So interestingly enough, we first tried letting GPT interact with pdb, through just a set of directed prompts, but we found that it kept hallucinating commands, not responding with the correct syntax and really struggling with line numbers. That's why we pivoted to just getting all the relevant data upfront GPT could need and letting GPT synthesize that data into a singular root cause.

I think we're going to explore the local model approach though - you raise some really great points about having more granular control over the state of the model.