svelte-rs

• Byte-exact JS output — ~3,000 cases (1,840 cluster + 1,182 legacy), each diffed against the reference compiler in all four modes (client/server × dev/prod) — 12,000+ passing comparisons.
• Diagnostics parity — 823 cases across 27 categories (a11y, runes, CSS, TypeScript, …), matched against svelte/compiler's own warnings and errors by code, severity, and span.
• Drop-in — same compile() / compileModule() API as svelte/compiler.
• ~9× faster on two anonymized real-world production codebases (8.6–10.2× across every mode, 3.7–4.9× over rsvelte's native binding).
• Ready to try — wired into a fork of vite-plugin-svelte, so it runs in a real Vite app today.

npm i -D @mrwaip/svelte-rs

import { compile } from '@mrwaip/svelte-rs/compiler';

const { js } = compile(
  `<script>let { name } = $props();</script><h1>hello {name}</h1>`,
  { filename: 'Hello.svelte', generate: 'client' }
);

console.log(js.code);

The API mirrors svelte/compiler; see packages/svelte-rs/compiler/index.d.ts. A few opt-in extras beyond the reference API:

• warningFilter: (warning) => boolean — matches Svelte 5's option; drops warnings the predicate rejects.
• suppress: WarningCode[] — a typed list of warning codes dropped at the source. Cheaper than filtering after the fact: suppressed warnings are never built, framed, or serialized.
• transformTypescript: true — transpiles TypeScript features that emit runtime code (enum, namespace, parameter properties, decorators) instead of reporting typescript_invalid_feature. Behaviourally equivalent to vitePreprocess({ script: true }), not byte-equal: esbuild inlines enum member reads and renames shadowed parameters, we don't.
• transformStyle: true (+ loadPaths) — compiles lang="scss" / lang="sass" with grass before the CSS pipeline. Semantically equal to dart-sass on our corpus (verified by lightningcss re-minification), but grass does not promise full dart-sass compatibility — keep a JS preprocessor fallback for unsupported constructs.
• cssTargets: string[] — browserslist queries; lowers modern CSS and adds vendor prefixes via lightningcss, replacing postcss + autoprefixer. Not byte-equal to autoprefixer: the prefix databases disagree (e.g. autoprefixer still emits -moz-user-select where lightningcss does not).
• withDiagnostics: true — returns { js, css, diagnostics } where each diagnostic carries severity: 'error' | 'warning', and never throws on error. Handy for editors, linters, and batch tooling.

Real input → output for every mode lives in tasks/compiler_tests/cluster_cases/: each leaf has case.svelte plus our output and the reference (case-rust.js / case-svelte.js, .dev.js, .server.js, .server.dev.js). The older flat cases2/ suite is legacy but still checked.

The fork of vite-plugin-svelte routes compile / compileModule through @mrwaip/svelte-rs automatically, falling back to svelte/compiler for options the Rust side doesn't support yet.

npm i -D @mrwaip/vite-plugin-svelte

import { defineConfig } from 'vite';
import { svelte } from '@mrwaip/vite-plugin-svelte';

export default defineConfig({
  plugins: [svelte()],
});

Source: https://github.com/MrWaip/vite-plugin-svelte.

Parity target: svelte@5.56.4. Check this before logging an issue.

Whole-corpus throughput on two anonymized real-world production codebases — every .svelte file compiled in every mode, single-threaded, one file at a time. Median of 5 runs, measured 2026-07-27 via just bench-compare <dir> on Apple Silicon (macOS arm64), against svelte@5.56.4, @rsvelte/vite-plugin-svelte-native 0.3.1 (native binding) and @rsvelte/compiler 0.9.4 (WASM).

Both projects are closed-source, so these runs are not reproducible from this repository — just bench-compare against the repo's own fixture corpus is. An explicit directory is scanned as-is, so the file counts include whatever .svelte files the checkout carries, vendored dependencies included.

On project A all four compilers compiled the identical file set (zero skips). On project B rsvelte rejects one file the other three accept (svelte 24 / ours 24 / rsvelte 25 skipped), so its columns there are measured on a set smaller by one file.

Reproduce with just bench-compare. Per-run instruction-count benchmarks (64 total) also run on every commit via CodSpeed. The rsvelte columns appear when its packages are installed. These are rsvelte's serial compile numbers, matching our single-file harness — its parallel compileBatch (rayon) is not measured here.

• rsvelte — another Rust port of the Svelte 5 compiler, also built on OXC. Ships a WASM build (@rsvelte/compiler), a native NAPI binding, and a multi-threaded compileBatch API.

Requires Rust, Node, and just (cargo install just or brew install just).

just playground               # build wasm + serve playground
just quick-check App.svelte   # diff one component against svelte/compiler
just test-compiler            # run the 12,000+ comparison e2e suite (all modes)
just test-diagnostics         # run the 823-case diagnostics parity suite

Standard PRs welcome — no AI required. Day-to-day work uses Claude Code with repo-specific skills under .claude/skills/. Before opening a PR: just test-compiler && just test-diagnostics && just lint must be green.

Bugs and questions: https://github.com/MrWaip/svelte-rs/issues.

• Svelte — the compiler this project mirrors; reference output is the source of truth.
• OXC — JS parser, AST, and codegen.
• vite-plugin-svelte — base of the Vite integration fork.

MIT © Lobkov Constantine