sync snapshot from LAN repo (source xuhd-new @ edb08220)

frontend-web/public/runtime-config.js

@@ -112,6 +112,9 @@ window.__CM_RUNTIME_CONFIG__ = {
   // A 系统跳转地址：人物、组织、事件等列表项打开链接使用。
   VITE_AIPI_PRESENT_URL: "http://present.a.uat.4.cn/present/domain",
 
+  // B 系统（XWT检索）头像代理；按部署环境设置为浏览器可访问的地址。
+  VITE_AIPI_MINIO_PROXY_URL: "http://minio-proxy.b.sit.4.cn/facebook-proxy",
+
   // B 系统 knowes 根地址。
   VITE_AIPI_KNOWES_URL: "http://ch1.b.uat.4.cn/knowes",
 

frontend-web/src/components/assistant-ui/document-upload-popup.tsx

@@ -599,8 +599,8 @@ const DocumentUploadPopupContent: FC<{
   };
 
   const yUserId = getYUserId();
-  const userToken = getStoredAuth()?.access_token;
-  const iframeTokenParam = userToken ? `&token=${encodeURIComponent(userToken)}` : "";
+  const notebookToken = getNotebookToken();
+  const iframeTokenParam = notebookToken ? `&token=${encodeURIComponent(notebookToken)}` : "";
 
   return (
     <Dialog open={isOpen} onOpenChange={onOpenChange}>

frontend-web/src/components/assistant-ui/network-actor.tsx

@@ -188,9 +188,12 @@ export function NetworkActor({ inputValue, searchSpaceId, onSuccess }: NetworkAc
   const [directionType, setDirectionType] = useState<string>('TW');
   const directionValue = useRef<string>('TW');
   const consumerUrl = readRuntimeEnv("VITE_AIPI_CONSUMER_URL") || 'https://ch1.b.uat.4.cn/consumer';
-  const backendUrl = typeof window !== 'undefined'
-    ? (readRuntimeEnv("VITE_BACKEND_BASE_URL") || 'http://localhost:8001')
-    : 'http://localhost:8001';
+  const minioProxyUrl = readRuntimeEnv("VITE_AIPI_MINIO_PROXY_URL") || 'http://minio-proxy.b.sit.4.cn/facebook-proxy';
+  const actorBaseUrl = readRuntimeEnv("VITE_AIPI_B2ACTOR_URL") || 'http://ty.b.uat.4.cn';
+  const avatarUrl = (profileImage: string | null) => profileImage
+    ? `${minioProxyUrl}?${new URLSearchParams({ url: profileImage, is_avatar: 'true' })}`
+    : `${actorBaseUrl}/manage/img/default.png`;
+  const backendUrl = readRuntimeEnv("VITE_AIPI_BACKENP_URL") || 'http://localhost:8001';
   const [selectedSearchSources, setSelectedSearchSources] = useState<UserDataItem[]>([]);
   const [isImporting, setIsImporting] = useState(false);
 
@@ -288,9 +291,7 @@ export function NetworkActor({ inputValue, searchSpaceId, onSuccess }: NetworkAc
         setIsSearching(false);
         const result: UserDataItem[] = data.data.list;
         result.forEach((item) => {
-          item.profile_image = item.profile_image
-            ? `http://minio-proxy.b.sit.4.cn/facebook-proxy?url=${item.profile_image}&is_avatar=true`
-            : 'http://ty.b.uat.4.cn/manage/img/default.png';
+          item.profile_image = avatarUrl(item.profile_image);
           item.classification = `${item.political_stand}-${item.history_key}-${item.history_title}`;
           item.title = item.user_name || item.screen_name || item.user_id;
           item.content = item.agentInfo['账号描述'];
@@ -330,9 +331,7 @@ export function NetworkActor({ inputValue, searchSpaceId, onSuccess }: NetworkAc
         const result: UserDataItem[] = data.data.data;
         result.forEach((item) => {
           item.hobbies = item.groupDiscussTags || [];
-          item.profile_image = item.profile_image
-            ? `http://minio-proxy.b.sit.4.cn/facebook-proxy?url=${item.profile_image}&is_avatar=true`
-            : 'http://ty.b.uat.4.cn/manage/img/default.png';
+          item.profile_image = avatarUrl(item.profile_image);
           item.classification = `${item.political_stand}`;
           item.title = item.group_name || item.groupName;
           if (item.hx_info?.['用户画像']) {

frontend-web/src/components/workspace/settings/context-window-settings-page.tsx

@@ -16,6 +16,7 @@ export function ContextWindowSettingsPage() {
   const update = useUpdateConversationContextSettings();
   const enabled = data?.recent_turn_window_enabled ?? false;
   const recentTurns = data?.recent_turns ?? 3;
+  const useQuestionAsTitle = data?.use_question_as_title ?? false;
 
   const handleChange = async (checked: boolean) => {
     try {
@@ -30,6 +31,19 @@ export function ContextWindowSettingsPage() {
     }
   };
 
+  const handleQuestionTitleChange = async (checked: boolean) => {
+    try {
+      await update.mutateAsync({ use_question_as_title: checked });
+      toast.success(
+        checked
+          ? "已开启：对话标题直接使用首个问题"
+          : "已关闭：恢复回答后的标题生成",
+      );
+    } catch (error) {
+      toast.error(error instanceof Error ? error.message : "保存失败");
+    }
+  };
+
   return (
     <section className="space-y-6">
       <div className="space-y-1">
@@ -62,6 +76,25 @@ export function ContextWindowSettingsPage() {
         )}
       </div>
 
+      <div className="flex items-center justify-between gap-6 rounded-lg border p-4">
+        <div className="space-y-1">
+          <div className="font-medium">标题直接使用首个问题</div>
+          <p className="text-muted-foreground text-sm">
+            开启后，首个问题会立即成为最终标题；回答结束后不再额外调用模型生成标题，也不轮询标题状态，避免影响下一轮提问。关闭后恢复原来的回答后标题生成方式。
+          </p>
+        </div>
+        {isLoading ? (
+          <Loader2Icon className="text-muted-foreground size-5 animate-spin" />
+        ) : (
+          <Switch
+            checked={useQuestionAsTitle}
+            disabled={isError || update.isPending}
+            onCheckedChange={handleQuestionTitleChange}
+            aria-label="标题直接使用首个问题"
+          />
+        )}
+      </div>
+
       {isError && (
         <div className="flex items-center justify-between gap-3">
           <p className="text-destructive text-sm">

frontend-web/src/components/workspace/settings/roundtable-room-settings-page.tsx

@@ -1,6 +1,6 @@
 "use client";
 
-import { Clock3Icon, ListOrderedIcon, Loader2Icon, UsersIcon } from "lucide-react";
+import { Clock3Icon, FileOutputIcon, ListOrderedIcon, Loader2Icon, UsersIcon } from "lucide-react";
 import { useEffect, useState } from "react";
 import { toast } from "sonner";
 
@@ -9,6 +9,7 @@ import { Switch } from "@/components/ui/switch";
 import {
   useRoundtableRoomLimitSettings,
   useUpdateRoundtableRoomLimitSettings,
+  useWriteFileCompatibilitySettings,
 } from "@/core/system-settings/hooks";
 import { loadRoundtableRoomSummary } from "@/roundtable-planning/api/rooms";
 
@@ -16,7 +17,10 @@ export function RoundtableRoomSettingsPage() {
   const { data, isLoading, isError, isFetching, refetch } =
     useRoundtableRoomLimitSettings();
   const update = useUpdateRoundtableRoomLimitSettings();
+  const globalFileStream = useWriteFileCompatibilitySettings();
   const enabled = data?.enabled ?? true;
+  const fileStreamEnabled = data?.write_file_compatibility_enabled ?? false;
+  const globalFileStreamEnabled = globalFileStream.data?.enabled ?? false;
   const [maxRooms, setMaxRooms] = useState(String(data?.max_rooms ?? 3));
   const [maxHoldMinutes, setMaxHoldMinutes] = useState(
     String(data?.max_hold_minutes ?? 30),
@@ -62,6 +66,7 @@ export function RoundtableRoomSettingsPage() {
     max_rooms?: number;
     max_hold_minutes?: number;
     max_cycles?: number;
+    write_file_compatibility_enabled?: boolean;
   }) => {
     try {
       await update.mutateAsync(next);
@@ -215,6 +220,37 @@ export function RoundtableRoomSettingsPage() {
         </Button>
       </div>
 
+      <div className="flex items-center justify-between gap-6 rounded-lg border p-4">
+        <div className="space-y-1">
+          <div className="flex items-center gap-2 font-medium">
+            <FileOutputIcon className="size-4" />
+            会商文件流式兼容
+          </div>
+          <p className="text-muted-foreground text-sm">
+            会商生成文件并打开沙箱时使用兼容写入。全局「文件流式兼容」开着时，会商自动使用；全局关着时，打开本开关后只有会商使用，普通对话保持原来的写入方式。
+          </p>
+          <p className="text-muted-foreground text-xs">
+            {globalFileStreamEnabled
+              ? "当前会商会使用兼容写入，因为全局开关已开启。"
+              : fileStreamEnabled
+                ? "当前只有会商使用兼容写入。"
+                : "当前会商使用各模型原来的写入方式。"}
+          </p>
+        </div>
+        {isLoading ? (
+          <Loader2Icon className="text-muted-foreground size-5 animate-spin" />
+        ) : (
+          <Switch
+            checked={fileStreamEnabled}
+            disabled={isError || update.isPending}
+            onCheckedChange={(checked) =>
+              void save({ write_file_compatibility_enabled: checked })
+            }
+            aria-label="会商启用文件流式兼容"
+          />
+        )}
+      </div>
+
       <p className="text-muted-foreground text-sm">
         当前全站占用：{activeCount == null ? "暂时无法读取" : `${activeCount} / ${data?.max_rooms ?? maxRooms}`}
       </p>

frontend-web/src/core/model-management/api.ts

@@ -1,6 +1,14 @@
 import { apiFetch } from "@/core/api";
 import { getBackendBaseURL } from "@/core/config";
 
+export type ModelWriteFileTransport =
+  | "native"
+  | "buffered_gateway"
+  | "artifact_renderer"
+  | "text_bridge";
+
+export type ModelWriteFileTransportToggle = "native" | "buffered_gateway";
+
 export interface ManagedModel {
   name: string;
   display_name: string | null;
@@ -23,6 +31,7 @@ export interface ManagedModel {
   thinking: Record<string, unknown> | null;
   use_responses_api: boolean | null;
   output_version: string | null;
+  write_file_transport: ModelWriteFileTransport;
   extra_settings: Record<string, unknown>;
   api_key_configured: boolean;
   api_key_source: string | null;
@@ -120,6 +129,22 @@ export async function updateManagedModelReasoningEffortCapability(
   return response.json() as Promise<ManagedModel>;
 }
 
+export async function updateManagedModelWriteFileTransport(
+  name: string,
+  writeFileTransport: ModelWriteFileTransportToggle,
+): Promise<ManagedModel> {
+  const response = await apiFetch(
+    `${getBackendBaseURL()}/api/admin/models/write-file-transport/${encodeURIComponent(name)}`,
+    {
+      method: "PUT",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ write_file_transport: writeFileTransport }),
+    },
+  );
+  if (!response.ok) return readError(response, "保存文件流式兼容设置失败");
+  return response.json() as Promise<ManagedModel>;
+}
+
 export async function deleteManagedModel(name: string): Promise<void> {
   const response = await apiFetch(`${getBackendBaseURL()}/api/admin/models/${encodeURIComponent(name)}`, {
     method: "DELETE",

frontend-web/src/core/model-management/hooks.ts

@@ -8,7 +8,9 @@ import {
   testManagedModel,
   updateManagedModel,
   updateManagedModelReasoningEffortCapability,
+  updateManagedModelWriteFileTransport,
   type ModelSavePayload,
+  type ModelWriteFileTransportToggle,
 } from "./api";
 
 export const MANAGED_MODELS_QUERY_KEY = ["admin", "model-config"] as const;
@@ -66,6 +68,25 @@ export function useUpdateManagedModelReasoningEffortCapability() {
   });
 }
 
+export function useUpdateManagedModelWriteFileTransport() {
+  const queryClient = useQueryClient();
+  return useMutation({
+    mutationFn: ({
+      name,
+      writeFileTransport,
+    }: {
+      name: string;
+      writeFileTransport: ModelWriteFileTransportToggle;
+    }) => updateManagedModelWriteFileTransport(name, writeFileTransport),
+    onSuccess: async () => {
+      await Promise.all([
+        queryClient.invalidateQueries({ queryKey: MANAGED_MODELS_QUERY_KEY }),
+        queryClient.invalidateQueries({ queryKey: ["models"] }),
+      ]);
+    },
+  });
+}
+
 export function useDeleteManagedModel() {
   const queryClient = useQueryClient();
   return useMutation({

frontend-web/src/core/system-settings/api.ts

@@ -728,10 +728,12 @@ export interface MemoryRuntimeSettings {
 export interface ConversationContextSettings {
   recent_turn_window_enabled: boolean;
   recent_turns: number;
+  use_question_as_title: boolean;
 }
 
 export interface ConversationContextUpdate {
-  recent_turn_window_enabled: boolean;
+  recent_turn_window_enabled?: boolean;
+  use_question_as_title?: boolean;
 }
 
 export interface RoundtableRoomLimitSettings {
@@ -739,6 +741,7 @@ export interface RoundtableRoomLimitSettings {
   max_rooms: number;
   max_hold_minutes: number;
   max_cycles: number;
+  write_file_compatibility_enabled: boolean;
 }
 
 export interface RoundtableRoomLimitUpdate {
@@ -746,6 +749,7 @@ export interface RoundtableRoomLimitUpdate {
   max_rooms?: number;
   max_hold_minutes?: number;
   max_cycles?: number;
+  write_file_compatibility_enabled?: boolean;
 }
 
 export interface OrdinaryKnowledgeRetrievalSettings {

frontend-web/src/core/threads/hooks.ts

@@ -1340,11 +1340,11 @@ export function useThreadStream({
             runId = inflight?.run_id ?? null;
             if (runId) {
               activeRunIdRef.current = runId;
-            } else {
-              const lastTerminal = runs.find((item) =>
-                TERMINAL_RUN_STATUSES.has(item.status),
-              );
-              if (lastTerminal) terminalStatus = lastTerminal.status;
+            } else {
+              const lastTerminal = runs.find((item) =>
+                TERMINAL_RUN_STATUSES.has(item.status),
+              );
+              if (lastTerminal) terminalStatus = lastTerminal.status;
             }
           } catch {
             if (!isCurrentRecovery()) return "unknown";
@@ -1357,45 +1357,45 @@ export function useThreadStream({
           }
         }
 
-        if (terminalStatus) {
-          try {
-            return (await applyTerminalState(terminalStatus))
-              ? "terminal"
-              : "unknown";
-          } catch {
-            if (isCurrentRecovery()) {
-              lastStreamActivityAtRef.current = Date.now();
-              setStreamHoldOpen(true);
-            }
-            return "unknown";
-          }
-        }
-
+        if (terminalStatus) {
+          try {
+            return (await applyTerminalState(terminalStatus))
+              ? "terminal"
+              : "unknown";
+          } catch {
+            if (isCurrentRecovery()) {
+              lastStreamActivityAtRef.current = Date.now();
+              setStreamHoldOpen(true);
+            }
+            return "unknown";
+          }
+        }
+
         if (!runId) {
-          // The backend confirms there is no in-flight run, but an SDK reader
-          // can remain locally open after a proxy/SSE disconnect. Detach that
-          // reader only after removing its reconnect marker so stop() cannot
-          // accidentally cancel a server run.
-          try {
-            const readerStopped = await stopRecoveredRunReader({
-              isCurrent: isCurrentRecovery,
-              clearReconnect: () => clearActiveRunId(tid),
-              stop: () => threadStreamRef.current?.stop(),
-            });
-            if (!readerStopped) return "unknown";
-            activeRunIdRef.current = null;
-            setStreamHoldOpen(false);
-            streamReconnectAttemptsRef.current = 0;
-            backgroundHistory.confirmTerminal();
-            void backgroundHistory.refresh(false);
-            return "terminal";
-          } catch {
-            if (isCurrentRecovery()) {
-              lastStreamActivityAtRef.current = Date.now();
-              setStreamHoldOpen(true);
-            }
-            return "unknown";
-          }
+          // The backend confirms there is no in-flight run, but an SDK reader
+          // can remain locally open after a proxy/SSE disconnect. Detach that
+          // reader only after removing its reconnect marker so stop() cannot
+          // accidentally cancel a server run.
+          try {
+            const readerStopped = await stopRecoveredRunReader({
+              isCurrent: isCurrentRecovery,
+              clearReconnect: () => clearActiveRunId(tid),
+              stop: () => threadStreamRef.current?.stop(),
+            });
+            if (!readerStopped) return "unknown";
+            activeRunIdRef.current = null;
+            setStreamHoldOpen(false);
+            streamReconnectAttemptsRef.current = 0;
+            backgroundHistory.confirmTerminal();
+            void backgroundHistory.refresh(false);
+            return "terminal";
+          } catch {
+            if (isCurrentRecovery()) {
+              lastStreamActivityAtRef.current = Date.now();
+              setStreamHoldOpen(true);
+            }
+            return "unknown";
+          }
         }
 
         // A second UI failure can pause recovery while /runs is in flight.

frontend-web/src/pages/ModelManagementPage.tsx

@@ -22,6 +22,7 @@ import {
   CircleAlertIcon,
   CpuIcon,
   EyeIcon,
+  FileOutputIcon,
   GripVerticalIcon,
   KeyRoundIcon,
   LoaderCircleIcon,
@@ -54,7 +55,9 @@ import {
   useSaveManagedModel,
   useTestManagedModel,
   useUpdateManagedModelReasoningEffortCapability,
+  useUpdateManagedModelWriteFileTransport,
 } from "@/core/model-management";
+import { useWriteFileCompatibilitySettings } from "@/core/system-settings/hooks";
 import {
   WorkspaceBody,
   WorkspaceContainer,
@@ -197,25 +200,44 @@ function SortableModelCard({
   dragEnabled,
   reasoningCapabilityDisabled,
   reasoningCapabilityPending,
+  globalWriteFileCompatEnabled,
+  writeFileCompatPending,
   onConfigure,
   onDelete,
   onReasoningCapabilityChange,
   onReasoningEffortChange,
+  onWriteFileCompatChange,
 }: {
   model: ManagedModel;
   isDefault: boolean;
   dragEnabled: boolean;
   reasoningCapabilityDisabled: boolean;
   reasoningCapabilityPending: boolean;
+  globalWriteFileCompatEnabled: boolean;
+  writeFileCompatPending: boolean;
   onConfigure: () => void;
   onDelete: () => void;
   onReasoningCapabilityChange: (enabled: boolean) => void;
   onReasoningEffortChange: (effort: "low" | "medium" | "xhigh") => void;
+  onWriteFileCompatChange: (enabled: boolean) => void;
 }) {
   const { attributes, listeners, setNodeRef, transform, transition, isDragging } = useSortable({
     id: model.name,
     disabled: !dragEnabled,
   });
+  const writeFileCompatEligible =
+    (model.use === "langchain_openai:ChatOpenAI" ||
+      model.use === "deerflow.models.patched_openai:PatchedChatOpenAI") &&
+    !model.use_responses_api;
+  const perModelWriteFileCompat = model.write_file_transport === "buffered_gateway";
+  const writeFileCompatActive = perModelWriteFileCompat || globalWriteFileCompatEnabled;
+  const writeFileCompatHint = globalWriteFileCompatEnabled
+    ? "全局文件流式兼容已开启，所有兼容模型均已强制启用；关闭全局开关后恢复此处的单独设置。"
+    : !writeFileCompatEligible
+      ? "仅支持 OpenAI 兼容接口（Chat Completions）模型。"
+      : perModelWriteFileCompat
+        ? "该模型已启用：文件正文改由普通文本流生成，适配缓存长工具参数的网关。"
+        : "针对会缓存长工具参数的内网模型网关；开启后文件正文改由普通文本流生成，对问答与文件卡片完全透明。";
 
   return (
     <article
@@ -288,6 +310,9 @@ function SortableModelCard({
         {model.supports_vision && (
           <ModelCapabilityBadge icon={EyeIcon} label="视觉" className="bg-amber-500/10 text-amber-700 dark:text-amber-300" />
         )}
+        {writeFileCompatActive && writeFileCompatEligible && (
+          <ModelCapabilityBadge icon={FileOutputIcon} label="文件流式兼容" className="bg-sky-500/10 text-sky-700 dark:text-sky-300" />
+        )}
         {model.supports_reasoning_effort && (
           <ModelCapabilityBadge label={`全局默认：${model.default_reasoning_effort === "low" ? "低" : model.default_reasoning_effort === "medium" ? "中" : "Ultra"}`} className="bg-muted text-muted-foreground" />
         )}
@@ -332,6 +357,22 @@ function SortableModelCard({
           />
         </div>
       )}
+      <div className="mt-2 flex items-center justify-between gap-4 rounded-lg border border-border px-3 py-2.5">
+        <div className="min-w-0">
+          <div className="flex flex-wrap items-center gap-2 text-sm font-medium">
+            <span>文件流式兼容</span>
+            <Tag theme="default" variant="light" size="small">config.yaml</Tag>
+            {globalWriteFileCompatEnabled && <Tag theme="primary" variant="light" size="small">全局已开启</Tag>}
+          </div>
+          <p className="mt-1 text-xs leading-5 text-muted-foreground">{writeFileCompatHint}</p>
+        </div>
+        <Switch
+          value={writeFileCompatActive}
+          loading={writeFileCompatPending}
+          disabled={globalWriteFileCompatEnabled || !writeFileCompatEligible}
+          onChange={(value) => onWriteFileCompatChange(Boolean(value))}
+        />
+      </div>
     </article>
   );
 }
@@ -446,6 +487,9 @@ export default function ModelManagementPage() {
   const reorder = useReorderManagedModels();
   const updateReasoningCapability =
     useUpdateManagedModelReasoningEffortCapability();
+  const updateWriteFileTransport = useUpdateManagedModelWriteFileTransport();
+  const { data: writeFileCompatSettings } = useWriteFileCompatibilitySettings();
+  const globalWriteFileCompatEnabled = writeFileCompatSettings?.enabled ?? false;
   const sourceModels = data?.models ?? [];
   const orderedModels = useMemo(() => {
     if (!pendingOrder || pendingOrder.length !== sourceModels.length) return sourceModels;
@@ -511,6 +555,26 @@ export default function ModelManagementPage() {
       showGlobalMessage("error", error instanceof Error ? error.message : "保存全局思考强度失败");
     }
   };
+  const handleWriteFileCompatChange = async (
+    model: ManagedModel,
+    enabled: boolean,
+  ) => {
+    try {
+      await updateWriteFileTransport.mutateAsync({
+        name: model.name,
+        writeFileTransport: enabled ? "buffered_gateway" : "native",
+      });
+      showGlobalMessage(
+        "success",
+        `模型“${model.display_name || model.name}”的文件流式兼容已${enabled ? "开启" : "关闭"}，从下一次发起的问答开始生效。`,
+      );
+    } catch (error) {
+      showGlobalMessage(
+        "error",
+        error instanceof Error ? error.message : "保存文件流式兼容设置失败",
+      );
+    }
+  };
   const dragEnabled = filter === "all" && !search.trim() && !reorder.isPending;
   const sensors = useSensors(
     useSensor(PointerSensor, { activationConstraint: { distance: 6 } }),
@@ -637,6 +701,11 @@ export default function ModelManagementPage() {
                             updateReasoningCapability.isPending &&
                             updateReasoningCapability.variables?.name === model.name
                           }
+                          globalWriteFileCompatEnabled={globalWriteFileCompatEnabled}
+                          writeFileCompatPending={
+                            updateWriteFileTransport.isPending &&
+                            updateWriteFileTransport.variables?.name === model.name
+                          }
                           onConfigure={() => setEditor(model)}
                           onDelete={() => setDeleteTarget(model)}
                           onReasoningCapabilityChange={(enabled) =>
@@ -645,6 +714,9 @@ export default function ModelManagementPage() {
                           onReasoningEffortChange={(effort) =>
                             void handleReasoningEffortChange(model, effort)
                           }
+                          onWriteFileCompatChange={(enabled) =>
+                            void handleWriteFileCompatChange(model, enabled)
+                          }
                         />
                       ))}
                     </div>

offline-backend-20260512/backend/CLAUDE.md

@@ -1489,7 +1489,7 @@ scheduler runs once per deployment. Per run:
 - **可检测的模型失败**：`LLMErrorHandlingMiddleware` 在同一模型重试 3 次后**不抛异常**，而是返回兜底 `AIMessage` —— 现在给它打 `additional_kwargs[LLM_ERROR_MARKER="deerflow_llm_error"]=reason`（model_dump 保留，经 checkpointer 序列化仍在）。`is_llm_error_message(msg)` 据标记（回退按已知兜底文案正文）识别，返回 reason。
 - **候选模型链**：保留真正模型故障后的自动换模，但“仅有思考 / 暂时无正文 / 空交付”不再判作模型故障，绝不据此切换用户选择的模型。`app/gateway/roundtable_model_fallback.py::fallback_model_chain(current)` 按 `config.yaml models[]` 顺序给出 `[当前模型, …其余]`（去重保序）。`reason_text(reason)` 翻译成简短中文。总开关 env `ROUNDTABLE_MODEL_FALLBACK`（默认开，`0/false/off` 关）；`ROUNDTABLE_MODEL_FALLBACK_MAX`（0=试完全部）。
 - **触发换模型的情形**：① 单轮 run 抛异常（start_run 失败 / idle 卡死 → `timeout`，其余 `exception`）；② 兜底错误文案（`is_llm_error_message`）。交付无效、仅有思考、空串以及普通总控轮既无正文又无派活**均不触发换模**；唯一例外是无工具的最终收口请求已经完整结束仍无正文，此时结果客观不可交付，才继续自动换下一个模型。席位仍只认剥掉 `<think>` 后的可见正文（`deerflow.agents.roundtable_orchestrator.delivery`：`visible_seat_text` / `classify_seat_delivery`）：前台 `_special_run` 对无效交付 yield `status:"invalid_delivery"` 且不广播；`GET /last-ai-message` 回 `{delivered, content, length, invalid_reason}`（思考-only 标 `thinking_only`，总控席位清单标 ⚠️）。后台引擎只把可见正文写入 `deliveries`，无效席位继续走既有未交付处理，不切换模型。Step3 单例（report/summary/dashboard/ingest/结构化）的产物是落盘文件，同样只对真模型错误换模型。
-- **Step2 总控工具与可配置收口轮数**：普通总控使用真正白名单，只允许 `agent_orchestration` / `ask_clarification`；`write_file`、`present_files`、技能目录与文件/命令工具均不可见。`system_settings.json -> roundtable_room_limit.max_cycles`（默认 8，范围 2..100；管理员在设置弹窗「会商并发与收口」中修改）控制 recommend 模式到第几轮切换为显式 `finalize_only`：空工具白名单、注入已有有效席位全文、要求直接输出最终总结（失败席位如实标缺口），不能表现为“已停止”。一次会商启动时锁定当时值，运行中修改从下一次研讨生效；前台实时会商与后台作业读取同一设置。chain / dag 的最终总结仍在既定席位完成后走同一纯正文路径；LangGraph 单轮 recursion limit 仍为 1000。
+- **Step2 总控工具与可配置收口轮数**：普通总控使用真正白名单，只允许 `agent_orchestration` / `ask_clarification`；`write_file`、`present_files`、技能目录与文件/命令工具均不可见。`system_settings.json -> roundtable_room_limit.max_cycles`（默认 8，范围 2..100；管理员在设置弹窗「会商并发与收口」中修改）控制 recommend 模式到第几轮切换为显式 `finalize_only`：空工具白名单、注入已有有效席位全文、要求直接输出最终总结（失败席位如实标缺口），不能表现为“已停止”。一次会商启动时锁定当时值，运行中修改从下一次研讨生效；前台实时会商与后台作业读取同一设置。同一页的「会商文件流式兼容」(`write_file_compatibility_enabled`，默认关) 只作用于会商总控、席位、第三步和后台作业：全局文件流式兼容开着时会商一并使用；全局关着而本开关开着时，只有这些会商 run 改走 `buffered_gateway`。普通对话不读这个开关。chain / dag 的最终总结仍在既定席位完成后走同一纯正文路径；LangGraph 单轮 recursion limit 仍为 1000。
 - **后台**：`InProcessRoundtableGateway._run_turn_with_fallback(...)` 包住 `_run_agent_turn`，逐候选模型试；换模型落 `model_switch` 诊断、全失败落 `model_fallback_exhausted`。`run_leader/run_seat/run_report/run_summary/run_dashboard_data` 全走它（report/summary/dashboard 的产物自愈重试保留，并记住已切到的可用模型用于后续续写轮）。各 Turn/Result 带回 `switches: list[ModelSwitch]`，引擎 `_emit_switches` → `JobProgressWriter.model_switch(...)` 往研讨时间线插一条 `role="system"` 系统提示气泡。
 - **前台**：`_special_run`/`_leader_run` stream 收尾后仅检测明确的模型错误标记；命中才 yield `{status:"model_switch", role, agent_name, failed_model, next_model, reason}` 帧并用下一个模型重新流式，思考-only/空正文不进入该分支；终态帧带 `model_used`。结构化抽取 `flow_extract.py` 的 round1 `model.astream` 也包了同款换模型循环（emit `event:"model_switch"` 帧）。
 - **Step2 单席位隔离**：`engine._run_recommend/_run_chain/_run_dag` 的每个席位执行包 try/except —— 一个席位彻底失败（落 `seat_failed` 诊断 + 置空交付）不拖垮整批/整链/整个作业。

offline-backend-20260512/backend/app/gateway/roundtable_inprocess_gateway.py

@@ -36,6 +36,7 @@ from app.gateway.routers._roundtable_seed import (
     SUMMARY_AGENT_ID,
     ensure_roundtable_functional_agents,
 )
+from app.gateway.routers.multi_agent import _roundtable_file_stream_context
 from app.gateway.services import RUN_SCOPE_BACKGROUND_CHILD
 from deerflow.agents.roundtable_orchestrator.delivery import classify_seat_delivery, visible_seat_text
 from deerflow.agents.roundtable_orchestrator.report_json import extract_report_json
@@ -365,6 +366,7 @@ async def _run_agent_turn(
             # 被上轮会话 memory.json 带偏(臆造席位 / 直接干活)。详见该处注释。
             "memory_injection_enabled": False,
             "allowed_tools": allowed_tools,
+            **_roundtable_file_stream_context(),
         },
         excluded_tools=excluded_tools,
         skill_stop_names=skill_stop_names,

offline-backend-20260512/backend/app/gateway/routers/model_management.py

@@ -224,6 +224,12 @@ class ReasoningEffortCapabilityRequest(BaseModel):
     default_reasoning_effort: Literal["low", "medium", "xhigh"] | None = None
 
 
+class WriteFileTransportToggleRequest(BaseModel):
+    """Per-model 文件流式兼容 switch: off keeps native, on selects buffered_gateway."""
+
+    write_file_transport: Literal["native", "buffered_gateway"]
+
+
 class ModelOrderRequest(BaseModel):
     """The complete ordered set of model stable names."""
 
@@ -313,6 +319,39 @@ def _validate_reasoning_capability(model: dict[str, Any], enabled: bool) -> None
         )
 
 
+# ``write_file_transport=buffered_gateway`` is implemented only by the patched
+# OpenAI-compatible chat-completions classes (see deerflow.models.factory);
+# every other driver is rejected there at construction time, so mirror the
+# same constraint here with a clearer message.
+_BUFFERED_GATEWAY_COMPATIBLE_USES = frozenset(
+    {
+        "langchain_openai:ChatOpenAI",
+        "deerflow.models.patched_openai:PatchedChatOpenAI",
+    }
+)
+
+
+def _validate_write_file_transport(model: dict[str, Any], transport: str) -> None:
+    if transport != WriteFileTransport.BUFFERED_GATEWAY.value:
+        return
+    if str(model.get("use") or "") not in _BUFFERED_GATEWAY_COMPATIBLE_USES:
+        raise HTTPException(
+            status_code=422,
+            detail=(
+                "write_file compatibility requires an OpenAI-compatible "
+                "langchain_openai:ChatOpenAI model"
+            ),
+        )
+    if bool(model.get("use_responses_api")):
+        raise HTTPException(
+            status_code=422,
+            detail=(
+                "write_file compatibility requires the chat-completions API, "
+                "not use_responses_api"
+            ),
+        )
+
+
 async def _restore_capability(
     store: ModelCapabilityStore,
     model_name: str,
@@ -672,6 +711,48 @@ async def update_reasoning_effort_capability(
     )
 
 
+@router.put(
+    "/write-file-transport/{model_name:path}",
+    response_model=ModelAdminItem,
+)
+async def update_write_file_transport(
+    model_name: str,
+    body: WriteFileTransportToggleRequest,
+    request: Request,
+) -> ModelAdminItem:
+    """Toggle one model's buffered-gateway write_file compatibility in config.yaml.
+
+    The global system-settings switch keeps overriding every compatible model
+    while it is enabled; this per-model value becomes effective as soon as the
+    global switch is turned off.
+    """
+
+    await _require_admin(request)
+    async with _MUTATION_LOCK:
+        path = _resolve_config_path()
+        raw_before, _, models = await asyncio.to_thread(_read_config_data, path)
+        index = next((i for i, item in enumerate(models) if str(item.get("name")) == model_name), None)
+        if index is None:
+            raise HTTPException(status_code=404, detail=f"Model '{model_name}' was not found")
+        updated = copy.deepcopy(models[index])
+        _validate_write_file_transport(updated, body.write_file_transport)
+        updated["write_file_transport"] = body.write_file_transport
+        try:
+            ModelConfig.model_validate(updated)
+        except Exception as exc:
+            raise HTTPException(status_code=422, detail=f"Invalid model configuration: {exc}") from exc
+        next_models = [*models]
+        next_models[index] = updated
+        await _save_models(request, raw_before, next_models)
+    capabilities = await _capabilities_by_name(request)
+    capability = capabilities.get(model_name, {})
+    return _to_admin_item(
+        updated,
+        supports_reasoning_effort=bool(capability.get("supports_reasoning_effort", False)),
+        default_reasoning_effort=str(capability.get("default_reasoning_effort") or "xhigh"),
+    )
+
+
 @router.put("/{model_name:path}", response_model=ModelAdminItem)
 async def update_model_configuration(model_name: str, body: ModelUpsertRequest, request: Request) -> ModelAdminItem:
     admin = await _require_admin(request)

offline-backend-20260512/backend/app/gateway/routers/multi_agent.py

@@ -742,6 +742,7 @@ def _run_payload(
     cancel_on_disconnect: bool = False,
     extra_context: dict[str, Any] | None = None,
     files: list[dict[str, Any]] | None = None,
+    roundtable_file_stream: bool = False,
 ) -> dict[str, Any]:
     """构造 ``POST /api/threads/{thread_id}/runs/stream`` 的请求体。
 
@@ -819,9 +820,20 @@ def _run_payload(
     # 供 read_peer_delivery 工具按需读取）。只进 context dict,不进 prompt → 不调工具就不耗 token。
     if extra_context:
         payload["context"].update(extra_context)
+    if roundtable_file_stream:
+        payload["context"].update(_roundtable_file_stream_context())
     return payload
 
 
+def _roundtable_file_stream_context() -> dict[str, bool]:
+    """会商专用文件流式兼容。关闭时不写 context，本轮仍只看全局开关。"""
+    from deerflow.config.system_settings import get_roundtable_room_limit_settings
+
+    if get_roundtable_room_limit_settings().write_file_compatibility_enabled:
+        return {"roundtable_write_file_compatibility": True}
+    return {}
+
+
 # ---------------------------------------------------------------------------
 # POST /api/multi-agent/init — Step 2 席位 + 总控初始化（SSE）
 # ---------------------------------------------------------------------------
@@ -1697,6 +1709,7 @@ async def _leader_run(
             cancel_on_disconnect=True,
             # 用户随干预上传的参考文件 → 总控可用 read_file 等工具读取后再派活。
             files=files,
+            roundtable_file_stream=True,
         )
 
     body = _mk_body(current_model)
@@ -1980,6 +1993,7 @@ async def _special_run(
             ),
             # 用户随干预上传的参考文件 → 席位可用 read_file 等工具读取。
             files=files,
+            roundtable_file_stream=True,
         )
 
     t_payload_built = time.perf_counter()

offline-backend-20260512/backend/app/gateway/routers/system_settings.py

@@ -456,10 +456,15 @@ class ConversationContextResponse(BaseModel):
         default=3,
         description="Fixed number of real user turns retained by the sliding window",
     )
+    use_question_as_title: bool = Field(
+        default=False,
+        description="Whether the first user question is the final title, without post-answer title generation",
+    )
 
 
 class ConversationContextUpdateRequest(BaseModel):
-    recent_turn_window_enabled: bool
+    recent_turn_window_enabled: bool | None = None
+    use_question_as_title: bool | None = None
 
 
 class RoundtableRoomLimitResponse(BaseModel):
@@ -467,6 +472,7 @@ class RoundtableRoomLimitResponse(BaseModel):
     max_rooms: int = 3
     max_hold_minutes: int = 30
     max_cycles: int = 8
+    write_file_compatibility_enabled: bool = False
 
 
 class RoundtableRoomLimitUpdateRequest(BaseModel):
@@ -474,6 +480,7 @@ class RoundtableRoomLimitUpdateRequest(BaseModel):
     max_rooms: int | None = Field(default=None, ge=1, le=20)
     max_hold_minutes: int | None = Field(default=None, ge=1, le=1440)
     max_cycles: int | None = Field(default=None, ge=2, le=100)
+    write_file_compatibility_enabled: bool | None = None
 
 
 class OrdinaryKnowledgeRetrievalResponse(BaseModel):
@@ -556,6 +563,7 @@ def _conversation_context_response(settings: ConversationContextSettings) -> Con
     return ConversationContextResponse(
         recent_turn_window_enabled=settings.recent_turn_window_enabled,
         recent_turns=3,
+        use_question_as_title=settings.use_question_as_title,
     )
 
 
@@ -573,13 +581,15 @@ async def update_conversation_context_setting(
     """Enable or disable the fixed recent-three-turn window. Admin only."""
     await _require_admin(request)
     current = load_system_settings()
-    updated = current.conversation_context.model_copy(
-        update={"recent_turn_window_enabled": body.recent_turn_window_enabled}
-    )
+    updated_fields = body.model_dump(exclude_unset=True, exclude_none=True)
+    if not updated_fields:
+        raise HTTPException(status_code=422, detail="至少需要提供一个上下文设置项")
+    updated = current.conversation_context.model_copy(update=updated_fields)
     saved = save_system_settings(current.model_copy(update={"conversation_context": updated}))
     logger.info(
-        "Conversation recent-turn window changed: enabled=%s retained_turns=3",
+        "Conversation runtime settings changed: recent_turn_window=%s use_question_as_title=%s",
         saved.conversation_context.recent_turn_window_enabled,
+        saved.conversation_context.use_question_as_title,
     )
     return _conversation_context_response(saved.conversation_context)
 
@@ -590,6 +600,7 @@ def _roundtable_room_limit_response(settings: RoundtableRoomLimitSettings) -> Ro
         max_rooms=settings.max_rooms,
         max_hold_minutes=settings.max_hold_minutes,
         max_cycles=settings.max_cycles,
+        write_file_compatibility_enabled=settings.write_file_compatibility_enabled,
     )
 
 
@@ -616,14 +627,17 @@ async def update_roundtable_room_limit_setting(
         updates["max_hold_minutes"] = body.max_hold_minutes
     if body.max_cycles is not None:
         updates["max_cycles"] = body.max_cycles
+    if body.write_file_compatibility_enabled is not None:
+        updates["write_file_compatibility_enabled"] = body.write_file_compatibility_enabled
     updated = current.roundtable_room_limit.model_copy(update=updates)
     saved = save_system_settings(current.model_copy(update={"roundtable_room_limit": updated}))
     logger.info(
-        "Roundtable runtime limits changed: enabled=%s max_rooms=%s max_hold_minutes=%s max_cycles=%s",
+        "Roundtable runtime limits changed: enabled=%s max_rooms=%s max_hold_minutes=%s max_cycles=%s write_file_compatibility=%s",
         saved.roundtable_room_limit.enabled,
         saved.roundtable_room_limit.max_rooms,
         saved.roundtable_room_limit.max_hold_minutes,
         saved.roundtable_room_limit.max_cycles,
+        saved.roundtable_room_limit.write_file_compatibility_enabled,
     )
     return _roundtable_room_limit_response(saved.roundtable_room_limit)
 

offline-backend-20260512/backend/app/gateway/services.py

@@ -577,6 +577,8 @@ _CONTEXT_CONFIGURABLE_KEYS: frozenset[str] = frozenset(
         "roundtable_peer_deliveries",
         # 岗位会商席位 / 方案总结：每轮 outputs 下 write_file 上限（行动规划不传此键）。
         "position_artifact_cap",
+        # 会商专用文件流式兼容。只由会商 run 写入；普通对话不带此键，仍只看全局开关。
+        "roundtable_write_file_compatibility",
         # 上下文压缩开关:由聊天框下方的按钮控制(默认关闭),透传到 run config 后
         # 在 _build_middlewares 里决定是否挂载 SummarizationMiddleware。
         "summarization_enabled",

offline-backend-20260512/backend/config.yaml

@@ -87,6 +87,7 @@ models:
   temperature: 1.0
   supports_vision: true
   supports_thinking: true
+  write_file_transport: buffered_gateway
 - name: MiniMax-M2.7-fash
   display_name: MiniMax-M2.7-fash
   use: langchain_openai:ChatOpenAI
@@ -107,9 +108,6 @@ models:
   base_url: https://ai.hzfzlz.com/v1
   request_timeout: 600.0
   max_retries: 2
-  # 不带 max_tokens 时该中转站按自身小默认值截断输出（实测 ~64 token），
-  # 检索词 JSON 会被截断导致解析失败走兜底。
-  # write_file 写整页 HTML 时常超过 4096，截断后只剩 path，沙箱无法流式预览。
   max_tokens: 32768
   stream_usage: true
   supports_thinking: false
@@ -125,6 +123,7 @@ models:
   stream_usage: true
   supports_thinking: false
   supports_vision: false
+  write_file_transport: native
 - name: zai-org/GLM-5-FP8
   display_name: zai-org/GLM-5-FP8
   use: langchain_openai:ChatOpenAI
@@ -195,6 +194,9 @@ models:
   temperature: 1.0
   supports_vision: false
   supports_thinking: true
+  # 不带 max_tokens 时该中转站按自身小默认值截断输出（实测 ~64 token），
+  # 检索词 JSON 会被截断导致解析失败走兜底。
+  # write_file 写整页 HTML 时常超过 4096，截断后只剩 path，沙箱无法流式预览。
   # - name: deepseek-v4-flash
   #   display_name: DeepSeek V4 Flash
   #   use: langchain_openai:ChatOpenAI

offline-backend-20260512/backend/packages/harness/deerflow/agents/lead_agent/agent.py

@@ -826,6 +826,8 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
     # create_chat_model(force_disable_thinking=...) so enable_thinking=false is
     # actually injected. Default off → no behavior change for normal chats.
     thinking_force_disabled = bool(cfg.get("thinking_force_disabled", False))
+    # Roundtable runs opt in explicitly. Missing/false keeps the global switch.
+    write_file_compatibility = True if cfg.get("roundtable_write_file_compatibility") else None
     reasoning_effort = cfg.get("reasoning_effort", None)
     requested_model_name: str | None = cfg.get("model_name") or cfg.get("model")
     is_plan_mode = cfg.get("is_plan_mode", False)
@@ -1025,7 +1027,7 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
 
         setup_tools = [setup_writing] if writing_setup_quick else [setup_writing, ask_clarification_tool]
         return create_agent(
-            model=create_chat_model(name=model_name, thinking_enabled=thinking_enabled, force_disable_thinking=thinking_force_disabled, app_config=resolved_app_config),
+            model=create_chat_model(name=model_name, thinking_enabled=thinking_enabled, force_disable_thinking=thinking_force_disabled, write_file_compatibility=write_file_compatibility, app_config=resolved_app_config),
             tools=setup_tools,
             middleware=_build_setup_middlewares(
                 config,
@@ -1062,7 +1064,7 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
             )
         )
         return create_agent(
-            model=create_chat_model(name=model_name, thinking_enabled=thinking_enabled, force_disable_thinking=thinking_force_disabled, app_config=resolved_app_config),
+            model=create_chat_model(name=model_name, thinking_enabled=thinking_enabled, force_disable_thinking=thinking_force_disabled, write_file_compatibility=write_file_compatibility, app_config=resolved_app_config),
             tools=setup_tools,
             middleware=_build_setup_middlewares(
                 config,
@@ -1087,7 +1089,7 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
     if is_bootstrap:
         # Special bootstrap agent with minimal prompt for initial custom agent creation flow
         return create_agent(
-            model=create_chat_model(name=model_name, thinking_enabled=thinking_enabled, force_disable_thinking=thinking_force_disabled, app_config=resolved_app_config),
+            model=create_chat_model(name=model_name, thinking_enabled=thinking_enabled, force_disable_thinking=thinking_force_disabled, write_file_compatibility=write_file_compatibility, app_config=resolved_app_config),
             tools=get_available_tools(
                 include_mcp=not classic_knowledge_space_strict,
                 model_name=model_name,
@@ -1097,6 +1099,7 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
                 allowed_tools=requested_allowed_tools,
                 allow_agent_orchestration=agent_id in _COORDINATOR_AGENT_IDS,
                 app_config=resolved_app_config,
+                write_file_compatibility=write_file_compatibility,
             )
             + [setup_agent],
             middleware=_build_middlewares(config, model_name=model_name, app_config=resolved_app_config),
@@ -1130,6 +1133,7 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
             allowed_tools=requested_allowed_tools,
             allow_agent_orchestration=agent_id in _COORDINATOR_AGENT_IDS,
             app_config=resolved_app_config,
+            write_file_compatibility=write_file_compatibility,
         )
     )
     if agent_id == FORCED_RESEARCH_AGENT_ID:
@@ -1243,7 +1247,7 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
     )
 
     return create_agent(
-        model=create_chat_model(name=model_name, thinking_enabled=thinking_enabled, reasoning_effort=reasoning_effort, force_disable_thinking=thinking_force_disabled, app_config=resolved_app_config),
+        model=create_chat_model(name=model_name, thinking_enabled=thinking_enabled, reasoning_effort=reasoning_effort, force_disable_thinking=thinking_force_disabled, write_file_compatibility=write_file_compatibility, app_config=resolved_app_config),
         tools=default_tools,
         middleware=_build_middlewares(config, model_name=model_name, agent_name=agent_id, app_config=resolved_app_config),
         system_prompt=system_prompt,

offline-backend-20260512/backend/packages/harness/deerflow/agents/middlewares/title_middleware.py

@@ -12,6 +12,7 @@ from langchain.agents import AgentState
 from langchain.agents.middleware import AgentMiddleware
 from langgraph.runtime import Runtime
 
+from deerflow.config.system_settings import get_conversation_context_settings
 from deerflow.config.title_config import get_title_config
 from deerflow.models import create_chat_model
 from deerflow.utils.time import now_iso
@@ -209,14 +210,15 @@ class TitleMiddleware(AgentMiddleware[TitleMiddlewareState]):
         return self._fallback_title(user_msg)
 
     def _should_set_provisional_title(self, state: TitleMiddlewareState) -> bool:
-        """Check if we should set an instant, question-derived placeholder title.
+        """Check if we should set an instant, question-derived title.
 
         Runs in ``before_model`` so the thread shows a sensible title the moment
-        the user sends their first message — eliminating the "untitled" window
-        while the (possibly long) first answer streams. No LLM call, no latency.
+        the user sends their first message — eliminating the "untitled" window.
+        When the administrator selects question-derived titles, this also works
+        when LLM title generation is disabled because no title model is needed.
         """
         config = self._get_title_config()
-        if not config.enabled:
+        if not config.enabled and not get_conversation_context_settings().use_question_as_title:
             return False
 
         # Already has a title (provisional or final) — nothing to do.
@@ -333,8 +335,13 @@ class TitleMiddleware(AgentMiddleware[TitleMiddlewareState]):
         if not user_msg.strip():
             return None
         title = self._validated_fallback_title(state, user_msg)
-        logger.debug("TitleMiddleware: set provisional title %r before first model call", title)
-        return {"title": title, "title_provisional": True}
+        use_question_as_title = get_conversation_context_settings().use_question_as_title
+        logger.debug(
+            "TitleMiddleware: set %s title %r before first model call",
+            "final question-derived" if use_question_as_title else "provisional question-derived",
+            title,
+        )
+        return {"title": title, "title_provisional": not use_question_as_title}
 
     def _generate_title_result(self, state: TitleMiddlewareState) -> dict | None:
         """Generate a local fallback title without blocking on an LLM call."""

offline-backend-20260512/backend/packages/harness/deerflow/config/system_settings.py

@@ -576,6 +576,13 @@ class ConversationContextSettings(BaseModel):
         default=False,
         description="Whether model requests use the fixed recent-three-turn sliding window.",
     )
+    use_question_as_title: bool = Field(
+        default=False,
+        description=(
+            "Whether the first user question is used as the final conversation title, "
+            "skipping post-answer LLM title generation."
+        ),
+    )
     model_config = ConfigDict(extra="ignore")
 
 
@@ -629,22 +636,12 @@ class RoundtableRoomLimitSettings(BaseModel):
             "The value is captured when an orchestration starts."
         ),
     )
-    model_config = ConfigDict(extra="ignore")
-
-
-class ConversationContextSettings(BaseModel):
-    """Deployment-wide short-context policy for chat model requests.
-
-    The recent-turn window is deliberately opt-in.  When enabled, the lead
-    agent keeps the persisted conversation intact but sends only the latest
-    three real user turns (plus their assistant/tool continuations) to the
-    model.  This avoids an extra summarization-model call while preserving the
-    complete transcript for the UI and checkpoint recovery.
-    """
-
-    recent_turn_window_enabled: bool = Field(
+    write_file_compatibility_enabled: bool = Field(
         default=False,
-        description="Whether model requests use the fixed recent-three-turn sliding window.",
+        description=(
+            "When true, roundtable leader/seat/step-3 runs use buffered-gateway "
+            "write_file streaming. Ordinary chat keeps the global switch."
+        ),
     )
     model_config = ConfigDict(extra="ignore")
 
@@ -714,7 +711,6 @@ class SystemSettings(BaseModel):
     write_file_compatibility: WriteFileCompatibilitySettings = Field(
         default_factory=WriteFileCompatibilitySettings
     )
-    conversation_context: ConversationContextSettings = Field(default_factory=ConversationContextSettings)
     roundtable_room_limit: RoundtableRoomLimitSettings = Field(default_factory=RoundtableRoomLimitSettings)
     local_wiki_encoder: LocalWikiEncoderSettings = Field(default_factory=LocalWikiEncoderSettings)
     strategy_projects: StrategyProjectSettings = Field(default_factory=StrategyProjectSettings)
@@ -851,15 +847,6 @@ def get_roundtable_room_limit_settings() -> RoundtableRoomLimitSettings:
         return RoundtableRoomLimitSettings()
 
 
-def get_conversation_context_settings() -> ConversationContextSettings:
-    """Return the admin-controlled recent-turn policy, failing closed."""
-    try:
-        return load_system_settings().conversation_context
-    except Exception:
-        logger.debug("Failed to read conversation context setting; keeping the recent-turn window disabled", exc_info=True)
-        return ConversationContextSettings()
-
-
 def save_system_settings(settings: SystemSettings) -> SystemSettings:
     """Persist the settings document atomically, returning the saved value."""
     global _settings_cache_key, _settings_cache_value

offline-backend-20260512/backend/packages/harness/deerflow/models/factory.py

@@ -136,7 +136,7 @@ def _enable_stream_usage_by_default(model_use_path: str, model_settings_from_con
         model_settings_from_config["stream_usage"] = True
 
 
-def create_chat_model(name: str | None = None, thinking_enabled: bool = False, *, app_config: AppConfig | None = None, force_disable_thinking: bool = False, **kwargs) -> BaseChatModel:
+def create_chat_model(name: str | None = None, thinking_enabled: bool = False, *, app_config: AppConfig | None = None, force_disable_thinking: bool = False, write_file_compatibility: bool | None = None, **kwargs) -> BaseChatModel:
     """Create a chat model instance from the config.
 
     Args:
@@ -187,10 +187,17 @@ def create_chat_model(name: str | None = None, thinking_enabled: bool = False, *
         },
     )
     configured_write_file_transport = model_config.write_file_transport
-    runtime_write_file_compatibility = get_write_file_compatibility_settings()
+    # ``None`` follows the global switch, so a roundtable run with its own
+    # switch off still uses compatibility when the global switch is on.
+    # ``True`` is the roundtable opt-in used when the global switch is off.
+    compatibility_enabled = (
+        get_write_file_compatibility_settings().enabled
+        if write_file_compatibility is None
+        else write_file_compatibility
+    )
     effective_write_file_transport = (
         WriteFileTransport.BUFFERED_GATEWAY
-        if runtime_write_file_compatibility.enabled
+        if compatibility_enabled
         else configured_write_file_transport
     )
     if model_config.tool_call_transport.value != "native":

offline-backend-20260512/backend/packages/harness/deerflow/subagents/executor.py

@@ -289,6 +289,8 @@ class SubagentExecutor:
         thread_data: ThreadDataState | None = None,
         thread_id: str | None = None,
         trace_id: str | None = None,
+        *,
+        write_file_compatibility: bool | None = None,
     ):
         """Initialize the executor.
 
@@ -317,6 +319,7 @@ class SubagentExecutor:
         self.sandbox_state = sandbox_state
         self.thread_data = thread_data
         self.thread_id = thread_id
+        self.write_file_compatibility = True if write_file_compatibility else None
         # Generate trace_id if not provided (for top-level calls)
         self.trace_id = trace_id or str(uuid.uuid4())[:8]
 
@@ -334,7 +337,12 @@ class SubagentExecutor:
         app_config = self.app_config or get_app_config()
         if self.model_name is None:
             self.model_name = resolve_subagent_model_name(self.config, self.parent_model, app_config=app_config)
-        model = create_chat_model(name=self.model_name, thinking_enabled=False, app_config=app_config)
+        model = create_chat_model(
+            name=self.model_name,
+            thinking_enabled=False,
+            app_config=app_config,
+            write_file_compatibility=self.write_file_compatibility,
+        )
 
         from deerflow.agents.middlewares.tool_error_handling_middleware import build_subagent_runtime_middlewares
 

offline-backend-20260512/backend/packages/harness/deerflow/tools/builtins/task_tool.py

@@ -157,6 +157,8 @@ async def task_tool(
     effective_model = resolve_subagent_model_name(config, parent_model, app_config=resolved_app_config)
 
     # Subagents should not have subagent tools enabled (prevent recursive nesting)
+    parent_context = runtime.context if runtime is not None and isinstance(getattr(runtime, "context", None), dict) else {}
+    roundtable_file_stream = bool(parent_context.get("roundtable_write_file_compatibility"))
     available_tools_kwargs = {
         "model_name": effective_model,
         "groups": parent_tool_groups,
@@ -164,6 +166,8 @@ async def task_tool(
     }
     if resolved_app_config is not None:
         available_tools_kwargs["app_config"] = resolved_app_config
+    if roundtable_file_stream:
+        available_tools_kwargs["write_file_compatibility"] = True
     tools = get_available_tools(**available_tools_kwargs)
 
     # Create executor
@@ -178,6 +182,8 @@ async def task_tool(
     }
     if resolved_app_config is not None:
         executor_kwargs["app_config"] = resolved_app_config
+    if roundtable_file_stream:
+        executor_kwargs["write_file_compatibility"] = True
     executor = SubagentExecutor(**executor_kwargs)
 
     # Start background execution (always async to prevent blocking)

offline-backend-20260512/backend/packages/harness/deerflow/tools/tools.py

@@ -71,6 +71,7 @@ def get_available_tools(
     allowed_tools: list[str] | None = None,
     allow_agent_orchestration: bool = False,
     app_config: AppConfig | None = None,
+    write_file_compatibility: bool | None = None,
 ) -> list[BaseTool]:
     """Get all available tools from config.
 
@@ -105,6 +106,8 @@ def get_available_tools(
     # requests entirely inside PatchedChatOpenAI.
     global_write_file_compatibility_enabled = (
         get_write_file_compatibility_settings().enabled
+        if write_file_compatibility is None
+        else write_file_compatibility
     )
     write_renderer_enabled = (
         not global_write_file_compatibility_enabled

offline-backend-20260512/backend/packages/harness/tests/test_patched_openai_write_file_text_bridge.py

@@ -4543,6 +4543,37 @@ def test_factory_global_switch_overrides_every_compatible_model(
     assert second.write_file_transport == "buffered_gateway"
 
 
+def test_roundtable_override_enables_buffered_gateway_while_global_is_off(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setattr(
+        model_factory,
+        "get_write_file_compatibility_settings",
+        lambda: SimpleNamespace(enabled=False),
+    )
+    config = AppConfig.model_construct(
+        models=[
+            ModelConfig(
+                name="native-model",
+                use="langchain_openai:ChatOpenAI",
+                model="native-model",
+                api_key="test-key",
+                base_url=None,
+            )
+        ]
+    )
+
+    ordinary = model_factory.create_chat_model("native-model", app_config=config)
+    roundtable = model_factory.create_chat_model(
+        "native-model",
+        app_config=config,
+        write_file_compatibility=True,
+    )
+
+    assert ordinary.write_file_transport == "native"
+    assert roundtable.write_file_transport == "buffered_gateway"
+
+
 def test_factory_rejects_bridge_for_non_patched_provider(monkeypatch: pytest.MonkeyPatch) -> None:
     config = AppConfig.model_construct(
         models=[

offline-backend-20260512/backend/tests/test_model_management.py

@@ -215,6 +215,121 @@ def test_reasoning_capability_route_precedes_generic_model_update_and_accepts_sl
     ]
 
 
+def test_write_file_transport_route_precedes_generic_model_update() -> None:
+    scope = {
+        "type": "http",
+        "method": "PUT",
+        "path": "/api/admin/models/write-file-transport/zai-org/GLM-5-FP8",
+    }
+    matching_paths = [route.path for route in router.router.routes if route.matches(scope)[0] is Match.FULL]
+
+    assert matching_paths[:2] == [
+        "/api/admin/models/write-file-transport/{model_name:path}",
+        "/api/admin/models/{model_name:path}",
+    ]
+
+
+def _admin_request():
+    return SimpleNamespace(
+        state=SimpleNamespace(user=SimpleNamespace(system_role="admin")),
+        app=SimpleNamespace(state=SimpleNamespace()),
+    )
+
+
+@pytest.mark.asyncio
+async def test_write_file_transport_toggle_persists_native_and_buffered_gateway(
+    tmp_path,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    config_path = tmp_path / "config.yaml"
+    config_path.write_text(
+        """models:
+  - name: flash
+    use: langchain_openai:ChatOpenAI
+    model: flash
+""",
+        encoding="utf-8",
+    )
+    saved: dict[str, object] = {}
+
+    async def save_models(_request, raw_before: str, models: list[dict[str, object]]) -> list[dict[str, object]]:
+        saved["raw_before"] = raw_before
+        saved["models"] = models
+        return models
+
+    monkeypatch.setattr(router, "_resolve_config_path", lambda: config_path)
+    monkeypatch.setattr(router, "_save_models", save_models)
+
+    enabled = await router.update_write_file_transport(
+        "flash",
+        router.WriteFileTransportToggleRequest(write_file_transport="buffered_gateway"),
+        _admin_request(),
+    )
+
+    assert enabled.write_file_transport == "buffered_gateway"
+    assert saved["models"] == [
+        {
+            "name": "flash",
+            "use": "langchain_openai:ChatOpenAI",
+            "model": "flash",
+            "write_file_transport": "buffered_gateway",
+        }
+    ]
+
+    disabled = await router.update_write_file_transport(
+        "flash",
+        router.WriteFileTransportToggleRequest(write_file_transport="native"),
+        _admin_request(),
+    )
+
+    assert disabled.write_file_transport == "native"
+    assert saved["models"][0]["write_file_transport"] == "native"
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("use", "use_responses_api"),
+    [
+        ("langchain_anthropic:ChatAnthropic", None),
+        ("langchain_openai:ChatOpenAI", True),
+    ],
+)
+async def test_write_file_transport_rejects_incompatible_models(
+    use: str,
+    use_responses_api: bool | None,
+    tmp_path,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    config_path = tmp_path / "config.yaml"
+    original = f"""models:
+  - name: incompatible
+    use: {use}
+    model: incompatible
+"""
+    if use_responses_api:
+        original = original.replace(
+            "    model: incompatible\n",
+            "    model: incompatible\n    use_responses_api: true\n",
+        )
+    config_path.write_text(original, encoding="utf-8")
+    monkeypatch.setattr(router, "_resolve_config_path", lambda: config_path)
+
+    async def fail_save_models(_request, _raw_before, _models):
+        raise AssertionError("incompatible models must not be persisted")
+
+    monkeypatch.setattr(router, "_save_models", fail_save_models)
+
+    with pytest.raises(HTTPException) as exc_info:
+        await router.update_write_file_transport(
+            "incompatible",
+            router.WriteFileTransportToggleRequest(write_file_transport="buffered_gateway"),
+            _admin_request(),
+        )
+
+    assert exc_info.value.status_code == 422
+    assert config_path.read_text(encoding="utf-8") == original
+
+
 def _reasoning_request(store, *, config: AppConfig | None = None):
     return SimpleNamespace(
         state=SimpleNamespace(

offline-backend-20260512/backend/tests/test_roundtable_room_limit.py

@@ -745,6 +745,7 @@ async def test_admin_can_change_roundtable_room_limit_without_restart(
     assert ss.get_roundtable_room_limit_settings().max_rooms == 3
     assert ss.get_roundtable_room_limit_settings().max_hold_minutes == 30
     assert ss.get_roundtable_room_limit_settings().max_cycles == 8
+    assert ss.get_roundtable_room_limit_settings().write_file_compatibility_enabled is False
 
     saved = await ss_router.update_roundtable_room_limit_setting(
         ss_router.RoundtableRoomLimitUpdateRequest(
@@ -762,6 +763,36 @@ async def test_admin_can_change_roundtable_room_limit_without_restart(
     assert ss.load_system_settings().roundtable_room_limit.max_rooms == 5
     assert ss.load_system_settings().roundtable_room_limit.max_hold_minutes == 45
     assert ss.load_system_settings().roundtable_room_limit.max_cycles == 12
+    assert saved.write_file_compatibility_enabled is False
+
+
+@pytest.mark.asyncio
+async def test_roundtable_file_stream_is_only_attached_to_roundtable_runs(
+    tmp_runtime_home: Path,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    _patch_admin(monkeypatch)
+    base = dict(
+        agent_name="roundtable-seat",
+        model_name="test-model",
+        thread_id="thread-1",
+        new_message="test",
+        excluded_tools=[],
+        skill_stop_names=[],
+    )
+    before = _run_payload(**base, roundtable_file_stream=True)
+    assert "roundtable_write_file_compatibility" not in before["context"]
+
+    saved = await ss_router.update_roundtable_room_limit_setting(
+        ss_router.RoundtableRoomLimitUpdateRequest(write_file_compatibility_enabled=True),
+        SimpleNamespace(),
+    )
+    assert saved.write_file_compatibility_enabled is True
+
+    opted_in = _run_payload(**base, roundtable_file_stream=True)
+    assert opted_in["context"]["roundtable_write_file_compatibility"] is True
+    ordinary = _run_payload(**base)
+    assert "roundtable_write_file_compatibility" not in ordinary["context"]
 
 
 @pytest.mark.asyncio

offline-backend-20260512/backend/tests/test_title_middleware.py

@@ -114,6 +114,46 @@ def _title_config() -> TitleConfig:
     )
 
 
+def test_admin_question_title_switch_finalizes_question_before_first_model_call(monkeypatch) -> None:
+    monkeypatch.setattr(
+        title_middleware_module,
+        "get_conversation_context_settings",
+        lambda: SimpleNamespace(use_question_as_title=True),
+    )
+    middleware = TitleMiddleware(
+        title_config=_title_config().model_copy(update={"enabled": False})
+    )
+    state = {
+        "messages": [SimpleNamespace(type="human", content="请帮我分析这份合同")],
+    }
+
+    result = middleware.before_model(state, runtime=SimpleNamespace())
+
+    assert result == {
+        "title": "请帮我分析这份合同",
+        "title_provisional": False,
+    }
+
+
+def test_admin_question_title_switch_off_keeps_async_title_generation(monkeypatch) -> None:
+    monkeypatch.setattr(
+        title_middleware_module,
+        "get_conversation_context_settings",
+        lambda: SimpleNamespace(use_question_as_title=False),
+    )
+    middleware = TitleMiddleware(title_config=_title_config())
+    state = {
+        "messages": [SimpleNamespace(type="human", content="请帮我分析这份合同")],
+    }
+
+    result = middleware.before_model(state, runtime=SimpleNamespace())
+
+    assert result == {
+        "title": "请帮我分析这份合同",
+        "title_provisional": True,
+    }
+
+
 @pytest.mark.asyncio
 async def test_finalize_thread_title_updates_checkpoint_and_display_name(monkeypatch) -> None:
     async def _fake_generate(self, state):

offline-backend-20260512/config.yaml

@@ -493,4 +493,3 @@ task_deeplink:
         taskDirection: "正向引导 + 风险预警"
         taskName: "任务01：暑期文旅舆情综合研判"
         taskContent: "围绕暑期文旅消费高峰，对重点城市、重点景区的公众情绪与舆情态势做综合研判，识别情绪拐点与潜在风险，提出正向引导与处置建议。"
-        taskType: 1