Thanks — you're asking how to signal the main LLM thread to wrap up, let it finish its final reply, and only then end the session (not forcibly cut generation). Below is the correct, minimal pattern to do that. 1) Send an urgent “wrap up” prompt to the agent’s main thread (so the model knows to produce a final concise closing turn). 2) Capture the returned SpeechHandle from the reply call. 3) Await the SpeechHandle’s completion (this waits for TTS synthesis + playback). 4) Call session shutdown/end only after the handle finishes. Example sequence (conceptual Python):

# 1) request final wrap-up
handle = await session.generate_reply(instructions="Please provide a concise closing statement and sign off — time is up.")

# 2) wait until the model's final output has finished streaming/playing
await handle.wait_for_playout()

# 3) end the session after playback completes
await session.shutdown(drain=False)

Key points • Use the SpeechHandle returned by

generate_reply()

or

say()

. •

wait_for_playout()

is the synchronization point that guarantees the LLM's final utterance (and audio) is fully delivered. • Call

shutdown()

(or

end_session

) only after

wait_for_playout()

completes. Setting

drain=False

is appropriate because you've already waited for the last output. If you want, tell me which SDK/language you’re using (or whether you need handling for text-only responses or streaming clients), and I’ll provide a concrete snippet for that environment. Sources: Agent speech and audio | LiveKit docs | Uninterruptable Agent | LiveKit docs | Job lifecycle | LiveKit docs