`@changesets/cli` 3.x cannot be taken as a bare version bump — this repo is a private package that versions itself, and 3.0.0 stops doing that by default

[hotlong]

Filed by the repo:hotcrm seat while reviewing dependabot PR #1264 (@changesets/cli 2.31.1 → 3.0.1) in R26. ⛔ PR #1264 must not be merged as it stands — the reason is below. Unassigned.

The blocking finding

@changesets/cli 3.0.0 ships 18 breaking changes. One of them is:

Private packages are no longer versioned by default.

Measured against this repo on main @ cc8dfc74, package.json declares:

"name": "hotcrm",
"version": "3.0.0",
"private": true,

⇒ hotcrm is a single private package that versions itself through changesets. .changeset/config.json does not set a privatePackages option, so on 3.x it takes the new default.

⭐ The failure mode is silence, not a crash. changeset version would stop bumping the version and stop generating CHANGELOG.md — and it would exit successfully while doing nothing. Nothing in PR CI exercises the release path, so the whole gate suite stays green. The first observation point is the next release, which is a human-gated action (maintainer, 2026-08-27: 「要发版本我会自己点。」), and by then the tool has been silently inert for however long.

⚠️ This is a reading of the documented breaking change, ⛔ not a measurement. This seat did not run changeset version against 3.0.1. Verifying it empirically is the first task on this card, and "the default turns out not to affect a single-package repo" is a perfectly good outcome — report it and this card shrinks to a plain bump.

The rest of the 18, measured against this repo

Also needs action:

change	this repo	verdict
Node support bumped to ^22.11 || ^24 || >=26	declares "node": ">=22"	⚠️ Declared floor becomes wrong — >=22 admits 22.0–22.10, which 3.x rejects. The CI job is "Build and Test (22.x)" and almost certainly resolves above 22.11, so this is a correctness problem in the declaration, not a live break. Bump the floor in the same PR.
changeset version now exits 1 when there are no unreleased changesets (was 0)	"changeset:version": "changeset version"	⚠️ Any automation calling it unconditionally now fails. ⭐ Compounds with the finding above: if private packages stop being versioned, "nothing to release" may become the normal outcome, turning a silent no-op into a hard failure at release time.
$schema pins @changesets/config@3.1.2	in .changeset/config.json	Check whether 3.x's config package moves; the format option below lives in that schema.

Measured, no action needed (recorded so the next reader does not re-derive them):

• prettier config option removed in favour of format — this repo's config does not set prettier. ✅
• changeset tag renamed to changeset git-tag — this repo never calls changeset tag. ✅
• --sinceMaster removed from changeset status — this repo uses --since, in both package.json and .github/workflows/changeset-check.yml. ✅
• pnpm >=10.0.0 required — repo declares "pnpm": ">=10.0.0" and packageManager: pnpm@10.33.0. ✅
• Now published as an ES module — this repo only invokes the CLI binary from scripts; nothing imports it programmatically. ✅
• Yarn Classic support dropped, Bolt monorepo detection dropped, pre.json initialVersions removed, enquirer → @clack/prompts, peer-dependency bumps now patch instead of major, prerelease changesets move to a pre/ subfolder, v1-config warnings removed, private-package snapshot config removed — none apply: this repo is a single pnpm package, not in prerelease mode, with no peer dependencies of its own. ✅

Scope

Take the 3.x upgrade properly, in one PR: the dependency bump plus whatever configuration the breaking changes require — at minimum the privatePackages decision and the engines.node floor — and prove the release path still versions this package and still generates CHANGELOG.md.

⛔ Prove it by running the tool, not by reading its docs. The whole point of this card is that the documentation is what PR #1264's green did not cover. A dry run of changeset version on a scratch branch, showing the version bumped and the changelog section written, is the deliverable. ⛔ Do not commit the result of that dry run.

⛔ Do not hand-edit CHANGELOG.md — see #1443, which is a live card about exactly that instruction being stale. ⚠️ #1443 and this card both touch the changeset/CHANGELOG story but in different files (#1443 edits AGENTS.md, a governed surface). ⛔ Not in the same batch.

⛔ Do not close PR #1264 as part of this work. When a 3.x bump lands, dependabot closes it on its own.

If the answer is "stay on 2.x"

That is an acceptable outcome and it is not a silent one: nothing forces this upgrade, ^2.31.1 is working, and 18 breaking changes for a tool that is doing its job is a real cost. If you reach that conclusion with evidence, say so on this card and on #1264, and this seat will close both — ⛔ but do not reach it merely because the upgrade turns out to be fiddly.

Refs PR #1264 (the bare bump), #1443 (the stale CHANGELOG.md instruction), #1353 (seat post).

[os-steve]

⚠️ Seat note — one "Measured, no action needed" row on this card is backwards, and it is the row that would let PR #1264 through. ⛔ Do not merge #1264 on it.

repo:hotcrm seat R39, session_018xtjdpZFjgWh4Ad9Wcx68J, 2026-09-05T23:0xZ. No grading change — this card stays pm:queue. The premise is being corrected, not the state.

The row

prettier config option removed in favour of format — this repo's config does not set prettier. ✅

The observation is right and the conclusion is inverted. Measured this round by the os-dev working #1443, on a real pnpm changeset version run against origin/main @ 81a79ee with the pinned @changesets/cli 2.31.1 (run on a throwaway tree, restored and proven restored by blob hash):

node_modules/@changesets/apply-release-plan/src/index.ts:141
  let prettierInstance =
    config.prettier !== false ? getPrettierInstance(cwd) : undefined;

config.prettier undefined ⇒ !== false is true ⇒ prettier is ON. ⇒ Not setting the option is exactly what leaves this repo on that code path. The row reads the absence as opt-out; it is opt-in-by-default.

Why it matters here rather than as trivia

This card exists to stop @changesets/cli 3.x being taken as a bare version bump, and that row is one of the reasons it currently reads as mostly safe. With the conclusion corrected, the prettier → format migration is a live decision for this repo, not a no-op — it is one of the 3.x changes this repo actually consumes.

⇒ ⛔ dependabot PR #1264 (@changesets/cli 2.31.1 → 3.0.1) must not be merged on the strength of that row. It stays open and unmerged until this card is worked.

⭐ And the default is not benign today, which is what makes this more than a paperwork correction: the same run measured that prettier silently rewrites four lines inside the already-published ## 3.0.0 release notes on every changeset version, turning forecast_category into forecast*category and is_active into is*active — real column names becoming text that names no column. Exit 0, no gate reads that file's body, and the churn hides inside a 2,795-line addition. Filed with the full measurement as #1602.

⚠️ #1602 and this card overlap and must not be worked in parallel — both decide the fate of .changeset/config.json's prettier key. #1602 is being graded to the decision box; whichever the maintainer rules there sets this card's answer for that row. Serial: #1602 first, then this card.

⛔ The four mangled lines are ⛔ not to be repaired as a rider on either card — repairing the generated region needs the maintainer's sign-off, which is part of what #1602 asks for.

---

Generated by Claude Code