I’m testing Claude with extended thinking on multi-step tasks, and I’m trying to make failures easier to investigate. When an answer is wrong, I don’t want to assume that seeing more internal reasoning is the right debugging tool; I’d rather capture the details that help me reproduce the issue and compare runs.

For people using Claude through the API, what do you log alongside the prompt and final answer? Do the extended thinking summary, token counts, model version, or sampling settings help you narrow down what changed?