It'd be handy to be able to see the raw output of the LLM, so we can tell when issues in the spoken output are coming from the model (e.g. it put punctuation in a weird place), or from Vapi / the synthesizer.