Streamed narration, surfaced to onProgress and never persisted as a step.
Streamed model reasoning/thinking tokens for live observability.
Structured tool invocation start emitted by CLI or cloud harnesses for live observability.
Structured tool execution result emitted by CLI or cloud harnesses for live observability.
Informational in-sandbox activity (file edit, shell command) for live view only.
The harness wants to do something the posture gates; becomes an MJ: AI Agent Requests row.
Token/cost usage for the turn. Without this the run cannot be accounted for — see below.
The turn ended. RawText is expected to carry the Loop next-step JSON envelope.
The session failed. Terminal for the turn; the run decides whether to retry.
Events an adapter emits while a turn is in flight.
Deliberately a small, closed set. Anything the harness does INSIDE its sandbox that does not cross one of these boundaries is governed by posture policy, not recorded as a run step — the opaque-super-step property of this design, which is intentional and must stay documented rather than quietly widened.