Changelog

Shipped: Failure classifier v2. 34 new error categories — up from 12. Includes tool-call-drift, context-pollution, and prompt-injection detection. The classifier now runs at ingest time with zero added latency. Shipped: LangGraph 0.3 auto-instrumentation. Drop-in support for LangGraph 0.3's new graph API. No code changes required — existing traces upgrade automatically. In progress: Deterministic replay seeds. We're rolling out reproducible replays for stochastic LLM calls. Early access available on request. Next up: Omium Desktop. A native app for inspecting traces locally. Beta signup is open.

Shipped: Checkpoint forks. Resume any failed run with alternative inputs — without affecting the original. Forks inherit their parent's context and can be diffed against the source run. Shipped: SAML SSO, GA. Okta, Auth0, Azure AD, Google Workspace. SCIM provisioning is included on Enterprise. Shipped: Audit log export. Every admin action on your workspace is now exportable to S3 or Snowflake. Available on Business and Enterprise plans. Fixed: A race condition that occasionally double-counted checkpoints during high-burst runs. Counts are now exactly-once across regions.