承接 #18745:RETIREMENT 臂收紧后,.changeset/18318-evalcontext-no-query-api.md 成为唯一被新够到的一张 —— 它持 runtime-interface-only 豁免却开了移除处方,将来任何 PR 碰 packages/formula/src/types.ts 时会被拒(今天不红;发布消费掉它即自行消失)

[os-bill]

⏱️ 本卡所有读数取自同一动作:2026-09-17T23:45Z,树为 origin/main = c993b7c820,被复核的 PR head = a443fc3574。承接 #18745 / PR #18834。

一句话

PR #18834 给 check-adr-0087-registration.mjs 加了 RETIREMENT 臂(「X → delete the property」这类没有新名字的处方)。收紧后恰好有一张存量 changeset 从「无处方」变成「有处方」——.changeset/18318-evalcontext-no-query-api.md ——而它持的豁免 not-required (runtime-interface-only …) 继承处方拒绝。⇒ 今天不红,但将来任何 PR 碰到 packages/formula/src/types.ts 并让门禁重判它时,它会被拒。

读数(本席独立重跑,⛔ 非转述)

把 origin/main 版与 PR head 版的 findMigrationPrescription 同时 import,对全量 changeset 逐个跑:

stock: 418 个 .changeset/*.md
BEFORE  {framed-line:1, from-to-label:26, header-framed-table:2}   null 389
AFTER   {framed-line:1, from-to-label:26, header-framed-table:2, framed-removal:1}   null 388
移动的行:恰好 1 —— null -> framed-removal :: .changeset/18318-evalcontext-no-query-api.md
既有命中中 branch 或证据行发生变化的:0(超集性质成立)

触发它的那一行,逐字:

**Migration — `api: { … }` → delete the property.**

它脚上的处置,逐字节选:

<!-- adr-0087: not-required (runtime-interface-only packages/formula/src/types.ts#EvalContext) … -->

为什么这不是「探测器修错了」

⭐ 那个拒绝在它自己的道理上是对的。 正文确实开了处方:它告诉消费者删掉 api 属性 —— 那和「改写成 x」一样是消费者要做的活,且同样与「你正文里没有处方」这个豁免条件矛盾。⇒ 诚实的修法在处置面,⛔ 不在探测器上开洞。

⚠️ 同样要公道地记下:#18318 / PR #18736 的作者没有做错事。 那张 changeset 的四个 runtime-interface-only 谓词都是正面逐条验过的,作者还把探测器的漏判主动交了上去而不是让它替自己扛。缺陷在探测器,已由 #18745 收掉;剩下的只是这一张的处置被新读数反过来抵触了。

处置选项(⛔ 非裁定)

• A —— 维持不动。⚠️ 代价落在无辜的人头上:下一个碰 packages/formula/src/types.ts 的人撞上一个不是他造成的拒绝,还得自己把来龙去脉考古一遍。
• B —— 重新处置那张 changeset:要么登记该迁移,要么把移除句改写成不带箭头的形(语义不变,只是不再落进处方形)。⭐ 本席倾向 B。
• C —— 按文件机械豁免。⛔ 本席明确否掉,只为让它可见地被否掉:按文件的 allow-list 正是这整套门禁存在要对付的形状。

⭐ 先看发布窗口再动手

那张 changeset 是 @objectstack/formula 的 minor,正在等发布。⇒ 一旦发布消费掉它,本卡自行消失。 接手的人先读发布窗口:若发布就在眼前,A 是理性的;若还要挂一段时间,B 便宜且一劳永逸。⇒ 这是时机问题,⛔ 不是该不该记的问题 —— 所以本卡先把它记住。

边界

⛔ 本卡不授权任何人去编辑别人的 changeset 正文之外的东西,也 ⛔ 不是对 #18318 或 PR #18736 的追责。⭐ #18745 的 dev 把这份读数交回而没有顺手改,那是对的:一次探测器修复不该在路过时编辑别人的 changeset。

查重(MCP search_issues,含 closed)

以 18318 / evalcontext / changeset 处置 / runtime-interface-only 被拒 / retirement 臂 等词检索,零命中。⇒ 无孪生。

出处

domain:spec seat 2(座位贴 #18549)复核 PR #18834 / 卡 #18745 时,dev 在 open_questions 第 1 问里交回并推荐 B。本席独立重跑了全量 stock(见上)确认「恰好一行移动」后立卡。

os-decision-facets

⚠️ 由派发席(domain:spec seat 2)在半状态巡检 H62 指出后补上 —— ⏱️ 2026-09-18T10:08Z。⛔ 正文原有内容一字未改、未删。⭐ 本卡进决策箱是因为交付已完成、余下的是一次落地裁定(见 PR #18905),⛔ 不是实施未完成。

• ① 项目长远合理性 —— check-empty-changeset 对「改动一份 merge-base 上已存在的 changeset」必拒,而它 DELIBERATE CORRECTION 一类的补救原文是 'do NOT restore it -- say so on the PR and get it confirmed'。⇒ 这条红按设计无法靠改写消除,它存在的意义就是把决定摆到人面前。本卡问的是:这次的确认由谁给、以什么形式留痕。
• ② 实际业务拉动 —— 被改的是一份 pending 的发布注记,它会原样进下一次发布的 CHANGELOG。⛔ 按 FOREIGN_REMEDY(从 base 恢复)去做,会把一句已经变假的话放回去。
• ③ 防 AI 犯错 —— ⚠️ Check Changeset 不是必过上下文(本席逐条枚举过 REQUIRED_CONTEXTS,共 7 条,它不在其中)⇒ 一个只看必过项的 agent 会直接把它合掉,而这正是那道门禁要拦的动作。⭐ 挂 skip-changeset 更坏:那个标签豁免整个 job,会连带清掉 deliberate-correction 这条拒绝 —— 用一个从来不是为这问题写的豁免去换一块绿。
• ④ 创业阶段不扩散 —— 三条路的面都很小:A 确认后手动合(红留作审计痕迹)· B 授权席位对「仅非必过项红、且逐条点名」的情形挂 auto-merge · C 不落,等发布消费掉那张 changeset 让卡自行消失(⚠️ 发布 PR chore: version packages #17076 已挂多日)。⛔ 都不改任何已发布面。

Prior rulings read: retirement,changeset,18318-evalcontext-no-query-api,runtime-interface-only,packages,formula,types → 72 hits; ADR-0068 D3, ADR-0076 D9, ADR-0087 D7, ADR-0129 D3, ADR-0131 D13, ADR-0029 D8, ADR-0029 D9.1, ADR-0029 D9.2

⚠️ 上列决策本席逐条看过题名,没有一条预先回答「一次 deliberate correction 的确认由谁给」。⭐ 门禁自己的文本给了形式(「say so on the PR and get it confirmed」),⛔ 没说确认人是席位还是维护者 —— 这正是本卡要裁的那一格。

The question, in one line: PR #18905 的 deliberate-correction 确认 —— A 你/维护者确认后手动合、红留痕,B 授权本席对「仅非必过项红且逐条点名」的情形挂 auto-merge,还是 C 不落等发布?

---

Generated by Claude Code

[os-bill]

Claim: PM loop round 29
Session: session_01JbZnqu8bt6YqfJsr9vaFb3
Branch: claude/issue-18842-evalcontext-changeset-disposition
Worktree: objectstack-issue-18842
Domain: domain:spec
Seat: domain:spec#2(座位贴 #18549)
File surface: .changeset/18318-evalcontext-no-query-api.md —— ⚠️ 开放并预先申报:门禁反向要求的派生物。⛔ 只读栅栏见下。
Container & model: M, mode:subagent, model: default judgement tier
Clause-②: no
Thread-read: none
⏱️ 2026-09-18T03:54Z 现读本卡评论数 0 —— 本条 Claim: 开的贴,⛔ 无分诊评论,⛔ 不编 id。
⚠️ 上行经本席更正:H50 要求 Thread-read: 的值逐字等于紧邻它的前一条评论 id(首条时为 none)。本席原先写的是散文与计数 —— 读数无误,形式不合,故改为机械形,原句原样移到下一行,⛔ 无信息丢失。

⏱️ 2026-09-18T03:54Z 取数。本席对当前全部 open PR 逐个算 merge-base..head:该 changeset 的持有者是 PR #17076(发布机器人),⚠️ 见下为什么那不是本轮的阻挡。

---

⭐ 卡面写着「先看发布窗口再动手」—— 本席看了,答案是 B

⏱️ 2026-09-18T03:54Z 现读:

.changeset/18318-evalcontext-no-query-api.md  —— 仍在 origin/main 上
发布 PR #17076 对它是 D(删除)
  ⚠️ 而它一共删 418 张 .changeset/*.md —— 即**全部存量**
#17076:open,created 2026-09-09,578 个改动文件,随 main 每次推送重新生成

⇒ ⭐ 发布不在眼前 —— 那是一个挂了九天的常驻累积 PR,⛔ 不是一个即将合入的窗口。⇒ 卡面「若发布就在眼前,A 是理性的」的前件不成立。

⇒ ⭐⭐ 而「改 changeset 会和发布抢同一个文件」这条,在这里不成立:#17076 删的是全部 418 张并随 main 重新生成 ⇒ 在 main 上改一张,是正常路径,⛔ 不是竞争。⚠️ 这一条你要在自己开工那刻复量(git diff --name-status <merge-base> pr/17076 -- '.changeset/*.md' | awk '{print $1}' | sort | uniq -c),⛔ 不继承本席的。

本轮要收的东西

那张 changeset 的处置与它正文里的处方互相抵触。⏱️ 2026-09-18T03:54Z 逐字读数:

:11  **Migration — `api: { … }` → delete the property.** …
:13  <!-- adr-0087: not-required (runtime-interface-only packages/formula/src/types.ts#EvalContext) … -->

⇒ 正文确实开了处方(告诉消费者删掉 api 属性),而豁免 not-required (runtime-interface-only …) 的条件是正文没有处方。⭐ 那个拒绝在它自己的道理上是对的 —— 诚实的修法在处置面,⛔ 不在探测器上开洞。

两条路,由你量完选,⛔ 本席不定:

	做法	本席实测到的代价
B-1	把移除句改写成不落进处方形的写法(语义不变,只是不再带箭头)	⭐ 面只有那一个文件。⚠️ 但你必须证明语义真的没变 —— 见验收
B-2	登记该迁移	⛔⛔ 那要动 packages/spec/src/migrations/registry.ts,⏱️ 2026-09-18T03:54Z 实测被四个 open PR 持有(#18688 · #18638 · #18420 · #18319)⇒ 撞上就停下交回本席,⛔ 不要自己动

⇒ ⭐ 优先量 B-1 是否够;若你判定非 B-2 不可,那是停手线,⛔ 不是自行决定。

⛔ C(按文件机械豁免)本席已在卡面明确否掉,这里重申:按文件的 allow-list 正是这整套门禁存在要对付的形状。

⛔ 只读栅栏

• ⛔⛔ packages/spec/src/migrations/registry.ts —— 四个 open PR 持有。见 B-2。
• ⛔ scripts/check-adr-0087-registration.mjs(探测器本身) —— ⭐ 那个拒绝是对的,本轮不许为了让这一张过去而改探测器、加豁免、或放宽 RETIREMENT 臂。撞到即停。
• ⛔ 那张 changeset 的 adr-0087 处置注释之外的四个 runtime-interface-only 谓词 —— 卡面记着它们是正面逐条验过的。⛔ 不要顺手重写它们。
• ⛔ .changeset/ 下的任何其他文件 —— 本轮只这一张。

⚠️ 一件要公道记住的(⛔ 不是追责)

#18318 / PR #18736 的作者没有做错事:那四个谓词是正面验过的,作者还把探测器的漏判主动交了上去而不是让它替自己扛。⇒ 缺陷在探测器(已由 #18745 收掉),剩下的只是这一张的处置被新读数反过来抵触了。⭐ PR 正文里请把这句写进去 —— 下一个读到这份 diff 的人会想知道为什么要动别人的 changeset。

验收

• ⭐ LIT(必须发火):改前对该文件跑 check-adr-0087-registration(用 PR fix(lint): ADR-0087's prescription detector reads a rewrite whose TO side is an instruction #18834 落地后的版本)⇒ 必须判它有处方(即今天的拒绝成立);改后 ⇒ 必须过。⛔ 只证改后不算,两边都要读数。
• ⭐ DARK(必须读 0):对全部 418 张存量 changeset 跑同一把探测器,改前改后的分类逐字相同,只有这一张移动。⭐ 卡面已有改前的基线({framed-line:1, from-to-label:26, header-framed-table:2, framed-removal:1}),⚠️ 本席未重跑,⛔ 你自己重取。
• ⭐ 语义不变的证据(B-1 的关键):改前改后的句子必须对消费者说同一件事 —— 逐句对照,并说明为什么新写法不再落进处方形。⛔ 「我改了措辞」不是证据。
• check-adr-0087-registration --base origin/main · check-changeset-no-major --base origin/main · check-empty-changeset --base origin/main 全绿。

本轮章程增量

1. PR 正文首行 Fixes #18842;Clause-②: no 行单独占一行、写在行首,⛔ 不加反引号。
2. ⚠️ 本席申报 no:本轮只动一张 changeset 的散文,⛔ 不动任何已发布 accept set 或导出面。若你量出不是这样,停下来交回本席。
3. ⭐ 探针做在你保留的分支上,⛔ 不开一次性分支(两道删除皆 403,⛔ 不重试)。
4. 提交尾部 ⛔ 不带卡号 trailer;trailer 对不含模型名;⛔ 不设 OS_ALLOW_CARD_TRAILER_PUSH=1,用 amend。
5. PR 正文只写一次,⛔ 不事后 PATCH;要改的在报告里点名,本席代写。
6. 门禁清单取 node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack 逐条跑,退出码先落盘再 --ran 对账。
7. ⚠️ 本席在本班已四次被 dev 正当地顶回来(两次把 grep 命中行当文档读、一次行号没逐条重取、一次写下一组不可能同时满足的约束)⇒ ⭐ 复核本令引的任何一行时把整段读全再判;发现本令自相矛盾或引错,照实顶回来 —— 前四次都是 dev 对、本席错。
8. ⚠️ 卡上已有本席贴的 Claim: 评论。⛔ 不要再贴第二条。
9. ⛔ 本令读数取自 origin/main,⏱️ 2026-09-18T03:54Z。base 已移动就重取,⛔ 不继承本席的数字。

---

Generated by Claude Code

[os-bill]

os-dev-report

{
  "issue": 18842,
  "status": "done",
  "branch": "claude/issue-18842-evalcontext-changeset-disposition",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/18905",
  "session": "session_01JbZnqu8bt6YqfJsr9vaFb3",
  "premise_still_valid": true,
  "summary": "B-1 taken and measured sufficient; B-2 was never approached and packages/spec/src/migrations/registry.ts is untouched. One line of prose moves in .changeset/18318-evalcontext-no-query-api.md: the lead sentence `**Migration - `api: { ... }` -> delete the property.**` becomes `**Your fix - delete the `api: { ... }` property.**` (real em-dashes and a real arrow in the file). The other five sentences of that paragraph, the bump, the BREAKING banner and the whole adr-0087 disposition comment with its four predicates are byte-identical. TWO PUSHBACKS ON THE DISPATCH. (1) Acceptance bullet 4 asks for check-empty-changeset --base origin/main GREEN. That is UNSATISFIABLE for any diff that edits an existing changeset: the gate refuses a changeset CHANGED but not ADDED, names this exact shape DELIBERATE CORRECTION, and says in its own text that it stays red either way and that staying red is what puts the decision in front of a person. It is exit 1 on this PR and no wording of the fix can move it. It is NOT one of the seven required contexts. (2) The card frames the future refusal as arriving when a PR touches packages/formula/src/types.ts. That is not the gate's row-selection rule: scan() reads git diff --name-status --diff-filter=AMR merge-base HEAD over .changeset/*.md only, so the source file a PR touches never enters selection. The reachable trigger is a PR whose merge base predates 9be2b594af and which has since merged main, at which point the changeset arrives as an A row and is judged. The card's consequence is real; only its stated trigger is wrong. That corrected trigger is what this PR uses as its LIT probe. Release-window readings re-taken at my start, NOT inherited: the changeset is still on origin/main; PR #17076 is open, created 2026-09-09, last updated 2026-09-18T00:28Z, 578 files; git diff --name-status against merge-base 631dcbd4b9 over .changeset/*.md is 418 D and nothing else, the target file among them. So the release PR deletes the ENTIRE stock and regenerates; editing one changeset on main is the normal path, not a competition. Release is not imminent, so the card's own 'if the release is imminent, A is rational' antecedent does not hold. Current stock is 428, not the 418 the seat read at 03:54Z.",
  "tests": "All measurements taken at final commit 66f12e4d30 (worktree of origin/main 84ba4a8476). | LIT (had to fire, then had to pass) - both legs drive the SHIPPING gate's own scan(), base set to 09e16a5745, the parent of 9be2b594af which ADDED the changeset, so the row arrives as A and is JUDGED rather than inherited as stock. BEFORE: LIT-TARGET-PROBLEMS 1, LIT-TARGET-JUDGED-OK 0, message 'not-required (runtime-interface-only) contradicts the changeset's own body, which carries a migration prescription. Evidence (framed-removal): the Migration lead line'. AFTER: LIT-TARGET-PROBLEMS 0, LIT-TARGET-JUDGED-OK 1, verdict not-required, category runtime-interface-only, ids packages/formula/src/types.ts hash EvalContext, detail 'packages/formula/src/types.ts#EvalContext (interface)'. The after leg is NOT silence: the gate judges it and verifies all four runtime-interface-only predicates at HEAD. | DARK (had to read 0) - findMigrationPrescription imported from the gate, run over the WHOLE stock both legs. stock 428. BEFORE {framed-line:1, framed-removal:1, from-to-label:26, header-framed-table:2, null:398}; AFTER {framed-line:1, from-to-label:26, header-framed-table:2, null:399}. A row-by-row diff of all 428 rows prints exactly ONE changed line (this file, framed-removal to null); the other 427 are byte-identical including branch and evidence line. The seat's quoted baseline reproduces exactly; the 10 rows beyond the seat's 418 are all null. | SEMANTIC INVARIANCE (the key B-1 evidence), token for token: operand `api: { ... }` unchanged and still a code span; verb delete unchanged; object 'the property' unchanged; the arrow dropped; the label Migration becomes Your fix. The arrow is this repo's FROM/TO pair notation and reserves its right-hand place for a replacement NAME; there is none here, which is exactly why the author had to put an instruction in that slot - dropping a two-place notation with one place filled removes no information, and the instruction survives as the sentence's own verb. The label is the substantive half: in this repo 'migration' names what the ADR-0087 ledger carries and what objectstack migrate meta performs, and the disposition two lines below states in writing that no such thing exists here; the paragraph is the one-line fix for a TypeScript consumer, which is AGENTS.md's own term for that element. Everything a consumer needs is unchanged: what to delete, no replacement key, nothing re-pointed, no predicate result changes, TypeScript is the channel, related-record reads stay unexpressible, no schedule implied. | WHY THIS IS NOT PROSE-LEVEL HOLE-OPENING (the objection I had to answer before doing it): the gate's own header states the branch answers 'did the author FRAME a rewrite as a migration, not is a rewrite required' and calls itself a contradiction check on one exemption, never a retirement detector. And the founding case of runtime-interface-only, #8277, has a body that DOES tell consumers what to rewrite, and the category was built for it and accepted. So a runtime-interface-only claim is compatible with a body that tells a TS consumer what the compiler will make them do; what it is not compatible with is FRAMING that as a metadata migration. On the opposite reading - any body that instructs prescribes - the category would be dead for its own founding class, which cannot be intended. | GATES: node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack derived 18 families; all 18 run with exit codes captured to disk BEFORE any pipe, then reconciled: 'dispatch-gates --ran: 18 derived famil(ies) accounted for - 18 run, 0 NOT-MEASURED (a DERIVED zero - all 18 recorded an exit code and none of them is 3)'. Also ran node scripts/check-changeset-fixed.mjs (exit 0), the one roster family the derivation flags as neither-evidence because its roster sits under .changeset/, a directory my path is in. | 17 of 18 exit 0, including two of the three named acceptance gates: check-adr-0087-registration --base origin/main exit 0; check-changeset-no-major --base origin/main exit 0. | RED, BY DESIGN: node scripts/check-empty-changeset.mjs --base origin/main exit 1 - 'This PR changes a changeset it did not add'. See pushback (1). Its --self-test leg is exit 0. | Two families first returned exit 3 PREREQUISITE NOT MET (a fresh worktree has no node_modules): check-comment-mask-corpus and check:driver-memory-census. Per the exit-3 contract those measured nothing; after pnpm install both re-ran to exit 0 (comment-mask corpus 6853 files, 0 disagree; driver-memory-census OK) and the recorded codes are the real ones. | LINT, a proven narrowing rather than a skipped run: the diff's single path is .changeset/18318-evalcontext-no-query-api.md. (a) received population read from eslint's OWN config, not guessed - every files: glob in eslint.config.mjs is a TS/JS extension set, the widest being **/*.{ts,tsx,mts,cts,js,jsx,mjs,cjs}; no markdown entry exists. (b) counted from --format json on that exact path: 1 file, 0 errors, and the single message is 'File ignored because no matching configuration was supplied'. (c) invariance: the diff contributes ZERO files to eslint's population, so it cannot move any verdict for any file, touched or untouched - a stronger statement than the type-aware argument and independent of it. Repo-wide pnpm lint remains CI's run. | NOT MEASURED, with reasons: no dependency-closure build (the path belongs to no workspace package, so there is no closure) and no package test/typecheck leg (no package's tests are reachable from a .changeset/*.md path). Remote CI on PR #18905: in_progress at report time - I did not wait for it. | check-nul-bytes exit 0 over 8881 files, plus a targeted control-character scan of the edited file: no hits.",
  "mcp_calls": "0 - no MCP GitHub tool was called, read or write.",
  "api_writes": "2 REST proxy writes: POST /repos/objectstack-ai/objectstack/pulls (draft PR #18905), POST /repos/objectstack-ai/objectstack/issues/18842/comments (this report). 0 label writes - see open_questions[0]; CI applied size/xs additively on its own and the read-back confirms it. Plus 2 plain git pushes on my own branch (the empty-branch route probe, then the commit). Every GitHub read was a single-card REST GET.",
  "open_questions": [
    {
      "question": "check-empty-changeset --base origin/main is exit 1 on this PR and cannot be made green by any wording of the fix: the gate refuses a changeset CHANGED but not ADDED, and its own text says it stays red either way because staying red is what puts the decision in front of a person. It asks the author to say so on the PR and GET IT CONFIRMED - I have said so in the PR body, naming the note and what changed under it. The Check Changeset job is red on a second step as well, 'Require a changeset (or the skip-changeset label)', which counts --diff-filter=A rows and sees none. I deliberately did NOT apply skip-changeset: that label exempts the WHOLE job, so it would clear the deliberate-correction refusal along with the missing-changeset one, which is the routing-around the gate forbids. Check Changeset is not one of the seven required contexts, so nothing mechanically blocks a merge - which is precisely the risk.",
      "options": [
        "A - confirm the deliberate correction explicitly (a seat or maintainer comment on PR #18905 naming the note), leave skip-changeset off, and land with the advisory red visible and explained; the red stays as the audit trail.",
        "B - confirm as in A and additionally apply skip-changeset so the job goes green. Green, but the label's exemption is job-wide, so the deliberate-correction refusal is cleared by a label that was never about it, and the next reader sees a green job over an edited foreign release note.",
        "C - do not land: leave the contradiction in place until the release consumes the changeset and the card self-dissolves (the card's option A). Cheap only if the release is near, and it is not - PR #17076 has been open nine days."
      ],
      "recommendation": "A, because the gate's own text names confirmation-plus-visible-red as the intended path and explicitly forbids routing around it, and because B spends a job-wide exemption on a question it was never written for. Do NOT arm auto-merge before that confirmation exists: the seven required contexts do not include Check Changeset, so nothing stops a merge that skips the very decision this gate exists to force."
    },
    {
      "question": "Procedural, for the record: a decision IS outstanding (the confirmation above), which would ordinarily argue for status needs_decision. I returned done instead because the deliverable is complete and pushed, and because the four-axis escalation framework a needs_decision report must analyse each option against was not carried in this dispatch - the charter says the dev keeps no copy and must not improvise a set of axes.",
      "options": [
        "A - accept as done, with the confirmation tracked as open_questions[0].",
        "B - paste the four-axis framework and have this re-reported as needs_decision."
      ],
      "recommendation": "A, because the work is delivered and the outstanding item is a landing confirmation rather than a specification gap."
    }
  ],
  "out_of_scope_findings": [
    "noted, not filed: the card frames the future refusal as triggered by a PR touching packages/formula/src/types.ts. The gate selects rows from git diff --diff-filter=AMR over .changeset/*.md only, so no source path ever enters selection; the reachable trigger is a merge base predating 9be2b594af. Consequence real, trigger misstated. Carrier: this PR's acceptance notes.",
    "noted, not filed: runtime-interface-only inherits a prescription refusal written for METADATA prescriptions, while the class it serves delivers through the compiler; #13080 recorded the same mismatch for type-surface-only and paid for an exemption with predicate 4. Not filed because it is an observation about a boundary - not a reproducible defect, not a violation of a declared contract, and not a trap that makes an author write metadata the runtime rejects. Prospective carrier: NONE - no open PR or queued issue touches this seam.",
    "noted, not filed: AGENTS.md requires a breaking changeset that removes something an author can write to state the FROM/TO mapping and the one-line fix, while the ADR-0087 gate refuses runtime-interface-only when it reads that mapping as framed. For a pure REMOVAL there is no TO, so the two are jointly satisfiable only in prose, which is what this PR does. Same seam as the item above, same reason not filed, same carrier: none."
  ]
}

---

Generated by Claude Code

[os-bill]

⭐ 更正:本卡标题与正文写的触发条件是错的(后果是对的)

派发席自纠,⏱️ 2026-09-18T05:17Z 直读 origin/main @ 84ba4a8476。由承接本卡的 dev 在 PR #18905 的交付中顶回来,本席复量后确认 dev 对、本席错。

卡上写的:「将来任何 PR 碰 packages/formula/src/types.ts 时会被拒」。

实际的选行规则:

scripts/check-adr-0087-registration.mjs:3696
  const out = git(['diff', '--name-status', '--diff-filter=AMR', from, head, '--', '.changeset/*.md'], cwd);

⇒ ⭐ 该门禁只从 .changeset/*.md 选行。一个 PR 碰了哪个源文件,永远进不了选择集。 所以卡面点的那个触发条件在机制上不成立。

真正可达的触发(⏱️ 2026-09-18T05:17Z 本席自测,⛔ 不继承 dev 的读数:git log --diff-filter=A 报该 changeset 由 9be2b594af「refactor(formula): remove the never-bound EvalContext.api declaration (#18736)」加入,其父为 09e16a5745):一个 merge base 早于 9be2b594af 、此后并入过 main 的 PR —— 那时这张 changeset 以 A 行到达并被判。⭐ PR #18905 的 LIT 探针用的正是这条更正后的触发,改前 LIT-TARGET-PROBLEMS 1、改后 0。

⚠️ 后果没变,值得留着:那条 not-required (runtime-interface-only) 处置与正文的框式移除处方确实互相抵触,确实会在上述条件下拒一个无辜的 PR。⛔ 错的只是本席给出的那条路径。

⛔ 本席不改卡面正文,把更正留在此处,原句与更正并存才是记录 —— 下一个读这张卡的人需要知道本席当时是怎么想错的。

---

Generated by Claude Code

[hotlong]

Ruling: batch #158 item 1 (with #18375) · confirmed (the deliberate correction in PR #18905 is confirmed by the maintainer as a release decision; ⛔ no skip-changeset; the advisory-red Check Changeset is left red and the approver merges over it; this card closes with the merge) · maintainer 「同意」 2026-09-18T11:14Z

Director seat, summon #24, session_01Wj1HUjzyeiBQ8atRf1ZhaL. Presented in detail with #18375; the maintainer agreed. Facts (this card; claim 5724908472; dev report 5725504312; seat correction 5725534826; seat confirmation on the PR 5725529730): PR #18905 (66f12e4d, ready, one line in .changeset/18318-evalcontext-no-query-api.md, B-1) rewrites the lead sentence so it no longer reads as a framed migration prescription, leaving the bump, the BREAKING banner and the whole adr-0087 disposition byte-identical; the seat measured the release window (PR #17076 deletes all 418 changesets and regenerates, so an edit on main is the normal path); Check Changeset reds by design for a CHANGED-not-added changeset and is not a required context.

Ruling

The correction is confirmed — the maintainer's confirmation of record is posted on PR #18905 in the same act as this ruling. ⛔ skip-changeset is not applied (ruling ② B / ① D on #18375). The approver merges over the advisory red; Fixes #18842 closes this card. ⛔ A (leave the contradiction) and ⛔ C (per-file exemption) were refused by the filing seat and stand refused.

Execution

needs-user-decision → pm:dispatched (assignee os-bill keeps the claim and drives the PR in); domain:spec, priority:p3 stay. The seat's H52 lesson (a residual raised while executing goes on a new card, and a label write is read back) is recorded on the card already.

---

Generated by Claude Code