@pushwoosh/rpc-gateway-ai-assistant 0.1.183 → 0.1.184

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@pushwoosh/rpc-gateway-ai-assistant",
-    "version": "0.1.183",
+    "version": "0.1.184",
     "description": "AI Assistant api gateway HTTP API Types and Data",
     "main": "index.js",
     "module": "index.js",

package/pushwoosh_accounts_assistant_v1.d.ts

@@ -40,13 +40,14 @@ export type Chat = {
     updatedAt: Date;
     /**
      * is_onboarding flags this chat as the account's onboarding chat.
-     * Frontend may render the chat with onboarding-specific UI. Inference for
-     * onboarding chats is routed exclusively to StreamInferenceLLM — other
-     * StreamInference* endpoints are not used while is_onboarding=true. The
-     * very first turn is initiated by the frontend calling StreamInferenceLLM
-     * with an empty message on a chat that has no message history; the server
-     * synthesises an invisible kickoff trigger and the onboarding agent emits
-     * the opening assistant message + suggested actions itself.
+     * Frontend may render the chat with onboarding-specific UI. The same
+     * onboarding agent runs on either inference provider — StreamInferenceLLM
+     * (self-hosted Qwen) or StreamInference (Claude); the frontend picks by
+     * which endpoint it calls for the chat. The opening turn is an ordinary
+     * turn: the frontend sends a real first user message (e.g. a "Get started"
+     * button prompt) and the onboarding agent replies with the greeting +
+     * suggested actions. An empty message is rejected (InvalidArgument) — there
+     * is no server-synthesised kickoff.
      */
     isOnboarding: boolean;
     /**
@@ -285,12 +286,11 @@ export interface AssistantService {
         signal?: AbortSignal;
     }): Promise<AsyncIterable<InferenceResponse>>;
     /**
-     * StreamInferenceLLM is the self-hosted LLM (Qwen) inference endpoint.
-     * Frontend routes onboarding chats (is_onboarding=true) here exclusively;
-     * other StreamInference* endpoints are reserved for regular chats. For
-     * the first turn on a fresh onboarding chat the frontend calls this RPC
-     * with an empty message; the server synthesises a kickoff trigger and
-     * the onboarding agent produces the opening assistant message itself.
+     * StreamInferenceLLM is the self-hosted LLM (Qwen) inference endpoint. It is
+     * one of two providers that can run the onboarding agent — the frontend
+     * routes an onboarding chat (is_onboarding=true) here or to StreamInference
+     * (Claude) by choosing the endpoint. Every turn, including the opening one,
+     * carries a real user message; an empty message is rejected. See is_onboarding.
      */
     StreamInferenceLLM(request: InferenceRequest, options?: {
         signal?: AbortSignal;