[eric] telemetry: envelopes get real concurrency via a sessions provider, every turn opens with a baseline crumb

backend/apps/agents/agent_manager.py

@@ -46,6 +46,8 @@ class AgentManager(SessionLifecycle, SessionPersistence, Messaging, SessionContr
     @typechecked
     def __init__(self):
         self.sessions: Dict[str, AgentSession] = {}
+        from backend.apps.agents.core.flight_recorder import set_sessions_provider
+        set_sessions_provider(lambda: self.sessions)
         self.tasks: Dict[str, asyncio.Task] = {}
         # Live mirror of the in-flight streamed assistant text per session, so a stop can persist the partial reply instantly instead of waiting out the multi-second SDK teardown the cancel handler sits behind.
         self.live_partial: Dict[str, PartialReply] = {}

backend/apps/agents/core/flight_recorder.py

@@ -13,6 +13,14 @@ from typeguard import typechecked
 P_RING_SIZE = 64
 p_lock = threading.Lock()
 p_rings: Dict[str, deque] = {}
+p_sessions_provider = None
+
+
+def set_sessions_provider(provider) -> None:
+    """agent_manager registers its live sessions dict once so envelopes built anywhere (error
+    handlers have no manager handle) still carry a real concurrency snapshot."""
+    global p_sessions_provider
+    p_sessions_provider = provider
 
 
 @typechecked
@@ -92,7 +100,7 @@ def build_envelope(
         "phase": phase,
         "attempts": attempts,
         "breadcrumbs": breadcrumbs(session_id),
-        "concurrency": concurrency_snapshot(sessions or {}),
+        "concurrency": concurrency_snapshot(sessions if sessions is not None else (p_sessions_provider() if p_sessions_provider else {})),
     }
 
 

backend/apps/agents/manager/run/TurnRunner.py

@@ -177,6 +177,8 @@ class TurnRunner(AgentManagerProtocol):
         p_use_persistent = persistent_client_enabled()
         capacity_retry_attempt = 0
         p_router_retry_attempt = 0
+        # Baseline crumb so even a first-call failure's envelope names the turn it died in.
+        flight_recorder.crumb(session_id, "turn-start", model=resolved_model, api=api_type)
         while True:
             try:
                 if p_use_persistent: