Record telemetry - Trajectory

Private beta. A telemetry event records what happened. A reward is the grader’s score.

Use log_event during the trajectory. Use client.telemetry.ingest_events for events that arrive later, with your organization API key.

During the trajectory

from trajectory import Client

with Client(api_key="<trajectory-api-key>") as runtime:
    runtime.trajectories.log_event(
        "<trajectory-id>",
        event_id="search-result-001",
        name="tool.result",
        payload={"tool": "search", "results": 3},
    )

Write runtime events before complete. Reuse event_id when retrying the same event. Read them back, newest first, with client.trajectories.list_events(tid). Events are not reward components. The source-grader guide logs grader details as an event; the diagnostic-artifact guide attaches a file to a trajectory.

Later application events

Required fields are event_id, event_type, timestamp, session_id, and source. Set trajectory_id when you have session.tid.

from datetime import datetime, timezone

from trajectory import Client

events = [
    {
        "event_id": "request-example-001:answer-accepted",
        "event_type": "answer.accepted",
        "timestamp": datetime.now(timezone.utc).isoformat(),
        "session_id": "session-example-001",
        "source": "support-app",
        "trajectory_id": "<trajectory-id>",
        "properties": {"surface": "chat"},
    }
]

with Client() as client:
    result = client.telemetry.ingest_events(events=events)

result reports ingested, skipped, and failures.