[eric] sessions: stopping, closing or deleting a chat stops every session it spawned, not only its browser agents; a SpawnAgent tree outlived a killed parent (Haik, exp.9)

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

backend/apps/agents/manager/SessionControl.py

@@ -21,13 +21,9 @@ from backend.apps.agents.manager.AgentManagerProtocol import AgentManagerProtoco
 class SessionControl(AgentManagerProtocol):
     @typechecked
     async def stop_agent(self, session_id: str):
-        """Stop a running agent and all its browser-agent children."""
-        # Stop children first so browser agents get cancelled before parent
-        children = [
-            s for s in self.sessions.values()
-            if s.parent_session_id == session_id and s.mode == "browser-agent"
-        ]
-        for child in children:
+        """Stop a running agent and every session it spawned, leaves first."""
+        from backend.apps.agents.manager.session.descendants import children_of
+        for child in children_of(self.sessions, session_id):
             await self.stop_agent(child.id)
 
         session = self.sessions.get(session_id)

backend/apps/agents/manager/session/SessionLifecycle.py

@@ -39,12 +39,9 @@ class SessionLifecycle(AgentManagerProtocol):
     @typechecked
     async def close_session(self, session_id: str) -> None:
         """Close a session: pause the agent if running, persist to JSON file,
-        and remove from in-memory state. Also stops browser-agent children."""
-        children = [
-            s for s in self.sessions.values()
-            if s.parent_session_id == session_id and s.mode == "browser-agent"
-        ]
-        for child in children:
+        and remove from in-memory state. Also stops every session it spawned."""
+        from backend.apps.agents.manager.session.descendants import children_of
+        for child in children_of(self.sessions, session_id):
             await self.stop_agent(child.id)
 
         task = self.tasks.get(session_id)
@@ -112,14 +109,11 @@ class SessionLifecycle(AgentManagerProtocol):
     @typechecked
     async def delete_session(self, session_id: str) -> None:
         """Permanently delete a session: remove from memory and JSON file.
-        Also stops browser-agent children first."""
+        Also stops every session it spawned first."""
         from backend.apps.agents.core.flight_recorder import drop_session
         drop_session(session_id)
-        children = [
-            s for s in self.sessions.values()
-            if s.parent_session_id == session_id and s.mode == "browser-agent"
-        ]
-        for child in children:
+        from backend.apps.agents.manager.session.descendants import children_of
+        for child in children_of(self.sessions, session_id):
             await self.stop_agent(child.id)
 
         task = self.tasks.get(session_id)

backend/apps/agents/manager/session/descendants.py

@@ -0,0 +1,13 @@
+"""Every session born from a parent, whatever kind it is. Stop, close and delete used to walk only
+browser-agent children, so a SpawnAgent tree outlived the parent that was killed (Haik, exp.9)."""
+
+from typing import Dict, List
+
+from typeguard import typechecked
+
+from backend.apps.agents.core.models import AgentSession
+
+
+@typechecked
+def children_of(sessions: Dict[str, AgentSession], parent_id: str) -> List[AgentSession]:
+    return [s for s in sessions.values() if s.parent_session_id == parent_id]

backend/tests/test_stop_cascades_to_every_descendant.py

@@ -0,0 +1,62 @@
+"""Haik, exp.9: killing a parent left its SpawnAgent children running. Stop, close and delete walked
+only browser-agent children, so a sub-agent tree (and anything under it) outlived the parent."""
+
+import asyncio
+import inspect
+
+from backend.apps.agents.agent_manager import agent_manager
+from backend.apps.agents.core.models import AgentSession
+from backend.apps.agents.manager import SessionControl
+from backend.apps.agents.manager.session import SessionLifecycle, descendants
+
+
+def p_tree() -> tuple[AgentSession, AgentSession, AgentSession, AgentSession]:
+    parent = AgentSession(name="parent", model="opus-4-8", status="running")
+    child = AgentSession(name="child", model="opus-4-8", status="running", mode="sub-agent", parent_session_id=parent.id)
+    grandchild = AgentSession(name="grandchild", model="opus-4-8", status="running", mode="browser-agent", parent_session_id=child.id)
+    stranger = AgentSession(name="stranger", model="opus-4-8", status="running", mode="sub-agent")
+    for s in (parent, child, grandchild, stranger):
+        agent_manager.sessions[s.id] = s
+    return parent, child, grandchild, stranger
+
+
+def p_park(session_id: str) -> asyncio.Task:
+    async def forever() -> None:
+        await asyncio.sleep(3600)
+    task = asyncio.get_running_loop().create_task(forever())
+    agent_manager.tasks[session_id] = task
+    return task
+
+
+def test_stopping_the_parent_stops_the_whole_tree_and_nobody_else():
+    parent, child, grandchild, stranger = p_tree()
+
+    async def run() -> None:
+        tasks = [p_park(s.id) for s in (parent, child, grandchild, stranger)]
+        await agent_manager.stop_agent(parent.id)
+        await asyncio.sleep(0)
+        # Checked inside the loop: closing it cancels every task that is left, which would fake the stranger's cancel.
+        assert [s.status for s in (parent, child, grandchild)] == ["stopped", "stopped", "stopped"]
+        assert stranger.status == "running", "an unrelated running chat is not the tree"
+        assert all(t.cancelled() or t.done() for t in tasks[:3])
+        assert not tasks[3].cancelled() and not tasks[3].done()
+        for s in (parent, child, grandchild):
+            assert s.id not in agent_manager.tasks
+        tasks[3].cancel()
+
+    asyncio.run(run())
+
+
+def test_children_of_walks_every_kind_not_just_browser_agents():
+    parent, child, grandchild, stranger = p_tree()
+    assert [s.id for s in descendants.children_of(agent_manager.sessions, parent.id)] == [child.id]
+    assert [s.id for s in descendants.children_of(agent_manager.sessions, child.id)] == [grandchild.id]
+    assert descendants.children_of(agent_manager.sessions, stranger.id) == []
+
+
+def test_close_and_delete_take_the_same_walk():
+    """Three doors, one walk: a mode filter in any of them is the bug coming back."""
+    for fn in (SessionControl.SessionControl.stop_agent, SessionLifecycle.SessionLifecycle.close_session, SessionLifecycle.SessionLifecycle.delete_session):
+        src = inspect.getsource(fn)
+        assert "children_of(self.sessions, session_id)" in src, fn.__name__
+        assert 'mode == "browser-agent"' not in src, fn.__name__