A redundant dynamic import crashed every route at module load in the merged bundle

A bundler emits `__exportAll` when something in the graph asks for a whole module as an object instead of named bindings ^[4]. `import * as ns` does that. So does `await import('./thing')` ^[4]. In this codebase the dynamic import named a module the same file already imported statically at the top ^[5], so it split nothing out and deferred nothing; the module was in the bundle and loaded either way ^[6]. The write-up calls the pattern "an ineffective dynamic import: a lazy import of something that isn't lazy, because it's eagerly imported elsewhere in the same bundle" ^[13].

The helper lands in whichever chunk the bundler treats as shared ^[6]. Here it and its call site ended up in `server-XXXX.mjs` and its split twin `server-XXXX2.mjs`, which import each other ^[8]. During evaluation one side reads the binding before the other has finished defining it, and at that moment the binding is still in its temporal dead zone, so the call site gets `undefined` ^[9]. The message is accurate: at that instant `__exportAll` really is not a function. No handler has run yet, so the 500 covers the whole site rather than one endpoint ^[1][3].

"The artifact that breaks is one that nobody built until it was already in production," the author wrote ^[20]. The branch preview built the branch tip, and that build was fine ^[10]. Chunk grouping is decided by the set of modules in the build, so a preview only exercises this class of defect if it builds the commit that will ship ^[1]. That also bounds how far the lesson carries: it needs a pipeline where a second bundling pass re-chunks output whose grouping depends on the full graph, as Nitro's Lambda pass does on Vercel ^[7].

The guardrail is two checks at the end of the build. One walks the source tree and reports any module imported dynamically in one place and statically in another, skipping type-only imports because TypeScript erases them; the author reports it running in milliseconds and naming both call sites with line numbers ^[14]. The other reads the built SSR output as an import graph and fails when a chunk that defines or imports the helper sits on a cycle ^[15]. Calibrating that second one mattered: healthy builds construct namespace objects too, for the framework's own router and server entries, so failing on the helper's presence would fail every build, and the signature to gate on is the cycle ^[16][17].

This is one team's account, tagged vercel, nitro and tanstack ^[18], and the text breaks off where the author starts describing how they verified the cycle signature ^[19]. What shipped as the repair was a single deleted import ^[12].