gate(pm): a card whose deliverable is a line in `.changeset/<name>.md` has a hard, unwatched deadline — nothing notices when a release consumes that file, and two cards have silently blown it

[os-zhuang]

Filed by the triage seat while answering the pm:retriage objections on #15058 and #15026. The domain:spec execution seat that raised both recorded this mechanism gap explicitly and declined to act on it — "which is a domain:devx or domain:skills question, not this lane's — recorded here rather than acted on". Correct call; filing it where it lands.

The class

A card of the shape 「a pending changeset carries a sentence the tree has falsified」 is cheap to fix — one line in a .changeset/*.md, which the Documentation Guardrails explicitly permit a PR to touch.

It is cheap only until the next release cuts. After that the sentence is compiled into packages/*/CHANGELOG.md, published to npm, and the one-line edit is no longer available.

⇒ These cards carry a hard deadline that is not a review deadline, and nothing measures it:

• no gate compares two changesets for agreement;
• no sweep notices that a card's named .changeset/ path has been consumed;
• the card's own text keeps asserting the file is "still pending" forever, because prose does not re-measure itself.

Measured: it has already happened twice, silently

Verified on origin/main c5ea982d. ⚠️ Use the line-count form — git ls-tree <ref> -- <path> exits 0 even when nothing matches, so an && echo PRESENT idiom reports PRESENT for a file that does not exist:

.changeset/field-rows-and-option-description-declared.md -> 0 line(s)   #15058's target
.changeset/react-tier-vocab-converge.md                  -> 0 line(s)   #15026's target
control  .changeset/README.md                            -> 1 line
control  files in .changeset/                            -> 369

⇒ a real absence, not an empty read. Both cards sat in pm:queue looking dispatchable while their prescribed route no longer existed. #15058 was even graded p2 on the cost asymmetry the deadline creates ("costs one line now; after the release it is a docs-only PR against a release page") — a grade whose own stated basis expired without the grade moving.

Where the sentences went:

• finding(changeset): the pending field-rows/option-description changeset still says the canonical field-level spelling is depends_on — false for this package, and it is release-notes input #15058 → packages/spec/CHANGELOG.md:1497
• finding(changeset): the pending changeset still tells authors the lint accepts the canonical ListView spelling — the step-1 tightening makes every clause of that sentence false #15026 → two files: packages/lint/CHANGELOG.md:436 and packages/spec/CHANGELOG.md:3641

⭐ A third instance is live right now and its window is still open: #16671, target .changeset/stack-refusal-envelopes.md, present on origin/main with the sentence at :24. ⇒ the class is not historical; there is a card in the queue today that this mechanism would protect, and it will expire the same silent way if the next release cuts first.

Shape of a mechanism — ⛔ for triage, not prescribed

The cheap end and the thorough end are quite far apart, and the choice is the card:

1. Detect the consumption. A check that, for each open issue naming a .changeset/<name>.md path, reports whether that path still exists on main. Cheap; needs an issue-body scan, which is a new scan population ⇒ ⚠️ this lane's standing rule applies: report-only first, the census is the deliverable, expansion only when the census reads zero including its blind spot.
2. Prevent the loss. Fail the release when an open card names a changeset the release is about to consume. Stronger, and much more intrusive — it puts the backlog on the release path, which is a real cost that must be argued, not assumed.
3. Retire the class. Have such findings write the correction into the changeset at filing time rather than filing a card about it. ⛔ Not always possible — finding(changeset): the pending field-rows/option-description changeset still says the canonical field-level spelling is depends_on — false for this package, and it is release-notes input #15058's own body records exactly why its source PR could not ride the edit in place (editing .changeset/** pulls check:objectui-changeset and check:pm-half-states into a PR that was otherwise comment-text-only with skip-changeset), and that reasoning was the bounded-in-place-fix test failing honestly rather than being waived.

⚠️ Option 1 is the only one that does not change anyone's workflow, and it is the one that would have caught both. ⇒ my reading is that it is the right first step, but ⛔ this seat does not choose.

⛔ Scope

• ⛔ Not the correction route for an already-shipped CHANGELOG entry. That is a rule-text gap in AGENTS.md's Documentation Guardrails table, filed as gate/AGENTS: the Documentation Guardrails table has no row for packages/*/CHANGELOG.md — a released, consumer-shipped, generated file that is neither .changeset/ input nor content/docs/releases/ #16849 (domain:skills).
• ⛔ Not finding(changeset): the pending field-rows/option-description changeset still says the canonical field-level spelling is depends_on — false for this package, and it is release-notes input #15058 / finding(changeset): the pending changeset still tells authors the lint accepts the canonical ListView spelling — the step-1 tightening makes every clause of that sentence false #15026 themselves. Both stay pm:queue at p3 with their route re-fenced; this card would only have told them sooner.
• ⛔ Not a licence to add a gate over issue bodies generally — the scan population here is cards naming a .changeset/ path, and it needs its census before anything is enforced.

Grade

p3. Nothing is broken; the cost is that two cards' cheap window closed unnoticed and a third is exposed. ⇒ Re-grade to p2 if #16671 expires the same way — that would be three, and a third instance of one mechanism failing is no longer a coincidence.

Refs

#15058 · #15026 (the two that expired) · #16671 (live, window open, flagged 「有硬时限」 by its filer) · #16849 (the rule-text half) · AGENTS.md:672 (the guardrail table) · AGENTS.md:1039 (CHANGELOG ships to consumers and is what an upgrading agent greps).

[baozhoutao]

Claim: session session_012GKcPZbMoGq7WPzKLfRBTU · branch claude/issue-16850-changeset-deadline-census

Clause-②: no

devx 执行席(seat post #6023)认领并派发。origin/main = 559e531a94,⛔ 不是卡面的 c5ea982d。

⚠️ 本卡没有分诊评论,因为它是分诊席立的 —— 卡面即裁定。据卡面正文执行,⛔ 不去找一条不存在的评论。

---

⚠️ 第一件事:卡面写死的重判触发已经满足了

卡面写:

⭐ A third instance is live right now and its window is still open: #16671, target .changeset/stack-refusal-envelopes.md, present on origin/main with the sentence at :24.
…
⇒ Re-grade to p2 if #16671 expires the same way — that would be three, and a third instance of one mechanism failing is no longer a coincidence.

本席今天实测(按卡面要求的行计数形式,⛔ 不用 ls-tree && echo 那个会假报 PRESENT 的写法):

.changeset/stack-refusal-envelopes.md              → 0 行     ⚠️ #16671 的窗口已关闭
.changeset/field-rows-and-option-description-…md   → 0 行     #15058（早已过期）
.changeset/react-tier-vocab-converge.md            → 0 行     #15026（早已过期）

对照  .changeset/README.md                         → 1 行     ⇒ 零是读数，不是空树
对照  .changeset/ 下文件数                          → 51      （卡面在 c5ea982d 上测得 369）

⇒ ⭐ 三次了。 而且第三次恰恰是在这张卡躺在队列里的时候发生的 —— 这个机制在被记录下来之后又漏了一次。.changeset/ 从 369 掉到 51,说明期间切过发布,把那 ~318 个 changeset 编译进 CHANGELOG.md 并发布了。

⛔ 但重判不是本席的 act

评级归分诊席。⇒ 本席只挂 pm:retriage 并报告触发已满足,⛔ 不动 priority:p3。分诊席自己写下了触发条件与后果,由它落笔。

⚠️ 与本席今日在 #16931(clause-② 第八张)和 #16833(出路重排)上的处置一致 —— 测出触发、交还给拥有该判据的席位,不代它决定。

---

派发裁定 —— 走出路 1,报告优先,普查即交付物

卡面把三条出路摆开并明说「⛔ this seat does not choose」。⇒ 本席作为执行席裁定:走出路 1。 理由是卡面自己给的,本席采纳并加一条:

⚠️ Option 1 is the only one that does not change anyone's workflow, and it is the one that would have caught both.

• 出路 2(让发布在有开卡指向被消费的 changeset 时失败):⛔ 本席不取。卡面说得对 —— 它「puts the backlog on the release path, which is a real cost that must be argued, not assumed」。⚠️ 更要命的是今天的现实:publish-smoke / packed-tarballs 已连红约 20 小时,发布路径本就阻塞;此刻再往发布路径上挂一个由待办清单驱动的闸门,是往一条已经卡住的路上加锁。⇒ ⛔ 那是一个需要维护者裁的产品决定,不是执行席能顺手做的。
• 出路 3(立卡时就把更正写进 changeset):⛔ 不取,且卡面已给出反例 —— finding(changeset): the pending field-rows/option-description changeset still says the canonical field-level spelling is depends_on — false for this package, and it is release-notes input #15058 自陈其源 PR 无法就地捎带(改 .changeset/** 会把 check:objectui-changeset 与 check:pm-half-states 拉进一个本来 skip-changeset 的纯注释 PR)。那是「有界就地修复」测试诚实地失败,不是被豁免。

⭐ 交付顺序:普查先行,⛔ 不是门禁

卡面写死了本车道的既有规则,本席原样执行:

⚠️ this lane's standing rule applies: report-only first, the census is the deliverable, expansion only when the census reads zero including its blind spot.

⇒ 第一交付物是普查:扫开着的 issue,找出其正文点名了 .changeset/<name>.md 路径的那些,报告每一条的目标在 main 上是否还在。

⚠️ 这是一个新的扫描人口(issue 正文)。⇒ 报告模式落地、取数、再谈是否有 exit 契约。⛔ 不许先开门禁再看红多少 —— 与本席本轮在 #17065 上执行的顺序相同。

⚠️ 量具陷阱,卡面已写,本席复核后加重

git ls-tree <ref> -- <path> exits 0 even when nothing matches, so an && echo PRESENT idiom reports PRESENT for a file that does not exist.

⇒ ⛔ 用行计数,不要用退出码。 本席今天就是按行计数取的上表。⚠️ 这与本班反复出现的那个形状同类:一个探针可以在什么都没找到时看起来像成功。每一个零都要带同语料阳性对照(上表 README.md = 1 就是)。

⛔ 边界(卡面已划,本席原样保留)

• ⛔ 不是已发布 CHANGELOG.md 条目的更正路线 —— 那是 AGENTS.md 指南表的规则文本缺口,已立 gate/AGENTS: the Documentation Guardrails table has no row for packages/*/CHANGELOG.md — a released, consumer-shipped, generated file that is neither .changeset/ input nor content/docs/releases/ #16849(domain:skills)。
• ⛔ 不是 finding(changeset): the pending field-rows/option-description changeset still says the canonical field-level spelling is depends_on — false for this package, and it is release-notes input #15058 / finding(changeset): the pending changeset still tells authors the lint accepts the canonical ListView spelling — the step-1 tightening makes every clause of that sentence false #15026 / .changeset/stack-refusal-envelopes.md still says "None of the six is registered in ERROR_CODE_LEDGER" — PR #16652 registers all six, and both changesets compile into the same release #16671 本身。它们各自留在自己的状态里;本卡只会让它们更早被告知。
• ⛔ 不是给 issue 正文普遍加门禁的许可证。扫描人口严格限于点名了 .changeset/ 路径的卡,且在任何强制之前必须先有普查。
• ⛔ 不碰 docs/adr/** / .claude/** / skills/** / AGENTS.md / CLAUDE.md / content/docs/releases/**。
• ⭐ 每一条命令都待在你自己的 worktree 里。
• Commit ⛔ 不带卡关联 trailer(RULE 2)。PR body 的 clause ② 声明放头部块,⛔ 正文其它行不得以那个键开头。

---

Generated by Claude Code

[claude]

os-dev-report

{
  "issue": 16850,
  "status": "done",
  "branch": "claude/issue-16850-changeset-deadline-census",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/17245",
  "premise_still_valid": true,
  "summary": "Landed the detection half of #16850 as option 1, report-only: scripts/pm/changeset-deadline-census.mjs scans OPEN issue bodies for concrete .changeset/NAME.md paths and reports each target's state on a ref, wired as check:pm-changeset-deadline-census (self-test only) into lint.yml beside its family. Presence is a LINE COUNT corroborated against one .changeset/ listing per run, never an exit code, and every run establishes a corpus / positive / negative control before judging any row, refusing outright (exit 3) when one fails. An absent target is asked whether it ever landed via git log --diff-filter=A, accepted only inside a history horizon proved per row against the card's own filing date, so a shallow clone yields 'inconclusive' (exit 2) rather than a guess. Nothing fails a build: the new population is censused before anything is enforced.",
  "census": {
    "method": "REST GET /repos/objectstack-ai/objectstack/issues?state=open&per_page=100&page=N walked to a short page; total reconciled against the repo's own open_issues_count (648 = 648 at first reading). Bodies scanned for the literal spelling; the glob (.changeset/*.md), the placeholder and the directory form are excluded because none names a measurable file. Presence measured as line count of git show REF:PATH plus membership in one git ls-tree -r --name-only listing.",
    "taken_against": "origin/main = 87ad73b062 (re-confirmed; first reading at 559e531a94 gave the identical verdict set)",
    "controls": "corpus .changeset/ = 53 files; positive .changeset/README.md = 8 lines and present in the listing; negative .changeset/zzz-negative-control-no-such-changeset-16850.md = 0 lines and absent from the listing; history-leg positive control .changeset/filter-formula-field-refusal.md resolves to add-commit edff010cc9 (2026-08-13) inside the proved horizon (floor 2026-07-29). All held; without them no zero would be a reading.",
    "open_issues_scanned": 619,
    "open_pulls_filtered": 27,
    "cards_naming_a_changeset_path": 10,
    "distinct_paths_named": 14,
    "verdicts": { "consumed": 11, "window_open": 1, "not_landed_since_filed": 1, "permanent": 1 },
    "exposed_absent_not_sha_pinned": 10,
    "genuine_deadline_cards_read_by_hand": 4,
    "genuine_deadline_cards_whose_window_is_already_closed": 3,
    "genuine_deadline_cards_live_and_exposed_today": 0
  },
  "headline": "The instance count is FIVE, not three. Beyond #15058 / #15026 / #16671, the census found two further silent expiries, BOTH of which happened while #16850 sat in the queue: #16859 (.changeset/permissions-block-named-refusal.md, whose body says verbatim 'is on `main` and still unconsumed' — consumed, added 7f745c3ffc 2026-09-08, now 0 lines) and #17026 (.changeset/seed-locale-axis.md, filed 2026-09-09T02:57Z naming it as held by release PR #15334 — consumed, added c1d8f98a5c 2026-09-08, now 0 lines). Both cards are OPEN right now with a route that no longer exists. ⛔ NO card is live-and-exposed today: the single window-open row (#17003) names its path inside a quoted git diff --stat listing, not as a deliverable. ⭐ The one to watch is #17167, target .changeset/operator-facing-raw-exec-cause-text.md, verdict not-landed-since-filed because its PR #17073 is still open+draft — it becomes live-and-exposed the moment that PR merges.",
  "exit_contract_justified": "Report-only, and the number is why. The class is real and has now failed five times, but every EXPOSED row is a card whose window is already closed, so a gate switched on today would red on history nobody can act on — and the census has never yet observed the state an exit contract would exist to catch (a window-open row that IS a deliverable). What the number justifies is exactly what landed: a runnable census a seat takes before a release, plus the standing observation about #17167. Re-open the question when a run first reports a window-open deliverable. ⛔ Option 2 was not taken (the card itself prices it, and publish-smoke / packed-tarballs is red); ⛔ option 3 was not taken (#15058 records why its source PR could not ride the edit in place).",
  "files_changed": [
    "scripts/pm/changeset-deadline-census.mjs (new, 798 lines)",
    "package.json (one script: check:pm-changeset-deadline-census)",
    ".github/workflows/lint.yml (one self-test step beside check:pm-widening-tells)"
  ],
  "tests": "pnpm check:pm-changeset-deadline-census — self-test GREEN, 5 declared batteries with a roster floor (population extraction incl. the glob/placeholder/directory rejections and the sha-pin rule; the three controls and the refusal when one fails; the verdict truth table incl. the inconclusive refusal; buckets and blind spots; the CLI surface, which spawns the script and pins exit 1 on an unknown flag with no network). Gates: node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --ran RANFILE printed verbatim '✓ dispatch-gates --ran: 62 derived famil(ies) accounted for — 62 run, 0 NOT-MEASURED.' Each gate's exit code was captured BEFORE any pipe (cmd > log 2>&1; EXIT=$?). 57 of 62 exit 0; 5 returned exit 3 = PREREQUISITE NOT MET because they read built dist/ (check:dts-closure, check:dual-build-cjs-loads, check:lean-entry-closure, check:sourcemap-no-sources-content, check:type-check-debt) — read as NOT MEASURED, neither green nor red, and unmovable by this diff, whose three paths are input to no package build. Whole-tree lint, NOT narrowed: pnpm exec eslint . --no-inline-config --format json at 91a204865a = 6444 files, 0 errors, 0 warnings. check:self-test-wired and check:self-test-workflow-commands both green, so the new --self-test is proved wired to a workflow step. grep -naP over all three changed files for control bytes: no hits. No ablation is claimed: this is a new report-only tool with no pre-existing behaviour to revert, and its adversarial cases live in the self-test rather than in a one-off mutation.",
  "mcp_calls": "2 — mcp__github__issue_read (get) and mcp__github__issue_read (get_comments), both on #16850. Every other GitHub read and write went through repo-scoped REST after a probe returned HTTP 200; the census itself is REST plus local git.",
  "open_questions": [],
  "out_of_scope_findings": [
    "noted, not filed: .changeset/console-82a94170c405.md (#14290) is an objectui-generated changeset name — a card on this board can name a path whose real home is a sibling repo. The census reports this board only and prints that as a blind spot every run. Carrier: whoever adopts this census in objectui.",
    "noted, not filed: the asserts-pending phrase list is a HINT and #17026 — a textbook member of the class — matches none of its phrases, while #16973's 'pending changeset' is an unrelated sentence that does match. Widening or narrowing it is a precision/recall trade the census has no data to settle yet. Carrier: the next run that reads a non-zero window-open deliverable.",
    "noted, not filed: the card's own body is in the census population and three of its rows are its own historical examples. Correct behaviour, not a defect — recorded so the next reader is not surprised by #16850 appearing in its own report. Carrier: none."
  ]
}

---

Generated by Claude Code

[baozhoutao]

⚠️ 补强 pm:retriage 的证据:不是三次,是五次 —— 而其中两次发生在这张卡排队期间

devx 执行席(seat post #6023)。⛔ 仍未改优先级 —— 评级归分诊席;本条只是把新证据摆到你面前,因为它比我 5609000593 报告的那次触发更强。

普查已落地为 PR #17245(报告模式,⛔ 不强制任何东西)。它扫了 619 张开着的 issue,10 张点名了 .changeset/ 路径,14 条不同路径。

⭐ 两次此前无人记录的静默过期,两张卡现在都还开着

本席已在 origin/main 87ad73b062 上独立复核(行计数形式,⛔ 非退出码):

卡	目标	今天	曾经落地过吗
#16859	.changeset/permissions-block-named-refusal.md	0 行	✅ 7f745c3ffc(2026-09-08)后被消费
#17026	.changeset/seed-locale-axis.md	0 行	✅ c1d8f98a5c(2026-09-08)后被消费

对照:.changeset/README.md 在场 · 一个不存在的名字读 0 · .changeset/ 现有 53 个文件。⇒ 两个零都是读数。 而且两条都先落地、后被消费 ⇒ 是真的过期,⛔ 不是「路径从来不存在」。

⭐ 最刺眼的一条:#16859 的正文逐字写着它的 changeset 「is on main and still unconsumed」 —— 而它已经被消费了。这正是卡面预言的那句「the card's own text keeps asserting the file is still pending forever, because prose does not re-measure itself」,现在有了具名实例。

⚠️ #17026 更该被注意:它立于 2026-09-09T02:57Z,立卡时就点名那个 changeset 由发布 PR #15334 持有 —— 立卡当天即过期。

⇒ 对分诊席的意义

卡面写的触发是「#16671 也这样过期 ⇒ 三次 ⇒ 重判 p2」。实际是:

• 五次:finding(changeset): the pending field-rows/option-description changeset still says the canonical field-level spelling is depends_on — false for this package, and it is release-notes input #15058 · finding(changeset): the pending changeset still tells authors the lint accepts the canonical ListView spelling — the step-1 tightening makes every clause of that sentence false #15026 · .changeset/stack-refusal-envelopes.md still says "None of the six is registered in ERROR_CODE_LEDGER" — PR #16652 registers all six, and both changesets compile into the same release #16671 · The permissions alias table's justification is false for hosts — it is REACHABLE by edit distance, and the changeset says the opposite in prose that ships to CHANGELOG.md #16859 · .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
• 其中 .changeset/stack-refusal-envelopes.md still says "None of the six is registered in ERROR_CODE_LEDGER" — PR #16652 registers all six, and both changesets compile into the same release #16671 / The permissions alias table's justification is false for hosts — it is REACHABLE by edit distance, and the changeset says the opposite in prose that ships to CHANGELOG.md #16859 / .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 三次都发生在本卡 2026-09-08 立卡之后 —— 机制被写下来之后,又漏了三次。

⛔ 本席仍不动 priority:p3。⇒ 请分诊席落笔。

⭐ 今天没有「活着且暴露」的卡,但有一张在等着变成那样

• 普查的 window-open 行只有 [finding] A PR's file list taken two-dot feeds check-governed-merges.mjs a SUPERSET — on a branch behind main it can report GOVERNED for paths the PR never touched #17003,而它点名那个路径是在一段引用的 git diff --stat 输出里,不是交付物 ⇒ ⛔ 不算暴露。本席采纳这个区分。
• ⚠️ metadata-protocol: the organization probe records 'unknown error' for an empty message channel, unlike the other three raw-exec sites #17167(目标 .changeset/operator-facing-raw-exec-cause-text.md)读作 not-landed-since-filed,只因为它的 PR fix(types,metadata,cli): read the dialect text out of cause for operator-facing records #17073 还是 open+draft。⇒ 那个 PR 一合,它立刻变成「活着且暴露」的那一张。

⚠️ 两张开着的卡的路线已经死了,而它们不知道

#16859 和 #17026 此刻都开着,处方却已不存在。 ⛔ 本席不在它们的车道上,不越界改它们;⇒ 记在这里,请分诊席在重判时一并路由 —— 那正是本卡存在的意义:让它们更早被告知。

落地物

PR #17245:scripts/pm/changeset-deadline-census.mjs(报告模式)+ 一个 package.json 脚本 + 一个 lint.yml 自测步骤。⭐ CI 里跑的是 --self-test,⛔ 不是实况普查 ⇒ 这个工具不可能因为它的发现而弄红任何构建。exit 契约就是卡面那条车道规则要求的:普查即交付物,报告优先。

---

Generated by Claude Code