diff --git a/frontend/src/api/chat.ts b/frontend/src/api/chat.ts
index 4331a79..592c908 100644
--- a/frontend/src/api/chat.ts
+++ b/frontend/src/api/chat.ts
@@ -1,6 +1,5 @@
-import request from './request'
 import { getToken } from '@/utils/token'
-import type { ApiResponse } from '@/types/api'
+import type { ChatResult } from '@/types/sse'
 
 const API_BASE = ''
 
@@ -28,13 +27,11 @@ export interface ChatOptions {
  * @param message 用户消息
  * @param chatId 会话 ID
  * @param options 可选参数
- * @param includeEnableRag 是否写入 enableRag 参数（sources 接口无意义，需排除）
  */
 function buildChatQueryParams(
   message: string,
   chatId: string,
   options?: ChatOptions,
-  includeEnableRag = true,
 ): URLSearchParams {
   const params = new URLSearchParams()
   params.set('message', message)
@@ -42,7 +39,7 @@ function buildChatQueryParams(
   if (options) {
     if (options.roleId) params.set('roleId', options.roleId)
     if (options.accountId) params.set('accountId', options.accountId)
-    if (includeEnableRag && options.enableRag === true) params.set('enableRag', 'true')
+    if (options.enableRag === true) params.set('enableRag', 'true')
     if (options.rewriteStrategy) params.set('rewriteStrategy', options.rewriteStrategy)
     if (options.categoryId !== undefined && options.categoryId !== '') params.set('categoryId', String(options.categoryId))
     if (options.categoryIds && options.categoryIds.length > 0) params.set('categoryIds', options.categoryIds.join(','))
@@ -51,10 +48,17 @@ function buildChatQueryParams(
   return params
 }
 
-/** 同步对话：GET /ai/chat */
-export function chatSync(message: string, chatId: string, options?: ChatOptions): Promise<string> {
-  const url = `${API_BASE}/ai/chat?${buildChatQueryParams(message, chatId, options).toString()}`
-  return fetch(url, { headers: authHeaders() }).then(res => res.text())
+/** 同次回答和引用来源：GET /ai/chat/result（直接 JSON，无 data 包裹）。 */
+export function chatSync(message: string, chatId: string, options?: ChatOptions, signal?: AbortSignal): Promise<ChatResult> {
+  const url = `${API_BASE}/ai/chat/result?${buildChatQueryParams(message, chatId, options).toString()}`
+  return fetchChatResult(url, authHeaders(), signal)
+}
+
+/** SDK 测试页也复用此响应处理，保留其独立域名与 SDK Token。 */
+export async function fetchChatResult(url: string, headers: Record<string, string>, signal?: AbortSignal): Promise<ChatResult> {
+  const res = await fetch(url, { headers, signal })
+  if (!res.ok) throw new Error((await res.text()) || 'HTTP ' + res.status)
+  return res.json()
 }
 
 /** 获取 SSE 流式对话 URL：GET /ai/chat/stream */
@@ -62,12 +66,6 @@ export function chatSSEUrl(message: string, chatId: string, options?: ChatOption
   return `${API_BASE}/ai/chat/stream?${buildChatQueryParams(message, chatId, options).toString()}`
 }
 
-/** 获取 RAG 引用来源：GET /ai/chat/sources（不传 enableRag，其余参数照传） */
-export function ragSources(message: string, chatId: string, options?: ChatOptions): Promise<ApiResponse> {
-  const path = `/ai/chat/sources?${buildChatQueryParams(message, chatId, options, false).toString()}`
-  return request.get(path).then(r => r.data)
-}
-
 /**
  * 获取 AI 推荐问题列表（管理后台 ChatPanel 用）。
  * 调用 /conversation/{id}/suggestions，由管理后台 JwtAuthFilter 守卫。
diff --git a/frontend/src/components/MessageSources.vue b/frontend/src/components/MessageSources.vue
index d76c7aa..92b1ab8 100644
--- a/frontend/src/components/MessageSources.vue
+++ b/frontend/src/components/MessageSources.vue
@@ -5,9 +5,9 @@
         <div v-for="(group, key) in grouped" :key="key" class="source-group">
           <div class="source-doc-title">📄 {{ group.title || '文档 ' + key }}</div>
           <div class="source-chunks">
-            <div v-for="chunk in group.chunks" :key="chunk.chunkIndex" class="source-chunk">
-              <span class="chunk-badge">#{{ chunk.chunkIndex }}</span>
-              <span v-html="renderMarkdown(chunk.content?.substring(0, 300) + (chunk.content?.length > 300 ? '...' : ''))"></span>
+            <div v-for="(chunk, index) in group.chunks" :key="index" class="source-chunk">
+              <span v-if="chunk.chunkIndex !== null" class="chunk-badge">#{{ chunk.chunkIndex }}</span>
+              <span v-html="renderMarkdown(chunk.snippet || '')"></span>
             </div>
           </div>
         </div>
@@ -19,17 +19,18 @@
 <script setup lang="ts">
 import { computed } from 'vue'
 import { renderMarkdown } from '@/utils/markdown'
+import type { SourceReference } from '@/types/sse'
 
-const props = defineProps<{ sources: any[] }>()
+const props = defineProps<{ sources: SourceReference[] }>()
 
 /** 按文档 ID 归并 chunk */
 const grouped = computed(() => {
-  const groups: Record<string, { title?: string; chunks: any[] }> = {}
+  const groups: Record<string, { title: string | null; chunks: SourceReference[] }> = {}
   for (const s of (props.sources || [])) {
-    const docId = s.documentId || s.metadata?.documentId || 'unknown'
+    const docId = s.documentId || s.sourceName || s.title || 'unknown'
     if (!groups[docId]) {
       groups[docId] = {
-        title: s.metadata?.title || s.sourceName,
+        title: s.title || s.sourceName,
         chunks: [],
       }
     }
diff --git a/frontend/src/sdk-test/SdkTestPanel.vue b/frontend/src/sdk-test/SdkTestPanel.vue
index 0b12b6d..ac6a5b6 100644
--- a/frontend/src/sdk-test/SdkTestPanel.vue
+++ b/frontend/src/sdk-test/SdkTestPanel.vue
@@ -282,13 +282,13 @@
 </template>
 
 <script setup lang="ts">
-import { ref, reactive, computed, onMounted, onBeforeUnmount, nextTick } from 'vue'
+import { ref, reactive, computed, watch, onMounted, onBeforeUnmount, nextTick } from 'vue'
 import { ChatList, ChatSender, ChatActionbar } from '@tdesign-vue-next/chat'
 import '@tdesign-vue-next/chat/es/style/index.css'
 import { MessagePlugin } from 'tdesign-vue-next'
-import { readSSEStream } from '@/utils/sse'
+import { readSSEStreamWithEvents } from '@/utils/sse'
+import { fetchChatResult } from '@/api/chat'
 import { renderMarkdown } from '@/utils/markdown'
-import { toast } from '@/utils/toast'
 import { toChatData, type ChatMessage } from '@/utils/chatAdapter'
 
 // ==================== 常量 ====================
@@ -309,11 +309,11 @@ const modeOptions = [
   { label: '同步调用', value: 'sync' },
 ]
 const strategyOptions = [
-  { label: '不重写', value: 'NONE' },
-  { label: '查询重写', value: 'REWRITE' },
-  { label: '翻译扩展', value: 'TRANSLATION' },
-  { label: '查询压缩', value: 'COMPRESSION' },
-  { label: '多路扩展', value: 'MULTI_QUERY' },
+  { label: '不重写（最快）', value: 'NONE' },
+  { label: '查询重写（额外耗时）', value: 'REWRITE' },
+  { label: '翻译扩展（额外耗时）', value: 'TRANSLATION' },
+  { label: '查询压缩（额外耗时）', value: 'COMPRESSION' },
+  { label: '多路扩展（额外耗时）', value: 'MULTI_QUERY' },
 ]
 
 // SDK 可选参数的默认值（与 SDK parseConfig 对齐），用于判断是否需要在生成代码中输出
@@ -609,7 +609,7 @@ function domAction(act: 'open' | 'close' | 'toggle'): void {
 const activeTab = ref<'chat' | 'sdk'>('chat')
 const demoMode = ref<'sse' | 'sync'>('sse')
 const demoRag = ref(true)
-const demoStrategy = ref('MULTI_QUERY')
+const demoStrategy = ref('NONE')
 const demoInput = ref('')
 const demoChatId = ref('')
 const isSending = ref(false)
@@ -626,7 +626,7 @@ function welcomeMessage(): ChatMessage {
   return {
     id: genId(),
     role: 'assistant',
-    content: '您好，这里是新接口（GET /ai/chat、/ai/chat/stream、/ai/chat/sources）对话演示区。请先在左侧获取 SDK Token，然后输入问题开始对话。',
+    content: '您好，这里是对话演示区（GET /ai/chat/result、/ai/chat/stream），回答会同时返回实际引用来源。请先在左侧获取 SDK Token，然后输入问题开始对话。',
     streaming: false,
     time: fmtTime(),
   }
@@ -636,7 +636,7 @@ const demoMessages = ref<ChatMessage[]>([welcomeMessage()])
 const demoData = computed(() => toChatData(demoMessages.value))
 
 function buildDemoUrl(
-  path: 'chat' | 'chat/stream' | 'chat/sources',
+  path: 'chat/result' | 'chat/stream',
   message: string,
   chatId: string,
   opts: { enableRag: boolean; rewriteStrategy?: string },
@@ -648,7 +648,7 @@ function buildDemoUrl(
   if (rid) p.set('roleId', rid)
   const uid = config.userId.trim()
   if (uid) p.set('accountId', uid)
-  if (path !== 'chat/sources' && opts.enableRag) p.set('enableRag', 'true')
+  if (opts.enableRag) p.set('enableRag', 'true')
   if (opts.enableRag && opts.rewriteStrategy) p.set('rewriteStrategy', opts.rewriteStrategy)
   return demoDomain.value + '/ai/' + path + '?' + p.toString()
 }
@@ -679,48 +679,45 @@ async function sendDemo(val?: string): Promise<void> {
   const requestMode = demoMode.value
   const opts = { enableRag: demoRag.value, rewriteStrategy: demoRag.value ? demoStrategy.value : undefined }
   // 提前构建 URL，固定本轮角色、账号、域名和检索配置。
-  const url = buildDemoUrl(requestMode === 'sync' ? 'chat' : 'chat/stream', text, cid, opts)
-  const sourcesUrl = opts.enableRag ? buildDemoUrl('chat/sources', text, cid, opts) : ''
+  const url = buildDemoUrl(requestMode === 'sync' ? 'chat/result' : 'chat/stream', text, cid, opts)
   const headers = sdkAuthHeaders()
+  const controller = new AbortController()
+  abortController?.abort()
+  abortController = controller
+  const isCurrent = () => abortController === controller
+    && !controller.signal.aborted && demoChatId.value === cid
+    && demoMessages.value.some(msg => msg.id === assistantMsg.id)
   await scrollDemoBottom()
 
   try {
-    if (abortController) abortController.abort()
-    abortController = new AbortController()
-
+    if (!isCurrent()) return
     if (requestMode === 'sync') {
-      const res = await fetch(url, { headers, signal: abortController.signal })
-      if (!res.ok) throw new Error('HTTP ' + res.status)
-      assistantMsg.content = await res.text()
+      const result = await fetchChatResult(url, headers, controller.signal)
+      if (!isCurrent()) return
+      assistantMsg.content = result.text
+      assistantMsg.sources = result.sources
     } else {
-      // 复用 readSSEStream（内部手写解析 OpenAI delta.content），逐 chunk 追加
-      await readSSEStream(
-        url,
-        (chunk: string) => {
+      await readSSEStreamWithEvents(url, {
+        onMessage: (chunk) => {
+          if (!isCurrent()) return
           assistantMsg.content += chunk
           demoMessages.value = [...demoMessages.value]
+          void scrollDemoBottom()
         },
-        undefined,
-        headers,
-        abortController.signal,
-      )
-    }
-
-    // 引用来源后台补齐，正文完成后立即结束发送状态。
-    if (opts.enableRag) {
-      void fetch(sourcesUrl, { headers }).then(async sres => {
-        const json = await sres.json()
-        if (!sres.ok || !json?.success) {
-          throw new Error(json?.message || (sres.ok ? '引用来源获取失败' : 'HTTP ' + sres.status))
-        }
-        if (demoChatId.value !== cid || !demoMessages.value.some(msg => msg.id === assistantMsg.id)) return
-        assistantMsg.sources = json.data || []
-        demoMessages.value = [...demoMessages.value]
-      }).catch((e: any) => {
-        toast('引用来源获取失败: ' + (e.message || e), 'warning')
-      })
+        onSources: (sources) => {
+          if (!isCurrent()) return
+          assistantMsg.sources = sources
+          demoMessages.value = [...demoMessages.value]
+        },
+        onError: (data) => {
+          if (!isCurrent()) return
+          assistantMsg.content += '\n\n' + (data.message || '工具调用出错')
+          demoMessages.value = [...demoMessages.value]
+        },
+      }, headers, controller.signal)
     }
   } catch (e: any) {
+    if (!isCurrent()) return
     if (e.name === 'AbortError') {
       assistantMsg.content = assistantMsg.content || '已取消'
     } else {
@@ -728,10 +725,12 @@ async function sendDemo(val?: string): Promise<void> {
       assistantMsg.error = true
     }
   } finally {
-    assistantMsg.streaming = false
-    isSending.value = false
-    demoMessages.value = [...demoMessages.value]
-    await scrollDemoBottom()
+    if (isCurrent()) {
+      assistantMsg.streaming = false
+      isSending.value = false
+      demoMessages.value = [...demoMessages.value]
+      await scrollDemoBottom()
+    }
   }
 }
 
@@ -752,6 +751,13 @@ function clearDemo(): void {
   demoMessages.value = [welcomeMessage()]
 }
 
+// Changing the SDK identity starts a new conversation and invalidates pending callbacks.
+watch(
+  [() => config.integrateId, () => config.userId, demoDomain, sdkToken],
+  clearDemo,
+  { flush: 'sync' },
+)
+
 async function onDemoAction(action: string, index: number): Promise<void> {
   const msg = demoMessages.value[index]
   if (!msg) return
@@ -922,41 +928,30 @@ const testResults = ref<TestCaseResult[]>([
       const url = buildDemoUrl('chat/stream', '你好', 'verify_sse', { enableRag: false })
       apiCount.value++
       const t0 = performance.now()
-      let res: Response
-      try {
-        res = await fetch(url, { headers: sdkAuthHeaders(), signal: AbortSignal.timeout(20000) })
-      } catch {
-        log('⚠ 请求失败', 'warn')
-        mark('skip', '后端不可用')
-        return
-      }
-      if (!res.ok) {
-        log('⚠ HTTP ' + res.status, 'warn')
-        mark('skip', '接口异常')
-        return
-      }
-      const reader = res.body!.getReader()
-      const decoder = new TextDecoder()
       let total = ''
       let chunks = 0
-      while (true) {
-        const { done, value } = await reader.read()
-        if (done) break
-        total += decoder.decode(value, { stream: true })
-        chunks++
-      }
+      let metadataChunks = 0
+      let doneCalls = 0
+      await readSSEStreamWithEvents(url, {
+        onMessage: (text) => { total += text; chunks++ },
+        onSources: (sources) => {
+          metadataChunks++
+          assert(sources.length === 0, '普通对话应返回空 sources')
+        },
+        onDone: () => { doneCalls++ },
+      }, sdkAuthHeaders(), AbortSignal.timeout(20000))
       const elapsed = Math.round(performance.now() - t0)
       apiDurations.value.push(elapsed)
-      assert(total.includes('choices') || total.includes('[DONE]'), '非 OpenAI Chat Completions 格式')
-      log('✓ OpenAI 格式（含 choices 或 [DONE]）', 'pass')
-      log('✓ SSE ' + chunks + ' chunks, ' + total.length + ' chars', 'pass')
+      assert(metadataChunks === 1 && doneCalls === 1, '应收到一次引用元数据和一次完成事件')
+      log('OpenAI 正文和来源元数据分别解析，完成事件只触发一次', 'pass')
+      log('SSE ' + chunks + ' chunks, ' + total.length + ' chars', 'pass')
       mark('pass', '通过 (' + elapsed + 'ms)')
     },
   },
   {
-    id: 'T8', name: 'API RAG 引用来源', desc: 'GET /ai/chat/sources', phase: 'p1', status: 'idle', summary: '', logs: [],
+    id: 'T8', name: 'API 回答引用来源', desc: 'GET /ai/chat/result：同次回答携带 sources', phase: 'p1', status: 'idle', summary: '', logs: [],
     fn: async (log, mark) => {
-      const url = buildDemoUrl('chat/sources', '请假', 'verify_sources', { enableRag: true, rewriteStrategy: 'REWRITE' })
+      const url = buildDemoUrl('chat/result', '请假', 'verify_sources', { enableRag: true, rewriteStrategy: 'NONE' })
       apiCount.value++
       let res: Response
       try {
@@ -972,9 +967,9 @@ const testResults = ref<TestCaseResult[]>([
         return
       }
       const json = await res.json()
-      log('返回 success=' + json.success + ' data=' + (json.data ? json.data.length : 0) + ' 条', 'info')
-      assert(json.success !== undefined, '应返回 success 字段')
-      log('✓ RAG 引用来源接口可用', 'pass')
+      log('返回回答 ' + (json.text?.length || 0) + ' 字，引用 ' + (json.sources?.length || 0) + ' 条', 'info')
+      assert(typeof json.text === 'string' && Array.isArray(json.sources), '应直接返回 text 和 sources 字段')
+      log('回答与引用来源由同一请求返回', 'pass')
       mark('pass', '通过')
     },
   },
diff --git a/frontend/src/types/sse.ts b/frontend/src/types/sse.ts
index a7f3f7d..13f77df 100644
--- a/frontend/src/types/sse.ts
+++ b/frontend/src/types/sse.ts
@@ -1,6 +1,24 @@
+/** 同次回答实际命中的知识库分块；ID 保持字符串精度。 */
+export interface SourceReference {
+  documentId: string | null
+  title: string | null
+  sourceName: string | null
+  chunkIndex: number | null
+  score: number | null
+  snippet: string | null
+}
+
+export interface ChatResult {
+  text: string
+  mcpEvents: unknown[]
+  suggestions: string[]
+  sources: SourceReference[]
+}
+
 /** SSE 事件类型定义 */
 export interface SSECallbacks {
   onMessage?: (chunk: string) => void
+  onSources?: (sources: SourceReference[]) => void
   onToolCallStart?: (data: any) => void
   onToolCallResult?: (data: any) => void
   onError?: (data: any) => void
diff --git a/frontend/src/utils/chatAdapter.ts b/frontend/src/utils/chatAdapter.ts
index 424a40a..2ca0b7e 100644
--- a/frontend/src/utils/chatAdapter.ts
+++ b/frontend/src/utils/chatAdapter.ts
@@ -7,6 +7,7 @@
  */
 
 import type { TdChatItemMeta, AIMessageContent, UserMessageContent } from '@tdesign-vue-next/chat'
+import type { SourceReference } from '@/types/sse'
 
 /** 附件信息（图片或文件） */
 export interface Attachment {
@@ -24,7 +25,7 @@ export interface ChatMessage {
   content: string
   streaming: boolean
   time: string
-  sources?: any[]
+  sources?: SourceReference[]
   toolCalls?: any[]
   error?: boolean
   feedback?: string | null
diff --git a/frontend/src/utils/sse.ts b/frontend/src/utils/sse.ts
index 47946f4..46b256a 100644
--- a/frontend/src/utils/sse.ts
+++ b/frontend/src/utils/sse.ts
@@ -1,206 +1,139 @@
-/**
- * SSE 流式读取工具 —— 从 utils.js 原封不动搬移
- *
- * 统一处理 Flux<String> / ServerSentEvent / SseEmitter 三种 SSE 接口。
- * 此文件使用原生 fetch + ReadableStream API，零框架依赖。
- */
-import type { SSECallbacks } from '@/types/sse'
+/** Shared SSE reader for plain text, OpenAI chunks and tool events. */
+import type { SSECallbacks, SourceReference } from '@/types/sse'
 import { getToken } from '@/utils/token'
 
-/** 构建带认证的请求头 */
+/** Preserve explicit SDK authorization instead of replacing it with the admin token. */
 function authHeaders(extra?: Record<string, string>): Record<string, string> {
   const token = getToken()
-  const base: Record<string, string> = extra ? { ...extra } : {}
-  // 仅当调用方未显式提供 Authorization 时才注入管理后台 token，避免覆盖调用方传入的鉴权头（如测试面板的 SDK Token）
-  if (token && !base['Authorization']) base['Authorization'] = `Bearer ${token}`
-  return base
+  const headers = { ...extra }
+  if (token && !headers['Authorization']) headers['Authorization'] = `Bearer ${token}`
+  return headers
 }
 
-/** 尝试从 OpenAI Chat Completions chunk 中提取 delta.content；返回 undefined 表示非 OpenAI 格式（回退纯文本） */
-function extractOpenAIDelta(text: string): string | undefined {
-  let obj: any
-  try { obj = JSON.parse(text) } catch { return undefined }
-  // OpenAI 错误 chunk（{"error":{"message":...}}）：提取错误信息作为内容，避免原始 JSON 泄漏到对话框
-  if (obj && obj.error) {
-    const msg = obj.error.message || obj.error.type
-    return typeof msg === 'string' && msg ? msg : '服务异常'
+function isSourceReference(value: unknown): value is SourceReference {
+  if (!value || typeof value !== 'object') return false
+  const nullableString = (v: unknown) => v === null || typeof v === 'string'
+  const nullableNumber = (v: unknown) => v === null || typeof v === 'number'
+  return 'documentId' in value && nullableString(value.documentId)
+    && 'title' in value && nullableString(value.title)
+    && 'sourceName' in value && nullableString(value.sourceName)
+    && 'chunkIndex' in value && nullableNumber(value.chunkIndex)
+    && 'score' in value && nullableNumber(value.score)
+    && 'snippet' in value && nullableString(value.snippet)
+}
+
+/** Returns false only for legacy plain-text payloads. Metadata never enters message text. */
+function dispatchOpenAI(text: string, handlers: SSECallbacks): boolean {
+  let value: unknown
+  try { value = JSON.parse(text) } catch { return false }
+  if (!value || typeof value !== 'object') return false
+  if ('error' in value && value.error && typeof value.error === 'object') {
+    const error = value.error
+    const message = 'message' in error ? error.message : 'type' in error ? error.type : undefined
+    handlers.onMessage?.(typeof message === 'string' && message ? message : '服务异常')
+    return true
   }
-  if (obj && Array.isArray(obj.choices)) {
-    const delta = obj.choices[0]?.delta
-    if (delta && typeof delta.content === 'string') return delta.content
-    return '' // OpenAI 形状但无 content（role/finish_reason 空 chunk）→ 跳过
+  if (!('choices' in value) || !Array.isArray(value.choices)) return false
+  if ('sources' in value && Array.isArray(value.sources)) {
+    if (!value.sources.every(isSourceReference)) throw new Error('无效的引用来源数据')
+    handlers.onSources?.(value.sources)
   }
-  return undefined // 是 JSON 但非 OpenAI 形状 → 回退纯文本
+  const content = value.choices[0]?.delta?.content
+  if (typeof content === 'string' && content) handlers.onMessage?.(content)
+  return true
 }
 
-/**
- * 通用 SSE 流式读取 —— 统一处理 Flux<String> / ServerSentEvent / SseEmitter 三种 SSE 接口
- *
- * @param url 请求地址
- * @param onChunk 每收到一段文本的回调
- * @param onDone 流结束的回调
- * @param headers 额外请求头
- * @param signal AbortSignal 用于取消请求（组件卸载时必须传入以释放网络资源）
- */
-export async function readSSEStream(
+/** Text-only callers share the same framing, completion and cleanup semantics. */
+export function readSSEStream(
   url: string,
   onChunk: (text: string) => void,
   onDone?: () => void,
   headers?: Record<string, string>,
-  signal?: AbortSignal
+  signal?: AbortSignal,
 ): Promise<void> {
-  const res = await fetch(url, { headers: authHeaders(headers), signal, credentials: 'include' })
-  if (!res.ok) throw new Error('HTTP ' + res.status)
-  const reader = res.body!.getReader()
-  const decoder = new TextDecoder()
-  let buffer = ''
-  // SSE 规范：同一事件内的多行 data: 字段用 \n 拼接，事件间用空行分隔
-  let eventDataLines: string[] = []
-  let currentEvent = 'message'
-
-  const flushEvent = () => {
-    if (eventDataLines.length === 0) return
-    const text = eventDataLines.join('\n')
-    eventDataLines = []
-    const ev = currentEvent
-    currentEvent = 'message'
-    if (text === '[DONE]') return
-    // 跳过 status / faq 等系统事件，不显示在对话框中
-    if (ev === 'status') return
-    // OpenAI Chat Completions 格式：提取 delta.content；无 content 的空 chunk 跳过
-    const extracted = extractOpenAIDelta(text)
-    if (extracted !== undefined) {
-      if (extracted) onChunk(extracted)
-      return
-    }
-    // 空事件视为 LLM 流式输出的换行符（Spring 将 "\n" 编码为单条空 data: 事件）
-    onChunk(text || '\n')
-  }
-
-  while (true) {
-    const { done, value } = await reader.read()
-    if (done) break
-    buffer += decoder.decode(value, { stream: true })
-    const lines = buffer.split('\n')
-    buffer = lines.pop() || ''
-    for (let line of lines) {
-      // 兼容 \r\n 行结束符
-      if (line.endsWith('\r')) line = line.slice(0, -1)
-      if (line === '') {
-        // 空行 = SSE 事件边界
-        flushEvent()
-      } else if (line.startsWith('event:')) {
-        // 记录事件类型，用于 flushEvent 时过滤系统事件
-        currentEvent = line.slice(6).trim()
-      } else if (line.startsWith('data:')) {
-        // 累积 data 字段；仅剥离 data: 后的一个可选空格，保留 markdown 列表缩进
-        let data = line.slice(5)
-        if (data.startsWith(' ')) data = data.slice(1)
-        eventDataLines.push(data)
-      } else if (!line.startsWith(':')) {
-        // Flux<String> 模式（非标准 SSE），先把已累积的 SSE 事件 flush 再处理
-        flushEvent()
-        if (line.trim()) onChunk(line)
-      }
-    }
-  }
-  // 流结束，flush 末尾未以空行收尾的事件
-  flushEvent()
-  if (onDone) onDone()
+  return readSSEStreamWithEvents(url, { onMessage: onChunk, onDone }, headers, signal)
 }
 
-/**
- * 增强版 SSE 流式读取（支持事件类型分发）
- * 解析 SSE 标准的 event: 字段，将不同类型事件分发到对应回调。
- *
- * @param url 请求地址
- * @param handlers 回调对象：
- *   - onMessage(chunk): 普通文本内容（event: message 或无 event 的 data）
- *   - onToolCallStart(data): 工具调用开始（event: tool_call_start）
- *   - onToolCallResult(data): 工具调用结果（event: tool_call_result）
- *   - onError(data): 错误事件（event: error）
- *   - onDone(): 流结束
- * @param headers 额外请求头
- * @param signal AbortSignal 用于取消请求（组件卸载时必须传入以释放网络资源）
- */
+/** Read complete SSE events; [DONE] terminates immediately without waiting for network EOF. */
 export async function readSSEStreamWithEvents(
   url: string,
   handlers: SSECallbacks,
   headers?: Record<string, string>,
-  signal?: AbortSignal
+  signal?: AbortSignal,
 ): Promise<void> {
-  const { onMessage, onToolCallStart, onToolCallResult, onError, onDone } = handlers
   const res = await fetch(url, { headers: authHeaders(headers), signal, credentials: 'include' })
-  if (!res.ok) throw new Error('HTTP ' + res.status)
-  const reader = res.body!.getReader()
+  if (!res.ok) throw new Error((await res.text()) || 'HTTP ' + res.status)
+  if (!res.body) throw new Error('响应流为空')
+  const reader = res.body.getReader()
   const decoder = new TextDecoder()
   let buffer = ''
-  let currentEvent = 'message'
-  // SSE 规范：同一事件内的多行 data: 字段用 \n 拼接
-  let eventDataLines: string[] = []
+  let event = 'message'
+  let data: string[] = []
+  let completed = false
+  let eof = false
 
   const flushEvent = () => {
-    if (eventDataLines.length === 0) return
-    const ev = currentEvent
-    currentEvent = 'message'
-    const raw = eventDataLines.join('\n')
-    eventDataLines = []
-    if (raw === '[DONE]') return
-    // 空事件视为换行符（仅 message 事件）；JSON 事件（工具调用等）空内容不应出现
-    const text = raw || '\n'
-
-    switch (ev) {
-      case 'tool_call_start':
-        if (onToolCallStart) onToolCallStart(JSON.parse(text))
-        break
-      case 'tool_call_result':
-        if (onToolCallResult) onToolCallResult(JSON.parse(text))
-        break
-      case 'error':
-        if (onError) onError(JSON.parse(text))
-        break
-      case 'status':
-        // 系统状态事件（generating / faq_hit 等），不显示在对话框中
-        break
-      case 'message':
-      default: {
-        // OpenAI Chat Completions 格式：提取 delta.content；无 content 的空 chunk 跳过
-        const extracted = extractOpenAIDelta(text)
-        if (extracted !== undefined) {
-          if (extracted && onMessage) onMessage(extracted)
-        } else if (onMessage) {
-          onMessage(text)
-        }
-        break
-      }
+    const currentEvent = event
+    event = 'message'
+    if (!data.length) return
+    const raw = data.join('\n')
+    data = []
+    if (raw === '[DONE]') {
+      completed = true
+      return
+    }
+    switch (currentEvent) {
+      case 'status': return
+      case 'tool_call_start': handlers.onToolCallStart?.(JSON.parse(raw)); return
+      case 'tool_call_result': handlers.onToolCallResult?.(JSON.parse(raw)); return
+      case 'error': handlers.onError?.(JSON.parse(raw)); return
+      default:
+        if (!dispatchOpenAI(raw, handlers)) handlers.onMessage?.(raw || '\n')
     }
   }
 
-  while (true) {
-    const { done, value } = await reader.read()
-    if (done) break
-    buffer += decoder.decode(value, { stream: true })
-    const lines = buffer.split('\n')
-    buffer = lines.pop() || ''
+  const consumeLine = (raw: string) => {
+    const line = raw.endsWith('\r') ? raw.slice(0, -1) : raw
+    if (!line) {
+      flushEvent()
+    } else if (line.startsWith('event:')) {
+      event = line.slice(6).trim()
+    } else if (line === 'data' || line.startsWith('data:')) {
+      const value = line === 'data' ? '' : line.slice(5)
+      data.push(value.startsWith(' ') ? value.slice(1) : value)
+    } else if (!line.startsWith(':') && !line.startsWith('id:') && !line.startsWith('retry:')) {
+      // Legacy Flux<String> bodies may contain unframed lines.
+      flushEvent()
+      if (!completed && line.trim()) handlers.onMessage?.(line)
+    }
+  }
 
-    for (let line of lines) {
-      if (line.endsWith('\r')) line = line.slice(0, -1)
-      if (line === '') {
-        // 空行 = SSE 事件边界
-        flushEvent()
-      } else if (line.startsWith('event:')) {
-        currentEvent = line.slice(6).trim()
-      } else if (line.startsWith('data:')) {
-        let data = line.slice(5)
-        if (data.startsWith(' ')) data = data.slice(1)
-        eventDataLines.push(data)
-      } else if (!line.startsWith(':')) {
-        // Flux<String> 模式（非标准 SSE）
-        flushEvent()
-        if (line.trim() && onMessage) onMessage(line)
+  try {
+    while (!completed) {
+      signal?.throwIfAborted()
+      const result = await reader.read()
+      signal?.throwIfAborted()
+      eof = result.done
+      buffer += eof ? decoder.decode() : decoder.decode(result.value, { stream: true })
+      let start = 0
+      let end: number
+      while (!completed && (end = buffer.indexOf('\n', start)) !== -1) {
+        consumeLine(buffer.slice(start, end))
+        start = end + 1
       }
+      buffer = buffer.slice(start)
+      if (eof) {
+        if (!completed && buffer) consumeLine(buffer)
+        if (!completed) flushEvent()
+        break
+      }
+    }
+    handlers.onDone?.()
+  } finally {
+    // Cancellation may itself reject or stall; it must not delay DONE or mask the original error.
+    if (!eof) {
+      try { void reader.cancel().catch(() => {}) } catch { /* Preserve the stream outcome. */ }
     }
+    try { reader.releaseLock() } catch { /* Preserve the stream outcome. */ }
   }
-  flushEvent()
-  if (onDone) onDone()
 }
diff --git a/frontend/src/views/ChatPanel.vue b/frontend/src/views/ChatPanel.vue
index c61c80c..eba12ec 100644
--- a/frontend/src/views/ChatPanel.vue
+++ b/frontend/src/views/ChatPanel.vue
@@ -235,14 +235,14 @@ import {
   ChatActionbar,
 } from '@tdesign-vue-next/chat'
 import '@tdesign-vue-next/chat/es/style/index.css'
-import { chatSync, chatSSEUrl, ragSources, fetchSuggestions, type ChatOptions } from '@/api/chat'
+import { chatSync, chatSSEUrl, fetchSuggestions, type ChatOptions } from '@/api/chat'
 import { getRoleList } from '@/api/role'
 import { getActiveModelConfig } from '@/api/model-config'
 import { truncateConversation } from '@/api/conversation'
 import { submitFeedback as submitFeedbackApi } from '@/api/feedback'
 import { uploadAttachment } from '@/api/upload'
 import { toast } from '@/utils/toast'
-import { readSSEStream, readSSEStreamWithEvents } from '@/utils/sse'
+import { readSSEStreamWithEvents } from '@/utils/sse'
 import { renderMarkdown } from '@/utils/markdown'
 import { useCategoryStore } from '@/stores/category'
 import { toChatData, isLastAssistant } from '@/utils/chatAdapter'
@@ -266,11 +266,11 @@ const QUICK_QUESTIONS = [
 ]
 
 const ragStrategyOptions = [
-  { label: '不重写', value: 'NONE' },
-  { label: '查询重写', value: 'REWRITE' },
-  { label: '翻译扩展', value: 'TRANSLATION' },
-  { label: '查询压缩', value: 'COMPRESSION' },
-  { label: '多路扩展', value: 'MULTI_QUERY' },
+  { label: '不重写（最快）', value: 'NONE' },
+  { label: '查询重写（额外耗时）', value: 'REWRITE' },
+  { label: '翻译扩展（额外耗时）', value: 'TRANSLATION' },
+  { label: '查询压缩（额外耗时）', value: 'COMPRESSION' },
+  { label: '多路扩展（额外耗时）', value: 'MULTI_QUERY' },
 ]
 
 const modeOptions = [
@@ -284,7 +284,7 @@ const mode = ref('sse')  // 默认 SSE 流式
 const selectedRole = ref('general')
 const roles = ref([FALLBACK_ROLE])
 const isRagMode = ref(false)
-const ragStrategy = ref('MULTI_QUERY')
+const ragStrategy = ref('NONE')
 const activeModel = ref<any>(null)
 const modelLoadError = ref('')
 const userInput = ref('')
@@ -376,6 +376,7 @@ function providerLabel(provider: string): string {
 
 // ==================== 角色管理 ====================
 function selectRole(roleKey: string): void {
+  abortChat()
   selectedRole.value = roleKey
   newChatId()
   currentSuggestions.value = []  // 切换角色时清空推荐问题
@@ -494,7 +495,7 @@ async function send(): Promise<void> {
   const cid = chatId.value || ('web_' + Date.now())
   chatId.value = cid
   const requestMode = mode.value
-  // 固定本轮参数，后台引用请求不读取后续切换的角色或 RAG 配置。
+  // 固定本轮参数，后续角色或配置切换不影响已发出的回答。
   const imageUrls = attachments.filter(a => a.type === 'image').map(a => a.url)
   const chatOptions: ChatOptions = {
     roleId: currentRoleId(),
@@ -516,64 +517,56 @@ async function send(): Promise<void> {
     streaming: true, time: formatTime(), sources: [], toolCalls: [],
   }
   messages.value.push(assistantMsg)
+  const controller = new AbortController()
+  sseAbortController?.abort()
+  sseAbortController = controller
+  const isCurrent = () => sseAbortController === controller
+    && !controller.signal.aborted && chatId.value === cid
+    && messages.value.some(msg => msg.id === assistantMsg.id)
   await scrollToBottom()
 
   try {
-    // 取消上一个 SSE 请求并创建新的 AbortController
-    if (sseAbortController) { sseAbortController.abort() }
-    sseAbortController = new AbortController()
-    const signal = sseAbortController.signal
-
+    if (!isCurrent()) return
     if (requestMode === 'sync') {
-      // 同步调用
-      assistantMsg.content = await chatSync(text, cid, chatOptions)
+      const result = await chatSync(text, cid, chatOptions, controller.signal)
+      if (!isCurrent()) return
+      assistantMsg.content = result.text
+      assistantMsg.sources = result.sources
     } else {
-      // SSE 流式
-      const url = chatSSEUrl(text, cid, chatOptions)
-      if (chatOptions.enableRag) {
-        await readSSEStreamWithEvents(url, {
-          onMessage: async (chunk: string) => {
-            assistantMsg.content += chunk
-            messages.value = [...messages.value]
-            await scrollToBottom()
-          },
-          onToolCallStart: (data: any) => {
-            assistantMsg.toolCalls!.push({ tool: data.tool, input: data.input, status: 'running', result: null })
-            scrollToBottom()
-          },
-          onToolCallResult: (data: any) => {
-            const tc = assistantMsg.toolCalls!.find(t => t.tool === data.tool && t.status === 'running')
-            if (tc) { tc.status = 'done'; tc.result = data.result; tc.latencyMs = data.latencyMs }
-            messages.value = [...messages.value]
-            scrollToBottom()
-          },
-          onError: (data: any) => {
-            assistantMsg.content += '\n\n⚠️ ' + (data.message || '工具调用出错')
-            messages.value = [...messages.value]
-          },
-          onDone: () => {},
-        }, undefined, signal)
-      } else {
-        await readSSEStream(url, async (chunk: string) => {
+      await readSSEStreamWithEvents(chatSSEUrl(text, cid, chatOptions), {
+        onMessage: (chunk) => {
+          if (!isCurrent()) return
           assistantMsg.content += chunk
           messages.value = [...messages.value]
-          await scrollToBottom()
-        }, () => {}, undefined, signal)
-      }
-    }
-
-    // 引用来源后台补齐，正文完成后立即结束发送状态。
-    if (chatOptions.enableRag) {
-      void ragSources(text, cid, chatOptions).then(sj => {
-        if (!sj?.success) throw new Error(sj?.message || '引用来源获取失败')
-        if (chatId.value !== cid || !messages.value.some(msg => msg.id === assistantMsg.id)) return
-        assistantMsg.sources = sj.data || []
-        messages.value = [...messages.value]
-      }).catch((e: any) => {
-        toast('引用来源获取失败: ' + (e.message || e), 'warning')
-      })
+          void scrollToBottom()
+        },
+        onSources: (sources) => {
+          if (!isCurrent()) return
+          assistantMsg.sources = sources
+          messages.value = [...messages.value]
+        },
+        onToolCallStart: (data) => {
+          if (!isCurrent()) return
+          assistantMsg.toolCalls!.push({ tool: data.tool, input: data.input, status: 'running', result: null })
+          messages.value = [...messages.value]
+          void scrollToBottom()
+        },
+        onToolCallResult: (data) => {
+          if (!isCurrent()) return
+          const tc = assistantMsg.toolCalls!.find(t => t.tool === data.tool && t.status === 'running')
+          if (tc) { tc.status = 'done'; tc.result = data.result; tc.latencyMs = data.latencyMs }
+          messages.value = [...messages.value]
+          void scrollToBottom()
+        },
+        onError: (data) => {
+          if (!isCurrent()) return
+          assistantMsg.content += '\n\n' + (data.message || '工具调用出错')
+          messages.value = [...messages.value]
+        },
+      }, undefined, controller.signal)
     }
   } catch (e: any) {
+    if (!isCurrent()) return
     // AbortError 不是真正的错误，不显示错误信息
     if (e.name === 'AbortError') {
       assistantMsg.content = assistantMsg.content || '已取消'
@@ -583,18 +576,18 @@ async function send(): Promise<void> {
       toast('对话失败：' + e.message, 'error')
     }
   } finally {
-    assistantMsg.streaming = false
-    isSending.value = false
-    messages.value = [...messages.value]
-
-    // 拉取推荐问题（suggest-message-list）；AbortError / 异常时不拉取
-    if (!sseAbortController?.signal.aborted && assistantMsg.content && !assistantMsg.error) {
-      fetchSuggestions(chatId.value).then(items => {
-        if (items.length) currentSuggestions.value = items
-      }).catch(() => {})
+    if (isCurrent()) {
+      assistantMsg.streaming = false
+      isSending.value = false
+      messages.value = [...messages.value]
+      // 推荐问题仍独立获取，但旧会话/旧回答的结果不得覆盖新对话。
+      if (assistantMsg.content && !assistantMsg.error) {
+        void fetchSuggestions(cid).then(items => {
+          if (isCurrent() && items.length) currentSuggestions.value = items
+        })
+      }
+      await scrollToBottom()
     }
-
-    await scrollToBottom()
   }
 }
 
diff --git a/frontend/src/views/PipelineFlow.vue b/frontend/src/views/PipelineFlow.vue
index 2469e6c..3bf0ecd 100644
--- a/frontend/src/views/PipelineFlow.vue
+++ b/frontend/src/views/PipelineFlow.vue
@@ -5,7 +5,7 @@
         <div>
           <span style="font-size:16px;font-weight:600;">🔀 AI 执行链</span>
           <p style="font-size:12px;color:var(--td-text-color-placeholder);margin:4px 0 0;">
-            下图展示从用户请求到 AI 回复的完整处理流程，包含意图路由、RAG 检索、熔断保护和 Advisor 链。
+            下图展示从用户请求到 AI 回复的完整处理流程，包含 FAQ 与本地路由、RAG 检索、熔断保护和 Advisor 链。
             菱形节点 = 决策分支 · 虚线框 = 独立子系统 · 虚线箭头 = 降级/异步路径。
           </p>
         </div>
@@ -76,14 +76,15 @@ mermaid.initialize({
 
 // Mermaid 流程图 DSL 定义
 // 节点类型: [矩形]=处理步骤, {菱形}=决策分支, subgraph=子系统
-// %%graph-meta: { updated: "2026-09-14", basedOn: "ChatPipeline v4, RagPipeline v2, AssistantApp v2", mermaidVersion: "flowchart-v2" }
+// %%graph-meta: { updated: "2026-09-14", basedOn: "ChatPipeline v5 direct retrieval, RagPipeline v3, AssistantApp sources metadata", mermaidVersion: "flowchart-v2" }
 const GRAPH_DEFINITION = `
 flowchart TD
   A["<b>用户请求</b><br/>message + roleId + accountId + chatId"]
 
   A --> B{"<b>Controller</b><br/>鉴权 / 角色解析 / KB 隔离判断<br/>构建 ChatContext"}
 
-  B --> C["<b>ChatPipeline.buildRequest</b><br/>编排决策入口"]
+  B --> CB{"<b>AssistantApp 熔断检查</b><br/>SimpleCircuitBreaker<br/>阈值: 连续 3 次失败 / 恢复: 5 分钟"}
+  CB -- "正常" --> C["<b>ChatPipeline.buildRequest</b><br/>编排决策入口"]
 
   C --> D{"enableRag ?"}
 
@@ -91,36 +92,36 @@ flowchart TD
 
   D -- "true" --> H["<b>FAQ 优先匹配</b><br/>FaqMatchEngine 完整三级匹配<br/>精确 → 关键词 → 向量语义<br/>沿用角色分类隔离"]
 
-  H -- "命中标准答案，跳过意图分类" --> T
-  H -- "未命中 / 异常" --> F["<b>IntentRouter</b><br/>寒暄词快速路径: 本地精确匹配（零 LLM）<br/>未命中则 LLM 意图分类<br/>FAQ / RAG / CHITCHAT"]
+  H -- "命中标准答案" --> T
+  H -- "未命中 / 异常" --> G{"<b>本地 isChitchat</b><br/>精确寒暄词匹配<br/>无意图分类 LLM"}
 
-  F --> G{"意图分类结果"}
+  G -- "寒暄命中" --> I["<b>模式: 纯对话</b><br/>跳过知识库检索<br/>不注入资料块"]
 
-  G -- "CHITCHAT<br/>confidence ≧ 0.6" --> I["<b>模式: 纯对话</b><br/>跳过知识库检索<br/>不注入资料块"]
-
-  G -- "FAQ / RAG / 降级<br/>其余情况" --> J["<b>RagPipeline.retrieve</b><br/>RAG 检索流水线入口"]
+  G -- "其余问题" --> J["<b>RagPipeline.retrieve</b><br/>默认零 LLM 预处理"]
 
   subgraph RAG["📚 RAG 检索流水线（当前: 纯向量检索）"]
     J --> K{"前置 FAQ 匹配<br/>completedCleanly ?"}
-    K -- "true: 跳过重复 FAQ" --> L["<b>2. 查询重写</b><br/>REWRITE / TRANSLATION<br/>COMPRESSION / MULTI_QUERY"]
+    K -- "true: 跳过重复 FAQ" --> L{"<b>2. rewriteStrategy</b><br/>默认 NONE"}
     K -- "false: 异常后重试" --> KR["<b>1. FAQ 匹配重试</b><br/>FaqMatchEngine 完整三级匹配"]
     KR -- "命中标准答案" --> T
     KR -- "未命中 / 异常" --> L
-    L --> M["<b>3. 向量检索</b><br/>PGVector similaritySearch<br/>topK=4 + 分类过滤"]
-    M --> S["<b>4. 构建资料块</b><br/>拼接检索文档<br/>注入 system prompt 末尾"]
+    L -- "NONE 原文直检索" --> M["<b>3. 向量检索</b><br/>PGVector similaritySearch<br/>topK=4 + 分类过滤"]
+    L -- "显式选择" --> LR["<b>查询重写 LLM</b><br/>REWRITE / TRANSLATION / MULTI_QUERY<br/>COMPRESSION: 最近 10 条历史补全指代"]
+    LR --> M
+    M --> S["<b>4. 构建资料块</b><br/>命中/未命中日志仅此层记录一次<br/>注入 system prompt + 保留命中文档"]
   end
 
   I --> T
   S --> T
   E --> T
 
-  T["<b>组装 ChatRequest</b><br/>finalMessage + finalSystemPrompt<br/>+ faqAnswer (可选)"]
+  T["<b>组装 ChatRequest</b><br/>finalMessage + finalSystemPrompt<br/>+ faqAnswer / 当次命中文档"]
 
-  T --> CB{"<b>🔌 AI 熔断检查</b><br/>SimpleCircuitBreaker<br/>阈值: 连续 3 次失败 / 恢复: 5 分钟"}
+  T -- "FAQ 标准答案" --> FAQ["<b>FAQ 直接回复</b><br/>内容安全检查 + 写入会话记忆<br/>不调用答案 LLM"]
 
   CB -- "熔断中" --> FALLBACK["<b>返回降级提示</b><br/>「AI 服务暂时不可用<br/>请稍后重试」"]
 
-  CB -- "正常" --> U["<b>AssistantApp</b><br/>chat / chatStream<br/>构建 ChatClient + MCP 工具"]
+  T -- "需要生成" --> U["<b>AssistantApp</b><br/>chat / chatStream<br/>构建 ChatClient + MCP 工具"]
 
   subgraph ADVISOR["🛡️ Advisor 链（环绕 LLM 调用）"]
     U --> V["<b>ContentSafetyAdvisor</b><br/>🔽 before: DFA 敏感词检测<br/>用户输入 BLOCK/MASK"]
@@ -131,8 +132,9 @@ flowchart TD
     Z --> AA["<b>ContentSafetyAdvisor</b><br/>🔼 after: AI 输出检测<br/>BLOCK/MASK 违规内容"]
   end
 
-  AA --> AB["<b>返回 AI 回复</b><br/>SSE 流式输出<br/>+ MCP 工具调用事件"]
+  AA --> AB["<b>返回 AI 回复 + 当次引用</b><br/>同步 JSON 或 SSE 正文 → sources metadata → stop / DONE<br/>引用不进入正文，不二次检索"]
 
+  FAQ --> AB
   FALLBACK --> AB
 
   AB -. "异步按需触发" .-> SG
diff --git a/frontend/tests/chat-protocol.test.mjs b/frontend/tests/chat-protocol.test.mjs
new file mode 100644
index 0000000..45610e1
--- /dev/null
+++ b/frontend/tests/chat-protocol.test.mjs
@@ -0,0 +1,164 @@
+import assert from 'node:assert/strict'
+import { afterEach, beforeEach, test } from 'node:test'
+import { fileURLToPath } from 'node:url'
+import { build } from 'esbuild'
+
+// Exercise the production TypeScript with Vite's existing esbuild dependency, no test framework.
+const root = fileURLToPath(new URL('../', import.meta.url))
+async function loadModule(entry) {
+  const result = await build({
+    entryPoints: [root + entry], bundle: true, write: false, format: 'esm', platform: 'node',
+    alias: { '@': root + 'src' },
+  })
+  return import('data:text/javascript;base64,' + Buffer.from(result.outputFiles[0].text).toString('base64'))
+}
+const { readSSEStream, readSSEStreamWithEvents } = await loadModule('src/utils/sse.ts')
+const { chatSync, fetchChatResult } = await loadModule('src/api/chat.ts')
+const originalFetch = globalThis.fetch
+const originalStorage = Object.getOwnPropertyDescriptor(globalThis, 'localStorage')
+
+beforeEach(() => {
+  Object.defineProperty(globalThis, 'localStorage', {
+    configurable: true, value: { getItem: () => 'admin-token' },
+  })
+})
+afterEach(() => {
+  globalThis.fetch = originalFetch
+  if (originalStorage) Object.defineProperty(globalThis, 'localStorage', originalStorage)
+  else delete globalThis.localStorage
+})
+
+const sources = [{
+  documentId: '9223372036854775806', title: '报销制度', sourceName: 'policy.pdf',
+  chunkIndex: 2, score: 0.125, snippet: '申请应在三十天内提交。',
+}]
+const metadata = JSON.stringify({ object: 'chat.completion.chunk', choices: [], sources })
+const answer = JSON.stringify({ choices: [{ delta: { content: '答复。' } }] })
+
+function responseFor(text, { open = false, fragment = false, cancel } = {}) {
+  const bytes = new TextEncoder().encode(text)
+  const body = new ReadableStream({
+    start(controller) {
+      if (fragment) for (const byte of bytes) controller.enqueue(Uint8Array.of(byte))
+      else controller.enqueue(bytes)
+      if (!open) controller.close()
+    },
+    cancel,
+  })
+  return new Response(body)
+}
+
+test('fragmented CRLF metadata preserves ID/snippet and DONE completes without EOF or cancellation settlement', { timeout: 1000 }, async () => {
+  let cancelled = 0
+  const response = responseFor(
+    `data: ${answer}\r\n\r\ndata: ${metadata}\r\n\r\ndata: [DONE]\r\n\r\ndata: ignored\r\n\r\n`,
+    { open: true, fragment: true, cancel() { cancelled++; return new Promise(() => {}) } },
+  )
+  const requests = []
+  globalThis.fetch = async (...args) => { requests.push(args); return response }
+  const events = []
+  await readSSEStreamWithEvents('/ai/chat/stream', {
+    onMessage: text => events.push(['text', text]),
+    onSources: value => events.push(['sources', value]),
+    onDone: () => events.push(['done']),
+  }, { Authorization: 'Bearer sdk-token' })
+  assert.deepEqual(events, [['text', '答复。'], ['sources', sources], ['done']])
+  assert.equal(requests.length, 1)
+  assert.equal(requests[0][1].headers.Authorization, 'Bearer sdk-token')
+  assert.equal(cancelled, 1)
+  assert.equal(response.body.locked, false)
+})
+
+test('text-only facade uses identical DONE handling and never leaks empty choices or sources', { timeout: 1000 }, async () => {
+  globalThis.fetch = async () => responseFor(
+    `data: ${metadata}\n\ndata: ${answer}\n\ndata: [DONE]\n\ndata: [DONE]\n\ndata: trailing\n\n`,
+    { open: true },
+  )
+  const text = []
+  let completed = 0
+  await readSSEStream('/ai/chat/stream', value => text.push(value), () => completed++)
+  assert.deepEqual(text, ['答复。'])
+  assert.equal(completed, 1)
+})
+
+test('EOF residuals, multiline data, tool events, status and legacy text retain their semantics', async () => {
+  globalThis.fetch = async () => responseFor(
+    ': heartbeat\r\nid: event-1\r\nretry: 1000\r\nevent: status\r\ndata: generating\r\n\r\n'
+      + 'event: tool_call_start\r\ndata: {"tool":"lookup"}\r\n\r\n'
+      + 'event: tool_call_result\r\ndata: {"tool":"lookup","result":"ok"}\r\n\r\n'
+      + 'event: error\r\ndata: {"message":"tool unavailable"}\r\n\r\n'
+      + 'data: first\r\ndata:   indented\r\n\r\ndata:\r\n\r\nlegacy\r\ndata: 尾部',
+    { fragment: true },
+  )
+  const events = []
+  await readSSEStreamWithEvents('/stream', {
+    onMessage: text => events.push(text),
+    onToolCallStart: value => events.push(value),
+    onToolCallResult: value => events.push(value),
+    onError: value => events.push(value),
+    onDone: () => events.push('done'),
+  })
+  assert.deepEqual(events, [
+    { tool: 'lookup' }, { tool: 'lookup', result: 'ok' }, { message: 'tool unavailable' },
+    'first\n  indented', '\n', 'legacy', '尾部', 'done',
+  ])
+})
+
+test('EOF flushes an unterminated plain-text line and an unterminated DONE event', async () => {
+  for (const body of ['legacy tail', 'data: [DONE]']) {
+    globalThis.fetch = async () => responseFor(body)
+    const text = []
+    let done = 0
+    await readSSEStream('/stream', value => text.push(value), () => done++)
+    assert.deepEqual(text, body.startsWith('data:') ? [] : ['legacy tail'])
+    assert.equal(done, 1)
+  }
+})
+
+test('cleanup failure cannot replace callback errors or call onDone on an error', async () => {
+  const original = new Error('consumer failed')
+  const response = responseFor(`data: ${answer}\n\n`, {
+    open: true, cancel() { throw new Error('cleanup failed') },
+  })
+  globalThis.fetch = async () => response
+  let done = 0
+  await assert.rejects(readSSEStream('/stream', () => { throw original }, () => done++), error => error === original)
+  assert.equal(done, 0)
+  assert.equal(response.body.locked, false)
+})
+
+test('empty sources for ordinary answers are delivered separately from content', async () => {
+  globalThis.fetch = async () => responseFor('data: {"choices":[],"sources":[]}\n\ndata: [DONE]\n\n')
+  const values = []
+  await readSSEStreamWithEvents('/stream', {
+    onMessage: () => assert.fail('metadata is not text'),
+    onSources: value => values.push(value),
+  })
+  assert.deepEqual(values, [[]])
+})
+
+test('synchronous chat makes one result request and retains direct answer, sources, explicit strategy and signal', async () => {
+  const result = { text: '答复。', mcpEvents: [], suggestions: [], sources }
+  const calls = []
+  globalThis.fetch = async (...args) => { calls.push(args); return Response.json(result) }
+  const controller = new AbortController()
+  assert.deepEqual(await chatSync('费用?', 'conversation-1', {
+    enableRag: true, roleId: '9223372036854775806', rewriteStrategy: 'MULTI_QUERY', categoryIds: ['123'],
+  }, controller.signal), result)
+  assert.equal(calls.length, 1)
+  const url = new URL(calls[0][0], 'https://test.invalid')
+  assert.equal(url.pathname, '/ai/chat/result')
+  assert.equal(url.searchParams.get('rewriteStrategy'), 'MULTI_QUERY')
+  assert.equal(url.searchParams.get('roleId'), '9223372036854775806')
+  assert.equal(calls[0][1].signal, controller.signal)
+  assert.equal(calls[0][1].headers.Authorization, 'Bearer admin-token')
+})
+
+test('synchronous SDK response uses the same direct contract and surfaces server errors', async () => {
+  const result = { text: '普通回答', mcpEvents: [], suggestions: [], sources: [] }
+  globalThis.fetch = async () => Response.json(result)
+  assert.deepEqual(await fetchChatResult('https://sdk.invalid/ai/chat/result', {}), result)
+  globalThis.fetch = async () => new Response('角色无访问权限', { status: 403 })
+  await assert.rejects(fetchChatResult('/ai/chat/result', {}), /角色无访问权限/)
+  await assert.rejects(readSSEStream('/ai/chat/stream', () => {}), /角色无访问权限/)
+})