Laminar

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.

The cross-cutting concepts shared by both SDKs — span context propagation, tags, metadata, sessions, privacy controls, custom LLM cost tracking, flushing — live in the two instrumentation files (one per language) so each is complete on its own.

Workflow for instrumenting a codebase

Copy this checklist and track progress:

- [ ] 1. Detect runtime (TS vs Python) and the package manager (lockfile-based).
- [ ] 2. Install the SDK with the detected package manager.
- [ ] 3. Wire LMNR_PROJECT_API_KEY from env; add it to .env.example/README (never commit secrets).
- [ ] 4. Call Laminar.initialize() once, at the earliest safe startup point.
- [ ] 5. Confirm auto-instrumentation captures the LLM/tool libraries in use.
- [ ] 6. Add first-party spans (observe / @observe) around orchestration; set user/session/metadata/tags early.
- [ ] 7. Apply privacy controls to sensitive inputs/outputs.
- [ ] 8. Run one representative flow; verify the trace tree, transcript view, and filters in the UI.

Ground rules:

• Use the repo's existing package manager (detect from lockfiles); never introduce a new one.
• Keep diffs minimal and aligned with existing code style. Don't refactor unless necessary.
• In a monorepo, install Laminar in the package(s) that actually run the traced code (server, worker, eval runner), not just the root.
• Don't guess APIs. If unsure, consult the reference file or ask. For exact, current package versions, query npmjs.com / pypi.org rather than relying on memory.
• Avoid double-instrumentation — never run two tracer SDKs that both instrument the same calls.

Prerequisites for any task

• A Laminar project API key (dashboard → Settings → Project API Keys), set as LMNR_PROJECT_API_KEY, or mint one automatically with the CLI setup command references/cli.md
• Self-hosted Laminar: also set the base URL. SDKs accept baseUrl/base_url (or LMNR_BASE_URL); local defaults are HTTP port 8000 and gRPC port 8001. The CLI uses --base-url http://localhost --port 8000.

Verify before declaring done

For instrumentation tasks, confirm in the UI: a single root span per trace boundary, child spans nested correctly, the transcript view reads cleanly (LLM/TOOL spans present), and tags/metadata are filterable. If traces are slow or missing in a short-lived script, flush — see the per-language reference.