Vite 8.3.3: what the 8.3 line changed for Rolldown-era builds

Vite 8.3.3 landed on October 6. On its own it is a four-fix patch, but it closes a month in which the 8.3 line went from a feature release (8.3.0, September 10) through three patches. Read together, the changelog shows where a Rolldown-based Vite still has rough edges: config merging, the experimental bundled dev mode, and the dependency optimizer.

What 8.3.0 added

The September 10 release carried the real features. The ones that matter for application code:

A top-level tsconfig option in the Vite config.
server.watch now accepts Rolldown watch options.
closeServer and closePreviewServer plugin hooks.
Subpath imports (#internal/...) work inside dynamic import() statements.
import.meta.ROLLDOWN_FILE_URL_* is used for assets referenced from JavaScript, and is extended to other plugins.
CSS style tags are minified, and worker search params are preserved.
A config warning for named imports from JSON modules.
--profile [name] on the CLI, so CPU profiles can be named.
A warning when a plugin returned from applyToEnvironment uses hooks that are not supported there.

The last two are the quiet wins for DX. Naming profiles makes before/after comparisons in CI artifacts traceable:

bash

vite build --profile baseline
# change something
vite build --profile after-chunk-split

The JSON warning is worth a look: grep for named imports from .json files, such as import { version } from "./package.json", and see whether your build now warns. The changelog does not say what replacement it recommends.

8.3.1 and 8.3.2: the merge and optimizer fixes

The patches are where the trade-offs show up. Several fixes are about mergeConfig and about Rolldown-specific options:

8.3.1 fixes merging of build.rolldownOptions.output.comments and handles server.ws: false correctly in mergeConfig.
8.3.2 fixes merging of build.rolldownOptions.output.minify.

If you keep a shared base config and extend it per app or per environment, this is the pattern that was affected:

import { defineConfig, mergeConfig } from "vite";

const base = defineConfig({
    build: { rolldownOptions: { output: { minify: true } } },
});

export default mergeConfig(
    base,
    defineConfig({
        build: { rolldownOptions: { output: { minify: false } } },
    }),
);

The changelog only says the merge was corrected; it does not describe the old failure mode. So the practical advice is to check the built output of any app that overrides these options through mergeConfig, rather than assume which direction the bug went.

The optimizer got three notable corrections across the two patches:

It no longer skips imports whose binding name starts with type (8.3.1).
Discovered dependencies that were still pending are processed before initialization (8.3.1).
Excluded optional peer require fallbacks are preserved (8.3.2), and warnings for unsupported browser: false mappings are avoided.

If you saw a missing export or an unexpected full reload after dependency discovery on a 8.3.0 install, these are the first entries to compare against.

Bundled dev mode keeps getting fixes

The changelog refers to a "bundled-dev" environment, which I understand to be the experimental mode from the 8.1 cycle (the Vite blog lists "Announcing Vite 8.1" on June 23; I did not read that post). 8.3.2 contains two fixes specific to it: lazy chunk sourcemaps are now served, and the Rolldown runtime is served from the installed Rolldown package. Treat that as evidence the mode is still maturing. If you opted in, pin the minor version and keep a way to switch back to the default dev server.

renderBuiltUrl and CSS preloading

8.3.2 changed two connected things. CSS preloading is corrected when renderBuiltUrl produces URLs that contain query parameters, and queries are now passed through to renderBuiltUrl. If you serve assets from a CDN and append a cache-busting or signing query in renderBuiltUrl, test a production build with a lazy-loaded route: the preload helper is the path that was affected. 8.3.2 also removed a quadratic link scan in that preload helper, which matters for apps with many preloaded CSS chunks.

Smaller items worth knowing

8.3.2 avoids encoding intermediate source maps, a build performance improvement. No numbers were published, so measure on your own project.
forwardConsole output of objects and arrays is now limited, to avoid excessive logging.
The SSR module runner sourceURL now encodes whitespace.
Worker URLs are aligned between client and server when using terser.
8.3.2 switches the repo from cross-spawn to tinyexec and updates its Vitest monorepo dependency to v5. Both are internal.
8.3.3 checks fs.serve for ?vite-wasm-instance requests, stores safeModulePaths as ids rather than URLs, and drops queries from the filename passed to transformIndexHtml.

That last one can bite plugin authors. If your transformIndexHtml hook parsed a query out of the filename, it will no longer see one:

const plugin = {
    name: "html-tweaks",
    transformIndexHtml(html, ctx) {
        // ctx.filename no longer carries a query string as of 8.3.3
        return html;
    },
};

The changelog states the filename should not include queries; it does not say which code paths used to include them, so grep your plugins for any ctx.filename parsing.

Backports

On October 6 the releases page also lists 8.2.4, 8.1.6, 7.3.7 and 6.4.4. The release notes I could read for 8.3.3 mention no security advisories or backports, and I could not see the contents of the other tags, so I will not guess what they fix. What the list does show is that Vite maintains patches across four older lines at once. If you are on 6.x or 7.x, there are patches available without a major upgrade.

Should you upgrade?

For most projects on 8.3.x, take 8.3.3: it is a patch, and the two fixes that touch config merging and HTML transforms are precisely the kind that are silent until a production build differs. Upgrade path:

bash

pnpm up vite@^8.3.3
pnpm build && pnpm preview

Diff the build output for apps that override build.rolldownOptions.output.* or define renderBuiltUrl. Skip bundled dev mode in production-critical workflows until its fix rate drops.

The caveat: the changelog entries are terse one-liners, so several of the behaviors above are inferred from titles. Read the linked pull requests before relying on any specific edge case.