Every release, newest first.

Set a pass rate, latency budget and cost ceiling. A version ships only when it clears them.

Exports over 2 MB were cut at the span boundary. They now stream in full.

Rolling back reset the canary slice to zero. It now returns to the last chosen value.

Set a budget per project or per version and get an alert before the invoice does.

Search across a million runs now returns in under 300 ms.

Timestamps were shown in UTC regardless of the profile setting.

Overlapping tool calls now sit on the same row of the waterfall with true timing.

Each failed rubric case now shows the grader's reason next to the score.

A retried webhook could post the same alert twice.

Run the collector inside your network so raw prompts never leave it.

The SDK now loads lazily and adds under 5 ms to a cold start.

Rules now walk arrays of arrays.

A faster console with a keyboard-first trace view, saved filters and a command palette.

Replays used the latest dataset instead of the pinned one.

Send a single run, redacted by your rules, to anyone without an account.

J and K move between spans, Enter opens one, Escape closes it.