Enhance LLM text streaming and message handling in backend and frontend

- Introduce event handlers in PassthroughLLMAssistantAggregator for managing LLM text streaming, including start, delta, and end events.
- Implement a new method to finalize text streams, ensuring proper handling of interruptions.
- Update useVoicePreview to support new message types for LLM text streaming, allowing real-time updates to chat messages.
- Enhance message sorting logic to maintain order based on timestamps and sequence numbers, improving user experience during voice interactions.

frontend/src/hooks/use-voice-preview.ts

@@ -29,6 +29,9 @@ export type ChatMessage = {
   content: string;
   /** 后端给的 ISO 时间戳 */
   timestamp: string;
+  sequence: number;
+  turnId?: string;
+  streaming?: boolean;
 };
 
 // http→ws、https→wss,自动跟随 API 基址(同源反代时也对)
@@ -59,6 +62,12 @@ function messageOrder(message: ChatMessage): number {
   return Number.isNaN(timestamp) ? Number.MAX_SAFE_INTEGER : timestamp;
 }
 
+function sortMessages(messages: ChatMessage[]): ChatMessage[] {
+  return messages.sort(
+    (a, b) => messageOrder(a) - messageOrder(b) || a.sequence - b.sequence,
+  );
+}
+
 function microphoneErrorMessage(error: unknown): string {
   if (error instanceof DOMException) {
     if (error.name === "NotAllowedError") {
@@ -303,6 +312,54 @@ export function useVoicePreview(assistantId: string | null) {
         try {
           const msg = JSON.parse(event.data);
           if (
+            msg?.type === "assistant-text-start" &&
+            typeof msg.turn_id === "string"
+          ) {
+            messageSeqRef.current += 1;
+            const next: ChatMessage = {
+              id: `assistant-${msg.turn_id}`,
+              role: "assistant",
+              content: "",
+              timestamp:
+                typeof msg.timestamp === "string"
+                  ? msg.timestamp
+                  : new Date().toISOString(),
+              sequence: messageSeqRef.current,
+              turnId: msg.turn_id,
+              streaming: true,
+            };
+            setMessages((prev) => sortMessages([...prev, next]));
+          } else if (
+            msg?.type === "assistant-text-delta" &&
+            typeof msg.turn_id === "string" &&
+            typeof msg.delta === "string"
+          ) {
+            setMessages((prev) =>
+              prev.map((message) =>
+                message.turnId === msg.turn_id
+                  ? { ...message, content: message.content + msg.delta }
+                  : message,
+              ),
+            );
+          } else if (
+            msg?.type === "assistant-text-end" &&
+            typeof msg.turn_id === "string"
+          ) {
+            setMessages((prev) =>
+              prev.map((message) =>
+                message.turnId === msg.turn_id
+                  ? {
+                      ...message,
+                      content:
+                        typeof msg.content === "string"
+                          ? msg.content
+                          : message.content,
+                      streaming: false,
+                    }
+                  : message,
+              ),
+            );
+          } else if (
             msg?.type === "transcript" &&
             (msg.role === "user" || msg.role === "assistant") &&
             typeof msg.content === "string" &&
@@ -317,15 +374,9 @@ export function useVoicePreview(assistantId: string | null) {
                 typeof msg.timestamp === "string"
                   ? msg.timestamp
                   : new Date().toISOString(),
+              sequence: messageSeqRef.current,
             };
-            setMessages((prev) =>
-              [...prev, next].sort(
-                (a, b) =>
-                  messageOrder(a) - messageOrder(b) ||
-                  Number(a.id.replace("msg-", "")) -
-                    Number(b.id.replace("msg-", "")),
-              ),
-            );
+            setMessages((prev) => sortMessages([...prev, next]));
           }
         } catch {
           /* 非 JSON / 未知消息,忽略 */