* feat: support the React Compiler via react: { compiler: true }
@vitejs/plugin-react 6.1.0 added an experimental `compiler` option backed by
oxc-transform-react. vinext already forwards its `react` option to that plugin,
but enabling the compiler broke the build for apps with JSX in plain `.js`
files: `vite:react-compiler` and `vinext:jsx-in-js` both run at `enforce: "pre"`
and the React plugin was registered first, so the compiler parsed those files as
plain JS and failed with `Unexpected JSX expression`.
Register `vinext:jsx-in-js` before `@vitejs/plugin-react`, bump the plugin-react
catalog entry to ^6.1.0, add oxc-transform-react as a dev dependency for the
tests, and document the option.
* fix: fail fast when react.compiler is set on plugin-react 6.0
Addresses review feedback.
@vitejs/plugin-react 6.0 accepts an unknown `compiler` key and drops it, and it
still satisfies vinext's published peer range, so a project that upgrades vinext
without upgrading the plugin would silently get no React Compiler. Probe the
plugin list the factory returns and throw an actionable error when the compiler
plugin is missing.
Also drops the hand-authored changeset: AGENTS.md says changesets are generated
from Conventional Commits in CI.
* fix: keep the compiler version error out of the import failure wrapper
Addresses review feedback.
The trailing `.catch()` wrapped every rejection from the plugin-react promise
chain, including the compiler version check, so the actionable message and its
install command were demoted to a nested `cause` behind
`vinext: Failed to load @vitejs/plugin-react.`. Attach the wrapper to the
dynamic import itself so only genuine import failures are relabeled.
* fix: move only the compiler plugin instead of the whole React group
Addresses review feedback.
Reordering the entire @vitejs/plugin-react group put `vinext:jsx-in-js` ahead of
`vite:react-babel`, so on plugin-react 5.x, which is still in the supported peer
range and does expose a `babel` option, a user-configured Babel transform would
stop seeing JSX in plain `.js` files. Split `vite:react-compiler` out of the
group and register just that plugin after `vinext:jsx-in-js`. Every other plugin
keeps the position it had before this PR.
Also documents that components using `<style jsx>` are not memoized, since
`vinext:styled-jsx` lowers their JSX through the Next.js SWC transform first.
Running the compiler ahead of it is not an option: the styled-jsx plugin selects
modules by matching `<style` in the source, and lowered JSX no longer matches,
which would silently disable styled-jsx entirely.
* fix(ci): account for React Compiler test dependency
---------
Co-authored-by: James <james@eli.cx>
* fix(shims): align public API with vendored Next types
* fix(shims): preserve revalidation and runtime behavior
* test(shims): align cache revalidation expectations
* fix(ci): package types for deploy suite
* fix(scss): preprocess SCSS CSS-module composes deps via SassAwareFileSystemLoader
postcss-modules' built-in FileSystemLoader reads files referenced by
`composes: className from './other.module.scss'` as raw text — without
Sass preprocessing. For .scss files this leaves SCSS variables and bare
@import directives in the CSS output, causing LightningCSS to crash with
"Invalid empty selector" during production builds.
Fix: introduce SassAwareFileSystemLoader that delegates to Vite's
`preprocessCSS` API for every `composes:` dependency, so Sass compilation
runs before CSS-module scoping. Register it via `css.modules.Loader` in
vinext's Vite config hook and thread the resolved config through via a new
`setSassLoaderResolvedConfig` call in the `configResolved` hook.
Fixes the following deploy-suite tests (both sass and sass-embedded variants):
- test/e2e/app-dir/scss/nm-module/nm-module.test.ts
- test/e2e/app-dir/scss/nm-module-nested/nm-module-nested.test.ts
- test/e2e/app-dir/scss/composes-external/composes-external.test.ts
Part of #1825 (non-tilde CSS-module/composes cases; complements #1881)
* fix(scss): address review feedback on SassAwareFileSystemLoader
- Guard the css.modules.Loader injection: respect a user's explicit
css.modules: false (mergeConfigRecursively would otherwise re-enable
CSS Modules with an object override) and never clobber a
user-provided custom Loader. The override is precomputed as a typed
const so the conditional spread doesn't disturb the config literal's
contextual typing.
- Surface composes preprocessing failures via config.logger.warn
instead of a bare catch {}, so missing composed classes are
observable; a missing/unreadable dependency file now rejects like
postcss-modules' FileSystemLoader did.
- Use Vite's exported preprocessCSS + PreprocessCSSResult type via a
top-level import instead of the lazy dynamic import and stringly
typed casts.
- Preserve parity for composes from non-*.module.* files: hand
preprocessCSS a virtual *.module.* filename so plain '.css' deps are
still scoped and export tokens extracted, matching the built-in
FileSystemLoader.
- Port the three failing Next.js fixtures (composes-external,
nm-module, nm-module-nested) as production-build regression tests in
tests/scss-composes.test.ts, plus a plain .module.css / non-module
.css parity test. All three SCSS tests reproduce the original
"[lightningcss minify] Invalid empty selector" crash when the
Loader injection is disabled.
- Add sass as a root devDependency so the SCSS tests (including the
previously always-skipped tests/scss.test.ts) actually run in CI;
drop the now-redundant knip ignore entry. Installing sass makes
Vite's SassPreprocessorOptions resolve to the concrete sass types
(previously any), so the tilde-importer's importers array now
carries an explicit SassPreprocessorOptions["importers"] cast.
* fix(scss): bind composes-loader config per plugin instance; scope extensionless deps
- Replace the module-level resolved-config singleton with a per-vinext-
instance binding (createSassAwareFileSystemLoader), so multiple builds in
one process never preprocess composes deps with another build's config.
- Cover extensionless composes deps in the virtual *.module.* rename
(append .module.css) so their tokens are scoped instead of silently
dropped; add a parity test case.
- Pin url() pass-through behaviour for composed deps in the parity test
(verbatim, never mangled into placeholders — matches the built-in
FileSystemLoader, verified against a vanilla Vite build).
* docs(scss): document the per-subtree recursion boundary of the composes loader
Notes the two structural consequences of delegating nested composes
recursion to preprocessCSS (per-subtree duplicate inlining collapsed by
LightningCSS, and no cross-boundary cycle short-circuit for invalid
circular composes), per review.
* fix(scss): fail the build on composed-dep preprocessing errors
Only a missing Sass implementation is downgraded to a logged warning;
any other preprocessCSS error (e.g. a Sass syntax error in the composed
dependency) now propagates and fails the build, matching postcss-modules'
built-in FileSystemLoader instead of shipping a green build with
silently-missing classes. Docstrings updated to accurately describe the
silent empty-token path when no resolved config is bound. Adds a
regression test for the fail-loudly path.
* chore(deps): move sass to the workspace catalog; drop the now-dead dynamic skip in scss.test.ts
Review feedback on #1882: sass belongs in the pnpm workspace catalog
rather than pinned inline in the root package.json. With sass now a
guaranteed root devDependency, the runtime describeIfSass skip in
tests/scss.test.ts is dead code — the suite always runs — so make the
unskip explicit by removing the conditional machinery.
* rebuild lockfile
* refactor(cache): extract Cloudflare cache adapters into @vinext/cloudflare
Move the Cloudflare KV data cache and edge CDN cache adapters out of
vinext into a new publishable @vinext/cloudflare package:
- cache/kv-data-adapter(.runtime).ts (KVCacheHandler, kvDataAdapter)
- cache/cdn-adapter(.runtime).ts (CloudflareCdnCacheAdapter, cdnAdapter)
tpr.ts stays in vinext. vinext now depends on @vinext/cloudflare
(workspace:*) and the package declares vinext as a peer dep; both build
from source via tsconfig paths so there is no build-order cycle. The
vinext/cloudflare barrel still re-exports KVCacheHandler for back-compat.
Wires up tsconfig paths, a vitest source alias, root build/postinstall,
and the preview/publish workflows for the new package. Updates internal
consumers (apps/web, examples/workers-cache), docs, and tests.
* ci(create-next-app): install @vinext/cloudflare from local tarball
vinext now depends on @vinext/cloudflare, which isn't published to npm
yet. The create-next-app smoke test packs vinext locally and resolves
its deps from the registry, so the install (and dev server) failed with
ERR_PNPM_FETCH_404 for @vinext/cloudflare.
Pack @vinext/cloudflare alongside vinext and add a pnpm override in the
scaffolded project pointing at the local tarball so the dependency
resolves offline.
* refactor(cloudflare): address review feedback
- Remove the root barrel export from @vinext/cloudflare; expose only the
./cache/* subpaths via a wildcard export (no root main/types).
- vinext/cloudflare re-exports KVCacheHandler from the full subpath.
- Drop the redundant .npmignore (the package.json "files" allowlist
already restricts the publish to dist).
- Remove the unsupported imperative setCacheHandler/KVCacheHandler usage
from both READMEs; the cache plugin config is the supported approach.
- Simplify test wiring: drop the now-unused @vinext/cloudflare tsconfig
path and dedupe the vitest source alias into a shared constant.
* chore(cloudflare): drop unused vite devDependency
The @vinext/cloudflare config uses vite-plus and nothing imports vite, so
the vite devDependency was unused. build/check/knip stay green without it.
* Apply suggestion from @james-elicx
* feat(release): commit-driven changesets with auto-generated changesets
Adopt stock Changesets for versioning/changelog/publishing, with the only
bespoke surface being a script that creates changesets automatically from
Conventional Commits. Auto-generated changeset files never live on `main` —
they are written to the CI working tree, consumed by changesets/action into a
rolling Version PR, and discarded. Manually authored changesets still work.
- scripts/create-changeset.mjs: derive per-package bumps from Conventional
Commits (paths -> package, type -> bump). Includes the version-vs-tag guard
so a merged Version PR publishes instead of re-opening a PR.
- scripts/version.mjs: `changeset version` + append a `## Contributors` list to
each bumped package's CHANGELOG.md (idempotent, pure rewrite unit-tested).
- .github/workflows/release-pr.yml: version-only changesets/action (no publish).
- .github/workflows/publish.yml: guarded OIDC publish (preserves
`vp pm publish --provenance`); version now comes from `changeset version`.
- Reconcile packages/vinext version (0.0.5 -> 0.0.55) so the guard is coherent.
Unit tests for both scripts pass (35). The release orchestration (Version PR
creation, OIDC publish, gh contributor resolution) can only be validated in a
live CI run.
* refactor(release): collapse to a single changesets workflow
Let changesets/action own as much as possible. Delete the separate publish.yml
(guard job, manual bump, manual OIDC/tag/release/notify steps) and the
release-pr.yml split. One workflow now:
- create-changeset.mjs writes auto changesets to the working tree (its
version-vs-tag guard yields nothing right after a Version PR merges).
- changesets/action maintains the Version PR and, when no changesets remain,
publishes via `changeset publish` with OIDC trusted publishing + provenance,
and creates the git tag + GitHub Release.
Removes the unused release:version script. Contributors list still handled by
scripts/version.mjs as the action's version command.
* refactor(release): convert release scripts to .mts, drop .d.mts files
Replace the .mjs + hand-written .d.mts declaration pairs with real TypeScript
(.mts) source. Node >=24 (the setup default) runs .mts directly via native type
stripping; .mts is unambiguously ESM so it needs no "type": "module" and emits
no MODULE_TYPELESS warning. The cross-import uses an explicit .mts specifier
(Node requires it), permitted in tsc via allowImportingTsExtensions (safe: the
project is noEmit).
- scripts/create-changeset.{mjs,d.mts} -> scripts/create-changeset.mts
- scripts/version.{mjs,d.mts} -> scripts/version.mts
- tsconfig: allowImportingTsExtensions
- release.yml: run node scripts/*.mts
vp check (format + lint + types) clean; 35/35 unit tests pass.
* refactor(release): trim release scripts (~690→~470 source lines)
Cut comment bloat, remove dead code, tighten without dropping behavior:
- Condense verbose JSDoc/@param blocks to one-line purpose comments; keep the
load-bearing "why" (correctness-rule header, insertContributors idempotency).
- Delete unused `newestChangelogVersion` (+ its tests) — exported/tested but
never called.
- Tighten run()/insertContributors/latestTagVersion without behavior change.
Multi-package machinery and the bottom `## Contributors` list are retained per
requirements. vp check clean; 33 unit tests pass.
* feat(release): grouped conventional changelog + filter bot contributors
The default changesets changelog groups by bump level (### Minor Changes) and
renders our changeset summary as a nested bullet dump. Replace it with a real
conventional-commits changelog: version.mts now regroups each release's commits
into ### Features / ### Bug Fixes / ### Performance sections (scope bolded, type
prefix dropped) and rewrites the newest CHANGELOG section, then appends the
## Contributors list with [bot] accounts filtered out.
- create-changeset.mts: extract conventionalParts + collectReleaseCommits +
releaseRangeStart and export the commit-walk so version.mts reuses it (no
duplicate git logic).
- version.mts: groupedChangelogBody + rewriteReleaseSection (idempotent: only
`## <digit>` is a section boundary); dedupeSortLogins drops `[bot]` logins.
- Tests updated for the new pure builders.
vp check clean; 33 unit tests pass. Verified end-to-end with a local dry run.
* refactor(release): simplify per PR review
- discoverPublishablePackages: scan only packages/* (publishable packages don't
live elsewhere), drop the apps/examples/benchmarks scan.
- Delete isReleaseCommit: the "chore: version packages" release commit is a
`chore`, already excluded by parseBumpFromSubject returning null. No need to
hardcode release-message formats.
vp check clean; 31 unit tests pass.
* docs(release): clarify why the version-vs-tag guard exists
Per PR review: the guard was confusing. Document inline that it only suppresses
the post-merge/pre-publish window (version bumped, tag not yet created), which
would otherwise re-open a Version PR instead of publishing. Keeping the
never-on-main design per review decision.
* feat(release): sub-group changelog areas and humanize area names
Within each type section (### Features / ### Bug Fixes / …), areas (commit
scopes) with more than 3 items now get their own `#### <Area>` sub-group with
the scope prefix dropped; smaller areas and scopeless commits fall under
`#### Other`. When no area qualifies the list stays flat. Area names are
humanized (app-router → "App Router", css → "CSS", ppr → "PPR", i18n → "i18n")
in headings and in the bold prefixes.
Also: groupedChangelogBody now renders only known release types (GROUPS) — it
no longer leaks non-release types into an "Other Changes" bucket.
vp check clean; 32 unit tests pass. Verified with a local dry run.
* feat(release): rename changelog Other sub-group to Misc
* feat(release): sub-group areas with 3+ items (was >3)
* fix(release): address PR review (stale refs, contributor API, doc)
- Fix stale references to deleted files: .mjs → .mts and release-pr.yml →
release.yml in .changeset/README.md and the release.yml header comment.
- resolveContributors: one paginated `gh api compare` call instead of one
`gh api commits/<sha>` per commit (N round-trips → 1). This also removes
version.mts's local git() helper, so no more `fatal:` probe noise from it.
No behavior change to the changelog/contributor output (verified via dry run:
same 4 contributors resolved). vp check clean; 32 unit tests pass.
The bot's "no git tags" bootstrap concern does not apply: tags v0.0.10..v0.0.55
exist and CI checks out with fetch-depth: 0, so the guard resolves 0.0.55.
* fix(release): run pinned changeset CLI; drop dead affectedPackages branch
Per second bonk review:
- Invoke the pinned, installed @changesets/cli via `vp exec changeset` instead
of `vp dlx @changesets/cli` (which fetches a floating latest at run time) — in
release.yml publish and version.mts. Removes the reproducibility gap vs the
lockfile-pinned 2.31.0.
- affectedPackages: remove the unreachable `dir === "."` / empty-prefix branch
now that discovery only returns `packages/*` keys.
vp check clean; 32 unit tests pass.
* fix(release): only emit valid GitHub logins in Contributors; doc fix
Per third bonk review:
- resolveContributors: `.author.login // empty` (was `// .commit.author.name`),
so commits with an unlinked email contribute no entry instead of a raw git
display name. dedupeSortLogins now keeps only `[a-zA-Z0-9-]+` shapes, which
also subsumes the previous `[bot]` filter — no more broken `- @Full Name`
mentions.
- .changeset/README.md: suggest `vp exec changeset` instead of the floating
`vp dlx @changesets/cli` for manual changeset authoring.
vp check clean; 32 unit tests pass; dry run resolves the same valid handles.
* test(msw): introduce MSW infrastructure and migrate font-google-build
Replaces ad-hoc `globalThis.fetch` hijacking in tests with Mock Service
Worker. MSW intercepts both `globalThis.fetch` and `node:http`/`https`,
so it covers fetches issued by in-process Vite servers and fixture pages
without each test having to manage its own stub.
The setup file enables `onUnhandledRequest: 'error'` so any future test
that issues an unmocked external request fails loudly instead of
silently hitting the network. Loopback hostnames (localhost, 127.0.0.1,
::1) are bypassed because the integration project deliberately hits
in-process test servers via fetch.
What this PR contains:
- `tests/_msw/{handlers,server,setup}.ts` — shared infra
- `setupFiles` wired into both vitest projects (unit + integration)
- `msw` + `@mswjs/interceptors` added to the workspace catalog
- `msw` postinstall skipped via `allowBuilds` (it only copies a browser
service worker; we exclusively use `msw/node`)
- `tests/font-google-build.test.ts` migrated as the proof of pattern
Temporary handlers for `httpbin.org/uuid` and
`example.com/not-cacheable` keep the existing fixture pages
(`revalidate-tag-test`, `dynamic-error-fetch`) working under the new
guard. These are placeholders flagged with `TODO(msw-migration)` — a
follow-up will replace those live URLs with test-local endpoints.
This is PR 1 of 3:
- PR 2 will migrate `tests/fetch-cache.test.ts`
- PR 3 will migrate `tests/font-google.test.ts` and the font block in
`tests/app-router.test.ts`
* test(msw): restrict MSW to the unit project + use passthrough for loopback
Two fixes to the initial MSW wiring that surfaced when CI ran the full
suite:
1. The integration project is now excluded from MSW. Integration tests
spin up in-process HTTP servers and fixture dev servers and exercise
them via `fetch("http://127.0.0.1:<port>/...")`. Even with a loopback
`passthrough()` handler installed, the @mswjs/interceptors layer
interferes with that traffic in subtle ways:
- `tests/prerender.test.ts > errors without writing .rsc when the
middleware short-circuit fallback RSC request fails` timed out at
30s — the 5xx response body never reached the caller.
- `tests/ecosystem.test.ts > nuqs` hung in `beforeAll` waiting for
its fixture process to come up, because the readiness fetch
stalled inside the interceptor.
Integration tests already talk to real local servers, not the
internet, so the unhandled-request guard buys little there. The
value of MSW for this repo is mocking external HTTP for unit tests
of fetch wrappers (font-google, font-google-build, fetch-cache),
which is preserved.
2. The `onUnhandledRequest` callback in `setup.ts` that tried to
silently let loopback through wasn't sufficient — MSW's interceptor
still ran for those requests and could stall them. Replaced with a
default `http.all(LOOPBACK_URL_PATTERN, () => passthrough())`
handler in `handlers.ts` and a plain `onUnhandledRequest: "error"`
in `setup.ts`. Unit tests like the middleware-rewrite proxy in
`tests/shims.test.ts` and the JSX-in-JS suite need this to keep
their in-process HTTP servers reachable.
* test(msw): address BigBonk review notes
- Update PR description in the commit body to drop the stale mention of
TODO(msw-migration) handlers — they were removed when the integration
project was excluded from MSW; the fixture pages run there.
- Tighten the loopback IPv4 octet regex from `\d+` to `\d{1,3}` so it
doesn't match absurdly long octet strings. `URL.hostname` normalises
real addresses, so this is purely a readability fix.
* test(msw): address remaining BigBonk notes from the initial review
- Add `0.0.0.0` to the loopback passthrough pattern so test servers that
bind to the unspecified address keep working without per-test mocks.
- Add a `TODO(msw-3)` comment in `server.ts` flagging the
`SetupServerApi` deprecation in MSW 2.14.6 — the `FetchInterceptor`-only
construction will need to migrate to `defineNetwork`'s interceptor
config when we bump to MSW 3.
Both raised by BigBonk on the first review of this PR (the second review
focused on the stale description and the loose `\d+` octet bound, which
were addressed in the previous commit).
* test(msw): scope server.use() inside the try block in font-google-build
Minor structural fix from BigBonk's review: moves the `server.use()`
handler registration inside the `try` block so all setup and the
matching `fs.unlink` cleanup share the same try/finally scope.
No behaviour change — `server.use()` is synchronous and infallible, so
this is purely a structural tidy-up.
* test(msw): correct comment in font-google-build to match interceptor config
BigBonk caught that the comment described the default `setupServer()`
behaviour (which uses both `FetchInterceptor` and `ClientRequestInterceptor`),
but this project's `server.ts` deliberately constructs `SetupServerApi`
with only `FetchInterceptor`. Update the comment to reflect that — it
intercepts `globalThis.fetch`, not `node:http`/`node:https`.
No code change, comment only.
* chore: bump pnpm to 10.32.1 and enable vitest agent reporter
* chore: upgrade GitHub Actions to Node 24 compatible versions
* perf: re-land unit/integration split and add CI sharding
- Restore unit/integration vitest project split (lost in vp migration)
- vitest 4 supports fileParallelism per-project (was global-only in v3),
so unit tests now run in parallel (~11s) while integration stays serial
- Add 3-shard matrix for integration tests in CI, cutting wall time from
~4.5min to ~1.5min
- Move kv-cache-handler to integration to avoid flakiness under parallelism
* chore: remove unused deploy-preview slash command workflow
* ci: add sentinel job for stable branch protection rule
* chore: migrate to vite plus
* Disable typeAware and typeCheck
* Update CI
* Fix CI
* Fix test
* Clean
* Run test with vp
* Try revert
* react: false In test
* Fix test
* Revert "Try revert"
This reverts commit 009da10473.
* Update
* Update
* Try revert ci changes
* revert
* Run vp migrate
* Disable typeAware and typeCheck for now
* Better resolve for test
* Use vp dev instead of vite
* Update expect
* Fix NormalizeManifestModuleId
* Try increase timeout
* Update to use vp
* Try new check
* Bring back npx vp
* Migrate CI
* Make next-intl resolvable
* Update
* Update
* Update
* chore: upgrade vitest to 4.1
* fix: clear console.warn spy state before asserting call counts
Vitest 4 changed vi.spyOn to return the same mock when called on an
already-intercepted function. This means console.warn call records from
earlier tests accumulate in the spy, causing toHaveBeenCalledTimes(2) to
see 4 calls instead of 2.
Adding mockClear() after each spy creation resets the call count so the
assertions only reflect calls made within the test itself.
* add oxfmt formatter: config, scripts, CI, editor setup, docs
* rebuild lockfile
* fix: add Format to required checks list, remove dead ignore pattern
* run fmt
* add format to agents.md again
* Skip deploy previews for fork PRs that lack Cloudflare secrets
Fork PRs don't have access to repository secrets (CLOUDFLARE_API_TOKEN,
CLOUDFLARE_ACCOUNT_ID), so deploy/smoke-test/comment steps always fail.
Add a fork detection condition to skip these steps gracefully. The build
steps still run, so example builds are still validated for fork PRs.
* fix: align Pages Router worker entry with prod-server request handling
The generated Cloudflare Worker entry for Pages Router apps was missing
several request handling steps that the Node.js production server
(prod-server.ts) already handled. This brings the two in sync:
- Run middleware (runMiddleware) before routing
- Apply next.config.js redirects, rewrites (before/after/fallback), and headers
- Handle basePath stripping and trailing slash normalization
- Merge middleware response headers with correct precedence
- Guard renderPage with typeof check
Also updates the two Pages Router examples (pages-router-cloudflare,
realworld-api-rest) and exports vinext/config/config-matchers so the
worker entry can import the shared matching utilities.
* chore: sync pnpm-lock.yaml with upstream package.json changes
João Lucas