Replay an agent run offline.

AgentTwin records one run of an LLM agent to a JSONL tape, then replays it without calling a model — so a failing run becomes a regression test you can commit. Below: one tape, replayed twice. The first replay is faithful and exits 1 with two failing assertions; the second changes a single strategy key and exits 0 with all three passing. No model call, no API key, no network, no side effects.

Left: the recorded replay, exit 1, two assertions fail. Right: the counterfactual replay, exit 0, three assertions pass.
Open the replay that fails (exit 1) → Open the replay that passes (exit 0) →
  1. Open the first viewer: max_repeats(shell, 3) fails at seq 12. Click the failing row and the viewer jumps to that frame.
  2. Open the second one: the same assertions, the same tape, one policy key (stop_on_failure: true) — every row is now PASS.
  3. Both pages are the ordinary single-file report.html this library writes. They work with file:// too, no server needed.