post_actions: FunctionActionFrame executes before TTS finishes due to concurrent audio context

[batuhanbadak]

pipecat version

v0.0.105

Python version

3.13

Operating System

Ubuntu

Issue description

After upgrading to pipecat v0.0.105, post_actions with {"type": "function"} execute while TTS is still speaking the LLM-generated response. The FE receives the event (e.g. "conversation ended") and closes the screen before the user hears the bot's closing message.

Reproduction steps

1. Create a pipecat-flows node where the LLM generates speech based on role_messages/task_messages
2. Add post_actions with {"type": "function", "handler": my_handler} — this handler sends a frontend event signaling the conversation is over
3. Use CartesiaTTSService on pipecat v0.0.105
4. Observe that my_handler executes while TTS is still speaking the LLM-generated response — the FunctionActionFrame passes through TTS without waiting for audio to finish

Expected behavior

Post-action FunctionActionFrame should be processed after TTS finishes speaking the LLM response, not concurrently. I fixed the issue by setting the private _pause_frame_processing, which resolved the issue on our end. However, this relies on a private API that could change without notice.

Actual behavior

After upgrading to pipecat v0.0.105, post_actions with {"type": "function"} execute while TTS is still speaking the LLM-generated response. The FE receives the event (e.g. "conversation ended") and closes the screen before the user hears the bot's closing message.

Logs