Improving LLM token usage when debugging