check:durability-log-level cannot represent a seam that propagates a durability failure by RETURNING an outcome object — its propagation lists are keyed on callee names

The gap

scripts/check-durability-degradation-log-level.mjs cannot represent a durability seam whose catch propagates by returning an outcome object, and this is a limitation of the instrument's shape rather than a missing entry in its lists.

The seam that exposed it: recordNotificationEventReceipt in
packages/metadata/src/migrations/migrate-sys-notification-to-event.ts (added by PR #16230).
It is a durability seam of exactly the persistSeedTenancyReceiptRow class — the data really was
rewritten, every other reading is clean, and the only durable record of it can be absent.

Why the gate cannot hold it:

DURABILITY_CRITICAL_CALLEES does not name it — that part is merely an entry.
⛔ But it cannot be added without a second change: this catch propagates by RETURNING an
outcome object, while both FAILURE_PROPAGATION_CALLEES and FAILURE_PROPAGATION_SITES are
keyed on callee NAMES. A returned object literal has none.

⇒ Any seam in this repo that reports a durability failure to its caller through a return value,
rather than by calling a named reporter, is invisible to this gate by construction.

What was measured, and what was not

Measured (by the metadata: the adr-0030-notification-event migration writes its sys_migration receipt row per the ruled ledger-claim matrix — last_run_at on every completed run, applied_at on migrated only, never verified_at (#15710 follow-on) #16100 round, on its own diff): the gate's write-side limb reports 29
seams both with and without the new file — i.e. adding a real return-propagating durability
seam moves that population by zero.
Measured, and it is the control that makes the zero above a reading: the gate's read-side
limb moves 68 → 67 when the same file is swapped back to origin/main. ⇒ the file is in the
swept population; the write-side silence is the instrument's blindness, not the file's absence.
⚠️ NOT a live defect. The seam is pinned per-file by that round's own failed receipt
case, so this is an instrument-expressiveness gap, ⛔ not an unprotected degradation. Hence
finding, not bug.
⚠️ NOT measured: how many other seams in the tree propagate this way. That census is the first
thing a round on this card should buy, because it decides whether the fix is worth its cost — a
gap with one member is a note, a gap with thirty is a hole.

Why the shape of the fix is not obvious

A name-keyed list cannot key on "returns an object with a failure-shaped field". The candidate
directions each cost something and none is free:

Key on the enclosing function's name instead of the callee's, for a declared set of seams —
cheap, but it moves the gate from "what does this call do" to "who is asking", which is a
different question and may weaken the existing 29.
Recognise a declared return shape (an outcome type carrying a failure member) — precise, but
it needs the gate to read types, which it currently does not.
Leave it name-keyed and accept the class as out of scope, recorded — honest, and it makes this
card's answer a documented boundary rather than a hole.

⛔ This card proposes none of them. It records the gap, its measurement, and its control.

⚠️ Context a maintainer may want more than this row

This is now the third open finding against this one script, alongside a decision and a blocked
triage card:

#	state	what it is about
#8897	`pm:on-hold`	`collectLoggedLevels` only records part of what it should
#12576	`pm:on-hold`	`readInventionKey` falls back where it should not
this card	new	return-propagating seams are unrepresentable
#9882	`pm:on-hold`	decide the shrink-only delta-ratchet for the same gate
#12753	`pm:blocked`	triage the parameterless-catch read seams, then decide the 2a criterion

⇒ Five open items on one instrument. Whether that is five small repairs or one rewrite is a
question this card does not answer and probably should not be decided one row at a time.

Provenance and dedup

Found by the #16100 execution round, which could not file it itself and said so rather than
guessing: filing without a duplicate check is a forbidden shape, and its dedup channel was gone in
both directions — REST /search/issues is not served on that channel (a same-session control query
that must have hit #16100 came back empty, so its zero was a channel artefact, not a reading),
and both search_issues MCP calls were refused by the user-level rate limit. It handed the row to
the PM seat instead. Its full reasoning:
#16100 (comment)

Dedup performed by this seat, over a complete enumeration rather than a search: all 652 open
issues fetched with bodies (651 of 652 carry one) and grepped locally.
FAILURE_PROPAGATION_CALLEES, FAILURE_PROPAGATION_SITES, persistSeedTenancyReceiptRow and
recordNotificationEventReceipt return 0 hits each; DURABILITY_CRITICAL_CALLEES returns 2
(#15166, #8897) and check-durability-degradation-log-level returns 7 — all read, none covering
this shape. Firing controls on the same command and scope: sys_migration → 4 (#16194, #16185,
#16100, #15207), NOTIFICATION_EVENT_MIGRATION_ID → 2, durability → 12. ⇒ the zeros are readings.

Filed unassigned and ungraded — domain:* and priority are triage's.

os-decision-facets

决策箱 —— 普查买回来了,数是 12,⇒ 路线选择越过了本卡口径,交维护者

domain:devx 执行席(座位贴 #6023,session session_017ef78bLdybu3AffehKkhfk,round 12)。⛔ 本席不代裁。普查与边界记录已随 PR #18523 落地(Part of,⛔ 不关本卡)。

先把「为什么现在必须有人裁」说清:⭐ 不是仪器缺一条腿,是治理文件自己写出了一条死路

AGENTS.md:938-943 同时规定三件事(本席在 origin/main 上逐字读的,⛔ 不是转述报告):

「a failure handed to the CALLER is not a degradation at all —— 第三个合法答案」,并明文把「a batch whose contract IS a per-item outcome report」算进去;
「Do not bolt a logger.error onto such a site」;
「declare how it delivers instead —— FAILURE_PROPAGATION_CALLEES 或 function-scoped 的 FAILURE_PROPAGATION_SITES」。

⚠️「function-scoped」这个词会让人以为第三条罩得住 return 式交付。本席去读了源码(scripts/check-durability-degradation-log-level.mjs:505、:513):SITES 的键确实是 FILE::FUNCTION,但每个条目还必须给出 callees: [[name, kind]],注释把话说死 ——「it supplies a name, and catchDeliversFailure() still has to prove every path out of the catch reaches it」;现存两个条目(migrateStoredMetadata 的 record()、duplicatePackage 的 failed.push())都是靠一次调用交付的。⇒ return { ok: false, error } 的 catch 里没有任何调用可供声明。

⇒ 在这 12 个成员中的任何一个上执行 AGENTS.md 的第 1 条(「发现新缝就在同一个 PR 里把它加进 DURABILITY_CRITICAL_CALLEES」),第 2、3 条同时不可用:正确的代码会把闸弄红,而唯二能弄绿的办法(挂 logger.error / 进 baseline)都是这份闸 header 自己拒绝过的。今天不触发,只因为 tier-1 是 0。读数时刻 2026-09-16T21:18Z。

选项 × 真实代价

选项	做什么	客户/开发者可感知的后果	实测
A 改键到外层函数名	把授权粒度从「这次调用干了什么」挪成「谁在问」	一个函数级许可会连同该函数里未来新增的吞掉写失败的 catch 一起放行	⚠️ 实测有风险:现存 4 条靠 `FAILURE_PROPAGATION_CALLEES` 表达的缝会被函数级许可覆盖,今天的绿会变松
B 认声明的返回类型	闸读返回类型,把「交付」判成一个结构而不是一个名字	新出现的同形缝自动可见,⛔ 不必有人记得去加一行	信号确实存在:12 个成员里 8 个返回具名 interface/type;代价是给一个刻意无类型的闸引入 program 级类型解析
C 记为范围外	把 PR #18523 落下的那段边界注记改写成「这类缝不归本闸管」	8 个 tier-3 成员成为永久盲区(补偿事务 unwind、`CREATE INDEX` 探测、`claimSuspension` 消费)	⛔ 且它与上面那条死路直接冲突 —— 要么同时改 `AGENTS.md`(治理面,维护者独有),要么留一条写在书上却做不到的规则
D 先买一次设计,把「交付渠道」这条轴与 #8897 合并	不动判据,先把「闸的词汇表是名字,而现实里的交付不总是名字」这条轴一次想清	短期什么都不变;⚠️ 第五个条目继续开着	本卡 12 个成员与 #8897 的 `port.warn` 接收者是同一条轴的两半

业务含义直译

A = 「谁签的字我就信谁,不看他签了什么」—— 像把审批权按部门授予而不是按事项授予。
B = 「把交付写进类型,机器自己认」—— 像把发票格式标准化,而不是维护一张「这些人开的票有效」的名单。
C = 「这类账我们不查了,写进制度」—— 诚实,但制度里另一条写着这类账必须查。
D = 「先开一次会把口径定下来,再改流程」。

四棱

① 项目长远合理性:B 缩小特例与契约增生 —— 名字词汇表是「每条缝一行、永久维护」的增生机制(今天 21+33+2+2 个名字),而返回形状是一条结构判据,新缝自动进。A 扩大(每个函数级许可都是一条永久特例,且它能覆盖未来新写的 catch)。C 把增生冻结成盲区,并留下一条书面上做不到的规则。
② 实际业务拉动:今天没有人撞上 —— tier-1 是 0,闸当前看见的 36 条缝全部靠日志或已声明的被调者作答。⇒ 按「零拉动默认 defer」,这件事不急;⛔ 但它不改字母,只改时序。
③ 防 AI 犯错:出错时谁看到什么?B 是响亮拒绝(闭合的结构判据:交付得到证明,或者闸红)。A 是静默泄露(函数级许可罩住它不该罩的 catch,没有人会看到)。C 是静默泄露的制度化(那 8 条缝以后连问都不会被问)。
④ 创业阶段不扩散:「每个已声明的键都是永久义务」—— A 增加永久义务(多一张按函数的表)。B 减少(判据一次写成,⛔ 不必逐条声明)。C 不增不减,但把义务转成了盲区。

Prior rulings read: failure_propagation_callees,failure_propagation_sites,return-propagating,outcome object,enclosing function name,durability → 0 hits; none

⚠️ 上行按 check-prior-rulings.mjs 的口径写:docs/adr 0/463,⇒ 无 ADR 裁定。⛔ 但 AGENTS.md 有 10 处命中,其中 :938-943 就是上面那条死路的出处 —— 它不是 ADR 裁定,是规则正文本身,本席按「不是裁定所以不写进 hits」的字面口径记 0,同时把它整段抄在上面,⛔ 不让它被这个 0 藏掉。

推荐:B(读声明的返回类型),时序按 D 走(与 #8897 合并成一次设计再动手),回退项 C。
自检行:「只看①选 B;②③④ 是否翻转:否 —— ② 零拉动只把它推后(合并进 D 那次设计),③④ 都指向同一个字母。」
置信缺口:⛔ 本席没有量 B 的实现成本(给这个闸引入 program 级类型解析要多久、会不会拖慢 CI),也没有读 #12576 —— 它可能在同一条轴上,若是,D 的设计面还要再大一圈。⚠️ 另:dev 席的建议是「④ 其次 ②」,⛔ 本席不沿用 —— D 不是一个方向,它是 B 该怎么设计的范围;字母必须落在一个真会改变仪器的选项上。

裁定之后怎么执行(维护者只裁方向,⛔ 不背执行)

裁 B ⇒ 本席开一张实现卡,口径:判据写死为「声明的返回类型里带一个失败成员」,阴性对照必测现存 36 条缝的绿一条不少;⛔ 不动 ACCEPTED_* 台账。
裁 C ⇒ 本席把 PR measure the return-propagating durability seams the log-level gate cannot name #18523 落下的第 4 条 limitation 从「已测量、路线未定」改写成「范围外」,并把 AGENTS.md:938-943 的冲突作为治理面改动单列一张卡交还维护者(⛔ 本席不改治理面)。
裁 A ⇒ 本席开卡,但验收第一条是「证明现存 4 条 FAILURE_PROPAGATION_CALLEES 缝不被函数级许可放松」,⛔ 证不出就退回。
裁 D(只买设计)⇒ 本卡转 pm:on-hold + Restart-when:,并把 check-durability-degradation-log-level: collectLoggedLevels only recognises a logger named logger/log/console, so a catch that reports through an injected logger reads as silent to BOTH rules #8897 一并挂上。

Generated by Claude Code

measure the return-propagating durability seams the log-level gate ca…

Ruling: batch #149 item 4 · letter C, plus a correction of AGENTS.md:938-943 (return-propagating seams are outside this gate's scope and stay pinned by per-file tests; the governance text is corrected to say so; the four sibling improvement cards on the same gate close not_planned) · maintainer 「16233 也帮我重新分析」, then 「同意」 on the re-asked form 2026-09-17T14:41Z

Director seat, summon #24, session_01Wj1HUjzyeiBQ8atRf1ZhaL. Presented in detail, then re-analysed under the maintainer's principle record (5716140776) — the AI-error axis is about platform users' AI, ⛔ not our fleet's CI; the maintainer agreed to C + the AGENTS.md correction, with the four cards on the same gate closed together.

Facts (this card; dev landing record 5704218834 and decision box 5704219265; origin/main read by the seat): check-durability-degradation-log-level guards the invariant 「the data really was rewritten and the only durable record of it can be absent ⇒ report it loudly」. Its vocabulary is callee names (FAILURE_PROPAGATION_CALLEES / _SITES); a catch that delivers the failure to its caller by return { ok: false, error } makes no call, so the gate cannot see it by construction. PR #18523 bought the census: 12 such seams (tier-1 — visible to the gate today — 0; tier-2 4; tier-3 8), every one pinned by its own file's test, ⛔ none an unprotected degradation. AGENTS.md:938-943 states a rule nobody can follow on those 12: a failure handed to the caller is 「the third legal answer」, 「do not bolt a logger.error onto such a site」, 「declare how it delivers — CALLEES or the function-scoped SITES」 — and every SITES entry must still supply a callee name (scripts/check-durability-degradation-log-level.mjs:505, :513), so the correct code reds the gate and the only two ways to green are the two the gate's own header refuses. It does not fire today only because tier-1 is 0. Fifth open card on one script.

Ruling — C + correction

The gate does not change. Return-propagating seams are outside its scope, by construction; the gate header's fourth limitation (landed by PR measure the return-propagating durability seams the log-level gate cannot name #18523) already records that boundary and points at scripts/measure-return-propagating-durability-seams.mjs as the census. ⛔ No re-keying (A), ⛔ no type-aware detection (B), ⛔ no design purchase (D) — all three are capability for an instrument, zero for a platform user.
AGENTS.md is corrected to match the instrument — one paragraph, in the block that currently reads lines 938-943: a catch that delivers by returning an outcome object has no callee to declare and is ⛔ not added to DURABILITY_CRITICAL_CALLEES / FAILURE_PROPAGATION_*; such a seam is pinned by its file's own test (the failed-receipt case is the reference shape), and the census script is where its population is read. The invariant itself (silent data loss must be loud) does not move. AGENTS.md is a governed surface ⇒ human merge by the approver; Clause-②: no; skip-changeset.
The four sibling cards close not_planned with a pointer to this ruling: check-durability-degradation-log-level: collectLoggedLevels only recognises a logger named logger/log/console, so a catch that reports through an injected logger reads as silent to BOTH rules #8897 (collectLoggedLevels partial), [finding] check:durability-log-level's readInventionKey falls back to one shared "anonymous" placeholder, so every anonymous-callback read seam in a file collapses onto a single baseline key #12576 (readInventionKey placeholder), Decide: shrink-only delta-ratchet for check-durability-degradation-log-level — measured at 0 net decreases/month over 3,195 commits #9882 (shrink-only delta-ratchet decision); Triage the parameterless-catch read seams, then decide the 2a criterion — the one surviving half of #9165, on numbers that need re-measuring first #12753 is already closed. The gate gets no further investment; if a real incident — silent data loss reaching a user — is ever traced to one of these gaps, that incident is the card.

Four-facet reading: ① zero — tier-1 is 0 and all 12 are pinned; ② the invariant stays, the instrument stays narrow, no rewrite; ③ zero under the corrected axis for the gate itself — the one real defect is a governance sentence that sends the next developer into a dead end, and that is a text fix; ④ zero mechanism, one paragraph.

Execution

needs-user-decision → pm:queue; domain:devx, priority:p3, tooling stand. The claiming seat edits only the AGENTS.md paragraph (and, if the header of the gate disagrees with the new text, the header sentence — nothing else in the script). Siblings: pm:on-hold removed, closed not_planned, one pointer comment each.

Generated by Claude Code

Claim: PM loop round 80
Session: session_017ef78bLdybu3AffehKkhfk
Branch: claude/issue-16233-agents-return-propagating-seams
Worktree: objectstack-issue-16233
Domain: domain:devx
Seat: domain:devx
File surface: AGENTS.md（仅 938-947 那一段；若门禁头部与新文字不一致，另加 scripts/check-durability-degradation-log-level.mjs 的那一句头部注释，⛔ 脚本其余一行不动）(stop on breach; explain in the report)
Container & model: S 级机械卡, mode:subagent, model: opus（引当次 --tier AGENTS.md 输出：「no path-derived mandate … floor sonnet · default opus · ceiling fable」⇒ 无路径强制，本席按卡判 default）
Clause-②: no（裁决原文即如此判）
Thread-read: 5716260390
Serial constraints cleared: none（本卡裁决点名的四张兄弟卡 #8897 / #12576 / #9882 / #12753 现读全部已 closed not_planned ⇒ 裁决的兄弟卡那一半已经做完，本卡只剩文字那一半；现读 17 个 open PR：触及 AGENTS.md 的 0 个）

⚠️ 本席必须先说清楚：这张卡在本席的取卡序里被漏了七十多轮，而漏的原因不是它

domain:devx 执行席，2026-09-19T23:31Z（本段及以下所有读数与本贴同一次动作；探针 worktree 停在 origin/main 24d622b94，读于 2026-09-19T23:31Z）。

本卡 priority:p3、立于 2026-09-06，是本席队列里最老的可派发卡。本席本轮才第一次按标签做全量枚举，此前一直在自己维护的清单上排序 —— 缺陷在取数，⛔ 不在本卡。已在轮报 5746106710 公开记为本席 self-error #35。

本席现读：裁决的哪一半已经做完，哪一半还欠着

维护者裁决在 5716260390（2026-09-17T14:41Z，批 #149 项 4，letter C + AGENTS.md 更正）。它派了两件事：

四张兄弟卡 closed not_planned —— 现读 check-durability-degradation-log-level: collectLoggedLevels only recognises a logger named logger/log/console, so a catch that reports through an injected logger reads as silent to BOTH rules #8897 / [finding] check:durability-log-level's readInventionKey falls back to one shared "anonymous" placeholder, so every anonymous-callback read seam in a file collapses onto a single baseline key #12576 / Decide: shrink-only delta-ratchet for check-durability-degradation-log-level — measured at 0 net decreases/month over 3,195 commits #9882 / Triage the parameterless-catch read seams, then decide the 2a criterion — the one surviving half of #9165, on numbers that need re-measuring first #12753 四张全部 closed not_planned ⇒ ✅ 已完成，⛔ 本卡不要再动它们。
AGENTS.md 那一段更正 —— ⛔ 未做。这是本卡的全部余量。

本席自己把裁决的事实基础重跑了一遍（⛔ 不是转述裁决）

普查脚本在树上：scripts/measure-return-propagating-durability-seams.mjs。本席现跑，读数与裁决逐字吻合：扫 2620 个非测试源文件 / 2968 个 try-catch / 250 个守着已声明持久调用的 / 21 个每条路径都返回构造失败对象的 ⇒ MEMBERS 12 个，其中 [1] 门禁今天看得见的 = 0、[2] write-shaped = 4、[3] 按实质判定为持久的 = 8；另有 ADJACENT 5、EXCLUDED 4。
门禁头部的第 4 条 honest limitation（scripts/check-durability-degradation-log-level.mjs:79-96，PR measure the return-propagating durability seams the log-level gate cannot name #18523 落地）已经把这个边界写清楚了，并且点名那个普查脚本。⇒ ⭐ 门禁这一侧不欠东西；欠的是治理文本。
AGENTS.md:938-947 现读仍是旧文字，逐字要求：「declare how it delivers instead — FAILURE_PROPAGATION_CALLEES（repo-wide names）or the function-scoped FAILURE_PROPAGATION_SITES」。⇒ 对那 12 个 seam，这条规则没人能照做（返回对象没有 callee 名可填），而门禁的两条 SITES 入口都仍要求一个 callee 名。裁决说的「a governance sentence that sends the next developer into a dead end」本席复核成立。

⚠️ 受管面：本卡交的是草稿 PR，落地不是本席的动作

AGENTS.md 是 Tier H 受管面（SKILL.md:605-606、:625）。据 SKILL.md:26-27：⛔ 本席永不对受管面 PR 提交批准 review、永不合并；⛔ 无授权批准时永不翻 ready、不入队、不挂 auto-merge —— 批准在案后席位落地。SKILL.md:613 另定：受管面 PR 的正文带 ## 维护者速读(草稿)，中文、业务角度、席位意见留空。

⇒ PR 保持 draft 交付，进维护者的受管草稿批次。⚠️ 代价要说在前面：本席另有一张只能人工合并的 PR #18414 已挂约 78 小时无人动 —— 本卡的 PR 大概率也要等。裁决点名了本席执行，所以本席照做，⛔ 但不假装它会自动落地。

给 dev 的边界

只改 AGENTS.md 现在 938-947 那一段（⚠️ 行号取自 24d622b94，读于 2026-09-19T23:31Z，你要在自己的 worktree 重新定位）。新文字要说清四件事，全部来自裁决原文：以返回结果对象交付失败的 catch 没有 callee 可声明；⛔ 不加进 DURABILITY_CRITICAL_CALLEES / FAILURE_PROPAGATION_*；这类 seam 由它自己文件的测试钉住（failed-receipt 那个 case 是参考形状）；它的population 在 scripts/measure-return-propagating-durability-seams.mjs 里读。⭐ 不变量本身（静默数据丢失必须吵）⛔ 不动。
只有在门禁头部与你写的新文字真的不一致时，才改 scripts/check-durability-degradation-log-level.mjs 的那一句头部注释。⚠️ 本席现读认为 :79-96 的第 4 条已经一致 ⇒ 预期不需要改脚本；若你判断需要，先在报告里说明哪一句不一致。⛔ 脚本的其余任何一行、任何词表、任何 baseline 都不许动。
⛔ 不碰那四张兄弟卡（已 closed）、⛔ 不碰 CLAUDE.md、⛔ 不碰 docs/adr/**、⛔ 不碰 content/docs/releases/**。
skip-changeset（裁决原文即如此）。PR 保持 draft，正文带 ## 维护者速读(草稿)：中文、业务角度、席位意见留空。⛔ 不要翻 ready、⛔ 不要入队、⛔ 不要挂 auto-merge。
跑 node scripts/pm/dispatch-gates.mjs --commands AGENTS.md 报出的门禁族，并在报告里贴 pnpm check:durability-log-level 的结果 —— 它是本段文字所治理的那个仪器。

Generated by Claude Code

os-dev-report

{
  "issue": 16233,
  "status": "blocked",
  "branch": "claude/issue-16233-agents-return-propagating-seams",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/19214",
  "premise_still_valid": true,
  "summary": "The ruling's remaining half is implemented: the AGENTS.md paragraph (origin/main 24d622b94 lines 934-943) now splits the declaration rule by how the failure leaves the catch — a call-delivered seam names its callee in FAILURE_PROPAGATION_CALLEES / _SITES as before, while a catch that delivers by RETURNING an outcome object has no callee to declare, is added to neither map nor to any baseline, is never given a bolted-on logger.error, is pinned by its own file's test asserting the returned failure outcome (the `failed`-receipt case as reference shape), and has its population read from scripts/measure-return-propagating-durability-seams.mjs. The invariant is untouched and the gate script is untouched. BLOCKED on one thing only a maintainer can rule: AGENTS.md sits at its line ceiling on main (1099, headroom 0) and this paragraph costs +10 lines (1109); raising a ceiling is maintainer-only under the #8435 remedy-authority convention, this card's ruling states no line budget, and the dev seat did not touch scripts/pm/check-skill-line-ratchet.mjs. Everything else is green; the PR is a draft with the 维护者速读(草稿) section and skip-changeset.",
  "tests": "Premise re-measured on the branch: `node scripts/measure-return-propagating-durability-seams.mjs --sites` :: exit 0 — 2620 non-test source files, 2968 try/catch, 250 guarding a declared durable call, 21 returning a constructed failure object on every path ⇒ MEMBERS 12 ([1] gate-visible 0, [2] write-shaped 4, [3] durable on the merits 8), ADJACENT 5, EXCLUDED 4 — reproduces the ruling's numbers exactly. The governed instrument: `pnpm check:durability-log-level` :: exit 0 — '36 durability-critical catch seam(s), all loud, rethrowing or propagating to the caller (4 propagating, declared)' and '68 read seam(s), none invents an unreported answer'; unchanged by this diff. `node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack` derived 14 families; `--ran` reconciliation: '14 derived famil(ies) accounted for — 14 run, 0 NOT-MEASURED'. 13 exit 0: check-closing-keyword-parity (and --self-test), check-comment-mask-corpus, check:agent-test-spelling, check:docs-audit-scope, check:driver-memory-census, check:nul-bytes, check:pm-governed-merges, check:pm-governed-prose, check:pm-skill-id-lint, check:refd-timer-probe, check:required-contexts, check:watch-hint-literal. 1 RED: `pnpm check:pm-skill-ratchet` :: exit 1 — '✗ check-skill-line-ratchet: AGENTS.md is 1109 lines; the ratchet ceiling is 1099. Raising a ceiling requires a maintainer ruling quoted in the PR.' Rewrap headroom measured, not assumed: the paragraph as it stands on main is 10 lines at 86, 88 and 90 bytes of prose width (0 lossless rewrap headroom); the rewritten paragraph is 21 / 20 / 20 at the same widths and landed at the surrounding block's 88-byte width ⇒ the +10 is bought by content, never by re-wrapping. Exit codes captured before any pipe (redirect-then-$?). No build or package test is owed: the diff touches no package, so the dependency closure is empty and no package's tests or typecheck are affected; no gate or tool script was edited. Ablation: not applicable — this is a prose change with no executable claim of its own.",
  "mcp_calls": "0 — no MCP GitHub tool was called at any point",
  "api_writes": "3 REST proxy writes — POST /repos/objectstack-ai/objectstack/pulls (draft PR 19214), POST /repos/objectstack-ai/objectstack/issues/19214/labels (skip-changeset; read back, present), POST /repos/objectstack-ai/objectstack/issues/16233/comments (this report). Plus git pushes on the feature branch only (empty-branch routing probe, the commit, and one --force-with-lease after amending the trailer pair; all five AGENTS.md §3 criteria held).",
  "open_questions": [
    {
      "question": "AGENTS.md is at its line ceiling (1099, headroom 0) and this ruled correction costs +10 lines (1109). The ruling directs the text change but states no line budget, and check-skill-line-ratchet requires a maintainer ruling quoted in the PR before a ceiling moves. How is the budget ruled?",
      "options": [
        "A — maintainer rules a +10 budget; the seat lands the ceiling raise 1099 → 1109 in this PR with the ruled-raise comment (the precedent shape used for 1058→1068, 1068→1074, 1075→1099), quoting the ruling verbatim and recording the measured 0 rewrap headroom",
        "B — maintainer rules a smaller budget and names what to drop; the candidates in cost order are the closing ⭐ invariant restatement (~2 lines, the invariant is already stated three times in this section), the `failed`-receipt reference-shape parenthesis (~1 line), and the `return { ok: false, error }` example (~1 line) — each drop removes something the ruling explicitly named, which is why the seat dropped none on its own",
        "C — maintainer rules the correction is not worth any line budget; then the dead end stays on the books and this card closes not_planned instead"
      ],
      "recommendation": "A, because the dead end is exactly what the ruling ordered removed and the three droppable pieces are all named in the ruling text itself; the +10 is measured, has 0 lossless rewrap headroom at three widths, and matches the size of prior ruled raises on this file."
    }
  ],
  "out_of_scope_findings": [
    "to file (class b, violates the declared contract this PR lands; dedupe words: FAILURE_PROPAGATION_CALLEES, remedy text, return-propagating, durability-log-level, outcome object) — the gate's own REMEDY STRING, scripts/check-durability-degradation-log-level.mjs:3882, is the same dead end in a second location: 'if this catch already HANDS THE FAILURE TO THE CALLER on every path ... declare how it delivers, in FAILURE_PROPAGATION_CALLEES or FAILURE_PROPAGATION_SITES in this script'. It is printed to the author of a FAILING seam, so the day any of the 12 census members becomes tier-1 visible (one vocabulary entry away), the gate tells that author to do the thing AGENTS.md now forbids. Latent today because tier-1 is 0. NOT fixed here: the dispatch boundary allows at most ONE header sentence of that script, and this is the remedy text, not the header.",
    "noted, not filed: scripts/check-durability-degradation-log-level.mjs:101-105 (inside the header's fourth limitation) still reads 'What to DO about it is deliberately not decided here — re-keying on the enclosing function, reading the declared return TYPE, and recording the shape as out of scope are all open'. After letter C those three are no longer all open — C was chosen. The operative half of limitation 4 (:79-99) agrees with the new AGENTS.md wording word for word (no entry in either map can name that seam; the census script is named), so the seat judged this NOT a genuine disagreement with the new text and left the script untouched, per the ruling's 'the gate does not change'. Carrier: the next PR that touches this script — and the ruling says the gate gets no further investment, so that may be none.",
    "noted, not filed: AGENTS.md's next paragraph (the 'It has teeth' block, main lines 951-954) still says 'a caller-propagating seam is declared via the propagation vocabulary, not baselined'. For a return-delivered seam only the second half holds — it is not baselined, but it is not declared via the vocabulary either. Outside the one-paragraph file surface this card grants, and the new paragraph immediately above states the exception, so the reader is not left wrong. Carrier: whoever gets the next ruled budget on this file."
  ]
}

_{⚠️ Attribution added by the dispatching PM after an H64 row: this os-dev-report was produced by the os-dev subagent dispatched for #16233 by the domain:devx execution seat, running inside session session_017ef78bLdybu3AffehKkhfk (the claim is comment 5746126007). ⛔ The GitHub author field (os-try-charles) is the TOKEN CLASS the write was made with, ⛔ not the seat that wrote it — attribution on this board is the session_ id the text carries, and this text carried none.}

Generated by Claude Code

⏳ `pm:dispatched` → `pm:awaiting-maintainer` —— 裁决点的那段文字写好了（草稿 PR #19214），但它要不到行

domain:devx 执行席，2026-09-19T23:47Z（本贴所有读数与本贴同一次动作；探针 worktree 停在 origin/main 24d622b94，读于 2026-09-19T23:47Z）。⛔ assignee 不摘：PR 在飞，本席跟到底。

Maintainer-action: 裁一个 AGENTS.md 的行预算（本次需要 +10 行：1099 → 1109），或改裁削减方案。 在此之前本卡在机制上落不了地。

本席复核过的阻塞（⛔ 不是采信 dev 的叙述）

读数	值
`AGENTS.md` 在 `origin/main` 的行数	1099
本仓行数棘轮对它的上限	1099 —— `scripts/pm/check-skill-line-ratchet.mjs:1438` `['AGENTS.md', 1099]`，旁注逐字写着「Landed count, headroom 0」
⇒ 余量	0
本次改写的成本	+10 行（dev 实测；并量了三个宽度下的重排余量均为 0 ⇒ 这 +10 是内容买的，⛔ 不是没排好版）
门禁读数	`pnpm check:pm-skill-ratchet` exit 1：「AGENTS.md is 1109 lines; the ratchet ceiling is 1099. Raising a ceiling requires a maintainer ruling quoted in the PR.」

⇒ 抬上限是维护者专属动作。⛔ 本席不抬、⛔ 不让 dev 抬、⛔ 也不靠删掉裁决点名的内容来凑行数。

⭐ 这不是本卡的偶然 —— 本席把它量成了一条普遍性质

本席在 origin/main 上跑了整个棘轮：32 个受上限文件中，31 个余量为 0；唯一的例外是 .claude/skills/dogfood-verification/SKILL.md（148/157，余量 9）。

⇒ 任何一条裁决，只要它要求往受管治理文件里写字，就按构造同时需要第二条裁决来批它的行 —— 因为棘轮把每个上限都钉在落地当时的行数上。⚠️ 本卡的裁决（5716260390）点了文字、⛔ 没点行预算，所以它不可自执行。这一点值得您知道，因为它对下一条同形裁决同样成立，⛔ 与本卡无关。

三条路（dev 呈上，本席复核后照呈，⛔ 不替您裁）

A —— 批 +10：本席在本 PR 里落 1099 → 1109，附裁决引文与实测重排余量 0（与 1058→1068、1068→1074、1075→1099 三次既有「ruled raise」同形）。
B —— 批一个更小的预算并点名删什么。按成本排序的候选：结尾那句 ⭐ 不变量复述（约 2 行，⚠️ 该不变量在本节已出现三次）、failed-receipt 参考形状的括号（约 1 行）、return { ok: false, error } 例子（约 1 行）。⚠️ 三样都是裁决原文点过名的，所以 dev 一样没敢自己删 —— 这一步本席同意是您的。
C —— 判这条更正不值这个预算：那么那条死路留在账上，本卡改 close not_planned。

本席荐 A，理由只有一条：那条死路正是裁决要求拿掉的东西，而 B 的三个候选全部写在裁决文本里。⛔ 本席不代裁。

另外两件（⛔ 都不在本 PR 里做）

⭐ 同一条死路还活在门禁自己的「修复建议」里，而那是印给正在失败的作者看的。 本席现读 scripts/check-durability-degradation-log-level.mjs 的 remedyFor()：propagationOption 逐字写着「do NOT bolt on a log — declare how it delivers, in FAILURE_PROPAGATION_CALLEES or FAILURE_PROPAGATION_SITES in this script（check:durability-log-level 分不清「故障已答给调用方」和「故障被吞掉」—— 词表加 saveMetaItem 后逼出 2 条基线,记的全是正确代码 #5241）」。⇒ 那 12 个普查成员里任何一个一旦变成 tier-1 可见（只差一条词表项），门禁就会当面叫那位作者去做 AGENTS.md 新文字禁止的事。今天是潜伏的（tier-1 = 0）。⛔ 超出本卡授权的「至多一句头部注释」，本席另立卡。
dev 回答了派发令问的那个问题：门禁头部第 4 条（:79-99）与新文字逐字一致、⛔ 不需要改；真正过时的是 :101-105 那句「What to DO about it is deliberately not decided here … are all open」—— letter C 之后它们不再全开着。⛔ 同样超出授权，随上条一并立卡。

已核通过的部分（一旦行预算到位即可武装）

12 个普查读数在分支上逐字复现裁决的数字（MEMBERS 12 / tier-1 0 / tier-2 4 / tier-3 8，ADJACENT 5，EXCLUDED 4）；pnpm check:durability-log-level exit 0 且本 diff 未改变它；14 个派生门禁族 --ran 对账 14/14、0 NOT-MEASURED，唯一的红就是上面那条行棘轮；PR #19214 是 draft、带 skip-changeset、正文含 ## 维护者速读(草稿)，⛔ 从未翻 ready、⛔ 未入队、⛔ 未挂 auto-merge —— 受管面纪律守住了。

Generated by Claude Code

更正上一条（`5746206377`）的一句承诺 —— 那两条不该另立卡，理由在裁决里

domain:devx 执行席，2026-09-19T23:49Z。⛔ 本条不改状态、不改标签；只撤回一个本席刚说出口、却与维护者裁决相抵的动作。

上一条里本席两次写了「⛔ 超出本卡授权，本席另立卡」。撤回。 本卡的裁决（5716260390）最后一段逐字写着：

The gate gets no further investment; if a real incident — silent data loss reaching a user — is ever traced to one of these gaps, that incident is the card.

并且这条裁决同时把四张同门禁的兄弟卡 closed not_planned。⇒ 为这个门禁再开一张卡，正是那句话刚刚拒绝过的事 —— 即便本席开卡的理由（remedy 文字会当面指错路）和前四张不同。⭐ 「我的理由不同」不是绕过裁决的许可，那是重开裁决的常见形状。

所以那两条改走哪里

它们不消失，只是不以卡的形式存在 —— 改为挂在本卡这条行预算问题下面，因为它们和那段 AGENTS.md 文字是同一件事的两半：

scripts/check-durability-degradation-log-level.mjs 的 remedyFor() → propagationOption 仍逐字说「declare how it delivers, in FAILURE_PROPAGATION_CALLEES or FAILURE_PROPAGATION_SITES」，而它是印给正在失败的作者看的；
同文件 :101-105 那句「re-keying …, reading the declared return TYPE, and recording the shape as out of scope are all open」在 letter C 之后不再为真。

⇒ 若您批了行预算（选项 A 或 B），这两句是同一次落地的自然同伴：各一句话，和那段 AGENTS.md 文字说同一件事，⛔ 不是给门禁加能力 —— 裁决拒绝的是能力投入（A/B/D 三条路），⛔ 不是让一段自相矛盾的提示文字继续印下去。是否一并批，仍由您定。
若您裁 C（不值这个预算），那么这两句和那段文字一起留在账上，⛔ 本席不再另起动作。

⚠️ 本席仍然荐 A；本条只把「另立卡」换成「同一次裁定里附带」，⛔ 不改推荐，⛔ 不改任何读数。

Generated by Claude Code

修 `Maintainer-action:` 的形状 —— 补上它缺的那一半「怎样算做完」

domain:devx 执行席，2026-09-20T00:11Z。⛔ 不改状态、不改标签；本笔只把上面那条 Maintainer-action: 重写成有可检查终点的形状，并公开一条本席选择不做的状态变更及其理由。

本席自己写出来的缺陷，由巡查当场抓到

本轮半态巡查对本卡报 H55：pm:awaiting-maintainer 的 Maintainer-action: 行「names an act but never its completion — no done when half — so the act has no checkable end」。⇒ 成立：5746206377 里本席只写了「裁一个行预算」，⛔ 没写怎样算裁完了、谁据此做什么、机器怎么看得出来。重写如下，取代那一行：

Maintainer-action: 裁 AGENTS.md 的行预算 —— A（批 +10：1099 → 1109）/ B（批更小的额度并点名删哪几样）/ C（不值这个预算 ⇒ 本卡 close not_planned），三选一。
done when: 本卡评论里出现维护者对 A/B/C 的一个字母。此后由本席在 PR #19214 内落对应改动（A/B 时同笔把 scripts/pm/check-skill-line-ratchet.mjs:1438 的 ['AGENTS.md', 1099] 按所批额度改写并引裁决原文；C 时关卡撤 PR），机器可读的完成判据是 pnpm check:pm-skill-ratchet 从 exit 1 转 exit 0。⛔ 在字母落下之前，本席不动那一行、不动上限、不合并。

⇒ 现在这条动作有终点，且终点不靠人读散文判断 —— 一个退出码就能答。

⛔ 本席不把本卡改成 `needs-user-decision`，理由公开记下

同一轮巡查另报 H52：本卡最新 os-dev-report（5746190353）带 1 条非空 open_questions，而卡上没有 needs-user-decision 标签。⚠️ 这条读得没错，但本席权衡后不改，两条理由：

本卡已经有维护者裁决了（5716260390，letter C）。现在缺的是那条裁决执行时的一个参数，⛔ 不是一个新的方向之争。这更像「已裁方向、等维护者补一个动作」，也就是 pm:awaiting-maintainer 的形状。
⚠️ 而且改过去会加深一个既有违规。 决策卡按协议必须在正文里带四棱块与其上的维护者速读。本席现读本卡正文：12611 字节，而 H6 的上限是 10000 ⇒ 本卡正文今天已经超限 2611 字节，⛔ 这不是本席造成的（是立卡时就有的）。把四棱块塞进去只会把超限拉得更远；而要腾出空间，就得把另一个席位实测写就的发现压掉四成以上 —— ⛔ 本席不单方面做这件事。

⇒ ⭐ 所以这里有一条真实的机制张力，本席把它记下而不是悄悄绕开：H52 要求的状态，其协议形状被 H6 的上限挡住，而挡住它的超限与本卡的内容无关。⚠️ 若分诊或维护者认为本卡就该是 needs-user-decision，那么随之要回答的是「谁来压这份正文、压掉哪一段」——那不是本席能自己决定的。

顺带：H64 那条已修并两向钉住

巡查报 5746190353（dev 的 os-dev-report）不含任何 session_ id。已补一行署名：该报告由本席为本卡派出的 os-dev 子代理产出，运行在 session_017ef78bLdybu3AffehKkhfk 内（认领见 5746126007）。两向对照：带 id ⇒ 无行；把 id 抹掉再读 ⇒ 行重新出现 ⇒ 清掉它的确实是那个 id，⛔ 不是巧合。

Generated by Claude Code

更正上一条（`5746316844`）的第二条理由 —— 它引了一条不存在的规则；结论不变，但换成一条真的规则

domain:devx 执行席，2026-09-20T00:12Z。⛔ 不改状态、不改标签。

本席错了什么

上一条为「不把本卡改成 needs-user-decision」给了两条理由，第二条是假的：

⚠️ 而且改过去会加深一个既有违规 …… 本卡正文 12611 字节，而 H6 的上限是 10000 ⇒ 本卡正文今天已经超限 2611 字节。

现读门禁源码：scripts/pm/check-half-states.mjs:2269 起的 H6 是 h6SeatBodyOversized —— 「soft size bound on seat-sticker bodies (#7583)」，常量 :2274 SEAT_BODY_SOFT_LIMIT = 10_000，只作用于 pm:seat 贴。⇒ 普通卡的正文根本不受它管。 本卡正文确实是 12611 字节，但没有任何门禁为此报行，本轮 H6 的 10 行也全是座位贴与旧卡，⛔ 不含本卡。

⇒ 本席是凭记忆引用了一个门禁的适用面，没有现读它的谓词，然后拿它去支撑一个状态判断。记为 self-error。⚠️ 这与本卡自身要治的毛病是同一族：一句没有被重新测量的断言，被当成了论据。

正确的理由，来自本席这次真的去跑的那个仪器

node scripts/pm/check-prior-rulings.mjs --card 16233 对本卡打印：

⚠️ A ruling ALREADY STANDS on this card's own thread (5716260390). Under SKILL.md 〈升级与决策〉③（⛔ 不推翻既有维护者裁决）the card is EXECUTION, not a decision: cite the ruling by comment id on a Ruling-ref: line, and ⛔ never re-present the card as unruled.

⇒ 把本卡挂成 needs-user-decision，正是把一张已裁的卡重新当作未裁呈上去 —— 协议明文禁止。⭐ 上一条的第一条理由（「已裁方向、缺的是执行参数」）因此是对的，只是当时本席是凭直觉说的，现在它有了出处。

⇒ 状态不变：pm:awaiting-maintainer。 并按该仪器的要求补上：

Ruling-ref: 5716260390（维护者裁决，批 #149 项 4，letter C + AGENTS.md 更正）

那 H52 那一行怎么算

本轮巡查的 H52（最新 os-dev-report 带非空 open_questions 而卡无 needs-user-decision）照旧报，⛔ 本席不靠改标签去消它 —— 消它的唯一正当做法是改成一个协议禁止的状态。⇒ 记为已知且有据的代价，理由即上一节。⚠️ 若分诊或维护者认为该行必须清掉，那要动的是 H52 与〈升级与决策〉③ 的交界，⛔ 不是本卡的标签。

⚠️ 上一条里的其余内容全部不变且仍然有效：重写后的 Maintainer-action: + done when（机器可读的完成判据是 pnpm check:pm-skill-ratchet 由 exit 1 转 exit 0）、以及 H64 的修复与两向钉住。

Generated by Claude Code

补一条与您那个行预算裁定相关的成本信息 —— 同一个文件、同一道棘轮，今天来了第二张卡

domain:devx 执行席，2026-09-20T02:03Z。⛔ 不改状态、不催；本笔只更新一个数字，好让那个待裁的问题带上它的真实代价。

本卡在等的是 AGENTS.md 的行预算（A/B/C，Ruling-ref: 5716260390）。本轮另一张卡 #19069（p2，AGENTS.md §11 的 driver-free 探针配方照抄会失败）也落在同一个文件上。

⚠️ 说清楚分寸，⛔ 不夸大：#19069 不一定要占行预算 —— 它的候选修法之一是把现有措辞替换掉（净零或净负行），那样就与棘轮无关。本席已在该卡上把「优先找净零/净负写法，做不到就停下回报，⛔ 不要压别人的段落腾行」写成给接手方的硬约束。

⇒ 所以现在的状态是：一张卡（本卡）确定需要 +10 行；第二张卡（#19069）可能需要，取决于它能不能写成净零。 若两张都要，那就是同一道棘轮上的两笔，而每一笔都只能由您批。

⭐ 顺带一条与您这个决定同源的普遍读数（本席本轮实测）：本仓 32 个受行棘轮管的文件里，31 个余量为 0（唯一例外是 .claude/skills/dogfood-verification/SKILL.md，148/157）。⇒ 这不是 AGENTS.md 的偶然，而是棘轮的设计：每个上限都钉在落地当时的行数。⚠️ 其后果是：任何要求往受管治理文件里写字的裁定，按构造都同时需要第二条裁定来批它的行 —— 否则它不可自执行。本卡就是第一个撞上这件事的，#19069 可能是第二个。

Generated by Claude Code

`Maintainer-action:` 补一条完整行 —— 本席此前那两条都不合格,原因写在明处

domain:devx 执行席(座位贴 #6023)。⛔ 本评论不改裁定、不改标、不催办,只把 H55 欠的那一行写成可被机器读的形状。

Maintainer-action: 裁 AGENTS.md 的行预算 —— A(批 +10:1099 → 1109)/ B(批更小额度并点名删哪几样)/ C(不值这个预算 ⇒ 本卡 close not_planned),三选一 — done when 维护者在本卡线程上留下一条点名 A / B / C(或等效裁定)的评论,本席即按该裁定执行。

⛔ 本席自己的两条错,记在这里

5746206377 —— 写了 Maintainer-action: 但没有 done when 半句。H55 的判据是 /\bdone when\b/,一个没有完成条件的行「名了一个动作却没有可检的终点」,H11 只能拿着一个年龄。
5746316844 —— 那一条写在 blockquote(>)里。directiveValues 容忍反引号、-/* 项目符号与 ** 粗体,⛔ 唯独不容忍 > ⇒ 它根本没被读到过。⚠️ 这与本席 9 月 19 日在 Restart-when: 上踩的是同一个坑(Four more @objectstack/* packages still reach the Console bundle from objectui's lockfile — the same publish-ordering trap #8134 closes for spec #9659 / [finding] The whole 520 KB @objectstack/lint entry rides the console's eagerly-loaded vendor chunk, to serve one advisory rule — and its module/fs/path imports are browser-stubbed #9707 两条修复当时也是发在 > 里,重发后才 fireable-in-any-channel=true)—— 同一个坑的第三次,所以这次先用仓里自己的读法验过再发。

上面那一行本席已用 maintainerActionValues() / maintainerActionComplete()(取自 origin/main 的 check-half-states.mjs)离线验过:值读出 1 条,complete=true。

⛔ 裁定本身没有变

本卡仍是 pm:awaiting-maintainer,仍挡着 AGENTS.md 的全部改动(含 #19069 的 driver-free 探针配方修正)。⛔ 本席不代裁行预算。

_{domain:devx 执行席 · 座位贴 #6023 · 读数 2026-09-20T06:22Z}

Generated by Claude Code

Maintainer ruling recorded — letter A, the `AGENTS.md` line budget is approved at +10 (1099 → 1109)

Recorded by domain:spec seat 5 (session_019srGWGCBBCBHqcDoRZpQRh, seat post #19357) at 2026-09-20T13:59Z. ⛔ This seat does not own domain:devx and did not rule anything — it is relaying a ruling under the maintainer's direct-dispatch channel, with all three provenance parts the protocol requires.

Whose instruction: the maintainer (elon@objectstack.ai), answering the decision this card's Maintainer-action: line (5748099287) put to him.

Where it was said: in this seat's live session, session_019srGWGCBBCBHqcDoRZpQRh — ⛔ not on a GitHub thread, which is why this comment exists: the done when half of that line asks for the letter on this card's thread, so the letter is carried here verbatim with its source named, ⛔ not paraphrased and ⛔ not translated.

The words, verbatim and untranslated:

docs(agents): say how a return-delivered durability seam is declared #19214 / 卡 check:durability-log-level cannot represent a seam that propagates a durability failure by RETURNING an outcome object — its propagation lists are keyed on callee names #16233 —— AGENTS.md 行预算, A 批 +10(1099→1109)

⇒ Ruling-ref: 5716260390 (the standing letter-C ruling this card executes) · line budget ruled separately, this comment.

What executes, and what does not

The remaining half of 5716260390 is already written: PR #19214 carries AGENTS.md +20/−10 and nothing else — this seat measured the branch itself (git diff 24d622b94..f2c3ebf6d, 1 file), and the head file is 1109 lines against origin/main's 1099. The ratchet entry was not touched by the dev, which was correct: scripts/pm/check-skill-line-ratchet.mjs:1438 still reads ['AGENTS.md', 1099] on the PR head.

So exactly one thing is owed, and it is now authorised:

['AGENTS.md', 1099] → ['AGENTS.md', 1109], with the raise recorded in this file's own established ruled-raise convention — the comment block above the entry, carrying the measured cost, the re-measured lossless-rewrap headroom, and the ruling verbatim with its source. The three existing raises on this entry (1058→1068, 1068→1074, 1075→1099) are the shape to copy; the newest of them quotes 「135 同意」 with its batch and comment id and closes with "AGENTS.md is not a CROSS_FILE_MOVES destination, so no ruledRaises record applies. Landed count, headroom 0, same convention." — all three of those sentences still hold and belong in the new block.
origin/main merged into the branch first: the merge base is 24d622b94 and main has moved a long way since 2026-09-19T23:44Z. ⛔ A merge commit, never a rebase or a force-push — this is another seat's branch and its checkout must stay valid.

done when, unchanged and machine-readable: pnpm check:pm-skill-ratchet goes from exit 1 to exit 0. The failing assertion this seat reproduced locally at the PR head, verbatim:

✗ check-skill-line-ratchet: AGENTS.md is 1109 lines; the ratchet ceiling is 1099. …
  Raising a ceiling requires a maintainer ruling quoted in the PR.

⛔ What this ruling does NOT authorise

⛔ Not a ready flip, not an enqueue, not an auto-merge arm. AGENTS.md is a governed surface — node scripts/pm/check-governed-merges.mjs --pr 19214 → exit 3, 1 of 1 path hit the register — so under the governance: human merge IS the review record for governed surfaces — retire the per-PR ADR approval check, add a report-only post-merge audit #9495 regime the human merge is the review record. The PR stays draft and lands by the maintainer's hand. A line budget is not a landing permission.
⛔ Not the two companion sentences from 5746211232 (the remedyFor() → propagationOption text in scripts/check-durability-degradation-log-level.mjs, and its :101-105 "are all open"). That comment put them to the maintainer as a separate yes/no and the ruling names only the budget ⇒ they stay unruled and out of this PR. ⚠️ Recorded as a live cost, not dropped: until one of them is ruled, that gate still prints, to an author who is failing it, the instruction the new AGENTS.md text forbids.
⛔ No content change to the ruled paragraph. 5716260390 named the text; the dev may not shrink it to buy lines, and this seat did not ask it to.

⚠️ One reading the maintainer should have, because it recurred tonight

This is the second governed PR red on this exact gate in one shift: #19303 (.claude/skills/pm-dispatch/references/lanes/director.md, +2/−1) needs 72 → 73, measured here on both refs. That is the general property this card's own thread already reported (5746910798): 31 of the 32 ratchet-capped files stand at headroom 0, so any ruling that requires writing text into a governed instruction file structurally needs a second ruling for its lines. ⛔ This seat does not act on that observation — raising a ceiling is the maintainer's floor, per instance, until he rules otherwise.

Generated by Claude Code

chore(pm): raise the AGENTS.md line ratchet to the ruled 1109

ACCEPT — the ruled raise is landed on the branch, `check:pm-skill-ratchet` is exit 0, and PR #19214 is now waiting on the maintainer's hand and nothing else

Posted by domain:spec seat 5 (session_019srGWGCBBCBHqcDoRZpQRh, seat post #19357) at 2026-09-20T14:57Z, under the maintainer's direct-dispatch channel (the authorisation and the ruling are quoted on this card at 5750261694). ⛔ This seat does not own domain:devx, does not take this card, and has not touched its labels, its assignee or its state — those stay exactly as os-try-charles left them.

① The `done when` is satisfied — this seat's own reading, ⛔ not the dev's

Run by this seat in its own worktree at head aa2e7408d, exit codes captured by redirect before any pipe:

node scripts/pm/check-skill-line-ratchet.mjs              →  TRUE EXIT=0
  ✓ check-skill-line-ratchet: AGENTS.md is 1109 lines (ceiling 1109; headroom 0).
node scripts/pm/check-skill-line-ratchet.mjs --self-test   →  TRUE EXIT=0
  ✓ 157 cases pass — including “every live ruled-raise record quotes its ruling, dates it,
    and names a positive line count”, i.e. the gate validating this very comment block.

⇒ pnpm check:pm-skill-ratchet goes exit 1 → exit 0, which is exactly the machine-readable completion criterion this card's own Maintainer-action: line named (5748099287).

② The diff, re-derived rather than read off the report

git diff --stat origin/main...aa2e7408d ⇒ two files, AGENTS.md +20/−10 and scripts/pm/check-skill-line-ratchet.mjs +29/−1.

what was checked	reading
the ruled paragraph was not shrunk to buy lines	`git diff f2c3ebf6d aa2e7408d -- AGENTS.md` is EMPTY ⇒ byte-identical to the previous round. ⭐ The budget existed so the text would not have to move, and it did not move
only the one ceiling moved	`['AGENTS.md', 1109]` at `:1466` (the line number shifted because the comment block went in above it); the widest-table-row pin at `:1938` is still `768`, untouched
the count is the ruled arithmetic, not an assumption	`AGENTS.md` is 1109 lines on the head, 1099 on `origin/main`, and `main` touched `AGENTS.md` in neither merge window ⇒ 1099 + the ruled 10 = 1109 exactly
no history was rewritten on someone else's branch	the previous head `f2c3ebf6d` is still an ancestor; two merge commits (`32d6e4775`, `aa2e7408d`) and one work commit (`cffced21f`). ⛔ No rebase, no amend, no force-push
nothing was weakened	`check:ratchet-remedy-authority` re-read and unmoved (this script classifies `excluded`, 0 live offers); 35 derived gate families, 35 run, 0 NOT-MEASURED, every one exit 0; repo-wide `pnpm lint` exit 0
the PR was not landed by a seat	still draft, labels unchanged (`documentation, size/s, skip-changeset`), no auto-merge, and `check-governed-merges --pr 19214` → exit 3 (GOVERNED, tier H)

② bis — the claim's declared file surface, amended in the same round

The claim of record (5746126007) declares File surface: AGENTS.md only. The maintainer's budget ruling authorises the edit to scripts/pm/check-skill-line-ratchet.mjs, which that claim does not list, so the dev explained rather than stopping and asked the seat to amend it. ⇒ Amended here: the declared file surface for this round is AGENTS.md + scripts/pm/check-skill-line-ratchet.mjs. ⛔ Nothing else; the two "companion" sentences in scripts/check-durability-degradation-log-level.mjs are not in this PR and stay unruled.

③ Three things this seat told the dev that it measured FALSE — all three stand corrected

⭐ This seat's commit-trailer instruction would have blocked the push. The dispatch prescribed Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>. scripts/check-commit-card-trailers.mjs's coauthorModelIdentifier() (:361) flags any harness-address co-author whose display name is not exactly the declared bare name, and .githooks/pre-push runs it ⇒ that spelling reds with "in a co-author trailer — the pair is model-free". The dev used the repo's own pair instead and the pre-push gate confirmed it on both pushes. This seat has re-read the predicate and confirms the dev is right: AGENTS.md rules the trailer pair model-free, and the repo enforces it. ⛔ This seat's future dispatches carry the repo's spelling, not the prescribed one.
"0 lossless rewrap headroom at three widths" — true only for widths ≤90 B. Re-measured with the gate's own wrapLine: the paragraph reflows to 18 lines at 100 B and 15 at 120 B. ⇒ the honest justification is not that no width saves lines, it is that re-wrap is not legal currency here (this file's 2026-08-17 rule refuses re-wrap funding, and the 120-byte cap documents itself as codifying the corpus's existing ≤91-byte ASCII wrap rather than mandating a reflow to it). ⭐ The dev wrote it that way in the comment block instead of inheriting the flat claim, which is the better record.
The PR body's rewrap table read 21/20/20 at 86/88/90 B; measured 21/21/20, and the paragraph as landed carries four lines above 88 B (max 90) ⇒ it sits at ≤90 B, not 88 B. Conclusion unaffected and slightly stronger.

④ The PR body was rewritten by this seat, and the seat broke something doing it

The body still said "the dev seat did not touch scripts/pm/check-skill-line-ratchet.mjs; the +10 is reported as measured and the budget is the approver's to rule" — the opposite of what the diff now does, on the body the maintainer reads as the landing record for a governed surface. A dev may not edit a PR body, so the seat did: the stale section is replaced by the ruling verbatim, the two corrected numbers, the Clause-②: no line the body lacked, and the landing constraint; the 维护者速读 items 2–3 are updated; everything else is byte-identical.

⚠️ The first attempt stacked a second footer, and the cause is a platform fact worth keeping. This seat preserved the body's existing footer line, believing that to be the safe shape. It is not: that footer carried a session-scoped URL (…/code/session_017ef78bLdybu3AffehKkhfk), which does not match the platform's canonical footer, so the platform read the body as having none and appended its own — read-back showed 2. ⇒ AGENTS.md:432 fires not only on re-sending a canonical footer but harder on preserving a non-canonical one, because that one cannot be matched. Repaired in the next act by stripping both blocks and sending none (read-back: exactly 1), with the authoring session preserved as prose so no attribution was lost.

⑤ What is left: one act, and it is not a seat's

⛔ Nothing on the AI side remains. AGENTS.md is on the governed register, so under the #9495 regime the human merge is the review record: the maintainer merges PR #19214 by hand, from draft. ⛔ No seat flips it ready, enqueues it or arms auto-merge — a line budget is a budget, not a landing permission. This card stays pm:awaiting-maintainer with its assignee intact until that merge closes it through Fixes #16233.

Generated by Claude Code

os-closed-card-sweep — machine-findable marker for this generated comment.

Removed the pm-loop state label(s) this closed card no longer claims: pm:awaiting-maintainer.

Closing pull request: docs(agents): say how a return-delivered durability seam is declared #19214, merged.
Closing commit 287eb4c2ea, merged into main.
Left untouched: tooling, domain:devx, priority:p3 — ownership, priority and outcome are not state claims.
The label set was read back after the write and matched.

A state label claims work is in flight. This card is closed on a merged delivery, so the claim
is stale; every other label is left exactly as it was found. Nothing here is a judgement about
the card, and no verdict-bearing label is ever touched by this sweep.

posted by half-state-patrol run 35552505373 · trigger schedule

Generated by Claude Code

fix(pm): read the ownership markers through the undecorated line, and…

fix(pm): the near-miss ownership census excludes seat posts and relea…

docs(agents): say how a return-delivered durability seam is declared (o…

check:durability-log-level cannot represent a seam that propagates a durability failure by RETURNING an outcome object — its propagation lists are keyed on callee names #16233

Description

The gap

What was measured, and what was not

Why the shape of the fix is not obvious

⚠️ Context a maintainer may want more than this row

Provenance and dedup

决策箱 —— 普查买回来了,数是 12,⇒ 路线选择越过了本卡口径,交维护者

先把「为什么现在必须有人裁」说清:⭐ 不是仪器缺一条腿,是治理文件自己写出了一条死路

选项 × 真实代价

业务含义直译

四棱

裁定之后怎么执行(维护者只裁方向,⛔ 不背执行)

Activity

hotlong commented on Sep 17, 2026

Ruling — C + correction

Execution

os-try-charles commented on Sep 19, 2026

⚠️ 本席必须先说清楚：这张卡在本席的取卡序里被漏了七十多轮，而漏的原因不是它

本席现读：裁决的哪一半已经做完，哪一半还欠着

本席自己把裁决的事实基础重跑了一遍（⛔ 不是转述裁决）

⚠️ 受管面：本卡交的是草稿 PR，落地不是本席的动作

给 dev 的边界

os-try-charles commented on Sep 19, 2026

os-try-charles commented on Sep 19, 2026

⏳ pm:dispatched → pm:awaiting-maintainer —— 裁决点的那段文字写好了（草稿 PR #19214），但它要不到行

本席复核过的阻塞（⛔ 不是采信 dev 的叙述）

⭐ 这不是本卡的偶然 —— 本席把它量成了一条普遍性质

三条路（dev 呈上，本席复核后照呈，⛔ 不替您裁）

另外两件（⛔ 都不在本 PR 里做）

已核通过的部分（一旦行预算到位即可武装）

os-try-charles commented on Sep 19, 2026

更正上一条（5746206377）的一句承诺 —— 那两条不该另立卡，理由在裁决里

所以那两条改走哪里

os-try-charles commented on Sep 20, 2026

修 Maintainer-action: 的形状 —— 补上它缺的那一半「怎样算做完」

本席自己写出来的缺陷，由巡查当场抓到

⛔ 本席不把本卡改成 needs-user-decision，理由公开记下

顺带：H64 那条已修并两向钉住

os-try-charles commented on Sep 20, 2026

更正上一条（5746316844）的第二条理由 —— 它引了一条不存在的规则；结论不变，但换成一条真的规则

本席错了什么

正确的理由，来自本席这次真的去跑的那个仪器

那 H52 那一行怎么算

os-try-charles commented on Sep 20, 2026

补一条与您那个行预算裁定相关的成本信息 —— 同一个文件、同一道棘轮，今天来了第二张卡

os-try-charles commented on Sep 20, 2026

Maintainer-action: 补一条完整行 —— 本席此前那两条都不合格,原因写在明处

⛔ 本席自己的两条错,记在这里

⛔ 裁定本身没有变

os-elon-musk commented on Sep 20, 2026

Maintainer ruling recorded — letter A, the AGENTS.md line budget is approved at +10 (1099 → 1109)

What executes, and what does not

⛔ What this ruling does NOT authorise

⚠️ One reading the maintainer should have, because it recurred tonight

os-elon-musk commented on Sep 20, 2026

ACCEPT — the ruled raise is landed on the branch, check:pm-skill-ratchet is exit 0, and PR #19214 is now waiting on the maintainer's hand and nothing else

① The done when is satisfied — this seat's own reading, ⛔ not the dev's

② The diff, re-derived rather than read off the report

② bis — the claim's declared file surface, amended in the same round

③ Three things this seat told the dev that it measured FALSE — all three stand corrected

④ The PR body was rewritten by this seat, and the seat broke something doing it

⑤ What is left: one act, and it is not a seat's

github-actions commented on Sep 21, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

⏳ `pm:dispatched` → `pm:awaiting-maintainer` —— 裁决点的那段文字写好了（草稿 PR #19214），但它要不到行

更正上一条（`5746206377`）的一句承诺 —— 那两条不该另立卡，理由在裁决里

修 `Maintainer-action:` 的形状 —— 补上它缺的那一半「怎样算做完」

⛔ 本席不把本卡改成 `needs-user-decision`，理由公开记下

更正上一条（`5746316844`）的第二条理由 —— 它引了一条不存在的规则；结论不变，但换成一条真的规则

`Maintainer-action:` 补一条完整行 —— 本席此前那两条都不合格,原因写在明处

Maintainer ruling recorded — letter A, the `AGENTS.md` line budget is approved at +10 (1099 → 1109)

ACCEPT — the ruled raise is landed on the branch, `check:pm-skill-ratchet` is exit 0, and PR #19214 is now waiting on the maintainer's hand and nothing else

① The `done when` is satisfied — this seat's own reading, ⛔ not the dev's