ryanbbrown/Retrodict ? reverse-engineered prompt

Build me an ARC AGI 3 agent that can solve the public games by reasoning from the game log instead of trying to interpret everything visually. I want it to watch each step, write what happened into a log, form hypotheses about the rules, test those ideas against past frames with Python, and only commit to a move plan when the prediction looks solid. It should keep a simple playbook file so it can resume after the context gets too long, and it should be able to recover when a plan fails by reading the diff of what changed.

Make the runner handle the loop, including checking predicted board states after each action, restarting when needed, and keeping the workspace files as the main memory. If helpful, look up any current docs online. The result should be something I can run locally from the repo and iterate on, with tests and a clean structure so the agent can be improved over time.