代码提交

Newer vLLM parsers put CoT in reasoning; older builds and DeepSeek still use reasoning_content. Canonicalize to the UI field and echo only the original wire field so other ChatOpenAI models stay unchanged. Point qwen3.6 at VllmChatModel.

Co-authored-by: Cursor <cursoragent@cursor.com>

.gitignore

@@ -100,6 +100,8 @@ offline-backend-20260512/backend/tests/*
 # Keep the model configuration regression coverage versioned even though
 # this deployment repository otherwise ignores newly added backend tests.
 !offline-backend-20260512/backend/tests/test_model_management.py
+# Keep vLLM reasoning/reasoning_content field compatibility regressions versioned.
+!offline-backend-20260512/backend/tests/test_reasoning_fields.py
 # Keep the forced-research agent runtime gate and sandbox-policy regressions versioned.
 !offline-backend-20260512/backend/tests/test_forced_research_agent.py
 # Keep fixed-question exact-match and streaming regression coverage versioned.

frontend-web/.env.example

@@ -59,6 +59,9 @@
 # 简介模式左侧菜单类型：与菜单管理页「简介模式」的类型 A / B 对应。两套网页分别填 A 或 B。
 # 不填或留空时按类型 A，继续使用原来的简介菜单数据。
 # VITE_COMPACT_MENU_TYPE=A
+# 问答页「极速」模式：true=在模式菜单显示并直连大模型；false 或未设=隐藏。
+# 部署后优先改 dist/runtime-config.js 的 VITE_TURBO_MODE_VISIBLE，刷新即可。
+# VITE_TURBO_MODE_VISIBLE=false
 # VITE_GITHUB_OAUTH_TOKEN=
 
 # --- 官方 DeerFlow 风格页面入口（仅 Vite，不引入 Next.js）---

frontend-web/public/runtime-config.js

@@ -151,6 +151,11 @@ window.__CM_RUNTIME_CONFIG__ = {
   // GitHub OAuth Token；不使用 GitHub 相关能力时留空。
   VITE_GITHUB_OAUTH_TOKEN: "",
 
+  // 是否在问答输入框「模式」菜单显示「极速」："true" 显示，"false" 隐藏。
+  // 默认关闭。开启后可直连大模型流式问答，不走智能体工具与技能。
+  // 部署后改 dist/runtime-config.js 再刷新页面即可，无需重新打包。
+  VITE_TURBO_MODE_VISIBLE: "true",
+
   // 是否在智能体管理页显示「舆情分析」类型："true" 显示，"false" 隐藏。
   // 关闭后，该智能体的专用对话地址也会自动跳回智能体管理页。
   VITE_SENTIMENT_AGENT_VISIBLE: "true",

frontend-web/src/components/workspace/input-box.tsx

@@ -7,6 +7,7 @@ import {
   CheckIcon,
   FileTextIcon,
   FolderTreeIcon,
+  GaugeIcon,
   GraduationCapIcon,
   LightbulbIcon,
   PaperclipIcon,
@@ -68,6 +69,7 @@ import { isHiddenFromUIMessage } from "@/core/messages/utils";
 import { useModels } from "@/core/models/hooks";
 import type { Model } from "@/core/models/types";
 import type { AgentThreadContext } from "@/core/threads";
+import { isTurboModeVisible } from "@/core/threads/use-direct-chat";
 import { textOfMessage } from "@/core/threads/utils";
 import { desensitize } from "@/lib/desensitize";
 import { cn } from "@/lib/utils";
@@ -223,7 +225,11 @@ export function InputBox({
     context.execution_profile === "single_round_rag";
   const multiRoundRagActive =
     context.execution_profile === "multi_round_rag";
+  const turboModeVisible = isTurboModeVisible();
+  const directChatActive =
+    turboModeVisible && context.execution_profile === "direct";
   const frontendRagActive = singleRoundRagActive || multiRoundRagActive;
+  const bypassAgentActive = frontendRagActive || directChatActive;
   const modelOptions = useMemo(() => {
     return singleRoundRagActive ? extraModels : models;
   }, [extraModels, models, singleRoundRagActive]);
@@ -253,11 +259,37 @@ export function InputBox({
     recommendedQuestions.length > 0;
 
   useEffect(() => {
-    // Multi-round RAG is a frontend execution profile, not a model option.
+    // Frontend-only execution profiles are not model options.
-    // Do not let the normal model fallback clear it on every render.
+    // Do not let the normal model fallback clear them on every render.
+    if (!turboModeVisible && context.execution_profile === "direct") {
+      onContextChange?.({
+        ...context,
+        execution_profile: undefined,
+      });
+      return;
+    }
     if (multiRoundRagActive) {
       return;
     }
+    if (directChatActive) {
+      if (models.length === 0) {
+        return;
+      }
+      const currentModel = models.find((m) => m.name === context.model_name);
+      const fallbackModel = currentModel ?? models[0]!;
+      if (
+        context.model_name === fallbackModel.name &&
+        context.execution_profile === "direct"
+      ) {
+        return;
+      }
+      onContextChange?.({
+        ...context,
+        model_name: fallbackModel.name,
+        execution_profile: "direct",
+      });
+      return;
+    }
     if (modelOptions.length === 0) {
       return;
     }
@@ -281,7 +313,7 @@ export function InputBox({
       mode: nextMode,
       execution_profile: fallbackModel.execution_profile,
     });
-  }, [context, modelOptions, multiRoundRagActive, onContextChange]);
+  }, [context, directChatActive, modelOptions, models, multiRoundRagActive, onContextChange, turboModeVisible]);
 
   const selectedModel = useMemo(() => {
     if (modelOptions.length === 0) {
@@ -314,7 +346,10 @@ export function InputBox({
         model_name,
         mode: getResolvedMode(context.mode, model.supports_thinking ?? false),
         reasoning_effort: context.reasoning_effort,
-        execution_profile: model.execution_profile,
+        execution_profile:
+          context.execution_profile === "direct"
+            ? "direct"
+            : model.execution_profile,
       });
       setModelDialogOpen(false);
     },
@@ -346,6 +381,16 @@ export function InputBox({
     [onContextChange, context, models, supportThinking],
   );
 
+  const handleDirectChatSelect = useCallback(() => {
+    const ordinaryModel =
+      models.find((model) => model.name === context.model_name) ?? models[0];
+    onContextChange?.({
+      ...context,
+      model_name: ordinaryModel?.name ?? context.model_name,
+      execution_profile: "direct",
+    });
+  }, [context, models, onContextChange]);
+
   const handleSingleRoundRagSelect = useCallback(() => {
     const ragModel =
       extraModels.find((model) => model.name === context.model_name) ??
@@ -749,7 +794,7 @@ export function InputBox({
               />
             </PromptInputActionMenuContent>
           </PromptInputActionMenu> */}
-            {!frontendRagActive && (
+            {!bypassAgentActive && (
               <AddAttachmentsButton
                 className={cn(
                   "rounded-full border-0 bg-transparent px-2.5! shadow-none 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",
@@ -758,7 +803,7 @@ export function InputBox({
                 )}
               />
             )}
-            {!frontendRagActive && !hideContextTools && <Tooltip
+            {!bypassAgentActive && !hideContextTools && <Tooltip
               content={
                 context.writing_mode
                   ? "写作模式已开启：报告由深度研究管线按「通用分析结构」撰写，优先使用对话中已检索的资料"
@@ -784,7 +829,7 @@ export function InputBox({
                 <span className="text-xs font-normal">写作</span>
               </PromptInputButton>
             </Tooltip>}
-            {!frontendRagActive && <Tooltip
+            {!bypassAgentActive && <Tooltip
               content={
                 artifactDirectoryModeEnabled
                   ? "文件模式已开启：右侧会显示当前对话生成的全部文件目录"
@@ -813,7 +858,7 @@ export function InputBox({
                 <span className="text-xs font-normal">文件</span>
               </PromptInputButton>
             </Tooltip>}
-            {!frontendRagActive && !hideContextTools && !hideMemoryTool && <Tooltip
+            {!bypassAgentActive && !hideContextTools && !hideMemoryTool && <Tooltip
               content={
                 memoryInjectionEnabled
                   ? "记忆注入已开启：对话前会注入历史记忆"
@@ -839,7 +884,7 @@ export function InputBox({
                 <span className="text-xs font-normal">记忆</span>
               </PromptInputButton>
             </Tooltip>}
-            {referenceModeEntryEnabled && !frontendRagActive && !hideContextTools && <Tooltip
+            {referenceModeEntryEnabled && !bypassAgentActive && !hideContextTools && <Tooltip
               content={
                 ragModeEnabled
                   ? "参考文献模式已开启：检索类结果会按编号引用展示"
@@ -869,7 +914,7 @@ export function InputBox({
                 <span className="text-xs font-normal">参考文献</span>
               </PromptInputButton>
             </Tooltip>}
-            {!frontendRagActive && <Tooltip
+            {!bypassAgentActive && <Tooltip
               content={
                 contextCompressionEnabled
                   ? "上下文压缩已开启：超长对话会自动压缩历史以节省 token"
@@ -909,7 +954,9 @@ export function InputBox({
                     : "flash"
                 }
                 tooltipContent={
-                  singleRoundRagActive
+                  directChatActive
+                    ? `${t.inputBox.turboMode}：${t.inputBox.turboModeDescription}`
+                    : singleRoundRagActive
                     ? "单轮 RAG 问答：基于知识库检索生成答案，并展示参考资料"
                     : multiRoundRagActive
                       ? "多轮 RAG 问答：多次检索并展示检索过程与参考资料"
@@ -921,7 +968,9 @@ export function InputBox({
                     "gap-1.5! rounded-full border-0 bg-transparent px-3! text-muted-foreground shadow-none hover:bg-muted/70 hover:text-foreground dark:text-slate-400 dark:hover:bg-slate-700/70 dark:hover:text-slate-100",
                     compact && "rounded-lg border px-2.5! py-1! text-[11px] font-medium",
                     compact &&
-                      (frontendRagActive
+                      (directChatActive
+                        ? "border-sky-200 bg-sky-50/60 text-sky-700 hover:bg-sky-100 dark:border-sky-700 dark:bg-sky-950/40 dark:text-sky-200"
+                        : frontendRagActive
                         ? "border-emerald-200 bg-emerald-50/60 text-emerald-700 hover:bg-emerald-100 dark:border-emerald-700 dark:bg-emerald-950/40 dark:text-emerald-200"
                         : context.mode === "thinking"
                         ? "border-indigo-200 bg-indigo-50/60 text-indigo-700 hover:bg-indigo-100 dark:border-indigo-700 dark:bg-indigo-950/40 dark:text-indigo-200"
@@ -933,7 +982,9 @@ export function InputBox({
                   )}
                   >
                     <div>
-                    {frontendRagActive ? (
+                    {directChatActive ? (
+                      <GaugeIcon className="size-3" />
+                    ) : frontendRagActive ? (
                       <BookMarkedIcon className="size-3" />
                     ) : context.mode === "flash" ? (
                       <ZapIcon className="size-3" />
@@ -948,10 +999,14 @@ export function InputBox({
                   <div
                     className={cn(
                       "text-xs font-normal",
-                      context.mode === "ultra" ? "golden-text" : "",
+                      context.mode === "ultra" && !directChatActive && !frontendRagActive
+                        ? "golden-text"
+                        : "",
                     )}
                   >
-                    {frontendRagActive
+                    {directChatActive
+                      ? t.inputBox.turboMode
+                      : frontendRagActive
                       ? multiRoundRagActive
                         ? "多轮 RAG"
                         : "单轮 RAG"
@@ -969,9 +1024,39 @@ export function InputBox({
                     {t.inputBox.mode}
                   </DropdownMenuLabel>
                   <PromptInputActionMenu>
+                    {turboModeVisible && (
                     <PromptInputActionMenuItem
                       className={cn(
-                        !frontendRagActive && context.mode === "flash"
+                        directChatActive
+                          ? "bg-sky-50 text-sky-800 hover:bg-sky-100 focus:bg-sky-100 dark:bg-sky-950/40 dark:text-sky-200 dark:hover:bg-sky-950/60 dark:focus:bg-sky-950/60"
+                          : "text-muted-foreground/65",
+                      )}
+                      onSelect={handleDirectChatSelect}
+                    >
+                      <div className="flex flex-col gap-2">
+                        <div className="flex items-center gap-1 font-bold">
+                          <GaugeIcon
+                            className={cn(
+                              "mr-2 size-4",
+                              directChatActive && "text-sky-700 dark:text-sky-200",
+                            )}
+                          />
+                          {t.inputBox.turboMode}
+                        </div>
+                        <div className="pl-7 text-xs">
+                          {t.inputBox.turboModeDescription}
+                        </div>
+                      </div>
+                      {directChatActive ? (
+                        <CheckIcon className="ml-auto size-4" />
+                      ) : (
+                        <div className="ml-auto size-4" />
+                      )}
+                    </PromptInputActionMenuItem>
+                    )}
+                    <PromptInputActionMenuItem
+                      className={cn(
+                        !bypassAgentActive && context.mode === "flash"
                           ? "text-accent-foreground"
                           : "text-muted-foreground/65",
                       )}
@@ -982,7 +1067,7 @@ export function InputBox({
                           <ZapIcon
                             className={cn(
                               "mr-2 size-4",
-                              !frontendRagActive && context.mode === "flash" &&
+                              !bypassAgentActive && context.mode === "flash" &&
                                 "text-accent-foreground",
                             )}
                           />
@@ -992,7 +1077,7 @@ export function InputBox({
                           {t.inputBox.flashModeDescription}
                         </div>
                       </div>
-                      {!frontendRagActive && context.mode === "flash" ? (
+                      {!bypassAgentActive && context.mode === "flash" ? (
                         <CheckIcon className="ml-auto size-4" />
                       ) : (
                         <div className="ml-auto size-4" />
@@ -1001,7 +1086,7 @@ export function InputBox({
                     {supportThinking && (
                       <PromptInputActionMenuItem
                         className={cn(
-                          !frontendRagActive && context.mode === "thinking"
+                          !bypassAgentActive && context.mode === "thinking"
                             ? "text-accent-foreground"
                             : "text-muted-foreground/65",
                         )}
@@ -1012,7 +1097,7 @@ export function InputBox({
                             <LightbulbIcon
                               className={cn(
                                 "mr-2 size-4",
-                                !frontendRagActive && context.mode === "thinking" &&
+                                !bypassAgentActive && context.mode === "thinking" &&
                                   "text-accent-foreground",
                               )}
                             />
@@ -1022,7 +1107,7 @@ export function InputBox({
                             {t.inputBox.reasoningModeDescription}
                           </div>
                         </div>
-                        {!frontendRagActive && context.mode === "thinking" ? (
+                        {!bypassAgentActive && context.mode === "thinking" ? (
                           <CheckIcon className="ml-auto size-4" />
                         ) : (
                           <div className="ml-auto size-4" />
@@ -1031,7 +1116,7 @@ export function InputBox({
                     )}
                     <PromptInputActionMenuItem
                       className={cn(
-                        !frontendRagActive && context.mode === "pro"
+                        !bypassAgentActive && context.mode === "pro"
                           ? "text-accent-foreground"
                           : "text-muted-foreground/65",
                       )}
@@ -1042,7 +1127,7 @@ export function InputBox({
                           <GraduationCapIcon
                             className={cn(
                               "mr-2 size-4",
-                              !frontendRagActive && context.mode === "pro" &&
+                              !bypassAgentActive && context.mode === "pro" &&
                                 "text-accent-foreground",
                             )}
                           />
@@ -1052,7 +1137,7 @@ export function InputBox({
                           {t.inputBox.proModeDescription}
                         </div>
                       </div>
-                      {!frontendRagActive && context.mode === "pro" ? (
+                      {!bypassAgentActive && context.mode === "pro" ? (
                         <CheckIcon className="ml-auto size-4" />
                       ) : (
                         <div className="ml-auto size-4" />
@@ -1060,7 +1145,7 @@ export function InputBox({
                     </PromptInputActionMenuItem>
                     <PromptInputActionMenuItem
                       className={cn(
-                        !frontendRagActive && context.mode === "ultra"
+                        !bypassAgentActive && context.mode === "ultra"
                           ? "text-accent-foreground"
                           : "text-muted-foreground/65",
                       )}
@@ -1071,14 +1156,14 @@ export function InputBox({
                           <RocketIcon
                             className={cn(
                               "mr-2 size-4",
-                              !frontendRagActive &&
+                              !bypassAgentActive &&
                                 context.mode === "ultra" &&
                                 "text-[#dabb5e]",
                             )}
                           />
                           <div
                             className={cn(
-                              !frontendRagActive &&
+                              !bypassAgentActive &&
                                 context.mode === "ultra" &&
                                 "golden-text",
                             )}
@@ -1090,7 +1175,7 @@ export function InputBox({
                           {t.inputBox.ultraModeDescription}
                         </div>
                       </div>
-                      {!frontendRagActive && context.mode === "ultra" ? (
+                      {!bypassAgentActive && context.mode === "ultra" ? (
                         <CheckIcon className="ml-auto size-4" />
                       ) : (
                         <div className="ml-auto size-4" />
@@ -1161,7 +1246,7 @@ export function InputBox({
                 </DropdownMenuGroup>
               </PromptInputActionMenuContent>
             </PromptInputActionMenu>
-            {!frontendRagActive && supportReasoningEffort && context.mode !== "flash" && (
+            {!bypassAgentActive && supportReasoningEffort && context.mode !== "flash" && (
               <PromptInputActionMenu>
                 <PromptInputActionMenuTrigger className={cn("gap-1.5! rounded-full border-0 bg-transparent px-3! text-muted-foreground shadow-none hover:bg-muted/70 hover:text-foreground dark:text-slate-400 dark:hover:bg-slate-700/70 dark:hover:text-slate-100", compact && "rounded-lg border border-border px-2.5! py-1! text-[13px]")}>
                   <div className="text-xs font-normal">
@@ -1279,7 +1364,7 @@ export function InputBox({
             )}
           </PromptInputTools>
           <PromptInputTools className={cn("gap-1.5", compact && "gap-2")}>
-            {!frontendRagActive && agentSelector && (
+            {!bypassAgentActive && agentSelector && (
               <AgentSelector
                 selectedAgentId={agentSelector.selectedAgentId}
                 onAgentChange={agentSelector.onAgentChange}
@@ -1287,8 +1372,8 @@ export function InputBox({
                 compact={compact}
               />
             )}
-            {!frontendRagActive && knowledgeSelector}
+            {!bypassAgentActive && knowledgeSelector}
-            {!frontendRagActive && promptPrefix && (
+            {!bypassAgentActive && promptPrefix && (
               <Tooltip
                 content={
                   promptPrefix.on

frontend-web/src/core/i18n/locales/en-US.ts

@@ -96,6 +96,9 @@ export const enUS: Translations = {
     mode: "Mode",
     flashMode: "Flash",
     flashModeDescription: "Fast and efficient, but may not be accurate",
+    turboMode: "Turbo",
+    turboModeDescription:
+      "Talk to the model directly with no agent tools or skills, for lower first-token latency",
     reasoningMode: "Reasoning",
     reasoningModeDescription:
       "Reasoning before action, balance between time and accuracy",

frontend-web/src/core/i18n/locales/types.ts

@@ -91,6 +91,8 @@ export interface Translations {
     mode: string;
     flashMode: string;
     flashModeDescription: string;
+    turboMode: string;
+    turboModeDescription: string;
     reasoningMode: string;
     reasoningModeDescription: string;
     proMode: string;

frontend-web/src/core/i18n/locales/zh-CN.ts

@@ -95,6 +95,8 @@ export const zhCN: Translations = {
     mode: "模式",
     flashMode: "快速",
     flashModeDescription: "快速且高效的完成任务，但可能不够精准",
+    turboMode: "极速",
+    turboModeDescription: "直连大模型流式问答，不走智能体工具与技能，首字更快",
     reasoningMode: "思考",
     reasoningModeDescription: "思考后再行动，在时间与准确性之间取得平衡",
     proMode: "专业问答",

frontend-web/src/core/messages/utils.ts

@@ -364,11 +364,16 @@ export function extractReasoningContentFromMessage(message: Message) {
   if (message.type !== "ai") {
     return null;
   }
-  if (
+  if (message.additional_kwargs) {
-    message.additional_kwargs &&
+    const reasoningContent = message.additional_kwargs.reasoning_content;
-    "reasoning_content" in message.additional_kwargs
+    if (typeof reasoningContent === "string") {
-  ) {
+      return reasoningContent;
-    return message.additional_kwargs.reasoning_content as string | null;
+    }
+    // Newer vLLM puts thinking in `reasoning` instead of `reasoning_content`.
+    const reasoning = message.additional_kwargs.reasoning;
+    if (typeof reasoning === "string") {
+      return reasoning;
+    }
   }
   if (Array.isArray(message.content)) {
     const part = message.content[0];
@@ -424,6 +429,9 @@ export function hasReasoning(message: Message) {
   if (typeof message.additional_kwargs?.reasoning_content === "string") {
     return true;
   }
+  if (typeof message.additional_kwargs?.reasoning === "string") {
+    return true;
+  }
   if (Array.isArray(message.content)) {
     const part = message.content[0];
     // Compatible with the Anthropic gateway

frontend-web/src/core/models/types.ts

@@ -7,7 +7,7 @@ export interface Model {
   supports_thinking?: boolean;
   supports_reasoning_effort?: boolean;
   /** Execution profiles handled by the frontend instead of DeerFlow. */
-  execution_profile?: "single_round_rag" | "multi_round_rag";
+  execution_profile?: "single_round_rag" | "multi_round_rag" | "direct";
 }
 
 export interface TokenUsageSettings {

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

@@ -1579,6 +1579,7 @@ export function useThreadStream({
             context: {
               ...mergedContext,
               thinking_enabled: mergedContext.mode !== "flash",
+              thinking_force_disabled: mergedContext.mode === "flash",
               is_plan_mode:
                 mergedContext.mode === "pro" || mergedContext.mode === "ultra",
               subagent_enabled: mergedContext.mode === "ultra",

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

@@ -14,6 +14,8 @@ export interface AgentThreadContext extends Record<string, unknown> {
   thread_id: string;
   model_name: string | undefined;
   thinking_enabled: boolean;
+  /** Flash mode: force vLLM/Qwen ``enable_thinking=false`` even without a per-model disable shape. */
+  thinking_force_disabled?: boolean;
   is_plan_mode: boolean;
   subagent_enabled: boolean;
   reasoning_effort?: "minimal" | "low" | "medium" | "high";
@@ -29,8 +31,8 @@ export interface AgentThreadContext extends Record<string, unknown> {
   llmwiki_knowledge_base_ids?: string[];
   /** Assistant knowledge-base ids selected for this conversation. */
   assistant_knowledge_base_ids?: string[];
-  /** Direct RAG modes handled in the frontend; these requests never enter Lead Agent. */
+  /** Frontend-only execution paths that never enter Lead Agent (RAG / 极速直连). */
-  execution_profile?: "single_round_rag" | "multi_round_rag";
+  execution_profile?: "single_round_rag" | "multi_round_rag" | "direct";
   agent_id?: string;
   /** @deprecated kept for reading legacy thread metadata; prefer agent_id */
   agent_name?: string;

frontend-web/src/core/threads/use-direct-chat.ts

@@ -0,0 +1,312 @@
+import type { Message } from "@langchain/langgraph-sdk";
+import { useCallback, useEffect, useRef, useState } from "react";
+
+import { apiFetch } from "@/core/api/fetch-client";
+import { getBackendBaseURL } from "@/core/config";
+import { readRuntimeFlag } from "@/env";
+import { createClientId } from "@/lib/client-id";
+
+import { textOfMessage } from "./utils";
+
+const STORAGE_PREFIX = "workspace.direct-chat.";
+const MAX_HISTORY_MESSAGES = 20;
+
+export function isDirectChatProfile(context: unknown): boolean {
+  return (
+    typeof context === "object" &&
+    context !== null &&
+    (context as { execution_profile?: unknown }).execution_profile === "direct"
+  );
+}
+
+/** 是否在模式菜单展示「极速」。仅 runtime-config.js 显式为 "true" 时开启，默认关闭。 */
+export function isTurboModeVisible(): boolean {
+  return readRuntimeFlag("VITE_TURBO_MODE_VISIBLE");
+}
+
+function storageKey(threadId: string) {
+  return `${STORAGE_PREFIX}${threadId}`;
+}
+
+function readStoredMessages(threadId: string): Message[] {
+  if (typeof window === "undefined" || !threadId) return [];
+  try {
+    const raw = window.sessionStorage.getItem(storageKey(threadId));
+    const parsed: unknown = raw ? JSON.parse(raw) : [];
+    return Array.isArray(parsed) ? (parsed as Message[]) : [];
+  } catch {
+    return [];
+  }
+}
+
+function writeStoredMessages(threadId: string, messages: Message[]) {
+  if (typeof window === "undefined" || !threadId) return;
+  try {
+    window.sessionStorage.setItem(storageKey(threadId), JSON.stringify(messages));
+  } catch {
+    // Quota errors must never interrupt an already-running answer.
+  }
+}
+
+function isAbortError(error: unknown): boolean {
+  return error instanceof DOMException
+    ? error.name === "AbortError"
+    : error instanceof Error && error.name === "AbortError";
+}
+
+function historyFromMessages(messages: Message[]): Array<{
+  role: "user" | "assistant";
+  content: string;
+}> {
+  const history: Array<{ role: "user" | "assistant"; content: string }> = [];
+  for (const message of messages) {
+    if (message.type !== "human" && message.type !== "ai") continue;
+    const content = textOfMessage(message)?.trim();
+    if (!content) continue;
+    history.push({
+      role: message.type === "human" ? "user" : "assistant",
+      content,
+    });
+  }
+  return history.slice(-MAX_HISTORY_MESSAGES);
+}
+
+function parseSseBlocks(buffer: string): {
+  events: Array<{ event?: string; data: string }>;
+  rest: string;
+} {
+  const parts = buffer.split("\n\n");
+  const rest = parts.pop() ?? "";
+  const events: Array<{ event?: string; data: string }> = [];
+  for (const part of parts) {
+    let event: string | undefined;
+    const dataLines: string[] = [];
+    for (const line of part.split("\n")) {
+      if (line.startsWith(":")) continue;
+      if (line.startsWith("event:")) event = line.slice(6).trim();
+      else if (line.startsWith("data:")) dataLines.push(line.slice(5).trimStart());
+    }
+    if (dataLines.length > 0) {
+      events.push({ event, data: dataLines.join("\n") });
+    }
+  }
+  return { events, rest };
+}
+
+async function readErrorDetail(response: Response): Promise<string> {
+  try {
+    const payload: unknown = await response.json();
+    if (
+      payload &&
+      typeof payload === "object" &&
+      "detail" in payload &&
+      typeof (payload as { detail: unknown }).detail === "string"
+    ) {
+      return (payload as { detail: string }).detail;
+    }
+  } catch {
+    // Fall through to status text.
+  }
+  return `极速问答失败（${response.status}）`;
+}
+
+/**
+ * Frontend-only adapter that talks to Gateway ``POST /api/direct-chat/stream``.
+ * It never submits a LangGraph run or touches the lead-agent stack.
+ */
+export function useDirectChatMessages(threadId: string | undefined) {
+  const [messages, setMessages] = useState<Message[]>(() =>
+    readStoredMessages(threadId ?? ""),
+  );
+  const messagesRef = useRef(messages);
+  const [isLoading, setIsLoading] = useState(false);
+  const [error, setError] = useState<string | null>(null);
+  const controllerRef = useRef<AbortController | null>(null);
+
+  useEffect(() => {
+    const stored = readStoredMessages(threadId ?? "");
+    messagesRef.current = stored;
+    setMessages(stored);
+    setError(null);
+  }, [threadId]);
+
+  useEffect(() => {
+    messagesRef.current = messages;
+  }, [messages]);
+
+  useEffect(() => () => controllerRef.current?.abort(), []);
+
+  const updateMessages = useCallback(
+    (targetThreadId: string, updater: (current: Message[]) => Message[]) => {
+      setMessages((current) => {
+        const next = updater(current);
+        messagesRef.current = next;
+        writeStoredMessages(targetThreadId, next);
+        return next;
+      });
+    },
+    [],
+  );
+
+  const stop = useCallback(() => controllerRef.current?.abort(), []);
+
+  const submit = useCallback(
+    async ({
+      targetThreadId,
+      query,
+      modelName,
+    }: {
+      targetThreadId: string;
+      query: string;
+      modelName?: string;
+    }) => {
+      if (isLoading) return;
+      const question = query.trim();
+      if (!question) return;
+
+      const humanId = createClientId("direct-human");
+      const assistantId = createClientId("direct-ai");
+      const clientTs = new Date().toISOString();
+      const history = historyFromMessages(messagesRef.current);
+      const userMessage: Message = {
+        id: humanId,
+        type: "human",
+        content: [{ type: "text", text: question }],
+        additional_kwargs: { client_ts: clientTs, direct_chat: true },
+      };
+      const initialAssistant: Message = {
+        id: assistantId,
+        type: "ai",
+        content: "",
+        additional_kwargs: { direct_chat: true },
+      };
+      updateMessages(targetThreadId, (current) => [
+        ...current,
+        userMessage,
+        initialAssistant,
+      ]);
+      setError(null);
+      setIsLoading(true);
+
+      const controller = new AbortController();
+      controllerRef.current = controller;
+      let answer = "";
+
+      const refreshAssistant = (content: string) => {
+        updateMessages(targetThreadId, (current) =>
+          current.map((item) =>
+            item.id === assistantId
+              ? ({
+                  ...item,
+                  content,
+                  additional_kwargs: {
+                    ...item.additional_kwargs,
+                    direct_chat: true,
+                  },
+                } as Message)
+              : item,
+          ),
+        );
+      };
+
+      try {
+        const response = await apiFetch(
+          `${getBackendBaseURL()}/api/direct-chat/stream`,
+          {
+            method: "POST",
+            headers: { "Content-Type": "application/json" },
+            signal: controller.signal,
+            body: JSON.stringify({
+              model_name: modelName,
+              messages: [...history, { role: "user", content: question }],
+            }),
+          },
+        );
+        if (!response.ok) {
+          throw new Error(await readErrorDetail(response));
+        }
+        if (!response.body) {
+          throw new Error("极速问答未返回数据流。");
+        }
+
+        const reader = response.body.getReader();
+        const decoder = new TextDecoder();
+        let buffer = "";
+        while (true) {
+          const { done, value } = await reader.read();
+          if (done) break;
+          buffer += decoder.decode(value, { stream: true });
+          const parsed = parseSseBlocks(buffer);
+          buffer = parsed.rest;
+          for (const block of parsed.events) {
+            let payload: unknown = null;
+            try {
+              payload = JSON.parse(block.data);
+            } catch {
+              continue;
+            }
+            if (block.event === "error") {
+              const detail =
+                payload &&
+                typeof payload === "object" &&
+                "detail" in payload &&
+                typeof (payload as { detail: unknown }).detail === "string"
+                  ? (payload as { detail: string }).detail
+                  : "极速问答失败。";
+              throw new Error(detail);
+            }
+            if (
+              block.event === "delta" &&
+              payload &&
+              typeof payload === "object" &&
+              "text" in payload &&
+              typeof (payload as { text: unknown }).text === "string"
+            ) {
+              answer += (payload as { text: string }).text;
+              refreshAssistant(answer);
+            }
+            if (
+              block.event === "result" &&
+              payload &&
+              typeof payload === "object" &&
+              "answer" in payload &&
+              typeof (payload as { answer: unknown }).answer === "string"
+            ) {
+              answer = (payload as { answer: string }).answer;
+              refreshAssistant(answer);
+            }
+          }
+        }
+        if (!answer) {
+          throw new Error("模型未返回任何文本。");
+        }
+      } catch (caught) {
+        if (!isAbortError(caught)) {
+          const message =
+            caught instanceof Error ? caught.message : "极速问答失败。";
+          setError(message);
+          updateMessages(targetThreadId, (current) =>
+            current.map((item) =>
+              item.id === assistantId
+                ? ({
+                    ...item,
+                    content: answer || message,
+                    additional_kwargs: {
+                      ...item.additional_kwargs,
+                      direct_chat: true,
+                    },
+                  } as Message)
+                : item,
+            ),
+          );
+        }
+      } finally {
+        if (controllerRef.current === controller) controllerRef.current = null;
+        setIsLoading(false);
+      }
+    },
+    [isLoading, updateMessages],
+  );
+
+  return { messages, isLoading, error, submit, stop };
+}

frontend-web/src/pages/AgentChatPage.tsx

@@ -13,6 +13,7 @@ import { useLocation } from "react-router-dom";
 import { useRouteBase } from "@/core/route-base";
 import { useCallback, useEffect, useMemo, useRef, useState, type ReactNode } from "react";
 import { useQuery } from "@tanstack/react-query";
+import { toast } from "sonner";
 
 import type { PromptInputMessage } from "@/components/ai-elements/prompt-input";
 import { Button } from "@/components/ui/button";
@@ -48,6 +49,7 @@ import { useNotification } from "@/core/notification/hooks";
 import { saveLastAgentId, useThreadSettings } from "@/core/settings";
 import { stopActiveThreadBeforeClarificationReply, stopThreadRun, useThreadStream } from "@/core/threads/hooks";
 import { useResearchWritingFeed } from "@/core/threads/use-research-writing-feed";
+import { useDirectChatMessages, isDirectChatProfile, isTurboModeVisible } from "@/core/threads/use-direct-chat";
 import { clearResearchReportStream } from "@/core/threads/use-research-report-stream";
 import { textOfMessage } from "@/core/threads/utils";
 import { resolveWritingModeReportPath } from "@/core/threads/writing-mode";
@@ -560,9 +562,26 @@ export default function AgentChatPage({
     },
   });
 
+  const directChat = useDirectChatMessages(threadId);
+  const directChatSelected =
+    isDirectChatProfile(settings.context) && isTurboModeVisible();
+
   // 写作模式深度研究进度：注入 displayThread 尾部（进度步骤条 + report.md
   // 流式 stub），成功且真实产物落地后自动撤出。
-  const { displayThread } = useResearchWritingFeed(threadId, thread);
+  const { displayThread: researchDisplayThread } = useResearchWritingFeed(threadId, thread);
+  const displayThread = useMemo(() => {
+    const base = researchDisplayThread;
+    if (directChat.messages.length === 0 && !directChat.isLoading) {
+      return base;
+    }
+    const messages = [...base.messages, ...directChat.messages];
+    return {
+      ...base,
+      values: { ...base.values, messages },
+      messages,
+      isLoading: base.isLoading || directChat.isLoading,
+    } as typeof base;
+  }, [directChat.isLoading, directChat.messages, researchDisplayThread]);
 
   // ── agentfx「入库」协助卡 ──────────────────────────────────────────────────
   // 任务工作区（taskBind）里，内置研判智能体写出的 task-reports.json 产物落地且
@@ -743,6 +762,62 @@ export default function AgentChatPage({
 
   const handleSubmit = useCallback(
     (message: PromptInputMessage, extraContext?: Record<string, unknown>) => {
+      if (directChatSelected) {
+        if (message.files && message.files.length > 0) {
+          toast.error("极速问答暂不支持附件，请移除附件后重试。");
+          return;
+        }
+        setHasStartedConversation(true);
+        saveLastAgentId(agent_id);
+        window.dispatchEvent(
+          new CustomEvent("workspace:sandbox-resume-autoscroll"),
+        );
+        void (async () => {
+          try {
+            let targetThreadId = threadId;
+            if (isNewThread) {
+              const created = await getAPIClient().threads.create({
+                metadata: {
+                  agent_id,
+                  ...(Object.keys(conversationMetadata).length > 0
+                    ? conversationMetadata
+                    : {}),
+                },
+              });
+              targetThreadId = created.thread_id;
+              knowledgeSelectionThread.current = targetThreadId;
+              activeThreadIdRef.current = targetThreadId;
+              setThreadId(targetThreadId);
+              setIsNewThread(false);
+              history.replaceState(
+                null,
+                "",
+                `${window.location.pathname}${window.location.search}#${agentChatsBase}/${targetThreadId}`,
+              );
+              const title = message.text.trim().replace(/\s+/g, " ").slice(0, 60);
+              if (title) {
+                void getAPIClient().threads
+                  .updateState(targetThreadId, { values: { title } })
+                  .catch(() => undefined);
+              }
+            }
+            if (!targetThreadId) {
+              toast.error("无法创建会话。");
+              return;
+            }
+            await directChat.submit({
+              targetThreadId,
+              query: message.text,
+              modelName: settings.context.model_name,
+            });
+          } catch (error) {
+            toast.error(
+              error instanceof Error ? error.message : "极速问答失败。",
+            );
+          }
+        })();
+        return;
+      }
       setHasStartedConversation(true);
       // Remember this agent so the new-chat landing screen pre-selects it next
       // time (covers both direct agent chats and the ChatPage handoff replay).
@@ -767,7 +842,7 @@ export default function AgentChatPage({
         ...pageReportThemeContext,
       });
     },
-    [sendMessage, threadId, agent_id, settings.context.writing_mode, writingArtifactPath, baseMode],
+    [sendMessage, threadId, agent_id, settings.context.writing_mode, settings.context.model_name, writingArtifactPath, baseMode, directChat, directChatSelected, isNewThread, conversationMetadata, agentChatsBase, setThreadId, setIsNewThread],
   );
 
   // 任务流程：切换 / 新建会话时，停掉「上一会话服务端在跑」的 run（老对话要停止）。
@@ -852,8 +927,12 @@ export default function AgentChatPage({
   );
 
   const handleStop = useCallback(async () => {
+    if (directChat.isLoading) {
+      directChat.stop();
+      return;
+    }
     await thread.stop();
-  }, [thread]);
+  }, [directChat, thread]);
 
   const isLandingState = isNewThread && !hasStartedConversation;
   // 任务闸门面板：仅 rwfx（taskGate）在新建对话且尚未开始时，于问答页顶部展示目的树 + 覆盖确认。
@@ -867,7 +946,7 @@ export default function AgentChatPage({
 
   const status = thread.error
     ? "error"
-    : thread.isLoading
+    : thread.isLoading || directChat.isLoading
       ? "streaming"
       : "ready";
 
@@ -1084,7 +1163,7 @@ export default function AgentChatPage({
                         autoFocus
                         status={status}
                         context={settings.context}
-                        disabled={isDemoMode || knowledgeSelectionLoading}
+                        disabled={isDemoMode || knowledgeSelectionLoading || directChat.isLoading}
                         onContextChange={(context) =>
                           setSettings("context", context)
                         }
@@ -1104,7 +1183,7 @@ export default function AgentChatPage({
                                 pendingKnowledgeSelection.current === null &&
                                 pendingAssistantKnowledgeSelection.current === null
                               }
-                              disabled={isDemoMode || knowledgeSelectionLoading}
+                              disabled={isDemoMode || knowledgeSelectionLoading || directChat.isLoading}
                               label="知识空间"
                             />
                           )
@@ -1175,7 +1254,7 @@ export default function AgentChatPage({
                         assistantId={agent_id}
                         status={status}
                         context={settings.context}
-                        disabled={isDemoMode || knowledgeSelectionLoading}
+                        disabled={isDemoMode || knowledgeSelectionLoading || directChat.isLoading}
                         onContextChange={(context) =>
                           setSettings("context", context)
                         }
@@ -1190,7 +1269,7 @@ export default function AgentChatPage({
                               assistantValue={selectedAssistantKnowledgeBaseIds}
                               onChange={changeKnowledgeSelection}
                               onAssistantChange={changeAssistantKnowledgeSelection}
-                              disabled={isDemoMode || knowledgeSelectionLoading}
+                              disabled={isDemoMode || knowledgeSelectionLoading || directChat.isLoading}
                               readOnly
                               label="知识空间"
                             />

frontend-web/src/pages/ChatPage.tsx

@@ -69,6 +69,7 @@ import { clearResearchReportStream } from "@/core/threads/use-research-report-st
 import { resolveWritingModeReportPath } from "@/core/threads/writing-mode";
 import { queueLlmWikiConversationDeposit } from "@/core/llmwiki/conversation-deposit";
 import { useSingleRoundRagMessages } from "@/core/threads/use-single-round-rag";
+import { useDirectChatMessages, isDirectChatProfile, isTurboModeVisible } from "@/core/threads/use-direct-chat";
 import { useMultiRoundRagMessages, mergeMessagesByTime, toLangGraphMessages } from "@/core/threads/use-multi-round-rag";
 import { useLayoutMode } from "@/core/system-settings/hooks";
 import { useRouteBase } from "@/core/route-base";
@@ -348,9 +349,12 @@ export default function ChatPage() {
   });
   const singleRoundRag = useSingleRoundRagMessages(threadId);
   const multiRoundRag = useMultiRoundRagMessages(threadId);
+  const directChat = useDirectChatMessages(threadId);
   const singleRoundRagSelected = isSingleRoundRagProfile(settings.context);
   const multiRoundRagSelected = isMultiRoundRagProfile(settings.context);
   const frontendRagSelected = isFrontendRagProfile(settings.context);
+  const directChatSelected =
+    isDirectChatProfile(settings.context) && isTurboModeVisible();
   // 写作模式深度研究进度：注入 displayThread 尾部（进度步骤条 + report.md
   // 流式 stub），成功且真实产物落地后自动撤出。
   const { displayThread: researchDisplayThread } = useResearchWritingFeed(threadId, thread);
@@ -360,13 +364,15 @@ export default function ChatPage() {
     if (
       singleRoundRag.messages.length === 0 &&
       ragMessages.length === 0 &&
+      directChat.messages.length === 0 &&
       !singleRoundRag.isLoading &&
-      !multiRoundRag.isLoading
+      !multiRoundRag.isLoading &&
+      !directChat.isLoading
     ) {
       return base;
     }
     const messages = mergeMessagesByTime(
-      [...base.messages, ...singleRoundRag.messages],
+      [...base.messages, ...singleRoundRag.messages, ...directChat.messages],
       ragMessages,
     );
     return {
@@ -374,9 +380,14 @@ export default function ChatPage() {
       values: { ...base.values, messages },
       messages,
       isLoading:
-        base.isLoading || singleRoundRag.isLoading || multiRoundRag.isLoading,
+        base.isLoading ||
+        singleRoundRag.isLoading ||
+        multiRoundRag.isLoading ||
+        directChat.isLoading,
     } as typeof base;
   }, [
+    directChat.isLoading,
+    directChat.messages,
     multiRoundRag.isLoading,
     multiRoundRag.messages,
     researchDisplayThread,
@@ -392,6 +403,52 @@ export default function ChatPage() {
 
   const handleSubmit = useCallback(
     (message: PromptInputMessage) => {
+      if (directChatSelected) {
+        if (message.files && message.files.length > 0) {
+          toast.error("极速问答暂不支持附件，请移除附件后重试。");
+          return;
+        }
+        setHasStartedConversation(true);
+        window.dispatchEvent(
+          new CustomEvent("workspace:sandbox-resume-autoscroll"),
+        );
+        void (async () => {
+          try {
+            let targetThreadId = threadId;
+            if (isNewThread) {
+              const created = await getAPIClient(isMock).threads.create({});
+              targetThreadId = created.thread_id;
+              setThreadId(targetThreadId);
+              setIsNewThread(false);
+              history.replaceState(
+                null,
+                "",
+                `${window.location.pathname}${window.location.search}#${routeBase}/chats/${targetThreadId}`,
+              );
+              const title = message.text.trim().replace(/\s+/g, " ").slice(0, 60);
+              if (title) {
+                void getAPIClient(isMock).threads
+                  .updateState(targetThreadId, { values: { title } })
+                  .catch(() => undefined);
+              }
+            }
+            if (!targetThreadId) {
+              toast.error("无法创建会话。");
+              return;
+            }
+            await directChat.submit({
+              targetThreadId,
+              query: message.text,
+              modelName: settings.context.model_name,
+            });
+          } catch (error) {
+            toast.error(
+              error instanceof Error ? error.message : "极速问答失败。",
+            );
+          }
+        })();
+        return;
+      }
       if (frontendRagSelected) {
         if (message.files && message.files.length > 0) {
           toast.error("RAG 问答暂不支持附件，请移除附件后重试。");
@@ -560,6 +617,8 @@ export default function ChatPage() {
       multiRoundRag,
       multiRoundRagSelected,
       frontendRagSelected,
+      directChat,
+      directChatSelected,
       routeBase,
       setIsNewThread,
       setThreadId,
@@ -655,6 +714,10 @@ export default function ChatPage() {
     saveLastAgentId(agentId);
   }, []);
   const handleStop = useCallback(async () => {
+    if (directChat.isLoading) {
+      directChat.stop();
+      return;
+    }
     if (singleRoundRag.isLoading) {
       singleRoundRag.stop();
       return;
@@ -664,7 +727,7 @@ export default function ChatPage() {
       return;
     }
     await thread.stop();
-  }, [multiRoundRag, singleRoundRag, thread]);
+  }, [directChat, multiRoundRag, singleRoundRag, thread]);
 
   const promptPrefixProp = useMemo(() => {
     if (!promptPrefixSettings?.enabled) return undefined;
@@ -914,7 +977,8 @@ export default function ChatPage() {
                         agentHandoffPending ||
                         knowledgeSelectionLoading ||
                         singleRoundRag.isLoading ||
-                        multiRoundRag.isLoading
+                        multiRoundRag.isLoading ||
+                        directChat.isLoading
                       }
                       onContextChange={(context) =>
                         setSettings("context", context)

offline-backend-20260512/backend/CLAUDE.md

@@ -1014,9 +1014,11 @@ scheduler runs once per deployment. Per run:
 
 ### vLLM Provider (`packages/harness/deerflow/models/vllm_provider.py`)
 
-- `VllmChatModel` subclasses `langchain_openai:ChatOpenAI` for vLLM 0.19.0 OpenAI-compatible endpoints
+- `VllmChatModel` subclasses `langchain_openai:ChatOpenAI` for vLLM OpenAI-compatible endpoints
-- Preserves vLLM's non-standard assistant `reasoning` field on full responses, streaming deltas, and follow-up tool-call turns
+- Thinking text is accepted from either `reasoning` (newer vLLM parsers) or `reasoning_content` (older vLLM / DeepSeek-style). The UI still reads canonical `additional_kwargs.reasoning_content`. Follow-up turns echo only the wire field that actually arrived, so other ChatOpenAI models are unchanged
-- Designed for configs that enable thinking through `extra_body.chat_template_kwargs.enable_thinking` on vLLM 0.19.0 Qwen reasoning models, while accepting the older `thinking` alias
+- Shared helpers live in `deerflow/models/reasoning_fields.py` and are also used by `PatchedChatOpenAI` (the default `langchain_openai:ChatOpenAI` patch)
+- Designed for configs that enable thinking through `extra_body.chat_template_kwargs.enable_thinking` on vLLM Qwen reasoning models, while accepting the older `thinking` alias
+- Per-model switch: set only that entry's `use: deerflow.models.vllm_provider:VllmChatModel`. Do not change SiliconFlow / DeepSeek entries that already use `langchain_openai:ChatOpenAI`
 
 ### IM Channels System (`app/channels/`)
 

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

@@ -42,6 +42,7 @@ from app.gateway.routers import (
     compact_menu_overrides,
     dashboard_sessions,
     deep_research,
+    direct_chat,
     external_llmwiki,
     feedback,
     fixed_questions,
@@ -1803,6 +1804,9 @@ This gateway provides custom endpoints for models, MCP configuration, skills, an
     # 开放问答接口 /api/open/chat —— 无需登录,指定模型 + 智能体做问答,可关闭思考模式
     app.include_router(open_chat.router)
 
+    # 极速问答 /api/direct-chat/stream —— 直连大模型流式输出,不走 lead agent / 工具 / 技能
+    app.include_router(direct_chat.router)
+
     # 开放接口 /api/open/3qfx/ask —— 无需登录,按 taskId 直接发起 3qfx 第一次问答(后台运行,落成真实会话)
     app.include_router(open_3qfx.router)
 

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

@@ -0,0 +1,149 @@
+"""Direct LLM chat — no lead agent, tools, skills, or memory.
+
+Used by the frontend 「极速」mode: stream tokens from ``create_chat_model``
+with thinking forced off so first-token latency matches a bare chat client.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+from collections.abc import AsyncIterator
+from typing import Any, Literal
+
+from fastapi import APIRouter, HTTPException
+from fastapi.responses import StreamingResponse
+from langchain_core.messages import AIMessage, BaseMessage, HumanMessage, SystemMessage
+from pydantic import BaseModel, Field
+
+from deerflow.config import get_app_config
+from deerflow.models import create_chat_model
+
+logger = logging.getLogger(__name__)
+
+router = APIRouter(prefix="/api/direct-chat", tags=["direct-chat"])
+
+_HEARTBEAT_SECONDS = 15.0
+_MAX_MESSAGES = 40
+
+
+class DirectChatMessage(BaseModel):
+    role: Literal["user", "assistant", "system"]
+    content: str = ""
+
+
+class DirectChatRequest(BaseModel):
+    messages: list[DirectChatMessage] = Field(..., min_length=1)
+    model_name: str | None = None
+
+
+def _chunk_text(chunk: Any) -> str:
+    content = getattr(chunk, "content", None)
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        parts: list[str] = []
+        for block in content:
+            if isinstance(block, str):
+                parts.append(block)
+            elif isinstance(block, dict) and isinstance(block.get("text"), str):
+                parts.append(block["text"])
+        return "".join(parts)
+    return ""
+
+
+def _build_lc_messages(body: DirectChatRequest) -> list[BaseMessage]:
+    built: list[BaseMessage] = []
+    for item in body.messages[-_MAX_MESSAGES:]:
+        text = (item.content or "").strip()
+        if not text:
+            continue
+        if item.role == "user":
+            built.append(HumanMessage(content=text))
+        elif item.role == "assistant":
+            built.append(AIMessage(content=text))
+        else:
+            built.append(SystemMessage(content=text))
+    if not any(isinstance(msg, HumanMessage) for msg in built):
+        raise HTTPException(status_code=400, detail="至少需要一条 user 消息")
+    return built
+
+
+def _sse(event: str, data: Any) -> str:
+    return f"event: {event}\ndata: {json.dumps(data, ensure_ascii=False)}\n\n"
+
+
+def _resolve_model_name(name: str | None) -> str:
+    config = get_app_config()
+    if not name:
+        if not config.models:
+            raise HTTPException(status_code=400, detail="未配置可用模型")
+        return config.models[0].name
+    if config.get_model_config(name) is None:
+        raise HTTPException(status_code=400, detail=f"模型不存在: {name}")
+    return name
+
+
+async def _direct_chat_event_stream(body: DirectChatRequest, model_name: str) -> AsyncIterator[str]:
+    lc_messages = _build_lc_messages(body)
+    queue: asyncio.Queue[tuple[str, Any]] = asyncio.Queue()
+    pieces: list[str] = []
+
+    async def produce() -> None:
+        try:
+            model = create_chat_model(
+                name=model_name,
+                thinking_enabled=False,
+                force_disable_thinking=True,
+            )
+            async for chunk in model.astream(lc_messages):
+                text = _chunk_text(chunk)
+                if text:
+                    pieces.append(text)
+                    await queue.put(("delta", {"text": text}))
+        except Exception as exc:  # noqa: BLE001 — stream errors become SSE error events
+            logger.exception("direct_chat 调用模型失败 (model=%s)", model_name)
+            await queue.put(("error", {"detail": f"模型调用失败:{exc}"}))
+        finally:
+            await queue.put(("__end__", None))
+
+    task = asyncio.create_task(produce())
+    try:
+        while True:
+            try:
+                event, data = await asyncio.wait_for(queue.get(), timeout=_HEARTBEAT_SECONDS)
+            except asyncio.TimeoutError:
+                yield ": keepalive\n\n"
+                continue
+            if event == "__end__":
+                break
+            if event == "error":
+                yield _sse("error", data)
+                return
+            yield _sse(event, data)
+
+        answer = "".join(pieces)
+        if not answer:
+            yield _sse("error", {"detail": "模型未返回任何文本。"})
+            return
+        yield _sse("result", {"answer": answer, "model_name": model_name})
+    finally:
+        if not task.done():
+            task.cancel()
+
+
+@router.post("/stream")
+async def direct_chat_stream(body: DirectChatRequest) -> StreamingResponse:
+    """Stream a plain chat completion. No agent graph, tools, or skills."""
+    model_name = _resolve_model_name(body.model_name)
+    _build_lc_messages(body)
+    return StreamingResponse(
+        _direct_chat_event_stream(body, model_name),
+        media_type="text/event-stream",
+        headers={
+            "Cache-Control": "no-cache",
+            "Connection": "keep-alive",
+            "X-Accel-Buffering": "no",
+        },
+    )

offline-backend-20260512/backend/config.yaml

@@ -33,6 +33,26 @@ concurrency:
     fail_open: true
 
 models: 
+- name: qwen3.6
+  display_name: qwen3.6
+  use: deerflow.models.vllm_provider:VllmChatModel
+  model: qwen3.6
+  api_key: sk-6AqP_UYpPYJKdclYWR-FSA
+  base_url: http://119.254.86.251:8001/v1
+  request_timeout: 30.0
+  max_retries: 2
+  stream_usage: true
+  supports_thinking: true
+  supports_reasoning_effort: false
+  supports_vision: false
+  when_thinking_enabled:
+    extra_body:
+      chat_template_kwargs:
+        enable_thinking: true
+  when_thinking_disabled:
+    extra_body:
+      chat_template_kwargs:
+        enable_thinking: false
 - name: MiniMax-M2.7-fash
   display_name: MiniMax-M2.7-fash
   use: langchain_openai:ChatOpenAI
@@ -59,16 +79,24 @@ models:
   supports_vision: false
 - name: qwen3.6
   display_name: qwen3.6
-  use: langchain_openai:ChatOpenAI
+  use: deerflow.models.vllm_provider:VllmChatModel
   model: qwen3.6
   api_key: sk-6AqP_UYpPYJKdclYWR-FSA
   base_url: http://119.254.86.251:8001/v1
   request_timeout: 30.0
   max_retries: 2
   stream_usage: true
-  supports_thinking: false
+  supports_thinking: true
   supports_reasoning_effort: false
   supports_vision: false
+  when_thinking_enabled:
+    extra_body:
+      chat_template_kwargs:
+        enable_thinking: true
+  when_thinking_disabled:
+    extra_body:
+      chat_template_kwargs:
+        enable_thinking: false
 - name: zai-org/GLM-5-FP8
   display_name: zai-org/GLM-5-FP8
   use: langchain_openai:ChatOpenAI
@@ -466,7 +494,7 @@ checkpointer:
 # 三层覆盖优先级(高->低):
 #   users/{user_id}/memory_config.json > .deer-flow/memory_user_overrides.json > 此处
 memory:
-  enabled: true
+  enabled: false # 临时关闭跨会话记忆（提取/注入/Hindsight 召回与 retain、memory 工具）
   version:
     v1 # 记忆版本: 'v1'(原版 DeerFlow, memory.json + LLM 自动提取)
     #           'v2'(USER.md/MEMORY.md + 可选 Hindsight)

offline-backend-20260512/backend/docs/CONFIGURATION.md

@@ -123,6 +123,27 @@ models:
 
 If your OpenRouter key lives in a different environment variable name, point `api_key` at that variable explicitly (for example `api_key: $OPENROUTER_API_KEY`).
 
+**vLLM (per-model, do not change other entries)**:
+
+`use` is per model. SiliconFlow / DeepSeek entries should stay on `langchain_openai:ChatOpenAI`. For a self-hosted vLLM reasoning model (thinking may arrive as `reasoning` on newer vLLM, or `reasoning_content` on older builds), set only that one entry to `VllmChatModel`:
+
+```yaml
+models:
+  - name: qwen3.6
+    display_name: qwen3.6
+    use: deerflow.models.vllm_provider:VllmChatModel
+    model: qwen3.6
+    api_key: $VLLM_API_KEY
+    base_url: http://127.0.0.1:8001/v1
+    supports_thinking: true
+    when_thinking_enabled:
+      extra_body:
+        chat_template_kwargs:
+          enable_thinking: true
+```
+
+The default `langchain_openai:ChatOpenAI` path also maps both thinking field names onto the UI's `reasoning_content`. Use `VllmChatModel` when follow-up turns must echo vLLM's original `reasoning` field.
+
 **Thinking Models**:
 Some models support "thinking" mode for complex reasoning:
 

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

@@ -433,7 +433,9 @@ def _build_middlewares(
 
     # Add MemoryMiddleware (after TitleMiddleware) — skip in bootstrap mode (agent creation flow) and ephemeral runs
     memory_injection_enabled = cfg.get("memory_injection_enabled", True)
-    if not is_bootstrap and not is_writing_setup and not is_report_structure_setup and not is_ephemeral:
+    memory_cfg = getattr(resolved_app_config, "memory", None)
+    memory_globally_enabled = memory_cfg is not None and getattr(memory_cfg, "enabled", False)
+    if memory_globally_enabled and not is_bootstrap and not is_writing_setup and not is_report_structure_setup and not is_ephemeral:
         middlewares.append(MemoryMiddleware(agent_name=agent_name, memory_config=resolved_app_config.memory, injection_enabled=memory_injection_enabled))
 
     # KnowledgeRagMiddleware — inject knowledge-base recall (phase 2). Gated

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

@@ -883,11 +883,16 @@ def _get_memory_context(agent_name: str | None = None, *, app_config: AppConfig
         return ""
     try:
         from deerflow.agents.memory.manager import build_memory_manager
-        from deerflow.config.memory_config import get_effective_memory_config
+        from deerflow.config.memory_config import get_effective_memory_config, get_memory_config
         from deerflow.runtime.user_context import get_effective_user_id
 
+        if not get_memory_config().enabled:
+            return ""
+
         user_id = get_effective_user_id()
         config = get_effective_memory_config(user_id)
+        if not config.enabled or not config.injection.enabled:
+            return ""
 
         manager = build_memory_manager(
             user_id=user_id,

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

@@ -16,7 +16,7 @@
 ============
 ``get_effective_memory_config(user_id)`` 按优先级合并 (高 -> 低):
   1. per-user 覆盖   数据库 ``user_memory_configs`` 表 (经进程内 cache 读取)
-     (有 per-user 记录时自动强制 enabled=True)
+     (仅白名单字段,不能覆盖 ``enabled``)
   2. config.yaml     ``memory:`` 段 (即 ``get_memory_config()``)
   3. Pydantic 默认值
 
@@ -254,13 +254,12 @@ def get_effective_memory_config(user_id: str) -> MemoryConfig:
     """返回某用户的生效记忆配置 (同步,从进程内 cache 读取 per-user 覆盖)。
 
     优先级 (高 -> 低):per-user DB 覆盖 (经 cache) > config.yaml > 默认值。
-    有 per-user 记录时自动强制 enabled=True。
+    ``enabled`` 始终以 config.yaml 为准,用户覆盖不能把它重新打开。
     """
     base = _memory_config.model_dump()
     recall_budget = _recall_budget_cache.get(user_id)
     if recall_budget is not None:
         base.setdefault("hindsight", {})["recall_budget"] = recall_budget
-        base["enabled"] = True
     try:
         return MemoryConfig(**base)
     except Exception as exc:  # noqa: BLE001 - 覆盖产生非法配置时降级

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

@@ -31,6 +31,96 @@ def _vllm_disable_chat_template_kwargs(chat_template_kwargs: dict) -> dict:
     return disable_kwargs
 
 
+# Qwen3.8 chat templates only accept these three effort names.
+_QWEN_REASONING_EFFORT = {
+    "minimal": "low",
+    "low": "low",
+    "medium": "medium",
+    "high": "xhigh",
+    "xhigh": "xhigh",
+}
+
+_OPENAI_THINKING_OFF_EXTRA_BODY = {
+    "enable_thinking": False,
+    "chat_template_kwargs": {"enable_thinking": False},
+    "thinking": {"type": "disabled"},
+}
+
+
+def _is_vllm_provider(model_use: str) -> bool:
+    lowered = (model_use or "").lower()
+    return "vllm_provider" in lowered or lowered.endswith(":vllmchatmodel")
+
+
+def _is_qwen38_model(model_config) -> bool:
+    blob = f"{getattr(model_config, 'name', '')} {getattr(model_config, 'model', '')}".lower()
+    return "qwen3.8" in blob
+
+
+def _should_control_openai_thinking(model_config) -> bool:
+    """Whether this OpenAI-compatible model needs explicit thinking on/off flags.
+
+    Qwen3.8 (and current vLLM reasoning parsers) default to thinking at ``xhigh``.
+    If DeerFlow does not send ``enable_thinking: false``, the first visible answer
+    token waits until that hidden reasoning finishes — third-party chat UIs that
+    disable thinking or stream the reasoning field therefore feel much faster.
+    """
+    if getattr(model_config, "supports_thinking", False):
+        return True
+    if _is_vllm_provider(getattr(model_config, "use", "")):
+        return True
+    return _is_qwen38_model(model_config)
+
+
+def _map_qwen_reasoning_effort(effort: object) -> str | None:
+    if effort is None:
+        return None
+    return _QWEN_REASONING_EFFORT.get(str(effort).strip().lower())
+
+
+def _apply_openai_thinking_control(
+    model_config,
+    model_class,
+    model_settings_from_config: dict,
+    *,
+    thinking_enabled: bool,
+    reasoning_effort: object,
+) -> None:
+    """Inject vLLM/Qwen thinking flags when the UI toggle is on or off."""
+    if "extra_body" not in getattr(model_class, "model_fields", {}):
+        return
+    qwen_vllm = _is_vllm_provider(getattr(model_config, "use", "")) or _is_qwen38_model(model_config)
+    if thinking_enabled:
+        if not qwen_vllm:
+            return
+        enable_body: dict = {
+            "enable_thinking": True,
+            "chat_template_kwargs": {"enable_thinking": True},
+        }
+        mapped_effort = _map_qwen_reasoning_effort(reasoning_effort)
+        if mapped_effort:
+            enable_body["chat_template_kwargs"]["reasoning_effort"] = mapped_effort
+        model_settings_from_config["extra_body"] = _deep_merge_dicts(
+            model_settings_from_config.get("extra_body"),
+            enable_body,
+        )
+        return
+    if not _should_control_openai_thinking(model_config):
+        return
+    disable_body = (
+        {
+            "enable_thinking": False,
+            "chat_template_kwargs": {"enable_thinking": False},
+        }
+        if qwen_vllm
+        else _OPENAI_THINKING_OFF_EXTRA_BODY
+    )
+    model_settings_from_config["extra_body"] = _deep_merge_dicts(
+        model_settings_from_config.get("extra_body"),
+        disable_body,
+    )
+
+
 def _enable_stream_usage_by_default(model_use_path: str, model_settings_from_config: dict) -> None:
     """Enable stream usage for OpenAI-compatible models unless explicitly configured.
 
@@ -127,10 +217,19 @@ def create_chat_model(name: str | None = None, thinking_enabled: bool = False, *
         elif has_thinking_settings and effective_wte.get("thinking", {}).get("type"):
             # Native langchain_anthropic: thinking is a direct constructor parameter
             model_settings_from_config["thinking"] = {"type": "disabled"}
+    requested_reasoning_effort = kwargs.get("reasoning_effort")
     if not model_config.supports_reasoning_effort:
         kwargs.pop("reasoning_effort", None)
         model_settings_from_config.pop("reasoning_effort", None)
 
+    _apply_openai_thinking_control(
+        model_config,
+        model_class,
+        model_settings_from_config,
+        thinking_enabled=thinking_enabled,
+        reasoning_effort=requested_reasoning_effort,
+    )
+
     # Hard override: force-disable thinking regardless of the per-model config shape.
     # Injected only for OpenAI-compatible classes (those declaring `extra_body`);
     # other providers would reject the kwarg, so they are skipped (they ignore it).

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

@@ -33,6 +33,7 @@ from langchain_core.outputs import ChatGeneration, ChatGenerationChunk, ChatResu
 from langchain_openai import ChatOpenAI
 
 from deerflow.models.message_ordering import normalize_chat_payload_messages
+from deerflow.models.reasoning_fields import apply_reasoning_from_payload, extract_reasoning_text, fill_ui_reasoning_content, restore_reasoning_fields
 
 
 class PatchedChatOpenAI(ChatOpenAI):
@@ -118,12 +119,12 @@ class PatchedChatOpenAI(ChatOpenAI):
         choices = chunk.get("choices", []) or chunk.get("chunk", {}).get("choices", [])
         choice = choices[0] if choices else {}
         delta = choice.get("delta") if isinstance(choice, dict) else None
-        reasoning_content = _extract_reasoning_content(delta if isinstance(delta, dict) else None)
+        if isinstance(generation_chunk.message, AIMessageChunk):
-        if reasoning_content and isinstance(generation_chunk.message, AIMessageChunk):
+            generation_chunk.message = apply_reasoning_from_payload(
-            generation_chunk.message = _append_reasoning_content(
                 generation_chunk.message,
-                reasoning_content,
+                delta if isinstance(delta, dict) else None,
             )
+            generation_chunk.message = fill_ui_reasoning_content(generation_chunk.message)
         return generation_chunk
 
     def _create_chat_result(
@@ -141,9 +142,15 @@ class PatchedChatOpenAI(ChatOpenAI):
             message = generation.message
             choice = choices[index] if index < len(choices) else {}
             choice_message = choice.get("message", {}) if isinstance(choice, dict) else {}
-            reasoning_content = _extract_reasoning_content(choice_message if isinstance(choice_message, dict) else None)
+            if isinstance(message, AIMessage):
-            if reasoning_content and isinstance(message, AIMessage):
+                updated = apply_reasoning_from_payload(
-                message = _append_reasoning_content(message, reasoning_content)
+                    message,
+                    choice_message if isinstance(choice_message, dict) else None,
+                    append=False,
+                )
+                updated = fill_ui_reasoning_content(updated)
+                if updated is not message:
+                    message = updated
                     generation = ChatGeneration(
                         message=message,
                         generation_info=generation.generation_info,
@@ -195,33 +202,22 @@ def _restore_tool_call_signatures(payload_msg: dict, orig_msg: AIMessage) -> Non
 
 
 def _restore_reasoning_content(payload_msg: dict, orig_msg: AIMessage) -> None:
-    """Re-inject ``reasoning_content`` onto assistant payload message.
+    """Re-inject thinking fields onto assistant payload message.
 
     Some thinking-enabled OpenAI-compatible providers require each historical
-    assistant turn to include the original ``reasoning_content`` in subsequent
+    assistant turn to include the original thinking field in subsequent
-    requests. LangChain stores it in ``AIMessage.additional_kwargs`` but does
+    requests. Newer vLLM puts that text in ``reasoning``; DeepSeek and older
-    not always serialize it back into ``messages``.
+    vLLM use ``reasoning_content``. Echo only the fields this turn originally
+    carried so other models are not given an extra field.
     """
-    reasoning_content = orig_msg.additional_kwargs.get("reasoning_content")
+    restore_reasoning_fields(payload_msg, orig_msg)
-    if reasoning_content is not None:
-        payload_msg["reasoning_content"] = reasoning_content
 
 
 def _extract_reasoning_content(payload: dict | None) -> str | None:
     """Extract reasoning text from provider payload delta/message objects."""
-    if not isinstance(payload, dict):
+    return extract_reasoning_text(payload)
-        return None
-    reasoning_content = payload.get("reasoning_content")
-    return reasoning_content if isinstance(reasoning_content, str) and reasoning_content else None
 
 
 def _append_reasoning_content(message: AIMessage | AIMessageChunk, reasoning_content: str) -> AIMessage | AIMessageChunk:
     """Append streamed reasoning chunks, dedupe complete response reasoning."""
-    additional_kwargs = dict(message.additional_kwargs)
+    return apply_reasoning_from_payload(message, {"reasoning_content": reasoning_content})
-    existing = additional_kwargs.get("reasoning_content")
-    if isinstance(existing, str) and existing:
-        if reasoning_content not in existing:
-            additional_kwargs["reasoning_content"] = f"{existing}{reasoning_content}"
-    else:
-        additional_kwargs["reasoning_content"] = reasoning_content
-    return message.model_copy(update={"additional_kwargs": additional_kwargs})

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

@@ -0,0 +1,140 @@
+"""Normalize OpenAI-compatible thinking fields across provider versions.
+
+Older vLLM / DeepSeek put thinking in ``reasoning_content``. Newer vLLM
+reasoning parsers put it in ``reasoning``. DeerFlow's UI and downstream code
+canonicalize on ``AIMessage.additional_kwargs.reasoning_content``.
+
+Only fields that actually appeared on a response are echoed back on later
+turns, so models that still use ``reasoning_content`` (or send neither) are
+unchanged.
+"""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+
+from langchain_core.messages import AIMessage, AIMessageChunk
+
+
+def reasoning_to_text(reasoning: Any) -> str:
+    """Best-effort extraction of readable reasoning text from provider payloads."""
+    if isinstance(reasoning, str):
+        return reasoning
+
+    if isinstance(reasoning, list):
+        parts = [reasoning_to_text(item) for item in reasoning]
+        return "".join(part for part in parts if part)
+
+    if isinstance(reasoning, dict):
+        for key in ("text", "content", "reasoning"):
+            value = reasoning.get(key)
+            if isinstance(value, str):
+                return value
+            if value is not None:
+                text = reasoning_to_text(value)
+                if text:
+                    return text
+        try:
+            return json.dumps(reasoning, ensure_ascii=False)
+        except TypeError:
+            return str(reasoning)
+
+    if reasoning is None:
+        return ""
+    try:
+        return json.dumps(reasoning, ensure_ascii=False)
+    except TypeError:
+        return str(reasoning)
+
+
+def extract_reasoning_text(payload: dict | None) -> str | None:
+    """Read thinking text from ``reasoning_content`` or ``reasoning``.
+
+    Prefer ``reasoning_content`` when both are present so older DeepSeek /
+    vLLM payloads keep their original string. Newer vLLM that only sends
+    ``reasoning`` (string or structured) is still captured.
+    """
+    if not isinstance(payload, dict):
+        return None
+    for key in ("reasoning_content", "reasoning"):
+        value = payload.get(key)
+        if value is None:
+            continue
+        text = reasoning_to_text(value)
+        if text:
+            return text
+    return None
+
+
+def _merge_streamed_value(existing: Any, incoming: Any) -> Any:
+    if isinstance(existing, str) and isinstance(incoming, str):
+        if incoming and incoming not in existing:
+            return f"{existing}{incoming}"
+        return existing or incoming
+    return incoming if incoming is not None else existing
+
+
+def apply_reasoning_from_payload(
+    message: AIMessage | AIMessageChunk,
+    payload: dict | None,
+    *,
+    append: bool = True,
+) -> AIMessage | AIMessageChunk:
+    """Copy provider thinking fields onto ``additional_kwargs`` for UI + echo-back."""
+    if not isinstance(payload, dict):
+        return message
+
+    additional_kwargs = dict(message.additional_kwargs)
+    changed = False
+
+    if "reasoning" in payload and payload.get("reasoning") is not None:
+        incoming = payload.get("reasoning")
+        existing = additional_kwargs.get("reasoning")
+        merged = _merge_streamed_value(existing, incoming) if append else incoming
+        if merged is not None and merged != existing:
+            additional_kwargs["reasoning"] = merged
+            changed = True
+
+    text = extract_reasoning_text(payload)
+    if text:
+        existing = additional_kwargs.get("reasoning_content")
+        if append and isinstance(existing, str) and existing:
+            if text not in existing:
+                additional_kwargs["reasoning_content"] = f"{existing}{text}"
+                changed = True
+        elif text != existing:
+            additional_kwargs["reasoning_content"] = text
+            changed = True
+
+    if not changed:
+        return message
+    return message.model_copy(update={"additional_kwargs": additional_kwargs})
+
+
+def fill_ui_reasoning_content(message: AIMessage | AIMessageChunk) -> AIMessage | AIMessageChunk:
+    """If LangChain stored ``reasoning`` but not the UI field, copy the text across."""
+    kwargs = message.additional_kwargs
+    reasoning = kwargs.get("reasoning")
+    if reasoning is None or kwargs.get("reasoning_content"):
+        return message
+    return apply_reasoning_from_payload(message, {"reasoning": reasoning}, append=False)
+
+
+def restore_reasoning_fields(payload_msg: dict, orig_msg: AIMessage) -> None:
+    """Re-inject the thinking field this assistant turn originally carried.
+
+    Newer vLLM stores thinking on ``reasoning`` and we also copy a UI-facing
+    ``reasoning_content`` string from it. Echo the wire field that actually
+    arrived: ``reasoning`` wins when present so we do not send a synthesized
+    ``reasoning_content`` to a parser that no longer accepts it. Older
+    DeepSeek / vLLM turns only have ``reasoning_content`` and stay on that
+    field.
+    """
+    reasoning = orig_msg.additional_kwargs.get("reasoning")
+    if reasoning is not None:
+        payload_msg["reasoning"] = reasoning
+        return
+    reasoning_content = orig_msg.additional_kwargs.get("reasoning_content")
+    if reasoning_content is not None:
+        payload_msg["reasoning_content"] = reasoning_content

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

@@ -6,15 +6,16 @@ from assistant messages and streaming deltas. That breaks interleaved
 thinking/tool-call flows because vLLM expects the assistant's prior reasoning to
 be echoed back on subsequent turns.
 
-This provider preserves ``reasoning`` on:
+This provider preserves ``reasoning`` (newer vLLM) and
+``reasoning_content`` (older vLLM / DeepSeek-style parsers) on:
     - non-streaming responses
     - streaming deltas
     - multi-turn request payloads
+UI code still reads the canonical ``additional_kwargs.reasoning_content``.
 """
 
 from __future__ import annotations
 
-import json
 from collections.abc import Mapping
 from typing import Any, cast
 
@@ -36,6 +37,7 @@ from langchain_openai import ChatOpenAI
 from langchain_openai.chat_models.base import _create_usage_metadata
 
 from deerflow.models.message_ordering import normalize_chat_payload_messages
+from deerflow.models.reasoning_fields import apply_reasoning_from_payload, restore_reasoning_fields
 
 
 def _normalize_vllm_chat_template_kwargs(payload: dict[str, Any]) -> None:
@@ -66,31 +68,9 @@ def _normalize_vllm_chat_template_kwargs(payload: dict[str, Any]) -> None:
 
 def _reasoning_to_text(reasoning: Any) -> str:
     """Best-effort extraction of readable reasoning text from vLLM payloads."""
-    if isinstance(reasoning, str):
+    from deerflow.models.reasoning_fields import reasoning_to_text
-        return reasoning
 
-    if isinstance(reasoning, list):
+    return reasoning_to_text(reasoning)
-        parts = [_reasoning_to_text(item) for item in reasoning]
-        return "".join(part for part in parts if part)
-
-    if isinstance(reasoning, dict):
-        for key in ("text", "content", "reasoning"):
-            value = reasoning.get(key)
-            if isinstance(value, str):
-                return value
-            if value is not None:
-                text = _reasoning_to_text(value)
-                if text:
-                    return text
-        try:
-            return json.dumps(reasoning, ensure_ascii=False)
-        except TypeError:
-            return str(reasoning)
-
-    try:
-        return json.dumps(reasoning, ensure_ascii=False)
-    except TypeError:
-        return str(reasoning)
 
 
 def _convert_delta_to_message_chunk_with_reasoning(_dict: Mapping[str, Any], default_class: type[BaseMessageChunk]) -> BaseMessageChunk:
@@ -106,12 +86,8 @@ def _convert_delta_to_message_chunk_with_reasoning(_dict: Mapping[str, Any], def
             function_call["name"] = ""
         additional_kwargs["function_call"] = function_call
 
-    reasoning = _dict.get("reasoning")
+    seed = AIMessageChunk(content="", additional_kwargs=additional_kwargs)
-    if reasoning is not None:
+    additional_kwargs = dict(apply_reasoning_from_payload(seed, dict(_dict)).additional_kwargs)
-        additional_kwargs["reasoning"] = reasoning
-        reasoning_text = _reasoning_to_text(reasoning)
-        if reasoning_text:
-            additional_kwargs["reasoning_content"] = reasoning_text
 
     tool_call_chunks = []
     if raw_tool_calls := _dict.get("tool_calls"):
@@ -151,11 +127,7 @@ def _convert_delta_to_message_chunk_with_reasoning(_dict: Mapping[str, Any], def
 
 def _restore_reasoning_field(payload_msg: dict[str, Any], orig_msg: AIMessage) -> None:
     """Re-inject vLLM reasoning onto outgoing assistant messages."""
-    reasoning = orig_msg.additional_kwargs.get("reasoning")
+    restore_reasoning_fields(payload_msg, orig_msg)
-    if reasoning is None:
-        reasoning = orig_msg.additional_kwargs.get("reasoning_content")
-    if reasoning is not None:
-        payload_msg["reasoning"] = reasoning
 
 
 class VllmChatModel(ChatOpenAI):
@@ -204,13 +176,12 @@ class VllmChatModel(ChatOpenAI):
             message = generation.message
             if not isinstance(message, AIMessage):
                 continue
-            reasoning = choice.get("message", {}).get("reasoning")
+            choice_message = choice.get("message", {}) if isinstance(choice, dict) else {}
-            if reasoning is None:
+            generation.message = apply_reasoning_from_payload(
-                continue
+                message,
-            message.additional_kwargs["reasoning"] = reasoning
+                choice_message if isinstance(choice_message, dict) else None,
-            reasoning_text = _reasoning_to_text(reasoning)
+                append=False,
-            if reasoning_text:
+            )
-                message.additional_kwargs["reasoning_content"] = reasoning_text
 
         return result
 

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

@@ -239,3 +239,67 @@ def test_force_disable_thinking_skips_generic_openai_models_without_thinking_sup
     )
 
     assert chat_model.extra_body is None
+
+
+def _extra_body(chat_model) -> dict:
+    body = chat_model.extra_body or {}
+    return body if isinstance(body, dict) else dict(body)
+
+
+def test_qwen38_flash_mode_disables_default_xhigh_thinking() -> None:
+    """Qwen3.8 defaults to thinking=xhigh; flash must send enable_thinking=false."""
+    config = AppConfig.model_construct(
+        models=[
+            ModelConfig(
+                name="qwen3.8-27b",
+                use="langchain_openai:ChatOpenAI",
+                model="Qwen/Qwen3.8-27B",
+                api_key="test-key",
+                base_url="http://vllm.example/v1",
+                supports_thinking=False,
+            )
+        ]
+    )
+
+    chat_model = create_chat_model("qwen3.8-27b", thinking_enabled=False, app_config=config)
+    extra_body = _extra_body(chat_model)
+
+    assert extra_body.get("enable_thinking") is False
+    assert extra_body.get("chat_template_kwargs", {}).get("enable_thinking") is False
+
+
+def test_qwen38_thinking_mode_sends_medium_reasoning_effort() -> None:
+    config = AppConfig.model_construct(
+        models=[
+            ModelConfig(
+                name="qwen3.8-27b",
+                use="langchain_openai:ChatOpenAI",
+                model="qwen3.8-27b",
+                api_key="test-key",
+                base_url="http://vllm.example/v1",
+                supports_thinking=True,
+            )
+        ]
+    )
+
+    chat_model = create_chat_model(
+        "qwen3.8-27b",
+        thinking_enabled=True,
+        reasoning_effort="medium",
+        app_config=config,
+    )
+    extra_body = _extra_body(chat_model)
+    kwargs = extra_body.get("chat_template_kwargs") or {}
+
+    assert extra_body.get("enable_thinking") is True
+    assert kwargs.get("enable_thinking") is True
+    assert kwargs.get("reasoning_effort") == "medium"
+
+
+def test_extract_reasoning_reads_vllm_reasoning_field() -> None:
+    from deerflow.models.patched_openai import _extract_reasoning_content
+
+    assert _extract_reasoning_content({"reasoning": "先分析问题"}) == "先分析问题"
+    assert _extract_reasoning_content({"reasoning_content": "逐步推理"}) == "逐步推理"
+    assert _extract_reasoning_content({"reasoning": {"content": "结构化思考"}}) == "结构化思考"
+    assert _extract_reasoning_content({"content": "答案"}) is None

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

@@ -0,0 +1,111 @@
+"""Reasoning field compatibility: vLLM `reasoning` vs older `reasoning_content`."""
+
+from __future__ import annotations
+
+from langchain_core.messages import AIMessage, AIMessageChunk
+
+from deerflow.models.patched_openai import _extract_reasoning_content, _restore_reasoning_content
+from deerflow.models.reasoning_fields import (
+    apply_reasoning_from_payload,
+    extract_reasoning_text,
+    fill_ui_reasoning_content,
+    restore_reasoning_fields,
+)
+from deerflow.models.vllm_provider import (
+    _convert_delta_to_message_chunk_with_reasoning,
+    _restore_reasoning_field,
+)
+
+
+def test_extract_reasoning_reads_vllm_reasoning_field() -> None:
+    assert extract_reasoning_text({"reasoning": "先分析问题"}) == "先分析问题"
+    assert extract_reasoning_text({"reasoning_content": "逐步推理"}) == "逐步推理"
+    assert extract_reasoning_text({"content": "答案"}) is None
+    assert _extract_reasoning_content({"reasoning": "先分析问题"}) == "先分析问题"
+
+
+def test_extract_reasoning_reads_structured_reasoning_object() -> None:
+    assert extract_reasoning_text({"reasoning": {"content": "结构化思考"}}) == "结构化思考"
+    assert extract_reasoning_text({"reasoning": {"text": "逐步分析"}}) == "逐步分析"
+
+
+def test_extract_prefers_reasoning_content_when_both_present() -> None:
+    assert (
+        extract_reasoning_text({"reasoning_content": "旧字段", "reasoning": "新字段"})
+        == "旧字段"
+    )
+
+
+def test_apply_new_vllm_reasoning_keeps_ui_field_and_wire_field() -> None:
+    message = apply_reasoning_from_payload(AIMessage(content="答案"), {"reasoning": "新版思考"})
+    assert message.additional_kwargs["reasoning"] == "新版思考"
+    assert message.additional_kwargs["reasoning_content"] == "新版思考"
+
+
+def test_apply_old_reasoning_content_does_not_invent_reasoning() -> None:
+    message = apply_reasoning_from_payload(
+        AIMessage(content="答案"),
+        {"reasoning_content": "旧版思考"},
+    )
+    assert "reasoning" not in message.additional_kwargs
+    assert message.additional_kwargs["reasoning_content"] == "旧版思考"
+
+
+def test_restore_new_vllm_echoes_reasoning_only() -> None:
+    orig = AIMessage(
+        content="答案",
+        additional_kwargs={"reasoning": "新版思考", "reasoning_content": "新版思考"},
+    )
+    payload: dict = {"role": "assistant", "content": "答案"}
+    restore_reasoning_fields(payload, orig)
+    assert payload["reasoning"] == "新版思考"
+    assert "reasoning_content" not in payload
+
+
+def test_restore_old_vllm_echoes_reasoning_content_only() -> None:
+    orig = AIMessage(content="答案", additional_kwargs={"reasoning_content": "旧版思考"})
+    payload: dict = {"role": "assistant", "content": "答案"}
+    restore_reasoning_fields(payload, orig)
+    assert payload["reasoning_content"] == "旧版思考"
+    assert "reasoning" not in payload
+
+
+def test_patched_openai_restore_does_not_add_reasoning_for_deepseek_turns() -> None:
+    orig = AIMessage(content="答案", additional_kwargs={"reasoning_content": "DeepSeek思考"})
+    payload: dict = {"role": "assistant", "content": "答案"}
+    _restore_reasoning_content(payload, orig)
+    assert payload == {"role": "assistant", "content": "答案", "reasoning_content": "DeepSeek思考"}
+
+
+def test_vllm_stream_delta_reasoning_maps_to_ui_field() -> None:
+    chunk = _convert_delta_to_message_chunk_with_reasoning(
+        {"role": "assistant", "content": "", "reasoning": "流式思考"},
+        AIMessageChunk,
+    )
+    assert isinstance(chunk, AIMessageChunk)
+    assert chunk.additional_kwargs["reasoning"] == "流式思考"
+    assert chunk.additional_kwargs["reasoning_content"] == "流式思考"
+
+
+def test_vllm_stream_delta_reasoning_content_does_not_copy_into_reasoning() -> None:
+    chunk = _convert_delta_to_message_chunk_with_reasoning(
+        {"role": "assistant", "content": "", "reasoning_content": "旧流式思考"},
+        AIMessageChunk,
+    )
+    assert isinstance(chunk, AIMessageChunk)
+    assert "reasoning" not in chunk.additional_kwargs
+    assert chunk.additional_kwargs["reasoning_content"] == "旧流式思考"
+
+
+def test_vllm_restore_old_field_stays_on_reasoning_content() -> None:
+    orig = AIMessage(content="答案", additional_kwargs={"reasoning_content": "旧版思考"})
+    payload: dict = {"role": "assistant", "content": "答案"}
+    _restore_reasoning_field(payload, orig)
+    assert payload["reasoning_content"] == "旧版思考"
+    assert "reasoning" not in payload
+
+
+def test_fill_ui_reasoning_content_from_langchain_kwargs() -> None:
+    message = AIMessage(content="答案", additional_kwargs={"reasoning": "网关已放入reasoning"})
+    filled = fill_ui_reasoning_content(message)
+    assert filled.additional_kwargs["reasoning_content"] == "网关已放入reasoning"

offline-backend-20260512/config.yaml

@@ -264,7 +264,7 @@ ai_writing:
 # 三层覆盖优先级(高->低):
 #   users/{user_id}/memory_config.json > .deer-flow/memory_user_overrides.json > 此处
 memory:
-  enabled: true
+  enabled: false # 临时关闭跨会话记忆（提取/注入/Hindsight 召回与 retain、memory 工具）
   version: v1
   provider: hindsight              # 'builtin' | 'hindsight'(外部 provider 名)
                                    # builtin 始终启用,这里只决定 external 选哪个