Replay verification

The most useful guarantee AgentBridge provides: the same inputs in the same world produce the same outputs. Every regression hunt benefits from this.

How it works

1. Start a session against the adapter with the replay capability negotiated. The adapter writes every command + response + event to <userdata>/agentbridge/<session_id>.ndjson.
2. Run any agent (mission / script / human) to drive a meaningful sequence. Periodically call snapshot_hash -- those hashes are the ground truth.
3. After the session, copy the ndjson somewhere stable.
4. Run agentbridge replay <ndjson> against a fresh adapter (with set_seed + snapshot_hash capabilities). The harness replays the inputs and asserts the snapshot hashes match.

Determinism is opt-in

The base AgentStateDump includes timestamps in time, which makes hashes drift across runs by design. For deterministic replay:

• Game uses set_seed to reset all PRNGs.
• State dump excludes wall-clock fields (override _time_dump in Godot, TimeDump in Unity).
• All non-seeded inputs (camera shake amplitudes, particle emit rates) are seeded.

Then snapshot_hash is byte-stable for the same inputs.

Catching regression

Drop the previous-known-good ndjson in your CI tree. Every PR runs agentbridge replay <recorded>.ndjson. If a code change breaks determinism, the divergence index points to the exact step.