Reproduction link or steps

Repo (starting point, see "partial repro" note below):
https://github.com/thedoublejay/rolldown-chunk-split-cjs-repro

Open in Stackblitz:
https://stackblitz.com/github/thedoublejay/rolldown-chunk-split-cjs-repro

git clone https://github.com/thedoublejay/rolldown-chunk-split-cjs-repro
cd rolldown-chunk-split-cjs-repro
npm install
npm run build       # vite 8.0.13, rolldown 1.0.1
npm run preview     # browse http://localhost:4173

The repo sets up the same conditions where a real-world Vite 8.0.13 build broke:

• 3 × React.lazy(() => import(...)) boundaries, all importing the same lodash/<sub> CJS submodules (forces rolldown to extract them to a shared vendor chunk)
• ESM libs (@mui/material, @tanstack/react-query, clsx, @emotion/react) mixed with the CJS modules in the shared chunk, matching the real failing chunk's shape
• tss-react/mui withStyles in a lazy component (matches the legacy component that surfaced the bug)
• vite.config.js mirrors the failing config's distinctive flags: react({ jsxImportSource: "@emotion/react" }) + rolldown.moduleTypes: { ".js": "jsx" }

Important — partial repro

This standalone repo does not yet trigger the broken emission on its own — rolldown emits a consistent interop shape here. The same Vite/rolldown versions + the same lodash subpaths + the same lazy boundary structure DO reproduce in a larger codebase (~5000 modules / ~1200-source vendor chunks). The trigger appears to depend on graph shape we couldn't isolate down to a small case.

I'm filing this rather than waiting because the production-side evidence below is concrete (minified output + bisect-narrowed regression range), and a maintainer hint at the chunking heuristic that changed between rolldown 1.0.0-rc.18 and 1.0.0 / 1.0.1 would let me push the repo over the edge. Happy to invest more time on shrinking the case.

What is expected?

When rolldown splits a CJS module into a separate output chunk, the import-site code and the export-side emission should agree on the interop shape:

• Either both sides treat the export as a __commonJSMin(...) lazy-wrapper function, or
• both sides treat the export as a static namespace value.

This is what rolldown 1.0.0-rc.18 (shipped in Vite 8.0.11) does, and what the same source produces correctly there.

What is actually happening?

In rolldown 1.0.0 / 1.0.1, when a CJS module ends up in a separate output chunk (not the main bundle), the importer and exporter disagree:

• The import-site emits a(c()) — i.e. calls c as a __commonJSMin() lazy-wrapper, then passes the result to _interopRequireDefault.
• The export-side emits the resolved module namespace directly as the export, not wrapped in a callable getter.

Result: TypeError: <minified> is not a function thrown synchronously at top-level of the lazy chunk (before any of its export bindings evaluate), which surfaces as a React.lazy rejection caught by the nearest ErrorBoundary.

Real-world emitted code (anonymized)

// Importing chunk (lazy)
import { Vt as t, fi as n, p as r, q as i, vi as a } from "./main.js";
import { n as s } from "./shared-A.chunk.js";
import { T as c } from "./shared-B.chunk.js";   // ← shared CJS+ESM vendor chunk

var l = a(n()),     // ← works: n() returns the __commonJSMin wrapper from main
    u = a(i()),     // ← works: i() returns the __commonJSMin wrapper from main
    d = a(c())      // ← throws: c is not callable, exporter emitted a value

Browser console

TypeError: c is not a function
    at app.<hash>.chunk.js:1:272
Caused by: React ErrorBoundary TypeError: c is not a function
    at Lazy (<anonymous>)
    at Suspense (<anonymous>)

Reproducible with both [email protected] and [email protected], both @vitejs/[email protected] and @vitejs/[email protected], so neither of those is the variable.

System Info

System:
  OS: macOS 26.5
  CPU: arm64 Apple Silicon
Binaries:
  Node: 24.14.1
  npm: 11.11.0
  Yarn: 4.13.0
npmPackages (broken configuration):
  vite: 8.0.13
  rolldown: 1.0.1 (transitive via vite)
  @vitejs/plugin-react: 6.0.2
  react / react-dom: 19.2.x

Any additional comments?

Bisect: which Vite/rolldown versions are affected?

Bisect was carried out against the real-world failing app: same source, same install, only the Vite/rolldown version changed.

Workaround in use

Pin Vite to exact 8.0.11.

Related

• #9407 — distinct bug, same rolldown 1.0.1 family: require("react") left as runtime __require even when react is inlined elsewhere in the same build. Different surface (SSR + dangling __require) but same theme of CJS-interop emission inconsistency across the bundle graph.
• #7449 — incorrect execution order with CJS + advancedChunks (different mechanism, but also CJS chunking-related).

Triage hint

If a maintainer can hint at what specifically changed in chunk emission between rolldown 1.0.0-rc.18 and 1.0.0 / 1.0.1 for shared CJS modules — particularly when a vendor chunk mixes ESM and __commonJSMin-wrapped CJS modules — I should be able to take the linked repo over the edge into a clean standalone repro.