Use when an LLM app needs production tracing, monitoring or debugging of model calls, or production traces should feed offline eval datasets. Adds Langfuse with @observe or startActiveObservation, records metadata, cost and latency, scores traces and manages prompts as code.