Langfuse v4.46.0 (2026-09-25) ships basic skill management (PR #17803) via unstable REST (list/create/get/patch/delete skills, versions, file content), streamlined experiment side-by-side comparison, formatted JSON on dataset items, optional 100-row tracing page size, assistant respect for data-retention, and v3 dataset-run-items reads on replicas.

Key Takeaways

  • ✓Shipped: Langfuse v4.46.0 (2026-09-25); skills PR #17803 merged same day
  • ✓Skills (unstable): /api/public/unstable/skills CRUD-ish endpoints for skill artifacts aimed at coding agents
  • ✓Experiments: streamlined side-by-side comparison; formatted JSON on dataset items
  • ✓Observability UX: optional 100-row tracing page size; plain-text header metrics; 300ms tooltips
  • ✓Compliance/perf: assistant respects data-retention; v3 dataset-run-items on read replica; docs langfuse.com/docs
🔬

In-Depth Technical Analysis

Background

LLM observability is moving from traces-only to evals plus agent skill assets. Pain points: skill files live outside the platform, experiment comparison is manual, large trace tables paginate poorly, assistants may read retained data. Langfuse v4.46.0 (2026-09-25) adds basic skill management and UX/perf fixes.

Architecture

PR #17803 exposes unstable REST under /api/public/unstable/skills (list/create/get/patch, versions, file content) so skill artifacts can live in-platform. Experiments get streamlined side-by-side comparison; dataset items support formatted JSON. v3 dataset-run-items reads go to a read replica. Assistant respects data-retention.

Benchmarks

No public latency leaderboard. Product metrics: optional 100-row tracing page size; 300ms tooltips; trace playback removed. Skill API is unstable—pin and wrap for production.

Getting started

Self-host: upgrade to v4.46.0. Docs: langfuse.com/docs and /docs/api. Note PyPI langfuse SDK versioning is independent of server 4.46.0—verify skill endpoints against the server release.