Trace events

Traces are ordered runtime facts. Event data is intentionally small and provider-neutral.

Event kind	Data	Meaning
execution.started	task	Configuration completed and execution began
model.requested	call	A provider call is about to run
model.failed	error	The provider raised an AgentError
tool.started	tool, call_id	Tool invocation began
tool.completed	tool, call_id	Tool invocation returned successfully
tool.failed	tool, call_id, error	Tool invocation failed
tool.denied	tool, optional call_id, reason or error	Tool was unavailable or unauthorized
result.rejected	error	Submitted result failed the contract
result.produced	none	A final result passed validation
artifact.created	path	A created workspace file was recorded
artifact.modified	path	A modified workspace file was recorded
artifact.deleted	path	A deleted workspace file was recorded
execution.completed	none	A successful execution ended
execution.cancelled	boundary	Caller cancellation was observed
execution.limit_exhausted	limit, optional boundary	A configured budget was exhausted

Consume events defensively

Event kinds and data are structured, but consumers should ignore fields they do not understand:

for event in execution.trace.events:
    if event.kind == "tool.failed":
        report_tool_failure(
            tool=event.data.get("tool"),
            error=event.data.get("error"),
            timestamp=event.timestamp,
        )

Do not treat a trace as model chain-of-thought. It records observable runtime transitions and tool outcomes.

See the generated Trace and TraceEvent API.