feat: integrate with Vite Task for zero-config build caching#22453
Conversation
b4ddf51 to
48ef6c3
Compare
@vitejs/plugin-legacy
vite
commit: |
07daf26 to
96f992f
Compare
| }, | ||
| "//": "READ CONTRIBUTING.md to understand what to put under deps vs. devDeps!", | ||
| "dependencies": { | ||
| "@voidzero-dev/vite-task-client": "github:voidzero-dev/vite-task#1aedd944&path:/packages/vite-task-client", |
There was a problem hiding this comment.
Temporarily pointed to git for easier debug. Will publish to npm after this PR is approved, before it's merged.
| // alongside inline inherited envs. Outside Vite Task this is a no-op (the | ||
| // client cannot connect). | ||
| for (const prefix of prefixes) { | ||
| getEnvs(`${prefix}*`, { tracked: true }) |
There was a problem hiding this comment.
Effect verified in vite_prefix_env_change_invalidates_cache.md.
| // `outDir`. Without this, those reads would be recorded as build inputs | ||
| // and mix with the writes that follow, tripping Vite Task's read-write | ||
| // overlap check. No-op outside Vite Task. | ||
| ignoreInput(outDir) |
There was a problem hiding this comment.
Effect verified in vite_build_caches_and_restores_outputs.md.
| // The dev server is a long-running, interactive process whose outputs | ||
| // (network responses, HMR updates) cannot be replayed from a cache, so | ||
| // tell Vite Task to skip storing this run. No-op outside Vite Task. | ||
| disableCache() |
There was a problem hiding this comment.
Effect verified in vite_dev_disables_cache.md.
| // Vite Task knows it. Vite's build output branches on the value below | ||
| // (`isProduction`), so changing it between runs must invalidate the | ||
| // cache. No-op outside Vite Task. | ||
| getEnv('NODE_ENV') |
There was a problem hiding this comment.
Effect verified in vite_node_env_change_invalidates_cache.md.
| const processEnv = { ...process.env } as DotenvPopulateInput | ||
| expand({ parsed, processEnv }) |
There was a problem hiding this comment.
The env file may reference a env var and I think those are not tracked now. I don't think we have to try to track them now though.
There was a problem hiding this comment.
You're right. Tracking them would require dotenv-expand to report which variables it referenced. Let me look into it outside of this PR.
When Vite runs under Vite Task (`vp run`), report build inputs, outputs, and tracked env vars to Vite Task so that caching of `vite build` works without manual input/output configuration: - `loadEnv`: ask Vite Task for every env matching each prefix, so the glob + match-set enter the build's cache fingerprint. - optimizer: mark the deps cache dir as neither input nor output (the lockfile hash already drives re-optimization). - `prepareOutDir`: ignore `outDir` as an input so `emptyDir`'s reads don't mix with the writes that follow. - `loadConfigFromBundledFile`: ignore `node_modules/.vite-temp` so the transient bundled-config file doesn't poison the cache. All calls go through `@voidzero-dev/vite-task-client`, which is a no-op when Vite is not running inside Vite Task.
`loadConfigFromBundledFile` previously told Vite Task to ignore the entire `node_modules/.vite-temp/` directory, but only when `nodeModulesDir` existed. When it didn't (e.g. configs outside a project with `node_modules`), the temp file was written next to the original config and still showed up as both an input and an output of the build, poisoning the Vite Task cache fingerprint. Move the `ignoreInput`/`ignoreOutput` calls below `tempFileName` and pass the exact path that's about to be written. The call is a no-op outside Vite Task, so this is safe in either case. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The previous commit dropped `path: /packages/vite-task-client` from the resolution metadata in the lockfile. The URL fragment in the resolution key still encodes the subdir, and pnpm 10.33.2 considers the inner field redundant (it rewrites the lockfile without it on `pnpm install`), but pnpm 10.33.4 — the version used by Vite CI — still needs the field to locate the subdir before deciding whether the package has a `prepare` script to run. Without it, CI fetches the tarball, sees the root `vite-task-monorepo` package's `prepare: husky` script, and aborts with `ERR_PNPM_GIT_DEP_PREPARE_NOT_ALLOWED`. Restore the field so the preview-release CI job (and consumers like the vite-task workspace) can install the tarball. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The build output branches on `process.env.NODE_ENV` (`isProduction`
controls default minification, dev-only warnings, esbuild target,
`import.meta.env.PROD/DEV`, etc.), so a different NODE_ENV produces a
different bundle. Add `getEnv('NODE_ENV')` from
`@voidzero-dev/vite-task-client` before the existing
`!!process.env.NODE_ENV` check in `resolveConfig`. The call asks the
runner for the env (populating `process.env` if Vite Task knows the
value and the local environment hasn't already set it) and registers
NODE_ENV as a tracked dependency of the post-run fingerprint, so
flipping NODE_ENV between runs invalidates the cache with `tracked env
'NODE_ENV' changed`.
No-op outside Vite Task.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The dev server is a long-running, interactive process whose outputs (network responses served on demand, HMR updates, optimizer state) are not replayable from a cache archive. Call `disableCache()` from `@voidzero-dev/vite-task-client` at the top of `_createServer` so the runner skips storing the run even when the user invokes `vp run --cache dev` or wires a dev task with `cache: true`. Without this, the runner would observe inputs, refuse to store on read/write overlap, or worse attempt to replay a useless archive on the next start. No-op outside Vite Task. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The new commit upstream of `@voidzero-dev/vite-task-client` refactors the `.node` addon so it exports a `load()` factory rather than the addon's operations directly, and updates the JS wrapper to `require(addon).load()`. The shape of the wrapper's public API is unchanged, but the bundled wrapper code that vite ships in `dist/` must move in lockstep with the addon shape the runner hands the process — so this PR's preview-release build needs to be regenerated against the new commit before vite-task can refresh its lockfile to consume it. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two reasons the bump can't stay at the abbreviated `1aedd944`: * pnpm 11 (Vite Task's local pin via `mise`, post-`vitejs#383`) refuses to resolve abbreviated GitHub spec SHAs with `[ERROR] Could not resolve 1aedd944 to a commit of …`. Switching to the 40-char form fixes that on both ends. * `137a0003` is the latest commit on the vite-task branch (`runner-aware-tools`) — it bumps the recorded vite-task-client SHA inside `vite-task`'s lockfile and allows `clippy::disallowed_macros` for `napi-derive`'s generated `std::format!`. The wrapper code shipped to vite is identical to `1aedd944`'s, but pointing at the head commit keeps the cross-reference between the two PRs unambiguous. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
`Build&Test: node-20, ubuntu-latest` on `0ed26a2ae` timed out twice in `playground/assets/__tests__/relative-base/assets-relative-base.spec.ts` (navigating to `http://localhost:5174/` after `createServer().listen()`) and `playground/fs-serve/__tests__/base/fs-serve-base.spec.ts` (navigating to `http://vite.dev/404`). The previous commit (`8f6364431`) ran the same wrapper code (the SHA bump in `0ed26a2ae` only swaps `1aedd944` → `137a00036d…` and those tree contents at `packages/vite-task-client/` are identical), and node-22, node-24 ubuntu, node-24 macOS, and node-24 Windows all pass. The retry just gives Playwright another shot on that runner. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The preview server is long-running and interactive, like the dev server, so its run should not be stored in the Vite Task cache. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2924189 to
5b54e07
Compare
`@voidzero-dev/vite-task-client`'s `getEnv`/`getEnvs` no longer mutate `process.env`; they return the value(s) instead. Populate `process.env` here: fall back to the returned `NODE_ENV` in `resolveConfig`, and merge prefixed envs into the resolved env map in `loadEnv`. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
88f2a9c to
614e0d2
Compare
- Adopt the side-effect-free `getEnv`/`getEnvs` shape; the `NODE_ENV`
fetch is now guarded by `process.env.NODE_ENV === undefined` so a
caller-set value is never overwritten.
- Drop the now-default `{ tracked: true }` argument from `getEnvs`.
- Bump vite-task-client to commit c1634234.
- Add the `path:` field to the lockfile `resolution` for the github
tarball; without it pnpm extracts the entire vite-task monorepo
into `node_modules/@voidzero-dev/vite-task-client/`.
- Allowlist `vite-task-monorepo` in `allowBuilds` so pnpm runs its
(now no-op) prepare script during the git-dep flow.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
614e0d2 to
23ded7b
Compare
Lays the groundwork for publishing `@voidzero-dev/vite-task-client` to npm. JS-side only — `src/index.js` (JSDoc as source of truth), `src/index.d.ts` regenerated from it, and the metadata required to publish. No Rust changes; no runner integration. The package is a thin wrapper with no `dependencies`: under a `vp run` task it loads the runner-provided napi addon at runtime, otherwise every method is a graceful no-op. ## Next steps 1. Publish `@voidzero-dev/vite-task-client@0.1.0` to npm. 2. Update the [Vite PR](vitejs/vite#22453) to depend on the published `@voidzero-dev/vite-task-client`. 3. Merge #410 — the remaining Rust implementation (IPC protocol + transport, napi binding, executor refactor, cache integration, e2e tests). Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
sapphi-red
left a comment
There was a problem hiding this comment.
LGTM (after the package is pointing to the published one)
Replaces the in-development git+subpath ref with the freshly published npm version (https://www.npmjs.com/package/@voidzero-dev/vite-task-client/v/0.1.0).
@sapphi-red Thank you for your time! The latest commit points |
…-integration # Conflicts: # pnpm-lock.yaml
9198994 to
bdf9921
Compare
vite reports its transient bundled-config file under node_modules/.vite-temp to the task runner at runtime (ignoreInput/ ignoreOutput via @voidzero-dev/vite-task-client, vitejs/vite#22453), so the synthetic input excludes for pack/test/check are redundant. Keeps the OXLINT_TSGOLINT_PATH env tracking, results.json, !dist/**, and .vite/task-cache excludes unchanged. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
vite reports its transient bundled-config file under node_modules/.vite-temp to the task runner at runtime (ignoreInput/ ignoreOutput via @voidzero-dev/vite-task-client, vitejs/vite#22453), so the synthetic input excludes for pack/test/check are redundant. Keeps the OXLINT_TSGOLINT_PATH env tracking, results.json, !dist/**, and .vite/task-cache excludes unchanged. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The vite-task-client integration (vitejs/vite#22453) has shipped in the tagged vite 8.1 release, so the e2e harness no longer needs to track vitejs/vite main via a pinned pkg.pr.new sha. Move the vite-task-tools>vite override to a normal ^8.1.0 version range. Co-authored-by: Claude <noreply@anthropic.com>
Release vite-plus v0.2.2: the Vite+ Beta release. **Vite+ is now in Beta**: stable and ready for production adoption, fully open source under MIT. Read the announcement to see what Vite+ is about and where it is headed: [Announcing Vite+ Beta](https://voidzero.dev/posts/announcing-vite-plus-beta). On top of the Beta milestone, this release brings cross-version upgrades via `vp migrate`, an official Docker toolchain image on GHCR, zero-config runner-aware `vp build` caching, and PGP-verified managed Node.js downloads. ### Highlights - **`vp migrate` upgrades existing Vite+ projects across versions**: previous release notes told users not to run `vp migrate` for upgrades. It now runs from the global CLI when the local one is older, re-pins `vite-plus` and the `vite` -> `@voidzero-dev/vite-plus-core` alias across dependencies, overrides/resolutions, and catalogs in every workspace package, aligns `vitest` / `@vitest/*` by actual usage, and defaults to a version-only upgrade (pass `--full` to also run the first-time setup bucket: hooks, editor, agent files, lint migration) ([#1891](#1891)), by @fengmk2 - **Official Vite+ Docker toolchain image**: `ghcr.io/voidzero-dev/vite-plus` bundles `vp` plus a native build toolchain on `debian:bookworm-slim` (amd64/arm64, non-root). Since `vp` provisions the exact Node.js from `.node-version`, one image builds any project, and a documented multi-stage build copies the resolved Node.js into a small vp-free runtime stage ([#1944](#1944)), by @fengmk2 - **Zero-config `vp build` caching via runner-aware Vite**: Vite reports its inputs, outputs, and tracked env reads to the `vp` runner over the new `@voidzero-dev/vite-task-client` IPC ([vite#22453](vitejs/vite#22453)), so `vp build` caches correctly with no hand-written cache config: outputs are tracked and restored automatically, and a changed `VITE_*` env var invalidates the cache and is named in the cache-miss message ([#1774](#1774)), by @wan9chi - **PGP-verified Node.js downloads**: installing a managed Node.js now verifies the release's clearsigned `SHASUMS256.txt.asc` against the vendored Node.js release keyring (pure Rust, no `gpg` required) before trusting any checksum, so a tampered archive is rejected before install; unsigned sources (musl builds, custom mirrors) fall back to checksum-only verification ([#1848](#1848)), by @fengmk2 ### Features - `vp check`: a `check` block in `vite.config.ts` (`check.fmt` / `check.lint`) can make plain `vp check` skip formatting or linting by default, mirroring `--no-fmt` / `--no-lint`; standalone `vp fmt` / `vp lint` and git hooks are unaffected, and a `note:` line keeps the config-based skip discoverable ([#1981](#1981)), by @fengmk2 - `vp env list-remote`: highlight installed versions (color, or a `*` prefix when piped) and label the project-resolved `current` and global `default` versions; `--json` gains `installed` / `current` / `default` fields ([#1907](#1907)), by @semimikoh - `vpr` ships as a `vite-plus` package bin, so the `vp run` shorthand works on clean installs without global PATH shims (Vercel build image, generic CI runners) ([#1988](#1988)), by @kvnwolf - Vite Task: `dependsOn` can select tasks from dependency packages, e.g. `dependsOn: [{ "task": "build", "from": "dependencies" }]` ([vite-task#479](voidzero-dev/vite-task#479)), by @wan9chi - Vite Task: a task's `env` / `untrackedEnv` glob patterns support `!` negation (e.g. `["VITE_*", "!VITE_SECRET"]`) ([vite-task#425](voidzero-dev/vite-task#425)), and an env-caused cache miss now names the variable inline, e.g. `cache miss: env 'NODE_ENV' changed` ([vite-task#438](voidzero-dev/vite-task#438)), by @wan9chi - Upgrade upstream dependencies: vite `8.0.16 -> 8.1.2`, rolldown `1.1.1 -> 1.1.4`, oxlint `1.70.0 -> 1.72.0`, oxfmt `0.55.0 -> 0.57.0`, oxlint-tsgolint `0.23.0 -> 0.24.0`, and the oxc toolchain `0.136.0 -> 0.138.0` ([#1924](#1924), [#1989](#1989), [#2000](#2000), [#2009](#2009)), by @voidzero-guard[bot] ### Fixes & Enhancements - Windows: `vp run` no longer hangs CI when a `node_modules/.bin` `.cmd` shim is routed through PowerShell; the npm/pnpm/yarn `.ps1` wrappers read stdin and block forever on a non-TTY pipe, so the PowerShell rewrite is now skipped when stdin is not an interactive terminal ([vite-task#491](voidzero-dev/vite-task#491), via [#1973](#1973)), by @fengmk2 - Vite Task: the task cache is stored in a per-schema-version directory (e.g. `node_modules/.vite/task-cache/v13/`), so switching between branches that pin different Vite+ versions no longer fails with `Unrecognized database version` ([vite-task#433](voidzero-dev/vite-task#433)), by @fengmk2 - Vite Task: env values in cache fingerprints are stored only as SHA-256 digests and env cache-miss details report names without values ([vite-task#455](voidzero-dev/vite-task#455)); prefix env assignments like `PATH=... command` now affect executable lookup during planning ([vite-task#440](voidzero-dev/vite-task#440)); `package.json` / `pnpm-workspace.yaml` files with a UTF-8 BOM parse correctly ([vite-task#424](voidzero-dev/vite-task#424)), by @wan9chi - `vp upgrade`: run the pinned pnpm with a managed Node.js LTS directly instead of re-entering `vp install`, so an incompatible session/project/system runtime can no longer make pnpm skip optional native binaries and leave the upgraded CLI broken ([#1900](#1900)), by @liangmiQwQ - Global package installs: each install writes to an immutable `packages/<name>#<uuid>` prefix that is activated via metadata after npm succeeds, so an interrupted reinstall can no longer leave the active package unavailable ([#1906](#1906)), and stale interrupted-install directories are swept with file-lock protection for concurrent installs ([#1945](#1945)), by @liangmiQwQ - `lazyPlugins()`: skip plugin factories only while config metadata is being resolved instead of keying off `VP_COMMAND`, so builds spawned under `vp run` / `vp exec` keep the user's plugins and `vp format` no longer loads them ([#1939](#1939)), by @fengmk2 - `vp migrate` (pnpm): add a direct `vite` devDep aliased to the core override wherever `vite-plus` is depended on, so vitest's `vite` peer binds to `@voidzero-dev/vite-plus-core` instead of pulling in a second upstream vite that broke the `vp test` cache ([#1933](#1933)), by @fengmk2 - `vp pack`: bundle `@tsdown/exe` and `@tsdown/css` into core so `--exe` and CSS bundling work without a resolvable top-level `tsdown`; the native `lightningcss` becomes an optional peer loaded lazily with an actionable error ([#1919](#1919)), by @fengmk2 - `vp env`: invalidate stale shim resolve cache entries when the project's Node.js version source changes ([#1951](#1951)), by @jong-kyung - Node shim: when the project declares npm via `packageManager` / `devEngines.packageManager`, child processes spawned from `node` resolve the managed npm instead of the Node-bundled one ([#1938](#1938)); `vp env which` reports bins linked by an intercepted `npm install -g` (e.g. `tsc`) instead of "not found" ([#1968](#1968)); bins with uppercase names (e.g. `vitePlus`) dispatch correctly ([#1963](#1963)), by @liangmiQwQ - `vp-setup`: pass the configured npm registry to the inner pnpm install so setup works behind custom registries ([#1795](#1795)), by @daflyinbed - Native binding: declare the platform packages' true ABI floor `engines.node >=20.0.0` so engine-strict package managers (pnpm) no longer skip the optional native dependency and fail with `Cannot find native binding` when a consumer's Node floor lands in a product-policy gap ([#1993](#1993)), by @fengmk2 - `vp create`: run `git init` without creating an initial commit, so commitlint-configured templates no longer reject the hardcoded message and template placeholders are not baked into history ([#2008](#2008)), by @forehalo - `vp staged --debug`: inline the bundled lint-staged version so debug logging no longer crashes reading a `package.json` that does not exist in the bundle ([#1925](#1925)), by @rokuosan - Installer: retry downloads truncated mid-body in `HttpClient::get_bytes` (the platform-tarball path for `vp upgrade` and the standalone installer) ([#1940](#1940)), and clean up the temp dir when a package-manager install fails instead of leaking `.tmpXXXX` directories ([#1949](#1949)), by @shulaoda - Windows/msys: normalize backslashes in the `env.fish` fallback path ([#1954](#1954)), by @Aalivexy - `install.ps1`: detect the missing VC++ runtime (`0xC0000135`) and print VC++ Redistributable guidance instead of a generic failure; interactive `irm | iex` installs keep the shell open ([#1962](#1962)), by @cheezone - `vp migrate`: preserve comments, key order, and trailing commas in existing `.vscode` / `.zed` JSONC configs by patching the original text instead of re-serializing it ([#1956](#1956)), by @fengmk2 - Migration: link the git hook warning to the migration guide ([#1902](#1902)), by @naokihaba - `vp info` / `vp view`: use package-manager-native commands (`pnpm view`, `bun info`, `yarn npm info`) instead of routing every lookup through `npm view` ([#1895](#1895)), by @jong-kyung - Correct overused `ErrorConfig` error types across the codebase ([#1934](#1934)), by @liangmiQwQ ### Refactor - `vite_install`: centralize Yarn v1/berry branching with `is_yarn_berry` ([#1897](#1897)), by @jong-kyung ### Docs - Document Vite Task automatic tracking (fs tracking and cache-reporting tools), reusing the task cache with GitHub Actions cache, and `dependsOn: [{ task, from: "dependencies" }]` ([#1992](#1992)), by @wan9chi - Rewrite the "Upgrading Vite+" guide: preview builds install through the registry bridge as ordinary `0.0.0-commit.<sha>` npm versions, and `vp migrate` is the recommended way to upgrade a project or move it onto a preview build ([#1965](#1965)), by @fengmk2 - Describe how to switch back to the release version from nightly ([#1887](#1887)), by @situ2001 - Clarify Git hook tool migration ([#1901](#1901)), by @naokihaba - Add a global installation explanation ([#1915](#1915)), update the `vp env` help output ([#1969](#1969)), and add liangmiQwQ as a team member ([#1911](#1911)), by @liangmiQwQ - Fix package manager command examples ([#1937](#1937)) and the `dependsOn` guide link ([#1883](#1883)), by @jong-kyung - Remove Fathom analytics from the uninstall docs ([#1946](#1946)), by @mdong1909 - Center the README logo and fix its size ([#1878](#1878)), by @hyf0 ### Chore - Prevent Vite beta upgrades in the upstream-deps script ([#1879](#1879)), stabilize flaky network-dependent tests ([#1923](#1923)), and bump the ecosystem-ci bun-vite-template to the oxlint jsx-a11y fix ([#1898](#1898)), by @fengmk2 - Pin snap-test install fixtures to a published `vite-plus` version so release-branch CI can pass before the new version is published ([#2017](#2017)), by @fengmk2 - Update the deprecated `VITE_PLUS_` env var prefixes to `VP_` in the RFCs ([#1984](#1984)), by @liangmiQwQ - CI: skip full CI for docs-only PRs ([#1991](#1991)) and for the docs deploy config ([#2015](#2015)), by @wan9chi - Update GitHub Actions ([#1904](#1904), [#1977](#1977)), claude-code-action to v1.0.158 ([#1979](#1979)), and pnpm to v10.34.4 ([#1996](#1996)), by @renovate[bot] - Update oxc-project/security-action to v1.0.8 ([#1918](#1918)), by @fengmk2 - Refresh trusted stack stats on the docs homepage ([#1913](#1913), [#1982](#1982)), by @voidzero-guard[bot] ### Bundled Versions | Tool | Version | Source | | --- | --- | --- | | vite | `8.1.2` | [`ba31193`](vitejs/vite@ba31193) | | rolldown | `1.1.4` | [`6cbd233`](rolldown/rolldown@6cbd233) | | tsdown | `0.22.3` | [npm](https://npmx.dev/package/tsdown/v/0.22.3) | | vitest | `4.1.9` | [npm](https://npmx.dev/package/vitest/v/4.1.9) | | oxlint | `1.72.0` | [npm](https://npmx.dev/package/oxlint/v/1.72.0) | | oxlint-tsgolint | `0.24.0` | [npm](https://npmx.dev/package/oxlint-tsgolint/v/0.24.0) | | oxfmt | `0.57.0` | [npm](https://npmx.dev/package/oxfmt/v/0.57.0) | ### Upgrade ```bash vp upgrade ``` New to Vite+? Start with the [Beta announcement](https://voidzero.dev/posts/announcing-vite-plus-beta), then create a project with `vp create` or bring an existing one over with `vp migrate`. ### New Contributors Welcome to our new contributors @rokuosan, @Aalivexy, @cheezone, @daflyinbed, @forehalo, @kvnwolf! 🎉 **Full Changelog**: v0.2.1...v0.2.2 --- Merging this PR will trigger the release workflow. --------- Co-authored-by: voidzero-guard[bot] <278573678+voidzero-guard[bot]@users.noreply.github.com> Co-authored-by: MK <fengmk2@gmail.com>
Integrates Vite with Vite Task so that
vite buildcan be cached with zero configuration when run undervp run: Vite reports its build inputs, outputs, and tracked env vars through@voidzero-dev/vite-task-client, which is a no-op when Vite runs outside Vite Task.Companion PR on the Vite Task side: voidzero-dev/vite-task#346.
See the proposal for background and design rationale.