Agent Replay

Records every action an AI agent takes so you can replay sessions step by step and catch regressions. Provides tools to start and stop recording, log individual actions with inputs, outputs, reasoning and timing, then replay the full sequence. The real utility is in compare_sessions, which diffs two runs and shows you exactly where behavior diverged, and find_divergence_point, which pinpoints the first step that deviated from expected output. Export sessions as JSON for tooling or markdown for human review. Useful when debugging non-deterministic agent behavior, comparing model versions, or proving that a workflow changed between runs. Works over stdio and stores sessions in memory keyed by session ID.