Do a better job of passing errors back to the model and make the conversation log easier to read.

I think right now we just take the most recent error and inject it into the chat history. But really we should show the entire history to the model, including all previous errors. Otherwise it might repeat mistakes.

It would be great if this was easier to visualize in the CLI output and eventual web UI.