trace history instead of chat response in toolloop

[comhar]

This PR updates the tool loop tracing logic so that it traces history instead of the chat response. This change allows us to trace all requests to/from the ai as they happen. This should make it easier to debug issues with function calls and results. Previously only assistant responses were traced.

Here's what the order cancellation example from the docs looks like now:

tl_trace_clean.mov

FWIW, here's the tr_hist method from the video

def unpack_kwargs(kwargs): return ', '.join(f'{k}={v}' for k, v in kwargs.items())
    
def tr_hist(h):
    if not isinstance(h, list): h = [h]
    for o in h:
        role, content = o["role"], o["content"]
        for m in content:
            if m['type'] == 'text': print(f"{role}: {m['text']}")
            if m['type'] == 'tool_result': print(f"{role}: tool result: {m['content']}")
            if hasattr(m, 'text'): print(f"{role}: {m.text}")
            if hasattr(m, 'id'):
                print(f"\n================ tool call ================")
                print(f"{m.name}({unpack_kwargs(m.input)})")
                print(f"===========================================\n")
    print("\n")

See #9 for more context.

[gitnotebooks]

Found 1 changed notebook. Review the changes at https://app.gitnotebooks.com/AnswerDotAI/claudette/pull/39

[jph00]

Great thanks @comhar ! BTW don't forget to at-mention me or add me as a reviewer otherwise I only notice these things if/when I happen past them :)