Anchorage API reference
    Preparing search index...

    Interface AgentThreadStartInput

    interface AgentThreadStartInput {
        agentId: string;
        dispatchId?: string;
        entryPath:
            | "approval.resume"
            | "http.start"
            | "signal.message"
            | "signal.queue"
            | "signal.reactive"
            | "signal.state"
            | "signal.notification"
            | "signal.wake"
            | "notification.dispatch"
            | "schedule.fire";
        idempotencyKey?: string;
        prompt: string;
        providerOptions?: Record<string, unknown>;
        requestContext?: Record<string, unknown>;
        resourceId?: string;
        runId?: string;
        scheduleId?: string;
        streamRequestContext?: Record<string, unknown>;
        threaded?: boolean;
        threadId?: string;
        topologyThreadId?: string;
    }
    Index

    Properties

    agentId dispatchId? entryPath idempotencyKey? prompt providerOptions? requestContext? resourceId? runId? scheduleId? streamRequestContext? threaded? threadId? topologyThreadId?

    Properties

    agentId: string
    dispatchId?: string

    Prepared schedule trigger authorizing this exact run.

    entryPath:
        | "approval.resume"
        | "http.start"
        | "signal.message"
        | "signal.queue"
        | "signal.reactive"
        | "signal.state"
        | "signal.notification"
        | "signal.wake"
        | "notification.dispatch"
        | "schedule.fire"
    idempotencyKey?: string

    Makes this start exactly-once for this principal: a retry carrying the same key converges onto the run the first call made, on whatever thread it made it, instead of starting a second one.

    TRUSTED-CALLER ONLY, and not by omission: the public agent router rejects every body field but prompt, so a key can only arrive from a host seam — a schedule adapter, a delegating agent, a host's own start path. That is the same posture requestContext and providerOptions take here, and for the same reason: a tenant able to name a key could converge onto, or collide with, a run it did not make. (Ownership makes the collision refuse rather than succeed, but a tenant-reachable key would still be a probe.)

    prompt: string
    providerOptions?: Record<string, unknown>

    Trusted, JSON-safe model provider options from an internal schedule.

    requestContext?: Record<string, unknown>

    Trusted stored context, used by schedule and wake adapters only. The topology strips runtime-owned keys before forwarding it to the host.

    resourceId?: string
    runId?: string
    scheduleId?: string

    Required provenance for a schedule.fire target.

    streamRequestContext?: Record<string, unknown>

    Trusted stream-call context layered over requestContext. Public HTTP starts do not expose either context field.

    threaded?: boolean
    threadId?: string
    topologyThreadId?: string

    Trusted host-only DO address for an ephemeral unthreaded execution.

    Settings

    Member Visibility

    On This Page

    Properties