fix: stop swallowing interrupts forever when an agent won't exit

src/lib/run.ts

@@ -95,16 +95,45 @@ export function runAgent(cmd: string, args: string[]): Promise<number> {
 					})
 				: spawner.spawn(cmd, args, { stdio: "inherit", env });
 
-		// The child shares our process group, so the terminal already delivers
-		// SIGINT/SIGTERM to it. Swallow them in the parent so we don't exit
-		// before the child finishes its own cleanup.
-		const swallow = () => {};
-		process.on("SIGINT", swallow);
-		process.on("SIGTERM", swallow);
+		// The child shares our process group (POSIX) / console (Windows), so the
+		// terminal already delivers SIGINT/SIGTERM to it. Swallow the first one
+		// in the parent so we don't exit before the child finishes its cleanup.
+		//
+		// Never swallow indefinitely, though: an unconditional swallow makes the
+		// hub unkillable whenever the child fails to exit, and the user is left
+		// with a frozen terminal and no way out but closing the window. Windows
+		// is where this bites — the launch chain runs through several cmd.exe
+		// batch shims (our own .cmd shim, npm's codevhub.cmd, the shell:true
+		// wrapper below, npm's codev.cmd), and a batch host that catches a
+		// console break stops at "Terminate batch job (Y/N)?" instead of
+		// exiting. A second interrupt hands the terminal back.
+		//
+		// This does not fire during a normal agent session: interactive agents
+		// hold the terminal in raw mode, where ctrl+c is delivered as input
+		// rather than as a signal, so the count only moves once the terminal is
+		// generating real interrupts.
+		let interrupts = 0;
+		const onInterrupt = () => {
+			interrupts += 1;
+			if (interrupts < 2) return;
+			cleanup();
+			process.stderr.write(
+				`\nStopped waiting for ${label}; it may still be shutting down in the background.\n`,
+			);
+			logWarn(`abandoned ${label} after repeated interrupts`, {
+				action: "process.exit",
+				eventType: "end",
+				outcome: "failure",
+				extra: { agent: cmd, exit_code: 130 },
+			});
+			resolve(130);
+		};
+		process.on("SIGINT", onInterrupt);
+		process.on("SIGTERM", onInterrupt);
 
 		const cleanup = () => {
-			process.off("SIGINT", swallow);
-			process.off("SIGTERM", swallow);
+			process.off("SIGINT", onInterrupt);
+			process.off("SIGTERM", onInterrupt);
 		};
 
 		child.once("error", (err: NodeJS.ErrnoException) => {

tests/lib/run.test.ts

@@ -104,6 +104,80 @@ describe("runAgent", () => {
 		},
 	);
 
+	// A fake child that never emits "exit" stands in for a wedged launch chain
+	// (on Windows, a cmd.exe shim sitting at "Terminate batch job (Y/N)?").
+	// `process.emit` invokes our listeners without asking the OS for a real
+	// signal, so these run identically on Windows.
+	function withWedgedChild(): {
+		child: ChildProcess;
+		restore: () => void;
+		stderr: string[];
+	} {
+		const child = new EventEmitter() as unknown as ChildProcess;
+		const spawnSpy = vi
+			.spyOn(spawner, "spawn")
+			.mockImplementation((() => child) as unknown as typeof spawner.spawn);
+		const original = process.stderr.write.bind(process.stderr);
+		const stderr: string[] = [];
+		process.stderr.write = ((chunk: string | Uint8Array) => {
+			stderr.push(typeof chunk === "string" ? chunk : chunk.toString());
+			return true;
+		}) as typeof process.stderr.write;
+		return {
+			child,
+			stderr,
+			restore: () => {
+				process.stderr.write = original;
+				spawnSpy.mockRestore();
+			},
+		};
+	}
+
+	test("swallows a single interrupt so the child can clean up", async () => {
+		const { child, restore } = withWedgedChild();
+		try {
+			const pending = runAgent("codev", []);
+			process.emit("SIGINT");
+			await new Promise((r) => setImmediate(r));
+			// Still waiting: the child's own exit code wins, not the interrupt.
+			child.emit("exit", 0, null);
+			expect(await pending).toBe(0);
+		} finally {
+			restore();
+		}
+	});
+
+	test("stops waiting for a wedged child after a second interrupt", async () => {
+		const { restore, stderr } = withWedgedChild();
+		try {
+			const pending = runAgent("codev", []);
+			process.emit("SIGINT");
+			await new Promise((r) => setImmediate(r));
+			process.emit("SIGINT");
+			expect(await pending).toBe(130);
+			expect(
+				stderr.some((m) => m.includes("Stopped waiting for CoDev Code")),
+			).toBe(true);
+		} finally {
+			restore();
+		}
+	});
+
+	test("removes its signal listeners once it stops waiting", async () => {
+		const before = process.listenerCount("SIGINT");
+		const { restore } = withWedgedChild();
+		try {
+			const pending = runAgent("codev", []);
+			process.emit("SIGINT");
+			await new Promise((r) => setImmediate(r));
+			process.emit("SIGINT");
+			await pending;
+			expect(process.listenerCount("SIGINT")).toBe(before);
+		} finally {
+			restore();
+		}
+	});
+
 	test("prints a startup banner to stderr before launching", async () => {
 		// vi.spyOn(process.stderr, 'write') doesn't reliably intercept under
 		// vitest's stdio handling, so swap the method directly. stdio:'inherit'