🐛 BUG: wrangler dev truncates deferred chunks from streaming ReadableStream responses

[roduyemi]

What versions & operating system are you using?

System:
  OS: macOS 15.1
  CPU: (8) arm64 Apple M1 Pro
  Memory: 104.33 MB / 16.00 GB
  Shell: 5.9 - /bin/zsh
Binaries:
  Node: 24.9.0
  npm: 11.6.0
  pnpm: 10.11.0
npmPackages:
  wrangler: ^4.75.0 => 4.85.0

Please provide a link to a minimal reproduction

https://stackblitz.com/edit/vitejs-vite-gwtpcbsd?file=README.md

Note: wrangler dev is not supported in StackBlitz WebContainers. Clone the repo locally and run npm install && npm run dev to reproduce.

Describe the Bug

wrangler dev truncates deferred chunks of a streaming ReadableStream response. When an outer stream wraps an inner stream and enqueues additional chunks via queueMicrotask after the inner stream ends, most deferred chunks are silently dropped.

Pattern (used by TanStack Router/Start's SSR streaming):

1. An inner ReadableStream sends ~500KB of HTML (simulating React's renderToReadableStream), then closes
2. An outer transform stream wraps it, stays open after the inner stream ends
3. Deferred data is enqueued via queueMicrotask with short delays (simulating deferred GraphQL queries resolving — matches TanStack Router's ScriptBuffer)
4. A stream end marker is enqueued and the outer stream closes

Expected: All chunks delivered.
Actual: Only 1 of 15 deferred chunks arrives. The stream end marker is missing.

npm run dev

# Should print 15, actually prints 1
curl -s http://localhost:8787/ | grep -c 'deferred-'

# Should print 1, actually prints 0
curl -s http://localhost:8787/ | grep -c 'STREAM_END'

Production Cloudflare Workers are NOT affected — streaming works correctly when deployed. This only affects wrangler dev.

This breaks TanStack Start/Router's defer() SSR streaming, where <script> tags are injected into the response after React's initial render stream completes via transformStreamWithRouter. Components that depend on the streamed deferred data stay in their Suspense fallback indefinitely.

Please provide any relevant error logs

No errors are logged. The chunks are silently dropped — the response simply closes before all chunks are delivered.

Related: #8004, #6577

[edmundhung]

Hi, I checked the linked StackBlitz repro, but it appears to be the default Vite React starter. Could you please share a minimal repro that actually runs through wrangler dev or vite dev using @cloudflare/vite-plugin?

Also, we support TanStack Start only through the Cloudflare Vite plugin, so it would be helpful if you can clarify whether the issue is with wrangler dev directly or with vite dev through the plugin.

[roduyemi]

Hi, I checked the linked StackBlitz repro, but it appears to be the default Vite React starter. Could you please share a minimal repro that actually runs through wrangler dev or vite dev using @cloudflare/vite-plugin?

Also, we support TanStack Start only through the Cloudflare Vite plugin, so it would be helpful if you can clarify whether the issue is with wrangler dev directly or with vite dev through the plugin.

Ah sorry, it didn't save! Can you see the minimal reproducer now? https://stackblitz.com/edit/vitejs-vite-gwtpcbsd?file=README.md. Also, yes we use the cloudflare vite plugin. As mentioned, all is fine apart from when using wrangler dev 🙂 but you'll see from the example repo. Thanks again!

[roduyemi]

@edmundhung, can you access the repro now: https://stackblitz.com/edit/vitejs-vite-gwtpcbsd?file=README.md? Thanks.

[edmundhung]

Thanks, I can see the updated repro now, which uses wrangler dev. Could you clarify your actual project setup?

wrangler and @cloudflare/vite-plugin share a lot of internal logic, but there can still be subtle differences in the integration path. I want to make sure we are reproducing the exact issue you are seeing.

[roduyemi]

Thanks, I can see the updated repro now, which uses wrangler dev. Could you clarify your actual project setup?

wrangler and @cloudflare/vite-plugin share a lot of internal logic, but there can still be subtle differences in the integration path. I want to make sure we are reproducing the exact issue you are seeing.

Hi @edmundhung, we use both in our workflow. We use wrangler dev to run our production build for e2e testing and vite development server for local development.

The vite plugin is added as follows:

cloudflare({
    viteEnvironment: { name: 'ssr' },
    configPath: resolve(__dirname, 'wrangler.jsonc'),
}),

We also use the tanstackStart() plugin from @tanstack/react-start/plugin/vite

However, the minimal repro has everything you need to replicate the bug. Thanks for helping to investigate and let me know if you need any more information.

[edmundhung]

I tried the updated repro locally with wrangler@4.85.0, but I'm not sure it is showing the exact issue you're seeing. The response body I get contains all 15 deferred- markers and the STREAM_END marker.

One thing I noticed is that this check can be misleading:

grep -c 'deferred-'

grep -c counts matching lines, not occurrences. The repro writes all deferred scripts on the same line, so this returns 1 even when all 15 markers are present.

Could you try this instead?

curl -s http://localhost:8787/ | grep -o 'deferred-' | wc -l
curl -s http://localhost:8787/ | grep -o 'STREAM_END' | wc -l

Also, this repro sets Content-Encoding: identity, so it doesn't appear to exercise the known compression issue from #6577 or #8004.

[roduyemi]

I tried the updated repro locally with wrangler@4.85.0, but I'm not sure it is showing the exact issue you're seeing. The response body I get contains all 15 deferred- markers and the STREAM_END marker.

One thing I noticed is that this check can be misleading:

grep -c 'deferred-'
grep -c counts matching lines, not occurrences. The repro writes all deferred scripts on the same line, so this returns 1 even when all 15 markers are present.

Could you try this instead?

curl -s http://localhost:8787/ | grep -o 'deferred-' | wc -l
curl -s http://localhost:8787/ | grep -o 'STREAM_END' | wc -l
Also, this repro sets Content-Encoding: identity, so it doesn't appear to exercise the known compression issue from #6577 or #8004.

We've done further investigation and this is not a wrangler bug. Apologies for the false report. We compared the actual HTML from a failing CI run against a working local response:

• Working (local wrangler dev): 14 deferred promises created, 14 resolution scripts present, $_TSR.e() end marker present — 1.1 MB
• Failing (CI/Docker/wrangler dev): 20 deferred promises created, 0 resolution scripts, no $_TSR.e() — 112 KB

Chunks aren't dropped. The resolution scripts are just never injected. We'll keep investigating.

Thanks for taking the time to investigate.

[edmundhung]

Happy to help :) Just let us know if you encounter any issue. I am gonna close the issue for now.