rig_author

ai.rokha/rokha · Verify this server

Build + save a multi-node Rig as a GRAPH the runtime executes for real — the agent-side mirror of the human Rig Builder canvas. `structure_type`: 'dag' (acyclic; parallel + real fan-in) or 'stateful_graph' (cyclic loops + conditional routing). `nodes`: each {id, skill (a Registry skill name — discover with registry_search), label?, instruction?, doc_input?, model?, model_policy?, endpoint?, tool?, params?, agent?}. `agent` names a SUB-AGENT persona handle as the step's ACTOR — the step's agentic/instruction execution runs AS that persona (your own agent always; someone else's only while public; a private/missing actor is a typed agent_unavailable step error). `endpoint` (+ `tool` + `params`) wires the node to a LIVE MCP server — the run performs a REAL tools/call there (empty endpoint = the step runs agentically). `doc_input: true` = a LIGHT LISTING step (the run does a REAL registry lookup + document fetch of the node's skill — no runtime needed). `model` pins the Anthropic model for the node's agentic execution (e.g. 'claude-sonnet-4-6'); `model_policy` 'preferred' (default — swap to the best usable model) or 'required' (the step refuses without it, typed model_unavailable; non-Haiku models need the owner's own API key). `edges`: each {from, to} node id (a dag rejects cycles; a stateful_graph must reach an exit). After authoring it becomes the caller's saved Working Rig, runnable via the run endpoint.