diff --git a/packages/coding-agent/src/modes/rpc/rpc-mode.ts b/packages/coding-agent/src/modes/rpc/rpc-mode.ts
index 0eb081d72..0d1d1e884 100644
--- a/packages/coding-agent/src/modes/rpc/rpc-mode.ts
+++ b/packages/coding-agent/src/modes/rpc/rpc-mode.ts
@@ -181,10 +181,17 @@ async function drainQueuedSfFeedbackCommands(cwd: string): Promise<void> {
 }
 
 function scheduleQueuedSfFeedbackDrain(cwd: string): void {
-	const timer = setTimeout(() => {
+	// Intentionally NOT unref'd: the drain must keep the process alive until it
+	// finishes. In a long-running rpc-mode child this is moot — the process has
+	// plenty of other handles. In a transient drainer subprocess (packaged-
+	// standalone path) the setTimeout was the only event-loop handle, so an
+	// unref'd timer let Node exit before the drain ran, stranding the renamed
+	// .draining file forever (sf-mpa6wuhm-wwddd1). The setTimeout(…, 0) is
+	// still desirable so the RPC response goes back to the caller first; we
+	// just don't want the timer to be release-eligible.
+	setTimeout(() => {
 		void drainQueuedSfFeedbackCommands(cwd);
 	}, 0);
-	timer.unref?.();
 }
 
 async function captureProcessWrites<T>(