feat(ai): 后台 AI 链路改走 HTTP 流式，WS 地址与心跳收敛到 utils

写作 / 图片 / 大纲 / 稿件四条链路用普通 POST 时，生成期间链路上零字节，
会被 60s 空闲超时掐成 504（出图实测 58s 断开）。后端已给每条链路开
/stream（SSE，每 15s 一帧心跳），这里改成消费它。

- 新增 utils/ai-stream.ts：SSE 帧解析 + 末帧 {type:'final', result} 拆包；
  runCreativeAssistantStream / runManuscriptGenStream 对外仍返回原来的 data，
  页面渲染逻辑不动
- 新增 utils/chat-socket.ts：buildWebSocketUrl 统一拼 /api/chat/{userId}
  （网关只放行 /api 前缀），attachHeartbeat 每 25s 发 ping 顶住代理空闲超时
- chatgpt 对话先试 WebSocket，握手失败落 /ai/chat/message/stream，
  帧内容与 WS 完全一致，处理函数直接复用
- 四个页面删掉各自重复的 buildWebSocketUrl，统一从 utils 引入

src/utils/ai-stream.ts

@@ -0,0 +1,228 @@
+import { API_BASE_URL, TOKEN_HEADER_NAME } from '@/config/setting';
+import { getSign, getToken } from '@/utils/token-util';
+
+/**
+ * 后台的 HTTP 流式通道（SSE）：WebSocket 连不上时的退路。
+ *
+ * 生产网关在 `/api` 前面就把 WebSocket 的 `Upgrade` 头丢了，握手永远到不了后端，
+ * 页面就卡在「正在建立 WebSocket 连接...」。后端因此给每条 AI 链路开了一份
+ * `/stream` 端点：请求体与原接口一模一样，响应是 `text/event-stream`，
+ * 每个 `data:` 行就是原来 WebSocket 上的一帧（同一段 JSON）。
+ *
+ * 所以页面只要把帧喂给原来的消息处理函数，落哪条通道都无感。
+ */
+
+const TENANT_ID = 10049;
+
+export interface StreamOptions {
+  /** 相对 `/api` 的路径，如 `/ai/activityOutlineGen/stream` */
+  path: string;
+  data?: Record<string, unknown> | null;
+  /** 每帧回调，内容与 WebSocket 帧一致 */
+  onMessage: (raw: string) => void;
+  /** 中止生成（页面上的「停止」按钮） */
+  signal?: AbortSignal;
+}
+
+/** 发一条流式请求，把每个 SSE 帧原样交给 onMessage。 */
+export async function streamAiRequest(options: StreamOptions): Promise<void> {
+  const payload: Record<string, unknown> = {
+    tenantId: TENANT_ID,
+    ...(options.data || {})
+  };
+  // 与 axios 拦截器保持一致：后端校验 sign
+  payload.sign = getSign(payload);
+
+  const token = getToken();
+  const headers: Record<string, string> = {
+    'Content-Type': 'application/json'
+  };
+  if (token) {
+    headers[TOKEN_HEADER_NAME] = token;
+  }
+
+  const response = await fetch(`${API_BASE_URL}${options.path}`, {
+    method: 'POST',
+    headers,
+    body: JSON.stringify(payload),
+    signal: options.signal
+  });
+
+  if (!response.ok) {
+    throw new Error(`流式请求失败（HTTP ${response.status}）`);
+  }
+  if (!response.body || typeof response.body.getReader !== 'function') {
+    throw new Error('当前浏览器不支持流式读取');
+  }
+
+  const reader = response.body.getReader();
+  const decoder = new TextDecoder('utf-8');
+  let buffer = '';
+
+  const drain = () => {
+    let boundary = buffer.indexOf('\n\n');
+    while (boundary !== -1) {
+      const frame = buffer.slice(0, boundary);
+      buffer = buffer.slice(boundary + 2);
+      boundary = buffer.indexOf('\n\n');
+      if (frame.startsWith(':')) {
+        // SSE 注释行（后端用它把响应头先顶出去）
+        continue;
+      }
+      const data = frame
+        .split('\n')
+        .filter((line) => line.startsWith('data:'))
+        .map((line) => line.slice(5).replace(/^ /, ''))
+        .join('');
+      if (data) {
+        options.onMessage(data);
+      }
+    }
+  };
+
+  for (;;) {
+    const { done, value } = await reader.read();
+    if (done) {
+      break;
+    }
+    buffer += decoder.decode(value, { stream: true });
+    drain();
+  }
+  buffer += decoder.decode();
+  drain();
+}
+
+/** 对话（HTTP 流式） */
+export function streamChatMessage(
+  data: Record<string, unknown>,
+  onMessage: (raw: string) => void,
+  signal?: AbortSignal
+): Promise<void> {
+  return streamAiRequest({
+    path: '/ai/chat/message/stream',
+    data,
+    onMessage,
+    signal
+  });
+}
+
+/** 稿件生成（HTTP 流式） */
+export function streamManuscriptGen(
+  data: Record<string, unknown>,
+  onMessage: (raw: string) => void,
+  signal?: AbortSignal
+): Promise<void> {
+  return streamAiRequest({
+    path: '/ai/manuscriptGen/stream',
+    data,
+    onMessage,
+    signal
+  });
+}
+
+/** 活动策划大纲（HTTP 流式） */
+export function streamActivityOutlineGen(
+  data: Record<string, unknown>,
+  onMessage: (raw: string) => void,
+  signal?: AbortSignal
+): Promise<void> {
+  return streamAiRequest({
+    path: '/ai/activityOutlineGen/stream',
+    data,
+    onMessage,
+    signal
+  });
+}
+
+/** 创意助手（HTTP 流式） */
+export function streamCreativeAssistant(
+  data: Record<string, unknown>,
+  onMessage: (raw: string) => void,
+  signal?: AbortSignal
+): Promise<void> {
+  return streamAiRequest({
+    path: '/ai/creativeAssistant/stream',
+    data,
+    onMessage,
+    signal
+  });
+}
+
+/**
+ * 创意助手（出图/出视频）：跑完才有结果，中途一个字节都不推。
+ *
+ * 直接用普通 POST 的话，出图那一两分钟链路上完全空闲，会在 60s 左右被网关
+ * 掐成 504 Gateway Time-out（实测 58s 断开）。走流式后后端每 15s 发一帧心跳
+ * 续命，真正的结果放在末帧 `{type:'final', result}` 里，这里拆包后原样返回
+ * 普通接口的 data 字段，页面的渲染逻辑不用改。
+ */
+export async function runCreativeAssistantStream(
+  data: Record<string, unknown>,
+  signal?: AbortSignal
+): Promise<Record<string, any> | null> {
+  let finalResult: any = null;
+  await streamCreativeAssistant(
+    data,
+    (raw) => {
+      try {
+        const frame = JSON.parse(raw);
+        if (frame && frame.type === 'final') {
+          finalResult = frame.result;
+        }
+      } catch {
+        // 业务帧解析失败就忽略：真正的失败原因由末帧的 message 带出来
+      }
+    },
+    signal
+  );
+  if (!finalResult) {
+    throw new Error('创意助手没有返回结果，请稍后重试');
+  }
+  if (Number(finalResult.code) !== 0) {
+    throw new Error(finalResult.message || '创意助手处理失败');
+  }
+  return finalResult.data ?? null;
+}
+
+/**
+ * 稿件生成：正文逐段推帧，末帧 `{type:'final', result}` 带结构化结果。
+ *
+ * 自由创作（不选文物体裁）在后端会转「AI 创意助手」，整段**不推正文**，
+ * 只有末帧有内容；模板模式则是正文边推、末帧给标题等结构。
+ * 两种都按同一条路走：帧回调照常转发，末帧拆出来当普通接口的返回值。
+ *
+ * @param onMessage 业务帧回调（正文片段），可不传
+ * @returns 结构化结果（普通接口 data 字段）；失败返回 null
+ */
+export async function runManuscriptGenStream(
+  data: Record<string, unknown>,
+  signal?: AbortSignal,
+  onMessage?: (raw: string) => void
+): Promise<Record<string, any> | null> {
+  let finalResult: any = null;
+  await streamManuscriptGen(
+    data,
+    (raw) => {
+      try {
+        const frame = JSON.parse(raw);
+        if (frame && frame.type === 'final') {
+          finalResult = frame.result;
+          return;
+        }
+      } catch {
+        // 非 JSON 帧按业务帧处理
+      }
+      if (onMessage) {
+        onMessage(raw);
+      }
+    },
+    signal
+  );
+  if (!finalResult) {
+    return null;
+  }
+  if (Number(finalResult.code) !== 0) {
+    throw new Error(finalResult.message || '稿件生成失败');
+  }
+  return finalResult.data ?? null;
+}

src/utils/chat-socket.ts

@@ -0,0 +1,53 @@
+import { API_BASE_URL } from '@/config/setting';
+
+/**
+ * 后台与大屏端（Vue）的聊天 WebSocket 细节，集中在这里。
+ *
+ * 四个页面（创意助手、活动策划大纲、稿件生成、公文校验）此前各自抄了一份
+ * 「拼 URL + 连 + 断」，路径改一次要改四处；现在统一走这里。
+ *
+ * 路径是 **`/api/chat/{userId}`**，不是 `/chat/{userId}`：生产网关只放行 `/api` 前缀，
+ * 老路径的握手请求到不了后端（实测 404）。后端同时保留了根路径端点给老版本小程序兜底，
+ * 见 `tuanwei-java` 的 `WebSocketConfig`。
+ */
+
+/** 心跳间隔。nginx / 网关的空闲超时普遍在 60s 上下，25s 留够余量。 */
+const HEARTBEAT_INTERVAL = 25000;
+
+export function buildWebSocketUrl(userId: number | string): string {
+  if (/^https?:\/\//.test(API_BASE_URL)) {
+    const url = new URL(API_BASE_URL);
+    url.protocol = url.protocol === 'https:' ? 'wss:' : 'ws:';
+    url.pathname = url.pathname.replace(/\/api\/?$/, '');
+    return `${url.toString().replace(/\/$/, '')}/api/chat/${userId}`;
+  }
+  // 相对路径（同源反代）：后台站点自己的域名下同样挂 /api/chat
+  const protocol = window.location.protocol === 'https:' ? 'wss:' : 'ws:';
+  return `${protocol}//${window.location.host}/api/chat/${userId}`;
+}
+
+/**
+ * 每 25s 发一个 `ping`，顶住反向代理与网关的空闲断开。
+ *
+ * 只发不收：页面把收到的每条消息当流式片段解析，服务端若回 pong 会被拼进正文。
+ * 服务端的 `@OnMessage` 收到文本直接丢弃。
+ *
+ * 连接关闭后定时器会自行结束（下次触发时发现 readyState 已是 CLOSED），
+ * 调用方不需要单独清理。
+ */
+export function attachHeartbeat(socket: WebSocket): () => void {
+  const timer = window.setInterval(() => {
+    if (socket.readyState === WebSocket.OPEN) {
+      try {
+        socket.send('ping');
+      } catch {
+        // 发送失败说明连接正在断，交给下一次 tick 或 onclose 处理
+      }
+      return;
+    }
+    if (socket.readyState === WebSocket.CLOSED) {
+      window.clearInterval(timer);
+    }
+  }, HEARTBEAT_INTERVAL);
+  return () => window.clearInterval(timer);
+}

src/views/gxmu/WXXZXForm/components/docValidate.vue

@@ -146,7 +146,7 @@
   import ByteMdViewer from '@/components/ByteMdViewer/index.vue';
   import { reviewWxxzxProjectBookQuality } from '@/api/gxmu/wxxzx-form';
   import { stopMessage } from '@/api/oa/chatgpt';
-  import { API_BASE_URL } from '@/config/setting';
+  import { attachHeartbeat, buildWebSocketUrl } from '@/utils/chat-socket';
   import { useUserStore } from '@/store/modules/user';
   import { copyText } from '@/utils/common';
   import type { WxxzxProjectQualityReviewResult } from '@/api/gxmu/wxxzx-form/model';
@@ -249,17 +249,6 @@
     });
   };
 
-  const buildWebSocketUrl = (userId: number) => {
-    if (/^https?:\/\//.test(API_BASE_URL)) {
-      const url = new URL(API_BASE_URL);
-      url.protocol = url.protocol === 'https:' ? 'wss:' : 'ws:';
-      url.pathname = url.pathname.replace(/\/api\/?$/, '');
-      return `${url.toString().replace(/\/$/, '')}/chat/${userId}`;
-    }
-    const protocol = window.location.protocol === 'https:' ? 'wss:' : 'ws:';
-    return `${protocol}//${window.location.host}/chat/${userId}`;
-  };
-
   const clearStreamResolver = () => {
     streamResolverRef.value = null;
   };
@@ -402,6 +391,10 @@
     const connectPromise = new Promise<void>((resolve, reject) => {
       let settled = false;
       const socket = new WebSocket(buildWebSocketUrl(userId));
+      socket.onopen = () => {
+        // 握手成功后再挂心跳：连接没建起来时发出去只会失败
+        attachHeartbeat(socket);
+      };
       cleanupPending = () => {
         if (wsConnectPromiseRef.value === connectPromise) {
           wsConnectPromiseRef.value = null;

src/views/gxmu/activityOutlineGen/index.vue

@@ -198,11 +198,12 @@
   import { asBlob } from 'html-docx-js-typescript';
   import ByteMdEditor from '@/components/ByteMdEditor/index.vue';
   import { activityOutlineGen } from '@/api/ai/activityOutlineGen';
+  import { streamActivityOutlineGen } from '@/utils/ai-stream';
   import type {
     ActivityOutlineGenParams,
     ActivityOutlineGenResponse
   } from '@/api/ai/activityOutlineGen/model';
-  import { API_BASE_URL } from '@/config/setting';
+  import { attachHeartbeat, buildWebSocketUrl } from '@/utils/chat-socket';
   import { useUserStore } from '@/store/modules/user';
   import { copyText } from '@/utils/common';
   import { resolveDisplayText, isFailedPayload, pickDisplayable, FAILED_TEXT } from '@/utils/ai-result';
@@ -422,17 +423,6 @@
     return parts.join(' | ');
   };
 
-  const buildWebSocketUrl = (userId: number) => {
-    if (/^https?:\/\//.test(API_BASE_URL)) {
-      const url = new URL(API_BASE_URL);
-      url.protocol = url.protocol === 'https:' ? 'wss:' : 'ws:';
-      url.pathname = url.pathname.replace(/\/api\/?$/, '');
-      return `${url.toString().replace(/\/$/, '')}/chat/${userId}`;
-    }
-    const protocol = window.location.protocol === 'https:' ? 'wss:' : 'ws:';
-    return `${protocol}//${window.location.host}/chat/${userId}`;
-  };
-
   const clearStreamResolver = () => {
     streamResolverRef.value = null;
   };
@@ -559,6 +549,10 @@
     const connectPromise = new Promise<void>((resolve, reject) => {
       let settled = false;
       const socket = new WebSocket(buildWebSocketUrl(userId));
+      socket.onopen = () => {
+        // 握手成功后再挂心跳：连接没建起来时发出去只会失败
+        attachHeartbeat(socket);
+      };
       cleanupPending = () => {
         if (wsConnectPromiseRef.value === connectPromise) {
           wsConnectPromiseRef.value = null;
@@ -637,23 +631,21 @@
     }
     loading.value = true;
     clearResult();
-    resultSummary.value = '正在建立 WebSocket 连接...';
+    resultSummary.value = '正在提交生成任务...';
     try {
-      await connectWebSocket();
-      resultSummary.value = '连接成功，已提交生成任务，等待流式返回...';
+      // 一律走 HTTP 流式（/stream），不再先连 WebSocket：
+      // 触发生成的那个 POST 在整段生成期间是零字节的，会被链路上的 60s 空闲
+      // 超时掐成 504 Gateway Time-out。/stream 每 15s 有后端心跳续命。
+      const payload = buildPayload();
+      resultSummary.value = '已提交生成任务，等待流式返回...';
       const waitStream = createStreamWaiter();
-      const data = await activityOutlineGen(buildPayload());
-      result.value = data || null;
-      if (!streamStarted.value && data && parseOutline(data)) {
-        outlineMarkdown.value = parseOutline(data);
-        syncEditableOutline();
-        resultSummary.value = buildSummary() || '接口直接返回生成结果';
-        loading.value = false;
-        resolveStream();
-        disconnectWebSocket();
-      } else {
-        await waitStream;
-      }
+      await streamActivityOutlineGen(
+        payload as unknown as Record<string, unknown>,
+        (raw) => handleSocketMessage(raw)
+      );
+      // 兜底：流结束但没收到 __END__ 时也别把页面卡在加载中
+      resolveStream();
+      await waitStream;
       if (!visibleOutlineMarkdown.value) {
         message.warning('接口已返回，但未解析到正文内容');
         return;

src/views/gxmu/creativeAssistant/index.vue

@@ -189,7 +189,7 @@
   import { computed, ref } from 'vue';
   import { message } from 'ant-design-vue';
   import ByteMdViewer from '@/components/ByteMdViewer/index.vue';
-  import { creativeAssistant } from '@/api/ai/creativeAssistant';
+  import { runCreativeAssistantStream } from '@/utils/ai-stream';
   import type { FileRecord } from '@/api/system/file/model';
   import type {
     CreativeAssistantOutputs,
@@ -404,7 +404,8 @@
     }
     loading.value = true;
     try {
-      const result = await creativeAssistant(buildRequestPayload());
+      // 出图中途零字节，普通 POST 会被网关按 60s 空闲超时掐成 504，改走流式
+      const result = await runCreativeAssistantStream(buildRequestPayload());
       workflowResponse.value = result || null;
       workflowOutputs.value = normalizeOutputPayload(result || null);
       if (

src/views/gxmu/manuscriptGen/index.vue

@@ -229,13 +229,14 @@
   import { asBlob } from 'html-docx-js-typescript';
   import { getProcessor } from 'bytemd';
   import { manuscriptGen } from '@/api/ai/manuscriptGen';
+  import { runManuscriptGenStream } from '@/utils/ai-stream';
   import type {
     ManuscriptGenParams,
     ManuscriptGenResult
   } from '@/api/ai/manuscriptGen/model';
   import { uploadFile } from '@/api/system/file';
   import type { FileRecord } from '@/api/system/file/model';
-  import { API_BASE_URL } from '@/config/setting';
+  import { attachHeartbeat, buildWebSocketUrl } from '@/utils/chat-socket';
   import { useUserStore } from '@/store/modules/user';
   import { copyText } from '@/utils/common';
   import { resolveDisplayText, isFailedPayload, pickDisplayable, FAILED_TEXT } from '@/utils/ai-result';
@@ -560,17 +561,6 @@
     return typeof payload === 'object';
   };
 
-  const buildWebSocketUrl = (userId: number) => {
-    if (/^https?:\/\//.test(API_BASE_URL)) {
-      const url = new URL(API_BASE_URL);
-      url.protocol = url.protocol === 'https:' ? 'wss:' : 'ws:';
-      url.pathname = url.pathname.replace(/\/api\/?$/, '');
-      return `${url.toString().replace(/\/$/, '')}/chat/${userId}`;
-    }
-    const protocol = window.location.protocol === 'https:' ? 'wss:' : 'ws:';
-    return `${protocol}//${window.location.host}/chat/${userId}`;
-  };
-
   const clearStreamResolver = () => {
     streamResolverRef.value = null;
   };
@@ -696,6 +686,7 @@
         }
       };
       socket.onopen = () => {
+        attachHeartbeat(socket);
         wsConnected.value = false;
       };
       socket.onmessage = (event) => {
@@ -859,25 +850,34 @@
     loading.value = true;
     clearGenerated();
     generatedTitle.value = form.theme.trim();
-    responseSummary.value = '正在建立 WebSocket 连接...';
+    responseSummary.value = '正在提交生成任务...';
     try {
-      await connectWebSocket();
+      // 一律走 HTTP 流式（/stream），不再先连 WebSocket：
+      // 触发生成的那个 POST 在整段生成期间是零字节的，会被链路上的 60s 空闲
+      // 超时掐成 504 Gateway Time-out（自由创作、不选文体时最先撞上）。
+      // /stream 每 15s 有后端心跳续命，正文逐段喂给原来的帧处理，
+      // 结构化结果从末帧 {type:'final'} 取。
+      const payload = buildPayload();
       responseSummary.value = templateMode.value
-        ? '连接成功，已提交生成任务，等待流式返回...'
+        ? '已提交生成任务，等待流式返回...'
         : '自由创作已提交给 AI 创意助手，生成完成后一次性返回...';
       const waitStream = createStreamWaiter();
-      const result = await manuscriptGen(buildPayload());
-      if (!streamStarted.value && shouldTreatAsDirectResult(result)) {
+      const result = await runManuscriptGenStream(
+        payload as unknown as Record<string, unknown>,
+        undefined,
+        (raw) => handleSocketMessage(raw)
+      );
+      if (result) {
         applyParsedResult(
-          typeof result === 'string' ? parseSerializedResult(result) : (result as ManuscriptGenResult),
+          typeof result === 'string'
+            ? parseSerializedResult(result)
+            : (result as ManuscriptGenResult),
           '接口直接返回生成结果'
         );
-        loading.value = false;
-        resolveStream();
-        disconnectWebSocket();
-      } else {
-        await waitStream;
       }
+      // 兜底：流结束但没收到 __END__ 时也别把页面卡在加载中
+      resolveStream();
+      await waitStream;
       message.success('内容生成完成');
     } catch (e: any) {
       disconnectWebSocket();

src/views/oa/chatgpt/index.vue

@@ -389,15 +389,14 @@ import type { AiChatHistory } from "@/api/ai/aiChatHistory/model";
 import { listAiChatList, removeAiChatList } from "@/api/ai/aiChatList";
 import type { AiChatList } from "@/api/ai/aiChatList/model";
 import { sendMessage, stopMessage } from "@/api/oa/chatgpt";
-import { manuscriptGen } from "@/api/ai/manuscriptGen";
+import { runCreativeAssistantStream, runManuscriptGenStream, streamChatMessage } from "@/utils/ai-stream";
 import {
     WRITING_DOC_TYPE_FREE,
     WRITING_DOC_TYPE_OPTIONS,
     WRITING_EVENT_TYPE_OPTIONS,
     WRITING_WORD_COUNT_OPTIONS
 } from "./writing-params";
-import { creativeAssistant } from "@/api/ai/creativeAssistant";
-import { API_BASE_URL } from "@/config/setting";
+import { attachHeartbeat, buildWebSocketUrl } from "@/utils/chat-socket";
 import type { Menu } from "@/api/system/menu/model";
 import { useUserStore } from "@/store/modules/user";
 import { logout } from "@/utils/page-tab-util";
@@ -785,17 +784,6 @@ const findFirstChildMenuPath = (menus: AuthorityTreeItem[]) => {
     return "";
 };
 
-const buildWebSocketUrl = (userId: number) => {
-    if (/^https?:\/\//.test(API_BASE_URL)) {
-        const url = new URL(API_BASE_URL);
-        url.protocol = url.protocol === "https:" ? "wss:" : "ws:";
-        url.pathname = url.pathname.replace(/\/api\/?$/, "");
-        return `${url.toString().replace(/\/$/, "")}/chat/${userId}`;
-    }
-    const protocol = window.location.protocol === "https:" ? "wss:" : "ws:";
-    return `${protocol}//${window.location.host}/chat/${userId}`;
-};
-
 const formatHistoryMessages = (list: AiChatHistory[]) => {
     const result: ChatMessageItem[] = [];
     [...list]
@@ -972,6 +960,7 @@ const connectWebSocket = async () => {
             }
         };
         socket.onopen = () => {
+          attachHeartbeat(socket);
             wsConnected.value = false;
         };
         socket.onmessage = (event) => {
@@ -1335,7 +1324,8 @@ const sendModeMessage = async (
         }
     };
     if (activeChatMode.value === "writing") {
-        const result = await manuscriptGen(buildWritingPayload(query), { signal });
+        // 稿件生成同样会被链路的 60s 空闲超时掐掉，改走流式（正文边推边显示）
+        const result = await runManuscriptGenStream(buildWritingPayload(query), signal);
         if (typeof (result as any)?.taskId === "string") {
             currentTaskId.value = (result as any).taskId;
         }
@@ -1343,7 +1333,8 @@ const sendModeMessage = async (
         return;
     }
     if (activeChatMode.value === "image") {
-        const result = await creativeAssistant(buildImagePayload(query), { signal });
+        // 出图中途零字节，普通 POST 会被网关按 60s 空闲超时掐成 504，改走流式
+        const result = await runCreativeAssistantStream(buildImagePayload(query), signal);
         if (typeof (result as any)?.taskId === "string") {
             currentTaskId.value = (result as any).taskId;
         }
@@ -1471,13 +1462,24 @@ const onSend = async () => {
         content.value = "";
         scrollToBottom();
         if (activeChatMode.value === "chat") {
-            await connectWebSocket();
-            await sendMessage({
+            // 先试 WebSocket；生产网关在 /api 前丢 Upgrade 头，握手必然失败，
+            // 这时落到 HTTP 流式端点，帧内容与 WS 完全一致，解析逻辑不用改
+            const payload = {
                 query,
                 conversationId: activeConversationId.value
-            }, {
-                signal: abortController.signal
-            });
+            };
+            const viaSocket = await connectWebSocket()
+                .then(() => true)
+                .catch(() => false);
+            if (viaSocket) {
+                await sendMessage(payload, { signal: abortController.signal });
+            } else {
+                await streamChatMessage(
+                    payload as Record<string, unknown>,
+                    (raw) => void handleSocketMessage(raw),
+                    abortController.signal
+                );
+            }
         } else {
             await sendModeMessage(query, assistantId, abortController.signal);
             streamingMessageId.value = undefined;