retrace

// prove the fix

Prove the fix actually worked

Re-run a proposed change against the run that actually failed and get a verdict (improved, regressed, unchanged, or inconclusive) so you can prove a fix works before you ship it.

Start freeRead the docs →

Fork → replay → diff → verdict

Retrace forks the failed run, replays your change, finds the first point of divergence, and a judge returns a verdict: improved, regressed, unchanged, or inconclusive.

Typed hypotheses

Test 'what if this prompt, tool result, model, or temperature were different' with typed substitutions, no changes to your agent code.

Only re-run what's affected

A dependency-aware replay re-executes just the steps your change touches and serves the rest from the recording, clearly labeled, so verification is fast and cheap.

Pinpoint the cause

retrace bisect binary-searches the run to find the exact first step whose change flips the verdict.

Corpus sweeps

Apply one fix across many failed runs at once to confirm it holds beyond a single example; partial results are kept and labeled if you hit a limit.

Regression gates

Mark known-good runs as golden and compare future runs against them for structural regressions, pairs directly with the CI eval gate.

terminal
# Re-run the failed trace and get a verdict
retrace traces verify-fix <trace-id>

# Find the first step whose change flips the result
retrace bisect --trace <id> --substitutions subs.json

Prove a fix, then bisect to the exact step that caused the regression.

REST API

  • POST /api/v1/forks/sweep: run one fix across N traces
  • GET /api/v1/forks/sweep/:id: sweep results
  • POST /api/v1/forks/sensitivity: sensitivity analysis

More of the platform

// recordRecordOne decorator turns every run into a regression test// fork & replayFork & ReplayFork the failed step and re-run, debug the cause, not the symptom// 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