diff --git a/frontend-web/src/components/workspace/chats/chat-box.tsx b/frontend-web/src/components/workspace/chats/chat-box.tsx
index 31ec10d9..20b6e434 100644
--- a/frontend-web/src/components/workspace/chats/chat-box.tsx
+++ b/frontend-web/src/components/workspace/chats/chat-box.tsx
@@ -12,7 +12,9 @@ import {
   ResizablePanelGroup,
 } from "@/components/ui/resizable";
 import { Tabs, TabsContent, TabsList, TabsTrigger } from "@/components/ui/tabs";
-import { useLocalSettings } from "@/core/settings";
+import { apiFetch } from "@/core/api";
+import { getBackendBaseURL } from "@/core/config";
+import { isThreadRagModeLocked, useLocalSettings } from "@/core/settings";
 import { useSkills } from "@/core/skills/hooks";
 import {
   buildSkillDisplayMap,
@@ -33,6 +35,33 @@ const CLOSE_MODE = { chat: 100, artifacts: 0 };
 const OPEN_MODE = { chat: 60, artifacts: 40 };
 const SHOW_REFERENCE_EVENT = "workspace:show-reference";
 
+type BackendReferenceSource = Parameters<typeof SkillDisplaySources>[0]["sources"][number];
+
+type BackendReferenceBatch = {
+  id: string;
+  assistant_message_id?: string;
+  sources: BackendReferenceSource[];
+};
+
+function logReferenceDebug(
+  event: string,
+  threadId: string,
+  payload: Record<string, unknown>,
+) {
+  console.info(`[RAG引用调试] ${event}`, payload);
+  void apiFetch(`${getBackendBaseURL()}/api/threads/reference-debug-log`, {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({
+      event,
+      thread_id: threadId,
+      payload,
+    }),
+  }).catch(() => {
+    // Debug logging must never affect the chat flow.
+  });
+}
+
 const ChatBox: React.FC<{ children: React.ReactNode; threadId: string }> = ({
   children,
   threadId,
@@ -42,10 +71,17 @@ const ChatBox: React.FC<{ children: React.ReactNode; threadId: string }> = ({
   const threadIdRef = useRef(threadId);
   const artifactsCountRef = useRef(0);
   const layoutRef = useRef<GroupImperativeHandle>(null);
+  const latestReferenceBatchIdRef = useRef<string | null>(null);
+  const manualReferenceBatchRef = useRef(false);
+  const referenceDebugSignatureRef = useRef<string>("");
   const [sideTab, setSideTab] = useState("document");
   const [referencesOpen, setReferencesOpen] = useState(true);
   const [activeReferenceBatchId, setActiveReferenceBatchId] = useState<string | null>(null);
+  const [pinnedReferenceId, setPinnedReferenceId] = useState<string | null>(null);
+  const [backendReferenceBatches, setBackendReferenceBatches] = useState<BackendReferenceBatch[]>([]);
   const [settings] = useLocalSettings();
+  const ragModeEnabled =
+    settings.display.ragModeEnabled || isThreadRagModeLocked(threadId);
   const { skills } = useSkills();
 
   const {
@@ -60,27 +96,84 @@ const ChatBox: React.FC<{ children: React.ReactNode; threadId: string }> = ({
 
   const [autoSelectFirstArtifact, setAutoSelectFirstArtifact] = useState(true);
   const skillDisplayMap = useMemo(
-    () => (settings.display.ragModeEnabled ? buildSkillDisplayMap(skills) : new Map()),
-    [settings.display.ragModeEnabled, skills],
+    () => (ragModeEnabled ? buildSkillDisplayMap(skills) : new Map()),
+    [ragModeEnabled, skills],
   );
-  const referenceBatches = useMemo(() => {
-    if (!settings.display.ragModeEnabled) {
+  const fallbackReferenceBatches = useMemo(() => {
+    if (!ragModeEnabled) {
       return [];
     }
-    const messages = thread.values.messages ?? [];
-    return messages
-      .filter((message) => message.type === "ai" && message.id)
-      .map((message) => ({
-        id: message.id!,
-        sources: collectRuntimeDisplaySources(messages, skillDisplayMap, {
-          afterMessageId: findPreviousHumanMessageId(messages, message.id),
-          beforeMessageId: message.id,
-          includeBeforeMessage: true,
-          includeModes: ["citation"],
-        }).filter((source) => source.mode === "citation"),
-      }))
-      .filter((batch) => batch.sources.length > 0);
-  }, [settings.display.ragModeEnabled, skillDisplayMap, thread.values.messages]);
+    return buildTurnReferenceBatches(
+      thread.values.messages ?? [],
+      skillDisplayMap,
+      thread.isLoading,
+    );
+  }, [ragModeEnabled, skillDisplayMap, thread.isLoading, thread.values.messages]);
+  useEffect(() => {
+    if (!ragModeEnabled) {
+      setBackendReferenceBatches([]);
+      return;
+    }
+    if (thread.isLoading) {
+      return;
+    }
+    const controller = new AbortController();
+    logReferenceDebug("开始拉取后端参考文献", threadId, {
+      messageCount: thread.values.messages?.length ?? 0,
+    });
+    apiFetch(
+      `${getBackendBaseURL()}/api/threads/${encodeURIComponent(threadId)}/reference-batches`,
+      { signal: controller.signal },
+    )
+      .then(async (response) => {
+        if (!response.ok) {
+          throw new Error(`Failed to load reference batches: ${response.status}`);
+        }
+        return (await response.json()) as { batches?: BackendReferenceBatch[] };
+      })
+      .then((data) => {
+        const batches = data.batches ?? [];
+        logReferenceDebug("后端参考文献返回", threadId, {
+          batchCount: batches.length,
+          sourceCounts: batches.map((batch) => batch.sources?.length ?? 0),
+          batchIds: batches.map((batch) => batch.id),
+        });
+        setBackendReferenceBatches(batches);
+      })
+      .catch((error) => {
+        if (!controller.signal.aborted) {
+          console.warn("Failed to load backend reference batches", error);
+        }
+      });
+    return () => controller.abort();
+  }, [ragModeEnabled, thread.isLoading, threadId, thread.values.messages]);
+  const referenceBatches =
+    backendReferenceBatches.length > 0
+      ? backendReferenceBatches
+      : fallbackReferenceBatches;
+  useEffect(() => {
+    if (!ragModeEnabled) {
+      return;
+    }
+    const signature = JSON.stringify({
+      backend: backendReferenceBatches.map((batch) => batch.sources?.length ?? 0),
+      fallback: fallbackReferenceBatches.map((batch) => batch.sources?.length ?? 0),
+      final: referenceBatches.map((batch) => batch.sources?.length ?? 0),
+    });
+    if (referenceDebugSignatureRef.current === signature) {
+      return;
+    }
+    referenceDebugSignatureRef.current = signature;
+    logReferenceDebug("前端参考文献来源选择", threadId, {
+      using: backendReferenceBatches.length > 0 ? "后端接口" : "前端兜底",
+      backendBatchCount: backendReferenceBatches.length,
+      backendSourceCounts: backendReferenceBatches.map((batch) => batch.sources?.length ?? 0),
+      fallbackBatchCount: fallbackReferenceBatches.length,
+      fallbackSourceCounts: fallbackReferenceBatches.map((batch) => batch.sources?.length ?? 0),
+      finalBatchCount: referenceBatches.length,
+      finalSourceCounts: referenceBatches.map((batch) => batch.sources?.length ?? 0),
+    });
+  }, [backendReferenceBatches, fallbackReferenceBatches, ragModeEnabled, referenceBatches, threadId]);
   const latestReferenceBatch =
     referenceBatches.length > 0
       ? referenceBatches[referenceBatches.length - 1]!
@@ -98,20 +191,32 @@ const ChatBox: React.FC<{ children: React.ReactNode; threadId: string }> = ({
   useEffect(() => {
     if (!latestReferenceBatch) {
       setActiveReferenceBatchId(null);
+      latestReferenceBatchIdRef.current = null;
+      manualReferenceBatchRef.current = false;
       return;
     }
+    const previousLatestId = latestReferenceBatchIdRef.current;
+    latestReferenceBatchIdRef.current = latestReferenceBatch.id;
     setActiveReferenceBatchId((current) =>
-      current && referenceBatches.some((batch) => batch.id === current)
+      current &&
+      referenceBatches.some((batch) => batch.id === current) &&
+      manualReferenceBatchRef.current &&
+      current !== previousLatestId
         ? current
         : latestReferenceBatch.id,
     );
   }, [latestReferenceBatch?.id, referenceBatches]);
   useEffect(() => {
     const handleShowReference = (event: Event) => {
-      const detail = (event as CustomEvent<{ batchId?: string }>).detail;
+      const detail = (event as CustomEvent<{ batchId?: string; referenceId?: string }>).detail;
       if (detail?.batchId) {
+        manualReferenceBatchRef.current = true;
         setActiveReferenceBatchId(detail.batchId);
       }
+      setPinnedReferenceId(null);
+      window.setTimeout(() => {
+        setPinnedReferenceId(detail?.referenceId ?? null);
+      }, 0);
       setReferencesOpen(true);
       setSideTab("references");
     };
@@ -125,7 +230,13 @@ const ChatBox: React.FC<{ children: React.ReactNode; threadId: string }> = ({
       threadIdRef.current = threadId;
       artifactsCountRef.current = 0;
       deselect();
-      setReferencesOpen(true);
+      setReferencesOpen(false);
+      setActiveReferenceBatchId(null);
+      setPinnedReferenceId(null);
+      setBackendReferenceBatches([]);
+      latestReferenceBatchIdRef.current = null;
+      manualReferenceBatchRef.current = false;
+      referenceDebugSignatureRef.current = "";
     }
 
     // Update artifacts from the current thread
@@ -257,7 +368,10 @@ const ChatBox: React.FC<{ children: React.ReactNode; threadId: string }> = ({
                 />
               </TabsContent>
               <TabsContent value="references" className="min-h-0 flex-1 overflow-y-auto p-0">
-                <ReferencePanel sources={referenceSources} />
+                <ReferencePanel
+                  highlightedReferenceId={pinnedReferenceId}
+                  sources={referenceSources}
+                />
               </TabsContent>
             </Tabs>
           ) : selectedArtifact ? (
@@ -280,7 +394,10 @@ const ChatBox: React.FC<{ children: React.ReactNode; threadId: string }> = ({
                   <XIcon />
                 </Button>
               </div>
-              <ReferencePanel sources={referenceSources} />
+              <ReferencePanel
+                highlightedReferenceId={pinnedReferenceId}
+                sources={referenceSources}
+              />
             </div>
           ) : (
             <div className="relative flex size-full justify-center">
@@ -324,10 +441,27 @@ const ChatBox: React.FC<{ children: React.ReactNode; threadId: string }> = ({
 };
 
 function ReferencePanel({
+  highlightedReferenceId,
   sources,
 }: {
+  highlightedReferenceId?: string | null;
   sources: Parameters<typeof SkillDisplaySources>[0]["sources"];
 }) {
+  const orderedSources = useMemo(() => {
+    if (!highlightedReferenceId) {
+      return sources;
+    }
+    const index = Number(highlightedReferenceId);
+    if (!Number.isFinite(index)) {
+      return sources;
+    }
+    const selected = sources.find((source) => source.index === index);
+    if (!selected) {
+      return sources;
+    }
+    return [selected, ...sources.filter((source) => source.index !== index)];
+  }, [highlightedReferenceId, sources]);
+
   return (
     <div className="flex size-full flex-col overflow-hidden">
       <header className="shrink-0 px-1 pb-3">
@@ -344,7 +478,10 @@ function ReferencePanel({
         </p>
       </header>
       <main className="min-h-0 flex-1 overflow-y-auto px-1 pt-1 pb-6">
-        <SkillDisplaySources sources={sources} />
+        <SkillDisplaySources
+          highlightedReferenceId={highlightedReferenceId}
+          sources={orderedSources}
+        />
       </main>
     </div>
   );
@@ -368,4 +505,101 @@ function findPreviousHumanMessageId(
   return undefined;
 }
 
+function buildTurnReferenceBatches(
+  messages: Message[],
+  skillDisplayMap: ReturnType<typeof buildSkillDisplayMap>,
+  isLoading: boolean,
+): BackendReferenceBatch[] {
+  const batches: BackendReferenceBatch[] = [];
+  for (let index = 0; index < messages.length; index++) {
+    const human = messages[index];
+    if (human?.type !== "human" || !human.id) {
+      continue;
+    }
+    const nextHumanIndex = findNextHumanIndex(messages, index + 1);
+    if (nextHumanIndex === -1 && isLoading) {
+      continue;
+    }
+    const endIndex = nextHumanIndex === -1 ? messages.length : nextHumanIndex;
+    const assistant = findLastAssistantAnswer(messages, index + 1, endIndex);
+    if (!assistant?.id) {
+      continue;
+    }
+    const sources = collectRuntimeDisplaySources(messages, skillDisplayMap, {
+      afterMessageId: human.id,
+      beforeMessageId:
+        nextHumanIndex === -1 ? undefined : messages[nextHumanIndex]?.id,
+      includeBeforeMessage: false,
+      includeModes: ["citation"],
+    }).filter((source) => source.mode === "citation");
+    if (sources.length > 0) {
+      batches.push({
+        id: assistant.id,
+        assistant_message_id: assistant.id,
+        sources,
+      });
+    }
+  }
+  return batches;
+}
+
+function findNextHumanIndex(messages: Message[], start: number): number {
+  for (let index = start; index < messages.length; index++) {
+    if (messages[index]?.type === "human") {
+      return index;
+    }
+  }
+  return -1;
+}
+
+function findLastAssistantAnswer(
+  messages: Message[],
+  start: number,
+  end: number,
+): Message | null {
+  for (let index = end - 1; index >= start; index--) {
+    const message = messages[index];
+    if (
+      message?.type === "ai" &&
+      message.id &&
+      hasAssistantAnswerContent(message) &&
+      !hasAssistantToolCalls(message)
+    ) {
+      return message;
+    }
+  }
+  return null;
+}
+
+function hasAssistantAnswerContent(message: Message): boolean {
+  const content = message.content;
+  if (typeof content === "string") {
+    return content.trim().length > 0;
+  }
+  if (!Array.isArray(content)) {
+    return false;
+  }
+  return content.some((part) => {
+    if (!part || typeof part !== "object") {
+      return false;
+    }
+    const block = part as Record<string, unknown>;
+    return block.type === "text" && typeof block.text === "string" && block.text.trim().length > 0;
+  });
+}
+
+function hasAssistantToolCalls(message: Message): boolean {
+  const direct = (message as { tool_calls?: unknown }).tool_calls;
+  if (Array.isArray(direct) && direct.length > 0) {
+    return true;
+  }
+  const chunks = (message as { tool_call_chunks?: unknown }).tool_call_chunks;
+  if (Array.isArray(chunks) && chunks.length > 0) {
+    return true;
+  }
+  const kwargs = (message as { additional_kwargs?: Record<string, unknown> }).additional_kwargs;
+  const extra = kwargs?.tool_calls;
+  return Array.isArray(extra) && extra.length > 0;
+}
+
 export { ChatBox };
diff --git a/frontend-web/src/components/workspace/input-box.tsx b/frontend-web/src/components/workspace/input-box.tsx
index 6e1ce3b3..966268fd 100644
--- a/frontend-web/src/components/workspace/input-box.tsx
+++ b/frontend-web/src/components/workspace/input-box.tsx
@@ -87,7 +87,11 @@ import {
 } from "../ui/dropdown-menu";
 
 import { useRecommendedQuestions } from "@/core/recommended-questions/hooks";
-import { useLocalSettings } from "@/core/settings";
+import {
+  isThreadRagModeLocked,
+  lockThreadRagMode,
+  useLocalSettings,
+} from "@/core/settings";
 
 
 import { useThread } from "./messages/context";
@@ -300,10 +304,14 @@ export function InputBox({
     });
   }, [context, memoryInjectionEnabled, onContextChange]);
 
-  const ragModeEnabled = localSettings.display.ragModeEnabled;
-  const handleRagModeToggle = useCallback(() => {
-    setLocalSettings("display", { ragModeEnabled: !ragModeEnabled });
-  }, [ragModeEnabled, setLocalSettings]);
+  const ragModeLocked = isThreadRagModeLocked(threadId);
+  const ragModeEnabled = localSettings.display.ragModeEnabled || ragModeLocked;
+  const handleRagModeToggle = useCallback(() => {
+    if (ragModeLocked) {
+      return;
+    }
+    setLocalSettings("display", { ragModeEnabled: !ragModeEnabled });
+  }, [ragModeEnabled, ragModeLocked, setLocalSettings]);
 
   const handleSubmit = useCallback(
     async (message: PromptInputMessage) => {
@@ -316,9 +324,12 @@ export function InputBox({
       }
       setFollowups([]);
       setFollowupsHidden(false);
-      setFollowupsLoading(false);
-
-      // Guard against submitting before the initial model auto-selection
+      setFollowupsLoading(false);
+      if (ragModeEnabled) {
+        lockThreadRagMode(threadId);
+      }
+
+      // Guard against submitting before the initial model auto-selection
       // effect has flushed thread settings to storage/state.
       if (resolvedModelName && context.model_name !== resolvedModelName) {
         onContextChange?.({
@@ -340,11 +351,13 @@ export function InputBox({
       onContextChange,
       onSubmit,
       onStop,
-      resolvedModelName,
-      selectedModel?.supports_thinking,
-      status,
-    ],
-  );
+      resolvedModelName,
+      ragModeEnabled,
+      selectedModel?.supports_thinking,
+      status,
+      threadId,
+    ],
+  );
 
   const requestFormSubmit = useCallback(() => {
     const form = promptRootRef.current?.querySelector("form");
@@ -638,11 +651,13 @@ export function InputBox({
               }
             >
               <PromptInputButton
-                type="button"
-                aria-pressed={ragModeEnabled}
-                className={cn(
-                  "gap-1.5! rounded-full px-3! shadow-none",
-                  ragModeEnabled
+                type="button"
+                aria-pressed={ragModeEnabled}
+                disabled={ragModeLocked || disabled}
+                className={cn(
+                  "gap-1.5! rounded-full px-3! shadow-none",
+                  ragModeLocked && "cursor-not-allowed opacity-80",
+                  ragModeEnabled
                     ? "border border-emerald-300 bg-emerald-50 text-emerald-700 hover:bg-emerald-100 dark:border-emerald-700 dark:bg-emerald-950/40 dark:text-emerald-200"
                     : "border-0 bg-transparent text-muted-foreground hover:bg-muted/70 hover:text-foreground dark:text-slate-400 dark:hover:bg-slate-700/70 dark:hover:text-slate-100",
                 )}
diff --git a/frontend-web/src/components/workspace/messages/markdown-content.tsx b/frontend-web/src/components/workspace/messages/markdown-content.tsx
index 1a67d6ec..dc0888c5 100644
--- a/frontend-web/src/components/workspace/messages/markdown-content.tsx
+++ b/frontend-web/src/components/workspace/messages/markdown-content.tsx
@@ -36,10 +36,13 @@ export function MarkdownContent({
 }: MarkdownContentProps) {
   const components = useMemo(() => {
     return {
-      a: (props: AnchorHTMLAttributes<HTMLAnchorElement>) => {
-        if (typeof props.children === "string") {
-          const match = /^citation:(.+)$/.exec(props.children);
-          if (match) {
+      a: (props: AnchorHTMLAttributes<HTMLAnchorElement>) => {
+        if (props.href?.startsWith("#reference-")) {
+          return <CitationLink {...props} />;
+        }
+        if (typeof props.children === "string") {
+          const match = /^citation:(.+)$/.exec(props.children);
+          if (match) {
             const [, text] = match;
             return <CitationLink {...props}>{text}</CitationLink>;
           }
diff --git a/frontend-web/src/components/workspace/messages/message-list-item.tsx b/frontend-web/src/components/workspace/messages/message-list-item.tsx
index f586f9de..23924482 100644
--- a/frontend-web/src/components/workspace/messages/message-list-item.tsx
+++ b/frontend-web/src/components/workspace/messages/message-list-item.tsx
@@ -184,10 +184,13 @@ function MessageContent_({
       img: (props: ImgHTMLAttributes<HTMLImageElement>) => (
         <MessageImage {...props} threadId={threadId} maxWidth="90%" />
       ),
-      a: ({ href, ...props }: AnchorHTMLAttributes<HTMLAnchorElement>) => {
-        if (typeof props.children === "string" && /^citation:/i.test(props.children)) {
-          return <CitationLink href={href} {...props} />;
-        }
+      a: ({ href, ...props }: AnchorHTMLAttributes<HTMLAnchorElement>) => {
+        if (href?.startsWith("#reference-")) {
+          return <CitationLink href={href} {...props} />;
+        }
+        if (typeof props.children === "string" && /^citation:/i.test(props.children)) {
+          return <CitationLink href={href} {...props} />;
+        }
         if (href?.startsWith("/mnt/")) {
           const url = resolveArtifact(href);
           return (
diff --git a/frontend-web/src/components/workspace/messages/message-list.tsx b/frontend-web/src/components/workspace/messages/message-list.tsx
index e659d72f..0523c307 100644
--- a/frontend-web/src/components/workspace/messages/message-list.tsx
+++ b/frontend-web/src/components/workspace/messages/message-list.tsx
@@ -21,7 +21,7 @@ import {
   hasToolCalls,
 } from "@/core/messages/utils";
 import { useRehypeSplitWordsIntoSpans } from "@/core/rehype";
-import { useLocalSettings } from "@/core/settings";
+import { isThreadRagModeLocked, useLocalSettings } from "@/core/settings";
 import { useSkills } from "@/core/skills/hooks";
 import {
   buildSkillDisplayMap,
@@ -73,10 +73,12 @@ export function MessageList({
   const updateSubtask = useUpdateSubtask();
   const messages = thread.messages;
   const [settings] = useLocalSettings();
+  const ragModeEnabled =
+    settings.display.ragModeEnabled || isThreadRagModeLocked(threadId);
   const { skills } = useSkills();
   const skillDisplayMap = useMemo(
-    () => (settings.display.ragModeEnabled ? buildSkillDisplayMap(skills) : new Map()),
-    [settings.display.ragModeEnabled, skills],
+    () => (ragModeEnabled ? buildSkillDisplayMap(skills) : new Map()),
+    [ragModeEnabled, skills],
   );
 
   // Track which clarification messages are still awaiting an answer.
@@ -198,7 +200,7 @@ export function MessageList({
           if (group.type === "human" || group.type === "assistant") {
             return group.messages.map((msg) => {
               const displaySources =
-                settings.display.ragModeEnabled &&
+                ragModeEnabled &&
                 group.type === "assistant" &&
                 msg.type === "ai"
                   ? collectRuntimeDisplaySources(messages, skillDisplayMap, {
diff --git a/frontend-web/src/components/workspace/settings/appearance-settings-page.tsx b/frontend-web/src/components/workspace/settings/appearance-settings-page.tsx
index b1498420..5970d17e 100644
--- a/frontend-web/src/components/workspace/settings/appearance-settings-page.tsx
+++ b/frontend-web/src/components/workspace/settings/appearance-settings-page.tsx
@@ -21,10 +21,12 @@ import { getStoredAuth } from "@/core/auth";
 import { enUS, isLocale, zhCN, type Locale } from "@/core/i18n";
 import { useI18n } from "@/core/i18n/hooks";
 import { useLocalSettings } from "@/core/settings";
-import {
-  useAppearanceSettings,
-  useUpdateAppearanceSettings,
-} from "@/core/system-settings/hooks";
+import {
+  useAppearanceSettings,
+  useCitationDisplaySettings,
+  useUpdateAppearanceSettings,
+  useUpdateCitationDisplaySettings,
+} from "@/core/system-settings/hooks";
 import { cn } from "@/lib/utils";
 
 import { SettingsSection } from "./settings-section";
@@ -40,9 +42,13 @@ export function AppearanceSettingsPage() {
   const currentTheme = theme;
   const [settings, setSettings] = useLocalSettings();
   const isAdmin = getStoredAuth()?.system_role === "admin";
-  const { data: appearance } = useAppearanceSettings();
-  const updateAppearance = useUpdateAppearanceSettings();
-  const globalDefault = appearance?.default_theme ?? "light";
+  const { data: appearance } = useAppearanceSettings();
+  const updateAppearance = useUpdateAppearanceSettings();
+  const { data: citationDisplay } = useCitationDisplaySettings();
+  const updateCitationDisplay = useUpdateCitationDisplaySettings();
+  const globalDefault = appearance?.default_theme ?? "light";
+  const cleanupWords =
+    citationDisplay?.cleanup_words ?? settings.display.citationCleanupWords ?? [];
 
   const handleSetGlobalDefault = async (value: "light" | "dark") => {
     if (value === globalDefault) return;
@@ -211,16 +217,16 @@ export function AppearanceSettingsPage() {
                 管理员配置。每行一个词或短句，参考文献卡片和详情弹框会自动移除这些广告词、导航词或免责声明。
               </p>
             </div>
-            <Textarea
-              value={(settings.display.citationCleanupWords ?? []).join("\n")}
-              onChange={(event) =>
-                setSettings("display", {
-                  citationCleanupWords: event.currentTarget.value
-                    .split(/\r?\n/)
-                    .map((item) => item.trim())
-                    .filter(Boolean),
-                })
-              }
+            <Textarea
+              value={cleanupWords.join("\n")}
+              onChange={(event) => {
+                const words = event.currentTarget.value
+                    .split(/\r?\n/)
+                    .map((item) => item.trim())
+                    .filter(Boolean);
+                setSettings("display", { citationCleanupWords: words });
+                updateCitationDisplay.mutate({ cleanup_words: words });
+              }}
               placeholder={"广告\n点击下载APP\n打开客户端\n免责声明\n责任编辑"}
               className="min-h-32 text-sm"
             />
diff --git a/frontend-web/src/components/workspace/skill-display-sources.tsx b/frontend-web/src/components/workspace/skill-display-sources.tsx
index 7bbf609d..080be3c0 100644
--- a/frontend-web/src/components/workspace/skill-display-sources.tsx
+++ b/frontend-web/src/components/workspace/skill-display-sources.tsx
@@ -10,13 +10,16 @@ import {
 } from "@/components/ui/dialog";
 import type { RuntimeDisplaySource } from "@/core/skill-display/runtime";
 import { useLocalSettings } from "@/core/settings";
+import { useCitationDisplaySettings } from "@/core/system-settings/hooks";
 import { cn } from "@/lib/utils";
 import { useState } from "react";
 
 export function SkillDisplaySources({
+  highlightedReferenceId,
   sources,
   className,
 }: {
+  highlightedReferenceId?: string | null;
   sources: RuntimeDisplaySource[];
   className?: string;
 }) {
@@ -29,6 +32,7 @@ export function SkillDisplaySources({
       {sources.map((source) => (
         <SourceCard
           key={`${source.skillName}-${source.index}`}
+          highlighted={highlightedReferenceId === String(source.index)}
           source={source}
           referenceTarget
         />
@@ -38,15 +42,19 @@ export function SkillDisplaySources({
 }
 
 export function SourceCard({
+  highlighted = false,
   source,
   referenceTarget = false,
 }: {
+  highlighted?: boolean;
   source: RuntimeDisplaySource;
   referenceTarget?: boolean;
 }) {
   const [open, setOpen] = useState(false);
   const [settings] = useLocalSettings();
-  const cleanupWords = settings.display.citationCleanupWords ?? [];
+  const { data: citationDisplay } = useCitationDisplaySettings();
+  const cleanupWords =
+    citationDisplay?.cleanup_words ?? settings.display.citationCleanupWords ?? [];
   const content = cleanReferenceText(source.content || source.snippet, cleanupWords);
   const snippet = compactReferenceText(content).slice(0, 280);
   const paragraphs = segmentReferenceText(content);
@@ -72,6 +80,8 @@ export function SourceCard({
         className={cn(
           "group bg-card/60 hover:bg-card hover:border-primary/40 focus-visible:ring-primary/30 w-full scroll-mt-12 rounded-xl border px-3.5 py-3 text-left shadow-sm transition-all focus-visible:ring-2 focus-visible:outline-none",
           "target:bg-primary/5 target:border-primary/60",
+          highlighted &&
+            "animate-[reference-flash_1.4s_ease-out] border-amber-400 bg-amber-50 shadow-[0_0_0_3px_rgba(251,191,36,0.22)]",
         )}
         onClick={() => setOpen(true)}
       >
@@ -168,10 +178,13 @@ function cleanReferenceText(
   cleanupWords: string[],
 ): string {
   let text = raw ?? "";
-  for (const word of cleanupWords) {
-    const trimmed = word.trim();
-    if (!trimmed) continue;
-    text = text.split(trimmed).join("");
+  const words = cleanupWords.map((word) => word.trim()).filter(Boolean);
+  if (words.length > 0) {
+    text = text
+      .replace(/\r\n/g, "\n")
+      .split("\n")
+      .filter((line) => !words.some((word) => line.includes(word)))
+      .join("\n");
   }
   return text
     .replace(/\r\n/g, "\n")
diff --git a/frontend-web/src/core/messages/search-results.ts b/frontend-web/src/core/messages/search-results.ts
index 7871d423..3ff6408e 100644
--- a/frontend-web/src/core/messages/search-results.ts
+++ b/frontend-web/src/core/messages/search-results.ts
@@ -7,6 +7,10 @@
  */
 
 export type SearchResultItem = {
+  id?: string;
+  recUuid?: string;
+  uuid?: string;
+  docId?: string;
   title?: string;
   url?: string;
   snippet?: string;
@@ -300,6 +304,10 @@ function normalizeSearchResultItem(raw: unknown): SearchResultItem | null {
   const date =
     pickFromLayers(layers, DATE_ALIASES) ??
     pickByKeyPattern(layers, DATE_KEY_PATTERN, { maxLength: 64 });
+  const id = pickFromLayers(layers, ["id", "taskId", "recordId", "record_id"]);
+  const recUuid = pickFromLayers(layers, ["recUuid", "rec_uuid"]);
+  const uuid = pickFromLayers(layers, ["uuid"]);
+  const docId = pickFromLayers(layers, ["docId", "doc_id", "documentId"]);
 
   const reserved = new Set<string>();
   for (const value of [title, content, url, source, date]) {
@@ -324,6 +332,10 @@ function normalizeSearchResultItem(raw: unknown): SearchResultItem | null {
   }
 
   return {
+    id,
+    recUuid,
+    uuid,
+    docId,
     title,
     url,
     snippet: content,
diff --git a/frontend-web/src/core/settings/index.ts b/frontend-web/src/core/settings/index.ts
index 12e80835..dc15f8cd 100644
--- a/frontend-web/src/core/settings/index.ts
+++ b/frontend-web/src/core/settings/index.ts
@@ -1,2 +1,3 @@
-export { useLocalSettings, useThreadSettings } from "./hooks";
-export type { LocalSettings } from "./local";
+export { useLocalSettings, useThreadSettings } from "./hooks";
+export { isThreadRagModeLocked, lockThreadRagMode } from "./local";
+export type { LocalSettings } from "./local";
diff --git a/frontend-web/src/core/settings/local.ts b/frontend-web/src/core/settings/local.ts
index 0e9080fe..e000e035 100644
--- a/frontend-web/src/core/settings/local.ts
+++ b/frontend-web/src/core/settings/local.ts
@@ -17,8 +17,9 @@ export const DEFAULT_LOCAL_SETTINGS: LocalSettings = {
   },
 };
 
-export const LOCAL_SETTINGS_KEY = "deerflow.local-settings";
-export const THREAD_MODEL_KEY_PREFIX = "deerflow.thread-model.";
+export const LOCAL_SETTINGS_KEY = "deerflow.local-settings";
+export const THREAD_MODEL_KEY_PREFIX = "deerflow.thread-model.";
+export const THREAD_RAG_MODE_KEY_PREFIX = "deerflow.thread-rag-mode.";
 
 function isBrowser(): boolean {
   return typeof window !== "undefined";
@@ -90,10 +91,10 @@ export function getThreadModelName(threadId: string): string | undefined {
   return localStorage.getItem(getThreadModelStorageKey(threadId)) ?? undefined;
 }
 
-export function saveThreadModelName(
-  threadId: string,
-  modelName: string | undefined,
-) {
+export function saveThreadModelName(
+  threadId: string,
+  modelName: string | undefined,
+) {
   if (!isBrowser()) {
     return;
   }
@@ -102,8 +103,26 @@ export function saveThreadModelName(
     localStorage.removeItem(key);
     return;
   }
-  localStorage.setItem(key, modelName);
-}
+  localStorage.setItem(key, modelName);
+}
+
+function getThreadRagModeStorageKey(threadId: string): string {
+  return `${THREAD_RAG_MODE_KEY_PREFIX}${threadId}`;
+}
+
+export function isThreadRagModeLocked(threadId: string): boolean {
+  if (!isBrowser() || !threadId) {
+    return false;
+  }
+  return localStorage.getItem(getThreadRagModeStorageKey(threadId)) === "true";
+}
+
+export function lockThreadRagMode(threadId: string) {
+  if (!isBrowser() || !threadId) {
+    return;
+  }
+  localStorage.setItem(getThreadRagModeStorageKey(threadId), "true");
+}
 
 export function applyThreadModelOverride(
   settings: LocalSettings,
diff --git a/frontend-web/src/core/skill-display/runtime.ts b/frontend-web/src/core/skill-display/runtime.ts
index 5d73b5f0..205f5f44 100644
--- a/frontend-web/src/core/skill-display/runtime.ts
+++ b/frontend-web/src/core/skill-display/runtime.ts
@@ -168,6 +168,7 @@ export function collectRuntimeDisplaySources(
   const seenCitationKeys = new Map<string, RuntimeDisplaySource>();
   const seenDisplayKeys = new Set<string>();
   const consumedToolMessageIds = new Set<string>();
+  const consumedToolMessages = new Set<Message>();
 
   for (let index = start; index < limit; index++) {
     const message = messages[index];
@@ -180,14 +181,24 @@ export function collectRuntimeDisplaySources(
         displays,
         toolCall.args,
       );
-      const result = findToolResultPayload(toolCall, messages, index);
+      const toolMessage = findToolResultMessage(
+        toolCall,
+        messages,
+        index,
+        consumedToolMessageIds,
+        consumedToolMessages,
+      );
+      if (!toolMessage) {
+        continue;
+      }
+      const result = parseToolMessagePayload(toolMessage);
       if (result === undefined) {
         continue;
       }
-      const linkedToolMessage = findToolMessageById(toolCall.id, messages);
-      if (linkedToolMessage?.id) {
-        consumedToolMessageIds.add(linkedToolMessage.id);
+      if (toolMessage.id) {
+        consumedToolMessageIds.add(toolMessage.id);
       }
+      consumedToolMessages.add(toolMessage);
       if (isDisplaySamplePayload(result)) {
         continue;
       }
@@ -223,7 +234,12 @@ export function collectRuntimeDisplaySources(
 
   for (let index = start; index < limit; index++) {
     const message = messages[index];
-    if (!message || message.type !== "tool" || consumedToolMessageIds.has(message.id ?? "")) {
+    if (
+      !message ||
+      message.type !== "tool" ||
+      consumedToolMessages.has(message) ||
+      consumedToolMessageIds.has(message.id ?? "")
+    ) {
       continue;
     }
     const result = parseToolMessagePayload(message);
@@ -238,7 +254,8 @@ export function collectRuntimeDisplaySources(
     }
     const normalized = match
       ? normalizeDisplaySources(result, match.display)
-      : shouldUseDefaultCitationMode(toolName, result)
+      : shouldUseDefaultCitationMode(toolName, result) ||
+          normalizeDefaultDisplaySources(result).length > 0
         ? normalizeDefaultDisplaySources(result)
         : [];
     appendDisplaySources(
@@ -268,16 +285,14 @@ function appendDisplaySources(
 ) {
   for (const draft of drafts) {
     if (draft.mode === "citation") {
-      const key = citationSourceKey(draft);
-      const existing = seenCitationKeys.get(key);
-      if (existing) {
-        continue;
-      }
       const source = {
         ...draft,
         index: nextCitationIndex(),
       };
-      seenCitationKeys.set(key, source);
+      const key = citationSourceKey(draft);
+      if (key) {
+        seenCitationKeys.set(`${key}:${source.index}`, source);
+      }
       target.push(source);
       continue;
     }
@@ -331,12 +346,19 @@ function citationSourceKey(source: RuntimeDisplaySourceDraft): string {
     stringifyMaybe(raw.docId) ||
     stringifyMaybe(raw.id);
   if (id) {
-    return normalizeName(`${source.skillName}:${id}`);
+    return normalizeCitationKey(String(id));
   }
-  const fallback = [source.skillName, source.title, source.source, source.time]
+  const fallback = [
+    source.skillName,
+    source.title,
+    source.source,
+    source.time,
+    source.content,
+    source.snippet,
+  ]
     .filter(Boolean)
     .join("|");
-  return normalizeName(fallback) || JSON.stringify(raw).slice(0, 300);
+  return normalizeCitationKey(fallback) || normalizeCitationKey(JSON.stringify(raw).slice(0, 300));
 }
 
 function nonCitationSourceKey(source: RuntimeDisplaySourceDraft): string {
@@ -543,7 +565,12 @@ export function normalizeDefaultDisplaySources(payload: unknown): DisplaySourceI
       : undefined,
   );
   return normalized.map((item, index) => ({
-    id: item.url ?? String(index + 1),
+    id:
+      item.recUuid ??
+      item.uuid ??
+      item.docId ??
+      item.id ??
+      item.url,
     title: item.title ?? item.source ?? `结果 ${index + 1}`,
     type: "result",
     snippet: item.snippet ?? item.content ?? "",
@@ -620,18 +647,79 @@ function replaceCitationTokens(
   byIndex: Map<number, RuntimeDisplaySource>,
   orderedSources: RuntimeDisplaySource[],
 ): string {
-  return content
+  const normalized = normalizeLooseCitationTokens(content, byIndex, orderedSources);
+  const withDoubleBracketCitations = normalized
     .replace(/\[\[(\d+)\]\]/g, (match, rawIndex: string) =>
       citationMarkdown(match, Number(rawIndex), byIndex, orderedSources),
-    )
-    .replace(/(^|[^:\[!])\[(\d+)\](?!\()/g, (match, prefix: string, rawIndex: string) =>
-      `${prefix}${citationMarkdown(
-        match.slice(prefix.length),
-        Number(rawIndex),
-        byIndex,
-        orderedSources,
-      )}`,
     );
+  return replaceBracketCitationTokens(
+    withDoubleBracketCitations,
+    byIndex,
+    orderedSources,
+  );
+}
+
+function replaceBracketCitationTokens(
+  content: string,
+  byIndex: Map<number, RuntimeDisplaySource>,
+  orderedSources: RuntimeDisplaySource[],
+): string {
+  const tokenPattern = /\\?\[(\d{1,3})\\?\](?!\()/g;
+  let cursor = 0;
+  let next = "";
+  for (const match of content.matchAll(tokenPattern)) {
+    const start = match.index ?? 0;
+    const previous = start > 0 ? content[start - 1] : "";
+    if (previous === ":" || previous === "[" || previous === "!") {
+      continue;
+    }
+    next += content.slice(cursor, start);
+    next += citationMarkdown(match[0], Number(match[1]), byIndex, orderedSources);
+    cursor = start + match[0].length;
+  }
+  next += content.slice(cursor);
+  return next;
+}
+
+function normalizeLooseCitationTokens(
+  content: string,
+  byIndex: Map<number, RuntimeDisplaySource>,
+  orderedSources: RuntimeDisplaySource[],
+): string {
+  const isValidIndex = (index: number) =>
+    Number.isInteger(index) &&
+    index > 0 &&
+    (byIndex.has(index) || !!orderedSources[index - 1]);
+  const toCitationRun = (raw: string) => {
+    const parts = raw.match(/\\?\[\d{1,3}\\?\]|\d{1,3}/g) ?? [];
+    const indexes = parts.map((part) => Number(part.replace(/\D/g, "")));
+    if (indexes.length === 0 || indexes.some((index) => !isValidIndex(index))) {
+      return raw;
+    }
+    return indexes.map((index) => `[${index}]`).join("");
+  };
+
+  let next = content.replace(
+    /(^|[^\d\[\]\\])((?:\d{1,3}|\\?\[\d{1,3}\\?\])(?:\\?\[\d{1,3}\\?\]|\d{1,3})+)/g,
+    (match, prefix: string, rawRun: string) => {
+      if (!rawRun.includes("[")) {
+        return match;
+      }
+      return `${prefix}${toCitationRun(rawRun)}`;
+    },
+  );
+
+  next = next.replace(
+    /(^|[\s\uFF0C\u3002\uFF1B\uFF1A\u3001,!?！？)])(\d{1,3})(?=([\s\uFF0C\u3002\uFF1B\uFF1A\u3001,!?！？)]|$))/g,
+    (match, prefix: string, rawIndex: string) => {
+      const index = Number(rawIndex);
+      if (!isValidIndex(index)) {
+        return match;
+      }
+      return `${prefix}[${index}]`;
+    },
+  );
+  return next;
 }
 
 function normalizeExistingCitationLinks(
@@ -654,16 +742,16 @@ function normalizeExistingCitationLinks(
         const index = Number(numberMatch[1]);
         const source = byIndex.get(index) ?? orderedSources[index - 1];
         if (source) {
-          return `[citation:[${source.index}]](#reference-${source.index})`;
+          return citationLinkMarkdown(source);
         }
       }
       const sourceByUrl = byUrl.get(url);
       if (sourceByUrl) {
-        return `[citation:[${sourceByUrl.index}]](#reference-${sourceByUrl.index})`;
+        return citationLinkMarkdown(sourceByUrl);
       }
       const sourceByTitle = byTitle.get(label);
       if (sourceByTitle) {
-        return `[citation:[${sourceByTitle.index}]](#reference-${sourceByTitle.index})`;
+        return citationLinkMarkdown(sourceByTitle);
       }
       return match;
     },
@@ -680,11 +768,11 @@ function citationMarkdown(
   if (!source) {
     return fallback;
   }
-  return `[citation:[${source.index}]](#reference-${source.index})`;
+  return citationLinkMarkdown(source);
 }
 
 function hasCitationMarkdown(content: string): boolean {
-  return /\[citation:\[[^\]]+\]\]\(#reference-[^)]+\)/.test(content);
+  return /\[citation:(?:\[[^\]]+\]|\d+)\]\(#reference-[^)]+\)/.test(content);
 }
 
 function appendCitationSummary(
@@ -693,7 +781,7 @@ function appendCitationSummary(
 ): string {
   const citations = citationSources
     .slice(0, 5)
-    .map((source) => `[citation:[${source.index}]](#reference-${source.index})`)
+    .map((source) => citationLinkMarkdown(source))
     .join(" ");
   if (!citations) {
     return content;
@@ -702,10 +790,22 @@ function appendCitationSummary(
   return `${trimmed}${trimmed ? "\n\n" : ""}参考来源：${citations}`;
 }
 
+function citationLinkMarkdown(source: RuntimeDisplaySource): string {
+  return `[citation:${source.index}](#reference-${source.index})`;
+}
+
 function normalizeName(name: string): string {
   return name.toLowerCase().replace(/[^a-z0-9]+/g, "");
 }
 
+function normalizeCitationKey(value: string): string {
+  return value
+    .toLowerCase()
+    .normalize("NFKC")
+    .replace(/\s+/g, " ")
+    .trim();
+}
+
 function isRecord(value: unknown): value is Record<string, unknown> {
   return !!value && typeof value === "object" && !Array.isArray(value);
 }
@@ -714,18 +814,28 @@ function stringifyMaybe(value: unknown): string | undefined {
   return typeof value === "string" && value.trim() ? value : undefined;
 }
 
-function findToolResultPayload(
+function findToolResultMessage(
   toolCall: RuntimeToolCall,
   messages: Message[],
   toolCallMessageIndex = -1,
-): unknown {
-  const toolMessage =
-    findToolMessageById(toolCall.id, messages) ??
-    findToolMessageByName(toolCall, messages, toolCallMessageIndex);
-  if (!toolMessage) {
-    return undefined;
+  consumedToolMessageIds?: Set<string>,
+  consumedToolMessages?: Set<Message>,
+): Message | undefined {
+  const byId = findToolMessageById(toolCall.id, messages);
+  if (
+    byId &&
+    !consumedToolMessages?.has(byId) &&
+    (!byId.id || !consumedToolMessageIds?.has(byId.id))
+  ) {
+    return byId;
   }
-  return parseToolMessagePayload(toolMessage);
+  return findToolMessageByName(
+    toolCall,
+    messages,
+    toolCallMessageIndex,
+    consumedToolMessageIds,
+    consumedToolMessages,
+  );
 }
 
 function parseToolMessagePayload(toolMessage: Message): unknown {
@@ -782,6 +892,8 @@ function findToolMessageByName(
   toolCall: RuntimeToolCall,
   messages: Message[],
   toolCallMessageIndex: number,
+  consumedToolMessageIds?: Set<string>,
+  consumedToolMessages?: Set<Message>,
 ): Message | undefined {
   const start = Math.max(toolCallMessageIndex + 1, 0);
   for (let index = start; index < messages.length; index++) {
@@ -795,6 +907,12 @@ function findToolMessageByName(
     if (message.type !== "tool") {
       continue;
     }
+    if (message.id && consumedToolMessageIds?.has(message.id)) {
+      continue;
+    }
+    if (consumedToolMessages?.has(message)) {
+      continue;
+    }
     const name =
       stringifyMaybe((message as { name?: unknown }).name) ??
       stringifyMaybe(
diff --git a/frontend-web/src/core/system-settings/api.ts b/frontend-web/src/core/system-settings/api.ts
index e6335706..265a2276 100644
--- a/frontend-web/src/core/system-settings/api.ts
+++ b/frontend-web/src/core/system-settings/api.ts
@@ -53,6 +53,14 @@ export interface PrivateSquareUpdate {
   password?: string;
 }
 
+export interface CitationDisplaySettings {
+  cleanup_words: string[];
+}
+
+export interface CitationDisplayUpdate {
+  cleanup_words?: string[];
+}
+
 export async function loadPrivateSquareStatus(): Promise<PrivateSquareStatus> {
   const response = await apiFetch(
     `${getBackendBaseURL()}/api/system-settings/private-square`,
@@ -107,6 +115,38 @@ export async function updateAppearanceSettings(
   return (await response.json()) as AppearanceSettings;
 }
 
+export async function loadCitationDisplaySettings(): Promise<CitationDisplaySettings> {
+  const response = await apiFetch(
+    `${getBackendBaseURL()}/api/system-settings/citation-display`,
+  );
+  if (!response.ok) {
+    throw new Error(
+      `Failed to load citation display settings: ${response.status}`,
+    );
+  }
+  return (await response.json()) as CitationDisplaySettings;
+}
+
+export async function updateCitationDisplaySettings(
+  body: CitationDisplayUpdate,
+): Promise<CitationDisplaySettings> {
+  const response = await apiFetch(
+    `${getBackendBaseURL()}/api/system-settings/citation-display`,
+    {
+      method: "PUT",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify(body),
+    },
+  );
+  if (!response.ok) {
+    const detail = await response
+      .json()
+      .catch(() => ({ detail: "Failed to update citation display" }));
+    throw new Error(detail.detail ?? "Failed to update citation display");
+  }
+  return (await response.json()) as CitationDisplaySettings;
+}
+
 export async function updatePrivateSquareSettings(
   body: PrivateSquareUpdate,
 ): Promise<PrivateSquareStatus> {
diff --git a/frontend-web/src/core/system-settings/hooks.ts b/frontend-web/src/core/system-settings/hooks.ts
index 903737a5..85cf0b41 100644
--- a/frontend-web/src/core/system-settings/hooks.ts
+++ b/frontend-web/src/core/system-settings/hooks.ts
@@ -2,12 +2,15 @@ import { useMutation, useQuery, useQueryClient } from "@tanstack/react-query";
 
 import {
   loadAppearanceSettings,
+  loadCitationDisplaySettings,
   loadPrivateSquareStatus,
   loadPromptPrefixSettings,
   updateAppearanceSettings,
+  updateCitationDisplaySettings,
   updatePrivateSquareSettings,
   updatePromptPrefixSettings,
   type AppearanceUpdate,
+  type CitationDisplayUpdate,
   type PrivateSquareUpdate,
   type PromptPrefixUpdate,
 } from "./api";
@@ -15,6 +18,7 @@ import {
 const PROMPT_PREFIX_KEY = ["systemSettings", "promptPrefix"] as const;
 const PRIVATE_SQUARE_KEY = ["systemSettings", "privateSquare"] as const;
 const APPEARANCE_KEY = ["systemSettings", "appearance"] as const;
+const CITATION_DISPLAY_KEY = ["systemSettings", "citationDisplay"] as const;
 
 export function usePromptPrefixSettings() {
   return useQuery({
@@ -77,3 +81,24 @@ export function useUpdateAppearanceSettings() {
     },
   });
 }
+
+export function useCitationDisplaySettings() {
+  return useQuery({
+    queryKey: CITATION_DISPLAY_KEY,
+    queryFn: () => loadCitationDisplaySettings(),
+    staleTime: 60_000,
+    refetchOnWindowFocus: false,
+    retry: false,
+  });
+}
+
+export function useUpdateCitationDisplaySettings() {
+  const queryClient = useQueryClient();
+  return useMutation({
+    mutationFn: (body: CitationDisplayUpdate) =>
+      updateCitationDisplaySettings(body),
+    onSuccess: (data) => {
+      queryClient.setQueryData(CITATION_DISPLAY_KEY, data);
+    },
+  });
+}
diff --git a/frontend-web/src/core/threads/hooks.ts b/frontend-web/src/core/threads/hooks.ts
index 55331139..e1687970 100644
--- a/frontend-web/src/core/threads/hooks.ts
+++ b/frontend-web/src/core/threads/hooks.ts
@@ -3,17 +3,19 @@ import type { ThreadsClient } from "@langchain/langgraph-sdk/client";
 import { useStream } from "@langchain/langgraph-sdk/react";
 import { useMutation, useQuery, useQueryClient } from "@tanstack/react-query";
 import { useCallback, useEffect, useRef, useState } from "react";
+import { useNavigate } from "react-router-dom";
 import { toast } from "sonner";
 
 import type { PromptInputMessage } from "@/components/ai-elements/prompt-input";
 
 import { getAPIClient } from "../api";
 import { apiFetch } from "../api";
+import { getStoredAuth } from "../auth";
 import { getBackendBaseURL } from "../config";
 import { useI18n } from "../i18n/hooks";
 import type { FileInMessage } from "../messages/utils";
-import type { LocalSettings } from "../settings";
-import { useLocalSettings } from "../settings";
+import type { LocalSettings } from "../settings";
+import { isThreadRagModeLocked, lockThreadRagMode, useLocalSettings } from "../settings";
 import { useUpdateSubtask } from "../tasks/context";
 import type { UploadedFileInfo } from "../uploads";
 import { promptInputFilePartToFile, uploadFiles } from "../uploads";
@@ -137,6 +139,42 @@ function getStreamErrorMessage(error: unknown): string {
   return "Request failed.";
 }
 
+/**
+ * Detect "thread not found" / forbidden errors from the LangGraph SDK.
+ *
+ * Backend returns 404 both when a thread truly doesn't exist and when the
+ * currently authenticated user doesn't own it — the SDK surfaces both as
+ * ``HTTP 404: {"detail":"Thread <uuid> not found"}``. The most common cause
+ * we see in practice is a stale tab whose URL still points at the previous
+ * login session's thread. Raw-toasting the 404 just confuses the user, so
+ * we re-frame it around the current login identity.
+ */
+function isThreadNotAccessibleError(message: string): boolean {
+  const lowered = message.toLowerCase();
+  if (/thread\s+[\w-]{8,}\s+not\s+found/i.test(message)) {
+    return true;
+  }
+  // Fallback: a bare 404/403 from the threads endpoint when the SDK has
+  // wrapped the detail differently.
+  return (
+    (lowered.includes("http 404") || lowered.includes("http 403")) &&
+    lowered.includes("thread")
+  );
+}
+
+function buildCurrentUserLabel(): string {
+  const auth = getStoredAuth();
+  if (!auth) {
+    return "未登录";
+  }
+  const email = auth.email?.trim();
+  if (email) {
+    return email;
+  }
+  const userId = auth.user_id?.trim();
+  return userId ? userId : "未知用户";
+}
+
 export function useThreadStream({
   threadId,
   context,
@@ -195,6 +233,7 @@ export function useThreadStream({
 
   const queryClient = useQueryClient();
   const updateSubtask = useUpdateSubtask();
+  const navigate = useNavigate();
   const runMetadataStorageRef = useRef<
     ReturnType<typeof getRunMetadataStorage> | undefined
   >(undefined);
@@ -294,7 +333,29 @@ export function useThreadStream({
     },
     onError(error) {
       setOptimisticMessages([]);
-      toast.error(getStreamErrorMessage(error));
+      const message = getStreamErrorMessage(error);
+      // A "thread not found" 404 almost always means the URL points at a
+      // thread from a previous login session — either this tab outlived a
+      // logout/re-login or the thread genuinely belongs to a different user.
+      // Re-frame the error around the *current* logged-in identity and bounce
+      // the user to a fresh chat instead of leaving them stuck on the stale
+      // route looking at a raw HTTP 404.
+      if (isThreadNotAccessibleError(message)) {
+        const userLabel = buildCurrentUserLabel();
+        toast.info(
+          `当前登录用户为 ${userLabel}，原对话不存在或无权访问，已为你打开新对话。`,
+        );
+        // Prefer staying inside an agent chat context when the user is on an
+        // agent thread URL — otherwise fall back to the global new chat.
+        const hash = typeof window !== "undefined" ? window.location.hash : "";
+        const agentMatch = hash.match(/#\/page\/workspace\/agents\/([^/]+)\//);
+        const target = agentMatch
+          ? `/page/workspace/agents/${agentMatch[1]}/chats/new`
+          : "/page/workspace/chats/new";
+        navigate(target, { replace: true });
+        return;
+      }
+      toast.error(message);
     },
     onFinish(state) {
       listeners.current.onFinish?.(state.values);
@@ -494,7 +555,13 @@ export function useThreadStream({
           }),
         );
 
-        await thread.submit(
+        const ragModeEnabled =
+          localSettings.display.ragModeEnabled || isThreadRagModeLocked(threadId);
+        if (ragModeEnabled) {
+          lockThreadRagMode(threadId);
+        }
+
+        await thread.submit(
           {
             messages: [
               {
@@ -537,10 +604,10 @@ export function useThreadStream({
                       ? "low"
                       : undefined),
               memory_injection_enabled: context.memory_injection_enabled ?? true,
-              rag_mode_enabled: localSettings.display.ragModeEnabled,
-              thread_id: threadId,
-            },
-          },
+              rag_mode_enabled: ragModeEnabled,
+              thread_id: threadId,
+            },
+          },
         );
         void queryClient.invalidateQueries({ queryKey: ["threads", "search"] });
       } catch (error) {
diff --git a/frontend-web/src/styles/globals.css b/frontend-web/src/styles/globals.css
index 3e08cac3..b91dcf12 100644
--- a/frontend-web/src/styles/globals.css
+++ b/frontend-web/src/styles/globals.css
@@ -567,6 +567,20 @@
       transform: translateY(-50%) scale(1);
     }
   }
+
+  @keyframes reference-flash {
+    0% {
+      background: rgba(254, 243, 199, 0.95);
+      box-shadow: 0 0 0 4px rgba(251, 191, 36, 0.28);
+    }
+    70% {
+      background: rgba(254, 243, 199, 0.58);
+      box-shadow: 0 0 0 2px rgba(251, 191, 36, 0.14);
+    }
+    100% {
+      box-shadow: none;
+    }
+  }
 }
 
 /* ─── Tiptap writing editor ─────────────────────────────────────────────── */
@@ -818,4 +832,4 @@
 * {
   border-color: var(--border);
   outline-color: var(--ring);
-}
\ No newline at end of file
+}
diff --git a/offline-backend-20260512/backend/app/gateway/routers/system_settings.py b/offline-backend-20260512/backend/app/gateway/routers/system_settings.py
index aadb877e..b2d8f90c 100644
--- a/offline-backend-20260512/backend/app/gateway/routers/system_settings.py
+++ b/offline-backend-20260512/backend/app/gateway/routers/system_settings.py
@@ -49,6 +49,14 @@ class PrivateSquareUpdateRequest(BaseModel):
     password: str | None = Field(default=None, description="New password in plain text. Pass an empty string to clear.")
 
 
+class CitationDisplayResponse(BaseModel):
+    cleanup_words: list[str] = Field(default_factory=list, description="Global citation cleanup words")
+
+
+class CitationDisplayUpdateRequest(BaseModel):
+    cleanup_words: list[str] | None = None
+
+
 def _to_response(settings: PromptPrefixSettings) -> PromptPrefixResponse:
     return PromptPrefixResponse(
         enabled=settings.enabled,
@@ -83,6 +91,34 @@ async def update_prompt_prefix(body: PromptPrefixUpdateRequest, request: Request
     return _to_response(saved.prompt_prefix)
 
 
+@router.get("/citation-display", response_model=CitationDisplayResponse)
+async def get_citation_display() -> CitationDisplayResponse:
+    """Return global citation display cleanup settings."""
+    settings = load_system_settings().citation_display
+    return CitationDisplayResponse(cleanup_words=settings.cleanup_words)
+
+
+@router.put("/citation-display", response_model=CitationDisplayResponse)
+async def update_citation_display(
+    body: CitationDisplayUpdateRequest,
+    request: Request,
+) -> CitationDisplayResponse:
+    """Update global citation display cleanup settings. Admin only."""
+    await _require_admin(request)
+
+    current = load_system_settings()
+    words = [
+        str(item).strip()
+        for item in (body.cleanup_words or [])
+        if str(item).strip()
+    ]
+    updated = current.citation_display.model_copy(update={"cleanup_words": words})
+    saved = save_system_settings(
+        current.model_copy(update={"citation_display": updated})
+    )
+    return CitationDisplayResponse(cleanup_words=saved.citation_display.cleanup_words)
+
+
 @router.get("/private-square", response_model=PrivateSquareStatusResponse)
 async def get_private_square_status() -> PrivateSquareStatusResponse:
     """Return whether the private square is enabled and a password is set."""
diff --git a/offline-backend-20260512/backend/app/gateway/routers/threads.py b/offline-backend-20260512/backend/app/gateway/routers/threads.py
index 48bf8953..8d3a0e90 100644
--- a/offline-backend-20260512/backend/app/gateway/routers/threads.py
+++ b/offline-backend-20260512/backend/app/gateway/routers/threads.py
@@ -12,9 +12,12 @@ matching the LangGraph Platform wire format expected by the
 
 from __future__ import annotations
 
-import logging
-import uuid
-from typing import Any
+import logging
+import json
+import os
+from pathlib import Path
+import uuid
+from typing import Any
 
 from fastapi import APIRouter, HTTPException, Request
 from langgraph.checkpoint.base import empty_checkpoint
@@ -23,13 +26,18 @@ from pydantic import BaseModel, Field, field_validator
 from app.gateway.authz import require_permission
 from app.gateway.deps import get_checkpointer
 from app.gateway.utils import sanitize_log_param
-from deerflow.config.paths import Paths, get_paths
-from deerflow.runtime import serialize_channel_values
-from deerflow.runtime.user_context import get_effective_user_id
-from deerflow.utils.time import coerce_iso, now_iso
+from deerflow.config.paths import Paths, get_paths
+from deerflow.runtime import serialize_channel_values
+from deerflow.runtime.references import (
+    build_reference_batches as build_runtime_reference_batches,
+    build_reference_debug_summary,
+)
+from deerflow.runtime.user_context import get_effective_user_id
+from deerflow.utils.time import coerce_iso, now_iso
 
-logger = logging.getLogger(__name__)
-router = APIRouter(prefix="/api/threads", tags=["threads"])
+logger = logging.getLogger(__name__)
+router = APIRouter(prefix="/api/threads", tags=["threads"])
+_RAG_DEBUG_LOGGER_NAME = "rag_citation_debug"
 
 
 # Metadata keys that the server controls; clients are not allowed to set
@@ -41,11 +49,38 @@ router = APIRouter(prefix="/api/threads", tags=["threads"])
 _SERVER_RESERVED_METADATA_KEYS: frozenset[str] = frozenset({"owner_id", "user_id"})
 
 
-def _strip_reserved_metadata(metadata: dict[str, Any] | None) -> dict[str, Any]:
-    """Return ``metadata`` with server-controlled keys removed."""
-    if not metadata:
-        return metadata or {}
-    return {k: v for k, v in metadata.items() if k not in _SERVER_RESERVED_METADATA_KEYS}
+def _strip_reserved_metadata(metadata: dict[str, Any] | None) -> dict[str, Any]:
+    """Return ``metadata`` with server-controlled keys removed."""
+    if not metadata:
+        return metadata or {}
+    return {k: v for k, v in metadata.items() if k not in _SERVER_RESERVED_METADATA_KEYS}
+
+
+def _get_rag_debug_logger() -> logging.Logger:
+    debug_logger = logging.getLogger(_RAG_DEBUG_LOGGER_NAME)
+    if getattr(debug_logger, "_rag_file_configured", False):
+        return debug_logger
+    log_path = Path(os.environ.get("RAG_CITATION_DEBUG_LOG", "logs/rag_citation_debug.log"))
+    if not log_path.is_absolute():
+        log_path = Path.cwd() / log_path
+    log_path.parent.mkdir(parents=True, exist_ok=True)
+    handler = logging.FileHandler(log_path, encoding="utf-8")
+    handler.setFormatter(logging.Formatter("%(asctime)s %(levelname)s %(message)s"))
+    debug_logger.addHandler(handler)
+    debug_logger.setLevel(logging.INFO)
+    debug_logger.propagate = False
+    setattr(debug_logger, "_rag_file_configured", True)
+    setattr(debug_logger, "_rag_file_path", str(log_path))
+    return debug_logger
+
+
+def _rag_debug_file_path() -> str:
+    return str(getattr(_get_rag_debug_logger(), "_rag_file_path", "logs/rag_citation_debug.log"))
+
+
+def _log_rag_debug(message: str, *args: Any) -> None:
+    logger.info(message, *args)
+    _get_rag_debug_logger().info(message, *args)
 
 
 # ---------------------------------------------------------------------------
@@ -132,11 +167,43 @@ class HistoryEntry(BaseModel):
     next: list[str] = Field(default_factory=list)
 
 
-class ThreadHistoryRequest(BaseModel):
+class ThreadHistoryRequest(BaseModel):
     """Request body for checkpoint history."""
 
     limit: int = Field(default=10, ge=1, le=100, description="Maximum entries")
-    before: str | None = Field(default=None, description="Cursor for pagination")
+    before: str | None = Field(default=None, description="Cursor for pagination")
+
+
+class ReferenceSourceResponse(BaseModel):
+    index: int
+    id: str | None = None
+    skillName: str = "backend-reference"
+    mode: str = "citation"
+    title: str
+    type: str = "result"
+    snippet: str = ""
+    source: str | None = None
+    time: str | None = None
+    url: str | None = None
+    content: str | None = None
+    score: float | None = None
+    raw: dict[str, Any] = Field(default_factory=dict)
+
+
+class ReferenceBatchResponse(BaseModel):
+    id: str
+    assistant_message_id: str
+    sources: list[ReferenceSourceResponse] = Field(default_factory=list)
+
+
+class ReferenceBatchesResponse(BaseModel):
+    batches: list[ReferenceBatchResponse] = Field(default_factory=list)
+
+
+class ReferenceDebugLogRequest(BaseModel):
+    event: str
+    thread_id: str | None = None
+    payload: dict[str, Any] = Field(default_factory=dict)
 
 
 # ---------------------------------------------------------------------------
@@ -163,7 +230,7 @@ def _delete_thread_data(thread_id: str, paths: Paths | None = None, *, user_id:
     return ThreadDeleteResponse(success=True, message=f"Deleted local thread data for {thread_id}")
 
 
-def _derive_thread_status(checkpoint_tuple) -> str:
+def _derive_thread_status(checkpoint_tuple) -> str:
     """Derive thread status from checkpoint metadata."""
     if checkpoint_tuple is None:
         return "idle"
@@ -179,7 +246,251 @@ def _derive_thread_status(checkpoint_tuple) -> str:
     if tasks:
         return "interrupted"
 
-    return "idle"
+    return "idle"
+
+
+def _message_type(message: Any) -> str:
+    if isinstance(message, dict):
+        return str(message.get("type") or message.get("role") or "")
+    return str(getattr(message, "type", "") or "")
+
+
+def _message_id(message: Any) -> str | None:
+    if isinstance(message, dict):
+        value = message.get("id")
+    else:
+        value = getattr(message, "id", None)
+    return str(value) if value else None
+
+
+def _message_content(message: Any) -> Any:
+    if isinstance(message, dict):
+        return message.get("content")
+    return getattr(message, "content", None)
+
+
+def _message_name(message: Any) -> str:
+    if isinstance(message, dict):
+        return str(message.get("name") or (message.get("additional_kwargs") or {}).get("name") or "tool")
+    return str(getattr(message, "name", None) or "tool")
+
+
+def _has_tool_calls(message: Any) -> bool:
+    if isinstance(message, dict):
+        direct = message.get("tool_calls")
+        chunks = message.get("tool_call_chunks")
+        extra = (message.get("additional_kwargs") or {}).get("tool_calls")
+    else:
+        direct = getattr(message, "tool_calls", None)
+        chunks = getattr(message, "tool_call_chunks", None)
+        extra = (getattr(message, "additional_kwargs", None) or {}).get("tool_calls")
+    return bool(direct or chunks or extra)
+
+
+def _has_answer_content(message: Any) -> bool:
+    content = _message_content(message)
+    if isinstance(content, str):
+        return bool(content.strip())
+    if isinstance(content, list):
+        for part in content:
+            if isinstance(part, dict) and part.get("type") == "text" and str(part.get("text") or "").strip():
+                return True
+    return False
+
+
+def _extract_tool_payload(message: Any) -> Any:
+    content = _message_content(message)
+    if isinstance(content, dict):
+        return _unwrap_tool_payload(content)
+    if isinstance(content, list):
+        texts: list[str] = []
+        for part in content:
+            if not isinstance(part, dict):
+                continue
+            if isinstance(part.get("json"), dict):
+                return _unwrap_tool_payload(part["json"])
+            if part.get("type") == "json" and isinstance(part.get("json"), dict):
+                return _unwrap_tool_payload(part["json"])
+            if part.get("type") == "text" and isinstance(part.get("text"), str):
+                texts.append(part["text"])
+            unwrapped = _unwrap_tool_payload(part)
+            if unwrapped is not None:
+                return unwrapped
+        content = "\n".join(texts)
+    if isinstance(content, str):
+        return _parse_json_like(content)
+    return None
+
+
+def _unwrap_tool_payload(value: dict[str, Any]) -> Any:
+    if any(isinstance(value.get(key), list) for key in ("results", "data", "items", "hits", "documents", "records", "sources")):
+        return value
+    for key in ("result", "output", "stdout", "content", "text", "payload", "response"):
+        nested = value.get(key)
+        if isinstance(nested, str):
+            parsed = _parse_json_like(nested)
+            if parsed is not None:
+                return parsed
+        if isinstance(nested, dict):
+            unwrapped = _unwrap_tool_payload(nested)
+            if unwrapped is not None:
+                return unwrapped
+    return value
+
+
+def _parse_json_like(text: str) -> Any:
+    trimmed = text.strip()
+    if not trimmed:
+        return None
+    try:
+        return json.loads(trimmed)
+    except Exception:
+        pass
+    first_obj = trimmed.find("{")
+    last_obj = trimmed.rfind("}")
+    if 0 <= first_obj < last_obj:
+        try:
+            return json.loads(trimmed[first_obj : last_obj + 1])
+        except Exception:
+            return None
+    return None
+
+
+def _first_list(payload: Any) -> list[Any]:
+    if isinstance(payload, list):
+        return payload
+    if not isinstance(payload, dict):
+        return []
+    for key in ("results", "data", "items", "hits", "documents", "records", "entries", "matches", "chunks", "sources"):
+        value = payload.get(key)
+        if isinstance(value, list):
+            return value
+    for key in ("result", "payload", "response", "output", "data"):
+        nested = payload.get(key)
+        found = _first_list(nested)
+        if found:
+            return found
+    return []
+
+
+def _nested_records(raw: Any) -> list[dict[str, Any]]:
+    if not isinstance(raw, dict):
+        return []
+    layers = [raw]
+    for key in ("metadata", "meta", "attrs", "attributes", "properties", "extra"):
+        nested = raw.get(key)
+        if isinstance(nested, dict):
+            layers.append(nested)
+    return layers
+
+
+def _pick(layers: list[dict[str, Any]], keys: tuple[str, ...]) -> str | None:
+    for layer in layers:
+        for key in keys:
+            value = layer.get(key)
+            if isinstance(value, (str, int, float)) and str(value).strip():
+                return str(value).strip()
+    return None
+
+
+def _normalize_reference_item(raw: Any, fallback_index: int) -> dict[str, Any] | None:
+    if not isinstance(raw, dict):
+        return None
+    layers = _nested_records(raw)
+    title = _pick(layers, ("title", "m_title", "name", "question", "keyword", "query", "text"))
+    content = _pick(layers, ("content", "page_content", "summary", "snippet", "description", "content_preview", "abstract", "body"))
+    url = _pick(layers, ("url", "link", "href", "source_url", "web_url"))
+    source = _pick(layers, ("source", "source1", "site", "provider", "from"))
+    time = _pick(layers, ("time", "date", "published_date", "publishedDate", "m_publish", "publish_time", "created_at"))
+    item_id = _pick(layers, ("recUuid", "rec_uuid", "uuid", "docId", "doc_id", "id", "recordId", "taskId"))
+    score_raw = _pick(layers, ("score", "relevance", "similarity"))
+    score = None
+    if score_raw is not None:
+        try:
+            score = float(score_raw)
+        except Exception:
+            score = None
+    if not title:
+        title = source or f"结果 {fallback_index}"
+    if not content:
+        content = ""
+    if not title and not content and not url:
+        return None
+    return {
+        "id": item_id or url,
+        "title": title,
+        "type": "result",
+        "snippet": content,
+        "content": content,
+        "source": source,
+        "time": time,
+        "url": url,
+        "score": score,
+        "raw": raw,
+    }
+
+
+def _reference_key(source: dict[str, Any]) -> str:
+    raw = source.get("raw") if isinstance(source.get("raw"), dict) else {}
+    item_id = (
+        source.get("id")
+        or source.get("url")
+        or raw.get("recUuid")
+        or raw.get("uuid")
+        or raw.get("docId")
+        or raw.get("id")
+    )
+    if item_id:
+        return str(item_id).casefold().strip()
+    return "|".join(
+        str(source.get(key) or "").casefold().strip()
+        for key in ("title", "source", "time", "content", "snippet")
+    )
+
+
+def _build_reference_batches(messages: list[Any]) -> list[ReferenceBatchResponse]:
+    batches: list[ReferenceBatchResponse] = []
+    for index, message in enumerate(messages):
+        if _message_type(message) != "human":
+            continue
+        end = len(messages)
+        for cursor in range(index + 1, len(messages)):
+            if _message_type(messages[cursor]) == "human":
+                end = cursor
+                break
+        assistant_message: Any | None = None
+        for cursor in range(end - 1, index, -1):
+            candidate = messages[cursor]
+            if (
+                _message_type(candidate) == "ai"
+                and _message_id(candidate)
+                and _has_answer_content(candidate)
+                and not _has_tool_calls(candidate)
+            ):
+                assistant_message = candidate
+                break
+        if assistant_message is None:
+            continue
+        seen: set[str] = set()
+        sources: list[ReferenceSourceResponse] = []
+        for tool_message in messages[index + 1 : end]:
+            if _message_type(tool_message) != "tool":
+                continue
+            payload = _extract_tool_payload(tool_message)
+            for raw_item in _first_list(payload):
+                normalized = _normalize_reference_item(raw_item, len(sources) + 1)
+                if not normalized:
+                    continue
+                key = _reference_key(normalized)
+                if key and key in seen:
+                    continue
+                if key:
+                    seen.add(key)
+                sources.append(ReferenceSourceResponse(index=len(sources) + 1, **normalized))
+        if sources:
+            mid = _message_id(assistant_message) or str(index)
+            batches.append(ReferenceBatchResponse(id=mid, assistant_message_id=mid, sources=sources))
+    return batches
 
 
 # ---------------------------------------------------------------------------
@@ -406,9 +717,9 @@ async def get_thread(thread_id: str, request: Request) -> ThreadResponse:
 
 
 # ---------------------------------------------------------------------------
-@router.get("/{thread_id}/state", response_model=ThreadStateResponse)
-@require_permission("threads", "read", owner_check=True)
-async def get_thread_state(thread_id: str, request: Request) -> ThreadStateResponse:
+@router.get("/{thread_id}/state", response_model=ThreadStateResponse)
+@require_permission("threads", "read", owner_check=True)
+async def get_thread_state(thread_id: str, request: Request) -> ThreadStateResponse:
     """Get the latest state snapshot for a thread.
 
     Channel values are serialized to ensure LangChain message objects
@@ -455,12 +766,84 @@ async def get_thread_state(thread_id: str, request: Request) -> ThreadStateRespo
         parent_checkpoint_id=parent_checkpoint_id,
         created_at=coerce_iso(metadata.get("created_at", "")),
         tasks=tasks,
-    )
-
-
-@router.post("/{thread_id}/state", response_model=ThreadStateResponse)
-@require_permission("threads", "write", owner_check=True, require_existing=True)
-async def update_thread_state(thread_id: str, body: ThreadStateUpdateRequest, request: Request) -> ThreadStateResponse:
+    )
+
+
+@router.get("/{thread_id}/reference-batches", response_model=ReferenceBatchesResponse)
+@require_permission("threads", "read", owner_check=True)
+async def get_thread_reference_batches(thread_id: str, request: Request) -> ReferenceBatchesResponse:
+    """Return backend-built citation batches for completed assistant turns."""
+    checkpointer = get_checkpointer(request)
+    config = {"configurable": {"thread_id": thread_id, "checkpoint_ns": ""}}
+    try:
+        checkpoint_tuple = await checkpointer.aget_tuple(config)
+    except Exception:
+        logger.exception("Failed to get reference batches for thread %s", sanitize_log_param(thread_id))
+        raise HTTPException(status_code=500, detail="Failed to get reference batches")
+
+    if checkpoint_tuple is None:
+        raise HTTPException(status_code=404, detail=f"Thread {thread_id} not found")
+
+    checkpoint = getattr(checkpoint_tuple, "checkpoint", {}) or {}
+    values = serialize_channel_values(checkpoint.get("channel_values", {}) or {})
+    messages = values.get("messages")
+    if not isinstance(messages, list):
+        _log_rag_debug(
+            "backend reference-batches thread=%s messages=missing batches=0 log_file=%s",
+            sanitize_log_param(thread_id),
+            _rag_debug_file_path(),
+        )
+        return ReferenceBatchesResponse(batches=[])
+    raw_batches = build_runtime_reference_batches(messages)
+    debug_summary = build_reference_debug_summary(messages, raw_batches)
+    _log_rag_debug(
+        "backend reference-batches thread=%s messages=%s turns=%s tool_messages=%s raw_items=%s "
+        "normalized_items=%s kept_items=%s batches=%s batch_sources=%s batch_ids=%s turns_detail=%s log_file=%s",
+        sanitize_log_param(thread_id),
+        debug_summary["messages"],
+        debug_summary["turns"],
+        debug_summary["tool_messages"],
+        debug_summary["raw_items"],
+        debug_summary["normalized_items"],
+        debug_summary["kept_items"],
+        len(raw_batches),
+        ",".join(str(batch.get("source_count", 0)) for batch in debug_summary["batches"]),
+        ",".join(str(batch.get("id") or "") for batch in debug_summary["batches"]),
+        json.dumps(debug_summary["turns_detail"], ensure_ascii=False),
+        _rag_debug_file_path(),
+    )
+    return ReferenceBatchesResponse(
+        batches=[
+            ReferenceBatchResponse(
+                id=str(batch.get("id") or ""),
+                assistant_message_id=str(batch.get("assistant_message_id") or batch.get("id") or ""),
+                sources=[
+                    ReferenceSourceResponse(**source)
+                    for source in batch.get("sources", [])
+                    if isinstance(source, dict)
+                ],
+            )
+            for batch in raw_batches
+        ]
+    )
+
+
+@router.post("/reference-debug-log")
+async def write_reference_debug_log(body: ReferenceDebugLogRequest) -> dict[str, str | bool]:
+    """Append frontend citation diagnostics to the backend RAG debug log."""
+    _get_rag_debug_logger().info(
+        "frontend %s thread=%s payload=%s log_file=%s",
+        body.event,
+        sanitize_log_param(body.thread_id or ""),
+        json.dumps(body.payload, ensure_ascii=False, default=str),
+        _rag_debug_file_path(),
+    )
+    return {"ok": True, "log_file": _rag_debug_file_path()}
+
+
+@router.post("/{thread_id}/state", response_model=ThreadStateResponse)
+@require_permission("threads", "write", owner_check=True, require_existing=True)
+async def update_thread_state(thread_id: str, body: ThreadStateUpdateRequest, request: Request) -> ThreadStateResponse:
     """Update thread state (e.g. for human-in-the-loop resume or title rename).
 
     Writes a new checkpoint that merges *body.values* into the latest
diff --git a/offline-backend-20260512/backend/app/gateway/services.py b/offline-backend-20260512/backend/app/gateway/services.py
index e29a3c93..46965a2c 100644
--- a/offline-backend-20260512/backend/app/gateway/services.py
+++ b/offline-backend-20260512/backend/app/gateway/services.py
@@ -100,7 +100,14 @@ _DEFAULT_ASSISTANT_ID = "lead_agent"
 
 # Built-in LangGraph graph IDs registered in langgraph.json.
 # These are NOT stored in the agents DB and must bypass the custom-agent access check.
-_BUILTIN_GRAPH_IDS: frozenset[str] = frozenset({"lead_agent", "canvas_agent", "ai_writing"})
+_BUILTIN_GRAPH_IDS: frozenset[str] = frozenset({"lead_agent", "canvas_agent", "ai_writing"})
+
+_RAG_CITATION_PROMPT = """当前对话已开启参考文献模式。
+如果本轮调用了搜索、知识库、检索类技能或工具，请基于本轮所有检索结果整体回答。
+你必须把同一轮内多次检索得到的资料视为同一批参考文献，按出现顺序从 [1] 开始连续编号，不要让每个检索词各自从 [1] 重新编号。
+回答正文中的关键结论必须标注对应编号，例如 [1]、[2]，多个来源可写为 [1][3][5]。
+不要引用不存在的编号，不要使用技能内部自带的编号，不要把参考文献完整列表堆在正文末尾。
+如果没有任何可用检索结果，请直接说明未检索到可靠参考资料。"""
 
 
 def _prepend_prompt_prefix(graph_input: dict[str, Any], prefix: str) -> dict[str, Any]:
@@ -129,7 +136,7 @@ def _prepend_prompt_prefix(graph_input: dict[str, Any], prefix: str) -> dict[str
     return {**graph_input, "messages": new_messages}
 
 
-def _prompt_prefix_opted_in(body: Any) -> bool:
+def _prompt_prefix_opted_in(body: Any) -> bool:
     """Detect whether the caller asked for the prompt prefix to be applied."""
     context = getattr(body, "context", None)
     if isinstance(context, Mapping) and context.get("prompt_prefix_enabled"):
@@ -137,7 +144,44 @@ def _prompt_prefix_opted_in(body: Any) -> bool:
     metadata = getattr(body, "metadata", None)
     if isinstance(metadata, Mapping) and metadata.get("prompt_prefix_enabled"):
         return True
-    return False
+    return False
+
+
+def _rag_mode_requested(body: Any) -> bool:
+    context = getattr(body, "context", None)
+    if isinstance(context, Mapping) and context.get("rag_mode_enabled"):
+        return True
+    metadata = getattr(body, "metadata", None)
+    if isinstance(metadata, Mapping) and (
+        metadata.get("rag_mode_enabled") or metadata.get("rag_mode")
+    ):
+        return True
+    return False
+
+
+def _thread_rag_mode_enabled(record: Any) -> bool:
+    if not isinstance(record, Mapping):
+        return False
+    metadata = record.get("metadata")
+    return isinstance(metadata, Mapping) and bool(
+        metadata.get("rag_mode_enabled") or metadata.get("rag_mode")
+    )
+
+
+def _append_invisible_prompt(graph_input: dict[str, Any], prompt: str) -> dict[str, Any]:
+    messages = graph_input.get("messages")
+    if not isinstance(messages, list) or not messages:
+        return graph_input
+    new_messages = list(messages)
+    for idx, msg in enumerate(new_messages):
+        if not isinstance(msg, HumanMessage):
+            continue
+        extra = dict(getattr(msg, "additional_kwargs", None) or {})
+        existing = str(extra.get("prompt_prefix") or "").strip()
+        extra["prompt_prefix"] = f"{existing}\n\n{prompt}".strip() if existing else prompt
+        new_messages[idx] = msg.model_copy(update={"additional_kwargs": extra})
+        break
+    return {**graph_input, "messages": new_messages}
 
 
 # Whitelist of run-context keys that the langgraph-compat layer forwards from
@@ -399,16 +443,29 @@ async def start_run(
     # Upsert thread metadata so the thread appears in /threads/search,
     # even for threads that were never explicitly created via POST /threads
     # (e.g. stateless runs).
-    try:
-        existing = await run_ctx.thread_store.get(thread_id)
-        if existing is None:
-            await run_ctx.thread_store.create(
-                thread_id,
-                assistant_id=body.assistant_id,
-                metadata=body.metadata,
-            )
-        else:
-            await run_ctx.thread_store.update_status(thread_id, "running")
+    requested_rag_mode = _rag_mode_requested(body)
+    effective_rag_mode = requested_rag_mode
+    try:
+        existing = await run_ctx.thread_store.get(thread_id)
+        effective_rag_mode = requested_rag_mode or _thread_rag_mode_enabled(existing)
+        if existing is None:
+            metadata = dict(body.metadata or {})
+            if effective_rag_mode:
+                metadata["rag_mode_enabled"] = True
+                metadata["rag_mode"] = True
+            await run_ctx.thread_store.create(
+                thread_id,
+                assistant_id=body.assistant_id,
+                metadata=metadata,
+            )
+        else:
+            await run_ctx.thread_store.update_status(thread_id, "running")
+            if requested_rag_mode and not _thread_rag_mode_enabled(existing):
+                await run_ctx.thread_store.update_metadata(
+                    thread_id,
+                    {"rag_mode_enabled": True, "rag_mode": True},
+                )
+                effective_rag_mode = True
             # SDK 的 ``client.threads.create()`` 不传 assistant_id —— 先建出来空的 thread，
             # 真正提交 run 时才知道用哪个 assistant。这里回填一次，避免历史列表 / 按
             # assistant 筛选时拿不到关联。
@@ -428,17 +485,21 @@ async def start_run(
     # applies once. We intentionally do NOT gate on thread-row existence: the
     # thread_meta row is frequently pre-created, which previously skipped the
     # prefix for genuine first turns.
-    if _prompt_prefix_opted_in(body):
-        prefix_settings = load_system_settings().prompt_prefix
-        if prefix_settings.enabled and prefix_settings.prompt_prefix.strip():
-            graph_input = _prepend_prompt_prefix(graph_input, prefix_settings.prompt_prefix.strip())
-    config = build_run_config(thread_id, body.config, body.metadata, assistant_id=body.assistant_id)
+    if _prompt_prefix_opted_in(body):
+        prefix_settings = load_system_settings().prompt_prefix
+        if prefix_settings.enabled and prefix_settings.prompt_prefix.strip():
+            graph_input = _prepend_prompt_prefix(graph_input, prefix_settings.prompt_prefix.strip())
+    if effective_rag_mode:
+        graph_input = _append_invisible_prompt(graph_input, _RAG_CITATION_PROMPT)
+    config = build_run_config(thread_id, body.config, body.metadata, assistant_id=body.assistant_id)
 
     # Merge DeerFlow-specific context overrides into both ``configurable`` and ``context``.
     # The ``context`` field is a custom extension for the langgraph-compat layer
     # that carries agent configuration (model_name, thinking_enabled, etc.).
     # Only agent-relevant keys are forwarded; unknown keys (e.g. thread_id) are ignored.
-    body_context = dict(getattr(body, "context", None) or {})
+    body_context = dict(getattr(body, "context", None) or {})
+    if effective_rag_mode:
+        body_context["rag_mode_enabled"] = True
     # Promote whitelisted top-level body fields into the context map so
     # callers can pass them either nested under ``context`` or as flat
     # top-level fields (``excluded_tools``, ``skill_stop_names``, …).
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/agents/lead_agent/agent.py b/offline-backend-20260512/backend/packages/harness/deerflow/agents/lead_agent/agent.py
index d567ce52..4b65a644 100644
--- a/offline-backend-20260512/backend/packages/harness/deerflow/agents/lead_agent/agent.py
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/agents/lead_agent/agent.py
@@ -7,9 +7,10 @@ from langchain_core.runnables import RunnableConfig
 from deerflow.agents.lead_agent.prompt import apply_prompt_template
 from deerflow.agents.middlewares.clarification_middleware import ClarificationMiddleware
 from deerflow.agents.middlewares.loop_detection_middleware import LoopDetectionMiddleware
-from deerflow.agents.middlewares.memory_middleware import MemoryMiddleware
-from deerflow.agents.middlewares.prompt_prefix_middleware import PromptPrefixMiddleware
-from deerflow.agents.middlewares.setup_agent_approval_middleware import SetupAgentApprovalMiddleware
+from deerflow.agents.middlewares.memory_middleware import MemoryMiddleware
+from deerflow.agents.middlewares.prompt_prefix_middleware import PromptPrefixMiddleware
+from deerflow.agents.middlewares.rag_citation_middleware import RagCitationMiddleware
+from deerflow.agents.middlewares.setup_agent_approval_middleware import SetupAgentApprovalMiddleware
 from deerflow.agents.middlewares.subagent_limit_middleware import SubagentLimitMiddleware
 from deerflow.agents.middlewares.summarization_middleware import BeforeSummarizationHook, DeerFlowSummarizationMiddleware
 from deerflow.agents.middlewares.title_middleware import TitleMiddleware
@@ -252,13 +253,16 @@ def _build_middlewares(
 
     Returns:
         List of middleware instances.
-    """
-    resolved_app_config = app_config or get_app_config()
-    middlewares = build_lead_runtime_middlewares(app_config=resolved_app_config, lazy_init=True)
+    """
+    resolved_app_config = app_config or get_app_config()
+    cfg = _get_runtime_config(config)
+    middlewares = build_lead_runtime_middlewares(app_config=resolved_app_config, lazy_init=True)
 
-    # Inject the admin prompt prefix into the model request (kept out of the
-    # persisted/displayed message so the user never sees it).
-    middlewares.append(PromptPrefixMiddleware())
+    # Inject the admin prompt prefix into the model request (kept out of the
+    # persisted/displayed message so the user never sees it).
+    middlewares.append(PromptPrefixMiddleware())
+    if cfg.get("rag_mode_enabled", False):
+        middlewares.append(RagCitationMiddleware())
 
     # Add summarization middleware if enabled
     summarization_middleware = _create_summarization_middleware(app_config=resolved_app_config)
@@ -266,8 +270,7 @@ def _build_middlewares(
         middlewares.append(summarization_middleware)
 
     # Add TodoList middleware if plan mode is enabled
-    cfg = _get_runtime_config(config)
-    is_plan_mode = cfg.get("is_plan_mode", False)
+    is_plan_mode = cfg.get("is_plan_mode", False)
     is_bootstrap = cfg.get("is_bootstrap", False)
     todo_list_middleware = _create_todo_list_middleware(is_plan_mode)
     if todo_list_middleware is not None:
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/agents/middlewares/rag_citation_middleware.py b/offline-backend-20260512/backend/packages/harness/deerflow/agents/middlewares/rag_citation_middleware.py
new file mode 100644
index 00000000..0801f047
--- /dev/null
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/agents/middlewares/rag_citation_middleware.py
@@ -0,0 +1,109 @@
+"""Inject stable citation numbering before RAG final answers."""
+
+from __future__ import annotations
+
+import logging
+import os
+from pathlib import Path
+from collections.abc import Awaitable, Callable
+from typing import Any, override
+
+from langchain.agents import AgentState
+from langchain.agents.middleware import AgentMiddleware
+from langchain.agents.middleware.types import ModelCallResult, ModelRequest, ModelResponse
+from langchain_core.messages import SystemMessage
+
+from deerflow.runtime.references import build_latest_reference_prompt, build_reference_debug_summary
+
+logger = logging.getLogger(__name__)
+_RAG_DEBUG_LOGGER_NAME = "rag_citation_debug"
+
+
+def _get_rag_debug_logger() -> logging.Logger:
+    debug_logger = logging.getLogger(_RAG_DEBUG_LOGGER_NAME)
+    if getattr(debug_logger, "_rag_file_configured", False):
+        return debug_logger
+    log_path = Path(os.environ.get("RAG_CITATION_DEBUG_LOG", "logs/rag_citation_debug.log"))
+    if not log_path.is_absolute():
+        log_path = Path.cwd() / log_path
+    log_path.parent.mkdir(parents=True, exist_ok=True)
+    handler = logging.FileHandler(log_path, encoding="utf-8")
+    handler.setFormatter(logging.Formatter("%(asctime)s %(levelname)s %(message)s"))
+    debug_logger.addHandler(handler)
+    debug_logger.setLevel(logging.INFO)
+    debug_logger.propagate = False
+    setattr(debug_logger, "_rag_file_configured", True)
+    setattr(debug_logger, "_rag_file_path", str(log_path))
+    return debug_logger
+
+
+def _log_rag_debug(message: str, *args: Any) -> None:
+    logger.info(message, *args)
+    _get_rag_debug_logger().info(message, *args)
+
+
+class RagCitationMiddleware(AgentMiddleware[AgentState]):
+    """Give the model the same numbered references the frontend will render."""
+
+    @staticmethod
+    def _enabled(request: ModelRequest) -> bool:
+        runtime = getattr(request, "runtime", None)
+        context = getattr(runtime, "context", None) if runtime is not None else None
+        if isinstance(context, dict) and context.get("rag_mode_enabled"):
+            return True
+        config = getattr(runtime, "config", None) if runtime is not None else None
+        if isinstance(config, dict):
+            configurable = config.get("configurable")
+            if isinstance(configurable, dict) and configurable.get("rag_mode_enabled"):
+                return True
+            runtime_context = config.get("context")
+            if isinstance(runtime_context, dict) and runtime_context.get("rag_mode_enabled"):
+                return True
+        return False
+
+    @staticmethod
+    def _patch(request: ModelRequest) -> ModelRequest | None:
+        if not RagCitationMiddleware._enabled(request):
+            return None
+        messages = list(request.messages)
+        prompt, batch = build_latest_reference_prompt(messages)
+        if not prompt:
+            _log_rag_debug(
+                "middleware reference-prompt enabled=true prompt=false messages=%s",
+                len(messages),
+            )
+            return None
+        summary = build_reference_debug_summary(messages, [batch] if batch else None)
+        _log_rag_debug(
+            "middleware reference-prompt enabled=true prompt=true messages=%s tool_messages=%s "
+            "raw_items=%s normalized_items=%s kept_items=%s latest_batch_sources=%s",
+            summary["messages"],
+            summary["tool_messages"],
+            summary["raw_items"],
+            summary["normalized_items"],
+            summary["kept_items"],
+            len((batch or {}).get("sources") or []),
+        )
+        patched_messages = [
+            *messages,
+            SystemMessage(content=prompt),
+        ]
+        return request.override(messages=patched_messages)
+
+    @override
+    def wrap_model_call(
+        self,
+        request: ModelRequest,
+        handler: Callable[[ModelRequest], ModelResponse],
+    ) -> ModelCallResult:
+        patched = self._patch(request)
+        return handler(patched or request)
+
+    @override
+    async def awrap_model_call(
+        self,
+        request: ModelRequest,
+        handler: Callable[[ModelRequest], Awaitable[ModelResponse]],
+    ) -> ModelCallResult:
+        patched = self._patch(request)
+        return await handler(patched or request)
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/config/system_settings.py b/offline-backend-20260512/backend/packages/harness/deerflow/config/system_settings.py
index 67e1a4d9..943f6ba0 100644
--- a/offline-backend-20260512/backend/packages/harness/deerflow/config/system_settings.py
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/config/system_settings.py
@@ -54,6 +54,16 @@ class AppearanceSettings(BaseModel):
     model_config = ConfigDict(extra="ignore")
 
 
+class CitationDisplaySettings(BaseModel):
+    """System-wide cleanup rules for citation cards and detail dialogs."""
+
+    cleanup_words: list[str] = Field(
+        default_factory=list,
+        description="Words or phrases used to remove noisy advertisement/disclaimer lines from citation display.",
+    )
+    model_config = ConfigDict(extra="ignore")
+
+
 class PrivateSquareSettings(BaseModel):
     """Admin-configurable settings for the agent private-square (私密广场).
 
@@ -74,6 +84,7 @@ class SystemSettings(BaseModel):
     prompt_prefix: PromptPrefixSettings = Field(default_factory=PromptPrefixSettings)
     private_square: PrivateSquareSettings = Field(default_factory=PrivateSquareSettings)
     appearance: AppearanceSettings = Field(default_factory=AppearanceSettings)
+    citation_display: CitationDisplaySettings = Field(default_factory=CitationDisplaySettings)
     model_config = ConfigDict(extra="ignore")
 
 
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/runtime/references.py b/offline-backend-20260512/backend/packages/harness/deerflow/runtime/references.py
new file mode 100644
index 00000000..9d9782f8
--- /dev/null
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/runtime/references.py
@@ -0,0 +1,366 @@
+"""Build stable per-turn citation batches from LangGraph messages."""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+
+
+def build_reference_batches(messages: list[Any]) -> list[dict[str, Any]]:
+    batches: list[dict[str, Any]] = []
+    for index, message in enumerate(messages):
+        if _message_type(message) != "human":
+            continue
+        end = len(messages)
+        for cursor in range(index + 1, len(messages)):
+            if _message_type(messages[cursor]) == "human":
+                end = cursor
+                break
+        assistant = _find_last_assistant_answer(messages, index + 1, end)
+        if assistant is None:
+            continue
+        sources = _collect_sources(messages[index + 1 : end])
+        if sources:
+            message_id = _message_id(assistant) or str(index)
+            batches.append(
+                {
+                    "id": message_id,
+                    "assistant_message_id": message_id,
+                    "sources": sources,
+                }
+            )
+    return batches
+
+
+def build_latest_reference_prompt(messages: list[Any]) -> tuple[str, dict[str, Any] | None]:
+    human_index = _find_latest_human_index(messages)
+    if human_index is None:
+        return "", None
+    sources = _collect_sources(messages[human_index + 1 :])
+    if not sources:
+        return "", None
+    assistant = _find_last_assistant_answer(messages, human_index + 1, len(messages))
+    message_id = _message_id(assistant) or _message_id(messages[human_index]) or str(human_index)
+    batch = {
+        "id": message_id,
+        "assistant_message_id": message_id,
+        "sources": sources,
+    }
+    lines = [
+        "本轮检索结果已由系统统一合并并重新编号。最终回答必须使用下面这套编号。",
+        "如果你仍需要补充检索，请先继续调用工具；最终回答只能引用最新合并后的编号。",
+        "",
+    ]
+    for source in sources:
+        title = source.get("title") or f"参考文献 {source.get('index')}"
+        meta = " · ".join(
+            str(item)
+            for item in (source.get("source"), source.get("time"))
+            if item
+        )
+        content = str(source.get("content") or source.get("snippet") or "").strip()
+        if len(content) > 520:
+            content = content[:517] + "..."
+        head = f"[{source.get('index')}] {title}"
+        if meta:
+            head += f"（{meta}）"
+        lines.append(head)
+        if content:
+            lines.append(content)
+        lines.append("")
+    lines.extend(
+        [
+            "引用规则：",
+            "1. 回答正文中的关键结论必须标注对应编号，例如 [1]、[2]。",
+            "2. 多个来源可写为 [1][3][5]。",
+            "3. 只能引用上方存在的编号，不要编造编号。",
+            "4. 不要使用工具原始编号，也不要重新排序编号。",
+            "5. 不要在正文末尾堆完整参考文献列表，前端会展示右侧参考文献。",
+        ]
+    )
+    return "\n".join(lines).strip(), batch
+
+
+def _find_latest_human_index(messages: list[Any]) -> int | None:
+    for index in range(len(messages) - 1, -1, -1):
+        if _message_type(messages[index]) == "human":
+            return index
+    return None
+
+
+def build_reference_debug_summary(messages: list[Any], batches: list[dict[str, Any]] | None = None) -> dict[str, Any]:
+    """Return compact counts for citation extraction logs."""
+    turn_summaries: list[dict[str, Any]] = []
+    total_tool_messages = 0
+    total_raw_items = 0
+    total_normalized_items = 0
+    total_kept_items = 0
+    for index, message in enumerate(messages):
+        if _message_type(message) != "human":
+            continue
+        end = len(messages)
+        for cursor in range(index + 1, len(messages)):
+            if _message_type(messages[cursor]) == "human":
+                end = cursor
+                break
+        tool_messages = 0
+        raw_items = 0
+        normalized_items = 0
+        kept_items = 0
+        for tool_message in messages[index + 1 : end]:
+            if _message_type(tool_message) != "tool":
+                continue
+            tool_messages += 1
+            items = _first_list(_extract_tool_payload(tool_message))
+            raw_items += len(items)
+            for raw_item in items:
+                normalized = _normalize_reference_item(raw_item, normalized_items + 1)
+                if not normalized:
+                    continue
+                normalized_items += 1
+                kept_items += 1
+        total_tool_messages += tool_messages
+        total_raw_items += raw_items
+        total_normalized_items += normalized_items
+        total_kept_items += kept_items
+        turn_summaries.append(
+            {
+                "human_index": index,
+                "tool_messages": tool_messages,
+                "raw_items": raw_items,
+                "normalized_items": normalized_items,
+                "kept_items": kept_items,
+            }
+        )
+    batch_summaries = [
+        {
+            "id": batch.get("id"),
+            "assistant_message_id": batch.get("assistant_message_id"),
+            "source_count": len(batch.get("sources") or []),
+        }
+        for batch in (batches or build_reference_batches(messages))
+    ]
+    return {
+        "messages": len(messages),
+        "turns": len(turn_summaries),
+        "tool_messages": total_tool_messages,
+        "raw_items": total_raw_items,
+        "normalized_items": total_normalized_items,
+        "kept_items": total_kept_items,
+        "turns_detail": turn_summaries,
+        "batches": batch_summaries,
+    }
+
+
+def _collect_sources(messages: list[Any]) -> list[dict[str, Any]]:
+    sources: list[dict[str, Any]] = []
+    for message in messages:
+        if _message_type(message) != "tool":
+            continue
+        payload = _extract_tool_payload(message)
+        for raw_item in _first_list(payload):
+            normalized = _normalize_reference_item(raw_item, len(sources) + 1)
+            if not normalized:
+                continue
+            normalized["index"] = len(sources) + 1
+            sources.append(normalized)
+    return sources
+
+
+def _find_last_assistant_answer(messages: list[Any], start: int, end: int) -> Any | None:
+    for index in range(end - 1, start - 1, -1):
+        candidate = messages[index]
+        if (
+            _message_type(candidate) == "ai"
+            and _message_id(candidate)
+            and _has_answer_content(candidate)
+            and not _has_tool_calls(candidate)
+        ):
+            return candidate
+    return None
+
+
+def _message_type(message: Any) -> str:
+    if isinstance(message, dict):
+        return str(message.get("type") or message.get("role") or "")
+    return str(getattr(message, "type", "") or "")
+
+
+def _message_id(message: Any) -> str | None:
+    value = message.get("id") if isinstance(message, dict) else getattr(message, "id", None)
+    return str(value) if value else None
+
+
+def _message_content(message: Any) -> Any:
+    return message.get("content") if isinstance(message, dict) else getattr(message, "content", None)
+
+
+def _has_tool_calls(message: Any) -> bool:
+    if isinstance(message, dict):
+        direct = message.get("tool_calls")
+        chunks = message.get("tool_call_chunks")
+        extra = (message.get("additional_kwargs") or {}).get("tool_calls")
+    else:
+        direct = getattr(message, "tool_calls", None)
+        chunks = getattr(message, "tool_call_chunks", None)
+        extra = (getattr(message, "additional_kwargs", None) or {}).get("tool_calls")
+    return bool(direct or chunks or extra)
+
+
+def _has_answer_content(message: Any) -> bool:
+    content = _message_content(message)
+    if isinstance(content, str):
+        return bool(content.strip())
+    if isinstance(content, list):
+        return any(
+            isinstance(part, dict)
+            and part.get("type") == "text"
+            and str(part.get("text") or "").strip()
+            for part in content
+        )
+    return False
+
+
+def _extract_tool_payload(message: Any) -> Any:
+    content = _message_content(message)
+    if isinstance(content, dict):
+        return _unwrap_tool_payload(content)
+    if isinstance(content, list):
+        texts: list[str] = []
+        for part in content:
+            if not isinstance(part, dict):
+                continue
+            if isinstance(part.get("json"), dict):
+                return _unwrap_tool_payload(part["json"])
+            if part.get("type") == "json" and isinstance(part.get("json"), dict):
+                return _unwrap_tool_payload(part["json"])
+            if part.get("type") == "text" and isinstance(part.get("text"), str):
+                texts.append(part["text"])
+            unwrapped = _unwrap_tool_payload(part)
+            if unwrapped is not None:
+                return unwrapped
+        content = "\n".join(texts)
+    if isinstance(content, str):
+        return _parse_json_like(content)
+    return None
+
+
+def _unwrap_tool_payload(value: dict[str, Any]) -> Any:
+    list_keys = ("results", "data", "items", "hits", "documents", "records", "sources")
+    if any(isinstance(value.get(key), list) for key in list_keys):
+        return value
+    for key in ("result", "output", "stdout", "content", "text", "payload", "response"):
+        nested = value.get(key)
+        if isinstance(nested, str):
+            parsed = _parse_json_like(nested)
+            if parsed is not None:
+                return parsed
+        if isinstance(nested, dict):
+            unwrapped = _unwrap_tool_payload(nested)
+            if unwrapped is not None:
+                return unwrapped
+    return value
+
+
+def _parse_json_like(text: str) -> Any:
+    trimmed = text.strip()
+    if not trimmed:
+        return None
+    try:
+        return json.loads(trimmed)
+    except Exception:
+        pass
+    first = trimmed.find("{")
+    last = trimmed.rfind("}")
+    if 0 <= first < last:
+        try:
+            return json.loads(trimmed[first : last + 1])
+        except Exception:
+            return None
+    return None
+
+
+def _first_list(payload: Any) -> list[Any]:
+    if isinstance(payload, list):
+        return payload
+    if not isinstance(payload, dict):
+        return []
+    for key in ("results", "data", "items", "hits", "documents", "records", "entries", "matches", "chunks", "sources"):
+        value = payload.get(key)
+        if isinstance(value, list):
+            return value
+    for key in ("result", "payload", "response", "output", "data"):
+        found = _first_list(payload.get(key))
+        if found:
+            return found
+    return []
+
+
+def _nested_records(raw: Any) -> list[dict[str, Any]]:
+    if not isinstance(raw, dict):
+        return []
+    layers = [raw]
+    for key in ("metadata", "meta", "attrs", "attributes", "properties", "extra"):
+        nested = raw.get(key)
+        if isinstance(nested, dict):
+            layers.append(nested)
+    return layers
+
+
+def _pick(layers: list[dict[str, Any]], keys: tuple[str, ...]) -> str | None:
+    for layer in layers:
+        for key in keys:
+            value = layer.get(key)
+            if isinstance(value, (str, int, float)) and str(value).strip():
+                return str(value).strip()
+    return None
+
+
+def _normalize_reference_item(raw: Any, fallback_index: int) -> dict[str, Any] | None:
+    if not isinstance(raw, dict):
+        return None
+    layers = _nested_records(raw)
+    title = _pick(layers, ("title", "m_title", "name", "question", "keyword", "query", "text"))
+    content = _pick(layers, ("content", "page_content", "summary", "snippet", "description", "content_preview", "abstract", "body"))
+    url = _pick(layers, ("url", "link", "href", "source_url", "web_url"))
+    source = _pick(layers, ("source", "source1", "site", "provider", "from"))
+    time = _pick(layers, ("time", "date", "published_date", "publishedDate", "m_publish", "publish_time", "created_at"))
+    item_id = _pick(layers, ("recUuid", "rec_uuid", "uuid", "docId", "doc_id", "id", "recordId", "taskId"))
+    score_raw = _pick(layers, ("score", "relevance", "similarity"))
+    score = None
+    if score_raw is not None:
+        try:
+            score = float(score_raw)
+        except Exception:
+            score = None
+    if not title:
+        title = source or f"结果 {fallback_index}"
+    if not content:
+        content = ""
+    if not title and not content and not url:
+        return None
+    return {
+        "id": item_id or url,
+        "skillName": "backend-reference",
+        "mode": "citation",
+        "title": title,
+        "type": "result",
+        "snippet": content,
+        "content": content,
+        "source": source,
+        "time": time,
+        "url": url,
+        "score": score,
+        "raw": raw,
+    }
+
+
+def _reference_key(source: dict[str, Any]) -> str:
+    raw = source.get("raw") if isinstance(source.get("raw"), dict) else {}
+    item_id = source.get("id") or source.get("url") or raw.get("recUuid") or raw.get("uuid") or raw.get("docId") or raw.get("id")
+    if item_id:
+        return str(item_id).casefold().strip()
+    return "|".join(
+        str(source.get(key) or "").casefold().strip()
+        for key in ("title", "source", "time", "content", "snippet")
+    )