[wrangler] fix: remove stale --experimental-vm-modules flag and prevent silent exits

.changeset/fix-node-v26-silent-exit.md

@@ -0,0 +1,11 @@
+---
+"wrangler": patch
+---
+
+Fix silent exit with no error message on Node.js v26
+
+`wrangler` was silently exiting with code 1 on Node.js v26 with no error message shown. This release fixes two independent issues that caused this behaviour:
+
+1. A stale Node.js flag that caused unexpected behaviour on Node.js v26 has been removed.
+
+2. If an error occurs in a situation where the normal error reporting path itself fails, `wrangler` now always prints the original error to stderr so the cause is visible rather than silently disappearing.

packages/miniflare/test/plugins/browser/index.spec.ts

@@ -34,7 +34,7 @@ async function sendMessage(ws: WebSocket, message: unknown) {
 		const chunks: Uint8Array[] = [firstChunk];
 		for (
 			let i = FIRST_CHUNK_DATA_SIZE;
-			i < data.length;
+			i < encodedUint8Array.length;
 			i += MAX_MESSAGE_SIZE
 		) {
 			chunks.push(encodedUint8Array.slice(i, i + MAX_MESSAGE_SIZE));
@@ -63,6 +63,7 @@ async function waitForClosedConnection(ws: WebSocket): Promise<void> {
 }
 
 const BROWSER_RENDERING_RETRY = {
+	timeout: 120_000,
 	retry: {
 		condition: /Chrome readiness probe .* timed out|Test timed out/i,
 		count: 3,
@@ -83,7 +84,7 @@ export default {
 
 // We need to run browser rendering tests in a serial manner to avoid a race condition installing the browser.
 // We set the timeout quite high here as one of these tests will need to download the Chrome headless browser.
-describe.sequential("browser rendering", { timeout: 20_000 }, () => {
+describe.sequential("browser rendering", { timeout: 120_000 }, () => {
 	// The CLI spinner outputs to stdout, so we mute it during tests
 	beforeEach(() => {
 		vi.spyOn(process.stdout, "write").mockImplementation(() => true);

packages/wrangler/AGENTS.md

@@ -11,7 +11,7 @@ Main CLI for Cloudflare Workers. ~2k-line yargs command tree in `src/index.ts`.
 - `src/` — CLI source
 - `src/__tests__/` — Unit tests, helpers in `src/__tests__/helpers/`
 - `e2e/` — E2E tests, requires Cloudflare credentials
-- `bin/wrangler.js` — Shim that spawns Node with `--experimental-vm-modules`
+- `bin/wrangler.js` — Shim that spawns Node to run `wrangler-dist/cli.js`, forwarding stdio and IPC
 - `bin/cf-wrangler.js` — `cf-wrangler` delegate entrypoint. Owns verb dispatch and argv parsing (`parseCfWranglerArgs`, `parseCfWranglerBuildArgs`); hands off to `runCfWranglerDev` / `runCfWranglerBuild` from `wrangler-dist/cli.js` in-process (no re-spawn — the parent tool owns the Node runtime)
 - `src/cf-wrangler/` — The `cf-wrangler` delegate entrypoint (see below)
 - `templates/` — Worker templates

packages/wrangler/bin/wrangler.js

@@ -25,7 +25,6 @@ Consider using a Node.js version manager such as https://volta.sh/ or https://gi
 		process.execPath,
 		[
 			"--no-warnings",
-			"--experimental-vm-modules",
 			...process.execArgv,
 			path.join(__dirname, "../wrangler-dist/cli.js"),
 			...process.argv.slice(2),

packages/wrangler/src/index.ts

@@ -2419,7 +2419,20 @@ export async function main(argv: string[]): Promise<void> {
 			if (dispatcher) {
 				dispatchGenericCommandErrorEvent(dispatcher, startTime, e);
 			}
-			await handleError(e, configArgs, argv);
+			try {
+				await handleError(e, configArgs, argv);
+			} catch (handleErrorErr) {
+				// handleError itself threw before it could log the error.
+				// Fall back to raw stderr so the user always sees something.
+				const message = e instanceof Error ? (e.stack ?? e.message) : String(e);
+				const handlerMessage =
+					handleErrorErr instanceof Error
+						? (handleErrorErr.stack ?? handleErrorErr.message)
+						: String(handleErrorErr);
+				process.stderr.write(
+					`\n${message}${handlerMessage ? `\n\n(error handler also failed: ${handlerMessage})` : ""}\n`
+				);
+			}
 			throw e;
 		}
 	} finally {