Langfuse

Recently shipped v4.17 through v4.24, with a new evaluation UX, stable evaluator endpoints, and a faster Python SDK.

Evaluation workflows got a major overhaul. The new evaluation UX landed alongside stable evaluator and evaluation rule endpoints, payload-backed rule filters, and the ability to restore previous evaluator versions. Rule filters now reuse the search bar, and evaluator execution SLO metrics are tracked. The scores table gained inline categories and a chart view.

Trace and session views keep tightening. The compact timeline is being readied as the default trace view, and the trace-table facet list collapses behind a "Show N more" control. Sessions display arbitrary metadata, and annotation queues can be created directly from the trace and events tables.

Self-hosted and enterprise controls expanded. Ask AI now works on self-hosted deployments, and instance AI features support OpenAI-compatible APIs. Custom OIDC profiles can be read from the userinfo endpoint, and the migration UI shows by default in dual-write mode. Dashboards and widgets joined the core S3 data export.

In-app agents gained operational depth. Tool approval sources are recorded in the event stream, step-limit truncation surfaces instead of failing silently, and sandbox lifecycle metrics are emitted for MicroVMs. Assistant worker surfaces now default to LANGFUSE_IN_APP_AGENT_ENABLED.

SDK updates. Python SDK v4.15.0 fixed an OTEL memory leak and added URL-safe Base64 media support; JS SDK v5.11.0 continues the runExperiment() migration. v4.24.0 requires LANGFUSE_AI_PROVIDER instead of defaulting to Bedrock.