Runs, events, checkpoints, approvals, and effects

A run is admitted before provider use. Its append-only events use monotonic sequences and idempotency identities. Checkpoints bind exact bytes by digest. Cancellation is a durable request, not proof that a provider, tool, or specialist already stopped.

Approvals, effects, and idempotency

Consequential tool calls create an effect intent before execution and a receipt afterward. An approval binds the run, tool, exact argument digest, consequence summary, risk class, data boundary, and expiry. Argument drift or a stale sequence is denied. Unknown effects require reconciliation rather than optimistic retry.

Every retryable mutation carries a stable idempotency key. Reusing that identity with different arguments is a conflict, not a second action. A crash after the external effect but before its receipt leaves the effect unknown until an adapter-specific reconciliation proves what happened.

Trace, replay, and fork

Trace export includes durable state, usage precision, budget enforcement, effects, approvals, artifacts, evaluations, and event digests. It never exports hidden model reasoning.

Replay re-evaluates retained evidence without repeating consequential effects. Fork creates a new run from an eligible checkpoint with a new identity and explicit authority evaluation. Neither operation rewrites the original event stream.