[finding] The `lane` job's declared mirror of `readChangesetState` names one ignored filename; the installed reader ignores four and also reads `.changeset/pre/*.md`

[os-zhuang]

Found while implementing #5442 (npm-presence publish predicate). Out of that card's scope — filing rather than fixing, and #5442 deliberately leaves the detector untouched.

The reading

.github/workflows/changeset-release.yml's Detect pending changesets step carries an explicit, load-bearing declaration:

⛔ This must stay a faithful mirror of readChangesetState in changesets/action's src/index.ts (it re-exports @changesets/read) […]

• every .changeset/*.md counts EXCEPT README.md;

Measured against the copy actually installed in this repository — node_modules/.pnpm/@changesets+read@1.0.0/node_modules/@changesets/read/dist/index.mjs — the reader does two things the mirror does not:

const ignoredMdFiles = [
 /^README\.md$/i,
 "AGENTS.md",
 "CLAUDE.md",
 "GEMINI.md"
];

and, separately, it also pulls in a whole second directory:

changesets.push(...(await fs.readdir(path.join(changesetBase, "pre"))).map((file) => `pre/${file}`));

So the two disagree in both directions:

case	@changesets/read@1.0.0	the mirror in the workflow
.changeset/AGENTS.md	ignored	counted as pending
.changeset/CLAUDE.md	ignored	counted as pending
.changeset/GEMINI.md	ignored	counted as pending
.changeset/ReadMe.MD	ignored (regex is /i)	counted as pending (exact 'README.md' compare)
.changeset/pre/*.md	counted as pending	not seen at all

Confirmed empirically: with a pre/probe.md plus AGENTS.md and CLAUDE.md added to a copy of the real .changeset/, readChangesets returned 161, counting pre/probe.md and ignoring the other two.

Why it is dormant right now, and why that is not reassuring

None of .changeset/AGENTS.md, CLAUDE.md, GEMINI.md, .changeset/pre/ or a non-canonically-cased README exists today, so the disagreement is currently unobservable. The first of those files anyone adds makes it live, and .changeset/AGENTS.md is a plausible thing to add — it is exactly where per-directory agent instructions go in this repo's conventions, and the changesets reader ignores it because people do that.

The blast radius depends on which predicate the release job is on:

• Before 16 of 90 versioned releases never reached npm — a release-PR merge that lands with changesets pending versions again instead of publishing (npm is currently a release behind: repo 17.6.0, npm 17.5.0) #5442 the mirror gated publishing, so a stray .changeset/AGENTS.md would have made pending_changesets permanently true and skipped the release job on every push — silently, and indistinguishably from the defect 16 of 90 versioned releases never reached npm — a release-PR merge that lands with changesets pending versions again instead of publishing (npm is currently a release behind: repo 17.6.0, npm 17.5.0) #5442 was filed for.
• After 16 of 90 versioned releases never reached npm — a release-PR merge that lands with changesets pending versions again instead of publishing (npm is currently a release behind: repo 17.6.0, npm 17.5.0) #5442 publishing is keyed on npm presence, so the mirror no longer decides whether a release exists. It now feeds the clear step, which is a deliberate superset and self-checks, so the disagreement degrades to a cosmetic wrong count in the step summary rather than a lost release.

That is a real reduction in severity, but it does not make the declaration true, and the comment still tells the next reader that one name is the whole exclusion list.

Also worth checking, and not checked here

The mirror is declared against changesets/action's bundled copy, not against this repository's @changesets/read@1.0.0. Those can be different versions with different ignoredMdFiles. The correct fix should establish which copy is authoritative for the mirror before matching it — matching the wrong one would replace a known drift with a confident one.

Not proposing a fix here

There is a real choice to make (restate the list in the comment and the shell; or factor the mirror into one shared script that both the workflow and a unit test can call, which would make it testable for the first time), and it is worth a look at whether the mirror still needs to be exact now that #5442 removed its publish-gating role.

⛔ No workflow was changed by this issue and no release was executed. Filed unassigned.

[claude]

Finding concentration round (triage seat, session session_01Kktexqp6uVuFMztvvTMf3V, 2026-08-24, maintainer-invoked): promoted → pm:queue + domain:devx, Bug — the detect step's declared load-bearing mirror is measurably unfaithful to the installed @changesets/read (three more ignored filenames + the whole .changeset/pre/ directory). Align the mirror with the installed reader and pin it against the dependency so the next @changesets/read bump re-fails honestly.

---

Generated by Claude Code