Skip to content

White-box Debugging

AgentCommander adopts a "White-box" philosophy. Unlike frameworks that hide agent logic inside opaque execution environments, we allow you to "step into" the agent's mind at any time.

💡 Pro Tip: Resuming Context

Since each experiment node runs in its own isolated directory, you can leverage the persistence features of the CLI tools for deep debugging.

Why is this powerful?

  • Inspect Final Prompts: See exactly what the CLI received after all template variables (like {{hint}} or {{parent_metric}}) were resolved. This is crucial for verifying your prompt engineering logic.
  • Review History: Access the full conversation log (e.g., GEMINI.md) to understand why the agent made a specific decision or error.
  • Manual Intervention: If an agent gets stuck in a loop, you can intervene manually and then let the workflow continue.

How to do it (Gemini Example):

  1. Open Terminal.
  2. Navigate to the experiment folder (e.g., cd my_project/Branch1/exp1.2.1).
  3. Run gemini -r (Resume mode).
  4. Interact:
    • Debug Prompt: Type "Show me the last prompt you received." to verify the input context.
    • Analyze Logic: Ask "Why did you choose this architecture?"
    • Manual Fix: Directly instruct "Fix the syntax error on line 45" to help it recover.

For Qwen CLI or OpenCode AI users, use qwen -c or opencode -c to achieve the same effect.