fix(changesets): stop re-formatting release notes that already shipped

[claude]

Description

changeset version re-printed the whole of CHANGELOG.md through Prettier on every run,
silently rewriting release notes that had already shipped. This sets prettier: false in
.changeset/config.json (option A of the ruling in #1602 comment 5565948014,
maintainer 「同意」), and repairs the one line the generator had already broken.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• CI/CD update

Related Issues

Fixes #1602
Related to #1503 · #1443

Changes Made

• .changeset/config.json: adds "prettier": false. The generated region becomes
  write-once — the generator no longer re-formats what it already wrote.
• CHANGELOG.md: repairs one line of the published ## 3.0.0 section back to the
  text the release actually said. This is the single authorised hand-edit of the
  generated region (ruling item 2).
• .changeset/changeset-version-stops-rewriting-published-notes.md: the changeset.

⚠️ Correction to the ruling's item 2 — the four named lines were not mangled on disk

Ruling item 2 asked for four lines to be repaired: CHANGELOG.md:4176, :4178, :5236,
:5237 on 81a79ee. Re-anchored before editing, they were already correct, and they
still are on this branch's base b760fb08 at the same four line numbers:

4176:     forecast_category map that actually classifies deals") and the English page has
4178:     `stage`, is the single boundary — the same one the _Closing This Quarter_ view
5236:   priced any line is still refused ("Set is_active=false to retire instead"), and
5237:   lines belonging to other deals are never touched. This is deliberately _not_ the

That is the before column of #1602's own table. Those four lines are what the next
run mangles — the card measured them by running changeset version and restoring — so
with prettier: false there is nothing to repair there: they can no longer be reached.
A repo-wide scan for the mangling signature confirms it. One line, and only one, carries
generator damage today:

CHANGELOG.md:1135 (inside ## 3.0.0)
- ... / status / due*date / owner_id / related_to*\*，没有任何重复字段。
+ ... / status / due_date / owner_id / related_to_*，没有任何重复字段。

Proved to be the generator's damage rather than the author's text: the changeset that
produced that paragraph, .changeset/activities-recurring-example.md as authored at
3ec9feba, says due_date / owner_id / related_to_*. The 3.0.0 generation emitted
due*date and related_to*\*. It does not appear in the card's table because it is
already a Prettier fixed point, so a further run leaves it alone — it is damage that has
shipped and stays shipped until repaired by hand.

The repair is squarely inside the ruling's authorisation and its stated reason ("published
release text is never silently rewritten — this repairs a silent rewrite back to what was
published"); only the line number moved. No other line of the generated region is touched.

Testing

Both probes ran on a throwaway worktree, were restored, and the restore is proven by blob
hash. ⛔ Neither run's output is committed.

Before — Prettier on (config sets no key), tree at b760fb08:

$ pnpm changeset version      -> exit 0
$ git diff --numstat CHANGELOG.md
4917    4    CHANGELOG.md
$ git diff -U0 CHANGELOG.md | grep '^-[^-]'
-    forecast_category map that actually classifies deals") and the English page has
-    `stage`, is the single boundary — the same one the _Closing This Quarter_ view
-  priced any line is still refused ("Set is_active=false to retire instead"), and
-  lines belonging to other deals are never touched. This is deliberately _not_ the

and what replaced them, at their new positions in the rewritten file:

9089:     forecast*category map that actually classifies deals") and the English page has
9091:     `stage`, is the single boundary — the same one the \_Closing This Quarter* view
10149:   priced any line is still refused ("Set is*active=false to retire instead"), and
10150:   lines belonging to other deals are never touched. This is deliberately \_not* the

Exactly the four rewrites #1602 tabulated, reproduced independently. Restore:
git hash-object CHANGELOG.md = 877ec8d4… = git rev-parse HEAD:CHANGELOG.md,
git diff HEAD empty.

After — prettier: false, tree at this branch's commit 465dad24:

$ pnpm changeset version      -> exit 0
$ git diff --numstat CHANGELOG.md
4806    0    CHANGELOG.md
$ git diff -U0 CHANGELOG.md | grep '^@@'
@@ -2,0 +3,4806 @@
$ git diff -U0 CHANGELOG.md | grep -c '^-[^-]'
0

One hunk, zero deletions — a pure prepend of the new section under the # Changelog
title. The pre-existing region is byte-identical, not merely diff-clean:

sha256(old file from line 3 on)          = d60fab5699e084561991cc0b37f9b77f9bfbf41985c6b105987e2ab135b29c0c
sha256(new file from line 4809 on)       = d60fab5699e084561991cc0b37f9b77f9bfbf41985c6b105987e2ab135b29c0c
cmp -> identical

Restore: git hash-object CHANGELOG.md = 4e301190… = git rev-parse HEAD:CHANGELOG.md,
git diff HEAD empty, git status clean.

pnpm verify — green, run in two locked calls on the shared box, in the chain's own
order (validate → typecheck → lint → lint:i18n-gate → hygiene → hygiene:tokens, then
build → test):

os-verify-lock: VERDICT command-exit 0 · held the lock 38s
os-verify-lock: VERDICT command-exit 0 · held the lock 238s (3m58s)

 Test Files  161 passed (161)
      Tests  3416 passed | 1 skipped (3417)

• Unit tests pass (pnpm test)
• Linting passes (pnpm lint)
• Build succeeds (pnpm build)
• pnpm validate, pnpm lint:i18n-gate, pnpm hygiene, pnpm hygiene:tokens pass

Ruling item 3 — the stranded preamble at :9007 stays where it is

The ruling permits moving it "if the dev judges it a one-hunk change". It is not one. The
preamble is a single line ~9,000 lines below the title, so moving it is a deletion hunk at
:9007 and an insertion hunk at :2 — two hunks by construction. It also sits inside the
generated ## 3.0.0 region (which ends at :9008, immediately above the
## Pre-3.0.0 hand-written history heading), so moving it would be a second hand-edit
of that region, and the ruling authorises exactly one. Left alone, and said so — the
remaining half of #1443's observation is still open.

Checklist — the changeset, and why 'hotcrm': patch rather than empty frontmatter

There is no scripts/check-changeset-* in this repo; the decision source is
.github/workflows/changeset-check.yml plus the split already in .changeset/ (148
entries: 76 with empty frontmatter, 72 versioned). The documented boundary in the
empty-frontmatter entries is "ships nothing to HotCRM users; no src/ metadata changed".

This one ships something to users: CHANGELOG.md is the artifact HotCRM publishes,
and the repaired line is text a reader of the 3.0.0 notes has already been given. HotCRM
is the reference implementation other marketplace apps fork, so a release note naming
due*date — a column that does not exist — is precisely the error that gets copied.
The release-notes reader is entitled to be told it was wrong and is now right. Hence a
versioned patch entry, matching the repo's other corrections-to-published-prose entries.

• I have added a changeset
• I have performed a self-review of my own code
• My changes generate no new warnings
• New and existing unit tests pass locally with my changes

Additional Notes

⚠️ #1503 / dependabot PR #1264 inherit this. @changesets/cli 3.x replaces the
prettier key with format, so the key this PR sets stops existing at the bump. #1264
must not merge until #1503 sets the off-equivalent under the new name; otherwise this
fix is silently undone and the next release re-mangles the four lines above.

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated
hotcrm	Ignored		Sep 7, 2026 11:34am UTC