A Run is the unit of agent execution. Every invocation of an agent — from the CLI, the API, an MCP tool, or a watcher — produces a Run with:Documentation Index
Fetch the complete documentation index at: https://docs.protean.peerislands.io/llms.txt
Use this file to discover all available pages before exploring further.
runId— opaque identifierseed— deterministic source for clock + RNGcassetteId— pinned LLM responses (null when running live)events[]— typed event log (turn.start,text.delta,tool_call.start,tool_call.end,finish,error)state—queued | running | succeeded | failed | cancelled
@protean/shared-types.
Why this shape
The event log is the source of truth, not in-memory state. The SDK, UI, and CLI all consume the same stream. Replay re-emits the events in order; the agent code is the same code path that ran the first time. This is the foundation everything else (replay, observability, time-travel debugging) sits on. Seereplay for the gate semantics.