agrag.eval.trajectory.read_trajectory
read_trajectory(spans:Sequence[ReadableSpan]) -> Trajectory
Read the tool and model steps from finished spans.
Keeps TOOL and LLM spans, drops CHAIN spans, and orders steps
by start time rather than export order. Skips spans agrag opens
itself and spans nested under an agrag.eval.judge span, so judge
calls and BAML request spans never read as planner steps. A step's
subagent is the subagent_type of its nearest ancestor task
span, or None for a planner step.
Parameters:
- spans (
Sequence[ReadableSpan]) – The finished spans of one traced agent run.
Returns:
Trajectory– The run's trajectory in start order.