retrace

// fork & replay

Fork from any step. Replay the future.

Branch a recorded run at the exact step that broke, change the input, and re-execute the entire agent down the new path, then compare the two runs side by side.

Start freeRead the docs →

Cascade replay

Forking re-runs the full function from the chosen step forward, so a changed input actually propagates into every downstream call, not just one isolated model call.

Reproducible re-execution

Server-side replays re-run the model with deterministic settings so results are as repeatable as providers allow, and recorded runs can be replayed back exactly from the tape.

Pay only for what changes

Steps before the fork point are served from the original recording at no token cost, you're billed only for the part of the run that actually diverges.

Side-by-side diff

Compare the original and forked runs step by step, with cost and latency deltas highlighted down to the first point where they diverge.

Verified replay

Accelerated replays are checked against a fresh execution; if the output diverges beyond a threshold, Retrace falls back to a full run so debugging fidelity is never silently lost.

Replay player & cost preview

Scrub any run like a video, and preview what a fork will cost (with a breakdown of where the savings come from) before you run it.

terminal
retrace forks create --trace <id> --span <id> --input "grounded prompt"
retrace forks replay <id> --wait
retrace forks diff <id> --json

Fork at a step, replay it, then inspect the diff.

REST API

  • POST /api/v1/forks: create a fork from a span
  • POST /api/v1/forks/:id/replay: re-execute the fork
  • GET /api/v1/forks/:id/diff: original vs fork diff

More of the platform

// recordRecordOne decorator turns every run into a regression test// prove the fixProve the FixRe-run a fix and get a verdict before you ship// eval gatesEval GatesFail the build when agent behavior regresses// guardrailsGuardrailsHalt a run when it blows a budget, loops, or stalls// enforcementEnforcementBlock a bad action before it runs, not diagnose it after// failure taxonomyFailure TaxonomyAuto-classify failures so you fix patterns, not one-offs// sessionsSessionsTrace multi-agent hand-offs so failures aren't a black box