agentsclimarketplace

Langfuse trace

Skill rczamor/rz-agent-team/skills/shared/langfuse-trace

Start and finalize a Langfuse trace tagged with app_id, agent_role, session_id. Wraps every LLM call, tool call, and session-level event. Required on every reasoning step per the non-negotiable full-reasoning-logs rule.From its SKILL.md

Install
npx -y skills add rczamor/rz-agent-team --skill langfuse-trace

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
  • 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

3.0 KB, as published. Nobody here has run it

langfuse-trace

Purpose

Emit observability so every LLM call and tool call is traceable back to app, role, and session. Enforces the three required labels — app_id, agent_role, session_id — on every trace. This is rule #1 in TEAM.md.

When to invoke

  • --action start at session top; returns a trace_id reused for child spans.
  • --action span around each LLM call (prompt, response, tokens, model, cost) and non-trivial tool call.
  • --action event for handoffs, memory writes, BLOCKER posts.
  • --action finalize at session end with status + summary.

Every agent, every session.

Required env vars

  • LANGFUSE_HOST — internal http://langfuse-web:3000 or external https://langfuse-<suffix>.srv1535988.hstgr.cloud.
  • LANGFUSE_PUBLIC_KEY / LANGFUSE_SECRET_KEY — API keys from /docker/langfuse/.env on VPS.
  • SESSION_APP_ID — app tag (required label).
  • AGENT_ROLE — one of the 11 roster roles (required label).
  • SESSION_ID — Conductor-assigned session UUID (required label).

Input

Flags to scripts/langfuse_trace.sh:

  • --action <start|span|event|finalize> (required).
  • --name <string> (required) — e.g. "session.conductor.dispatch", "llm.plan-review".
  • --trace-id <id> (required for span/event/finalize; returned by start).
  • --parent-id <id> (optional) — nest under another span.
  • --input <json> / --output <json> (optional) — payloads for LLM/tool spans.
  • --metadata <json> (optional) — model, tokens, cost, latency_ms, tool_name.
  • --status <success|error> (finalize only).

Output

JSON on stdout: { "trace_id": "...", "span_id": "...", "url": "https://langfuse.../trace/..." }. Non-zero exit on API failure.

Example invocation

TRACE=$(bash scripts/langfuse_trace.sh --action start --name "session.backend-eng.sia-CAR-198")
TRACE_ID=$(echo "$TRACE" | jq -r .trace_id)

bash scripts/langfuse_trace.sh --action span --trace-id "$TRACE_ID" \
  --name "llm.generate-migration" \
  --input '{"prompt":"..."}' --output '{"response":"..."}' \
  --metadata '{"model":"claude-opus-4-7","input_tokens":2100,"output_tokens":540,"cost_usd":0.042}'

bash scripts/langfuse_trace.sh --action finalize --trace-id "$TRACE_ID" --status success

Implementation notes

  • Attach tags: [app_id, agent_role, session_id] on the root trace; propagate to children.
  • Use Langfuse SDK if available, else curl against /api/public/ingestion.
  • Batch spans and flush every 5s or on finalize.
  • Retry 5xx with backoff 1s/2s/4s; fail loud after 3 attempts.
  • Prefer internal langfuse-web:3000 from inside Docker; external URL only from outside the VPS.
  • Redact authorization headers in stderr logs.
  • Always include cost_usd when known — drives DevOps cost reports.

Keep looking

Skills are one crate of 608,475. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.