[issue-3064][slice-5/6] Lifecycle-aware monitoring: health-monitor...

orchestrator/concurrent_executor.py

@@ -494,6 +494,8 @@ def _agent_free(*, action: str, role: str, payload: Any = None) -> None:
             on_exhausted=self._teardown_exhausted_session,
         )
 
+        # #3064 slice-5: convergence-stall notifier re-uses the same
+        # OVERSEER_ALERT surface wired for the supervisor.
         loop = OrchestratorEventLoop(
             tracker,
             spawner_adapter,
@@ -504,6 +506,8 @@ def _agent_free(*, action: str, role: str, payload: Any = None) -> None:
             roles=make_role_list(roles),
             job_supervisor=supervisor,
             job_status_view=self._event_status_view,
+            convergence_stall_notifier=self._emit_supervision_alert,
+            active_roles_notifier=self._publish_active_roles,
         )
         self._event_loop = loop
         loop.start()
@@ -514,8 +518,67 @@ def _agent_free(*, action: str, role: str, payload: Any = None) -> None:
             phase=phase,
             roles=[r.value for r in roles],
         )
+
+        # #3064 slice-5: set orchestrator mode on the health monitor and
+        # heartbeat coordinator so their tripwire/refresh behavior reflects
+        # the ownership mode (roles with no active Job are normal in
+        # orchestrator mode; gateway-session refresh via heartbeat fan-out
+        # is suppressed).
+        self._enable_orchestrator_mode_surfaces()
+
         return loop
 
+    def _publish_active_roles(self, roles: set[str]) -> None:
+        """Publish the event loop's live-Job role set to the health monitor.
+
+        Wired as the event loop's ``active_roles_notifier`` so the monitor's
+        ``_active_jobs`` reflects which roles currently have an in-flight
+        one-shot Job on every poll tick.  This is what makes orchestrator-mode
+        active-Job scoping (and silent-mid-event-pod coverage) actually take
+        effect in production — without it ``_active_jobs`` stays empty and
+        ``_orchestrator_skip_tripwire`` suppresses every role.
+
+        Best-effort: a missing health monitor (unit tests stand up only the
+        component under test) is tolerated.
+        """
+        try:
+            from health_monitor import get_health_monitor
+
+            hm = get_health_monitor()
+            if hm is not None:
+                hm.set_active_roles(roles)
+        except Exception:  # noqa: BLE001 — best-effort
+            pass
+
+    def _enable_orchestrator_mode_surfaces(self) -> None:
+        """Propagate orchestrator mode to downstream surfaces.
+
+        In orchestrator mode:
+        - The health monitor suppresses alerts for roles with no active Job.
+        - The heartbeat coordinator suppresses gateway-session fan-out
+          (refresh happens at spawn time from slice-4 worktree re-attach).
+        - Absent-sender heartbeats between events trip nothing.
+
+        Best-effort: a missing health monitor or coordinator is tolerated
+        (unit tests often stand up only the component under test).
+        """
+        try:
+            from health_monitor import get_health_monitor
+
+            hm = get_health_monitor()
+            if hm is not None:
+                hm.set_orchestrator_mode(True)
+        except Exception:  # noqa: BLE001 — best-effort
+            pass
+
+        try:
+            from heartbeat import get_heartbeat_coordinator
+
+            hc = get_heartbeat_coordinator()
+            hc.set_orchestrator_mode(True)
+        except Exception:  # noqa: BLE001 — best-effort
+            pass
+
     def owns_event_loop(self) -> bool:
         """True when the orchestrator-owned BRC event loop drives this phase.
 

orchestrator/consensus_wrapper.py

@@ -98,6 +98,13 @@ def _event_loop_owner() -> str:
 # default -- well above the WS7-observed 10-13 min legitimate-idle ceiling.
 EVENT_PUMP_IDLE_BUDGET_MIN_DEFAULT = 30
 
+# OVERSEER_ALERT anomaly name raised when the idle budget is exceeded. The
+# orchestrator-side re-homed convergence-stall check (``event_loop.py``) must
+# raise the SAME name so the overseer classifies the in-pod and orchestrator
+# variants identically. Single-sourced here and interpolated into the wrapper
+# template below; imported by ``event_loop`` for its OVERSEER_ALERT emission.
+EVENT_PUMP_IDLE_BUDGET_ANOMALY = "stuck-phase-transition"
+
 # Heartbeat cadence for the wrapper-owned background heartbeat emitter
 # (#2908 task-2-2). Migrated from
 # ``sandbox/egg_agent_tools/handlers/message.py:_WAIT_LOOP_HEARTBEAT_INTERVAL_SECS``
@@ -139,6 +146,7 @@ def _event_loop_owner() -> str:
 # Placeholders interpolated by ``str.format``:
 #   {agent_command_prefix}   -- ``python3 -m egg_agent --model X --max-turns N``
 #   {idle_budget_min_default}, {hb_interval_default}, {wait_timeout_default}
+#   {idle_budget_anomaly}    -- OVERSEER_ALERT anomaly name on budget breach
 _EVENT_PUMP_WRAPPER_TEMPLATE = r"""#!/bin/bash
 set -uo pipefail
 
@@ -692,7 +700,7 @@ def _event_loop_owner() -> str:
 " 2>/dev/null || echo "(snapshot unavailable)")
     timeout 5 egg-orch overseer alert "${{EGG_PIPELINE_ID:-unknown}}" \
         --role "${{EGG_AGENT_ROLE:-agent}}" \
-        --anomaly stuck-phase-transition \
+        --anomaly {idle_budget_anomaly} \
         --priority "$priority" \
         --summary "BRC event-pump idle for ${{idle}}s$summary_extra" \
         --detail "Event-pump for role=${{EGG_AGENT_ROLE:-agent}} slice=${{EGG_SLICE_ID:-none}} has seen no actionable BRC event for ${{idle}}s (configured budget ${{IDLE_BUDGET_SECS}}s). The loop continues blocking; no FAILED transition is forced. BRC state: $brc_snapshot" \
@@ -1108,6 +1116,7 @@ def build_event_pump_wrapped_command(
     script = _EVENT_PUMP_WRAPPER_TEMPLATE.format(
         agent_command_prefix=agent_command_prefix,
         idle_budget_min_default=idle_budget_min,
+        idle_budget_anomaly=EVENT_PUMP_IDLE_BUDGET_ANOMALY,
         hb_interval_default=heartbeat_interval_secs,
         wait_timeout_default=wait_timeout_secs,
         spvr_backoff_factor=SUPERVISION_BACKOFF_FACTOR,