Diagnose Failure Patterns in Agent Traces

Implement diagnose_trace(steps) that analyzes a list of agent loop step dicts and returns a diagnostic report dict. Each step dict has keys: - step_number (int) - thought (str) - action (str — tool name or "Final Answer") - arguments (dict) - observation (str) - duration_ms (int) Return a dict: { "total_steps": len(steps), "total_duration_ms": <sum of all duration_ms>, "tool_call_count": <number of steps where action is not "Final Answer">, "error_count": <number of steps where observation contains "error" (case-insensitive) or action is "error">, "repeated_actions": <list of action names that appear 3+ times consecutively, deduplicated>, "stuck_reasoning": <True if any thought appears 3+ times consecutively, else False>, "slow_steps": <list of step numbers where duration_ms > 2000>, "has_final_answer": <True if any action is "Final Answer", else False> } Edge cases: - If steps is None or empty, return a zeroed report (total_steps: 0, all counts 0, empty lists). - A step with missing keys should be skipped (not counted in any metric). - repeated_actions only catches consecutive repeats (e.g. ["search", "search", "search"]), not non-consecutive ones.