What we shipped

Replay: re-run production traces against a new prompt

Pick any set of traces, change the prompt, model or temperature, and run them again. Tool responses are served from the recording, so replays never call Stripe, your CRM or your customers. Every evaluator is scored on both versions and shown side by side.

Replay is available on Team and Enterprise. Read the guide →

• ImprovedTrace search is 3× faster on workspaces with more than 100M spans.
• FixedStreaming spans from the Vercel AI SDK no longer show a duplicate first token.

Command menu

Press ⌘ K anywhere to jump to a trace, an agent, a dataset or a setting. Paste a trace ID to open it directly.

• NewKeyboard shortcuts: J/K to move between spans, E to expand all.
• ImprovedToken counts now show cached and uncached input separately.

Judge calibration

Label a sample of traces yourself and see how often each LLM judge agrees with you. Sightline reports agreement, Cohen’s κ and a confusion matrix, and links every disagreement back to its trace.

• NewInvite reviewers who only have access to labelling, not to raw traces.
• FixedEval scores on retried spans are no longer counted twice.
• FixedThe cost chart now respects the workspace time zone.

EU data residency

Enterprise workspaces can now store all trace data in Frankfurt. Existing workspaces can migrate from the settings page with no downtime.

• ImprovedThe Python SDK adds under 0.4ms of overhead per span, down from 1.1ms.