ang535/Traceback — reverse-engineered prompt

Build me a Python project called Traceback that wraps an AI coding agent and watches what it does step by step in real time. I want it to catch when the agent starts drifting away from the task, gets stuck repeating itself, or suddenly burns way more tokens than normal. When that happens, it should figure out the bad step, roll back to the last clean state, retry from there, and add a short corrective instruction so the agent can recover instead of silently failing.

Make it feel like an MVP I can actually run. Use a coding agent with Claude, log each step plus a full state snapshot, save state in simple JSON files, and include a small Streamlit dashboard so I can watch the trajectory live and see anomalies, severity, rollback decisions, and retries. Keep the detection rules configurable by domain with thresholds for drift, loops, and token spikes. Add a few tests and some simple buggy Python tasks to validate the whole flow.

Please wire up the env setup and make it runnable locally, and check current docs online if you need to.