分诊章程没有「发布是否已上膛」的判据 —— 一个席位据此手打了一个两仓都永不命中的查询，读成了假放行

[os-musk]

由分诊座位（#6015 · R+194，会话 session_015WpYyzhX8x2kEhouLBidt8）立卡，自 objectui#9134 分流而来。⛔ 未认领、未派发。

objectui#9134 已定级（pm:queue · priority:p2 · domain:devx），它承载的是本仓外的那一半：objectui 自己的两份文档教了一个错 title。这一张承载章程的那一半，按 anchoring rule 落点在 .claude/skills/pm-dispatch/** ⇒ domain:skills，与那张不同仓、不同面，⛔ 不可合并。

缺口，实测

「这张卡会不会被下一次发布关掉窗口」是一个反复出现的定级输入——priority 和 pm:blocked 都可能挂在它上面。章程里没有任何地方说该怎么测它：

grep -rn "发布窗口\|release window\|changeset 窗口\|changeset-window" .claude/skills/pm-dispatch/
  -> 0
对照  grep -rc "分诊" .claude/skills/pm-dispatch/SKILL.md
  -> 44                      ⇒ 仪器在读这些文件

Version Packages 在章程里只出现 2 次（SKILL.md:34 / :683），两次都是禁令（⛔ 不合并、⛔ 不催），⛔ 没有一次是检测指令。

⇒ 每个席位在需要这个判断时都在现场自己发明一个，而且发明出来的东西没有任何东西会复核。

那次实例

objectui#9065 的 watch 复核（评论 5628564370，2026-09-11T02:31Z）记下：

「Second control, for the thing that actually closes the window: open Version Packages PRs in this repo = 0（11 open PRs total）。⇒ no release is staged, so the six cannot ship this minute.」

它被明写成一条 control——即那位席位知道零需要对照，只是对照本身是坏的。当时：

仓	常驻 Changesets 发布 PR	真实 title	head
objectui	objectui#5400（开了 23 天）	chore: release packages	changeset-release/main
objectstack	#17076	chore: version packages	changeset-release/main

两仓都不叫 Version Packages，而且两仓彼此也不同名（release vs version）。⇒ 那个查询在任何状态下、任何时刻、两个仓里都恒返 0。它不是读数过期，它是一个永远打不响的查询，而它渲染出来恰好就是提问者想要的那个答案。

正确的判据，以及一个更锋利的形式

✅ head 分支 changeset-release/main —— Changesets action 自己的约定，两仓逐字相同，⛔ 不受 title 配置影响：

is:pr is:open head:changeset-release        # 两仓各返 1

⭐ 但更值得写进章程的是：「发布是否已上膛」本身就是个信息量很低的问题。 只要有任何 pending changeset，那个 PR 就恒开着——它的存在几乎不携带信息。真正携带信息的是本张卡的那条 changeset 是否已经被折进该 PR 的 head，那是对 head 上 CHANGELOG.md 的一次直读，回答的是「离永久还有多远」而不是「发布开始了没有」。

⇒ 建议章程写后者，并把前者标成不充分。

为什么值一张卡而不是一次改词

1. 它是假放行方向：错的那一侧总是「窗口没关，慢慢来」。
2. 这个窗口关掉的代价有实测先例：.changeset/seed-locale-axis.md ships two present-tense claims that PR #17013 makes false, and it is queued in the pending release PR #15334 — the correction cannot be made from the PR that falsifies it #17026 —— 发布在立卡 47 分钟后跑掉，两条现在时的假陈述永久留在 packages/spec/CHANGELOG.md 与 packages/metadata-protocol/CHANGELOG.md 里。
3. ⚠️ 那位席位做对了方法论（零配对照），错的是对照的内容。这说明缺的不是纪律而是判据文本——纪律已经在了，它指向了一个不存在的仪器。

验收

• 章程（SKILL.md 或 references/ 内合适的一篇）写入「发布是否已上膛」的判据，以 head 分支为准，并明写 ⛔ 不得以 title 判定、⛔ 两仓 title 不同且由 action 配置决定。
• 同处写入上面那个更锋利的形式：对该 PR head 上 CHANGELOG.md 的直读，才回答「这条 changeset 离永久还有多远」。
• 给出一条可照抄的查询与一条点亮对照（本卡正文两行即可用）。
• ⛔ 不改任何禁令（⛔ 不合并、⛔ 不催发布 PR 原样保留）。
• ⛔ 不建台账、不加门。

⛔ 明确未断言

• ⛔ 未断言那次假放行导致了任何损失——objectui#9065 的六张卡是否真的差点被发走，本席未测。
• ⛔ 未断言 .changeset/seed-locale-axis.md ships two present-tense claims that PR #17013 makes false, and it is queued in the pending release PR #15334 — the correction cannot be made from the PR that falsifies it #17026 的损失由这一类检查造成；它只作为「窗口关掉值多少钱」的实测先例被引用。
• ⛔ 未断言任何发布 PR 该合并或该held——那是维护者动作。

查重：release staged window changeset detection predicate 在本仓返 2（#9500 · #17026），两张都不承载这一类；仪器活（有命中）⇒ 无重复。

Refs：objectui#9134（同源的另一半，本仓外）· objectui#9065（假放行发生处）· objectui#5400 · #17076 · #17026（先例）

[claude]

Type set by the skills seat at 2026-09-12T12:15Z: Task — the card arrived from the triage seat with pm:queue · domain:skills · priority:p3 and no type; a charter clause that does not exist yet (the release-staged criterion, head-branch based, plus the sharper CHANGELOG-on-head reading) is a Task, not a Bug. The grade and routing stand as filed. Skills seat, session session_01MCLBsUgfykL74aU716rzVK.

---

Generated by Claude Code

[claude]

Claim: PM loop round 1
Session: session_01MCLBsUgfykL74aU716rzVK (GitHub os-sales, skills seat), claimed at 2026-09-12T12:16Z
Branch: claude/issue-17830-release-staged-criterion
Worktree: objectstack-issue-17830
Domain: domain:skills (filed and graded p3 by the triage seat at 09:39Z; type Task set by this seat 12:16Z)
File surface (region-declared): ONE references/ file of the pm-dispatch skill — the dev picks the one that already carries landing / release-window reading discipline (references/landing-operations.md 69/69 or references/dispatch-runbook.md 241/241 are the candidates; both ratcheted at headroom 0 ⇒ paid by density inside the same section) — gaining the 「发布是否已上膛」 criterion: head branch changeset-release/main is the test (is:pr is:open head:changeset-release, one hit per repo), ⛔ never the PR title (objectui chore: release packages vs objectstack chore: version packages, both action-configured), plus the sharper reading: the card's own changeset folded into that PR head's CHANGELOG.md is what answers 「离永久还有多远」; one copyable query and one lit control. ⛔ Not SKILL.md (PR #17837 at the human terminal on it; and :34 / :683 are prohibitions that stay as they are), ⛔ no ledger, no gate, no change to any prohibition.
Container & model: S, mode:subagent, model: opus — default tier (references/**-only); skills-seat review at the contract-review tier; governed (.claude/**) ⇒ draft PR at the human terminal with the four-piece, ⛔ never flipped ready by the seat
Clause-②: no
Thread-read: 5645818504 (this seat's type comment, 2026-09-12T12:16Z) — the card has no other comments
Serial constraints cleared: landing-operations.md and dispatch-runbook.md free — no open PR touches either (all open PRs' file lists scanned at 12:14Z); SKILL.md excluded from the face; verify lock free; H17 index ∩ this face = ∅. Card-reference face re-taken on the live board at 12:14Z: objectstack's standing release PR is #17076 (chore: version packages, head changeset-release/main, open); objectui#5400 not re-read from this session (cited from the card); git grep "Version Packages" origin/main -- .claude/skills/pm-dispatch/ → SKILL.md :34 / :683 only, both prohibitions, zero criterion (control: 分诊 hits the same files). Read-only, ⛔ not touched.

---

Generated by Claude Code

[claude]

os-dev-report

{
"issue": 17830,
"status": "done",
"branch": "claude/issue-17830-release-staged-criterion",
"pr": "#17850",
"head_sha": "df83eb3c1c8b0e6b1214fe62a02682ef5b7719e6",
"files": [
".claude/skills/pm-dispatch/references/dispatch-runbook.md"
],
"premise_still_valid": true,
"summary": "A, B and C all landed. A: the release-staged criterion is written into references/dispatch-runbook.md section 落卡与裁决记录细则, immediately before the 卡上 re-check 命令两条硬线 bullets — four lines carrying (i) head branch changeset-release/main is the test with the copyable query is:pr is:open head:changeset-release, (ii) one lit control git ls-remote origin changeset-release/main, (iii) the explicit refusal of the PR title with both action-configured spellings named (objectui chore: release packages, objectstack chore: version packages), (iv) the sharper reading — whether this card changeset is already folded into that head CHANGELOG.md — as what answers 离永久还有多远, with the staged bit marked insufficient. Carrier chosen over landing-operations.md because that file declares its own scope as ACCEPT-to-MERGED operations while this question is a grading input taken BEFORE dispatch and recorded on the card, and the chosen section is the one place in the charter that governs a card-recorded reading and its mandatory lit positive control — exactly the discipline the failing seat honoured while the instrument it pointed at did not exist. B: no SKILL.md edit; the one-line pointer is written up in the PR body as a follow-up for the PM to fold into the next SKILL.md PR. C: pnpm check:pm-skill-ratchet green — dispatch-runbook.md is 241 lines (ceiling 241; headroom 0), every edited line measured and reported. No prohibition, ledger or gate touched. Density paid by DELETION only, never a re-wrap: two rationale tails (级序理由 115 B, whose order is stated at SKILL.md:458-460; 理由两条 100 B, same section as the addition) and the three-line 三分区 dev-obligation block (64+119+73 B) collapsed to one 94 B pointer, its single source being SKILL.md:531 + references/core-rules.md:114 + the dev-agent definition. Five lines out, five in (four criterion + one pointer): net 0, file exactly at ceiling, bytes 471 out against 545 in (stated in the PR body). Assignee was already set by the PM; not touched. The PR is DRAFT and stays so — .claude/** is governed.",
"tests": "No package build or test is owed: the diff touches one .md file under .claude/ and no package. Gate sweep derived in this worktree with node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack (no paths; stderr names objectstack-ai/objectstack at 758ac40, change set 1 path). All 14 derived commands run, $? captured before any pipe, all exit 0 (table in the PR body). Key verdict lines, quoted from the gates themselves: "✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/references/dispatch-runbook.md is 241 lines (ceiling 241; headroom 0)."; "✓ check-skill-id-lint: 27 file(s) clean (pattern /#[0-9]{3,}/g)."; "check-nul-bytes: OK (scanned 8476 text file(s) ... no raw ASCII control bytes)." One gate first exited 3 — pnpm --filter @objectstack/lint run check:doc-formula-expressions, PREREQUISITE NOT MET, two unbuilt workspace packages, nothing measured; NOT recorded as a failure. The two packages were built through the shared verify lock (OS_VERIFY_LOCK_SLOT=issue-17830-dev; "os-verify-lock: VERDICT command-exit 0 · held the lock 1s · waited 0s") and the gate re-run to a real exit 0. Reconciliation node scripts/pm/dispatch-gates.mjs --ran ... prints "14 derived, 14 run, 0 NOT-MEASURED, 0 UNRUN" and "a DERIVED zero — all 14 recorded an exit code and none of them is 3". The sweep was taken on df83eb3, the final commit on the branch. No ablation and no reverse verification apply: the deliverable is charter prose, there is no guard to mutate. Instead the two instruments the text hands a seat were exercised live: git ls-remote origin changeset-release/main prints c26a619 refs/heads/changeset-release/main (control lights), and the sharper reading was performed — git fetch origin changeset-release/main:refs/os-dev-17830/relhead -f then git show refs/os-dev-17830/relhead:packages/spec/CHANGELOG.md returns the staged 17.5.0 entries, so the CHANGELOG-on-head read is operable. Repo-wide scans (pnpm lint and the 52 artifact-roster / 11 wide-population families the derivation names) are CI-owned and were deliberately not run; the derivation also notes 14 families that apply only once a changeset path exists, and there is none. Byte discipline beyond the gate: grep -naP over the candidate lines and the PR body for the ASCII control range hit nothing (exit 1).",
"gates": {
"derived_with": "node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack (no paths, in the worktree)",
"reconciliation": "14 derived, 14 run, 0 NOT-MEASURED, 0 UNRUN",
"ran": [
{
"command": "node scripts/check-closing-keyword-parity.mjs",
"exit": 0
},
{
"command": "node scripts/check-closing-keyword-parity.mjs --self-test",
"exit": 0
},
{
"command": "node scripts/check-comment-mask-corpus.mjs",
"exit": 0
},
{
"command": "pnpm --filter @objectstack/lint run check:doc-formula-expressions",
"exit": 0,
"note": "first run exited 3 = PREREQUISITE NOT MET (unbuilt @objectstack/formula + @objectstack/lint), nothing measured; built under the verify lock, re-run to 0"
},
{
"command": "pnpm check:agent-test-spelling",
"exit": 0
},
{
"command": "pnpm check:doc-authoring",
"exit": 0
},
{
"command": "pnpm check:driver-memory-census",
"exit": 0
},
{
"command": "pnpm check:nul-bytes",
"exit": 0
},
{
"command": "pnpm check:pm-governed-merges",
"exit": 0
},
{
"command": "pnpm check:pm-skill-id-lint",
"exit": 0
},
{
"command": "pnpm check:pm-skill-ratchet",
"exit": 0
},
{
"command": "pnpm check:refd-timer-probe",
"exit": 0
},
{
"command": "pnpm check:skill-frame-sync",
"exit": 0
},
{
"command": "pnpm check:watch-hint-literal",
"exit": 0
}
],
"ci": "in_progress — the full farm (pnpm lint and the repo-wide families) is CI-owned; not waited on, per the report-on-local-completion clause"
},
"premises": {
"1_charter_gap": "HOLDS. git grep -n -E -i \"changeset-release|release window|发布窗口|上膛\" origin/main -- .claude/skills/pm-dispatch/ exits 1 with no output; git grep -n \"Version Packages\" origin/main -- .claude/skills/pm-dispatch/ returns SKILL.md:34 and SKILL.md:683 only, both prohibitions. Control: git grep -l \"分诊\" origin/main -- .claude/skills/pm-dispatch/ lists 15 files including references/dispatch-runbook.md, so the instrument reads the corpus the zero was taken over.",
"2_standing_release_pr": "HOLDS. REST GET /repos/objectstack-ai/objectstack/pulls?state=open&head=objectstack-ai:changeset-release/main -> HTTP 200, exactly one element: PR 17076, title "chore: version packages", head changeset-release/main, state open, draft false, author github-actions[bot], created 2026-09-09T06:17:15Z.",
"3_ceilings": "HOLDS, read from the gate own CEILINGS map rather than from the card: references/landing-operations.md 69 lines / ceiling 69; references/dispatch-runbook.md 241 / 241; MAX_LINE_BYTES = 120; SKILL.md 812 / 812, left untouched. A third candidate the map exposes and the dispatch did not name, references/release-aftercare.md 50 / 50, was read and rejected on register (it declares itself as post-release obligations).",
"4_serial_face": "HOLDS. All 22 open PRs file lists read at 2026-09-12T12:2xZ: the pm-dispatch paths in flight are SKILL.md + references/core-rules.md (PR 17837), references/decision-analysis.md (PR 17515), references/platform-readings.md (PR 17803). references/dispatch-runbook.md and references/landing-operations.md are both free."
},
"mcp_calls": "0 — zero MCP GitHub calls for the whole run. The card body and all 9 timeline items came from the public single-card embedded JSON payload (hasNextPage false, 9/9 edges); every board read and every write (branch push, draft PR create, label POST, read-backs) went through repo-scoped REST with curl, and every tree read through git.",
"open_questions": [],
"out_of_scope_findings": [
"noted, not filed: PR 17837 (open, at the human terminal) sharpens the GENERAL half of the same discipline — a zero-hit control must share its claim subject and failure mode — in SKILL.md and references/core-rules.md. Complementary, not overlapping: it constrains what a control must be, this PR supplies the missing instrument; no text is duplicated between the two diffs. Carrier: PR 17837 author.",
"noted, not filed: references/dispatch-runbook.md carries three rules already stated verbatim in substance in SKILL.md (:158 vs SKILL.md:287, :62 vs SKILL.md:80, :167 vs SKILL.md:583), against this file own landing convention ("no rule already stated in SKILL.md"). They are payment available to a future addition in this file; :167 was deliberately NOT taken here because it is rung 3 of an enumerated ladder that would read truncated without it. Carrier: the next PR needing headroom in this file.",
"noted, not filed: the charter clause this card asked for is reachable only from references/dispatch-runbook.md — SKILL.md carries no pointer to it, and adding one is out of scope on this card (PR 17837 holds SKILL.md at the human terminal). Written up as work item B in the PR body for the PM to fold into the next SKILL.md PR. Carrier: the PM."
]
}

---

Generated by Claude Code

[claude]

ACCEPT — contract-tier review of PR #17850 (head df83eb3c, report 5645945788) by the skills seat at 2026-09-12T12:43Z.

Scope against the claim (5645825394): one file, references/dispatch-runbook.md (+5/−5), section 「落卡与裁决记录细则」, placed immediately before the two hard lines on zero readings so the reach for a lit control lands on a real instrument: (i) 「已上膛 = 开着的 head changeset-release/main PR」 with the copyable is:pr is:open head:changeset-release; (ii) the lit control git ls-remote origin changeset-release/main and 「⛔ title 永不是判据」; (iii) both action-configured titles named; (iv) the sharper reading — this card's changeset folded into that head's CHANGELOG.md answers 「离永久还有多远」. Density paid by deletion only: two rationale tails (级序理由 — the order itself stays at SKILL.md :458–460; 理由两条 — the rule above it stands) and the three-line 三分区 dev-obligation block collapsed to a pointer (its single sources verified by the seat: SKILL.md :531, os-dev.md :182–184). Prohibitions SKILL.md :34 / :683 untouched; no ledger, no gate; SKILL.md not edited.

Read on the head, not the report: 241 → 241 lines; the five edited lines measure 117 / 96 / 119 / 119 / 94 bytes (the file's widest line, 120 B, is pre-existing); no three-digit card id in the file (check:pm-skill-id-lint is why the lit control is the ls-remote hit and not a PR number). Both instruments were exercised live by the dev: the control lights (c26a6190 refs/heads/changeset-release/main), and the CHANGELOG-on-head read returns the staged 17.5.0 entries. Carrier choice accepted: landing-operations.md declares itself ACCEPT→MERGED operations, while this question is a grading input recorded on the card before dispatch. Commit trailers model-free, no card relation; the PR body carries Fixes #17830. Gates: 14 derived / 14 run / all exit 0 (one prerequisite exit 3 re-run green under the verify lock).

Follow-up the seat carries: a reader of SKILL.md alone will not find the clause — one pointer line beside SKILL.md :34 rides the next SKILL.md PR (#17728 is next on that chain once PR #17837 lands); recorded here and on the seat post, ⛔ not this PR's.

Recorded from the dev, not filed: dispatch-runbook.md :158 / :62 / :167 each restate a rule SKILL.md already carries — payment available to the next addition in this file (:167 deliberately not taken: rung ③ of an enumerated ladder).

Governed surface (.claude/**) ⇒ PR #17850 stays a draft at the human terminal: needs-user-decision on the PR, the seat's 维护者速读 comment, reviewers os-zhuang + hotlong requested; ⛔ the seat never flips it ready, never arms auto-merge. On MERGED: landing record here, pm:dispatched + assignee cleared in one write. Skills seat, session session_01MCLBsUgfykL74aU716rzVK, 2026-09-12T12:43Z.

---

Generated by Claude Code

[claude]

Landed — PR #17850 MERGED 2026-09-12T15:50:31Z as 2f1fec6f6 (merged by os-zhuang through the merge queue after approval at the human terminal). Skills seat stay-behind record, session session_01MCLBsUgfykL74aU716rzVK (GitHub os-sales), 2026-09-12T15:51Z.

Verified on origin/main by git (not on the PR): the merge commit is an ancestor of origin/main (ancestor=yes) and single-parent (parents field words=2: commit + 1 parent); references/dispatch-runbook.md is 241 lines (ceiling 241, no raise) and carries the release-window criterion lines (「已上膛」 with the changeset-release/main lit control) in 「落卡与裁决记录细则」. ACCEPT 5645955481, 速读 5645955635.

Residue cleared in this stroke: pm:dispatched off, assignee cleared (read back). domain:skills and type stay (ownership, not state). GitHub closes the card on the Fixes keyword.

---

Generated by Claude Code