[objectstack-fleet]

Requested by Tesla · project thread

Fixes #27848

Clause-②: no
Metadata-protocol: none

维护者速读

• 改了什么:rest-channel.md 反向两 op 那句「未实调」换成三行实测读数(反向 F 实调通过、行即回读、出队未实测),并把只剩一个指针的「第三桶」小节折进队列读法的指针行;80 → 79 行,每行 ≤120 字节。
• 为什么改:复核 FAIL 时 automerge_disable/pr_draft 已是席位的强制动作,席位读到「未实调」无法判断它要发的 op 是否可靠;中继日志里两 op 已于 2026-10-25 在一张已挂臂未入队的 PR 上走通。
• 风险与回滚:只改规则文本,无门禁、无代码、无发布面;回滚 = revert 本提交。
• 先看:.claude/skills/pm-dispatch/references/rest-channel.md 第 57–60 行(改句 + 三行新读数),再看第 72 行(折并的指针)。
• 两分钟验证:pnpm check:pm-skill-ratchet 绿;打开 fleet-write run 40164374680 的日志,对照 ✓ action 1 automerge_disable … auto-merge off · not in the merge queue 与 ✓ action 2 pr_draft … draft 两行。
• 需要你做的:Tier H 受管面(.claude/**),席位契约复审 PASS 后请 os-zhuang 批准,席位落地。

What changes

One file: .claude/skills/pm-dispatch/references/rest-channel.md (80 → 79 lines, ceiling 80; every touched line ≤ 120 bytes).

• The reverse-ops clause on the auto-merge line (「反向两 op 未实调」) now ends 「反向实调如下」 and is followed by three rule lines:
  1. ✓ 反向 F = automerge_disable 再 pr_draft — measured on an armed, unqueued pull; both ops answered HTTP 200; the approval on the pull was not dismissed.
  2. The row is the read-back: autoMergeRequest null and isDraft true are the landed states; an answer still armed or still ready is a FAILED action (exit 5) that stops the stroke.
  3. Dequeue not measured: the one pr_draft sent to a queued pull answered no pullRequest (row FAILED, exit 5) while the conversion landed and the pull kept its queue place; queue membership is read from the timeline, not from the row.
• Paid for by deleting the three-line 第三桶 —— git 零配额等价物 section (heading, blank line, one pointer line) and carrying its pointer into the existing queue-reading pointer line (「①② 与 git 零配额四条的拼写与边界住 platform-readings.md」). The four readings it pointed at live in platform-readings.md (quota-starvation section ①–④), so no rule is lost.
• platform-readings.md is NOT touched: the order allows that edit only on a measured dequeue, and none was observed.

The rules-only convention of this file (no provenance dates, no incident narrative) keeps the run ids out of the rule text; they are recorded here.

Why

PR #27845 made automerge_disable and pr_draft the mandatory disarm on a contract-review FAIL. A seat reading 「未实调」 could not tell whether the ops it must send are known to work.

Tests

No code changed; the gates below are the test surface.

Measured strokes (evidence for the three lines)

The clean stroke, both ops — 2026-10-25. fleet-write run 40164374680 (job 121082282055, request fw-20261025T050717Z-17ca1b, head f9a1e93) sent [{"op":"automerge_disable","pull":25745},{"op":"pr_draft","pull":25745}]. Log rows:

✓ action 1 automerge_disable: POST /graphql disablePullRequestAutoMerge -> HTTP 200 · #25745 · auto-merge off · not in the merge queue · its base has a merge queue
✓ action 2 pr_draft: POST /graphql convertPullRequestToDraft -> HTTP 200 · #25745 · draft
fleet-write/execute: ✓ 2 request(s) landed, every action done.

Those rows are the mutation answers judged by PR_LANDED_STATE (ops.mjs): autoMergeRequest: null, isInMergeQueue: false, isMergeQueueEnabled: true, then isDraft: true. PR #25745's timeline (GET /issues/25745/timeline, 105 events) confirms the stored state: auto_merge_enabled by os-zhuang at 05:06:41Z, then auto_merge_disabled (05:07:45Z) and convert_to_draft (05:07:48Z), both by objectstack-fleet[bot]; the approving review 5641652106 carries no review_dismissed event; the PR has no added_to_merge_queue event at any point (it was armed on a red battery, never queued), so a dequeue is not measured here.

A pr_draft on a queued pull — 2026-10-24. Run 40042854158 (job 120717690554, request fw-20261024T000127Z-f103d0): [{"op":"pr_draft","pull":25550}] →

✗ action 1 pr_draft: POST /graphql convertPullRequestToDraft -> GraphQL: the answer carries no pullRequest
fleet-write/execute: ✗ stopped at action 1 (pr_draft); 0 later action(s) NOT attempted.
Process completed with exit code 5.

PR #25550's timeline: added_to_merge_queue 2026-10-23T23:55:28Z; convert_to_draft by objectstack-fleet[bot] 2026-10-24T00:01:34Z (so the conversion landed although the row read FAILED); ready_for_review 00:09:55Z; the only removed_from_merge_queue is 00:27:36Z by github-merge-queue[bot], followed one second later by merged. No dequeue: the queued pull kept its place through draft and re-ready and merged from the queue. This is the third observation of the reading already on platform-readings.md line 42 (2026-08-08, 09-29); not added there, see Acceptance notes.

A pr_draft after a queue ejection — 2026-10-27. Run 40434252633 (job 121959027306, request fw-20261027T014328Z-71eed1): pr_draft on #26343 answered HTTP 200 · #26343 · draft after the queue had already ejected the pull (removed_from_merge_queue 01:43:37Z by github-merge-queue[bot], convert_to_draft 01:44:07Z). Not a dequeue either; cited as the newer pr_draft reading.

Search method. PR timelines of the three pulls named in the seat's records, and GET /actions/workflows/fleet-write.yml/runs?created=… on the minute windows around each timeline event (12, 3 and 4 runs respectively), each candidate's job log read in full. Not an exhaustive scan of every fleet-write run since 2026-10-24 (hundreds per hour), so a later stroke carrying automerge_disable may exist unrecorded; the 2026-10-25 run is the only one found that sends it.

Acceptance notes

• platform-readings.md line 42 (「已入队者 2026-08-08、09-29 照合」) gained a third observation on 2026-10-24 (PR feat(lint)!: membership key declarations, and the row policies that read them, are judged when authored #25550). Not added: the dispatch order limits platform-readings.md edits to a measured dequeue. Carrier: skills seat 1.
• The seat's lead dated the revert 2026-10-26; the stroke is 2026-10-25 05:07Z (timeline + run log). Correction only, no card.
• node scripts/pm/dispatch-gates.mjs --tier .claude/skills/pm-dispatch/references/rest-channel.md prints "no path-derived mandate: the surface hits none of the 4 declared glob(s)", while the FRAME and the Claim call the tier MANDATORY for this surface. The run was on claude-fable-5-1 either way; noted for the seat, not filed (gate-vocabulary observation, no wrong answer to a public door).
• check-skill-line-ratchet.mjs row for this file stays at 80 with the file at 79 lines. The convention lowers the row to the landed count, but scripts/pm/ is outside this card's declared file surface; the seat may fold the lowering into its next touch of the ratchet table.
• .claude/skills/pm-dispatch/ has no evals/ directory, so there is no skill evals check to run for this surface.

Verification

Run at head f194f8f in a fresh linked worktree of this repository after pnpm install --frozen-lockfile and pnpm --filter '@objectstack/lint...' build (under os-verify-lock.sh, VERDICT command-exit 0, 75 s).

• node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack → 17 families for the one changed path. Every line run with stdin from /dev/null, exit code captured before any pipe:

17 derived gates, all exit 0

node scripts/check-closing-keyword-parity.mjs :: exit 0
node scripts/check-closing-keyword-parity.mjs --self-test :: exit 0
node scripts/check-comment-mask-corpus.mjs :: exit 0
node scripts/pm/check-harness-current.mjs --self-test :: exit 0
pnpm --filter @objectstack/lint run check:doc-formula-expressions :: exit 0
pnpm check:agent-test-spelling :: exit 0
pnpm check:cross-package-test-inputs :: exit 0
pnpm check:doc-authoring :: exit 0
pnpm check:driver-memory-census :: exit 0
pnpm check:gitlink-declared :: exit 0
pnpm check:nul-bytes :: exit 0
pnpm check:pm-governed-merges :: exit 0
pnpm check:pm-skill-id-lint :: exit 0
pnpm check:pm-skill-ratchet :: exit 0
pnpm check:refd-timer-probe :: exit 0
pnpm check:skill-frame-sync :: exit 0
pnpm check:watch-hint-literal :: exit 0

• node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --ran gates-ran.list → ✓ 17 derived famil(ies) accounted for — 17 run, 0 NOT-MEASURED; changed lines 11 (+5 / −6), under the 3000 threshold.
• Line widths: awk 'length($0) > 120' over the file prints nothing; touched lines are 114 / 113 / 109 / 113 / 120 bytes.
• Changeset: .claude/** publishes nothing (fast-track class) → skip-changeset, declared on the pr_create stroke.
• Relay writes: 1 (pr_create, this PR) + 1 (os-dev-report comment on pm-dispatch(rest-channel): record the first measured automerge_disable / pr_draft stroke in place of 「反向两 op 未实调」, now that the disarm rule mandates both ops #27848). MCP GitHub write calls: 0.

Human Minimum

• 1. Schemas start as Zod; packages/spec holds no runtime logic. (n/a — docs only)
• 2. Every name I added uses the casing for its kind. (n/a)
• 3. No formatter ran; pnpm lint is CI's run (repo-wide scan owned by CI).
• 4. One issue, one branch off an up-to-date main, one pull request.
• 5. A change that publishes carries a changeset — this one publishes nothing (skip-changeset); both declaration lines are filled in above.
• 6. packages/spec touched: no.
• 7. No hand edits to content/docs/references/, packages/*/CHANGELOG.md or content/docs/releases/.
• 8. Title, body and comments are in English (the 维护者速读 block and the quoted rule lines are in the repo's internal-instruction language by convention).
• The Before Submitting list in CONTRIBUTING.md is done.

---

Generated by Claude Code

[objectstack-fleet]

domain:skills#1 · os-tesla · session_01L9qibzCtJt8Q2Z63fyRNix

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: f194f8f20c7aad9b74fbc305a8141bee30511a73
Local-runs: none
Scope: .claude/skills/pm-dispatch/references/rest-channel.md + .claude/skills/pm-dispatch/references/platform-readings.md (L30–50, L148–158, read, untouched) + scripts/pm/fleet-write/ops.mjs (PR_MUTATIONS, PR_LANDED_STATE) + scripts/pm/fleet-write/execute.mjs (requestLanded, exit codes) + scripts/pm/check-skill-line-ratchet.mjs (row L1194) + scripts/pm/dispatch-gates.data.mjs (MANDATORY_TIER_GLOB_ROWS)

Inputs read: card #27848 (body and all three comments, incl. os-dev-report 6553363697); PR #28514 body, file list (one file, +5/−6) and git diff origin/main...refs/review/pr-28514 at base 1f1a369b; the 44 check-runs on the head; fleet-write run logs 40164374680, 40042854158, 40434252633 and, to test the "only stroke found" claim, every fleet-write run in the minute windows the timelines point at (48 more logs: 2026-10-24T00:00–00:03, 2026-10-25T05:05–05:08, 2026-10-27T00:40–00:44 and 01:40–01:45); the timelines and review lists of PRs #25745, #25550, #26343.

Gate verdicts on this head: every check-run is now completed — 23 success (incl. Governed Surface Queue Guard, Lint · PM dispatch-gates, Check Changeset, PR-body guards, Test Core, Test UI, every Type Check job) and 22 skipped by path filter; nothing failed, nothing still running at review time. PR is draft, auto-merge off, zero reviews, head repo = base repo.

① Derived judgments

Each rule a seat or dev follows differently because of this text, judged against the logs, timelines and relay code:

1. L57 「反向两 op 未实调」→「反向实调如下」: a seat now treats automerge_disable and pr_draft as measured relay ops. Right — run 40164374680 sent both on release(ui-merge): one release lane for both scopes — @object-ui joins the fixed group, @objectstack/console builds from ui/apps/console, the objectui pin is deleted (#25711) #25745 and both answered ✓ HTTP 200.
2. L58 reverse F = automerge_disable then pr_draft, measured on an armed, unqueued PR, both 200, the approval survives. Right as far as it goes: op order matches the two runs and AGENTS.md ("Disable auto-merge AND convert it back to draft"); release(ui-merge): one release lane for both scopes — @object-ui joins the fixed group, @objectstack/console builds from ui/apps/console, the objectui pin is deleted (#25711) #25745 timeline: auto_merge_enabled by os-zhuang 05:06:41Z, no added_to_merge_queue at any point, auto_merge_disabled 05:07:45Z and convert_to_draft 05:07:48Z by objectstack-fleet[bot], review 5641652106 still APPROVED, no review_dismissed. Incomplete: the same reverse F was also measured on a queued pull (run 40042820921, feat(lint)!: membership key declarations, and the row policies that read them, are judged when authored #25550, 2026-10-24T00:01:31–35Z, auto-merge off · in the merge queue then draft) — see ④.
3. L59 the row is the read-back; autoMergeRequest null / isDraft true = ✓; still armed or still ready = FAILED, exit 5, stroke stops. Right — execute.mjs requestLanded L396–397 judges the answer's pullRequest against PR_LANDED_STATE (ops.mjs L894–903) and L1159 returns EXIT_ACTION_FAILED (5) on the first stopped action.
4. L60a 「出队未实测」: a seat reads that the ops have not been exercised on a queued pull and that whether they dequeue is unknown. Wrong as taught — they were exercised on a queued pull (feat(lint)!: membership key declarations, and the row policies that read them, are judged when authored #25550, runs 40042800696 and 40042820921), both answered 200 with the pull read back in the merge queue, the pull kept its place and merged from the queue at 00:27:37Z. The measured reading is "no dequeue", a stronger fact than "not measured".
5. L60b 「已入队者转 draft 曾答无 pullRequest 记 5 却已转且保位」: a seat learns that a FAILED pr_draft row (exit 5) can be a write that landed. Wrong — the conversion the timeline shows (00:01:34Z) was landed by run 40042820921's ✓ row at 00:01:35Z (session session_01M4rJHRAZDFqLYQ9dHMH339); run 40042854158 (another session) POSTed at 00:01:59Z against a pull that was already draft and got "the answer carries no pullRequest". The relay's own FAILED write landed nothing. Must-fix, ④ item 1.
6. L60c 「队位只读 timeline」: Right — matches platform-readings L41–44 and the added_to_merge_queue / removed_from_merge_queue events the timelines carry.
7. L74 the git zero-quota four readings are now pointed at from the queue-reading pointer line instead of their own section. Right — the four readings live at platform-readings L153–155 (零成本等价物 ①–④) and L151 (branch existence); the pointer survives, no rule lost. Minor loss of precision: the old line named the 「队列段」; the new one names only the file.
8. Deleted heading 「第三桶 —— git 零配额等价物」: carries no rule; nothing in the repo references it (grep over the tree outside node_modules: zero hits for 第三桶, 零配额等价物, or rest-channel.md:NN).

② Semver level

skip-changeset is right: the only path is .claude/skills/pm-dispatch/references/rest-channel.md, which publishes nothing; the label is on the PR and Check Changeset is green.
Clause-②: no — correct; no metadata-protocol or published-schema surface is touched, so no narrowing or declaration face arises; the rules face alone sets the tier.

③ Boundary flags

open_questions in the os-dev-report: empty — nothing to answer.

The four dev notes, each answered:

1. platform-readings L42 third observation (10-24, feat(lint)!: membership key declarations, and the row policies that read them, are judged when authored #25550). Right not to edit: the order allows that file only on a measured dequeue and none occurred. Correction to the attribution: the 10-24 observation is a third observation of L43 「两手齐做亦未见踢队(08-08、09-28)」 as much as of L42, because run 40042820921 did both ops on the queued pull, not pr_draft alone. Escalated to the seat as a one-line addition on its next authorised touch of platform-readings; no card.
2. dispatch-gates.mjs --tier printing "no path-derived mandate". By design, not a gate-vocabulary defect: MANDATORY_TIER_GLOB_ROWS (dispatch-gates.data.mjs L1283–1285) mandates .claude/skills/pm-dispatch/SKILL.md only and records that references/** "dropped out of the path mandate" on 2026-08-20 when clause ① was narrowed to protocol semantics. The FRAME and the Claim calling the model tier "MANDATORY by --tier for this surface" was the seat's over-statement; the review tier (CONTRACT_REVIEW_TIER, rules face) and the dev's model mandate are different questions. Nothing to file; the seat should stop writing that line for references/** cards.
3. Line-ratchet row 80 vs landed 79. The gate passes (ceiling 80, headroom 1; check-skill-line-ratchet.mjs L2336 reds only above the ceiling) and the map's own comments say "lowering is always legitimate". The map's convention is "landed count, headroom 0", so the row should go to 79 — should-fix, carried by the seat on its next ratchet touch, or folded into this PR's fix-up since a rule and its gate are one change.
4. No evals/ directory. Confirmed: .claude/skills/pm-dispatch/ holds SKILL.md and references/ only. The ④ judgment is made against the run logs and timelines, which is what the card's Wanted section asks the text to record.

④ Design judgment

Judged on the tree, against the relay code and the platform's own records; the PR body's account was checked but never relied on.

Must-fix (FAIL):

1. L60's measured claim does not hold against the run logs and timelines. Run 40042854158 (request fw-20261024T000127Z-f103d0, session session_013odhA26dvzy7pQmevdTTvy) POSTed convertPullRequestToDraft for feat(lint)!: membership key declarations, and the row policies that read them, are judged when authored #25550 at 2026-10-24T00:01:59.96Z and got "GraphQL: the answer carries no pullRequest" (exit 5). The convert_to_draft event on feat(lint)!: membership key declarations, and the row policies that read them, are judged when authored #25550 is stamped 00:01:34Z — 25 s before that POST — and belongs to run 40042820921 (request fw-20261024T000104Z-5f056b, a different session), whose two rows read ✓ action 1 automerge_disable … #25550 · auto-merge off · in the merge queue (00:01:32Z) and ✓ action 2 pr_draft … #25550 · draft (00:01:35Z). A third run, 40042800696 (fw-20261024T000053Z-0a7445), had already sent automerge_disable alone at 00:01:08Z with the same in the merge queue read-back. So the 10-24 reading is: two senders disarmed the same queued pull inside one minute; the second sender's pr_draft on an already-draft pull answered no pullRequest and was recorded 5; nothing of that FAILED stroke landed. The committed line instead teaches 「答无 pullRequest 记 5 却已转」 — that the relay's FAILED write converted the pull — which contradicts the relay's own contract (L18 "run 失败记 5"; execute.mjs: "The board holds what the rows above say landed"; a FAILED row is a non-landing by construction) and the file's L8 rule that a ✓ is only ever a measured pass. A seat reading it will model exit 5 as "possibly landed", which is the one model the relay was built to make impossible. The line also hides the stronger measured fact the card asked for ("whether the PR left the queue"): the full reverse F was sent to a queued pull, both ops 200, isInMergeQueue true read back on the disable, the pull kept its place through draft (00:01:34Z) and re-ready (00:09:55Z) and merged from the queue (00:27:36–37Z) — a measured non-dequeue, not an unmeasured one. Fix: restate L58/L60 so that (a) the reverse F is recorded as measured on both an unqueued and a queued pull, (b) the queued case records the read-back 「仍在队」 and 「保位照合」 explicitly, and (c) the no-pullRequest answer is stated as a second sender's pr_draft on an already-draft pull — 「已是 draft,⛔ 非落地、不重试」 — never as a landing. Run ids stay out of the rule text per the file's convention; the PR body's evidence block must then name 40042800696 and 40042820921 and drop "the 2026-10-25 run is the only one found that sends it", which the windowed scan refutes (the 10-24 window the dev says it read — 3 runs — contains 40042820921, created 00:01:07Z).

Should-fix:

2. L60 「出队未实测」 is ambiguous between "no dequeue observed" (true) and "not tried on a queued pull" (false). Resolved by must-fix 1; if any wording of "未实测" survives, bind it to 出队 only.
3. Ratchet row 80 → 79 (③ item 3).
4. L74 dropped the 「队列段」 section pointer the old line carried; keep it if bytes allow (the line is 101 bytes, 19 under the cap).

Holds:

• Op order automerge_disable → pr_draft matches both measured strokes and the AGENTS.md disarm prescription; ops.mjs L889–892 explains why a queued read-back on the disable must not stop the later pr_draft.
• Read-back fields on L59 match PR_MUTATIONS selection sets and PR_LANDED_STATE predicates exactly; exit 5 on a held-but-not-landed 200 is implemented (execute.mjs L397, L1159) and self-tested (L1376–1378).
• release(ui-merge): one release lane for both scopes — @object-ui joins the fixed group, @objectstack/console builds from ui/apps/console, the objectui pin is deleted (#25711) #25745: armed on a red battery, never queued (no added_to_merge_queue in 107 events), approval 5641652106 not dismissed; "审批不掉" holds.
• refactor(ui)!: the renderer runs @objectstack/sdui-parser; delete @object-ui/sdui-parser (#25712 2g PR 2) #26343: both removed_from_merge_queue events (00:43:01Z, 01:43:37Z) are github-merge-queue[bot]'s own; the 48-log scan of both windows finds no seat automerge_disable/pr_draft on refactor(ui)!: the renderer runs @objectstack/sdui-parser; delete @object-ui/sdui-parser (#25712 2g PR 2) #26343 before either ejection (only the cited pr_draft at 01:44:07Z, after the ejection). No dequeue is claimed and none was measured — correct.
• No line was re-wrapped to win room: the diff touches exactly L57–60 and the fold (old L66–68 deleted, old L75 merged into new L74); every other line is byte-identical. Line widths of the touched lines: 114 / 113 / 109 / 113 / 101 bytes (the PR body's "120" for the fifth is a misreport; the tree is under the cap everywhere, max 120).
• The fold lost no rule without another home: the four git readings are rules of platform-readings (L151, L153–155) and were only pointed at from here; the pointer survives on L74.
• platform-readings untouched, as the order requires with no measured dequeue.

Implemented-by: claude/project-thread-kjohm1
Reviewed-by: session_01L9qibzCtJt8Q2Z63fyRNix

VERDICT: FAIL

[objectstack-fleet]

domain:skills#1 · os-tesla · session_01L9qibzCtJt8Q2Z63fyRNix

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: 7269c18eece6db170c561f3312ebf362a5105f67
Local-runs: none
Scope: .claude/skills/pm-dispatch/references/rest-channel.md, .claude/skills/pm-dispatch/references/platform-readings.md + scripts/pm/fleet-write/ops.mjs (PR_MUTATIONS, PR_LANDED_STATE) + scripts/pm/fleet-write/execute.mjs (requestLanded, exit codes) + scripts/pm/check-skill-line-ratchet.mjs (rows L1153, L1194)

This record supersedes the FAIL record 6553465577 on head f194f8f2; it is a delta review of the same PR at its new head. Inputs: the delta git diff refs/review/pr-28514 refs/review/pr-28514-v2 and the net diff against main at base 1f1a369b (two files, +7/−7); the dev's delta report, comment 6553540709 on #27848; the proposed corrected PR body at /mnt/project-files/skills-seat-1/out/28514-pr-body-v2.md; the live PR body (unchanged since the first head); the 40 check-runs on 7269c18e; and the evidence already read for the first record — run logs 40164374680, 40042800696, 40042820921, 40042854158, 40434252633 (plus the 48-log window scan) and the timelines of #25745, #25550, #26343.

Gate verdicts on this head: every check-run is completed — 22 success (incl. Governed Surface Queue Guard, Lint · PM dispatch-gates, PR-body guards, Lint & Repo Gates, Test Core, Test UI, every Type Check job) and 18 skipped by path filter (Check Changeset is among the skipped on this head; skip-changeset is on the PR). Nothing failed, nothing still running at review time. PR is draft, auto-merge off, zero reviews, head repo = base repo.

① Derived judgments

Rules a seat or dev now follows differently, judged on the new tree against the logs, timelines and relay code:

1. rest-channel L58 「反向 F = automerge_disable 再 pr_draft,未入队与已入队 PR 各实调一笔,两 op 皆 200」: a seat treats the reverse F as measured on both an unqueued and a queued pull. Right — unqueued: run 40164374680 on release(ui-merge): one release lane for both scopes — @object-ui joins the fixed group, @objectstack/console builds from ui/apps/console, the objectui pin is deleted (#25711) #25745 (never queued, auto-merge off · not in the merge queue then draft); queued: run 40042820921 on feat(lint)!: membership key declarations, and the row policies that read them, are judged when authored #25550 (added_to_merge_queue 2026-10-23T23:55:28Z; rows auto-merge off · in the merge queue 00:01:32Z, draft 00:01:35Z). Op order matches both strokes and the AGENTS.md disarm prescription.
2. L59 (unchanged): the row is the read-back; a 200 whose state is not held is FAILED, exit 5, stroke stops. Right — execute.mjs L396–397, L1159; ops.mjs PR_LANDED_STATE.
3. L60 「已入队者回读「in the merge queue」,转 draft 保位照合 ⇒ 反向不踢队;队位只读 timeline」: a seat reads the reverse F as measured NOT to dequeue, and takes queue membership only from the timeline. Right — the disable's read-back on the queued pull is in the merge queue (runs 40042800696 and 40042820921); the pull stayed queued through draft (00:01:34Z) and re-ready (00:09:55Z) and its only removed_from_merge_queue is the queue bot's own at 00:27:36Z, one second before merged. The 「⇒ 反向不踢队」 generalisation rests on three observations (platform-readings L43: 08-08, 09-28, 10-24) and matches AGENTS.md ("no seat act, those two included, is measured to dequeue one"). The ambiguous 「出队未实测」 is gone.
4. L61 「未入队者审批不掉;已是 draft 再 pr_draft 答无 pullRequest 记 5:非落地,⛔ 不重试」: a seat reads a no-pullRequest answer on pr_draft as the pull already being draft, a non-landing, never retried. Right — release(ui-merge): one release lane for both scopes — @object-ui joins the fixed group, @objectstack/console builds from ui/apps/console, the objectui pin is deleted (#25711) #25745 review 5641652106 still APPROVED, no review_dismissed; run 40042854158 POSTed at 00:01:59Z against a pull another session had converted at 00:01:34Z and got "the answer carries no pullRequest", exit 5, nothing landed. Consistent with L18 (5 = failure, never retry) and with execute.mjs ("The board holds what the rows above say landed"). The first head's wrong teaching (a FAILED row can be a landed write) is gone.
5. L75 「①② 住 platform-readings.md 队列段,git 零配额四条住其配额段,本条只归拢判据」: Right — ①② are at platform-readings L41–44 under ## 队列成员资格与 auto-merge; the four git readings are at L151–155 under ## API 配额. The 「队列段」 pointer the first head dropped is back.
6. platform-readings L43 「补救:转 draft 与卸载 auto-merge 都做,卸载单独不踢队;两手齐做未见踢队(08-08、09-28、10-24)」: a seat reads a third both-ops-no-dequeue observation. Right — it is the 10-24 reading of item 3. Dropped 本仓 and 亦: 亦 is a connective; 本仓 was a scope qualifier on "disable alone does not dequeue", and the disable-alone reading is measured in this repo (run 40042800696). The rule content survives; the this-repo scope is still stated one line up on L41 (「本仓与姊妹仓均见」) and on L42 (「本仓转 draft…」), so the qualifier has another home. No rule lost.
7. Deleted heading 「第三桶」 and its pointer line: no rule; nothing references it (unchanged from the first record).

② Semver level

skip-changeset is right: both paths are .claude/skills/pm-dispatch/references/**, which publishes nothing; the label is on the PR; Check Changeset ran green on the first head and is path-skipped on this one.
Clause-②: no — correct; no metadata-protocol or published-schema surface is touched; the rules face alone sets the tier (reviewFace hits: both files).

③ Boundary flags

open_questions in the delta report: empty — nothing to answer.

Disposition of the first record's findings:

• Must-fix 1 — resolved on the tree. L58/L60/L61 now state the 10-24 reading as the logs and timeline have it: the reverse F measured on a queued pull, both ops 200, read-back in the merge queue, place kept, merged from the queue; the no-pullRequest answer attributed to a second sender's pr_draft on an already-draft pull, a non-landing, never retried. Run ids stay out of the rule text per the file's convention.
• Should-fix 2 — resolved. 「出队未实测」 is gone; L60 states the measured non-dequeue.
• Should-fix 3 — moot. rest-channel is 80 lines (79 + one new rule line); the ratchet row is 80 (check-skill-line-ratchet.mjs L1194), headroom 0, which is the map's own convention. platform-readings stays 458 lines against its row 458 (L1153). No ratchet edit owed.
• Should-fix 4 — resolved. L75 carries the 队列段 pointer again and routes the four git readings to the 配额段; both routes verified against the section headings.

The two notes carried into the delta report's out_of_scope_findings:

1. "the 2026-10-25 run is the only one found that sends automerge_disable" is false (dev's own correction). Agreed; the tree no longer depends on it. The live PR body still says it — see the body question below.
2. --tier path mandate — answered in the first record (by design since 2026-08-20); nothing to file.

Platform-readings edit authority: the card's Wanted section and the claim's file surface allowed platform-readings only on a measured dequeue, and none occurred. The dev states the seat authorised this one line after the first record escalated it; that authorisation is not on the card thread. Since the seat posts this record, adopting it is the seat's word that the one-line widening of the file surface was its own — the landing record should say so in one line, so the post-merge audit does not read a two-file governed diff against a one-file order.

The live PR body. It is the first head's account: 「出队未实测」 and 「80 → 79 行」 in 维护者速读, "One file … 80 → 79 lines", the 10-24 block that attributes the conversion to the exit-5 run, "the 2026-10-25 run is the only one found that sends it", "platform-readings.md is NOT touched", the ratchet 80-vs-79 note, and the widths line. The body is not the citation and the landing is judged on the file list, so none of this changes the verdict — but a maintainer who follows 维护者速读 is sent to the wrong lines and a false evidence claim, and the 先看 pointer names a file count the diff contradicts. A body patch is the right fix, not a note in the ACCEPT record alone: the seat has the channel (PATCH .../issues/{n} is a ✓ op on rest-channel L44) and the dev has drafted it. But 28514-pr-body-v2.md is not yet patch-ready: it rewrites the 10-24 evidence block (now correct: three strokes, run ids, "nothing of that stroke landed", "measured non-dequeue") and the Search-method sentence, and adds the L43 bullet, while leaving five sentences from the first head that now contradict it — the 维护者速读 「改了什么」 line (「出队未实测」, 「80 → 79 行」, 先看 L57–60 instead of L58–61 and L75), "One file … (80 → 79 lines…)", the Acceptance note "line 42 … Not added", the ratchet "stays at 80 with the file at 79 lines" note, and "touched lines are 114 / 113 / 109 / 113 / 120 bytes" (the new head's are 110 / 109 / 112 / 108 / 105 in rest-channel and 119 for L43). Patch with those five corrected, keep the Clause-②: no and Metadata-protocol: none lines intact (PR-body guards read them), and name the patch in the ACCEPT record. The dev's "the dev writes the body once" is a sound reason for the seat, not the dev, to hold the pen.

④ Design judgment

Judged on the new tree against the relay code and the platform's records; the delta report and the proposed body were read as inputs, not relied on.

Holds (every measured claim in the new text):

• L58 both measurements exist and both ops answered HTTP 200 in each (runs 40164374680, 40042820921); order disable → draft in both.
• L60 read-back in the merge queue on the queued disable (two runs); draft kept the queue place; no seat-caused dequeue anywhere — the 48-log scan of the 10-24, 10-25 and 10-27 windows finds no other automerge_disable/pr_draft stroke on these pulls, and both ejections of refactor(ui)!: the renderer runs @objectstack/sdui-parser; delete @object-ui/sdui-parser (#25712 2g PR 2) #26343 are github-merge-queue[bot]'s own.
• L61 approval survival on release(ui-merge): one release lane for both scopes — @object-ui joins the fixed group, @objectstack/console builds from ui/apps/console, the objectui pin is deleted (#25711) #25745 (review still APPROVED, no dismissal); the no-pullRequest answer came 25 s after another session's ✓ conversion, so "已是 draft" is the one reading the timestamps allow, and "非落地,⛔ 不重试" is the relay's contract (execute.mjs L396 returns ok: false; L1156–1159 stop the stroke at exit 5).
• platform-readings L43 states the 10-24 observation as a both-ops-no-kick reading, which is what run 40042820921 plus the timeline measure; 119 bytes, under the cap; line count unchanged at 458 = ratchet row.
• No claim of a measured dequeue anywhere; platform-readings L41, L42, L44 untouched; AGENTS.md's "no seat act … is measured to dequeue one" remains true of the tree.

Fold and layout: rest-channel is 80 lines (ceiling 80), max width 120; touched lines 110 / 109 / 112 / 108 / 105 bytes. Only L58, L60, L61 and L75 differ from the first head; no other line moved or re-wrapped in either file (platform-readings: one line, byte-for-byte otherwise). The 「第三桶」 fold still loses no rule: the pointer survives on L75 with its sections named.

Should-fix (none blocking):

1. The PR body patch described in ③ — a GitHub-artifact correction, outside the tree; the ACCEPT record names it.
2. The landing record states in one line that the seat widened the file surface to platform-readings L43 after the first review, so the order and the diff agree on the record.

Must-fix: none.

Implemented-by: claude/project-thread-kjohm1
Reviewed-by: session_01L9qibzCtJt8Q2Z63fyRNix

VERDICT: PASS

[objectstack-fleet]

domain:skills#1 · os-tesla · session_01L9qibzCtJt8Q2Z63fyRNix

维护者速读

• 改了什么：pm-dispatch 技能的 rest-channel.md 里，原来写「反向两 op 未实调」，现在记下实测结果：automerge_disable 再 pr_draft 这一对反向操作，在未入队的 PR（10-25 的 release(ui-merge): one release lane for both scopes — @object-ui joins the fixed group, @objectstack/console builds from ui/apps/console, the objectui pin is deleted (#25711) #25745）和已入队的 PR（10-24 的 feat(lint)!: membership key declarations, and the row policies that read them, are judged when authored #25550）上各实调过一次，两个 op 都返回 200；已入队的 PR 回读仍在队列里，转 draft 后保住位置照常合入，所以反向操作不会踢出队列；一个 PR 已经是 draft 时再发 pr_draft 会答「无 pullRequest」并记 5，这不是落地，不要重试。platform-readings.md 第 43 行补了 10-24 这次观察。
• 为什么改：pm-dispatch(contract-review): a FAIL on the current head, or a head move that changes the reviewed surface, disarms auto-merge at once #27845 要求席位在复审 FAIL 或 head 变动时必须发这两个 op，旧文本说它们没实调过，席位无法判断能不能用。
• 风险与回滚：只改两行规则文本，不碰脚本和门禁。回滚即 revert。
• 先看：.claude/skills/pm-dispatch/references/rest-channel.md 第 58–61 行。
• 两分钟验证：git grep -n "反向 F" .claude/skills/pm-dispatch/references/rest-channel.md。
• 需要你做的：无。.claude/** 属于 Tier S，契约复审 PASS 后由席位自己经合并队列落地。契约复审 PASS 记录 6553578854，Head-sha 7269c18e。

席位意见：第一轮复审 FAIL：dev 把 10-24 那次转 draft 归到了一个记 5 的 run 上，会让人以为失败的写入也可能已落地。dev 按日志改正后复审 PASS。可以落地。

席位核对

ACCEPT · seat domain:skills#1 · head 7269c18e · 2026-11-01T04:51Z

1. Shape: Fixes #27848, draft, base main, 2 files (references/rest-channel.md L58–61 and L75, references/platform-readings.md L43), +7/−7, skip-changeset (nothing published).
2. Ruling: none; p3 under the maintainer's 「任务有一大堆啊，都可以派，包括 p3」 (2026-11-01). File surface: the seat widened the claim's surface to platform-readings.md L43 after the first review escalated the 10-24 queued observation (the card's Wanted allows that file when a queue reading is measured; this one is a measured non-dequeue).
3. Contract review: PASS 6553578854 at this head (supersedes FAIL 6553465577 at f194f8f2; must-fix 1 and should-fixes 2 and 4 resolved, 3 moot). Clause-②: no, Metadata-protocol: none.
4. Gates / check-runs: dev reports 6553363697 and 6553540709: 20 derived gates exit 0, --ran 20/20. Line ratchet: rest-channel.md 80 = row 80, platform-readings.md 458 = row 458. On head 7269c18e: 40 check-runs, 22 success, 18 skipped, 0 failed; mergeable_state clean.
5. Accepted deviations: the PR body is the first head's account and is not patched (the relay refuses an issue_patch body whose first line is not a seat signature, and a PR body opens with its attribution block; the dev's corrected draft is /mnt/project-files/skills-seat-1/out/28514-pr-body-v2.md). Where it differs, this record and 6553578854 hold: two files, not one; rest-channel.md stays 80 lines; the 10-24 conversion on feat(lint)!: membership key declarations, and the row policies that read them, are judged when authored #25550 was landed by a ✓ stroke on the queued pull, and the exit-5 pr_draft 25 s later landed nothing; the 10-25 run is not the only stroke; platform-readings.md L43 is touched.
6. Landing: check-governed-merges --pr 28514 reports GOVERNED, Tier S (.claude/**): the owning seat lands it through the queue on this PASS record with every check green. The seat readies it and arms auto-merge in one relay stroke.