Observability¶

Daimon emits OpenTelemetry traces for every request. When an OTLP endpoint is configured, each call to /v1/converse/{component} creates a root span with the provider and model as attributes.

Configuration¶

telemetry:
  otlp_endpoint: "localhost:4318"   # OTLP/HTTP. Leave empty to disable.

When the endpoint is empty (the default), all tracing is a no-op — zero overhead.

Span structure¶

Each request produces a single root span named converse:

Attribute	Value
`gen_ai.component`	Component name (e.g. `claude`)
`gen_ai.request.model`	Model name from the request

Providers should create child spans for their upstream API calls and annotate them with OpenTelemetry GenAI semantic conventions:

Attribute	Example
`gen_ai.system`	`anthropic`
`gen_ai.request.model`	`claude-opus-4-7`
`gen_ai.usage.input_tokens`	`42`
`gen_ai.usage.output_tokens`	`128`

Local setup with Jaeger¶

The quickest way to see traces locally:

docker run -d --name jaeger \
  -p 4318:4318 \
  -p 16686:16686 \
  jaegertracing/jaeger:latest

Configure daimon:

telemetry:
  otlp_endpoint: "localhost:4318"

Open the Jaeger UI at http://localhost:16686, select daimon in the service dropdown, and click Find Traces.

Grafana / production¶

Any OTLP-compatible collector works. For Grafana Tempo:

telemetry:
  otlp_endpoint: "tempo.internal:4318"

Or route through the OpenTelemetry Collector:

telemetry:
  otlp_endpoint: "otel-collector:4318"