finding: random changeset filenames collide silently across parallel agents — a round overwrote a sibling PR's minor changeset and every gate stayed green

Filed by the domain:spec execution seat, 2026-09-11T16:5xZ, from a NEAR MISS that a round self-caught in #17400 and reported rather than quietly repairing. ⛔ No domain:* and no priority from me.

What happened, measured

A round wrote its changeset with a changesets-style random word-pair filename — .changeset/plain-donkeys-repeat.md. That file already existed on origin/main: it is a sibling PR's minor changeset. The round overwrote it and committed the overwrite.

It was caught by one thing only: git status printed M where an A was expected.

Repaired completely and provably — the sibling file restored from origin/main and shown byte-identical by git hash-object (ee8dfe6a85… == ee8dfe6a85…), the round's own changeset renamed to an issue-scoped name, and the sibling path absent from the final diff vs origin/main (0 matches). ⇒ ⛔ Nothing was actually lost. This card is about the mechanism, not the incident.

⭐ Why this is worth a card

Nothing mechanical would have stopped it. Walk the failure that almost shipped:

A changeset filename is content-free — plain-donkeys-repeat.md says nothing about what it releases, so a collision looks like nothing.
Overwriting it produces a valid changeset file. check:empty-changeset and check-changeset-no-major both still pass — the file parses, it just now says something else.
The sibling PR's release note is silently replaced by this PR's. The sibling stays green; its own CI never re-runs on this change.
The loss surfaces at release time, in the generated CHANGELOG, by which point the authoring PR is merged and the evidence is gone.

⇒ A minor release note for a shipped package could be deleted by an unrelated PR, with every gate green on both sides. That is a release-integrity hazard, not a working-style preference.

⚠️ And the exposure is structural here: this repo runs many agents in parallel, each creating changesets, all drawing from the same small random-word space. The collision rate is not hypothetical — it happened today on the first round that used a random name.

The cheap mechanical fix (⛔ a direction, not a prescription)

Every other round today named its changeset issue-scoped — .changeset/17187-…md, .changeset/16875-…md — and those cannot collide by construction. Options a taker might weigh:

A — a gate refusing a changeset filename that is not issue-scoped (or refusing a diff whose changeset entry is M rather than A). ⭐ The M-not-A shape is exactly what caught this one, so it is already known to be detectable.
B — an os-dev rule in .claude/agents/os-dev.md mandating the issue-scoped name. Cheaper, but prose, so it binds only rounds that read it.
C — nothing: treat it as working discipline.

⚠️ A and B are not equivalent, and the difference is the whole point of this card: B is the same class of protection that just failed. The round that collided was following its dispatch faithfully — the dispatch simply had not told it to scope the name. Another prose rule protects only the rounds that happen to be told.

⛔ I am not choosing: a new gate is gate strength, which is the maintainer's floor.

An adjacent hazard from the same round, recorded not filed

On this container's shallow clone, git log -- PATH attributes every file to the graft-boundary commit dea214f818, so path provenance is unusable without --unshallow. That round noticed and declined to rely on it. ⛔ Not a repo defect — a property of the checkout — but it is the kind of thing that turns into a false "this file has not changed since" reading.

Duplicate check — completed to the limit this session allows

⛔ REST /search/* is refused for this session and MCP search_issues is rate-limited. Completed: enumeration of 569 open issues grepped by TITLE — random 0, overwrite 0, plain-donkeys 0; changeset 6 and collision 3, each a different subject on inspection (#17296 an ADR-0049 question, #17281 a retirement's trailing edges, #17229 a false gate line, #17026 stale claims inside a queued changeset, #14290 a dispatch-gates traversal gap). Lit control spec = 75. ⚠️ Titles only; bodies were not searched. Close as duplicate if one surfaces.

Related: #17400 / PR #17711 (where it happened and was repaired) · #17026 (a different way a changeset's content goes wrong after it is written)

Generated by Claude Code

Population measured — the exposure is 81% of the directory, and my own dispatch rule introduced a THIRD spelling

Follow-up by the filing seat, 2026-09-12T01:30Z, on origin/main @ 9bd4344e4b. ⚠️ This corrects the scale I implied when I filed this card.

The census

git ls-tree --name-only origin/main .changeset/ | grep '\.md$' | grep -v README     → 209 changesets

shape	count	example
`<card>-<slug>.md` — card-scoped, the prevailing scoped form	39	`15052-search-fields-docblock-icontains.md`
`issue-<card>-<slug>.md`	0	— (⚠️ see below)
no card number at all — descriptive or generated	170	`analytics-compareto-kind-refusal.md`, `adr-0112-envelope-refusal-declaration.md`

⇒ ⭐ 170 of 209 (81%) carry no issue scope whatsoever. When I filed this card I framed the hazard as 「随机 changeset 文件名并行碰撞」 — a random two-word default. That understates it: the collision surface is not the occasional generated name, it is the majority of the directory, including hand-written descriptive names, which collide exactly as readily when two agents pick the same topic word.

⚠️ And I introduced a second scoped spelling today

My dispatches this round mandated .changeset/issue-<n>-<slug>.md. Measured just now: zero pre-existing entries use that form, while 39 use the bare <n>-<slug> form. ⇒ the mandate was followed as written by every round that received it (PR #17740's own report flagged the divergence, correctly, as 「not any of the three filing classes」), but it means the tree now grows two scoped conventions instead of converging on one.

⛔ Not a defect — both are collision-free and both are issue-scoped, which is the property this card is about. ⭐ But the fix, when someone writes it, should adopt the prevailing <n>-<slug> spelling, ⛔ not mine. I will use that spelling in subsequent dispatch orders.

What no gate does

grep -n 'isChangesetFile' scripts/check-changeset-no-major.mjs
  → const isChangesetFile = (p) => p.startsWith('.changeset/') && p.endsWith('.md') && !p.endsWith('/README.md')

⇒ the only filename predicate in the changeset gates is prefix + extension. Neither check-changeset-no-major.mjs nor check-changeset-fixed.mjs reads the name for scope, so ⛔ nothing today can notice two agents writing the same filename — which is exactly how the original collision passed with both PRs green.

⚠️ Declared limit: I measured the filename population and the filename predicates. ⛔ I did not measure how many of the 170 unscoped names were generated by changeset add versus hand-written, and that distinction may matter to whoever picks a fix.

_{domain:spec execution seat · session_01MkQhmuuJAVDjmeWNixwDDH · readings taken 2026-09-12T01:30Z on origin/main @ 9bd4344e4b}

Generated by Claude Code

Ruling recorded — A′ + B: a gate refuses a PR that MODIFIES a `.changeset/*.md` that exists on the base and is not the PR's own; new changesets are named issue-scoped by rule, with no gate on names (director seat, decision batch #130 item 3, 2026-09-13)

Maintainer, verbatim (live PM chat, reply received 2026-09-13T07:1xZ), to decision batch #130 presented as 9308 B · 1A′ · 2A′ · 3(A′+B): 「同意」.

Derived first from the long-term axis: changesets' random word-pair names were designed for one human running the CLI at a time; this repository runs many agents in parallel drawing from the same small name space, and the collision happened on the first round that used a random name. The durable protection is mechanical and content-blind — "you may not change someone else's release note" — which is exactly the M-not-A signal that caught this one; the naming rule makes new collisions impossible by construction without touching the 171 legacy files. B alone is the class of protection that just failed, so A′ is not optional. ②③④ do not overturn ①.

What is ruled

A′ — gate (scripts/, wired into the changeset gate family): a PR whose diff MODIFIES or DELETES a .changeset/*.md that exists on the merge base and was not added by this PR is refused by name, with the remedy "rename yours; restore theirs from base". Definition of "added by this PR" is the file's absence on the merge base — re-evaluated against the current base in the merge queue, so a base move cannot make a PR's own file read as foreign. A PR editing its own changeset across commits is unaffected. Self-test covers: foreign M (red), foreign D (red), own M (green), new A (green).
B — rule: new changesets are named <issue>-<slug>.md; one line in the dev-round guidance (governed), carried by the skills PR objectstack#17748 is already assembling, ⛔ not its own card and ⛔ not a gate.
Legacy names are untouched; the gate looks at diff shape only.

State

needs-user-decision → pm:queue; bug / domain:devx / priority:p2 kept (the gate is the landing point); the B line is noted on #17748.

Generated by Claude Code

Claim: PM loop round R11
Session: session_012GKcPZbMoGq7WPzKLfRBTU
Branch: claude/issue-17712-foreign-changeset-guard
Worktree: objectstack-issue-17712
Domain: domain:devx
File surface: scripts/check-changeset-*.mjs(优先扩写既有家族成员);⚠️ 若确需新文件才动 .github/workflows/** (stop on breach; explain in the report)
Container & model: M, mode:subagent, model: default (opus)
Clause-②: no
Thread-read: 5651909449
Serial constraints cleared: none

_{本卡由 domain:devx 执行 PM 席派发;assignee 与本 Claim: 由 PM 写,承接的 dev ⛔ 不再补第二条 claim、⛔ 不动 assignee。}

Generated by Claude Code

fix(scripts): refuse a PR that modifies or deletes a changeset it did…

交班状态标注 —— `domain:devx` 执行 PM 席停席

_{session_012GKcPZbMoGq7WPzKLfRBTU · 2026-09-14T06:0xZ · 维护者指令:转交给其它 session。交班简报在座位贴 #6023。}

⚠️ 本卡的 dev 是本会话的子进程,随本会话一同终止。 下面是停席那一刻的读数;⛔ 分支上若有晚于该时刻的提交,那是一个已经死掉的 dev 推的最后一笔,⛔ 不要当作"还有人在做"。

状态:分支已推,⚠️ 还没有 PR

分支 claude/issue-17712-foreign-changeset-guard @ eeb4012a(2026-09-14T05:59:44Z)
一个提交:fix(scripts): refuse a PR that modifies or deletes a changeset it did not add (#17712)
diff:只有 scripts/check-empty-changeset.mjs(+469 / -6) ⇒ dev 采纳了简报的偏好,扩写既有门禁家族成员、零 .github/workflows/** diff ⇒ 这张卡的 PR 是可以被 PM 武装的(不会撞 HTTP 422)

⚠️ ⛔ 本席既未收到 dev 报告,也未复核过那 469 行。 接手席位必须自己复核,重点是裁决(总监 5651909449,批次 #130 项 3,维护者「同意」)点名的四个自测用例是否都在、且真的跑到:foreign M → 红、foreign D → 红、own M → 绿、new A → 绿;以及 "added by this PR" 是否定义为在 merge base 上不存在、且对队列里的当前 base 重新求值。

⛔ 范围提醒:只做 A′ 门禁。裁决的 B 项(新 changeset 命名规则)是治理面的一行,由 #17748 承载,⛔ 不属本卡;⛔ 也不许改任何历史 changeset 文件名(裁决第 3 项:legacy 不动,门禁只看 diff 形状)。

给接手席位的规矩

本席的 Claim: 留在卡上作为历史,⛔ 不表示本席仍在占卡。接手席位可以接这张卡:按本仓规矩,发你自己的 Claim:(含独立成行的 Branch: 与 Clause-②:),⛔ 不要改本席那条,也 ⛔ 不要沿用本席的分支名以外的旧假设。
⛔ 别重做已经做完的那部分 —— 上面写清了分支上已有什么。

Generated by Claude Code

状态收口:标签改回 `pm:queue`,assignee 已清 —— 接手席位可直接认领

_{session_012GKcPZbMoGq7WPzKLfRBTU · 2026-09-14T06:1xZ}

承上一条(5659821268)。本席停席,本卡没有在席 PM、也没有活着的 dev,继续挂 pm:dispatched 会把一个半态说成"正在做"。⇒ 剥回 pm:queue 并清掉 assignee。

分支状态复测未变:claude/issue-17712-foreign-changeset-guard @ eeb4012a,一笔提交,只改 scripts/check-empty-changeset.mjs(+469 / -6),无 PR。

⛔ 接手人注意:上一条里的复核要点与范围边界(只做 A′;B 在 #17748;legacy 名不动)仍然成立,⛔ 不要因为标签变了就当作没做过的新卡。

Generated by Claude Code

维护者裁决 —— D(本轮 `AskUserQuestion` 在场通道,2026-09-14T08:06Z)

_{domain:devx 执行席 session_017ef78bLdybu3AffehKkhfk 录裁 · 承本卡评论 5660507653 呈报的 A/B/C/D}

维护者选定项,逐字:

D — 只改补救语(推荐)
门禁强度一分不减,只把拒绝时那句话改对:点名「你的改动可能让这条待发布说明变成假的」,并指向已有的人工确认路径。不新增标签、不放宽接受集,一次 revert 可回退。四棱同向 D。

裁定的作用范围

✅ 改 scripts/check-empty-changeset.mjs 中 scanForeign() 的拒绝文案与 ::error 注解文案。
⛔ 不动判定逻辑:--diff-filter=MD、--no-renames、merge-base 语义、changeset-release/main 豁免、skip-changeset 的整 job 豁免,一律不动。
⛔ 不新增标签(否决 B)、⛔ 不按 diff 形状放行(否决 C)、⛔ 不新增任何接受集或公开面。
⇒ 这是拉回已声明契约方向之外的纯文案修正,Clause-②: no。

裁决所依据的实测(本席取数,已在 `5660507653` 与 #18146 评论 `5660165662` 留档)

30 天窗口(scripts/pm/git-history.mjs 判定完整,exit 0;population 429,对照 --diff-filter=A = 431):9 个 PR 会被新门禁拦,4 个带 skip-changeset 实际豁免,5 个硬拦(#18011 · #17845 · #17709 · #16750 · #16636)。承重实例 ed7243d52:该 PR 落地 boolean 聚合支持,并同笔改对了描述该行为的待发布说明;按现文案 "rename yours; restore theirs from base" 执行,会把一条已被该 PR 改成假的发布说明恢复回去。

⛔ 诚实边界:本席实测了这 5 个会被拒,并读了其中 2 个的 diff;未逐个裁定 5 个都正当。裁决 D 修的是补救语的适配,不是给这 5 个发通行证。

承接

门禁已在 main 生效(0ffb4963e),⇒ 这条错误指引从现在起每命中一次就误导一次。本席据此立卡并派发实施,卡号随后追评在此。

Generated by Claude Code

⚠️ 更正:本席早前在本线程贴的普查数字是错的 —— 真实规模大约三倍

_{domain:devx 执行席 session_017ef78bLdybu3AffehKkhfk · 更正读数 2026-09-14T09:04Z}

本席早前在本线程报的「30 天窗口 429 个 changeset 提交 / 9 个 foreign / 5 个硬红」是错的。#18160 的 dev 证否了它,本席独立重测,确认 dev 对、本席错。

git-history.mjs count --days=30 --path=.changeset --ref=origin/main → 2020
git-history.mjs count --days=7  --path=.changeset --ref=origin/main →  443   ← 本席报的 429 是这一档

	原报(错)	真 30 天窗口
changeset-touching 提交	429	2020
foreign `M`/`D` 提交	9	28
Version Packages(结构性豁免)	—	4
可用 `skip-changeset` 豁免	4	10
无法豁免 ⇒ 硬红	5	14

发火对照:同窗口 --diff-filter=A 的 changeset 行 2070。

结论方向不变,只是更强

裁决 D(只改补救语)不受影响 —— 被那句错误补救语误导的不是每月 5 个 PR,而是 14 个。承重实例 ed7243d52 也不受影响,dev 用直接读取(git cat-file -e ed7243d52^:.changeset/aggregate-field-type-compatibility.md 成功)替换了本席的推断,结论相同。

根因

本席第一次跑的是 git-history.mjs log 而非 count,而那一次调用同时在给浅检出做 deepen。工具 exit 0,但本席收到的输出是截断的。⚠️ 征兆当场可见却没追:末行是 afe1c4e0a1794a5681method: git log … —— 提交 hash 从中间断掉、method: 行黏在其后。

⇒ ⛔ exit 0 不保证你收到的行是全部,浅检出上触发 deepen 的那一次调用尤其如此。先 count 定量级,再 log 取清单。

⛔ 本更正不改本卡任何标签与状态。

Generated by Claude Code

fix(scripts): refuse a PR that modifies or deletes a changeset it did…

fix(scripts): the foreign-changeset refusal names both classes, not j…

fix(metadata): close out the adr-0030-notification-event retirement r…

fix(changeset, spec): cut the envelope spans the lowering refusal mak…

pm-dispatch: three landings the seat decides on its own record (rulin…

docs(changeset): the pending callerParamKeys note no longer lists the…

fix(plugin-security, objectql): a row-level check holds for every row…

fix(spec): the edge-condition upgrade entry and its pending changeset…

fix(formula, driver-mongodb): refuse == / != against a list literal a…

fix(cli)!: retire `os generate schema` — it refuses, and points at `o…

fix(service-analytics): an inferred cube lives only in its request — …

fix(plugin-security,driver-sql,driver-turso): lower type-blind at the…

fix(service-analytics)!: refuse a JSON-stored dimension or count_dist…

Navigation Menu

Repository navigation

finding: random changeset filenames collide silently across parallel agents — a round overwrote a sibling PR's `minor` changeset and every gate stayed green #17712

What happened, measured

⭐ Why this is worth a card

The cheap mechanical fix (⛔ a direction, not a prescription)

An adjacent hazard from the same round, recorded not filed

Duplicate check — completed to the limit this session allows

Population measured — the exposure is 81% of the directory, and my own dispatch rule introduced a THIRD spelling

The census

⚠️ And I introduced a second scoped spelling today

What no gate does

Ruling recorded — A′ + B: a gate refuses a PR that MODIFIES a `.changeset/*.md` that exists on the base and is not the PR's own; new changesets are named issue-scoped by rule, with no gate on names (director seat, decision batch #130 item 3, 2026-09-13)

What is ruled

State

交班状态标注 —— `domain:devx` 执行 PM 席停席

状态:分支已推,⚠️ 还没有 PR

给接手席位的规矩

状态收口:标签改回 `pm:queue`,assignee 已清 —— 接手席位可直接认领

8 remaining items

维护者裁决 —— D(本轮 `AskUserQuestion` 在场通道,2026-09-14T08:06Z)

裁定的作用范围

裁决所依据的实测(本席取数,已在 `5660507653` 与 #18146 评论 `5660165662` 留档)

承接

⚠️ 更正:本席早前在本线程贴的普查数字是错的 —— 真实规模大约三倍

结论方向不变,只是更强

根因

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

finding: random changeset filenames collide silently across parallel agents — a round overwrote a sibling PR's minor changeset and every gate stayed green #17712

Description

What happened, measured

⭐ Why this is worth a card

The cheap mechanical fix (⛔ a direction, not a prescription)

An adjacent hazard from the same round, recorded not filed

Duplicate check — completed to the limit this session allows

Activity

os-bill commented on Sep 12, 2026

Population measured — the exposure is 81% of the directory, and my own dispatch rule introduced a THIRD spelling

The census

⚠️ And I introduced a second scoped spelling today

What no gate does

os-tesla commented on Sep 13, 2026

Ruling recorded — A′ + B: a gate refuses a PR that MODIFIES a .changeset/*.md that exists on the base and is not the PR's own; new changesets are named issue-scoped by rule, with no gate on names (director seat, decision batch #130 item 3, 2026-09-13)

What is ruled

State

claude commented on Sep 14, 2026

claude commented on Sep 14, 2026

交班状态标注 —— domain:devx 执行 PM 席停席

状态:分支已推,⚠️ 还没有 PR

给接手席位的规矩

claude commented on Sep 14, 2026

状态收口:标签改回 pm:queue,assignee 已清 —— 接手席位可直接认领

8 remaining items

claude commented on Sep 14, 2026

维护者裁决 —— D(本轮 AskUserQuestion 在场通道,2026-09-14T08:06Z)

裁定的作用范围

裁决所依据的实测(本席取数,已在 5660507653 与 #18146 评论 5660165662 留档)

承接

claude commented on Sep 14, 2026

⚠️ 更正:本席早前在本线程贴的普查数字是错的 —— 真实规模大约三倍

结论方向不变,只是更强

根因

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

finding: random changeset filenames collide silently across parallel agents — a round overwrote a sibling PR's `minor` changeset and every gate stayed green #17712

Ruling recorded — A′ + B: a gate refuses a PR that MODIFIES a `.changeset/*.md` that exists on the base and is not the PR's own; new changesets are named issue-scoped by rule, with no gate on names (director seat, decision batch #130 item 3, 2026-09-13)

交班状态标注 —— `domain:devx` 执行 PM 席停席

状态收口:标签改回 `pm:queue`,assignee 已清 —— 接手席位可直接认领

维护者裁决 —— D(本轮 `AskUserQuestion` 在场通道,2026-09-14T08:06Z)

裁决所依据的实测(本席取数,已在 `5660507653` 与 #18146 评论 `5660165662` 留档)