agent/sessions: keep channel sticky to active sessions on hang

nerve/agent/sessions.py

@@ -205,9 +205,9 @@ async def get_active_session(
     ) -> str:
         """Get or create the active session for a channel.
 
-        If the channel has a mapped session with activity within the sticky
-        period, reuse it. Otherwise, create a fresh session and map the
-        channel to it.
+        Reuses the channel's mapped session when it is currently active
+        (a turn is in flight) or when its last activity falls within the
+        sticky period. Otherwise creates a fresh session and remaps.
         """
         row = await self.db.get_channel_session(channel_key)
         if row:
@@ -222,7 +222,25 @@ async def get_active_session(
         return session_id
 
     def _is_within_sticky_period(self, session: dict) -> bool:
-        """Check if a session had activity within the sticky period."""
+        """Check whether a session is still the channel's owner.
+
+        Active sessions are always sticky regardless of the timestamp.
+        A turn that hangs never reaches mark_active() at engine.run's
+        end, so last_activity_at freezes at turn-start; without this
+        carve-out, a hang lasting longer than sticky_period_minutes
+        would orphan the session and route the user's follow-up
+        message into a fresh, empty one. The engine's idle-message
+        timeout in receive_response (cli_idle_timeout_seconds) bounds
+        how long a truly hung session can hold the channel before it
+        flips to idle/error and the next message can mint a fresh
+        session.
+
+        Idle sessions fall back to the time-based check so channels
+        that have been quiet for longer than sticky_period_minutes
+        roll over to a new session as before.
+        """
+        if session.get("status") == SessionStatus.ACTIVE.value:
+            return True
         ts = session.get("last_activity_at") or session.get("updated_at")
         if not ts:
             return False

tests/test_sessions.py

@@ -127,6 +127,55 @@ async def test_auto_session_rotated_after_sticky_period(self, sm: SessionManager
         sid2 = await sm.get_active_session("telegram:222", source="telegram")
         assert sid1 != sid2
 
+    async def test_auto_session_reused_when_active_despite_old_timestamp(
+        self, sm: SessionManager, db: Database,
+    ):
+        """An active session keeps the channel even if last_activity_at is stale.
+
+        A hung turn never reaches mark_active() at engine.run's end, so
+        last_activity_at freezes at turn-start. Without the active-status
+        carve-out in _is_within_sticky_period, a hang lasting longer than
+        sticky_period_minutes would orphan the session and route the
+        user's follow-up message into a fresh, empty one.
+        """
+        sid1 = await sm.get_active_session("telegram:333", source="telegram")
+        # Mark active and back-date timestamps to look like a hung turn that
+        # started long before the sticky-period cutoff.
+        await sm.mark_active(sid1, sdk_session_id="sdk-stuck")
+        await db.update_session_fields(sid1, {
+            "last_activity_at": "2020-01-01T00:00:00+00:00",
+        })
+        await db.db.execute(
+            "UPDATE sessions SET updated_at = '2020-01-01T00:00:00' WHERE id = ?",
+            (sid1,),
+        )
+        await db.db.commit()
+        sid2 = await sm.get_active_session("telegram:333", source="telegram")
+        assert sid1 == sid2
+
+    async def test_auto_session_rotated_when_idle_after_sticky_period(
+        self, sm: SessionManager, db: Database,
+    ):
+        """Idle sessions still roll over after the sticky period.
+
+        Once a hung session has been recovered (status flipped to idle by
+        the engine's exception path), the time-based cutoff applies again
+        and a new follow-up message mints a fresh session.
+        """
+        sid1 = await sm.get_active_session("telegram:444", source="telegram")
+        await sm.mark_active(sid1, sdk_session_id="sdk-x")
+        await sm.mark_idle(sid1)
+        await db.update_session_fields(sid1, {
+            "last_activity_at": "2020-01-01T00:00:00+00:00",
+        })
+        await db.db.execute(
+            "UPDATE sessions SET updated_at = '2020-01-01T00:00:00' WHERE id = ?",
+            (sid1,),
+        )
+        await db.db.commit()
+        sid2 = await sm.get_active_session("telegram:444", source="telegram")
+        assert sid1 != sid2
+
     async def test_set_and_get_active_session(self, sm: SessionManager, db: Database):
         await sm.get_or_create("ch-1")
         await sm.set_active_session("telegram:123", "ch-1")