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

.github/workflows/ci.yml

@@ -108,6 +108,11 @@ jobs:
               # reached CI without running the Action's own suite.
               - 'scripts/**'
               - 'action/**'
+              # A patched dependency is application behaviour, not metadata:
+              # `patches/wrangler@*.patch` is the only thing keeping `wrangler
+              # dev` alive under the e2e suite. Editing it must run the suite
+              # that proves it still works.
+              - 'patches/**'
             e2e:
               - *js
               - 'e2e/**'
@@ -371,6 +376,12 @@ jobs:
       # The dmesg grep is the other half: a SIGKILL with no exit code is the OOM
       # killer, and that is a different fix from a crash.
       #
+      # It was neither. `code=1 signal=null`, no OOM kill recorded, 8 GB free --
+      # `wrangler dev` exited on ONE failed proxied request, reporting it as
+      # `✘ [ERROR]` with an empty message. That bug is patched; see
+      # `pnpm-workspace.yaml`. This step stays because it is what made the answer
+      # readable from the job page, and the next death will be a different one.
+      #
       # ON PUBLISHING THIS LOG, since `playwright.config.ts` turns screenshots
       # off over exactly this concern: the e2e Worker holds no secret. Its
       # `MASTER_KEY` and every seeded value are committed constants in

e2e/harness/processes.ts

@@ -257,6 +257,17 @@ export async function startDevServer(options: {
    * So record the death where the evidence already is, with the two fields that
    * decide the diagnosis: a `signal` of SIGKILL is the OOM killer, a non-zero
    * `code` is workerd deciding to stop on its own.
+   *
+   * That death had ONE cause, and it is now fixed rather than watched for. A
+   * single proxied request that failed -- `Network connection lost.`, raised
+   * when a client disconnects mid-upload, which is what Playwright does every
+   * time it closes a page while a form POST is in flight -- was fatal to the
+   * whole dev server, and was reported as `✘ [ERROR]` with an empty message.
+   * `pnpm-workspace.yaml` carries the upstream patch and the full explanation.
+   *
+   * This listener stays regardless. It cost nothing, it is what turned seventy
+   * identical `ERR_CONNECTION_REFUSED` failures into one legible line, and the
+   * next reason the Worker dies will not be that one.
    */
   let exited: { code: number | null; signal: NodeJS.Signals | null } | null = null;
   child.once("exit", (code, signal) => {

patches/wrangler@4.121.0.patch

@@ -0,0 +1,28 @@
+diff --git a/wrangler-dist/cli.js b/wrangler-dist/cli.js
+index 9345fc99a8922769799e1a3008feb34d83400b81..893aaf2e91ae6f00a7e1e2e5f6ff21b3bb68a2ee 100644
+--- a/wrangler-dist/cli.js
++++ b/wrangler-dist/cli.js
+@@ -183957,7 +183957,10 @@ var init_dist12 = __esm({
+             this.onPreviewTokenExpired();
+             break;
+           case "error":
+-            this.emitErrorEvent("Error inside ProxyWorker", message.error);
++            logger3.error(
++              "Error proxying request to the local Worker:",
++              message.error.message
++            );
+             break;
+         }
+       }
+@@ -279576,7 +279579,10 @@ var init_ProxyController = __esm({
+             this.emitPreviewTokenExpiredEvent(message.proxyData);
+             break;
+           case "error":
+-            this.emitErrorEvent("Error inside ProxyWorker", message.error);
++            logger2.error(
++              "Error proxying request to the local Worker:",
++              message.error.message
++            );
+             break;
+           case "debug-log":
+             logger2.debug("[ProxyWorker]", ...message.args);

pnpm-workspace.yaml

@@ -107,3 +107,41 @@ catalog:
   vite-plus: 0.2.9
   vitest: 4.1.10
   wrangler: 4.121.0
+
+# -----------------------------------------------------------------------------
+# One upstream bug, patched here until it ships.
+#
+# `wrangler dev` treats a SINGLE failed proxied request as fatal to the whole
+# dev server. Every request the browser makes is forwarded by wrangler's own
+# ProxyWorker to the Worker runtime; if that inner fetch REJECTS, the
+# ProxyWorker reports it to the ProxyController, `DevEnv.handleErrorEvent` finds
+# no matching non-fatal case, and re-emits it as a top-level `error`. 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.
+#
+# It exits SILENTLY. The rejection crosses a Worker boundary, so the cause is a
+# plain object rather than an Error; `castErrorCause` wraps it in `new Error()`,
+# whose message is "". The entire report is `✘ [ERROR]` with nothing after it.
+#
+# The inner fetch rejects with `Network connection lost.` when a client
+# disconnects while its request body is still being uploaded -- which a browser
+# does routinely, e.g. Playwright closing a page while a form POST is in flight.
+# That is a client event, not a server fault, and it killed `e2e` mid-run:
+# `wrangler dev` died about twenty tests in and every remaining spec failed with
+# `ERR_CONNECTION_REFUSED`, three attempts each, none of them naming the cause.
+# Re-running the job "fixed" it because the race simply did not recur.
+#
+# The patch is upstream's own fix, applied to the shipped bundle: report the
+# failure and keep serving. It is two call sites because wrangler bundles the
+# ProxyController twice (its own, and the remote-bindings copy).
+#
+#   issue  https://github.com/cloudflare/workers-sdk/issues/15203
+#   fix    https://github.com/cloudflare/workers-sdk/pull/15207
+#
+# REMOVE THIS -- both the patch file and the entry below -- as soon as a
+# wrangler release contains 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.
+# -----------------------------------------------------------------------------
+patchedDependencies:
+  wrangler@4.121.0: patches/wrangler@4.121.0.patch