Tracing

Every Program::run and Program::execute runs inside a typedlm.execute span at level INFO. Field names follow the OpenTelemetry GenAI conventions where they exist.

Field	Value
typedlm.program	name of the signature
gen_ai.operation.name	chat
typedlm.strategy	native_schema, tool_call, json_mode or prompt_only
gen_ai.response.model	model that answered, on success
gen_ai.usage.input_tokens, gen_ai.usage.output_tokens	sum over all attempts
typedlm.attempts	number of model calls
typedlm.outcome	valid, repaired or failed
error.type	Error::kind(), on failure

Each attempt adds a DEBUG event typedlm attempt with typedlm.attempt, typedlm.latency_ms, token usage and typedlm.result (ok or the error kind).

Inputs and answers

They are not recorded by default, because they often contain personal data. Enable them per program where your traces may hold them:

let classify = Program::<ClassifyTicket, _>::new(provider).record_content(true);

The input and each answer then appear as DEBUG events with the fields typedlm.input and typedlm.response.

Any tracing subscriber works, for example tracing-subscriber for local output or an OpenTelemetry exporter.