Replay
Re-run a captured branch of a production trace against a new prompt, model, or configuration, with lineage preserved.
Replay re-runs a captured branch of a production trace under a configuration you choose. The original run stays immutable, and the replay keeps a reference back to it.
#What replay needs
A span is replayable when it captured the prompt, the model settings, and the tool boundaries around it. Spans recorded without payloads show structure and cost but cannot be replayed.
#Selecting a branch
Choose the smallest branch that still contains the behavior you are investigating: the parent decision, the affected span, and its output. A narrow selection makes the comparison readable.
#Dependency policy
Every replay declares what happens to each external call.
| Policy | Behavior | Use when |
|---|---|---|
captured | Replays the recorded response | The dependency is nondeterministic or has side effects |
live | Calls the dependency again | Fresh data is part of what you are testing |
mock | Uses a supplied fixture | You are testing a failure path |
#From replay to experiment
A single replay is evidence, not a rollout. Save the configuration as a candidate, attach the source trace, and evaluate it against a dataset of representative runs before it reaches traffic.
Lineage stays attached through evaluation and canary, so the eventual production change is traceable back to the run that motivated it.