Useful that the rollback also does git checkout, because pruning the chat history alone would leave the failed patch sitting in the workspace. Two checks I'd add to the 20-case suite: run the same task where the first failure is a misleading error (a test that fails for an environment reason, not because of the patch) and see whether the 927-character lesson teaches the agent a wrong rule that then hurts the retry. Also compare against a task needing three or more attempts, where lessons accumulate, to see if the lesson log itself starts poisoning the prompt. A single run per mode on gemini flash is a thin sample, so I'd repeat each mode ten times and report the spread.