diff --git a/app/workers/listener.py b/app/workers/listener.py
index eeeb627..abce4a1 100644
--- a/app/workers/listener.py
+++ b/app/workers/listener.py
@@ -3,10 +3,13 @@
 The handler is invoked on every notification AND every sweep_interval seconds,
 so a missed NOTIFY (e.g. connection blip) never wedges the worker. Under
 normal operation the timer fires on an empty queue and is a no-op.
+
+Uses psycopg 3's built-in notifies(timeout=...) instead of wrapping anext()
+in asyncio.wait_for — the wait_for-cancellation path corrupts the async
+generator's state and produces spurious StopAsyncIteration errors.
 """
 
 import asyncio
-import contextlib
 from collections.abc import Awaitable, Callable
 from typing import Any
 
@@ -49,12 +52,15 @@ async def _run_once(
             await cur.execute(f"LISTEN {channel}")
         log.info("listener.subscribed", channel=channel)
 
+        # Drain anything already pending on subscription
         await _safe_call(handler, channel)
 
-        notifies = conn.notifies()
+        # Stream notifications with a built-in timeout. When the timeout elapses
+        # without a notification, the inner loop body simply doesn't run; we
+        # then call the handler (sweep) and re-enter the outer loop.
         while True:
-            with contextlib.suppress(TimeoutError):
-                await asyncio.wait_for(anext(notifies), timeout=sweep_interval)
+            async for _ in conn.notifies(timeout=sweep_interval, stop_after=1):
+                pass  # notification received — handler runs unconditionally below
             await _safe_call(handler, channel)
     finally:
         await conn.close()