Skip to main content
sessions get returns the conversation. The audit log returns the run: an immutable, time-ordered record of the model calls, the tool calls and their output, and how the run ended. It is the debugging surface in alpha — there are no traces or dashboards yet.
Page with --after/--limit; pass the response’s next_cursor as the next after.

Event types

llm_call and tool_result are the earlier names for model_call and tool_call; events recorded before the rename still carry them, so handle both. Every event carries session_id, the idempotency_key of the run it belongs to, the agent_revision that ran, event_time, and a payload whose shape depends on type.
Model and tool output is stored up to 32 KiB per value. Alongside it the event carries the complete value’s byte count, SHA-256 digest, and a truncation flag (contentBytes, contentSha256, contentTruncated), so you can tell a short answer from a truncated one. Payload keys are camelCase (mcpServerUrl, toolCalls).

What it answers

Filtering by run is a client-side filter on idempotency_key — the log is per-session, and the key is what identifies one invocation within it.