels documentation · size/s · skip-changeset. Untouched.

① Derived judgments

Prime Directive #14, AGENTS.md:285-:288 at head — right. Clause by clause against the timelines and the code:

• "Disable auto-merge AND convert it back to draft, which disarms an unqueued PR" — right. The card's three unqueued cases (feat(spec)!: 移除 ./automation 的孤儿 EventSchema —— C6 双源清账,基线 13 → 12 (#4658) #4745, fix(cli): os login --json 声明为 NDJSON 事件流，每行一份可解析文档 (#6531) #6727, docs(rest): openapi-endpoints 的 requestBody 注释改写为实测状态——服务出的文档里 $ref 为 0 (#6797) #6829) log auto_merge_disabled within 1 s of convert_to_draft; feat(spec)!: the ADR-0087 migration chain leaves the root entry for @objectstack/spec/migrations (#20646) #20695's own timeline shows the same pair on an unqueued PR at 18:31:59Z / 18:32:02Z. The disable is a measured relay act (ops.mjs:213 disablePullRequestAutoMerge, op automerge_disable). The order swap from the base ("draft AND disable") is immaterial: both acts are required, and on an unqueued PR either one alone disarms.
• "no seat act, those two included, is measured to dequeue one" — right, and honestly worded (see the measurement judgment below). docs(adr),chore(devx): reconstruct ADR-0079 + make every cited ADR number resolve #6732: added_to_merge_queue 2026-08-08T14:01:50Z, convert_to_draft 14:04:32Z, merged 14:38:56Z with removed_from_merge_queue by github-merge-queue[bot] in the same second — merged in draft. feat(spec)!: the ADR-0087 migration chain leaves the root entry for @objectstack/spec/migrations (#20646) #20695: queued 2026-09-29T23:32:51Z, draft 23:57:02Z, merged 2026-09-30T00:04:21Z — merged in draft. No removed_from_merge_queue follows either conversion.
• "(platform-readings.md, queue membership)" — right as a citation form: the section ## 队列成员资格与 auto-merge exists at that file's :5 and holds :41-:44; AGENTS.md:443-:444 already cites the same file the same way; no issue number, date or narrative enters the rule text (header :11-:15). Consistency of the cited section is a sequencing matter, judged below.
• "so ask the maintainer at once to remove a queued PR by hand" — right, and the one route named is the one measured: docs(governed): the retirement route is retiredKey() on any shape, with the retiredAfter stamp and the step-18 registry shape #20797 added_to_merge_queue 2026-09-30T07:32:47Z by os-zhuang, removed_from_merge_queue 07:33:36Z by os-zhuang (a human account, in GOVERNED_APPROVERS), no convert_to_draft anywhere in that timeline, no red build (re-queued by the fleet bot 07:56:08Z, merged 08:20:18Z). This is the only measured removal without a red build.
• "then confirm from the remote that it is in neither the queue nor origin/main" and "Draft is no barrier by itself — the barrier is this directive" — kept verbatim from the base. The barrier still holds; the rule still forbids the seat from treating draft as the safeguard.
• Nothing true is lost: the base's "disabling alone drops only the arming" survives as "disarms an unqueued PR"; the base's "draft is what removes queue membership" is the falsified half and is the only thing removed.

"State on your PR", AGENTS.md:429-:430 — right. "flipped back destroys an unqueued PR's auto-merge at once … invisibly" is the measured unqueued reading; the paragraph's rule (do not "correct" a flag another actor set) is unchanged; the pointer "(§7's draft-flip re-arm note)" still resolves to the rewritten §7 sentence at :533. Nothing true lost.

Multi-agent §7, AGENTS.md:533-:534 — right. "flipping back to draft drops an unqueued PR's auto-merge at once, and it does not return by itself — ready first, arm second": #4745 and #6727 were re-armed only by a fresh act after ready (01:30:52Z, 13:46:02Z). "neither returns" → "it does not return" is the correct follow-through once queue membership leaves the sentence. Nothing true lost.

.claude/skills/pm-dispatch/SKILL.md:202 — right, with one named residual.
Base: 「已入队才读到本条 ⇒ 转 draft 与 disable 都做;出队以阳性探针答,ref 缺席只旁证。」 (107 bytes)
Head: 「已入队 ⇒ 转 draft 并 disable,即请维护者手动出队(席位无实测手段);出队以阳性探针答。」 (116 bytes)

• The dropped 「ref 缺席只旁证」 does live at platform-readings.md:44, verified at both origin/main and the docs(pm-dispatch): the draft-conversion reading splits by queue state — no queued PR left the queue on a draft conversion #20845 tip, in its stronger form: 「出队按队列 ref 阳性探针答不在队向,加未落地;ref 缺席只作旁证,⛔ 永不承载结论。」 — :44 is a context line in docs(pm-dispatch): the draft-conversion reading splits by queue state — no queued PR left the queue on a draft conversion #20845's diff, untouched. lanes/cli.md:17 routes queue readers to that file. The line's positive-probe half 「出队以阳性探针答」 is kept. No rule is lost.
• Dropping 「才读到本条」 widens the literal trigger from "queued by the time you first read this clause" to "queued". In its block (## 路径分叉, :180) the bullet sits after :182 (take the path surface BEFORE ready / arm / queue) and :186 (the Tier H fork) and its ①-④ items, and the legitimate queue entries carry their own preconditions (:196 批准后, :204 PASS 后), so in context "已入队" still names the same case: a governed PR found queued ahead of its landing key. AGENTS.md PD14 keeps the exact trigger "Already armed or queued when you read this?", and AGENTS.md wins on conflict (:6). Residual: read in isolation the new line is literally satisfied by a Tier S PR the seat itself just queued after its PASS record, where the cost of a misread is one unnecessary manual-dequeue request to the maintainer. Restoring 「才读到本条」 costs 15 bytes (116 → 131, over the 120 cap); the only way to fit it is to drop the reason 「(席位无实测手段)」, which the triage direction requires ("says why"). The trade is forced by the pins and the direction; judged acceptable, flagged in ③ for the maintainer's eye.

The dequeue-act measurement — the relay half verified at the code; the sentence is an honest statement. scripts/pm/fleet-write/ops.mjs at 05a7547c9 (unchanged on main since): OPS (:253) has exactly the 15 keys the dev lists — comment, comment_edit, labels_add, labels_remove, assign, unassign, issue_patch, issue_create, pr_create, pr_request_reviewers, pr_ready, pr_draft, automerge_enable, automerge_disable, transfer; PR_MUTATIONS (:209-:213) carries only ready / draft / auto-merge enable / disable; ALLOWED_MUTATIONS (:393) is exactly markPullRequestReadyForReview, convertPullRequestToDraft, enablePullRequestAutoMerge, disablePullRequestAutoMerge, transferIssue; the strings dequeue, queue, enqueue do not occur in the file. The three external legs (published GraphQL schema declares dequeuePullRequest; live introspection 403 from the session egress; REST OpenAPI has no queue path; github/docs removal reasons) are the dev's readings of sources outside this checkout and were not re-fetched here. The rule text does not depend on them: "no seat act … is measured to dequeue one" rests on the in-repo evidence alone (four queued drafts never dequeued by the conversion; the relay carries no dequeue). The wording claims "not measured", never "does not exist", which is exactly what the evidence supports, and the PR body discloses that the platform's schema does declare one. Honest.

The cross-file citation and sequencing. At origin/main (72f8c3820, file identical to 05a7547c9), platform-readings.md:42 still reads 「本仓转 draft 同时掉 auto-merge 与队列成员资格,不自动恢复,转正后重挂;姊妹仓曾保位照合。」, while :41 already says 「转 draft 不是可靠的踢队手段」 and :43 already prescribes both acts. PR #20845 (head 02ea176a1, base 30839063b; the file is byte-identical between that base and current main, so it applies cleanly) rewrites :41-:43 to the split-by-queue-state reading and leaves :44 alone. So: this PR's sentence is true on its own; its citation is consistent with the cited section only once #20845 has landed — until then the rule and :42 disagree, with AGENTS.md winning in the window. The order that keeps the two texts consistent at every moment is #20845 first (or in the same batch), then #20854. #20845 is Tier S and still a draft (needs its own PASS record, then the queue); #20854 is Tier H and waits on the maintainer's word anyway, so the owning seat can hold this order.

Pins — right, measured at the head. AGENTS.md 1107 → 1108 lines against ceiling 1116 (check-skill-line-ratchet.mjs:1531), headroom 9 → 8; pm-dispatch/SKILL.md 319 → 319 against ceiling 319 (:374), :202 edited in place; every added line measured 116 / 118 / 110 / 120 / 118 / 83 / 87 / 88 / 88 bytes, max 120 against the 120-byte cap. Lint & Repo Gates is green on the head.

Survivor grep at the head (AGENTS.md + .claude/**: queue membership, dequeue, evict, 转 draft, 出队, 踢队, 踢出, and draft beside queue/auto-merge): no copy of the falsified claim remains except platform-readings.md:41-:43, which is #20845's. The four sites the PR body names are all of them.

② Semver level

skip-changeset is right. The diff touches AGENTS.md and .claude/skills/pm-dispatch/SKILL.md only; neither is in any released package's files[], so nothing publishes (AGENTS.md:1072-:1076: the label is for a diff that publishes nothing from any released package). Clause-②: no is on the PR body's second line and on the claim (5911100852). Check Changeset is success on the head.

③ Boundary flags

1. Dev deviation — SKILL.md:202 drops 「ref 缺席只旁证」 and 「才读到本条」. Answered above: the first has its home at platform-readings.md:44 (verified at main and at docs(pm-dispatch): the draft-conversion reading splits by queue state — no queued PR left the queue on a draft conversion #20845's tip); the second widens the literal trigger but not the practical one, bounded by the block's :182 / :186 / :196 / :204 and by AGENTS.md precedence. Acceptable; the maintainer may weigh whether 「已入队才读到本条」 should return in a later shrink of the line.
2. Dev deviation — AGENTS.md net +1, not at most 0. The claim stated an aim, not a pin; 1108 / 1116 holds. Acceptable.
3. Dev deviation — the introspection is a published-schema reading, live endpoint 403. Answered: the rule's wording is "not measured", which the in-repo evidence alone establishes; the external readings are disclosed as such in the PR body. Acceptable.
4. Dev deviation — wording departs from the PM's suggested route. The suggestion is not in this review's inputs; judged against the card's triage direction (5910732998) instead, every requirement is met: the falsified claim removed at all sites, the unqueued half kept, only the measured removal act named, a queued pull-back routed to the maintainer at once with the reason, readings cited not restated, skill copies grep-found and corrected in the same PR. Acceptable.
5. The dequeue measurement and the tooling-child decision — escalated to the triage seat, as the direction reserves it. The direction's condition for a tooling child ("the claim measures that the platform exposes one") is met by the dev's schema-document reading, not a live one. One observation for that decision, from the code: the relay already executes GraphQL mutations server-side (PR_MUTATIONS, run by the fleet-write dispatch workflow), so the session egress's GraphQL refusal is not by itself the constraint on a relay dequeuePullRequest op; that op would be a new capability and a maintainer ruling, exactly as the card says. Not this PR's scope; no card filed by this review.
6. The queue-guard refusal text — out of this PR's scope; filed as [finding] the Governed Surface Queue Guard's refusal text tells a seat to dequeue a governed PR by converting it to DRAFT — measured not to dequeue in this repo (same family as #20847) #20855. Verified at 05a7547c9 (unchanged on main since): check-governed-queue-guard.mjs:1451-:1452 and :1911-:1912 both print "convert it back to DRAFT (disarming auto-merge alone does NOT dequeue it)". The claim excluded scripts and scripts/pm/** is not a governed surface; the guard's --self-test is green on this head, so nothing in this PR breaks it. [finding] the Governed Surface Queue Guard's refusal text tells a seat to dequeue a governed PR by converting it to DRAFT — measured not to dequeue in this repo (same family as #20847) #20855 (open, unlabelled, bare for triage) carries it and should land with or after docs(agents,pm-dispatch): a draft conversion drops only an unqueued PR's auto-merge; a queued governed pull-back goes to the maintainer #20854 so the printed remedy and the rule agree. Escalated via [finding] the Governed Surface Queue Guard's refusal text tells a seat to dequeue a governed PR by converting it to DRAFT — measured not to dequeue in this repo (same family as #20847) #20855, already filed.
7. The close-PR candidate act — escalated to the triage seat, correctly not named in the rule. Verified: ops.mjs:292-:302 issue_patch accepts state, so a seat can close a PR through the relay; .github/workflows/cross-repo-issue-closer.yml:399-:401 asserts closing "drops its merge-queue membership and any armed auto-merge in the same step". No timeline reading of that on a queued PR exists in the inputs, and the dev's reading of GitHub's removal-reason list does not name it. The dev rightly did not attempt it (a mutation). The rule names only measured acts, so its silence on closing is correct. Caution for any future measurement: closing a governed PR is not a reversible pull-back (review state, head-ref handling), so it needs a maintainer ruling before any seat measures it.
8. Sequencing with docs(pm-dispatch): the draft-conversion reading splits by queue state — no queued PR left the queue on a draft conversion #20845 — escalated to the owning seat as a landing-order constraint: land docs(pm-dispatch): the draft-conversion reading splits by queue state — no queued PR left the queue on a draft conversion #20845 (Tier S, draft, needs its own PASS record) first or in the same batch as docs(agents,pm-dispatch): a draft conversion drops only an unqueued PR's auto-merge; a queued governed pull-back goes to the maintainer #20854 (Tier H, maintainer's word). Until docs(pm-dispatch): the draft-conversion reading splits by queue state — no queued PR left the queue on a draft conversion #20845 lands, platform-readings.md:42 contradicts the rule that cites its section, with AGENTS.md winning in the window.

This PR is Tier H (AGENTS.md, :272-:275): this record is not its landing key. The maintainer's authorized APPROVED review by an account in GOVERNED_APPROVERS remains the key, and the PR stays a draft until then.

Implemented-by: claude/issue-20847-agents-draft-dequeue
Reviewed-by: session_01KTZmMfzVzjNvyaLyQ8mHvg

VERDICT: PASS