Summary

The cloudflare-node wrapper builds the streamed response body from a JS-backed ReadableStream with a manually captured controller, and Writable.write acknowledges chunks immediately without backpressure (controller.desiredSize is never consulted).

On deployed Workers (production runtime, not reproducible with wrangler dev/local workerd) we observed this intermittently stalling mid-stream: the final flush(es) of chunked SSR/RSC responses are never delivered to the client and the terminating chunk is never sent, so the connection stays open indefinitely. Browsers keep waiting on the incomplete response, gradually exhaust their per-origin connection pool, and the whole site eventually freezes for that browser profile (every request stuck in "Stalled", even sign-out).

This PR switches the response body to the Workers-native (C++-backed) IdentityTransformStream and awaits writer.write(), so the runtime pumps the stream itself with real end-to-end backpressure.

Reproduction / evidence

Observed on a production app (Next.js 16.2.10, @opennextjs/cloudflare 1.20.1, wrangler 4.107.0, compatibility_date 2025-12-01, free plan, custom domain). No auth required — a plain RSC request to the login page reproduces it:

$ for i in $(seq 1 20); do curl --http1.1 -s -o /dev/null -m 8 \
    -w "%{time_total}s %{size_download}B\n" -H "RSC:1" https://<app>/login; done
0.14s 8599B      # ok — complete flight payload, stream closed
8.00s 6934B      # HUNG — stalls at the same byte offset every time,
...              #        terminating chunk never arrives (7/20 in this run)

• Hangs affect both HTML documents and RSC payloads (any streamed SSR response); hang rate was 20–35% across runs.
• The stall point is deterministic (always the same byte offset for a given page) — the tail that never arrives corresponds to the last flush(es).
• A hung response stays open at least 180s (test timeout) — the terminating chunk never comes.
• wrangler tail shows no event at all for hung requests (the invocation never completes); completed requests log outcome: ok with 8–22ms CPU, so this is not exceededCpu.
• Static assets (ASSETS binding) never hang (0/30).
• Not reproducible locally via wrangler dev/miniflare (0/30 with the same build), so it appears to be an interaction between the JS-backed stream pump and the production runtime — possibly related to recent workerd stream regressions (e.g. 🐛 Bug Report — Runtime APIs: client disconnect no longer cancels async ReadableStream response body (regressed in 1.20260619.1, persists in 1.20260701.1) cloudflare/workerd#6832, Errored response-body streams terminate cleanly instead of aborting, silently truncating the response cloudflare/workerd#6818), but the wrapper-side pattern makes it invisible: writes are acked without backpressure, so nothing ever notices the pump stalled.

Validation

Measured on the same production deployment, patched adapter only (no other changes):

• pnpm build (tsc) passes, pnpm lint (biome) clean, tests-unit 600/600 pass.
• IdentityTransformStream is declared locally in the file (it is a Workers-specific global); this wrapper only runs on the Workers runtime.

Notes

• writer.write() resolving only when the chunk is accepted also gives the wrapper real backpressure, which the previous implementation lacked (unbounded queue growth in the JS stream).
• Non-Uint8Array chunks are converted with Buffer.from(chunk, encoding) since IdentityTransformStream only accepts bytes.
• The null-body-status fast path is unchanged.

🤖 Generated with Claude Code