Tracing And Runs Api

Agent wraps a multi-step pipeline with @traceable on the orchestrator and child @traceable on each step. A trace tree is expected with the orchestrator as parent_run and three child runs under it.

Decorate the orchestrator and each step with @traceable; rely on LANGSMITH_TRACING=true and Python contextvars (or run_tree propagation) to auto-attach child runs to the active parent. Do NOT pass parent_run_id manually unless crossing a thread/process boundary — incorrect manual ids create orphane…

An agent issues three sub-operations: an LLM call, a vector search retriever, and a tool execution. Operator wants per-type cost and latency aggregates in the LangSmith Projects view.

Set run_type explicitly on each @traceable: 'llm' for model calls, 'retriever' for retrieval, 'tool' for tool invocations, 'chain' for orchestration. LangSmith's per-type filters and cost rollups depend on run_type being correct — defaulting everything to 'chain' loses the breakdown.

Service A starts a trace, calls service B over HTTP. Service B is expected to show as a child run under A's trace in LangSmith.

Service A serializes get_current_run_tree().to_headers() and adds them to the outbound HTTP request. Service B reads langsmith-trace and baggage headers via TracingMiddleware (or with tracing_context(parent=headers)) and resumes the trace context. Both services use the same LangSmith project.