fix(ci): stop one aborted upload from killing wrangler dev

[yashau]

The e2e job died mid-run and passed on a re-run. It was not flaky infrastructure and it was not the OOM killer.

What actually happened

wrangler dev exits on a single failed proxied request, and reports it as ✘ [ERROR] with an empty message.

Every request goes through wrangler's own ProxyWorker. When that inner fetch rejects:

1. the ProxyWorker reports it to the ProxyController;
2. DevEnv.handleErrorEvent finds no matching non-fatal case and re-emits it as a top-level error;
3. the dev command is await events.once(devEnv, "teardown") — and Node's events.once rejects on error, so the command throws and the process exits 1;
4. the cause crossed a Worker boundary, so it is a plain object rather than an Error; castErrorCause wraps it in new Error(), whose message is "".

Hence a fatal exit that names nothing. From the failing run:

[wrangler:info] GET /users/0199e2e0-0000-7000-8000-000000000023 200 OK (94ms)
✘ [ERROR]

# wrangler dev EXITED — code=1 signal=null

code=1 signal=null, no OOM kill recorded, 8 GB free.

The rejection is Network connection lost., raised when a client disconnects while its request body is still uploading. Playwright does that every time it closes a page with a form POST in flight — which is why it landed roughly twenty tests in, and why re-running "fixed" it.

Confirmed by reduction

Five aborted uploads against a bare Worker (no assets, no app code) reproduce the exact CI signature on wrangler 4.121.0, and survive it patched:

	stock	patched
aborted upload	✘ [ERROR] (empty), exit 1, port unbound	Error proxying request to the local Worker: Network connection lost., still serving

Aborting mid-response is survivable; only mid-request-body is fatal.

The fix

patches/wrangler@4.121.0.patch applies upstream's own fix — cloudflare/workers-sdk#15207 for #15203: report the failed request and keep serving. Two call sites, because wrangler bundles the ProxyController twice (its own, and the remote-bindings copy).

Both upstream PRs for this are still open and main is still unfixed, so there is no release to upgrade to. Remove the patch when one carries the fix — a version bump that leaves it here fails the install rather than applying quietly, which is the intended prompt to go and check.

patches/** joins the js paths filter: a patched dependency is application behaviour, and editing it must run the suite that proves it still works.

Testing

• Full suite locally: 100 passed, 1 flaky (an unrelated rbac-ui visibility assertion that passed on retry), wrangler dev alive throughout.
• Targeted repro above, stock vs patched.
• pnpm install --frozen-lockfile reapplies the patch from a clean store; verified 2 patched call sites, 0 fatal ones.
• pnpm fmt:check clean; pre-commit hooks (typos, zizmor, actionlint, lint-js-repo, fmt-js) all pass.

🤖 Generated with Claude Code