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

[os-bill]

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:

1. A changeset filename is content-free — plain-donkeys-repeat.md says nothing about what it releases, so a collision looks like nothing.
2. 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.
3. 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.
4. 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

[os-bill]

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