agrag.eval.trajectory.SpanCapture

SpanCapture() -> None

Capture one agent run's spans for read_trajectory.

Use as a context manager around agent.ainvoke and read the run with trajectory() after. Each capture has its own provider and exporter, so captures never share spans and the global provider is unchanged.

Example

with SpanCapture() as capture:
    agent = build_agent(engine, settings, tracer=capture.tracer)
    result = await agent.ainvoke({"messages": [...]})
trajectory = capture.trajectory()

Functions:

• trajectory – Read the captured spans as a trajectory.

Attributes:

• tracer (Tracer) – The tracer to pass as tracer= to build_agent.

tracer

tracer: Tracer

The tracer to pass as tracer= to build_agent.

Returns:

• Tracer – A tracer bound to this capture's private provider.

trajectory

trajectory() -> Trajectory

Read the captured spans as a trajectory.

Returns:

• Trajectory – The trajectory read from the spans captured so far.