Observability (OpenLit)

CUGA can emit OpenTelemetry traces, metrics, and logs for every LLM call using OpenLit — a drop-in OTel instrumentation for popular LLM SDKs.

Install

OpenLit ships as an optional extra:

pip install "cuga[observability]"
# or with uv:
uv sync --group observability

Configure

[observability]
openlit = true

Point OpenLit at your OTLP collector via environment:

export OTEL_EXPORTER_OTLP_ENDPOINT="http://localhost:4318"

Common service-identifying env vars (OTEL_SERVICE_NAME, OTEL_RESOURCE_ATTRIBUTES) work as usual — CUGA does not override them.

Local testing stack

The cuga-agent repo ships a docker-compose stack under deployment/docker-compose/openlit/ containing an OTel Collector, Tempo (traces), Prometheus (metrics), and Grafana (UI). Start it, point CUGA at the collector, run a task, and you'll see per-call traces with prompt, model, token counts, latency, and cost.

What gets captured

For each LLM invocation OpenLit records:

• Trace span — start/end, duration, parent-child relationships across planner / shortlister / coder / reflection nodes.
• Attributes — model name, provider, temperature, prompt/response content (off by default — configure per OpenLit's docs), token usage (input/output/total), and any raised exception.
• Metrics — request counts, token counts, and latency histograms exported via OTLP.

Combining with Langfuse

langfuse_tracing = true under [advanced_features] is independent of OpenLit and can be enabled in parallel — useful when you want both an OTel-native pipeline and a Langfuse dashboard.