AutoProver

Generate a fake-LLM replay tape (composer/testing/ui_harness_<name>.py) for the AutoProve smoke harness by RECORDING a real run and curating it. Use this whenever the user wants to create, record, or regenerate a tape / ui_harness script for a scenario so the pipeline can be replayed end-to-end with no real LLM calls — e.g. "make a tape for the Answer smoketest", "record a tape from this run", "generate a ui_harness for scenario X", "I need a deterministic replay of the autoprove pipeline". The tape is keyed by run_task task_id and replayed by composer.testing.harness_tape.HarnessFakeLLM. Recording yields a draft; a clean replay needs a hand-clean pass. Pairs with the inspect-run skill for debugging a recorded run.

Read the full LLM message history (inputs, outputs, tool calls, errors, token usage, stop reasons) for an AIAutoProver run — addressable by its log folder OR by run_id. Use this skill ANY time the user asks something about a specific run that produced an autoProve log — e.g. "why did that run loop", "what error did the model see on retry 5", "how many tokens did it burn", "show me the prompt for the system-analysis call", "what did the code_explorer subagent for X return", "what was the LLM's last response before it died". The events.jsonl only contains metadata; the actual messages live in Postgres and this skill is the way to reach them. Runs are also registered in a run index keyed by run_id — pass a run_id (or use `runs` to list them) when there's no events.jsonl handy. If the user asks about a run but gives you neither a log path nor a run_id, proactively ask which run (point them at ".certora_internal/autoProve/<timestamp>.events.jsonl", the autoProve folder, or a run_id from `inspect_run.py runs`) befo