Thanks — I’m sorry you’re seeing inconsistent state-change logs. You’re asking why you sometimes get the transition from

initializing

→

listening

, the agent speaks, but you don’t see the later

agent_state_changed

events (for

speaking

→

listening

or similar). I’ll walk through the most likely causes and give focused troubleshooting steps so we can narrow this down. Likely causes to check • Exceptions inside the event handler: If your

on_agent_state_changed

callback raises an exception (synchronous or asynchronous), it can interrupt event processing and stop further callbacks from running. • Side-effects in the handler that affect the session/room: Deleting the room or otherwise closing the session while the agent is speaking can prevent later state transitions / events from being emitted or delivered. • Fire-and-forget tasks not awaited or not observed: Creating background tasks (like

asyncio.create_task(lkapi.room.delete_room(...))

) without proper error handling can hide exceptions from those operations and lead to race conditions with the agent lifecycle. • TTS / audio playback timing: If the agent’s speaking finishes and the room is removed immediately after (or during) playback, events for the subsequent state may not reach your handler. Quick checks you can run now 1. Add robust error logging inside the handler so any exception is visible: ◦ Wrap the handler body in try/except and log exceptions with stack traces. ◦ Log the entire

event

object on every call (old_state, new_state, timestamp if available). 1. Temporarily remove the room deletion call to see if the missing events stop: ◦ Comment out the

lkapi.room.delete_room(...)

call and confirm whether you then consistently receive the post-speech state changes. 1. Avoid deleting the room synchronously from the handler: ◦ If you must delete the room, delay the delete until after the agent session fully completes (for example, schedule deletion after observing

new_state == "idle"

or after a short await). 1. Observe the background task’s exceptions: ◦ When you create

disconn_task = asyncio.create_task(...)

, add a done callback or await it eventually to surface exceptions: ▪︎

disconn_task.add_done_callback(lambda t: log(t.exception()))

1. Add more granular logs around lifecycle boundaries: ◦ Log when you call

session.start(...)

, when

CustomAgent

is created/entered, and when TTS playback starts/ends (if accessible). 1. Run a reproducible minimal test: ◦ Create a minimal flow that only starts the session, triggers a single speak, and logs all state events. Confirm whether the issue reproduces in isolation. Safe code-pattern suggestion • Do not perform destructive operations (room deletion, session close) directly inside the

agent_state_changed

handler while the agent may still be speaking. • Wait for a clear terminal/idle state before deleting the room: a. Observe

new_state == "idle"

(or whichever final state your flow uses). b. Then call

lkapi.room.delete_room(...)

(and handle/await the response). Questions to help me narrow this down • Do you see any exceptions or error logs when the missing events occur? If so, please paste them. • Does removing/commenting out the

delete_room

call make the missing events go away? • Which agent states do you observe consistently (list them in order when you do see full events)? • Are you running this in a single-threaded asyncio loop or in a framework that manages tasks differently (e.g., an ASGI server)? If you share the logs from a single failing run (including any exception tracebacks and the printed events around the speaking period), I’ll use those to point to the exact race or exception and suggest a minimal code change. Sources: Turns overview | LiveKit Documentation | Agent session | LiveKit Documentation, with interactive elements