Execution

Status: Stable.

Normative home: selfHostedRunner, subWorkflow, multiAgent.

Why this exists

A step can run on a user's machine, in a child run, or under a supervisor loop; the host stays the orchestration, persistence and replay authority.

selfHostedRunner

A runner is a user-operated process that dials out, holds credentials the host cannot reach, and executes single model or tool steps. It receives dispatch frames over SSE and POSTs result frames, on host-defined paths. A host MUST NOT advertise selfHostedRunner unless it accepts registrations, routes matching dispatch and delivers results.

• Records. A registration (schemas/v2/self-hosted-runner-registration.schema.json) MUST NOT appear in discovery. dispatchKinds lists model, tool or both.
• Subject isolation. A host MUST NOT route a step to a runner its run's subject does not own, and MUST match on subject before capability.
• Credentials. A runner credential MUST NOT transit the host or appear in any frame, event, result, debug bundle or log. The runner bearer MUST NOT be a provider credential.
• Untrusted output. Runner output re-entering an agent loop MUST be fenced as untrusted.
• At most once. A frame's seq is a per-runner cursor, resumed by Last-Event-ID, and MUST NOT be conflated with event sequence. The host MUST drop, not re-dispatch, a {runId, stepId} already persisted; the runner MUST answer under the pair it received.
• Liveness. Dispatch to a departed runner MUST fail with retriable runner_unavailable, never hang.
• Replay and fork. Replay MUST NOT re-dispatch. A fork MUST NOT pin later dispatch to the original runnerId.

subWorkflow

core.subWorkflow starts a child run of another workflow and waits for its terminal status.

• workflowId. A host MUST refuse the parent run when that workflow is not loaded.
• waitForCompletion defaults to true; a host MAY refuse false with validation_error.
• onChildFailure. fail-parent (default) fails the node and the run; absorb records the failure and continues.
• Output. node.completed MUST carry outputs.childRunId and outputs.childStatus (completed | failed | cancelled), and MAY add fields.
• inputMapping (childVar → parentVar) seeds the child once, at creation, after and over its variables[].defaultValue, which MUST seed first. An unset parent variable MUST arrive undefined, never an error or null. A host not advertising subWorkflow.inputMapping MUST refuse a non-empty inputMapping at registration with validation_error, naming it in details.requiredCapability.
• outputMapping (parentVar → childVar). After the child completes, the host MUST copy each mapped variable into the parent, without throwing on or copying an undefined one.
• propagateCancellation (default true) cancels the child with its parent.

multiAgent

multiAgent.executionModel.version is cumulative: a host advertising N MUST implement levels 1 through N.

Levels 2–6 are not yet restated in v2 prose. Their rules are those of their owning RFCs: level 2 RFC 0039, level 3 RFC 0040, level 4 RFC 0041, level 5 RFC 0061 and RFC 0111, level 6 RFC 0090.

At level 1, a workflow whose core.orchestrator.supervisor feeds core.dispatch runs this loop:

• Each turn records one decision as runOrchestrator.decided. terminate completes the run; clarify and escalate suspend on a clarification or approval interrupt; next-worker dispatches each of nextWorkerIds[] as a child run, and the next turn waits for all of them.
• The loop MUST be re-entrant: replay from any iteration MUST reproduce its state.
• A worker moves pending → dispatching → running → harvested, or ends failed or cancelled. Each transition that occurs MUST emit core.workflowChain.event whose causationId names the prior transition, first runOrchestrator.decided. A host MUST NOT synthesize a phase it never produces.
• output.harvested fires exactly when a child completes with a non-empty outputMapping.
• A host not advertising executionModel MUST NOT emit core.workflowChain.event.

Sources: RFCs 0007, 0022, 0037, 0122.