[eric] agents: delegation watchdog only counts children born after its own tool call, ending the mid-run sidecar shootings (ENG-327)

backend/apps/agents/manager/streaming/unwedge_sidecar.py

@@ -167,12 +167,25 @@ def is_delegation_core_tool(tool_name: str) -> bool:
 
 
 @typechecked
-def delegation_children_settled(session_id: str) -> bool:
-    """True when this session HAS delegated children and every one of them is terminal. No children
-    yet is NOT settled: a run queued behind the admission cap can wait minutes legitimately."""
+def delegation_children_settled(session_id: str, since: float) -> bool:
+    """True when this session has delegated children born AFTER this tool call started and every one
+    of them is terminal. No children yet is NOT settled: a run queued behind the admission cap can
+    wait minutes legitimately. The `since` scope is load-bearing: a parent's SECOND delegation used
+    to read its first run's terminal children as 'settled' while the new run was still queued, and
+    the watchdog shot a healthy sidecar mid-run (39 kills + 40 force-ended turns in one afternoon
+    of concurrent load, measured 2026-08-16 on the packaged build)."""
     from backend.apps.agents.agent_manager import agent_manager
-    kids = [s for s in agent_manager.sessions.values()
-            if getattr(s, "parent_session_id", None) == session_id and getattr(s, "mode", "") == "browser-agent"]
+    kids = []
+    for s in agent_manager.sessions.values():
+        if getattr(s, "parent_session_id", None) != session_id or getattr(s, "mode", "") != "browser-agent":
+            continue
+        born = getattr(s, "created_at", None)
+        try:
+            if born is None or born.timestamp() < since:
+                continue
+        except Exception:
+            continue
+        kids.append(s)
     if not kids:
         return False
     return all(getattr(s, "status", "") in ("completed", "error", "failed", "stopped") for s in kids)
@@ -201,7 +214,7 @@ def arm_delegation_watchdog(ctx: object, tool_use_id: str, tool_name: str) -> No
         if not session_id:
             return
         try:
-            settled = delegation_children_settled(session_id)
+            settled = delegation_children_settled(session_id, started)
         except Exception:
             settled = False
         settled_streak["n"] = settled_streak["n"] + 1 if settled else 0

backend/apps/help/changelog.py

@@ -98,6 +98,7 @@ P_RELEASES: List[ReleaseNote] = [
             "The mouse-wheel Zoom/Scroll setting works on real mice now. Accelerated wheels (Magic Mouse, Logitech smooth scrolling) report fractional scroll amounts that were being mistaken for a trackpad, so the wheel always panned no matter what the setting said.",
             "A browser helper that talks itself into refusing (\"I\u2019m a text-based AI\") no longer poisons its browser for every later task. Refused and fabricated runs are forgotten instead of remembered, and agents can ask for a completely fresh browser when one misbehaves.",
             "Agents no longer claim \"the user told me to stop\" when it was OpenSwarm\u2019s own housekeeping talking. Internal wrap-up and retry messages now identify themselves, so an agent\u2019s explanation of why it stopped reflects what actually happened.",
+            "Agents no longer lose their tools partway through a run. Under heavy multi-agent load, a safety timer could mistake a queued browser or app task for a lost one and restart the agent's tool connection mid-task; it now checks the right task before acting.",
             "Heavy sessions no longer vanish without a trace. When memory climbs past the safe line the app now sheds weight itself: preview thumbnails pause and refetchable caches drop, instead of growing until the operating system kills it mid-task.",
         ],
     ),

backend/tests/test_unwedge_sidecar.py

@@ -132,3 +132,65 @@ def test_a_blocking_tool_arms_nothing_at_all():
         arm_wedge_watchdog(ctx, "tu-2", "mcp__openswarm-core__AskUI")
         assert len(loop._scheduled) == before, "an exempt tool must not even schedule a timer"
     asyncio.run(run())
+
+
+# ------------------------------------------------- stale children must never read as settled
+
+
+def p_fake_kid(parent: str, status: str, born_ts: float):
+    from datetime import datetime
+
+    class Kid:
+        parent_session_id = parent
+        mode = "browser-agent"
+
+    k = Kid()
+    k.status = status
+    k.created_at = datetime.fromtimestamp(born_ts)
+    return k
+
+
+def test_stale_terminal_children_do_not_settle_a_new_delegation(monkeypatch):
+    # The packaged-log failure (2026-08-16): a parent's second AppAgent call queued behind the
+    # admission cap while its FIRST run's children sat terminal; the old check read that as
+    # settled and shot a healthy sidecar. Children born before `since` must be invisible.
+    import time as t
+    from backend.apps.agents import agent_manager as am_mod
+    now = t.time()
+    kids = {"a": p_fake_kid("parent1", "completed", now - 600),
+            "b": p_fake_kid("parent1", "stopped", now - 300)}
+    monkeypatch.setattr(am_mod.agent_manager, "sessions", kids)
+    assert unwedge_sidecar.delegation_children_settled("parent1", since=now - 10) is False
+    # Negative control: with the scope disabled (since=0 sees every child), the old verdict comes
+    # back, proving the filter is the thing doing the work.
+    assert unwedge_sidecar.delegation_children_settled("parent1", since=0.0) is True
+
+
+def test_fresh_terminal_child_settles_and_fresh_running_child_does_not(monkeypatch):
+    import time as t
+    from backend.apps.agents import agent_manager as am_mod
+    now = t.time()
+    call_started = now - 120
+    monkeypatch.setattr(am_mod.agent_manager, "sessions",
+                        {"a": p_fake_kid("parent1", "completed", now - 60)})
+    assert unwedge_sidecar.delegation_children_settled("parent1", since=call_started) is True
+    monkeypatch.setattr(am_mod.agent_manager, "sessions",
+                        {"a": p_fake_kid("parent1", "running", now - 60)})
+    assert unwedge_sidecar.delegation_children_settled("parent1", since=call_started) is False
+
+
+def test_no_children_is_not_settled(monkeypatch):
+    # A run queued behind the admission cap has no child yet; waiting is legitimate.
+    from backend.apps.agents import agent_manager as am_mod
+    monkeypatch.setattr(am_mod.agent_manager, "sessions", {})
+    assert unwedge_sidecar.delegation_children_settled("parent1", since=0.0) is False
+
+
+def test_mixed_stale_terminal_and_fresh_running_is_not_settled(monkeypatch):
+    import time as t
+    from backend.apps.agents import agent_manager as am_mod
+    now = t.time()
+    monkeypatch.setattr(am_mod.agent_manager, "sessions",
+                        {"old": p_fake_kid("parent1", "completed", now - 900),
+                         "new": p_fake_kid("parent1", "running", now - 30)})
+    assert unwedge_sidecar.delegation_children_settled("parent1", since=now - 60) is False