Laminar

Laminar is an observability platform for LLM and agent applications. You instrument code so production runs become traces (trees of spans), then explore them in the UI or query them with SQL. This skill covers the SDKs (@lmnr-ai/lmnr for TypeScript, lmnr for Python), the lmnr-cli command-line tool, and the SQL Query API.

Mental model

Read this before instrumenting anything — it drives every decision below.

• One trace = one unit of work you want to analyze end-to-end (a request, a turn, a job, a pipeline run).
• A trace is a tree of spans. Spans are typed (LLM, TOOL, DEFAULT, etc.). Types drive UI behavior — the transcript view is high-signal and focuses on LLM and TOOL spans.
• Auto-instrumentation captures common LLM/tool libraries (OpenAI, Anthropic, Vercel AI SDK, LangChain, and more). You still add first-party spans around your own orchestration so traces are readable.
• Great traces have: one clear root span per trace boundary; a few meaningful child spans for major steps; stable, low-cardinality names; context for filtering (userId, sessionId, metadata, tags); and privacy controls so secrets/PII are never recorded.

Cardinality rule (applies everywhere): never put dynamic IDs (request IDs, user IDs, document IDs) in span names or tags. Span names and tags stay stable and low-cardinality; identifiers go in metadata or the dedicated userId/sessionId fields.

Choose your task

Read only the reference file(s) for the task at hand. Each is self-contained.