2025.4 introduced the ability to start conversations on a voice satellite using assist_satellite.start_conversation. I was hoping to be able to use this to run a script if I respond yes, but it appears to just be limited to what your LLM can do. Is there any way to define the actions that are taken when responding yes or no, or is this something we need to wait for to be developed?
For anyone else trying to do this, I found a way. Make sure your script is exposed to Assist and the key thing is to make sure you use extra_system_prompt to tell the LLM what you want it to do. In my case, this is what I used:
- action: assist_satellite.start_conversation
target:
entity_id: assist_satellite.{{ states('input_text.last_voice_satellite') }}
data:
start_message: Would you like me to delete the message?
extra_system_prompt: The user is being asked if they would like to run script.delete_noticeboard_note.
preannounce: false
Hopefully this saves someone a few hours of trying to figure this out for themselves. 
6 Likes
Thanks alot for this it works very well with the PE. This does not seem to work with the implementation on the Atom Echo, I have even set them back to default to try them. It does not look like the firmware was updated after Feb though so this is likely part of it. Don’t suppose anyone got it working with them?
The atom echo does not yet support the required firmware. Sorry
1 Like
Yep checked on github and looks like there is a pull request on the topic Modify M5 Stack ATOM ECHO to support start and continue conversations by kahrendt · Pull Request #99 · esphome/wake-word-voice-assistants · GitHub
This should now be available