From 3907034a1df6f31cc52cfa99386c7a9bf95e7c6a Mon Sep 17 00:00:00 2001
From: fengxiangdebaozi <10845145+fengxiangdebaozi@user.noreply.gitee.com>
Date: Mon, 15 Jun 2026 18:20:20 +0800
Subject: [PATCH 1/9] =?UTF-8?q?feat(roundtable):=20=E7=AC=AC=E4=BA=8C?=
 =?UTF-8?q?=E6=AD=A5=E5=B8=AD=E4=BD=8D=E6=89=A7=E8=A1=8C=E6=A8=A1=E5=BC=8F?=
 =?UTF-8?q?=EF=BC=88=E5=BF=AB=E9=80=9F/=E6=80=9D=E8=80=83/=E4=B8=93?=
 =?UTF-8?q?=E4=B8=9A=E9=97=AE=E7=AD=94/=E5=A4=9A=E6=99=BA=E8=83=BD?=
 =?UTF-8?q?=E4=BD=93=EF=BC=89+=20=E5=AD=90=E6=99=BA=E8=83=BD=E4=BD=93?=
 =?UTF-8?q?=E4=BB=BB=E5=8A=A1=E5=8D=A1=E7=89=87=20+=20=E5=90=8E=E5=8F=B0?=
 =?UTF-8?q?=E5=AE=8C=E6=88=90=E8=B7=A8=E9=A1=B5=E9=80=9A=E7=9F=A5=20+=20?=
 =?UTF-8?q?=E9=93=BE=E6=9D=A1=E9=BB=98=E8=AE=A4=E6=A8=A1=E5=BC=8F=E6=8C=81?=
 =?UTF-8?q?=E4=B9=85=E5=8C=96?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- 席位执行模式：选择参会智能体弹窗加模式选择器，映射 thinking/reasoning/subagent
  透传 recommend/chain/dag 三条编排；后端 _special_run 覆盖 seat policy（仅席位，
  总控不变）；ultra 解锁席位 task 子代理（custom 流回传 task_running 实时进度）。
- 子智能体任务卡片：复用主聊天 SubtaskCard + SubtaskContext，从席位流解析 task
  调用/结果，在 Step2 气泡渲染多张任务卡片（覆盖业务链条路径）。
- 后台挂起也按所选模式跑（StartParams→inprocess gateway→run_seat 覆盖 policy）。
- 后台作业完成全工作区跨页通知：RoundtableJobNotifier 挂在 PageLayout 常驻层，
  作业终态弹「达成共识」同款浮窗。
- 业务链条可配「默认执行模式」并持久化（roundtable_chains.seat_mode；选链时弹窗
  默认用它，仍可临时改）。
- 修复后台 Step3 HTML 看板概率截断：deepseek max_tokens=4096 单轮写不完长 HTML，
  产物自愈重试上限 2→5，让 append 续写写完整（同惠及 summary/dashboard）。

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---
 frontend-web/src/pages/PageRoutes.tsx         |   3 +
 .../src/roundtable-planning/api/chains.ts     |   9 +
 .../roundtable-planning/api/multi-agent.ts    |  13 ++
 .../api/roundtable-jobs.ts                    |  11 +
 .../components/BusinessChainPicker.tsx        |  14 +-
 .../components/MessageStepsCard.tsx           |   2 +
 .../components/RecommendAgentsDialog.tsx      |  55 ++++-
 .../components/RoundtableJobNotifier.tsx      | 125 ++++++++++
 .../components/Step2Panel.tsx                 |  28 +++
 .../hooks/useStep2Orchestration.ts            | 218 +++++++++++++++++-
 .../roundtable-planning/lib/business-chain.ts |  16 +-
 .../src/roundtable-planning/lib/job-types.ts  |   7 +
 .../src/roundtable-planning/lib/seat-mode.ts  |  77 +++++++
 .../pages/BusinessChainEditorPage.tsx         |  43 ++++
 .../pages/RoundtablePlanningPage.tsx          |  61 +++--
 .../backend/app/gateway/deps.py               |   9 +-
 .../gateway/roundtable_inprocess_gateway.py   |  43 +++-
 .../app/gateway/roundtable_job_executor.py    |   6 +
 .../app/gateway/routers/multi_agent.py        |  43 +++-
 .../app/gateway/routers/roundtable_chains.py  |  19 ++
 .../app/gateway/routers/roundtable_jobs.py    |   9 +
 .../versions/20260615_01_chain_seat_mode.py   |  50 ++++
 .../persistence/roundtable_chains/model.py    |   3 +
 .../persistence/roundtable_chains/sql.py      |   6 +
 .../scripts/sql/column_additions_mysql.sql    |   4 +
 25 files changed, 834 insertions(+), 40 deletions(-)
 create mode 100644 frontend-web/src/roundtable-planning/components/RoundtableJobNotifier.tsx
 create mode 100644 frontend-web/src/roundtable-planning/lib/seat-mode.ts
 create mode 100644 offline-backend-20260512/backend/packages/harness/deerflow/persistence/migrations/versions/20260615_01_chain_seat_mode.py

diff --git a/frontend-web/src/pages/PageRoutes.tsx b/frontend-web/src/pages/PageRoutes.tsx
index 178c1544..95b6ac62 100644
--- a/frontend-web/src/pages/PageRoutes.tsx
+++ b/frontend-web/src/pages/PageRoutes.tsx
@@ -15,6 +15,7 @@ import Header from "@/strategy-components/Header";
 import OpenCanvasRoutes from "@/open-canvas/routes/OpenCanvasRoutes";
 import RoundtableRoutes from "./RoundtableRoutes";
 import NotebookRoutes from "./NotebookRoutes";
+import { RoundtableJobNotifier } from "@/roundtable-planning/components/RoundtableJobNotifier";
 
 function PageLayout() {
   const [collapsed, setCollapsed] = useState(false);
@@ -35,6 +36,7 @@ function PageLayout() {
   // 高度链,根布局跟随容器高度。全屏下 100% 解析到视口,与 h-screen 等价;无界等嵌套
   // 下容器不占满整屏时,高度才能正确收缩而非锁死整屏。
   return (
+    <RoundtableJobNotifier>
     <div className="h-full flex flex-col bg-muted/20">
       {pageSidebarAllowed && <Header collapsed={collapsed} />}
       <div className="flex flex-1 overflow-hidden">
@@ -86,6 +88,7 @@ function PageLayout() {
         </main>
       </div>
     </div>
+    </RoundtableJobNotifier>
   );
 }
 
diff --git a/frontend-web/src/roundtable-planning/api/chains.ts b/frontend-web/src/roundtable-planning/api/chains.ts
index 4be2479f..57df4440 100644
--- a/frontend-web/src/roundtable-planning/api/chains.ts
+++ b/frontend-web/src/roundtable-planning/api/chains.ts
@@ -14,6 +14,7 @@ import { apiFetch } from "@/core/api/fetch-client";
 import { getBackendBaseURL } from "@/core/config";
 
 import type { BusinessChain, ChainSeat } from "@/roundtable-planning/lib/business-chain";
+import type { SeatMode } from "@/roundtable-planning/lib/seat-mode";
 
 /** Fields accepted when creating a chain. */
 export interface ChainCreatePayload {
@@ -28,6 +29,8 @@ export interface ChainCreatePayload {
   stageGoals?: (string | null)[] | null;
   /** Optional coordinator orchestration prompt (dag-only). null/"" = none. */
   coordinatorPrompt?: string | null;
+  /** 默认席位执行模式（flash/thinking/pro/ultra）；null = 未设。 */
+  seatMode?: SeatMode | null;
 }
 
 /** Partial update — send only what changed. */
@@ -41,6 +44,8 @@ export interface ChainUpdatePayload {
   stageGoals?: (string | null)[] | null;
   /** 写给总控的编排提示；send `null`/"" to clear, omit to leave. */
   coordinatorPrompt?: string | null;
+  /** 默认席位执行模式；send `null` to clear, omit to leave. */
+  seatMode?: SeatMode | null;
   /** 发布开关：true 发布为公共 / false 取消发布；omit 不改。仅 owner 生效。 */
   isPublic?: boolean;
 }
@@ -70,6 +75,7 @@ interface ServerChain {
   stages?: string[][] | null;
   stage_goals?: (string | null)[] | null;
   coordinator_prompt?: string | null;
+  seat_mode?: string | null;
   is_public?: boolean;
   is_mine?: boolean;
   owner_name?: string | null;
@@ -86,6 +92,7 @@ function toChain(s: ServerChain): BusinessChain {
     stages: Array.isArray(s.stages) ? s.stages : null,
     stageGoals: Array.isArray(s.stage_goals) ? s.stage_goals : null,
     coordinatorPrompt: s.coordinator_prompt ?? null,
+    seatMode: (s.seat_mode as SeatMode | null) ?? undefined,
     isPublic: s.is_public ?? false,
     // 后端缺省时按「我的」处理(向后兼容旧响应)。
     isMine: s.is_mine ?? true,
@@ -104,6 +111,7 @@ function fromCreate(p: ChainCreatePayload) {
     ...(p.stages !== undefined ? { stages: p.stages } : {}),
     ...(p.stageGoals !== undefined ? { stage_goals: p.stageGoals } : {}),
     ...(p.coordinatorPrompt !== undefined ? { coordinator_prompt: p.coordinatorPrompt } : {}),
+    ...(p.seatMode !== undefined ? { seat_mode: p.seatMode } : {}),
   };
 }
 
@@ -117,6 +125,7 @@ function fromUpdate(p: ChainUpdatePayload) {
   // Forward stage goals even when explicitly null (clears goals server-side).
   if (p.stageGoals !== undefined) out.stage_goals = p.stageGoals;
   if (p.coordinatorPrompt !== undefined) out.coordinator_prompt = p.coordinatorPrompt;
+  if (p.seatMode !== undefined) out.seat_mode = p.seatMode;
   if (p.isPublic !== undefined) out.is_public = p.isPublic;
   return out;
 }
diff --git a/frontend-web/src/roundtable-planning/api/multi-agent.ts b/frontend-web/src/roundtable-planning/api/multi-agent.ts
index bdb833f2..581f3499 100644
--- a/frontend-web/src/roundtable-planning/api/multi-agent.ts
+++ b/frontend-web/src/roundtable-planning/api/multi-agent.ts
@@ -202,6 +202,15 @@ export interface StreamMultiAgentRequest {
   displayName?: string;
   newMessage: string;
   model?: string;
+  /**
+   * 仅 special 用。席位执行模式（快速/思考/专业问答/多智能体问答）派生的运行参数，
+   * 覆盖后端 `roundtable_run_policy` 的 thinking/reasoning 默认值。省略时后端按 policy
+   * 默认执行（向后兼容，旧前端不传不破坏）。三者通常一起传（见 `lib/seat-mode.ts`）。
+   */
+  thinkingEnabled?: boolean;
+  reasoningEffort?: "minimal" | "low" | "medium" | "high";
+  /** 仅 special 用；ultra 档置真，给席位解锁 `task` 子代理工具。 */
+  subagentEnabled?: boolean;
   skillStopNames?: string[];
   /**
    * 仅 leader 用。置 true 表示这是「派活后让总控判断继续/综合」的轮次:后端会把各已交付
@@ -444,6 +453,10 @@ export async function streamMultiAgent(
       display_name: req.displayName,
       new_message: req.newMessage,
       model: req.model,
+      // 席位模式派生参数（仅 special）；undefined 时不带上 → 后端按 policy 默认执行。
+      thinking_enabled: req.thinkingEnabled,
+      reasoning_effort: req.reasoningEffort,
+      subagent_enabled: req.subagentEnabled,
       skill_stop_names: req.skillStopNames ?? [],
       // 仅 leader 用;综合轮让后端注入各席位完整交付正文。省略 → 后端按 false 处理。
       synthesis_mode: req.synthesisMode ?? false,
diff --git a/frontend-web/src/roundtable-planning/api/roundtable-jobs.ts b/frontend-web/src/roundtable-planning/api/roundtable-jobs.ts
index b0f44067..d766d0c3 100644
--- a/frontend-web/src/roundtable-planning/api/roundtable-jobs.ts
+++ b/frontend-web/src/roundtable-planning/api/roundtable-jobs.ts
@@ -15,6 +15,7 @@ import { apiFetch } from "@/core/api/fetch-client";
 import { getBackendBaseURL } from "@/core/config";
 
 import { getAvatarTypeFor } from "../lib/roundtable-constants";
+import { seatModeToRunParams } from "../lib/seat-mode";
 import {
   isTerminalJobStatus,
   type DispatchChainNode,
@@ -75,6 +76,9 @@ function mapJob(raw: Record<string, unknown>): RoundtableJob {
 }
 
 function toStartBody(p: StartJobPayload) {
+  // 席位模式 → thinking/reasoning/subagent，与前端实时路径(/run/stream)同一份映射；
+  // 缺省(未选模式)时不带这三字段 → 后端按 policy 默认(≈pro)，向后兼容。
+  const seat = p.seatMode ? seatModeToRunParams(p.seatMode) : null;
   return {
     draftId: p.draftId ?? null,
     intent: p.intent,
@@ -86,6 +90,13 @@ function toStartBody(p: StartJobPayload) {
     seedMessage: p.seedMessage,
     threadIds: p.threadIds ?? null,
     coordinatorName: p.coordinatorName ?? "roundtable-coordinator",
+    ...(seat
+      ? {
+          thinking_enabled: seat.thinking_enabled,
+          reasoning_effort: seat.reasoning_effort,
+          subagent_enabled: seat.subagent_enabled,
+        }
+      : {}),
   };
 }
 
diff --git a/frontend-web/src/roundtable-planning/components/BusinessChainPicker.tsx b/frontend-web/src/roundtable-planning/components/BusinessChainPicker.tsx
index bb431011..f484cc6f 100644
--- a/frontend-web/src/roundtable-planning/components/BusinessChainPicker.tsx
+++ b/frontend-web/src/roundtable-planning/components/BusinessChainPicker.tsx
@@ -13,7 +13,7 @@
  *
  * 详见 `docs/multi-agent-business-chain-dev.md` §6.2 / §7 阶段 3。
  */
-import { useEffect, useMemo, useState, type ReactNode } from "react";
+import { useEffect, useMemo, useRef, useState, type ReactNode } from "react";
 import {
   ArrowRightIcon,
   CheckIcon,
@@ -44,12 +44,15 @@ export interface BusinessChainPickerProps {
   onCancel: () => void;
   /** 跳转到业务链条配置页（同时关闭弹窗）。 */
   onGoConfig: () => void;
+  /** 选中链条变化时回调（含其 seatMode），父弹窗据此把执行模式默认设为链条的配置。 */
+  onSelectedChainChange?: (chain: BusinessChain | null) => void;
 }
 
 export function BusinessChainPicker({
   onConfirm,
   onCancel,
   onGoConfig,
+  onSelectedChainChange,
 }: BusinessChainPickerProps) {
   const [chains, setChains] = useState<BusinessChain[]>([]);
   const [loading, setLoading] = useState(true);
@@ -102,6 +105,15 @@ export function BusinessChainPicker({
     () => chains.find((c) => c.id === selectedId) ?? null,
     [chains, selectedId],
   );
+  // 选中链条变化 → 通知父弹窗（用 ref 持有回调，只在 selectedChain 真正变化时触发，
+  // 避免父级每次 render 重置用户已手动改过的执行模式）。
+  const onSelectedChainChangeRef = useRef(onSelectedChainChange);
+  useEffect(() => {
+    onSelectedChainChangeRef.current = onSelectedChainChange;
+  }, [onSelectedChainChange]);
+  useEffect(() => {
+    onSelectedChainChangeRef.current?.(selectedChain);
+  }, [selectedChain]);
   const survivingSeats = useMemo(
     () => (selectedChain ? selectedChain.seats.filter((s) => isExisting(s.agent_id)) : []),
     [selectedChain, isExisting],
diff --git a/frontend-web/src/roundtable-planning/components/MessageStepsCard.tsx b/frontend-web/src/roundtable-planning/components/MessageStepsCard.tsx
index df5f9795..37656fb8 100644
--- a/frontend-web/src/roundtable-planning/components/MessageStepsCard.tsx
+++ b/frontend-web/src/roundtable-planning/components/MessageStepsCard.tsx
@@ -52,6 +52,8 @@ export interface MessageStepsCardProps {
 
 /** 工具调用阶段的 step 才进入折叠卡;非工具步骤(connecting/streaming/tool_result)归到「思考」段。 */
 function isToolStep(step: BubbleStep): boolean {
+  // `task`(派生子智能体)单独渲染成任务卡片(SubtaskCard),不在步骤折叠卡里重复展示。
+  if (step.key === "tool_calling:task") return false;
   return step.key.startsWith("tool_calling:") && step.key.length > "tool_calling:".length;
 }
 
diff --git a/frontend-web/src/roundtable-planning/components/RecommendAgentsDialog.tsx b/frontend-web/src/roundtable-planning/components/RecommendAgentsDialog.tsx
index c70d5e7c..61bb1de5 100644
--- a/frontend-web/src/roundtable-planning/components/RecommendAgentsDialog.tsx
+++ b/frontend-web/src/roundtable-planning/components/RecommendAgentsDialog.tsx
@@ -34,6 +34,11 @@ import { MessageStepsCard } from "./MessageStepsCard";
 import { BusinessChainPicker } from "./BusinessChainPicker";
 import type { IntentSummary } from "../api/intent";
 import type { SeatSelectionResult } from "../lib/business-chain";
+import {
+  DEFAULT_SEAT_MODE,
+  SEAT_MODE_OPTIONS,
+  type SeatMode,
+} from "../lib/seat-mode";
 import {
   filterRecommendCandidates,
   listRecommendHistory,
@@ -200,6 +205,9 @@ export function RecommendAgentsDialog({
   const navigate = useNavigate();
   // 席位来源模式：智能推荐 / 业务链条。默认**业务链条**。
   const [mode, setMode] = useState<"recommend" | "chain">("chain");
+  // 席位执行模式（快速/思考/专业问答/多智能体问答）——推荐 / 业务链条两种来源通用，
+  // 决定第二步每个子智能体席位的推理深度。默认 pro（= 改造前席位固定行为）。
+  const [seatMode, setSeatMode] = useState<SeatMode>(DEFAULT_SEAT_MODE);
   const [state, setState] = useState<InternalState>(INITIAL_STATE);
   const [selectedIds, setSelectedIds] = useState<Set<string>>(new Set());
 
@@ -515,6 +523,14 @@ export function RecommendAgentsDialog({
     return `已选 ${selectedCount} / ${MAX_SELECTED}`;
   })();
 
+  // 统一在回传前合并席位模式，推荐 / 业务链条两条确认路径共用，保证 seatMode 总被带上。
+  const handleConfirmWithMode = useCallback(
+    (result: SeatSelectionResult) => {
+      onConfirm({ ...result, seatMode });
+    },
+    [onConfirm, seatMode],
+  );
+
   const handleConfirm = useCallback(() => {
     if (!canConfirm) return;
     // Order: recommended-and-selected first (in recommender order), then
@@ -530,8 +546,8 @@ export function RecommendAgentsDialog({
     for (const c of otherCards) {
       if (selectedIds.has(c.agent_id)) ordered.push({ agent_id: c.agent_id, name: c.name });
     }
-    onConfirm({ agents: ordered, mode: "recommend" });
-  }, [canConfirm, recommendedCards, otherCards, selectedIds, onConfirm]);
+    handleConfirmWithMode({ agents: ordered, mode: "recommend" });
+  }, [canConfirm, recommendedCards, otherCards, selectedIds, handleConfirmWithMode]);
 
   const handleGoConfig = useCallback(() => {
     onCancel();
@@ -595,6 +611,35 @@ export function RecommendAgentsDialog({
           </button>
         </div>
 
+        {/* 席位执行模式：快速 / 思考 / 专业问答 / 多智能体问答。推荐 & 业务链条通用，
+            决定第二步每个子智能体席位的推理深度（thinking/reasoning/subagent）。 */}
+        <div className="shrink-0 space-y-1.5">
+          <div className="flex flex-wrap items-center gap-2">
+            <span className="text-xs font-medium text-muted-foreground">执行模式</span>
+            <div className="inline-flex rounded-lg border border-border bg-muted/40 p-0.5 text-xs font-medium">
+              {SEAT_MODE_OPTIONS.map((opt) => (
+                <button
+                  key={opt.value}
+                  type="button"
+                  onClick={() => setSeatMode(opt.value)}
+                  title={opt.description}
+                  className={cn(
+                    "rounded-md px-2.5 py-1 transition-colors",
+                    seatMode === opt.value
+                      ? "bg-card text-primary shadow-sm"
+                      : "text-muted-foreground hover:text-foreground",
+                  )}
+                >
+                  {opt.label}
+                </button>
+              ))}
+            </div>
+          </div>
+          <p className="text-[11px] leading-relaxed text-muted-foreground">
+            {SEAT_MODE_OPTIONS.find((o) => o.value === seatMode)?.description}
+          </p>
+        </div>
+
         {/* ===== 智能推荐模式（现状，零行为变更）===== */}
         {mode === "recommend" && (
           <>
@@ -750,9 +795,13 @@ export function RecommendAgentsDialog({
         {/* ===== 业务链条模式 ===== */}
         {mode === "chain" && (
           <BusinessChainPicker
-            onConfirm={onConfirm}
+            onConfirm={handleConfirmWithMode}
             onCancel={onCancel}
             onGoConfig={handleGoConfig}
+            onSelectedChainChange={(chain) => {
+              // 选中链条若配了默认执行模式 → 把顶部模式选择器默认设为它（用户仍可临时改）。
+              if (chain?.seatMode) setSeatMode(chain.seatMode);
+            }}
           />
         )}
       </DialogContent>
diff --git a/frontend-web/src/roundtable-planning/components/RoundtableJobNotifier.tsx b/frontend-web/src/roundtable-planning/components/RoundtableJobNotifier.tsx
new file mode 100644
index 00000000..4d61dbc4
--- /dev/null
+++ b/frontend-web/src/roundtable-planning/components/RoundtableJobNotifier.tsx
@@ -0,0 +1,125 @@
+/**
+ * 圆桌「后台挂起作业」全工作区跨页完成通知器。
+ *
+ * 挂在 `PageLayout`（`/page/*` 的常驻层，导航切页不卸载）里：圆桌页点「后台挂起」后
+ * 调 `watchJob(jobId)` 注册监听，本 Provider 用 `streamJob` 订阅该作业 SSE；作业跑到
+ * 终态时弹一个**与「达成共识」同款**的顶部浮窗（bg-primary + Sparkles + bounce），
+ * 因此用户切到聊天/任务/其它工作区页面也能收到完成提示。
+ *
+ * 自带浮窗（不走 sonner）—— 各工作区页各自挂了一个 sonner `<Toaster>`，若这里也用
+ * sonner 会出现多 Toaster 重复弹；自渲染浮窗由本常驻 Provider 独占，天然不重复。
+ *
+ * 不含「刷新恢复」：watchedJobIds 是内存态，整页刷新后丢失（本期范围）。Provider 缺失
+ * 时（如 embed 路由在 PageRoutes 之外）`watchJob` 为 no-op，调用方不报错。
+ */
+import {
+  createContext,
+  useCallback,
+  useContext,
+  useEffect,
+  useRef,
+  useState,
+  type ReactNode,
+} from "react";
+import { Sparkles } from "lucide-react";
+
+import { streamJob } from "@/roundtable-planning/api/roundtable-jobs";
+import type { RoundtableJob } from "@/roundtable-planning/lib/job-types";
+
+interface RoundtableJobNotifierValue {
+  /** 注册一个后台作业 id，作业完成（done/error）时弹全局提示。重复 id 自动去重。 */
+  watchJob: (jobId: string) => void;
+}
+
+const RoundtableJobNotifierContext = createContext<RoundtableJobNotifierValue>({
+  watchJob: () => {
+    /* 无 Provider（如 embed 路由）时 no-op，不报错 */
+  },
+});
+
+export function useRoundtableJobNotifier(): RoundtableJobNotifierValue {
+  return useContext(RoundtableJobNotifierContext);
+}
+
+/** 单个作业的 SSE 监听器（无 UI）：到终态回调 onTerminal；传输中断回调 onStreamError。 */
+function JobWatcher({
+  jobId,
+  onTerminal,
+  onStreamError,
+}: {
+  jobId: string;
+  onTerminal: (job: RoundtableJob) => void;
+  onStreamError: (jobId: string) => void;
+}) {
+  const onTerminalRef = useRef(onTerminal);
+  const onStreamErrorRef = useRef(onStreamError);
+  useEffect(() => {
+    onTerminalRef.current = onTerminal;
+    onStreamErrorRef.current = onStreamError;
+  }, [onTerminal, onStreamError]);
+  useEffect(() => {
+    const stop = streamJob(jobId, {
+      onProgress: () => {},
+      onDone: (job) => onTerminalRef.current(job),
+      onError: () => onStreamErrorRef.current(jobId),
+    });
+    return () => stop();
+  }, [jobId]);
+  return null;
+}
+
+/** 提示浮窗存活时长（ms）—— 比页内 toast(3.5s) 略长，给跨页用户更多时间看到。 */
+const NOTICE_DURATION_MS = 6000;
+
+export function RoundtableJobNotifier({ children }: { children: ReactNode }) {
+  const [watchedJobIds, setWatchedJobIds] = useState<string[]>([]);
+  const [notice, setNotice] = useState<string | null>(null);
+
+  const watchJob = useCallback((jobId: string) => {
+    setWatchedJobIds((prev) => (prev.includes(jobId) ? prev : [...prev, jobId]));
+  }, []);
+
+  const handleTerminal = useCallback((job: RoundtableJob) => {
+    setWatchedJobIds((prev) => prev.filter((id) => id !== job.id));
+    if (job.status === "done") {
+      setNotice("🎉 后台研讨已完成！可在「历史记录」查看结论与报告");
+    } else if (job.status === "error") {
+      setNotice("❌ 后台任务执行出错，请在「历史记录」中查看");
+    }
+    // cancelled：用户主动取消，不提示。
+  }, []);
+
+  const handleStreamError = useCallback((jobId: string) => {
+    // SSE 传输中断（非作业终态）：静默停止监听，不误报完成。
+    setWatchedJobIds((prev) => prev.filter((id) => id !== jobId));
+  }, []);
+
+  useEffect(() => {
+    if (!notice) return;
+    const timer = setTimeout(() => setNotice(null), NOTICE_DURATION_MS);
+    return () => clearTimeout(timer);
+  }, [notice]);
+
+  return (
+    <RoundtableJobNotifierContext.Provider value={{ watchJob }}>
+      {children}
+      {watchedJobIds.map((id) => (
+        <JobWatcher
+          key={id}
+          jobId={id}
+          onTerminal={handleTerminal}
+          onStreamError={handleStreamError}
+        />
+      ))}
+      {notice && (
+        <div
+          className="fixed top-4 left-1/2 z-[100] flex -translate-x-1/2 transform items-center gap-2 rounded-xl border border-primary/40 bg-primary px-4 py-2.5 text-xs font-medium text-white shadow-lg animate-bounce"
+          role="status"
+        >
+          <Sparkles className="h-4 w-4 shrink-0" />
+          <span>{notice}</span>
+        </div>
+      )}
+    </RoundtableJobNotifierContext.Provider>
+  );
+}
diff --git a/frontend-web/src/roundtable-planning/components/Step2Panel.tsx b/frontend-web/src/roundtable-planning/components/Step2Panel.tsx
index 128d2b05..7835841a 100644
--- a/frontend-web/src/roundtable-planning/components/Step2Panel.tsx
+++ b/frontend-web/src/roundtable-planning/components/Step2Panel.tsx
@@ -57,6 +57,8 @@ import {
   StreamingCursor,
 } from "@/roundtable-planning/components/MessageBubble";
 import { MessageStepsCard } from "@/roundtable-planning/components/MessageStepsCard";
+import { SubtaskCard } from "@/components/workspace/messages/subtask-card";
+import { useSubtask } from "@/core/tasks/context";
 import { ParallelSeatStrip } from "@/roundtable-planning/components/ParallelSeatStrip";
 import { DagFlowMiniPanel } from "@/roundtable-planning/components/DagFlowMiniPanel";
 import { useArtifacts } from "@/components/workspace/artifacts";
@@ -909,6 +911,15 @@ export function Step2Message({
           resolveAgentName={resolveAgentName}
         />
       )}
+      {/* 「多智能体问答」模式:本席位派生的子智能体任务卡片(复用主聊天 SubtaskCard,
+          从共享 SubtaskContext 取状态)。同时发起多个 task → 多张卡片纵向堆叠。 */}
+      {Array.isArray(msg.subtaskIds) && msg.subtaskIds.length > 0 && (
+        <div className="flex w-full flex-col gap-2">
+          {msg.subtaskIds.map((taskId) => (
+            <RoundtableSubtaskCard key={taskId} taskId={taskId} isLoading={isStreaming} />
+          ))}
+        </div>
+      )}
       {reasoning && (
         <ReasoningBlock streaming={isStreaming}>{reasoning}</ReasoningBlock>
       )}
@@ -932,6 +943,23 @@ export function Step2Message({
   );
 }
 
+/**
+ * 子智能体任务卡片包裹层：子任务内容是**实时态**,存放在共享 SubtaskContext(主聊天同一份);
+ * 草稿回显 / 历史记录等场景气泡上虽留有 subtaskId,但 context 里已无对应 Subtask —— 此时
+ * 返回 null 不渲染(而非让 SubtaskCard 的 `useSubtask(id)!` 取到 undefined 崩溃)。
+ */
+function RoundtableSubtaskCard({
+  taskId,
+  isLoading,
+}: {
+  taskId: string;
+  isLoading: boolean;
+}) {
+  const task = useSubtask(taskId);
+  if (!task) return null;
+  return <SubtaskCard taskId={taskId} isLoading={isLoading} />;
+}
+
 /**
  * 收集气泡里所有"写了文件"的产出路径,按出现顺序去重,聚合到一行文件卡片。
  *
diff --git a/frontend-web/src/roundtable-planning/hooks/useStep2Orchestration.ts b/frontend-web/src/roundtable-planning/hooks/useStep2Orchestration.ts
index 84f98a36..70bfb38f 100644
--- a/frontend-web/src/roundtable-planning/hooks/useStep2Orchestration.ts
+++ b/frontend-web/src/roundtable-planning/hooks/useStep2Orchestration.ts
@@ -27,6 +27,7 @@ import {
   type ThreadIdMap,
 } from "@/roundtable-planning/api/multi-agent";
 import { uploadFiles } from "@/core/uploads/api";
+import { useUpdateSubtask } from "@/core/tasks/context";
 
 /** 干预消息携带的上传文件引用（与 streamMultiAgent 的 files 字段同形）。 */
 type RunFileRef = NonNullable<StreamMultiAgentRequest["files"]>[number];
@@ -37,12 +38,56 @@ function takeOnce<T>(ref: { current: T | null }): T | undefined {
   ref.current = null;
   return v;
 }
+
+/**
+ * 从一帧里识别子智能体 custom 事件(task_started/running/completed/failed/cancelled/
+ * timed_out)。custom 流帧通常是裸 payload 对象 `{type:"task_*", task_id, ...}`;
+ * 开启 stream_subgraphs 时也可能被包成 `[namespace, payload]`,这里两种都兼容。
+ */
+function asSubagentTaskEvent(frame: unknown): {
+  type: string;
+  task_id: string;
+  message?: unknown;
+  result?: unknown;
+  error?: unknown;
+  description?: unknown;
+} | null {
+  const cand =
+    frame && typeof frame === "object" && !Array.isArray(frame)
+      ? (frame as Record<string, unknown>)
+      : Array.isArray(frame) && frame.length >= 2 && frame[1] && typeof frame[1] === "object"
+        ? (frame[1] as Record<string, unknown>)
+        : null;
+  if (!cand) return null;
+  const type = typeof cand.type === "string" ? cand.type : "";
+  const taskId = typeof cand.task_id === "string" ? cand.task_id : "";
+  if (!type.startsWith("task_") || !taskId) return null;
+  return { type, task_id: taskId, message: cand.message, result: cand.result, error: cand.error, description: cand.description };
+}
+
+/** 从 LangGraph 消息 chunk 的 content（string 或结构化 block 数组）里抽纯文本。 */
+function extractFrameText(content: unknown): string {
+  if (typeof content === "string") return content;
+  if (Array.isArray(content)) {
+    let out = "";
+    for (const item of content) {
+      if (typeof item === "string") out += item;
+      else if (item && typeof item === "object") {
+        const block = item as Record<string, unknown>;
+        if (typeof block.text === "string") out += block.text;
+      }
+    }
+    return out;
+  }
+  return "";
+}
 import {
   buildDagStageLeaderTask,
   pickDagSeatTask,
   type OrchestrationPlan,
 } from "@/roundtable-planning/lib/dag-types";
 import type { ChainNodeState } from "@/roundtable-planning/lib/job-types";
+import { seatModeToRequestFields, type SeatMode } from "@/roundtable-planning/lib/seat-mode";
 import {
   DEFAULT_SELECTED_AGENTS,
   MAX_CYCLES,
@@ -65,6 +110,12 @@ export type Step2Dialogue = {
   streaming?: boolean;
   /** special 席位气泡对应的 LangGraph thread,用于沙箱预览。 */
   agentThreadId?: string;
+  /**
+   * 「多智能体问答」模式下,本席位用 `task` 工具派生的子智能体任务 id 列表(=task 工具
+   * 调用的 tool_call_id)。渲染时据此从共享 SubtaskContext 取每个 Subtask 渲染任务卡片
+   * （复用主聊天 `SubtaskCard`,见 Step2Panel.Step2Message）。非 ultra 模式恒为空。
+   */
+  subtaskIds?: string[];
   steps?: Array<{
     id: string;
     key: string;
@@ -222,6 +273,12 @@ export interface UseStep2OrchestrationOptions {
   currentStep: number;
   /** 全局选中模型(顶栏「推理模型」下拉)。 */
   selectedModel: string;
+  /**
+   * 席位执行模式（快速/思考/专业问答/多智能体问答）。决定第二步**每个子智能体席位**
+   * 的推理深度（thinking/reasoning_effort/subagent）；总控派活不受影响。由「选择参会
+   * 智能体」弹窗确认时的 `SeatSelectionResult.seatMode` 决定。缺省 `pro`。
+   */
+  selectedMode: SeatMode;
   /**
    * 编排模式:
    *  - `recommend`(默认)—— 总控自由派活循环(`runOrchestration`);
@@ -376,6 +433,7 @@ export function useStep2Orchestration(
   const {
     currentStep,
     selectedModel,
+    selectedMode,
     orchestrationMode,
     orchestrationPlan,
     dagExecMode = "fast",
@@ -507,6 +565,8 @@ export function useStep2Orchestration(
   // runOrchestration 的 useCallback deps 大幅收窄,避免每次模型切换、agentConfigs
   // 改写都重建一遍循环函数 —— 之前没收窄时 useEffect 会因 deps 变化重复 init。
   const selectedModelRef = useRef<string>(selectedModel);
+  // 席位执行模式同样打到 ref，让席位调用在编排循环里读到最新值而不重建回调。
+  const selectedModeRef = useRef<SeatMode>(selectedMode);
   const agentConfigsRef = useRef<Record<number, AgentConfig>>(agentConfigs);
   const agentNameToRoleRef = useRef(agentNameToRole);
   const threadIdsRef = useRef<ThreadIdMap | null>(threadIds);
@@ -543,6 +603,9 @@ export function useStep2Orchestration(
   useEffect(() => {
     selectedModelRef.current = selectedModel;
   }, [selectedModel]);
+  useEffect(() => {
+    selectedModeRef.current = selectedMode;
+  }, [selectedMode]);
   useEffect(() => {
     agentConfigsRef.current = agentConfigs;
   }, [agentConfigs]);
@@ -945,6 +1008,137 @@ export function useStep2Orchestration(
     [upsertSeatToolCall],
   );
 
+  // ── 「多智能体问答」子智能体任务卡片(参照主聊天 message-list 的 task 识别)──────────
+  // ultra 模式下席位用 `task` 工具派生子智能体。从席位流的 messages-tuple 帧里:
+  //  (1) AI chunk 的 tool_calls 里 name==="task" → 建/更新一条 Subtask(in_progress);
+  //  (2) type==="tool" 且 tool_call_id 命中已知 task id → 据「Task Succeeded./failed./
+  //      timed out」前缀更新为 completed/failed(与主聊天 message-list.tsx 同源解析)。
+  // Subtask 存进共享 SubtaskContext(主聊天同一份),气泡只记 id 列表并渲染 <SubtaskCard>。
+  const updateSubtask = useUpdateSubtask();
+  // 已识别的 task tool_call_id —— 用于在收到 tool 结果帧时只认领「子智能体任务」结果,
+  // 不误把 write_file 等其它工具的结果当成子任务处理。
+  const knownTaskIdsRef = useRef<Set<string>>(new Set());
+
+  /** 把某个子任务 id 记到对应席位气泡上(新引用 → 触发气泡重渲染,卡片随之刷新状态)。 */
+  const touchBubbleSubtaskId = useCallback((bubbleId: number, taskId: string) => {
+    setStep2RoundtableDialogues((prev) =>
+      prev.map((d) => {
+        if (d.id !== bubbleId) return d;
+        const ids = d.subtaskIds ?? [];
+        // 每次都换新数组引用,确保状态(in_progress→completed/failed)变化时卡片重渲染。
+        return { ...d, subtaskIds: ids.includes(taskId) ? [...ids] : [...ids, taskId] };
+      }),
+    );
+  }, []);
+
+  /** 处理一条消息(values 快照里的整条 message,或 messages-tuple 的 chunk)的子任务语义。 */
+  const ingestSubagentMessage = useCallback(
+    (msg: Record<string, unknown>, bubbleId: number) => {
+      // (1) AI 消息：抓 task 工具调用 → 建/更新 in_progress 子任务。
+      const toolCalls = msg.tool_calls;
+      if (Array.isArray(toolCalls)) {
+        for (const tc of toolCalls) {
+          if (!tc || typeof tc !== "object") continue;
+          const t = tc as Record<string, unknown>;
+          if (t.name !== "task") continue;
+          const id = typeof t.id === "string" ? t.id : "";
+          if (!id) continue;
+          const a = (t.args && typeof t.args === "object" ? t.args : {}) as Record<string, unknown>;
+          updateSubtask({
+            id,
+            status: "in_progress",
+            subagent_type: typeof a.subagent_type === "string" ? a.subagent_type : "",
+            description: typeof a.description === "string" ? a.description : "",
+            prompt: typeof a.prompt === "string" ? a.prompt : "",
+          });
+          knownTaskIdsRef.current.add(id);
+          touchBubbleSubtaskId(bubbleId, id);
+        }
+      }
+      // (2) tool 结果消息：命中已知 task id 时按前缀判完成/失败（与主聊天解析一致）。
+      const type = typeof msg.type === "string" ? msg.type.toLowerCase() : "";
+      if (type === "tool") {
+        const taskId = typeof msg.tool_call_id === "string" ? msg.tool_call_id : "";
+        if (taskId && knownTaskIdsRef.current.has(taskId)) {
+          const result = extractFrameText(msg.content);
+          if (result.startsWith("Task Succeeded. Result:")) {
+            updateSubtask({
+              id: taskId,
+              status: "completed",
+              result: result.split("Task Succeeded. Result:")[1]?.trim(),
+            });
+          } else if (result.startsWith("Task failed.")) {
+            updateSubtask({
+              id: taskId,
+              status: "failed",
+              error: result.split("Task failed.")[1]?.trim(),
+            });
+          } else if (result.startsWith("Task timed out")) {
+            updateSubtask({ id: taskId, status: "failed", error: result });
+          }
+          touchBubbleSubtaskId(bubbleId, taskId);
+        }
+      }
+    },
+    [updateSubtask, touchBubbleSubtaskId],
+  );
+
+  const ingestSubagentTaskFrame = useCallback(
+    (frame: unknown, bubbleId: number) => {
+      // custom 事件帧(ultra 席位子智能体进度)：task_running 带子智能体最新消息 → 回填
+      // latestMessage(卡片实时显示「正在调用 X 工具」);完成/失败即时定状态(比 values 更快)。
+      const ev = asSubagentTaskEvent(frame);
+      if (ev) {
+        knownTaskIdsRef.current.add(ev.task_id);
+        if (ev.type === "task_running") {
+          // latestMessage 非空才会触发 SubtaskContext setTasks 重渲染(见 useUpdateSubtask)。
+          updateSubtask({ id: ev.task_id, status: "in_progress", latestMessage: ev.message as never });
+        } else if (ev.type === "task_started") {
+          updateSubtask({
+            id: ev.task_id,
+            status: "in_progress",
+            description: typeof ev.description === "string" ? ev.description : undefined,
+          });
+        } else if (ev.type === "task_completed") {
+          updateSubtask({
+            id: ev.task_id,
+            status: "completed",
+            result: typeof ev.result === "string" ? ev.result : extractFrameText(ev.result),
+          });
+        } else if (
+          ev.type === "task_failed" ||
+          ev.type === "task_cancelled" ||
+          ev.type === "task_timed_out"
+        ) {
+          updateSubtask({
+            id: ev.task_id,
+            status: "failed",
+            error: typeof ev.error === "string" ? ev.error : "子智能体任务失败",
+          });
+        }
+        touchBubbleSubtaskId(bubbleId, ev.task_id);
+        return;
+      }
+
+      // values 快照帧：{messages:[...]} —— 权威来源，按整条消息列表对账(子任务完成/失败)。
+      if (frame && typeof frame === "object" && !Array.isArray(frame)) {
+        const msgs = (frame as Record<string, unknown>).messages;
+        if (Array.isArray(msgs)) {
+          for (const m of msgs) {
+            if (m && typeof m === "object") ingestSubagentMessage(m as Record<string, unknown>, bubbleId);
+          }
+        }
+        return;
+      }
+      // messages-tuple 帧：[chunk, metadata] —— 增量，派活时即时建卡。
+      if (!Array.isArray(frame) || frame.length === 0) return;
+      const chunk = frame[0];
+      if (!chunk || typeof chunk !== "object") return;
+      ingestSubagentMessage(chunk as Record<string, unknown>, bubbleId);
+    },
+    [ingestSubagentMessage, updateSubtask, touchBubbleSubtaskId],
+  );
+
   // ─── Dialogue / Speaker helpers ────────────────────────────────────
   const appendDialogue = useCallback(
     (card: {
@@ -1406,6 +1600,8 @@ export function useStep2Orchestration(
               newMessage: fullTask,
               peerDeliveries: priorDeliveries,
               model: perSeatModel || selectedModelRef.current || undefined,
+              // 席位执行模式（快速/思考/专业问答/多智能体问答）→ thinking/reasoning/subagent。
+              ...seatModeToRequestFields(selectedModeRef.current),
               signal: subAborter.signal,
               onTextDelta: (text) =>
                 a.isParallel ? ingestParallelDelta(a.bubbleId, text) : appendToDialogue(a.bubbleId, text),
@@ -1415,7 +1611,10 @@ export function useStep2Orchestration(
                 appendStepIfNew(a.bubbleId, label, `${phase}:${toolName ?? ""}`, args);
                 handleArtifactToolPhase(phase, toolName, args, a.seatThreadId);
               },
-              onUpstreamFrame: (frame) => handleUpstreamFrame(frame, a.seatThreadId),
+              onUpstreamFrame: (frame) => {
+                handleUpstreamFrame(frame, a.seatThreadId);
+                ingestSubagentTaskFrame(frame, a.bubbleId);
+              },
             });
             activeAbortersRef.current.delete(subAborter);
             if (isStale()) {
@@ -1690,6 +1889,7 @@ export function useStep2Orchestration(
       triggerToast,
       handleArtifactToolPhase,
       handleUpstreamFrame,
+      ingestSubagentTaskFrame,
       resolveSeatThreadId,
     ],
   );
@@ -1829,6 +2029,8 @@ export function useStep2Orchestration(
               displayName: seatName,
               newMessage: task,
               model: perSeatModel || selectedModelRef.current || undefined,
+              // 席位执行模式 → thinking/reasoning/subagent（与推荐/DAG 路径一致）。
+              ...seatModeToRequestFields(selectedModeRef.current),
               signal: subAborter.signal,
               onTextDelta: (text) => appendToDialogue(subBubbleId, text),
               onPhaseChange: (phase, toolName, args) => {
@@ -1837,7 +2039,10 @@ export function useStep2Orchestration(
                 appendStepIfNew(subBubbleId, label, `${phase}:${toolName ?? ""}`, args);
                 handleArtifactToolPhase(phase, toolName, args, seatThreadId);
               },
-              onUpstreamFrame: (frame) => handleUpstreamFrame(frame, seatThreadId),
+              onUpstreamFrame: (frame) => {
+                handleUpstreamFrame(frame, seatThreadId);
+                ingestSubagentTaskFrame(frame, subBubbleId);
+              },
             });
             if (isStale()) {
               finalizeDialogue(subBubbleId, specRes.content || "(已停止)", "已停止");
@@ -1936,6 +2141,7 @@ export function useStep2Orchestration(
       triggerToast,
       handleArtifactToolPhase,
       handleUpstreamFrame,
+      ingestSubagentTaskFrame,
       resolveSeatThreadId,
     ],
   );
@@ -2163,6 +2369,8 @@ export function useStep2Orchestration(
                   // 前序各阶段的完整交付 → 供本席位 read_peer_delivery 按需取全文(§12.9)。
                   peerDeliveries: priorDeliveries,
                   model: perSeatModel || selectedModelRef.current || undefined,
+                  // 席位执行模式 → thinking/reasoning/subagent（与推荐/链条路径一致）。
+                  ...seatModeToRequestFields(selectedModeRef.current),
                   signal: subAborter.signal,
                   // §6：并行席位走 ref 全量累积 + 节流 flush(只渲染选中的);单节点
                   // stage 实际串行,1 路流直接 setState 不卡,沿用 appendToDialogue。
@@ -2174,7 +2382,10 @@ export function useStep2Orchestration(
                     appendStepIfNew(bubbleId, label, `${phase}:${toolName ?? ""}`, args);
                     handleArtifactToolPhase(phase, toolName, args, seatThreadId);
                   },
-                  onUpstreamFrame: (frame) => handleUpstreamFrame(frame, seatThreadId),
+                  onUpstreamFrame: (frame) => {
+                    handleUpstreamFrame(frame, seatThreadId);
+                    ingestSubagentTaskFrame(frame, bubbleId);
+                  },
                 });
                 if (isStale()) {
                   finalizeDialogue(bubbleId, specRes.content || "(已停止)", "已停止");
@@ -2276,6 +2487,7 @@ export function useStep2Orchestration(
       triggerToast,
       handleArtifactToolPhase,
       handleUpstreamFrame,
+      ingestSubagentTaskFrame,
       resolveSeatThreadId,
       ingestParallelDelta,
       resetParallelStreams,
diff --git a/frontend-web/src/roundtable-planning/lib/business-chain.ts b/frontend-web/src/roundtable-planning/lib/business-chain.ts
index 414ea3cf..a8a824f4 100644
--- a/frontend-web/src/roundtable-planning/lib/business-chain.ts
+++ b/frontend-web/src/roundtable-planning/lib/business-chain.ts
@@ -10,6 +10,7 @@
  */
 
 import { buildPlanFromStages } from "./dag-types";
+import type { SeatMode } from "./seat-mode";
 
 /** 链条里的一个有序席位。`seats[0]` 是第一个发言者。 */
 export interface ChainSeat {
@@ -48,6 +49,11 @@ export interface BusinessChain {
    * 空 / null = 不注入。
    */
   coordinatorPrompt?: string | null;
+  /**
+   * 默认席位执行模式（快速/思考/专业问答/多智能体）。选入研讨时弹窗默认用它（仍可临时改）。
+   * 缺省 = 未设，调用方回退 `DEFAULT_SEAT_MODE`(pro)。
+   */
+  seatMode?: SeatMode;
   /** 是否已发布为公共（所有账号可在「公共」筛选查看 / 选入研讨）。 */
   isPublic?: boolean;
   /** 是否当前用户所有（仅本人可编辑 / 删除 / 发布；公共筛选里别人的链条为 false）。 */
@@ -85,6 +91,12 @@ export interface SeatSelectionResult {
    * 缺省 `fast`。
    */
   dagExecMode?: "fast" | "sandbox";
+  /**
+   * 席位执行模式（快速/思考/专业问答/多智能体问答）——决定第二步**每个子智能体席位**
+   * 的推理深度（thinking/reasoning_effort/subagent）。三种编排（recommend/chain/dag）通用。
+   * 缺省由调用方按 `DEFAULT_SEAT_MODE`（pro）处理。
+   */
+  seatMode?: SeatMode;
 }
 
 /** 链条席位数硬约束，与圆桌推荐弹窗的 `MIN_SELECTED`/`MAX_SELECTED` 对齐。 */
@@ -132,12 +144,12 @@ export function buildChainSeatSelection(
       const agents = plan.stages
         .flatMap((st) => st.agentIds)
         .map((id) => ({ agent_id: id, name: nameOf(id) }));
-      return { agents, mode: "dag", chain: chainMeta, plan, dagExecMode: "fast" };
+      return { agents, mode: "dag", chain: chainMeta, plan, dagExecMode: "fast", seatMode: chain.seatMode };
     }
   }
 
   const agents = survivingSeats.map((s) => ({ agent_id: s.agent_id, name: nameOf(s.agent_id) }));
-  return { agents, mode: "chain", chain: chainMeta };
+  return { agents, mode: "chain", chain: chainMeta, seatMode: chain.seatMode };
 }
 
 /**
diff --git a/frontend-web/src/roundtable-planning/lib/job-types.ts b/frontend-web/src/roundtable-planning/lib/job-types.ts
index 9d995c34..d86c286b 100644
--- a/frontend-web/src/roundtable-planning/lib/job-types.ts
+++ b/frontend-web/src/roundtable-planning/lib/job-types.ts
@@ -9,6 +9,7 @@
  */
 
 import type { OrchestrationPlan } from "./dag-types";
+import type { SeatMode } from "./seat-mode";
 
 /** 作业状态机（§2.1）。 */
 export type JobStatus =
@@ -120,6 +121,12 @@ export interface StartJobPayload {
   /** 预建好的 thread 映射（Phase 3b 真网关用；模拟网关忽略）。 */
   threadIds?: Record<string, string> | null;
   coordinatorName?: string;
+  /**
+   * 席位执行模式（快速/思考/专业问答/多智能体）。后台续跑时席位按此模式执行（与前端实时
+   * 路径一致）；缺省由后端按 policy 默认(≈pro)。在 `toStartBody` 里派生成 thinking/
+   * reasoning/subagent 三字段发给后端。
+   */
+  seatMode?: SeatMode;
 }
 
 /** 终态集合——SSE/轮询判定作业是否结束。 */
diff --git a/frontend-web/src/roundtable-planning/lib/seat-mode.ts b/frontend-web/src/roundtable-planning/lib/seat-mode.ts
new file mode 100644
index 00000000..78d75aae
--- /dev/null
+++ b/frontend-web/src/roundtable-planning/lib/seat-mode.ts
@@ -0,0 +1,77 @@
+/**
+ * 圆桌第二步「席位执行模式」共享定义。
+ *
+ * 与主聊天输入框（`components/workspace/input-box.tsx`）的四档模式同源：
+ * 用户在「选择参会智能体」弹窗里选一档，第二步派活的**每个子智能体（special 席位）**
+ * 都按这一档的推理深度执行。模式只决定 `thinking_enabled / reasoning_effort /
+ * subagent_enabled` 三个运行参数；席位的工具裁剪仍由后端 `roundtable_run_policy`
+ * 决定（见 `app/gateway/roundtable_run_policy.py`）。
+ *
+ * 总控（leader）派活是短决策任务，**不**受本模式影响（沿用其快速派活策略）。
+ *
+ * 映射逻辑刻意对齐主聊天 `core/threads/hooks.ts` 的 mode → context 转换，保持一致体验：
+ *  - flash    快速      —— 不思考、最快
+ *  - thinking 思考      —— 思考 + low
+ *  - pro      专业问答  —— 思考 + medium（= 现有席位默认行为，向后兼容）
+ *  - ultra    多智能体  —— 思考 + high + 席位可派子代理（task 工具）
+ */
+
+export type SeatMode = "flash" | "thinking" | "pro" | "ultra";
+
+/** 默认档位：专业问答（pro）—— 与改造前席位的固定行为（thinking + medium）一致。 */
+export const DEFAULT_SEAT_MODE: SeatMode = "pro";
+
+/** 席位单轮 run 的派生参数（透传给 `/api/multi-agent/run/stream`，覆盖后端 policy 默认）。 */
+export interface SeatRunParams {
+  thinking_enabled: boolean;
+  reasoning_effort: "minimal" | "low" | "medium" | "high";
+  subagent_enabled: boolean;
+}
+
+/** 把一个席位模式映射成具体的运行参数。 */
+export function seatModeToRunParams(mode: SeatMode): SeatRunParams {
+  switch (mode) {
+    case "flash":
+      return { thinking_enabled: false, reasoning_effort: "minimal", subagent_enabled: false };
+    case "thinking":
+      return { thinking_enabled: true, reasoning_effort: "low", subagent_enabled: false };
+    case "ultra":
+      return { thinking_enabled: true, reasoning_effort: "high", subagent_enabled: true };
+    case "pro":
+    default:
+      return { thinking_enabled: true, reasoning_effort: "medium", subagent_enabled: false };
+  }
+}
+
+/**
+ * 把席位模式映射成 `streamMultiAgent` 请求字段（camelCase），方便在席位调用处直接 spread。
+ * 与 `seatModeToRunParams` 同一份逻辑，只是改成 API 客户端期望的字段名。
+ */
+export function seatModeToRequestFields(mode: SeatMode): {
+  thinkingEnabled: boolean;
+  reasoningEffort: SeatRunParams["reasoning_effort"];
+  subagentEnabled: boolean;
+} {
+  const p = seatModeToRunParams(mode);
+  return {
+    thinkingEnabled: p.thinking_enabled,
+    reasoningEffort: p.reasoning_effort,
+    subagentEnabled: p.subagent_enabled,
+  };
+}
+
+/** 弹窗模式选择器用的展示元数据（标签 + 一句话说明），顺序即展示顺序。 */
+export const SEAT_MODE_OPTIONS: ReadonlyArray<{
+  value: SeatMode;
+  label: string;
+  description: string;
+}> = [
+  { value: "flash", label: "快速", description: "快速且高效地完成任务，但可能不够精准" },
+  { value: "thinking", label: "思考", description: "思考后再行动，在时间与准确性之间取得平衡" },
+  { value: "pro", label: "专业问答", description: "思考、计划再执行，获得更精准的结果，可能需要更多时间" },
+  {
+    value: "ultra",
+    label: "多智能体问答",
+    description: "在专业问答基础上，每个席位可调用子代理分工协作，能力最强、更慢更耗时",
+  },
+];
diff --git a/frontend-web/src/roundtable-planning/pages/BusinessChainEditorPage.tsx b/frontend-web/src/roundtable-planning/pages/BusinessChainEditorPage.tsx
index 5587e1fa..4eabfe46 100644
--- a/frontend-web/src/roundtable-planning/pages/BusinessChainEditorPage.tsx
+++ b/frontend-web/src/roundtable-planning/pages/BusinessChainEditorPage.tsx
@@ -85,6 +85,11 @@ import {
   type BusinessChain,
   type ChainSeat,
 } from "@/roundtable-planning/lib/business-chain";
+import {
+  DEFAULT_SEAT_MODE,
+  SEAT_MODE_OPTIONS,
+  type SeatMode,
+} from "@/roundtable-planning/lib/seat-mode";
 import {
   addEmptyStage,
   addSeatAsNewStage,
@@ -153,6 +158,8 @@ export default function BusinessChainEditorPage() {
   const [description, setDescription] = useState("");
   // 写给总控的「编排提示」：告诉总控该怎么派活、做什么（仅分层 DAG 模式注入）。
   const [coordinatorPrompt, setCoordinatorPrompt] = useState("");
+  // 默认席位执行模式（快速/思考/专业问答/多智能体）—— 选入研讨时弹窗默认用它。
+  const [seatMode, setSeatMode] = useState<SeatMode>(DEFAULT_SEAT_MODE);
   // 分层编排是唯一数据源；落库时展平为 seats + 折叠为 stages。
   const [stageGroups, setStageGroups] = useState<StageGroups>([]);
   const [saving, setSaving] = useState(false);
@@ -277,6 +284,7 @@ export default function BusinessChainEditorPage() {
     setTitle(chain.title);
     setDescription(chain.description ?? "");
     setCoordinatorPrompt(chain.coordinatorPrompt ?? "");
+    setSeatMode(chain.seatMode ?? DEFAULT_SEAT_MODE);
     setStageGroups(chainToStageGroups(chain));
     setStep(1);
   }, []);
@@ -288,6 +296,7 @@ export default function BusinessChainEditorPage() {
     setTitle("");
     setDescription("");
     setCoordinatorPrompt("");
+    setSeatMode(DEFAULT_SEAT_MODE);
     setStageGroups([]);
     setStep(1);
   }, []);
@@ -560,6 +569,7 @@ export default function BusinessChainEditorPage() {
           stages,
           stageGoals,
           coordinatorPrompt: coordPrompt,
+          seatMode,
         });
         if (!updated) {
           toast.error("链条已不存在，可能被删除");
@@ -578,6 +588,7 @@ export default function BusinessChainEditorPage() {
           stages,
           stageGoals,
           coordinatorPrompt: coordPrompt,
+          seatMode,
         });
         setCurrentChainId(created.id);
         toast.success("已创建业务链条");
@@ -592,6 +603,7 @@ export default function BusinessChainEditorPage() {
     title,
     description,
     coordinatorPrompt,
+    seatMode,
     stageGroups,
     seatCountInvalid,
     currentChainId,
@@ -699,6 +711,37 @@ export default function BusinessChainEditorPage() {
         />
       </div>
 
+      {/* 默认执行模式：选入研讨时弹窗默认用它（仍可临时改）。 */}
+      <div className="space-y-2">
+        <label className="flex items-center gap-1 text-xs font-semibold text-foreground">
+          默认执行模式
+          <span className="text-[10px] font-normal text-muted-foreground">
+            （选入研讨时默认用它，进研讨前仍可临时切换）
+          </span>
+        </label>
+        <div className="inline-flex flex-wrap gap-0.5 rounded-lg border border-border bg-muted/40 p-0.5 text-xs font-medium">
+          {SEAT_MODE_OPTIONS.map((opt) => (
+            <button
+              key={opt.value}
+              type="button"
+              onClick={() => setSeatMode(opt.value)}
+              title={opt.description}
+              className={cn(
+                "rounded-md px-2.5 py-1 transition-colors",
+                seatMode === opt.value
+                  ? "bg-card text-primary shadow-sm"
+                  : "text-muted-foreground hover:text-foreground",
+              )}
+            >
+              {opt.label}
+            </button>
+          ))}
+        </div>
+        <p className="text-[11px] leading-relaxed text-muted-foreground">
+          {SEAT_MODE_OPTIONS.find((o) => o.value === seatMode)?.description}
+        </p>
+      </div>
+
       {/* 总控提示词：label + AI 优化（配色沿用现有 primary）+ 大号等宽编辑区 */}
       <div className="space-y-2">
         <div className="flex items-center justify-between gap-2">
diff --git a/frontend-web/src/roundtable-planning/pages/RoundtablePlanningPage.tsx b/frontend-web/src/roundtable-planning/pages/RoundtablePlanningPage.tsx
index 1ef6d6ac..4f168b1d 100644
--- a/frontend-web/src/roundtable-planning/pages/RoundtablePlanningPage.tsx
+++ b/frontend-web/src/roundtable-planning/pages/RoundtablePlanningPage.tsx
@@ -100,6 +100,7 @@ import {
   buildChainSeatSelection,
   type SeatSelectionResult,
 } from "@/roundtable-planning/lib/business-chain";
+import { DEFAULT_SEAT_MODE, type SeatMode } from "@/roundtable-planning/lib/seat-mode";
 import type { OrchestrationPlan } from "@/roundtable-planning/lib/dag-types";
 import type { IntentSummary } from "@/roundtable-planning/api/intent";
 import {
@@ -108,6 +109,7 @@ import {
   startJob,
   startJobKeepalive,
 } from "@/roundtable-planning/api/roundtable-jobs";
+import { useRoundtableJobNotifier } from "@/roundtable-planning/components/RoundtableJobNotifier";
 import { getChain } from "@/roundtable-planning/api/chains";
 import { filterRecommendCandidates } from "@/roundtable-planning/api/recommend";
 import { listAgents } from "@/core/agents";
@@ -169,6 +171,8 @@ function readEmbedHashQueryPrefill(): RoundtablePrefill | null {
 
 export default function RoundtablePlanningPage() {
   const { theme } = useTheme();
+  // 全工作区跨页的后台作业完成通知器（Provider 挂在 PageLayout 常驻层）。
+  const { watchJob } = useRoundtableJobNotifier();
   const isDark = theme === "dark";
   const navigate = useNavigate();
   const location = useLocation();
@@ -214,6 +218,9 @@ export default function RoundtablePlanningPage() {
   // 「导入会商」弹窗（输入分享码把别人分享的会商复制到本账号）。
   const [isImportOpen, setIsImportOpen] = useState<boolean>(false);
   const [selectedModel, setSelectedModel] = useState<string>("");
+  // Step 2 席位执行模式（快速/思考/专业问答/多智能体问答）——决定每个子智能体席位的
+  // 推理深度。由「选择参会智能体」弹窗确认时的 SeatSelectionResult.seatMode 决定。
+  const [seatMode, setSeatMode] = useState<SeatMode>(DEFAULT_SEAT_MODE);
   // Step 2 编排模式：recommend = 总控自由派活；chain = 总控按业务链条顺序派活；
   // dag = 分层 DAG（stage 内并行、stage 间串行，配合 orchestrationPlan）。
   // 由推荐弹窗确认时的 SeatSelectionResult.mode 决定（见 handleConfirmAgents）。
@@ -329,6 +336,7 @@ export default function RoundtablePlanningPage() {
   const step2 = useStep2Orchestration({
     currentStep,
     selectedModel,
+    selectedMode: seatMode,
     orchestrationMode,
     orchestrationPlan,
     dagExecMode,
@@ -739,6 +747,8 @@ export default function RoundtablePlanningPage() {
       // result.mode 决定 Step 2 编排：recommend = 总控自由派活；chain = 按链条顺序派活；
       // dag = 分层并行（来源链条带 stages，携带 result.plan）。
       setOrchestrationMode(result.mode);
+      // 席位执行模式：弹窗未带（旧调用）时回退默认 pro，保证行为可预期。
+      setSeatMode(result.seatMode ?? DEFAULT_SEAT_MODE);
       setChainMeta(result.mode === "recommend" ? null : result.chain ?? null);
       setOrchestrationPlan(result.mode === "dag" ? result.plan ?? null : null);
       setDagExecMode(result.mode === "dag" ? result.dagExecMode ?? "fast" : "fast");
@@ -1118,9 +1128,11 @@ export default function RoundtablePlanningPage() {
   // 「后台挂起」核心（手动按钮 / 离开页面自动挂起共用）：幂等保证草稿存在 →
   // 同草稿已有在跑作业则跳过 → 启动后台作业。后端另有「同草稿未终态作业幂等返回」
   // 兜底（覆盖跨标签页 / beforeunload 与本函数并发触发），绝不重复建作业。
-  const startBackgroundJob = useCallback(async (): Promise<boolean> => {
+  // 返回本次后台作业的 jobId（已在跑则返回既有作业 id），供「完成时提示」监听；
+  // 未发起（无席位）返回 null。
+  const startBackgroundJob = useCallback(async (): Promise<string | null> => {
     const agents = step2.selectedAgents.map((a) => ({ agentId: a.agent_id, name: a.name }));
-    if (agents.length === 0) return false;
+    if (agents.length === 0) return null;
     // 后台挂起发起前**立刻强制存一次**当前聊天（保住发起时刻的本地研讨记录，防丢）。
     const draftId = await drafts.saveDraftNow(2);
     // 防重复：该草稿已有在跑作业就不再新建（避免重复作业 / 前后端双跑）。
@@ -1130,28 +1142,31 @@ export default function RoundtablePlanningPage() {
       (existing.status === "running" ||
         existing.status === "awaiting_input" ||
         existing.status === "queued");
-    if (!alreadyRunning) {
-      await startJob({
-        draftId: draftId ?? undefined,
-        intent: buildInitialIntent(),
-        agents,
-        model: selectedModel || undefined, // 模型沿用第二步选择
-        // 后台作业引擎支持 dag（后端 _run_dag：stage 串行 / stage 内并行）——
-        // 直接把当前编排模式透传，业务链的分层 / 并行 / 阶段目标 / 总控提示词全部保留。
-        orchestrationMode,
-        chain: chainMeta,
-        // dag 模式：把整份分层计划带给后端（含 stages / 阶段目标 / 总控提示词 / finalSynthesis）。
-        // 非 dag 时为 null（后端按 mode 走 chain/recommend）。
-        orchestrationPlan: orchestrationMode === "dag" ? orchestrationPlan : null,
-        // 不传 threadIds：后台作业由真网关自建全新线程，绝不复用前端本地研讨的线程
-        // （那些线程可能有活跃 run，复用会撞 409）。
-      });
+    if (alreadyRunning) {
+      return existing?.id ?? null;
     }
-    return true;
+    const job = await startJob({
+      draftId: draftId ?? undefined,
+      intent: buildInitialIntent(),
+      agents,
+      model: selectedModel || undefined, // 模型沿用第二步选择
+      seatMode, // 席位执行模式沿用第二步选择（后台续跑同样按此模式跑席位）
+      // 后台作业引擎支持 dag（后端 _run_dag：stage 串行 / stage 内并行）——
+      // 直接把当前编排模式透传，业务链的分层 / 并行 / 阶段目标 / 总控提示词全部保留。
+      orchestrationMode,
+      chain: chainMeta,
+      // dag 模式：把整份分层计划带给后端（含 stages / 阶段目标 / 总控提示词 / finalSynthesis）。
+      // 非 dag 时为 null（后端按 mode 走 chain/recommend）。
+      orchestrationPlan: orchestrationMode === "dag" ? orchestrationPlan : null,
+      // 不传 threadIds：后台作业由真网关自建全新线程，绝不复用前端本地研讨的线程
+      // （那些线程可能有活跃 run，复用会撞 409）。
+    });
+    return job.id;
   }, [
     step2,
     drafts,
     selectedModel,
+    seatMode,
     orchestrationMode,
     orchestrationPlan,
     chainMeta,
@@ -1165,17 +1180,20 @@ export default function RoundtablePlanningPage() {
       triggerToast("⚠️ 请先选择研讨席位再后台挂起");
       return;
     }
+    let jobId: string | null = null;
     try {
-      await startBackgroundJob();
+      jobId = await startBackgroundJob();
     } catch (err) {
       triggerToast(`❌ 后台挂起失败：${(err as Error).message}`);
       return;
     }
+    // 注册到全工作区通知器，作业完成时**跨页**弹「达成共识同款」提示。
+    if (jobId) watchJob(jobId);
     // 重置回第一步（同新建任务）；后台作业在历史记录里继续跑、可点开看进度。
     resetToStep1();
     triggerToast("🚀 已转入后台，任务会自动跑完研讨、总结报告与结果绘制；可在「历史记录」查看进度");
     void drafts.refresh();
-  }, [step2, drafts, startBackgroundJob, resetToStep1, triggerToast]);
+  }, [step2, drafts, startBackgroundJob, resetToStep1, triggerToast, watchJob]);
 
   // ── 自动后台挂起：研讨进行中离开本页面（路由卸载）或刷新 / 关标签页时，任务自动转后台 ──
   // 最新状态进 ref：卸载 cleanup 与 beforeunload 都是「拍快照时刻」的同步回调，
@@ -1205,6 +1223,7 @@ export default function RoundtablePlanningPage() {
           intent: buildInitialIntent(),
           agents,
           model: selectedModel || undefined,
+          seatMode,
           orchestrationMode,
           chain: chainMeta,
           orchestrationPlan: orchestrationMode === "dag" ? orchestrationPlan : null,
diff --git a/offline-backend-20260512/backend/app/gateway/deps.py b/offline-backend-20260512/backend/app/gateway/deps.py
index 109d32d2..7813ffc2 100644
--- a/offline-backend-20260512/backend/app/gateway/deps.py
+++ b/offline-backend-20260512/backend/app/gateway/deps.py
@@ -259,7 +259,14 @@ async def langgraph_runtime(app: FastAPI) -> AsyncGenerator[None, None]:
 
                 _app = app  # 闭包捕获 app，供网关访问 app.state（checkpointer/thread_store）
                 gateway_factory = lambda p: InProcessRoundtableGateway(  # noqa: E731
-                    _app, user_id=p.user_id, agents=p.agents, model=p.model
+                    _app,
+                    user_id=p.user_id,
+                    agents=p.agents,
+                    model=p.model,
+                    # 席位执行模式沿用 Step2 选择，后台席位 run 据此覆盖 policy 的 thinking/reasoning + 开子代理。
+                    seat_thinking_enabled=p.seat_thinking_enabled,
+                    seat_reasoning_effort=p.seat_reasoning_effort,
+                    seat_subagent_enabled=p.seat_subagent_enabled,
                 )
                 logger.info("[roundtable-jobs] gateway = inprocess (real models, default)")
 
diff --git a/offline-backend-20260512/backend/app/gateway/roundtable_inprocess_gateway.py b/offline-backend-20260512/backend/app/gateway/roundtable_inprocess_gateway.py
index 948075c8..b0c494d4 100644
--- a/offline-backend-20260512/backend/app/gateway/roundtable_inprocess_gateway.py
+++ b/offline-backend-20260512/backend/app/gateway/roundtable_inprocess_gateway.py
@@ -164,6 +164,7 @@ async def _run_agent_turn(
     skill_stop_names: list[str],
     thinking_enabled: bool,
     reasoning_effort: str,
+    subagent_enabled: bool = False,
 ) -> list[dict[str, Any]]:
     """进程内驱动 agent_name 在 thread_id 上跑一轮（发 message），阻塞到完成，返回最终 messages。
 
@@ -184,7 +185,8 @@ async def _run_agent_turn(
             "reasoning_effort": reasoning_effort,
             "thinking_enabled": thinking_enabled,
             "is_plan_mode": False,
-            "subagent_enabled": False,
+            # ultra 席位才置真 → 解锁 task 子代理工具（席位 policy 未禁 task）。leader 恒 False。
+            "subagent_enabled": subagent_enabled,
             "thread_id": thread_id,
             "memory_recall_disabled": True,
             # 与 multi_agent._run_payload 对齐:关 builtin 记忆注入,避免单例总控/席位
@@ -262,7 +264,13 @@ def _read_summary_md(app, thread_id: str, user_id: str) -> str | None:
 # it all at once」后发出超长 write_file，单次输出被 max_tokens 截断 → 工具调用从未执行、
 # run 以 success 收尾但 outputs 下没有 .html → step3.html 回退成开场白文字。
 # 自愈：检测产物缺失/截断后，同线程追加纠偏消息（要求 write_file 分段 + append 续写）重试。
-_ARTIFACT_RETRY_LIMIT = 2
+#
+# 取值依据（2026-06-15 修复「后台 HTML 看板只生成一半、概率渲染失败」）：模型单轮输出受
+# ``config.yaml`` 的 ``max_tokens``（deepseek-chat = 4096）硬截。一份完整看板 HTML 几十 KB
+# ≈ 上万 token，单轮写不完；旧值 2（最多 3 轮 ≈ 12k token）对长报告不够 → 落库半截 HTML →
+# 渲染失败。调到 5（最多 6 轮 append ≈ 24k token，足够绝大多数看板）显著降低截断概率。
+# 循环命中完整产物（``</html>`` / 非空 md / 合法 json）即 break，短报告不会白跑额外轮次。
+_ARTIFACT_RETRY_LIMIT = 5
 
 _REPORT_RETRY_MISSING = (
     f"检测到 /mnt/user-data/outputs/{_REPORT_FILE} 还没有生成——你上一轮的 write_file 没有执行成功"
@@ -332,11 +340,26 @@ def _build_seat_intro(agents: list[SeatRef]) -> str:
 class InProcessRoundtableGateway:
     """进程内真网关。executor 先调 ``init_threads`` 建好线程，再跑编排引擎。"""
 
-    def __init__(self, app, *, user_id: str, agents: list[SeatRef], model: str | None = None) -> None:
+    def __init__(
+        self,
+        app,
+        *,
+        user_id: str,
+        agents: list[SeatRef],
+        model: str | None = None,
+        seat_thinking_enabled: bool | None = None,
+        seat_reasoning_effort: str | None = None,
+        seat_subagent_enabled: bool = False,
+    ) -> None:
         self._app = app
         self._user_id = user_id
         self._agents = list(agents)
         self._model = model
+        # 席位执行模式（沿用 Step2 选择）：覆盖 seat policy 的 thinking/reasoning（None → policy 默认），
+        # subagent 仅 ultra 档为真。只作用 special 席位；leader 派活不受影响。
+        self._seat_thinking_enabled = seat_thinking_enabled
+        self._seat_reasoning_effort = seat_reasoning_effort
+        self._seat_subagent_enabled = seat_subagent_enabled
         self._seat_intro = ""
         self._first_leader = True
 
@@ -400,6 +423,7 @@ class InProcessRoundtableGateway:
 
     async def run_seat(self, *, thread_ids, agent_id, task, model) -> SeatTurn:
         policy = roundtable_run_policy("seat")
+        # 席位模式覆盖 policy 的 thinking/reasoning（None → policy 默认），与前端 _special_run 一致。
         messages = await _run_agent_turn(
             self._app,
             thread_id=thread_ids[agent_id],
@@ -408,8 +432,17 @@ class InProcessRoundtableGateway:
             model=model or self._model,
             excluded_tools=policy.excluded_tools,
             skill_stop_names=policy.skill_stop_names,
-            thinking_enabled=policy.thinking_enabled,
-            reasoning_effort=policy.reasoning_effort,
+            thinking_enabled=(
+                policy.thinking_enabled
+                if self._seat_thinking_enabled is None
+                else self._seat_thinking_enabled
+            ),
+            reasoning_effort=(
+                policy.reasoning_effort
+                if self._seat_reasoning_effort is None
+                else self._seat_reasoning_effort
+            ),
+            subagent_enabled=self._seat_subagent_enabled,
         )
         return SeatTurn(final_text=_latest_ai_text(messages), tool_calls=_tool_calls_since_last_human(messages))
 
diff --git a/offline-backend-20260512/backend/app/gateway/roundtable_job_executor.py b/offline-backend-20260512/backend/app/gateway/roundtable_job_executor.py
index c21d6aec..2fb39e8f 100644
--- a/offline-backend-20260512/backend/app/gateway/roundtable_job_executor.py
+++ b/offline-backend-20260512/backend/app/gateway/roundtable_job_executor.py
@@ -179,6 +179,12 @@ class StartParams:
     # 是否自动产出 Step3（方案总结报告 + 结果绘制）。默认沿用全局开关（现已恢复 True，
     # 见 ROUNDTABLE_BACKGROUND_REPORT_ENABLED）。测试可显式覆盖。
     enable_report: bool = ROUNDTABLE_BACKGROUND_REPORT_ENABLED
+    # 席位执行模式（快速/思考/专业问答/多智能体）派生参数，沿用 Step2 选择。仅 special 席位用：
+    # 覆盖 roundtable_run_policy("seat") 的 thinking/reasoning（None → policy 默认）；
+    # subagent 仅 ultra 档为真。透传给进程内网关(InProcessRoundtableGateway)的席位 run。
+    seat_thinking_enabled: bool | None = None
+    seat_reasoning_effort: str | None = None
+    seat_subagent_enabled: bool = False
 
 
 # 网关工厂：给定一次启动参数，返回一个 RoundtableGateway 实例。
diff --git a/offline-backend-20260512/backend/app/gateway/routers/multi_agent.py b/offline-backend-20260512/backend/app/gateway/routers/multi_agent.py
index a26f5742..ee12623f 100644
--- a/offline-backend-20260512/backend/app/gateway/routers/multi_agent.py
+++ b/offline-backend-20260512/backend/app/gateway/routers/multi_agent.py
@@ -637,6 +637,8 @@ def _run_payload(
     skill_stop_names: list[str],
     thinking_enabled: bool = True,
     reasoning_effort: str = "medium",
+    subagent_enabled: bool = False,
+    stream_custom: bool = False,
     multitask_strategy: str = "reject",
     extra_context: dict[str, Any] | None = None,
     files: list[dict[str, Any]] | None = None,
@@ -676,7 +678,9 @@ def _run_payload(
             "reasoning_effort": reasoning_effort,
             "thinking_enabled": thinking_enabled,
             "is_plan_mode": False,
-            "subagent_enabled": False,
+            # 默认 False；圆桌「多智能体问答(ultra)」档会让席位 _special_run 传 True，
+            # 给席位解锁 task 子代理工具（席位 policy 未禁 task，故可用）。
+            "subagent_enabled": subagent_enabled,
             "thread_id": thread_id,
             # 圆桌四个 router(intent / recommend / leader / special)都是临时任务,
             # 用户个人记忆与本次研讨无关。MemoryMiddleware 的 Hindsight recall 是
@@ -691,7 +695,12 @@ def _run_payload(
             # 圆桌是一次性研讨,本就不该吃历史会话记忆,这里统一关注入根治该污染。
             "memory_injection_enabled": False,
         },
-        "stream_mode": ["messages-tuple", "values"],
+        # ultra 席位会派子智能体(task 工具),其 task_started/running/completed 进度走
+        # **custom** 流(StreamWriter)。仅这种场景加 "custom",让前端任务卡片能显示子智能体
+        # 实时进度;其余角色不加,避免无谓帧。
+        "stream_mode": (
+            ["messages-tuple", "values", "custom"] if stream_custom else ["messages-tuple", "values"]
+        ),
         "stream_subgraphs": True,
         "stream_resumable": True,
         "assistant_id": "lead_agent",
@@ -1570,6 +1579,9 @@ async def _special_run(
     no_file: bool = False,
     peer_deliveries: list[dict[str, Any]] | None = None,
     files: list[dict[str, Any]] | None = None,
+    thinking_enabled: bool | None = None,
+    reasoning_effort: str | None = None,
+    subagent_enabled: bool = False,
 ) -> AsyncIterator[str | dict[str, Any]]:
     """子智能体（special）单轮：执行总控下发的 task 文本，交付 content。
 
@@ -1584,6 +1596,12 @@ async def _special_run(
     ``peer_deliveries``：前序席位的**完整交付** ``[{name, content}, ...]``。注入到
     run context 的 ``roundtable_peer_deliveries``,供席位用 ``read_peer_delivery`` 工具
     **按需**取全文(平时只看广播摘要,§12.9 传摘要按需传全文)。不进 prompt,不耗 token。
+
+    ``thinking_enabled`` / ``reasoning_effort`` / ``subagent_enabled``：圆桌第二步「席位
+    执行模式」（快速/思考/专业问答/多智能体问答）派生的覆盖值。**仅 special 席位**接收：
+    前两者为 None 时回退到 policy 默认（向后兼容，旧前端不传 → 维持 thinking + medium）；
+    ``subagent_enabled`` 仅 ultra 档为 True，给席位解锁 ``task`` 子代理工具。**总控
+    （leader）不受影响**，始终用其快速派活策略。
     """
     t0 = time.perf_counter()
     thread_id = agent_threads[agent_name]
@@ -1613,8 +1631,13 @@ async def _special_run(
         new_message=new_message,
         excluded_tools=seat_excluded,
         skill_stop_names=merge_skill_stop_names(policy.skill_stop_names, skill_stop_names),
-        thinking_enabled=policy.thinking_enabled,
-        reasoning_effort=policy.reasoning_effort,
+        # 席位执行模式覆盖 policy 的 thinking/reasoning（None → 回退 policy 默认）；
+        # subagent 仅 ultra 档置真。工具裁剪(excluded_tools)仍完全由 policy 决定。
+        thinking_enabled=(policy.thinking_enabled if thinking_enabled is None else thinking_enabled),
+        reasoning_effort=(policy.reasoning_effort if reasoning_effort is None else reasoning_effort),
+        subagent_enabled=subagent_enabled,
+        # ultra 席位才会派子智能体 → 才需要 custom 流回传子任务实时进度(task_running 等)。
+        stream_custom=subagent_enabled,
         # 干预/重新派活会把同一席位 thread 再跑一次;抢占掉没结束的旧 run,避免 409。
         multitask_strategy="rollback",
         # 前序席位完整交付 → run context,供 read_peer_delivery 工具按需取全文(§12.9)。
@@ -1756,6 +1779,15 @@ async def run_multi_agent_stream(request: Request, payload: dict = Body(...)) ->
     # 不写死任何默认模型;空串会让 lead_agent._resolve_model_name() 回退到
     # config.yaml 的 models[0],与 /api/ai-writing 等其它路由一致。
     model_name: str = str(payload.get("model") or "")
+    # 圆桌第二步「席位执行模式」(快速/思考/专业问答/多智能体问答)派生参数。**仅 special
+    # 席位消费**:覆盖 seat policy 的 thinking/reasoning。thinking_enabled/reasoning_effort
+    # 缺省(None/空)→ 后端按 policy 默认执行(向后兼容,旧前端不传维持 thinking + medium);
+    # subagent_enabled 仅 ultra 档为真,给席位解锁 task 子代理工具。leader 不消费这些字段。
+    _te_raw = payload.get("thinking_enabled")
+    seat_thinking_enabled: bool | None = bool(_te_raw) if _te_raw is not None else None
+    _re_raw = payload.get("reasoning_effort")
+    seat_reasoning_effort: str | None = _re_raw if isinstance(_re_raw, str) and _re_raw else None
+    seat_subagent_enabled: bool = bool(payload.get("subagent_enabled") or False)
 
     if not agent_name or agent_name not in agent_threads:
         raise HTTPException(
@@ -1803,6 +1835,9 @@ async def run_multi_agent_stream(request: Request, payload: dict = Body(...)) ->
                         no_file=no_file,
                         peer_deliveries=peer_deliveries,
                         files=files,
+                        thinking_enabled=seat_thinking_enabled,
+                        reasoning_effort=seat_reasoning_effort,
+                        subagent_enabled=seat_subagent_enabled,
                     )
                 async for chunk in gen:
                     if isinstance(chunk, dict):
diff --git a/offline-backend-20260512/backend/app/gateway/routers/roundtable_chains.py b/offline-backend-20260512/backend/app/gateway/routers/roundtable_chains.py
index 2f2db415..2f6a6363 100644
--- a/offline-backend-20260512/backend/app/gateway/routers/roundtable_chains.py
+++ b/offline-backend-20260512/backend/app/gateway/routers/roundtable_chains.py
@@ -57,6 +57,8 @@ class ChainResponse(BaseModel):
     stage_goals: list[str | None] | None = None
     # 写给总控的「编排提示」（dag-only）：告诉总控怎么派活、做什么。None/空 = 不注入。
     coordinator_prompt: str | None = None
+    # 默认席位执行模式（flash/thinking/pro/ultra）；None = 未设，选入研讨时回退默认(pro)。
+    seat_mode: str | None = None
     # 发布状态：True = 公共（所有账号可在「公共」筛选查看 / 选用）。
     is_public: bool = False
     # 是否当前用户所有（仅本人可编辑 / 删除 / 发布；公共筛选里别人的链条 is_mine=False）。
@@ -83,6 +85,7 @@ class ChainCreateRequest(BaseModel):
     stages: list[list[str]] | None = None
     stage_goals: list[str | None] | None = None
     coordinator_prompt: str | None = Field(default=None, max_length=4000)
+    seat_mode: str | None = Field(default=None, max_length=16)
 
 
 class ChainUpdateRequest(BaseModel):
@@ -93,6 +96,7 @@ class ChainUpdateRequest(BaseModel):
     stages: list[list[str]] | None = None
     stage_goals: list[str | None] | None = None
     coordinator_prompt: str | None = Field(default=None, max_length=4000)
+    seat_mode: str | None = Field(default=None, max_length=16)
     # 发布开关（True 发布为公共 / False 取消发布）；仅 owner 可改（update 本就 owner 限定）。
     is_public: bool | None = None
 
@@ -224,6 +228,18 @@ def _validate_stage_goals(
             )
 
 
+_VALID_SEAT_MODES = {"flash", "thinking", "pro", "ultra"}
+
+
+def _validate_seat_mode(seat_mode: str | None) -> None:
+    """席位执行模式只能是 flash/thinking/pro/ultra 或 None（未设）。"""
+    if seat_mode is not None and seat_mode not in _VALID_SEAT_MODES:
+        raise HTTPException(
+            status_code=400,
+            detail=f"非法席位执行模式: {seat_mode}（应为 {'/'.join(sorted(_VALID_SEAT_MODES))} 之一）",
+        )
+
+
 # ── routes ───────────────────────────────────────────────────────────────
 
 
@@ -249,6 +265,7 @@ async def create_chain(request: Request, body: ChainCreateRequest) -> ChainRespo
     _validate_seats(body.seats)
     _validate_stages(body.seats, body.stages)
     _validate_stage_goals(body.stages, body.stage_goals)
+    _validate_seat_mode(body.seat_mode)
     store = _get_store(request)
     user_id = _current_user_id(request)
     row = await store.create_chain(
@@ -261,6 +278,7 @@ async def create_chain(request: Request, body: ChainCreateRequest) -> ChainRespo
             "stages": body.stages,
             "stage_goals": body.stage_goals,
             "coordinator_prompt": body.coordinator_prompt,
+            "seat_mode": body.seat_mode,
         },
     )
     await _enrich_owner_names([row])
@@ -290,6 +308,7 @@ async def update_chain(request: Request, chain_id: str, body: ChainUpdateRequest
         _validate_stages(body.seats, body.stages)
     # Goals align to stages; reject goals sent without a layering, enforce length/size.
     _validate_stage_goals(body.stages, body.stage_goals)
+    _validate_seat_mode(body.seat_mode)
     store = _get_store(request)
     user_id = _current_user_id(request)
     # Distinguish "omitted" from "explicitly sent"; only forward provided fields.
diff --git a/offline-backend-20260512/backend/app/gateway/routers/roundtable_jobs.py b/offline-backend-20260512/backend/app/gateway/routers/roundtable_jobs.py
index 5c344e7c..e5b37c27 100644
--- a/offline-backend-20260512/backend/app/gateway/routers/roundtable_jobs.py
+++ b/offline-backend-20260512/backend/app/gateway/routers/roundtable_jobs.py
@@ -68,6 +68,12 @@ class StartJobRequest(BaseModel):
     # 预建好的 thread 映射（Phase 3b 真网关用；模拟网关忽略）。
     threadIds: dict[str, str] | None = None
     coordinatorName: str = "roundtable-coordinator"
+    # 席位执行模式（快速/思考/专业问答/多智能体）派生参数，由前端 toStartBody 下发。
+    # 仅 special 席位用：覆盖 roundtable_run_policy("seat") 的 thinking/reasoning；
+    # None/缺省 → 后端按 policy 默认(≈pro)，向后兼容旧前端。subagent 仅 ultra 档为真。
+    thinking_enabled: bool | None = None
+    reasoning_effort: str | None = None
+    subagent_enabled: bool = False
 
 
 class ResumeJobRequest(BaseModel):
@@ -219,6 +225,9 @@ async def start_job(request: Request, body: StartJobRequest) -> RoundtableJobRes
             draft_id=body.draftId,
             chain=body.chain,
             orchestration_plan=body.orchestrationPlan,
+            seat_thinking_enabled=body.thinking_enabled,
+            seat_reasoning_effort=body.reasoning_effort,
+            seat_subagent_enabled=body.subagent_enabled,
         )
     )
     return _to_response(row)
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/persistence/migrations/versions/20260615_01_chain_seat_mode.py b/offline-backend-20260512/backend/packages/harness/deerflow/persistence/migrations/versions/20260615_01_chain_seat_mode.py
new file mode 100644
index 00000000..f833ff75
--- /dev/null
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/persistence/migrations/versions/20260615_01_chain_seat_mode.py
@@ -0,0 +1,50 @@
+"""Add seat_mode column to roundtable_chains.
+
+Per-chain「默认席位执行模式」: an optional short string (flash/thinking/pro/ultra, see
+frontend ``lib/seat-mode.ts``) the user picks in the chain editor. When the chain is
+selected into a roundtable discussion the picker defaults to it (still overridable per
+run). NULL = unset → callers fall back to the default (pro). VARCHAR(16) on MySQL, plain
+String elsewhere — matching ``String(16)`` on ``RoundtableChainRow.seat_mode``.
+
+``create_all`` adds the column for brand-new DBs; the startup ``_ensure_orm_columns_sync``
+sweep adds it to already-existing tables (sqlite/mysql/postgres). This revision keeps the
+alembic graph consistent with the rest of the chain columns and documents the change.
+
+Revision ID: 20260615_01
+Revises: 20260612_01
+Create Date: 2026-06-15
+"""
+
+from __future__ import annotations
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy import inspect
+
+
+revision = "20260615_01"
+down_revision = "20260612_01"
+branch_labels = None
+depends_on = None
+
+
+def _has_table(table_name: str) -> bool:
+    inspector = inspect(op.get_bind())
+    return table_name in inspector.get_table_names()
+
+
+def _has_column(table_name: str, column_name: str) -> bool:
+    if not _has_table(table_name):
+        return False
+    inspector = inspect(op.get_bind())
+    return any(col["name"] == column_name for col in inspector.get_columns(table_name))
+
+
+def upgrade() -> None:
+    if _has_table("roundtable_chains") and not _has_column("roundtable_chains", "seat_mode"):
+        op.add_column("roundtable_chains", sa.Column("seat_mode", sa.String(length=16), nullable=True))
+
+
+def downgrade() -> None:
+    if _has_column("roundtable_chains", "seat_mode"):
+        op.drop_column("roundtable_chains", "seat_mode")
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/persistence/roundtable_chains/model.py b/offline-backend-20260512/backend/packages/harness/deerflow/persistence/roundtable_chains/model.py
index 36d54713..eb2d5baf 100644
--- a/offline-backend-20260512/backend/packages/harness/deerflow/persistence/roundtable_chains/model.py
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/persistence/roundtable_chains/model.py
@@ -49,6 +49,9 @@ class RoundtableChainRow(Base):
     # 怎么派活、做什么（整体编排策略 / 风格 / 约束）。**仅分层 DAG 模式生效**：注入到总控
     # 每阶段派活提示（``buildDagStageLeaderTask``）。NULL = 不注入。纯文本，存为 LONGTEXT。
     coordinator_prompt: Mapped[str | None] = mapped_column(PortableLongText, nullable=True)
+    # 可选「默认席位执行模式」：flash/thinking/pro/ultra（见前端 lib/seat-mode.ts）。选入研讨时
+    # 弹窗默认用它（仍可临时改）。NULL = 未设，调用方回退默认(pro)。短串，VARCHAR(16)。
+    seat_mode: Mapped[str | None] = mapped_column(String(16), nullable=True)
     # 发布开关：True = 公共（所有账号可在「公共」筛选里查看 / 选用）；False = 仅本人可见。
     # 编辑 / 删除始终只限 owner；发布只是放开「查看 + 选入研讨」。索引以支持「公共」列表筛选。
     is_public: Mapped[bool] = mapped_column(
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/persistence/roundtable_chains/sql.py b/offline-backend-20260512/backend/packages/harness/deerflow/persistence/roundtable_chains/sql.py
index 1b86b9cd..f3506589 100644
--- a/offline-backend-20260512/backend/packages/harness/deerflow/persistence/roundtable_chains/sql.py
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/persistence/roundtable_chains/sql.py
@@ -56,6 +56,8 @@ class RoundtableChainRepository:
             "stage_goals": _loads(getattr(row, "stage_goals", None)),
             # 总控编排提示（写给总控的派活指引，dag-only）；None/空 = 不注入。纯文本,不 json。
             "coordinator_prompt": getattr(row, "coordinator_prompt", None),
+            # 默认席位执行模式（flash/thinking/pro/ultra）；None = 未设。
+            "seat_mode": getattr(row, "seat_mode", None),
             # 发布状态 + 是否本人所有（前端据此显示「公共」徽标、决定能否编辑 / 发布）。
             "is_public": bool(getattr(row, "is_public", False)),
             "is_mine": (current_user_id is None) or (row.user_id == current_user_id),
@@ -110,6 +112,7 @@ class RoundtableChainRepository:
             stages=_dumps(data.get("stages")),
             stage_goals=_dumps(data.get("stage_goals")),
             coordinator_prompt=(data.get("coordinator_prompt") or None),
+            seat_mode=(data.get("seat_mode") or None),
             is_public=bool(data.get("is_public") or False),
             created_at=now,
             updated_at=now,
@@ -131,6 +134,7 @@ class RoundtableChainRepository:
         stages=_UNSET,
         stage_goals=_UNSET,
         coordinator_prompt=_UNSET,
+        seat_mode=_UNSET,
         is_public=_UNSET,
     ) -> dict[str, Any] | None:
         async with self._sf() as session:
@@ -149,6 +153,8 @@ class RoundtableChainRepository:
                 row.stage_goals = _dumps(stage_goals)
             if coordinator_prompt is not _UNSET:
                 row.coordinator_prompt = coordinator_prompt or None
+            if seat_mode is not _UNSET:
+                row.seat_mode = seat_mode or None
             if is_public is not _UNSET and is_public is not None:
                 row.is_public = bool(is_public)
             row.updated_at = datetime.now(UTC)
diff --git a/offline-backend-20260512/backend/scripts/sql/column_additions_mysql.sql b/offline-backend-20260512/backend/scripts/sql/column_additions_mysql.sql
index 330b6c41..ae02f476 100644
--- a/offline-backend-20260512/backend/scripts/sql/column_additions_mysql.sql
+++ b/offline-backend-20260512/backend/scripts/sql/column_additions_mysql.sql
@@ -82,6 +82,10 @@ CALL _add_column_if_missing('roundtable_drafts', 'task_id', 'VARCHAR(128) NULL D
 --                     ORM: RoundtableJobRow.orchestration_plan = PortableLongText(LONGTEXT), nullable。
 CALL _add_column_if_missing('roundtable_jobs', 'orchestration_plan', 'LONGTEXT NULL DEFAULT NULL');
 
+-- seat_mode: 业务链条「默认席位执行模式」(flash/thinking/pro/ultra),对应 migration 20260615_01。
+--            选入研讨时弹窗默认用它(仍可临时改)。ORM: RoundtableChainRow.seat_mode = String(16), nullable。
+CALL _add_column_if_missing('roundtable_chains', 'seat_mode', 'VARCHAR(16) NULL DEFAULT NULL');
+
 -- ---- 清理临时存储过程 --------------------------------------------------------
 DROP PROCEDURE IF EXISTS _add_column_if_missing;
 

From 88b1a330a6ca3e7c408402894cb9084ec848c08a Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E8=91=A3=E5=AE=8F=E5=87=AF?= <hkdong@kland.com.cn>
Date: Mon, 15 Jun 2026 18:31:50 +0800
Subject: [PATCH 2/9] =?UTF-8?q?fix(roundtable):=20=E4=BF=AE=E5=A4=8D?=
 =?UTF-8?q?=E5=BC=B9=E7=AA=97=E8=A2=AB=E9=A1=B6=E9=83=A8=E5=AF=BC=E8=88=AA?=
 =?UTF-8?q?=E6=A0=8F=E9=81=AE=E6=8C=A1=E7=9A=84=20z-index=20=E9=97=AE?=
 =?UTF-8?q?=E9=A2=98?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

将全局 Header z-[2101] 降为 z-[45]、圆桌页内部 header z-[2000] 降为 z-[40]，
使两者低于 Shadcn Dialog 遮罩层 z-50，弹窗打开时可正常覆盖导航栏。

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 .../src/roundtable-planning/pages/RoundtablePlanningPage.tsx    | 2 +-
 frontend-web/src/strategy-components/Header.tsx                 | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/frontend-web/src/roundtable-planning/pages/RoundtablePlanningPage.tsx b/frontend-web/src/roundtable-planning/pages/RoundtablePlanningPage.tsx
index 9163c070..c0b01eb6 100644
--- a/frontend-web/src/roundtable-planning/pages/RoundtablePlanningPage.tsx
+++ b/frontend-web/src/roundtable-planning/pages/RoundtablePlanningPage.tsx
@@ -1366,7 +1366,7 @@ export default function RoundtablePlanningPage() {
       )}
 
       {/* HEADER */}
-      <header className="relative z-[2000] flex shrink-0 items-center justify-between gap-4 border-b border-slate-100 bg-white px-6 py-3 shadow-[0_1px_3px_rgba(0,0,0,0.02)] dark:border-border dark:bg-card dark:shadow-none">
+      <header className="relative z-[40] flex shrink-0 items-center justify-between gap-4 border-b border-slate-100 bg-white px-6 py-3 shadow-[0_1px_3px_rgba(0,0,0,0.02)] dark:border-border dark:bg-card dark:shadow-none">
         <StepNavigation
           currentStep={currentStep}
           intentReady={step1.intentReady}
diff --git a/frontend-web/src/strategy-components/Header.tsx b/frontend-web/src/strategy-components/Header.tsx
index 0b2e12f7..5e507e62 100644
--- a/frontend-web/src/strategy-components/Header.tsx
+++ b/frontend-web/src/strategy-components/Header.tsx
@@ -294,7 +294,7 @@ const Header = ({ collapsed }: HeaderProps) => {
         open={notificationOpen}
         onOpenChange={setNotificationOpen}
       />
-      <header className="strategy-app-header theme-transition relative z-[2101] flex h-16 items-center justify-between px-4">
+      <header className="strategy-app-header theme-transition relative z-[45] flex h-16 items-center justify-between px-4">
         {showTopNav ? (
           <>
             {/* 左分区（flex-1）：菜单靠右贴近中间基座 */}

From 69c8c14c91ad5f56dbccd890ad2ef5b2c46e8e8e Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E8=91=A3=E5=AE=8F=E5=87=AF?= <hkdong@kland.com.cn>
Date: Mon, 15 Jun 2026 20:50:19 +0800
Subject: [PATCH 3/9] =?UTF-8?q?fix(nav):=20=E4=BF=AE=E5=A4=8D=E9=97=AE?=
 =?UTF-8?q?=E7=AD=94=E6=A2=B3=E7=90=86=E7=AD=89=E5=B8=A6=E5=AD=90=E8=B7=AF?=
 =?UTF-8?q?=E5=BE=84=E9=A1=B5=E9=9D=A2=E7=9A=84=E5=AF=BC=E8=88=AA=E9=AB=98?=
 =?UTF-8?q?=E4=BA=AE=E4=B8=A2=E5=A4=B1=E9=97=AE=E9=A2=98?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

KnowledgeBasePage 加载后自动跳转至 /knowledge/:noteId，原精确路径匹配
导致顶栏和侧边栏选中态丢失。将 hasActiveChild / isItemActive /
menuItemContainsPath 等 8 处判断统一改为前缀匹配
（pathname === path || pathname.startsWith(path + "/")），
覆盖浅色侧边栏、深色侧边栏、顶部导航栏三套组件。

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 .../src/components/page-sidebar-dark.tsx      |   4 +-
 frontend-web/src/components/page-sidebar.tsx  |   8 +-
 .../workspace/messages/message-list.tsx       |  36 +++-
 .../workspace/workspace-nav-menu.tsx          |  11 +-
 .../workspace/workspace-sidebar.tsx           |   2 +-
 .../src/core/page-layout/sidebar-menu.ts      | 164 ++++++++++--------
 .../src/pages/LightAppManagementPage.tsx      |   2 +-
 frontend-web/src/pages/StrategyRoutes.tsx     |   2 +
 frontend-web/src/pages/WorkspaceRoutes.tsx    |  10 +-
 .../src/strategy-components/Header.tsx        |  23 +--
 10 files changed, 161 insertions(+), 101 deletions(-)

diff --git a/frontend-web/src/components/page-sidebar-dark.tsx b/frontend-web/src/components/page-sidebar-dark.tsx
index 15b6f8c8..af274aff 100644
--- a/frontend-web/src/components/page-sidebar-dark.tsx
+++ b/frontend-web/src/components/page-sidebar-dark.tsx
@@ -54,7 +54,7 @@ export function PageSidebarDark({ collapsed, onToggleCollapsed }: PageSidebarDar
   };
 
   const hasActiveChild = (item: MenuItem): boolean => {
-    if (item.path && pathname === item.path) {
+    if (item.path && (pathname === item.path || pathname.startsWith(item.path + "/"))) {
       return true;
     }
 
@@ -153,7 +153,7 @@ export function PageSidebarDark({ collapsed, onToggleCollapsed }: PageSidebarDar
     inWujieEnv
       ? item.id === activeTopId
       : item.path
-        ? pathname === item.path
+        ? pathname === item.path || pathname.startsWith(item.path + "/")
         : hasActiveChild(item);
   const handleSidebarItemClick = (item: MenuItem) => {
     if (inWujieEnv) {
diff --git a/frontend-web/src/components/page-sidebar.tsx b/frontend-web/src/components/page-sidebar.tsx
index 42b556fe..4ad89043 100644
--- a/frontend-web/src/components/page-sidebar.tsx
+++ b/frontend-web/src/components/page-sidebar.tsx
@@ -42,7 +42,7 @@ type PageSidebarProps = {
 type SettingsSection = "account" | "appearance" | "memory" | "tools" | "about";
 
 const hasActiveChild = (item: MenuItem, pathname: string): boolean => {
-  if (item.path && pathname === item.path) {
+  if (item.path && (pathname === item.path || pathname.startsWith(item.path + "/"))) {
     return true;
   }
   return item.children?.some((child) => hasActiveChild(child, pathname)) ?? false;
@@ -116,7 +116,9 @@ function SecondaryNavSidebar() {
   const secondaryItems = activeTopItem?.children ?? [];
 
   const isItemActive = (item: MenuItem) =>
-    item.path ? pathname === item.path : hasActiveChild(item, pathname);
+    item.path
+      ? pathname === item.path || pathname.startsWith(item.path + "/")
+      : hasActiveChild(item, pathname);
 
   const handleItemClick = (item: MenuItem) => {
     const path = item.path ?? findFirstRoutePath(item.children);
@@ -210,7 +212,7 @@ function SecondaryNavSidebar() {
             {thirdItems.map((child) => {
               const ChildIcon = child.icon;
               const childActive = child.path
-                ? pathname === child.path
+                ? pathname === child.path || pathname.startsWith(child.path + "/")
                 : hasActiveChild(child, pathname);
               return (
                 <button
diff --git a/frontend-web/src/components/workspace/messages/message-list.tsx b/frontend-web/src/components/workspace/messages/message-list.tsx
index 9b7badbf..5eb2ddde 100644
--- a/frontend-web/src/components/workspace/messages/message-list.tsx
+++ b/frontend-web/src/components/workspace/messages/message-list.tsx
@@ -1,7 +1,8 @@
 import type { Message } from "@langchain/langgraph-sdk";
 import type { BaseStream } from "@langchain/langgraph-sdk/react";
 import { CheckIcon, SendIcon, XIcon } from "lucide-react";
-import { useCallback, useContext, useEffect, useMemo, useState } from "react";
+import { useCallback, useContext, useEffect, useMemo, useRef, useState } from "react";
+import { useStickToBottomContext } from "use-stick-to-bottom";
 
 import {
   Conversation,
@@ -52,6 +53,38 @@ import { SubtaskCard } from "./subtask-card";
 export const MESSAGE_LIST_DEFAULT_PADDING_BOTTOM = 160;
 export const MESSAGE_LIST_FOLLOWUPS_EXTRA_PADDING_BOTTOM = 80;
 
+/** 用户发送新问题时强制滚到底部（即使用户此前上滑看过历史）。 */
+function ScrollToBottomOnUserMessage({
+  threadId,
+  messages,
+}: {
+  threadId: string;
+  messages: Message[];
+}) {
+  const { scrollToBottom } = useStickToBottomContext();
+  const prevHumanCountRef = useRef<number | null>(null);
+
+  useEffect(() => {
+    prevHumanCountRef.current = null;
+  }, [threadId]);
+
+  useEffect(() => {
+    const humanCount = messages.reduce(
+      (count, message) => count + (message.type === "human" ? 1 : 0),
+      0,
+    );
+    const prev = prevHumanCountRef.current;
+    prevHumanCountRef.current = humanCount;
+    if (prev !== null && humanCount > prev) {
+      requestAnimationFrame(() => {
+        requestAnimationFrame(() => scrollToBottom());
+      });
+    }
+  }, [messages, scrollToBottom]);
+
+  return null;
+}
+
 export function MessageList({
   className,
   contentClassName,
@@ -266,6 +299,7 @@ export function MessageList({
     <Conversation
       className={cn("flex size-full flex-col justify-center", className)}
     >
+      <ScrollToBottomOnUserMessage threadId={threadId} messages={messages} />
       <ConversationContent
         className={cn(
           "mx-auto w-full max-w-(--container-width-lg) gap-8 pt-12",
diff --git a/frontend-web/src/components/workspace/workspace-nav-menu.tsx b/frontend-web/src/components/workspace/workspace-nav-menu.tsx
index 73cf8a72..fe14ee4f 100644
--- a/frontend-web/src/components/workspace/workspace-nav-menu.tsx
+++ b/frontend-web/src/components/workspace/workspace-nav-menu.tsx
@@ -165,7 +165,7 @@ export function WorkspaceNavMenu() {
                       Checkpoint 迁移
                     </DropdownMenuItem>
                   )}
-                  {isAdmin && (
+                  {isAdmin ? (
                     <DropdownMenuItem
                       onClick={() =>
                         navigate("/page/strategy/admin/user-leaderboard")
@@ -174,6 +174,15 @@ export function WorkspaceNavMenu() {
                       <TrophyIcon />
                       用户排行榜
                     </DropdownMenuItem>
+                  ) : (
+                    <DropdownMenuItem
+                      onClick={() =>
+                        navigate("/page/strategy/user-leaderboard")
+                      }
+                    >
+                      <TrophyIcon />
+                      用户排行榜
+                    </DropdownMenuItem>
                   )}
                   {isAdmin && (
                     <DropdownMenuItem
diff --git a/frontend-web/src/components/workspace/workspace-sidebar.tsx b/frontend-web/src/components/workspace/workspace-sidebar.tsx
index a721a9c9..ae5c6fd0 100644
--- a/frontend-web/src/components/workspace/workspace-sidebar.tsx
+++ b/frontend-web/src/components/workspace/workspace-sidebar.tsx
@@ -44,7 +44,7 @@ export function WorkspaceSidebar({
           variant="sidebar"
           collapsible="none"
           className={cn(
-            "h-full w-(--sidebar-width)",
+            "h-full w-(--sidebar-width) border-r border-border",
             useWhiteSidebar &&
               "[&_[data-slot=sidebar-inner]]:bg-white [&_[data-slot=sidebar-container]]:bg-white",
             className,
diff --git a/frontend-web/src/core/page-layout/sidebar-menu.ts b/frontend-web/src/core/page-layout/sidebar-menu.ts
index 7e3ccbf5..25f2a51f 100644
--- a/frontend-web/src/core/page-layout/sidebar-menu.ts
+++ b/frontend-web/src/core/page-layout/sidebar-menu.ts
@@ -43,6 +43,8 @@ export type MenuItem = {
   label: string;
   icon: ElementType;
   path?: string;
+  /** 非 admin 用户使用的路由（菜单对普通用户可见时生效，与简洁模式公开排行榜一致） */
+  userPath?: string;
   adminOnly?: boolean;
   children?: MenuItem[];
 };
@@ -81,18 +83,18 @@ export const sidebarMenuItems: MenuItem[] = [
         label: "XK中任务",
         icon: ClipboardList,
         children: [
-          {
-            id: "xk-task-jc",
-            label: "任务JC",
-            icon: ListTodo,
-            path: "/page/strategy/task-management/xk-task-jc",
-          },
-          {
-            id: "xk-task-dispatch",
-            label: "任务调度",
-            icon: ListTodo,
-            path: "/page/strategy/task-management/xk-task-dispatch",
-          },
+          // {
+          //   id: "xk-task-jc",
+          //   label: "任务JC",
+          //   icon: ListTodo,
+          //   path: "/page/strategy/task-management/xk-task-jc",
+          // },
+          // {
+          //   id: "xk-task-dispatch",
+          //   label: "任务调度",
+          //   icon: ListTodo,
+          //   path: "/page/strategy/task-management/xk-task-dispatch",
+          // },
         ],
       },
       {
@@ -125,24 +127,24 @@ export const sidebarMenuItems: MenuItem[] = [
         label: "PG中任务",
         icon: Gauge,
         children: [
-          {
-            id: "pg-task-pg",
-            label: "任务PG",
-            icon: ListTodo,
-            path: "/page/strategy/task-management/effect-task-pg",
-          },
-          {
-            id: "pg-task-xdpg",
-            label: "XDPG",
-            icon: ListTodo,
-            path: "/page/strategy/task-management/effect-action-pg",
-          },
-          {
-            id: "pg-task-info-pg",
-            label: "信息PG",
-            icon: ListTodo,
-            path: "/page/strategy/task-management/effect-info-pg",
-          },
+          // {
+          //   id: "pg-task-pg",
+          //   label: "任务PG",
+          //   icon: ListTodo,
+          //   path: "/page/strategy/task-management/effect-task-pg",
+          // },
+          // {
+          //   id: "pg-task-xdpg",
+          //   label: "XDPG",
+          //   icon: ListTodo,
+          //   path: "/page/strategy/task-management/effect-action-pg",
+          // },
+          // {
+          //   id: "pg-task-info-pg",
+          //   label: "信息PG",
+          //   icon: ListTodo,
+          //   path: "/page/strategy/task-management/effect-info-pg",
+          // },
         ],
       },
       {
@@ -150,24 +152,24 @@ export const sidebarMenuItems: MenuItem[] = [
         label: "TY中任务",
         icon: Radar,
         children: [
-          {
-            id: "ty-task-xdty",
-            label: "XDTY任务",
-            icon: ListTodo,
-            path: "/page/strategy/task-management/mnty-task-mechanism",
-          },
-          {
-            id: "ty-task-plan",
-            label: "方案TY任务",
-            icon: ListTodo,
-            path: "/page/strategy/task-management/mnty-task-plan",
-          },
-          {
-            id: "ty-task-xxty",
-            label: "XXTY任务",
-            icon: ListTodo,
-            path: "/page/strategy/task-management/mnty-task-zsty",
-          },
+          // {
+          //   id: "ty-task-xdty",
+          //   label: "XDTY任务",
+          //   icon: ListTodo,
+          //   path: "/page/strategy/task-management/mnty-task-mechanism",
+          // },
+          // {
+          //   id: "ty-task-plan",
+          //   label: "方案TY任务",
+          //   icon: ListTodo,
+          //   path: "/page/strategy/task-management/mnty-task-plan",
+          // },
+          // {
+          //   id: "ty-task-xxty",
+          //   label: "XXTY任务",
+          //   icon: ListTodo,
+          //   path: "/page/strategy/task-management/mnty-task-zsty",
+          // },
         ],
       },
     ],
@@ -197,6 +199,13 @@ export const sidebarMenuItems: MenuItem[] = [
     label: "综合管理",
     icon: Settings,
     children: [
+      {
+        id: "mgmt-user-leaderboard",
+        label: "用户排行榜",
+        icon: Trophy,
+        path: "/page/strategy/admin/user-leaderboard",
+        userPath: "/page/strategy/user-leaderboard",
+      },
       { id: "mgmt-agents", label: "智能体管理", icon: Bot, path: "/page/strategy/agents" },
       {
         id: "mgmt-business-chain-mapping",
@@ -232,13 +241,6 @@ export const sidebarMenuItems: MenuItem[] = [
         path: "/page/strategy/admin/users",
         adminOnly: true,
       },
-      {
-        id: "mgmt-user-leaderboard",
-        label: "用户排行榜",
-        icon: Trophy,
-        path: "/page/strategy/admin/user-leaderboard",
-        adminOnly: true,
-      },
       {
         id: "mgmt-admin-scheduled-tasks",
         label: "定时任务管理",
@@ -260,12 +262,12 @@ export const sidebarMenuItems: MenuItem[] = [
     label: "资源管理",
     icon: Database,
     children: [
-      {
-        id: "resource-network-xwt",
-        label: "网络xwt资源管理",
-        icon: Network,
-        path: "/page/strategy/resource-management/resource-network-actor",
-      },
+      // {
+      //   id: "resource-network-xwt",
+      //   label: "网络xwt资源管理",
+      //   icon: Network,
+      //   path: "/page/strategy/resource-management/resource-network-actor",
+      // },
       {
         id: "resource-knowledge-native",
         label: "问答梳理",
@@ -278,12 +280,12 @@ export const sidebarMenuItems: MenuItem[] = [
         icon: Library,
         path: "/page/strategy/resource-management/resource-knowledge-base",
       },
-      {
-        id: "resource-evaluation-index",
-        label: "作战资源管理",
-        icon: FileBox,
-        path: "/page/strategy/resource-management/resource-evaluation-index",
-      },
+      // {
+      //   id: "resource-evaluation-index",
+      //   label: "作战资源管理",
+      //   icon: FileBox,
+      //   path: "/page/strategy/resource-management/resource-evaluation-index",
+      // },
     ],
   },
 ];
@@ -298,9 +300,9 @@ export const defaultExpandedIds: ReadonlySet<string> = new Set([
   "comprehensive-management",
 ]);
 
-/** 该菜单项自身或任意后代的 path 是否等于当前路由 */
+/** 该菜单项自身或任意后代的 path 是否匹配当前路由（支持子路径，如 /knowledge/:id） */
 export function menuItemContainsPath(item: MenuItem, pathname: string): boolean {
-  if (item.path && item.path === pathname) {
+  if (item.path && (pathname === item.path || pathname.startsWith(item.path + "/"))) {
     return true;
   }
   return item.children?.some((child) => menuItemContainsPath(child, pathname)) ?? false;
@@ -308,7 +310,8 @@ export function menuItemContainsPath(item: MenuItem, pathname: string): boolean
 
 /**
  * 在「一级 → 二级 → 三级」菜单中，取当前路由所在的「激活二级项」的三级子项。
- * 浅色侧边栏据此在二级右侧展开三级列；无三级则返回空数组（侧边栏保持窄栏）。
+ * 浅色侧边栏据此在二级右侧展开三级列；深色顶栏据此渲染三级横条。
+ * 全量模式：三级仅一项时不返回（无选择必要，侧边栏保持窄栏 / 不渲染横条）。
  */
 export function getActiveThirdLevelItems(
   items: MenuItem[],
@@ -320,16 +323,23 @@ export function getActiveThirdLevelItems(
   const activeSecond = secondary.find(
     (item) => (item.children?.length ?? 0) > 0 && menuItemContainsPath(item, pathname),
   );
-  return activeSecond?.children ?? [];
+  const third = activeSecond?.children ?? [];
+  return third.length <= 1 ? [] : third;
 }
 
-/** 按 admin 角色过滤菜单：非 admin 不展示 ``adminOnly: true`` 的项及其子项 */
+/** 按 admin 角色过滤菜单：非 admin 不展示 ``adminOnly: true`` 的项；``userPath`` 替换 ``path``。 */
 export function filterMenuItemsByRole(items: MenuItem[], isAdmin: boolean): MenuItem[] {
   return items
     .filter((item) => !item.adminOnly || isAdmin)
-    .map((item) =>
-      item.children
-        ? { ...item, children: filterMenuItemsByRole(item.children, isAdmin) }
-        : item,
-    );
+    .map((item) => {
+      const children = item.children
+        ? filterMenuItemsByRole(item.children, isAdmin)
+        : undefined;
+      const base = children ? { ...item, children } : item;
+      if (!isAdmin && base.userPath) {
+        const { userPath: _userPath, ...rest } = base;
+        return { ...rest, path: base.userPath };
+      }
+      return base;
+    });
 }
diff --git a/frontend-web/src/pages/LightAppManagementPage.tsx b/frontend-web/src/pages/LightAppManagementPage.tsx
index 0180389d..b613735a 100644
--- a/frontend-web/src/pages/LightAppManagementPage.tsx
+++ b/frontend-web/src/pages/LightAppManagementPage.tsx
@@ -650,7 +650,7 @@ export default function LightAppManagementPage() {
             onClick={() => setDialogOpen(false)}
           />
 
-          <div className="lam-dialog relative z-[110] m-4 flex max-h-[90vh] w-full max-w-[580px] flex-col rounded-[5px] border border-[#E8E8E8] bg-white p-6 shadow-[0_8px_32px_rgba(34,34,34,0.16)]">
+          <div className="lam-dialog relative z-[110] m-4 flex max-h-[90vh] w-full max-w-[900px] flex-col rounded-[5px] border border-[#E8E8E8] bg-white p-6 shadow-[0_8px_32px_rgba(34,34,34,0.16)]">
             <div className="lam-dialog-header flex flex-shrink-0 items-center justify-between border-b border-[#E8E8E8] pb-4">
               <div className="flex items-center space-x-2.5">
                 <span className="lam-dialog-icon rounded-[5px] bg-rose-50 p-1.5 text-[#A8071A]">
diff --git a/frontend-web/src/pages/StrategyRoutes.tsx b/frontend-web/src/pages/StrategyRoutes.tsx
index 469d9a0d..af149f3a 100644
--- a/frontend-web/src/pages/StrategyRoutes.tsx
+++ b/frontend-web/src/pages/StrategyRoutes.tsx
@@ -25,6 +25,7 @@ import ChatPage from "./ChatPage";
 import ChatsPage from "./ChatsPage";
 import AdminUserLeaderboardPage from "./AdminUserLeaderboardPage";
 import AdminUserLeaderboardSettingsPage from "./AdminUserLeaderboardSettingsPage";
+import PublicLeaderboardPage from "./PublicLeaderboardPage";
 import LightAppIframeView from "./LightAppIframeView";
 import LightAppListPage from "./LightAppListPage";
 import LightAppManagementPage from "./LightAppManagementPage";
@@ -253,6 +254,7 @@ export default function StrategyRoutes() {
                 )
               }
             />
+            <Route path="user-leaderboard" element={<PublicLeaderboardPage />} />
             <Route
               path="admin/user-leaderboard"
               element={
diff --git a/frontend-web/src/pages/WorkspaceRoutes.tsx b/frontend-web/src/pages/WorkspaceRoutes.tsx
index 31e27de3..30ea1f35 100644
--- a/frontend-web/src/pages/WorkspaceRoutes.tsx
+++ b/frontend-web/src/pages/WorkspaceRoutes.tsx
@@ -70,6 +70,7 @@ export function WorkspaceLayout({
   workspaceSidebarHidden = false,
   headerHidden = false,
   compactWorkspaceSidebar = false,
+  defaultHistorySidebarOpen,
   wrapperOverflowHidden = false,
   insetOverflowHidden = false,
   pageSidebarDark = false,
@@ -80,6 +81,8 @@ export function WorkspaceLayout({
   workspaceSidebarHidden?: boolean;
   headerHidden?: boolean;
   compactWorkspaceSidebar?: boolean;
+  /** 全量模式下通用问答等页：聊天记录侧栏默认展开；未指定时紧凑模式仍默认收起。 */
+  defaultHistorySidebarOpen?: boolean;
   wrapperOverflowHidden?: boolean;
   insetOverflowHidden?: boolean;
   pageSidebarDark?: boolean;
@@ -110,8 +113,7 @@ export function WorkspaceLayout({
       }
       key={String(compactWorkspaceSidebar)}
       data-page-sidebar-collapsed={pageSidebarCollapsed}
-      // 紧凑模式下历史侧栏默认收起（聊天区更宽），由聊天页右上角图标钮展开
-      defaultOpen={!compactWorkspaceSidebar}
+      defaultOpen={defaultHistorySidebarOpen ?? !compactWorkspaceSidebar}
     >
       {!workspaceSidebarHidden && (
         <WorkspaceSidebar compactChrome={compactWorkspaceSidebar} />
@@ -153,7 +155,7 @@ function NewChatEntryRoute({
   }
 
   return (
-    <WorkspaceLayout {...layoutProps}>
+    <WorkspaceLayout {...layoutProps} defaultHistorySidebarOpen>
       <ChatRuntime>
         <ChatPage />
       </ChatRuntime>
@@ -201,7 +203,7 @@ export default function WorkspaceRoutes({
         <Route
           path="chats/:thread_id"
           element={
-            <WorkspaceLayout {...layoutProps}>
+            <WorkspaceLayout {...layoutProps} defaultHistorySidebarOpen>
               <ChatRuntime>
                 <ChatPage />
               </ChatRuntime>
diff --git a/frontend-web/src/strategy-components/Header.tsx b/frontend-web/src/strategy-components/Header.tsx
index 5e507e62..fae3d1a1 100644
--- a/frontend-web/src/strategy-components/Header.tsx
+++ b/frontend-web/src/strategy-components/Header.tsx
@@ -8,7 +8,10 @@ import { NotificationDialog } from "@/components/workspace/notification-dialog";
 import { getStoredAuth } from "@/core/auth";
 import { isWujieEnv } from "@/core/embed/wujie";
 import { useUnreadCount } from "@/core/notifications";
-import type { MenuItem } from "@/core/page-layout/sidebar-menu";
+import {
+  getActiveThirdLevelItems,
+  type MenuItem,
+} from "@/core/page-layout/sidebar-menu";
 import { useSidebarMenuItems } from "@/core/page-layout/use-sidebar-menu";
 import { cn } from "@/lib/utils";
 
@@ -145,7 +148,7 @@ const Header = ({ collapsed }: HeaderProps) => {
   const showLightTopNav = !isDark && !isWujieEnv();
 
   const hasActiveChild = (item: MenuItem): boolean => {
-    if (item.path && pathname === item.path) {
+    if (item.path && (pathname === item.path || pathname.startsWith(item.path + "/"))) {
       return true;
     }
 
@@ -170,9 +173,7 @@ const Header = ({ collapsed }: HeaderProps) => {
     return null;
   };
 
-  const activeTopItem = visibleMenuItems.find(hasActiveChild);
-  const activeSecondItem = activeTopItem?.children?.find(hasActiveChild);
-  const thirdSubNavItems = activeSecondItem?.children ?? [];
+  const thirdSubNavItems = getActiveThirdLevelItems(visibleMenuItems, pathname);
 
   const leftMenuItems = visibleMenuItems.filter((i) => LEFT_MENU_IDS.includes(i.id));
   const rightMenuItems = visibleMenuItems.filter((i) => RIGHT_MENU_IDS.includes(i.id));
@@ -185,7 +186,7 @@ const Header = ({ collapsed }: HeaderProps) => {
   };
 
   const renderTopItem = (item: MenuItem) => {
-    const isSelected = item.path ? pathname === item.path : hasActiveChild(item);
+    const isSelected = item.path ? (pathname === item.path || pathname.startsWith(item.path + "/")) : hasActiveChild(item);
     return (
       <button
         key={item.id}
@@ -199,7 +200,7 @@ const Header = ({ collapsed }: HeaderProps) => {
   };
 
   const renderLightTopItem = (item: MenuItem) => {
-    const isSelected = item.path ? pathname === item.path : hasActiveChild(item);
+    const isSelected = item.path ? (pathname === item.path || pathname.startsWith(item.path + "/")) : hasActiveChild(item);
     return (
       <button
         key={item.id}
@@ -310,7 +311,7 @@ const Header = ({ collapsed }: HeaderProps) => {
             </div>
 
             {/* 中：CM助手（落在背景图发光基座上，纯文字无图标） */}
-            <div className="header-center-title">CM助手</div>
+            <div className="header-center-title">CM业务中心</div>
 
             {/* 右分区（flex-1）：菜单靠左贴近中间基座，操作区靠最右 */}
             <div className="flex h-full min-w-0 flex-1 items-center justify-between">
@@ -336,7 +337,7 @@ const Header = ({ collapsed }: HeaderProps) => {
                   <FileText className="h-4 w-4 text-white" />
                 </div>
                 <span className="whitespace-nowrap text-base font-semibold text-white">
-                  CM助手
+                  CM业务中心
                 </span>
               </div>
               <nav className="light-topnav" aria-label="一级菜单">
@@ -363,7 +364,7 @@ const Header = ({ collapsed }: HeaderProps) => {
               </div>
               {!collapsed && (
                 <span className="text-sm font-semibold text-white">
-                  CM助手
+                  CM业务中心
                 </span>
               )}
             </div>
@@ -379,7 +380,7 @@ const Header = ({ collapsed }: HeaderProps) => {
         <nav className="third-submenu-bar" aria-label="三级菜单">
           <div className="third-submenu-list">
             {thirdSubNavItems.map((item) => {
-              const isSelected = item.path ? pathname === item.path : hasActiveChild(item);
+              const isSelected = item.path ? (pathname === item.path || pathname.startsWith(item.path + "/")) : hasActiveChild(item);
 
               return (
                 <button

From 9ed5420381eb9e0f83149ef048d9f650116c5d87 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E8=91=A3=E5=AE=8F=E5=87=AF?= <hkdong@kland.com.cn>
Date: Mon, 15 Jun 2026 22:33:18 +0800
Subject: [PATCH 4/9] =?UTF-8?q?feat(light-app):=20=E5=BA=94=E7=94=A8?=
 =?UTF-8?q?=E7=AE=A1=E7=90=86=E9=A1=B5=E5=8A=A0=E7=B1=BB=E5=9E=8B=E5=88=86?=
 =?UTF-8?q?=E7=B1=BB=E7=AD=9B=E9=80=89=20+=20=E4=BF=AE=E5=A4=8D=E6=B7=B1?=
 =?UTF-8?q?=E8=89=B2=E9=A1=B6=E6=A0=8F=E6=A0=87=E9=A2=98=E9=AB=98=E5=88=86?=
 =?UTF-8?q?=E5=B1=8F=E5=86=85=E8=BE=B9=E8=B7=9D?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- 应用管理工具栏新增「全部/新页跳转/iframe嵌套」分类切换，选中 iframe
  嵌套时出现挂载菜单下拉，可按 mountLocation 进一步筛选（深色/浅蓝主题适配）
- 深色顶栏 .header-center-title 内边距由固定 155px 改为 8.07vw，随视口
  等比缩放以匹配 background cover，修复 4K 等高分屏两侧菜单贴近中间的问题

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---
 .../src/pages/LightAppManagementPage.tsx      | 56 ++++++++++++++++++-
 .../src/pages/light-app-management.css        | 15 +++--
 frontend-web/src/styles/globals.css           |  5 +-
 3 files changed, 68 insertions(+), 8 deletions(-)

diff --git a/frontend-web/src/pages/LightAppManagementPage.tsx b/frontend-web/src/pages/LightAppManagementPage.tsx
index b613735a..2e28b2c0 100644
--- a/frontend-web/src/pages/LightAppManagementPage.tsx
+++ b/frontend-web/src/pages/LightAppManagementPage.tsx
@@ -133,6 +133,8 @@ export default function LightAppManagementPage() {
   const [form, setForm] = useState<FormState>(emptyForm);
   const [activeParamPopover, setActiveParamPopover] = useState<string | null>(null);
   const [viewMode, setViewMode] = useState<"table" | "cards">("table");
+  const [categoryFilter, setCategoryFilter] = useState<"all" | LightAppType>("all");
+  const [mountFilter, setMountFilter] = useState<string>("");
 
   const saving = actions.create.isPending || actions.update.isPending;
 
@@ -220,7 +222,22 @@ export default function LightAppManagementPage() {
     toast.success("已复制地址");
   };
 
-  const tableApps = useMemo(() => apps, [apps]);
+  const selectCategory = (cat: "all" | LightAppType) => {
+    setCategoryFilter(cat);
+    if (cat !== "iframe") setMountFilter("");
+  };
+
+  const tableApps = useMemo(
+    () =>
+      apps.filter((app) => {
+        if (categoryFilter !== "all" && app.appType !== categoryFilter) return false;
+        if (categoryFilter === "iframe" && mountFilter && app.mountLocation !== mountFilter) {
+          return false;
+        }
+        return true;
+      }),
+    [apps, categoryFilter, mountFilter],
+  );
 
   return (
     <div className="light-app-management-page h-full overflow-y-auto bg-neutral-50 dark:bg-[#0C1836] dark:text-[#C5CFE0]">
@@ -252,7 +269,7 @@ export default function LightAppManagementPage() {
           </button>
         </div>
 
-        <div className="lam-toolbar flex flex-col items-center gap-3.5 rounded-[5px] bg-white p-3 shadow-[0_1px_3px_rgba(0,0,0,0.01)] md:flex-row">
+        <div className="lam-toolbar flex flex-col items-center gap-3.5 rounded-[5px] bg-white p-3 shadow-[0_1px_3px_rgba(0,0,0,0.01)] md:flex-row md:flex-wrap">
           <div className="relative w-full flex-1">
             <Search className="absolute left-3.5 top-1/2 h-4 w-4 -translate-y-1/2 text-neutral-400" />
             <input
@@ -272,6 +289,41 @@ export default function LightAppManagementPage() {
             )}
           </div>
 
+          <div className="lam-cat-toggle flex shrink-0 items-center space-x-0.5 self-end rounded-[5px] border border-neutral-200/50 bg-neutral-100/60 p-0.5 md:self-auto">
+            {(
+              [
+                { value: "all", label: "全部" },
+                { value: "window_open", label: "新页跳转" },
+                { value: "iframe", label: "iframe 嵌套" },
+              ] as const
+            ).map((c) => (
+              <button
+                key={c.value}
+                onClick={() => selectCategory(c.value)}
+                className={`lam-cat-toggle-btn cursor-pointer rounded-[5px] px-4 py-1.5 text-xs font-bold transition-all ${
+                  categoryFilter === c.value
+                    ? "is-active bg-white text-[#F5222D] shadow-sm"
+                    : "text-neutral-500 hover:text-neutral-700"
+                }`}
+              >
+                {c.label}
+              </button>
+            ))}
+          </div>
+
+          {categoryFilter === "iframe" && (
+            <div className="lam-mount-filter w-full shrink-0 animate-in fade-in self-end md:w-[180px] md:self-auto">
+              <Select
+                className="w-full"
+                value={mountFilter || undefined}
+                placeholder="按挂载菜单筛选"
+                clearable
+                options={MOUNT_OPTIONS}
+                onChange={(v) => setMountFilter((v as string) ?? "")}
+              />
+            </div>
+          )}
+
           <div className="lam-view-toggle flex shrink-0 items-center space-x-0.5 self-end rounded-[5px] border border-neutral-200/50 bg-neutral-100/60 p-0.5 md:self-auto">
             <button
               onClick={() => setViewMode("table")}
diff --git a/frontend-web/src/pages/light-app-management.css b/frontend-web/src/pages/light-app-management.css
index 18a97f05..8f0fa408 100644
--- a/frontend-web/src/pages/light-app-management.css
+++ b/frontend-web/src/pages/light-app-management.css
@@ -65,22 +65,26 @@
   outline: none;
 }
 
-.dark .light-app-management-page .lam-view-toggle {
+.dark .light-app-management-page .lam-view-toggle,
+.dark .light-app-management-page .lam-cat-toggle {
   background: var(--color-bg-page);
   border-color: var(--color-border);
 }
 
-.dark .light-app-management-page .lam-view-toggle-btn.is-active {
+.dark .light-app-management-page .lam-view-toggle-btn.is-active,
+.dark .light-app-management-page .lam-cat-toggle-btn.is-active {
   background: var(--color-bg-card);
   color: var(--color-link);
   box-shadow: var(--shadow-1);
 }
 
-.dark .light-app-management-page .lam-view-toggle-btn:not(.is-active) {
+.dark .light-app-management-page .lam-view-toggle-btn:not(.is-active),
+.dark .light-app-management-page .lam-cat-toggle-btn:not(.is-active) {
   color: var(--color-text-secondary);
 }
 
-.dark .light-app-management-page .lam-view-toggle-btn:not(.is-active):hover {
+.dark .light-app-management-page .lam-view-toggle-btn:not(.is-active):hover,
+.dark .light-app-management-page .lam-cat-toggle-btn:not(.is-active):hover {
   color: var(--color-text);
 }
 
@@ -592,7 +596,8 @@
 }
 
 /* ── 视图切换激活态 ──────────────────────────────────────────────────── */
-[data-theme="lightblue"] .light-app-management-page .lam-view-toggle-btn.is-active {
+[data-theme="lightblue"] .light-app-management-page .lam-view-toggle-btn.is-active,
+[data-theme="lightblue"] .light-app-management-page .lam-cat-toggle-btn.is-active {
   color: #0062D9;
 }
 
diff --git a/frontend-web/src/styles/globals.css b/frontend-web/src/styles/globals.css
index 3494b65c..a032c9c7 100644
--- a/frontend-web/src/styles/globals.css
+++ b/frontend-web/src/styles/globals.css
@@ -950,7 +950,10 @@ html.dark .strategy-app-header {
   display: flex;
   align-items: center;
   justify-content: center;
-  padding: 0 155px;
+  /* 左右内边距随视口宽度等比缩放，与背景图基座的 background-size:cover 同步，
+     使两侧菜单在 1080p / 2K / 4K 下都贴齐发光基座边缘。
+     基准：1920 宽时 155px → 155/1920×100 ≈ 8.07vw。 */
+  padding: 0 8.07vw;
   /* 文字略微上移，贴合基座视觉中心 */
   transform: translateY(-6px);
   font-family: "Source Han Sans SC", "Noto Sans SC", "Noto Sans S Chinese",

From 6b49741e690cda3df6b749faa6e7bb692bfdd6e4 Mon Sep 17 00:00:00 2001
From: fengxiangdebaozi <10845145+fengxiangdebaozi@user.noreply.gitee.com>
Date: Tue, 16 Jun 2026 10:05:54 +0800
Subject: [PATCH 5/9] =?UTF-8?q?=E4=BB=A3=E7=A0=81=E4=BF=AE=E6=94=B9?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 frontend-web/docs/multi-agent-api.md          | 265 ------------------
 ...agent-background-run-遗留问题与下次继续.md | 125 ---------
 frontend-web/docs/xvqiu-0616.md               | 161 +++++++++++
 .../workspace/frequent-agents-list.tsx        | 127 +++++++++
 .../components/workspace/recent-chat-list.tsx |  33 ++-
 .../workspace/workspace-sidebar.tsx           |  13 +-
 offline-backend-20260512/backend/config.yaml  |   2 +-
 7 files changed, 332 insertions(+), 394 deletions(-)
 delete mode 100644 frontend-web/docs/multi-agent-api.md
 delete mode 100644 frontend-web/docs/multi-agent-background-run-遗留问题与下次继续.md
 create mode 100644 frontend-web/docs/xvqiu-0616.md
 create mode 100644 frontend-web/src/components/workspace/frequent-agents-list.tsx

diff --git a/frontend-web/docs/multi-agent-api.md b/frontend-web/docs/multi-agent-api.md
deleted file mode 100644
index a493801e..00000000
--- a/frontend-web/docs/multi-agent-api.md
+++ /dev/null
@@ -1,265 +0,0 @@
-# 多智能体协调 API 接口文档
-
-后端路由：`offline-backend-20260512/backend/app/gateway/routers/multi_agent.py`
-路由前缀：`/api/multi-agent`
-
-## 概览
-
-提供一个"总控 + 多个子智能体"的协调编排接口。
-
-- **总控（leader）** 智能体：通过 `agent_orchestration` 技能给子智能体派活
-- **子智能体（special）**：执行被派到的具体任务，完成后把交付内容广播给其他线程
-
-调用方按下面顺序使用：
-
-1. 调 `POST /api/multi-agent/init` 一次性创建协调智能体 + 给每个子智能体建 `thread`，拿到 `thread_ids` 映射
-2. 客户端把 `thread_ids` 保管好，后续每次调 `POST /api/multi-agent/run/stream` 时回传
-3. 先调用 `agent_type="leader"` 拿派活清单，再针对清单里每个子智能体调用 `agent_type="special"`
-
-
-## 1. 初始化多智能体集群
-
-### `POST /api/multi-agent/init`
-
-创建协调智能体并为列表中的每个子智能体建立独立的 `thread`，同时把"职责说明"广播到所有子智能体线程作为初始上下文。
-
-#### 请求体
-
-```json
-{
-  "agent_names": ["taiwan-intelligence-gathering-agent", "us-intelligence-gathering-agent"],
-  "main_agent_name": "main-controller",
-  "model": "deepseek-chat"
-}
-```
-
-| 字段 | 类型 | 必填 | 说明 |
-|---|---|---|---|
-| `agent_names` | `string[]` | 是 | 已有子智能体的 id 列表（kebab-case，区分大小写不敏感，自动转小写） |
-| `main_agent_name` | `string` | 否 | 要创建的协调智能体名，默认 `main-agent`。重名时会复用已存在的同名智能体 |
-| `model` | `string` | 否 | 协调智能体使用的模型名。不传则用后端默认模型 |
-
-#### 响应（200）
-
-```json
-{
-  "status": "success",
-  "main_agent_name": "main-controller",
-  "thread_ids": {
-    "main-controller": "thread-uuid-xxx",
-    "taiwan-intelligence-gathering-agent": "thread-uuid-yyy",
-    "us-intelligence-gathering-agent": "thread-uuid-zzz"
-  }
-}
-```
-
-`thread_ids` 是一个 `{agent_name: thread_id}` 字典。**前端必须把这个对象完整保管下来**，后续调用 stream 接口需要原样回传。
-
-#### 错误
-
-| 状态码 | 触发条件 |
-|---|---|
-| `400` | `agent_names` 为空 |
-| `401` | 未携带 `Authorization` 头 |
-| `500` | 协调智能体创建失败且不是"已存在"错误 |
-
----
-
-## 2. 流式调度
-
-### `POST /api/multi-agent/run/stream`
-
-通过 SSE（Server-Sent Events）流式驱动一个 leader 或 special 智能体跑一轮对话。
-
-#### 请求体
-
-```json
-{
-  "agent_type": "leader",
-  "d_agent_thread_id": {
-    "main-controller": "thread-uuid-xxx",
-    "taiwan-intelligence-gathering-agent": "thread-uuid-yyy"
-  },
-  "agent_name": "main-controller",
-  "new_message": "请分析特朗普的支持率",
-  "skill_stop_names": [],
-  "model": "deepseek-chat"
-}
-```
-
-| 字段 | 类型 | 必填 | 说明 |
-|---|---|---|---|
-| `agent_type` | `"leader"` \| `"special"` | 否 | 默认 `leader`。`leader` 调用协调智能体，`special` 调用子智能体 |
-| `d_agent_thread_id` | `object` | 是 | 由 `/init` 接口返回的 `thread_ids` 映射，原样回传 |
-| `agent_name` | `string` | 是 | 本轮要驱动的智能体名（必须存在于 `d_agent_thread_id` 中） |
-| `new_message` | `string` | 是 | 这轮发给智能体的用户消息 |
-| `skill_stop_names` | `string[]` | 否 | 触发后中断执行的技能名列表（默认空） |
-| `model` | `string` | 否 | 本轮 LLM 模型名，默认 `deepseek-chat` |
-
-#### 响应
-
-`Content-Type: text/event-stream` 的 SSE 长连接，按到达顺序产出多种数据：
-
-##### a. 透传上游事件
-
-原样转发底层 `POST /api/threads/{thread_id}/runs/stream` 的事件帧（LangGraph 协议），包含 `event:`、`data:` 等行。前端按 LangGraph SDK 既有方式渲染即可。
-
-##### b. 最后一帧：调度状态
-
-流结束前会额外产出一帧 `data:` 行，结构因 `agent_type` 而异：
-
-**`agent_type == "leader"`：**
-
-```
-data: {"status": [["us-intelligence-gathering-agent", "收集美国民众对特朗普的看法"], ["taiwan-intelligence-gathering-agent", "..."]]}
-```
-
-`status` 是一个 `[agent_name, task_text]` 二元组数组，代表协调智能体本轮派出去的任务清单。**前端拿到后应当对每个元素再调用一次本接口（`agent_type="special"`）来执行子智能体**。如果数组为空，说明 leader 觉得已有信息足够，不再派活，可以直接展示 leader 的最终文本回答。
-
-**`agent_type == "special"`：**
-
-```
-data: {"status": "子智能体xxx完成yyy工作"}
-```
-
-代表子智能体完成本轮工作并已把交付内容广播给其他线程。
-
-#### 副作用（重要）
-
-- **leader 调用**：自动把"总控智能体收到用户消息: …"广播到所有非自己的线程；派活后再把"总控智能体给 X 派活, 内容为: …"广播到所有非派活双方的线程
-- **special 调用**：完成后把"子智能体 X 完成 Y 工作, 交付内容为: …"广播到所有非自己的线程
-
-这意味着每个线程的对话历史里会逐步累积出全局上下文，无需前端再做合并。
-
-#### 错误
-
-| 状态码 | 触发条件 |
-|---|---|
-| `400` | `agent_name` 缺失或未在 `d_agent_thread_id` 中 |
-| `401` | 未携带 `Authorization` 头 |
-| `502` | 上游 `runs/stream` 返回非 200 / SSE 解析失败 |
-
----
-
-## 3. 前端调用流程示例
-
-```ts
-// 1. 拿 token
-const loginResp = await fetch("/api/v1/auth/login/username", {
-  method: "POST",
-  headers: { "Content-Type": "application/json" },
-  body: JSON.stringify({ username: "yourname" }),
-});
-const { access_token } = await loginResp.json();
-const authHeaders = {
-  Authorization: `Bearer ${access_token}`,
-  "Content-Type": "application/json",
-};
-
-// 2. 初始化
-const initResp = await fetch("/api/multi-agent/init", {
-  method: "POST",
-  headers: authHeaders,
-  body: JSON.stringify({
-    agent_names: [
-      "taiwan-intelligence-gathering-agent",
-      "us-intelligence-gathering-agent",
-    ],
-    main_agent_name: "main-controller",
-    model: "deepseek-chat",
-  }),
-});
-const { thread_ids } = await initResp.json();
-
-// 3. 流式调用 leader
-async function streamLeader(message: string, model = "deepseek-chat") {
-  const resp = await fetch("/api/multi-agent/run/stream", {
-    method: "POST",
-    headers: authHeaders,
-    body: JSON.stringify({
-      agent_type: "leader",
-      d_agent_thread_id: thread_ids,
-      agent_name: "main-controller",
-      new_message: message,
-      skill_stop_names: [],
-      model,
-    }),
-  });
-  const reader = resp.body!.getReader();
-  const decoder = new TextDecoder();
-  let buf = "";
-  let dispatched: [string, string][] = [];
-
-  while (true) {
-    const { done, value } = await reader.read();
-    if (done) break;
-    buf += decoder.decode(value, { stream: true });
-    const lines = buf.split("\n");
-    buf = lines.pop() ?? "";
-    for (const line of lines) {
-      if (!line.startsWith("data:")) continue;
-      try {
-        const payload = JSON.parse(line.slice(5).trim());
-        if (Array.isArray(payload.status)) {
-          dispatched = payload.status;
-        }
-        // 其余事件按 LangGraph SDK 协议自行渲染
-      } catch {
-        /* 非 JSON 帧，可忽略或交给现有渲染器 */
-      }
-    }
-  }
-  return dispatched;
-}
-
-// 4. 对派活清单逐个调用 special
-async function streamSpecial(agentName: string, task: string, model = "deepseek-chat") {
-  return fetch("/api/multi-agent/run/stream", {
-    method: "POST",
-    headers: authHeaders,
-    body: JSON.stringify({
-      agent_type: "special",
-      d_agent_thread_id: thread_ids,
-      agent_name: agentName,
-      new_message: task,
-      skill_stop_names: [],
-      model,
-    }),
-  });
-  // 同样按 SSE 解析
-}
-
-// 5. 编排：先 leader 派活 → 依次执行 special → 再让 leader 汇总
-const dispatched = await streamLeader("请分析特朗普的支持率");
-for (const [agentName, task] of dispatched) {
-  await streamSpecial(agentName, task);
-}
-const next = await streamLeader("根据上面收集到的信息出最终结论");
-// 当 leader 返回 status: [] 时，意味着汇报完毕，可结束循环
-```
-
----
-
-## 4. 注意事项
-
-- **`model` 字段的取值**：必须是后端 `config.yaml` 中 `models[]` 已配置的模型名（如 `deepseek-chat`、`glm-4`、`gpt-4o` 等）。传错会导致上游 502
-- **`thread_ids` 必须每次回传**：后端不存储这个映射，是无状态接口
-- **协调智能体必须装有 `agent_orchestration` 技能**：`/init` 创建时会自动指定 `skills=["agent_orchestration"]`。后端 `skills/` 目录里必须存在该技能定义，否则 leader 无法派活
-- **同名协调智能体冲突**：`main_agent_name` 重名时后端不会报错（"already exists" 走兼容分支），会复用既有智能体——所以同一个用户多次 init 用同一个名字是安全的，但 SOUL 不会被覆盖
-- **SSE 断连**：本接口底层 `on_disconnect: "continue"`，断连后上游 run 会继续跑到结束（不会取消）
-- **并发广播**：广播消息使用 `asyncio.gather` 并发更新所有线程 state，性能上不会随子智能体数线性变慢
-
----
-
-## 5. 相关接口（底层依赖）
-
-本接口内部转发以下 Gateway 原生接口，了解细节可查：
-
-| 接口 | 路由 | 用途 |
-|---|---|---|
-| 创建 thread | `POST /api/threads` | 为每个智能体建立独立会话 |
-| 读 thread state | `GET /api/threads/{id}/state` | 广播前取当前消息列表 |
-| 写 thread state | `POST /api/threads/{id}/state` | 把广播消息追加到 messages |
-| 取 agent 详情 | `GET /api/agents/{id}` | init 时读取子智能体的 name/description |
-| 建 agent | `POST /api/agents` | init 时创建协调智能体 |
-| 流式 run | `POST /api/threads/{id}/runs/stream` | leader/special 的实际执行 |
diff --git a/frontend-web/docs/multi-agent-background-run-遗留问题与下次继续.md b/frontend-web/docs/multi-agent-background-run-遗留问题与下次继续.md
deleted file mode 100644
index 4bdb11f6..00000000
--- a/frontend-web/docs/multi-agent-background-run-遗留问题与下次继续.md
+++ /dev/null
@@ -1,125 +0,0 @@
-# 多智能体「后台挂起」—— 当前状态、遗留问题与下次继续指南
-
-> 配套设计文档：`multi-agent-background-run-dev.md`（分 Phase 0-9 的总方案）。
-> 本文记录**截至本次会话的实际进展 + 还没解决的遗留问题**，方便下次直接续接。
-
----
-
-## 一、整体状态：跑通到哪了
-
-「后台挂起」= Step2 把研讨交给后端**进程内驱动真模型**（leader→各席位→共识→报告），
-survive 刷新/关页，对话与报告作为一个 **run 写进草稿**（和手动跑同一存取），可在历史回看。
-
-**真网关已在本地验证可跑通**（2026-06-04 日志确认：roundtable-coordinator + 各席位 +
-roundtable-report 真调 deepseek、真搜索、跑到 report run success）。但还有几个**内容/产物层面的遗留问题**（见第三节）。
-
-### 已完成（可用）
-- ✅ 后端持久化：`roundtable_jobs` 表（迁移 `20260604_01` / `20260605_01` dialogues / `20260605_02` step3）。
-- ✅ 编排引擎 `packages/harness/deerflow/agents/roundtable_orchestrator/`（recommend + chain + 报告，纯逻辑可单测）。
-- ✅ 控制面 `app/gateway/routers/roundtable_jobs.py`：start/get/by-draft/stream(SSE)/resume/cancel/admin-cleanup。
-- ✅ 执行器 `app/gateway/roundtable_job_executor.py`：后台任务 + `set_current_user` + 生命周期 + **每轮把对话作为 `job-<id>` run 写进 `draft.step2.runs[]`（同手动存取）** + done 写 step3。
-- ✅ 真网关 `app/gateway/roundtable_inprocess_gateway.py`：`SimpleNamespace(app)` + `start_run` 进程内驱动 agent；复用 `multi_agent._extract_dispatch` 解析派活；checkpointer 读最终 messages。
-- ✅ 模拟网关 `roundtable_orchestrator/simulated_gateway.py`（不调真模型，前端联调用）。
-- ✅ 前端：`api/roundtable-jobs.ts`、`lib/job-types.ts`、Step2「后台挂起」按钮、历史下拉 loading、`DispatchChainModal`/`DispatchChainJobModal`（流程图 + 对话面板 markdown + 加宽 + 澄清作答 + 取消 + 查看）。
-- ✅ 上下文「广播」（inline）：席位/总控 prompt 带上「任务背景 + 前序各席位交付 + 别提问」约束（修「席位拿不到关键信息、反问用户、总控瞎编」）。
-- ✅ 后端测试：`tests/test_roundtable_jobs.py` / `test_roundtable_orchestrator.py` / `test_roundtable_job_executor.py` / `test_roundtable_jobs_api.py` 全过。
-
----
-
-## 二、怎么启动 / 验证（每次开工先做）
-
-1. **开真网关**（默认是模拟网关）：已写进 `offline-backend-20260512/.env`：
-   ```
-   ROUNDTABLE_JOB_GATEWAY=inprocess
-   ```
-   启动日志应出现 `[roundtable-jobs] gateway = inprocess (real models)`。
-
-2. **必须用 `make gateway`（不带 --reload）跑后端！**
-   `make dev` 的 `--reload` 监视 `backend/` 目录，agent 运行时往 `.deer-flow/` 写文件
-   （memory.json / checkpoint / 沙箱输出）会触发热重载，**把长时后台任务杀掉**。
-   （改进项：给 uvicorn 加 `--reload-exclude '.deer-flow/*'`，或始终用 `make gateway`。）
-
-3. **直接查库验证**（最有效，PowerShell 控制台是 GBK，导出到文件看中文）：
-   ```bash
-   cd offline-backend-20260512/backend
-   # 查最新作业的 dialogues / step3
-   PYTHONPATH=. uv run python -c "import sqlite3,json; con=sqlite3.connect('.deer-flow/data/deerflow.db'); ..."
-   ```
-   - 作业表：`.deer-flow/data/deerflow.db` → 表 `roundtable_jobs`（`dialogues`/`step3` 列）。
-   - 草稿表：`roundtable_drafts` → `step2` 列里应有 `runs[ {id:"job-<jobId>", jobId, jobStatus, step2RoundtableDialogues:[...], step3:{...}} ]`。
-
-4. **前端验证**：后台挂起 → 等跑完 → 从历史加载该草稿 → Step2 应显示后台研讨完整对话（走现有 `normalizeStep2Runs`/`hydrateStep2Run`）。
-
----
-
-## 三、遗留问题（下次重点，已按优先级排）
-
-### P0 —— 报告 agent 不写 HTML 文件（结果绘制拿不到产物）✅ 已于 2026-06-12 修复并恢复
-- **根因（2026-06-12 定位）**：`dc8fca6`（2026-06-11 智能体可见性对齐）给 `services._enforce_agent_access`
-  加 admin 判断时**裸访问 `request.state`** —— 后台进程内虚拟 request（`SimpleNamespace(app=app)`）
-  没有 `state` 属性 → `AttributeError` 杀死整个编排（前端真 HTTP 请求有 auth 中间件写的 `state.user`
-  所以手动路径一直正常）。修复：getattr 全链防御 + 虚拟 request 补空 `state`（含 scheduler 同款），
-  回归测试 `tests/test_roundtable_inprocess_request.py`。
-- **已恢复**：`ROUNDTABLE_BACKGROUND_REPORT_ENABLED = True`。后台收口后自动跑 Step3 双产物：
-  ① `roundtable-summary` 写「方案总结报告.md」（新增 `gateway.run_summary` + `engine._run_step3`，
-  prompt 对齐 `useStep3Summary.buildSummaryPrompt`）→ ② `roundtable-report` 画「方案总览.html」。
-  两段故障隔离（一段失败保留另一段产物）；step3 形状对齐前端 `DraftStep3Snapshot`
-  （`summaryReport` 嵌套），写草稿 `draft.step3` 与手动调用同款。测试
-  `tests/test_roundtable_step3_pipeline.py`。
-- **待真机验证**：早期（06-05，回归引入前）曾观察到 deepseek-chat 在后台不调 `write_file`，
-  当时根因可能就是更早版本的同类访问崩溃，也可能另有 prompt 层问题——恢复后跑一单真模型确认
-  outputs 下两份产物都落盘；若报告仍不写 HTML，按旧排查清单对比前端 run 参数（run policy 现已
-  由 `roundtable_run_policy.py` 单源统一，差异面只剩 prompt/上下文）。
-
-### P1 —— 内容质量（本次刚加 inline 广播，待真机验证）
-- **已做**：席位/总控 prompt 注入「任务背景 + 前序交付 + 别提问」。
-- **待验证**：真机跑一次，看席位是否基于真实上下文产出（不再反问/凭记忆瞎编）。
-- **若 inline 不够**：考虑实现**真·thread 广播**（像 `multi_agent._append_thread_message` 那样把
-  「用户意图 / 派活 / 各席位交付」追加进各 thread 历史）。进程内要避开 HTTP loopback 鉴权，
-  得用 checkpointer 直接 append human 消息（参考 `threads` 路由 POST `/state` 的内部逻辑）。这块较重，先看 inline 够不够。
-
-### P2 —— 对话「没有步骤（steps）」
-- **现象**：手动跑的气泡带工具调用步骤卡（`onPhaseChange` → `appendStepIfNew` → `Step2Dialogue.steps`）；
-  后台对话只存了最终正文，没有 steps。
-- **下次怎么做**：后台 `_run_agent_turn` 跑完后，从 checkpoint 的 messages 里抽出 tool_calls，
-  映射成 `Step2Dialogue.steps`（`{id,key,label,ts,args}`），在执行器 `_to_step2_dialogue` 里带上。
-  纯展示增强，不影响内容/报告。
-
-### P3 —— 双跑浪费 + --reload 坑（工程健壮性）
-- 进 Step2 时前端会自动跑一遍本地研讨（真模型、真线程）；点后台挂起后前端那条 run 在后端还会跑完
-  （`stopOrchestration` 只断客户端、停不了后端 run）→ 和后台作业短暂双跑、双倍 token。
-  - 优化方向：进 Step2 改成「显式开始」而非自动跑；或后台挂起时真正 cancel 前端那条 run。
-- `make dev` 的 `--reload` 会杀后台任务（见第二节）；改进项：`--reload-exclude '.deer-flow/*'`。
-
----
-
-## 四、关键文件清单（下次定位用）
-
-**后端**
-| 文件 | 作用 |
-|------|------|
-| `app/gateway/roundtable_inprocess_gateway.py` | 真网关：进程内驱动 leader/seat/re 编排循环 + `build_report_prompt` + inline 广播（`_seat_message`/`_deliveries_block`） |
-| `.../roundtable_orchestrator/progress.py` | 进度落库 + `on_dialogue` 每轮回调 |
-| `.../persistence/roundtable_jobs/` | job 表 model/sql |
-| `app/gateway/deps.py` | `ROUNDTABLE_JOB_GATEWAY` 选网关 + 注入 executor（draft_store） |
-
-**前端**
-| 文件 | 作用 |
-|------|------|
-| `src/roundtable-planning/api/roundtable-jobs.ts` | 作业 API + SSE |
-| `src/roundtable-planning/lib/job-types.ts` | 类型（RoundtableJob/JobDialogue/...） |
-| `src/roundtable-planning/components/DispatchChainModal.tsx` / `DispatchChainJobModal.tsx` | 流程图弹窗 + 对话面板 |
-| `src/roundtable-planning/pages/RoundtablePlanningPage.tsx` | 后台挂起按钮、历史下拉、**runs[] 多运行模型**（`hydrateStep2Run`/`getStep2SnapshotWithRuns`/`normalizeStep2Runs`） |
-| `src/roundtable-planning/api/drafts.ts` | `DraftStep2RunSnapshot`（带 jobId/jobStatus/step3） |
-
----
-
-## 五、关键架构事实（避免重复踩坑）
-
-1. **真网关进程内驱动**：`start_run(body, thread_id, SimpleNamespace(app=app))` + `set_current_user`，
-   不走 HTTP（loopback 调 `/api/multi-agent/*` 不可行：`_auth_headers` 只转发用户 token + owner check）。
-2. **后台作业首启自建全新线程**（`StartParams.is_resume`），绝不复用前端本地研讨的线程（会撞 409）。
-3. **存取统一**：后台作业 = `draft.step2.runs[]` 里一个带 jobId 的 run，和手动跑同一存取/hydrate 路径，
-   **不要再开 observe/回显分支**（之前试过，被否）。
-4. **聊天记录确实存上了**（job.dialogues + draft 的 run）；之前"没存上"是测了旧代码 / 看的是前端本地残缺数据。
-5. 多 worker / SSE：状态进 DB、SSE 走 DB 轮询；真网关进程内无 loopback。
diff --git a/frontend-web/docs/xvqiu-0616.md b/frontend-web/docs/xvqiu-0616.md
new file mode 100644
index 00000000..ed3544d1
--- /dev/null
+++ b/frontend-web/docs/xvqiu-0616.md
@@ -0,0 +1,161 @@
+# 需求文档 · 问答中心左侧栏改造（0616）
+
+> 整理自 2026-06-16 需求草稿。本批需求统一围绕「问答中心」菜单下 5 个页面入口的**左侧侧边栏**进行改造，目标是让每个业务页面的左栏从「通用导航」变为「贴合该业务场景的快捷入口 + 历史记录」。
+>
+> 菜单结构参考：`frontend-web/src/core/page-layout/sidebar-menu.ts` 中 `qa-center`（问答中心）分组，含 5 个子项，正好对应下文 5 条需求。
+
+## 0. 背景与共性约定
+
+| 需求 | 菜单项 | id | 现有路由 | 对应页面/组件 |
+|------|--------|----|----------|--------------|
+| 1 | 通用问答 | `general-qa` | `/page/workspace/chats/new` | Workspace Chat + 左栏（`recent-chat-list.tsx` 等） |
+| 2 | 多智能体会商 | `multi-agent-roundtable` | `/page/strategy/qa/multi-agent-roundtable` | `roundtable-planning/pages/RoundtablePlanningPage.tsx` |
+| 3 | zzJC支持 | `qa-zzjc` | `/page/strategy/agents/6bf/chats/new` | `pages/AgentChatPage.tsx` |
+| 4 | zzXD应用 | `qa-zzxd` | `/page/strategy/agents/10d69686718349688e58a8790cb61024/chats/new` | `pages/AgentChatPage.tsx` |
+| 5 | 课题研究 | `qa-research` | `/page/canvas/ai-writing` | `open-canvas/pages/AIWritingPage.tsx` |
+
+**共性约定（适用于全部 5 条，除非单条另有说明）：**
+
+- 「更多」一律为可点击入口；点击后跳转到对应的完整列表/管理页（具体目标见各条「待确认」）。
+- 列表为空时的兜底：参考各条「现状」中的级联/空态规则；无内容时该区块整体隐藏或显示占位文案（待 UI 设计确认，默认隐藏，与现有 `recent-chat-list.tsx` 一致）。
+- 左栏宽度、折叠交互沿用现有侧边栏体系（`SidebarGroup` 系列组件）。
+- 所有「智能体」相关数据复用 `core/agents`（`Agent` 类型已含 `is_pinned` / `is_favorite` / `featured_order`，scope 支持 `mine` / `square` / `builtin`）。
+
+---
+
+## 1. 通用问答 — 左侧栏改造
+
+**页面**：`/page/workspace/chats/new`
+
+### 1.1 现状
+当前左栏主要展示「最近会话」列表（`recent-chat-list.tsx`，已过滤 `scheduler` / `bootstrap` / `roundtable` 类型线程，置顶会话优先）。
+
+### 1.2 目标布局（自上而下）
+左栏从上到下依次为三块：**新建回答按钮 → 常用智能体 → 聊天记录**。
+
+#### ① 顶部：新建回答
+- 左栏最顶部放置「**新建回答**」按钮（主操作）。
+- 点击 = 开启一次全新的通用问答会话（等价于跳转 `/page/workspace/chats/new`）。
+
+#### ② 常用智能体（最多 3 个 + 更多）
+- 展示「常用智能体」卡片/列表项，**最多 3 个**。
+- 数据来源采用**级联兜底**逻辑（取到即用，不足再降级）：
+  1. 若用户有**置顶智能体**（`is_pinned = true`）→ 展示置顶的；
+  2. 否则展示**用户自己的智能体**（`scope = "mine"`）；
+  3. 仍没有 → 展示**默认广场智能体**（`scope = "square"` / `builtin`，即广场默认推荐）。
+- 三者按上述优先级取，最终最多取 3 个。
+- 下方显示「**更多**」入口。
+- 点击某个常用智能体 → 进入该智能体的对话（待确认：是否为该智能体新建会话）。
+
+#### ③ 聊天记录（最多 10 条 + 更多）
+- 展示用户的问答聊天记录，**最多 10 条**（沿用现有 `useThreads` + 过滤规则）。
+- 下方显示「**更多**」入口。
+- 点击「更多」→ **跳转到问答记录页面**（完整聊天记录列表页）。
+
+### 1.3
+- 「常用智能体」点击后「打开该智能体主页」
+- 「常用智能体」的「更多」跳转目标页 智能体管理 `/page/strategy/agents`。
+- 「聊天记录」的「问答记录页面」具体路由http://localhost:5173/#/page/workspace/chats。
+- 级联兜底是（置顶组够 3 个就完全不展示我的/广场）如果不够「补足到 3 个
+
+---
+
+## 2. 多智能体会商 — 左侧栏重做
+
+**页面**：`/page/strategy/qa/multi-agent-roundtable`（`RoundtablePlanningPage`）
+
+### 2.1 现状
+圆桌会商页已有业务链选择、分析草稿等能力（`roundtable-planning/api/chains.ts`、`hooks/useRoundtableDrafts.ts`）。需求要求**左栏整体重做**。
+
+### 2.2 目标布局（自上而下）
+
+#### ① 顶部：业务链（公共，前 3 个）
+- 顶部展示**业务链**，仅取**公共**链条（`listChains("public")`，即所有已发布 `is_public = true` 的链条）。
+- 仅展示**前 3 个**。
+- 点击某条业务链 → 以该业务链启动一次会商（沿用现有"会商自动串联圆桌"流程，见记忆 `business-chain-mapping`）。
+
+#### ② 下方：分析记录
+- 业务链下方展示「**分析记录**」列表。
+- 数据来源：圆桌分析草稿/历史（`useRoundtableDrafts` / `listDrafts`）。
+- 点击某条分析记录 → 打开对应的历史会商分析。
+
+### 2.3 待确认
+- 「业务链」是否需要「更多」入口跳转到业务链条配置页（`/page/strategy/business-chains`）？需求未提及，默认仅展示前 3 个、无更多。  需要更多入口
+- 「分析记录」是否需要分页/「更多」？需求未提及，默认展示全部或合理上限 分析记录默认全部吧。
+- 公共业务链「前 3 个」的排序口径（创建时间倒序？已有 `listChains` 默认 newest first） 按照现在默认的显示来吧，然后要支持用户单选，如果用户选择后，点击进入多智能体研讨按钮时，直接按照选择的业务链进去执行，如果用户取消勾选，或没有勾选就还是打开弹窗。
+
+---
+
+## 3. zzJC支持 — 左侧栏改造（按类型聚合智能体）
+
+**页面**：`/page/strategy/agents/6bf/chats/new`（当前为单一智能体 `6bf` 的对话页）
+
+### 3.1 现状
+当前 zzJC支持直接进入某个固定智能体（id=`6bf`）的对话页（`AgentChatPage`）。
+
+### 3.2 目标布局
+- **左栏按「类型」分组**，展示该类型下的智能体列表（"类型"待确认，推测为智能体 `tags` / 分类）。
+- 智能体分组**下方**展示**聊天记录**。
+- 左栏**宽度与通用问答（需求 1）保持一致**（"左侧是问答宽"）。
+
+### 3.3 Admin 设置入口
+- 在 **admin 账号权限下**，页面**右侧顶部显示「设置」图标按钮**（齿轮图标）。
+- 点击设置按钮 → 打开设置面板，可对该模块进行配置。
+- 非 admin 账号不显示设置按钮（沿用 `sidebar-menu.ts` 中 `adminOnly` 同款角色判断）。
+
+### 3.4 待确认
+- 「类型」的确切含义：是智能体的 `tags`，还是 zzJC 业务自定义的一组分类？
+- 现 zzJC 入口是单个智能体 `6bf`；改为「按类型展示多个智能体」后，**进入该菜单时默认落地页是什么**（智能体列表？还是仍默认进 `6bf` 对话）？是否需要新建一个 zzJC 聚合页。
+- Admin「设置」具体可配置项：配置「该模块纳入哪些类型/哪些智能体」？还是其他（如默认智能体、欢迎语）？需明确设置项清单。
+- 「聊天记录」过滤范围：仅 zzJC 相关会话，还是全部会话？
+
+---
+
+## 4. zzXD应用 — 左侧栏改造
+
+**页面**：`/page/strategy/agents/10d69686718349688e58a8790cb61024/chats/new`
+
+### 4.1 需求
+**与需求 3（zzJC支持）完全一致**：
+- 左栏按类型展示该类型下的智能体 + 下方聊天记录；
+- 左栏宽度同通用问答；
+- admin 账号下右侧顶部显示设置图标按钮，可设置。
+
+### 4.2 待确认
+- 同需求 3 全部待确认项。
+- zzXD 与 zzJC 是否共用同一套「按类型 + 设置」组件（仅数据源/默认智能体不同），还是各自独立配置？建议**抽象为同一可复用组件**，通过参数区分。
+
+---
+
+## 5. 课题研究 — 左侧栏改造（研究模板 + 笔记本）
+
+**页面**：`/page/canvas/ai-writing`（`AIWritingPage`）
+
+### 5.1 目标布局（自上而下）
+
+#### ① 上半部分：研究模板
+- 左栏上半部分展示「**研究模板**」。
+- 研究模板 = **原来的「文章类型」**（`open-canvas/hooks/useArticleTypes.ts` 的 `articleTypes`，即 AI 写作表单里的「文章类型」选择项）。
+- 点击某个研究模板 → 以该模板/文章类型发起课题研究（沿用现有 `articleType` 设置逻辑）。
+
+#### ② 下半部分：笔记本
+- 左栏下半部分展示「**笔记本**」。
+- 笔记本内容 = **「简洁模式」左侧菜单下方展示的笔记本内容**（`useStudioNotebooks`，即 `workspace-nav-studio.tsx` 中渲染的 studio notebooks 列表）。
+- 点击某个笔记本 → 打开该笔记本（沿用现有 `/page/workspace/studio/notebooks/{id}` 行为，或作为课题研究的素材来源 `materialSource = 'notebook'`，待确认）。
+
+> 备注：草稿原文写作「简介模式」，结合代码应为「**简洁模式**」（`appearance-settings-page.tsx` 中的全局界面模式：全量 / 简洁；简洁模式下笔记本在左栏菜单展示）。
+
+### 5.2 待确认
+- 「研究模板」点击行为：仅预填表单 `articleType`，还是直接开始生成？
+- 「笔记本」点击行为：跳转笔记本详情页，还是将笔记本设为当前课题研究的素材来源（`materialSource='notebook'`）？
+- 上/下两部分是否各自可折叠、是否需要「更多」入口与数量上限？需求未提及。
+
+---
+
+## 6. 整体待确认与建议
+
+1. **复用性**：需求 3 / 4（zzJC / zzXD）布局完全相同，建议实现为同一套「按类型分组 + 聊天记录 + admin 设置」可复用组件，通过配置区分模块。
+2. **左栏宽度统一**：需求 1 / 3 / 4 都要求"问答宽"，建议统一左栏宽度常量，避免各页面各写一套。
+3. **「更多」目标页**：需求 1 涉及「智能体更多」「聊天记录更多→问答记录页」两个跳转目标需先确定路由是否已存在。
+4. **Admin 设置项清单**：需求 3 / 4 的设置面板具体能配置什么，需要产品给出字段清单后才能落地。
+5. **入口语义变更**：需求 3 / 4 把原「单一智能体对话」入口改为「按类型聚合」，菜单 `path` 与默认落地页可能需要调整，需同步确认。
diff --git a/frontend-web/src/components/workspace/frequent-agents-list.tsx b/frontend-web/src/components/workspace/frequent-agents-list.tsx
new file mode 100644
index 00000000..e84213aa
--- /dev/null
+++ b/frontend-web/src/components/workspace/frequent-agents-list.tsx
@@ -0,0 +1,127 @@
+"use client";
+
+import { Bot, ChevronRight } from "lucide-react";
+import Link from "next/link";
+import { usePathname } from "next/navigation";
+import { useMemo } from "react";
+
+import {
+  SidebarGroup,
+  SidebarGroupContent,
+  SidebarGroupLabel,
+  SidebarMenu,
+  SidebarMenuButton,
+  SidebarMenuItem,
+} from "@/components/ui/sidebar";
+import { useAgentsPaginated } from "@/core/agents";
+import type { Agent } from "@/core/agents";
+import { desensitize } from "@/lib/desensitize";
+
+/** 常用智能体最多展示个数（需求 1.2 ②）。 */
+const MAX_FREQUENT = 3;
+
+/** 「更多」跳转目标：智能体管理页（需求 1.3）。 */
+const MORE_HREF = "/page/strategy/agents";
+
+/** 打开某个智能体主页（其对话落地页，thread_id=new = 新会话）。 */
+function agentHomeHref(agentId: string): string {
+  return `/page/strategy/agents/${agentId}/chats/new`;
+}
+
+/**
+ * 「常用智能体」侧栏区块（通用问答左栏，全量模式）。
+ *
+ * 级联兜底取数（需求 1.2 ② / 1.3）——按优先级补足到最多 3 个、去重：
+ *   1. 置顶智能体（``is_pinned``，可来自「我的」或「广场」）；
+ *   2. 不足 3 个 → 用「我的」智能体（``scope="mine"``）补足；
+ *   3. 仍不足 → 用「广场」默认推荐（``scope="square"``）补足。
+ *
+ * 取不到任何智能体时整块隐藏（与 ``RecentChatList`` 的空态约定一致）。
+ */
+export function FrequentAgentsList() {
+  const pathname = usePathname();
+  const { agents: mine } = useAgentsPaginated({
+    scope: "mine",
+    page: 1,
+    pageSize: 20,
+  });
+  const { agents: square } = useAgentsPaginated({
+    scope: "square",
+    page: 1,
+    pageSize: 20,
+  });
+
+  const frequent = useMemo<Agent[]>(() => {
+    const out: Agent[] = [];
+    const seen = new Set<string>();
+    const add = (agent: Agent | undefined) => {
+      if (!agent || out.length >= MAX_FREQUENT || seen.has(agent.id)) return;
+      seen.add(agent.id);
+      out.push(agent);
+    };
+    // 1. 置顶组优先（够 3 个则下面我的/广场都不再补）。
+    [...mine, ...square].filter((a) => a.is_pinned).forEach(add);
+    // 2. 我的智能体补足。
+    mine.forEach(add);
+    // 3. 广场默认推荐补足。
+    square.forEach(add);
+    return out;
+  }, [mine, square]);
+
+  if (frequent.length === 0) {
+    return null;
+  }
+
+  return (
+    <SidebarGroup className="px-2 pt-3 mt-2 border-t border-sidebar-border/60">
+      <SidebarGroupLabel className="text-muted-foreground/70 px-3 mb-1 text-[10px] font-semibold tracking-[0.12em] uppercase select-none">
+        常用智能体
+      </SidebarGroupLabel>
+      <SidebarGroupContent>
+        <SidebarMenu>
+          <div className="flex w-full flex-col gap-0.5">
+            {frequent.map((agent) => {
+              const href = agentHomeHref(agent.id);
+              const isActive = pathname === href;
+              return (
+                <SidebarMenuItem key={agent.id}>
+                  <SidebarMenuButton
+                    isActive={isActive}
+                    asChild
+                    className={
+                      "rounded-lg h-7 px-3 text-[13px] text-muted-foreground hover:text-foreground hover:bg-sidebar-accent/60 " +
+                      "data-[active=true]:bg-white data-[active=true]:text-foreground data-[active=true]:shadow-[0_1px_2px_rgba(15,23,42,0.05)] " +
+                      "dark:data-[active=true]:bg-slate-700/80 dark:data-[active=true]:text-slate-100 dark:data-[active=true]:shadow-none"
+                    }
+                  >
+                    <Link
+                      className="flex w-full items-center gap-2 whitespace-nowrap"
+                      href={href}
+                    >
+                      <Bot className="size-3.5 shrink-0 text-muted-foreground/70" />
+                      <span className="truncate">{desensitize(agent.name)}</span>
+                    </Link>
+                  </SidebarMenuButton>
+                </SidebarMenuItem>
+              );
+            })}
+            <SidebarMenuItem>
+              <SidebarMenuButton
+                asChild
+                className="rounded-lg h-7 px-3 text-[12px] text-muted-foreground/70 hover:text-foreground hover:bg-sidebar-accent/60"
+              >
+                <Link
+                  className="flex w-full items-center justify-between gap-1"
+                  href={MORE_HREF}
+                >
+                  <span>更多</span>
+                  <ChevronRight className="size-3.5 shrink-0" />
+                </Link>
+              </SidebarMenuButton>
+            </SidebarMenuItem>
+          </div>
+        </SidebarMenu>
+      </SidebarGroupContent>
+    </SidebarGroup>
+  );
+}
diff --git a/frontend-web/src/components/workspace/recent-chat-list.tsx b/frontend-web/src/components/workspace/recent-chat-list.tsx
index 457c8806..702bcaa7 100644
--- a/frontend-web/src/components/workspace/recent-chat-list.tsx
+++ b/frontend-web/src/components/workspace/recent-chat-list.tsx
@@ -2,6 +2,7 @@
 
 import {
   ChevronDown,
+  ChevronRight,
   Download,
   FileJson,
   FileText,
@@ -74,7 +75,17 @@ import { desensitize } from "@/lib/desensitize";
 import { isIMEComposing } from "@/lib/ime";
 import { cn } from "@/lib/utils";
 
-export function RecentChatList({ compact = false }: { compact?: boolean }) {
+export function RecentChatList({
+  compact = false,
+  limit,
+  moreHref,
+}: {
+  compact?: boolean;
+  /** 最多展示的会话条数；省略则不限制（需求 1.3：通用问答最多 10 条）。 */
+  limit?: number;
+  /** 「更多」入口跳转目标；省略则不展示「更多」行。 */
+  moreHref?: string;
+}) {
   const { t } = useI18n();
   const router = useRouter();
   const pathname = usePathname();
@@ -190,6 +201,8 @@ export function RecentChatList({ compact = false }: { compact?: boolean }) {
   if (visibleThreads.length === 0) {
     return null;
   }
+  const shownThreads =
+    limit != null ? visibleThreads.slice(0, limit) : visibleThreads;
   return (
     <>
       <SidebarGroup className="px-2 pt-3 mt-2 border-t border-sidebar-border/60">
@@ -225,7 +238,7 @@ export function RecentChatList({ compact = false }: { compact?: boolean }) {
         >
           <SidebarMenu>
             <div className="flex w-full flex-col gap-0.5">
-              {visibleThreads.map((thread) => {
+              {shownThreads.map((thread) => {
                 const isActive = pathOfThread(thread) === pathname;
                 const pinned = isThreadPinned(thread);
                 const favorited = isThreadFavorite(thread);
@@ -368,6 +381,22 @@ export function RecentChatList({ compact = false }: { compact?: boolean }) {
                   </SidebarMenuItem>
                 );
               })}
+              {moreHref && (
+                <SidebarMenuItem>
+                  <SidebarMenuButton
+                    asChild
+                    className="rounded-lg h-7 px-3 text-[12px] text-muted-foreground/70 hover:text-foreground hover:bg-sidebar-accent/60"
+                  >
+                    <Link
+                      className="flex w-full items-center justify-between gap-1"
+                      href={moreHref}
+                    >
+                      <span>{t.common.more}</span>
+                      <ChevronRight className="size-3.5 shrink-0" />
+                    </Link>
+                  </SidebarMenuButton>
+                </SidebarMenuItem>
+              )}
             </div>
           </SidebarMenu>
         </SidebarGroupContent>
diff --git a/frontend-web/src/components/workspace/workspace-sidebar.tsx b/frontend-web/src/components/workspace/workspace-sidebar.tsx
index a721a9c9..8712f9d0 100644
--- a/frontend-web/src/components/workspace/workspace-sidebar.tsx
+++ b/frontend-web/src/components/workspace/workspace-sidebar.tsx
@@ -10,7 +10,9 @@ import {
 } from "@/components/ui/sidebar";
 import { useTheme } from "@/contexts/ThemeContext";
 import { cn } from "@/lib/utils";
+import { usePathname } from "next/navigation";
 
+import { FrequentAgentsList } from "./frequent-agents-list";
 import { RecentChatList } from "./recent-chat-list";
 import { RecentRoundtableList } from "./recent-roundtable-list";
 import { RecentWritingList } from "./recent-writing-list";
@@ -25,7 +27,11 @@ export function WorkspaceSidebar({
 }: React.ComponentProps<typeof Sidebar> & { compactChrome?: boolean }) {
   const { open: isSidebarOpen } = useSidebar();
   const { theme } = useTheme();
+  const pathname = usePathname();
   const useWhiteSidebar = compactChrome && theme === "light";
+  // 通用问答（全量模式）左栏：新建回答 → 常用智能体 → 聊天记录（需求 0616 #1）。
+  // 仅在通用问答（/page/workspace/chats*）下启用三块式布局；其余 workspace 页保持原样。
+  const isGeneralQa = pathname.startsWith("/page/workspace/chats");
 
   // 紧凑模式：用「外层定宽容器 + transition-[width]」做展开/收起过渡动画。
   // collapsible="none" 的 Sidebar 是普通 in-flow 定宽块（非 absolute，故不受本布局对
@@ -53,7 +59,12 @@ export function WorkspaceSidebar({
         >
           <SidebarContent>
             <WorkspaceNavChatList showCollapseTrigger compact />
-            <RecentChatList compact />
+            {isGeneralQa && <FrequentAgentsList />}
+            <RecentChatList
+              compact
+              limit={isGeneralQa ? 10 : undefined}
+              moreHref={isGeneralQa ? "/page/workspace/chats" : undefined}
+            />
           </SidebarContent>
           <SidebarRail />
         </Sidebar>
diff --git a/offline-backend-20260512/backend/config.yaml b/offline-backend-20260512/backend/config.yaml
index 4f1c6569..f2971f53 100644
--- a/offline-backend-20260512/backend/config.yaml
+++ b/offline-backend-20260512/backend/config.yaml
@@ -14,7 +14,7 @@ models:
     display_name: DeepSeek Chat
     use: langchain_openai:ChatOpenAI
     model: deepseek-chat
-    api_key: sk-3fee562f834a4c08bb0d471ef44cb445
+    api_key: sk-35345154983047aba9e12faa86c5f62c
     base_url: https://api.deepseek.com
     request_timeout: 600.0
     max_retries: 2

From 2526304394a05f2eee8395b853c9f20744084846 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E8=91=A3=E5=AE=8F=E5=87=AF?= <hkdong@kland.com.cn>
Date: Tue, 16 Jun 2026 10:09:30 +0800
Subject: [PATCH 6/9] sd

---
 offline-backend-20260512/backend/config.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/offline-backend-20260512/backend/config.yaml b/offline-backend-20260512/backend/config.yaml
index 4f1c6569..f2971f53 100644
--- a/offline-backend-20260512/backend/config.yaml
+++ b/offline-backend-20260512/backend/config.yaml
@@ -14,7 +14,7 @@ models:
     display_name: DeepSeek Chat
     use: langchain_openai:ChatOpenAI
     model: deepseek-chat
-    api_key: sk-3fee562f834a4c08bb0d471ef44cb445
+    api_key: sk-35345154983047aba9e12faa86c5f62c
     base_url: https://api.deepseek.com
     request_timeout: 600.0
     max_retries: 2

From da3d38013e3062f023580222a320bf6859ef736f Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E8=91=A3=E5=AE=8F=E5=87=AF?= <hkdong@kland.com.cn>
Date: Tue, 16 Jun 2026 14:15:49 +0800
Subject: [PATCH 7/9] =?UTF-8?q?feat(ai-writing):=20=E7=B4=A0=E6=9D=90?=
 =?UTF-8?q?=E6=94=B6=E9=9B=86=E6=8A=80=E8=83=BD=E8=B0=83=E7=94=A8=E5=A4=B1?=
 =?UTF-8?q?=E8=B4=A5=E6=97=B6=E5=AE=B9=E9=94=99=E5=9B=9E=E9=80=80=E5=88=B0?=
 =?UTF-8?q?=E5=85=B6=E4=BB=96=E6=8A=80=E8=83=BD?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

默认/选中技能调用失败（超时、报错、全部检索词失败）时不再直接报错，
改用其他技能继续检索：节点 emit skill_error 步骤（带真实原因），配置技能
全部失败且无新素材时跑 fallback 兜底（内置知识库检索优先，再到目录其它已
启用技能）。前端 SkillStepsCard 渲染 skill_error（玫红）/ fallback（琥珀），
时间线展示"技能报错→改用其他技能搜集"。

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---
 .../components/ai-writing/SkillStepsCard.tsx  |  33 +++--
 .../src/open-canvas/types/ai-writing.ts       |   4 +
 offline-backend-20260512/backend/CLAUDE.md    |   2 +-
 .../agents/ai_writing/nodes/researcher.py     | 114 +++++++++++++++---
 4 files changed, 126 insertions(+), 27 deletions(-)

diff --git a/frontend-web/src/open-canvas/components/ai-writing/SkillStepsCard.tsx b/frontend-web/src/open-canvas/components/ai-writing/SkillStepsCard.tsx
index cb8c349c..0faf8d74 100644
--- a/frontend-web/src/open-canvas/components/ai-writing/SkillStepsCard.tsx
+++ b/frontend-web/src/open-canvas/components/ai-writing/SkillStepsCard.tsx
@@ -1,5 +1,5 @@
 import { useState } from 'react';
-import { ChevronRight, Sparkles, ListChecks, Zap, Loader2, CheckCircle2, CircleStop, Search } from 'lucide-react';
+import { ChevronRight, Sparkles, ListChecks, Zap, Loader2, CheckCircle2, CircleStop, Search, AlertTriangle, Shuffle } from 'lucide-react';
 import { cn } from '@/lib/utils';
 import type { SkillSearchStep } from '../../types/ai-writing';
 import { CollapseTransition } from './CollapseTransition';
@@ -16,6 +16,8 @@ const STEP_ICON: Record<SkillSearchStep['type'], typeof Sparkles> = {
   select_skills: Sparkles,
   invoke_skill: Zap,
   skill_result: CheckCircle2,
+  skill_error: AlertTriangle,
+  fallback: Shuffle,
   stop_early: CircleStop,
 };
 
@@ -25,6 +27,8 @@ const STEP_LABEL: Record<SkillSearchStep['type'], string> = {
   select_skills: '筛选相关技能',
   invoke_skill: '调用技能',
   skill_result: '检索结果',
+  skill_error: '技能调用失败',
+  fallback: '改用其他技能',
   stop_early: '提前停止',
 };
 
@@ -71,6 +75,9 @@ export function SkillStepsCard({ steps, inProgress }: SkillStepsCardProps) {
             const label = STEP_LABEL[step.type] ?? step.type;
             const isStopEarly = step.type === 'stop_early';
             const isResult = step.type === 'skill_result';
+            // 容错：技能调用失败标红，回退其它技能标琥珀，提示用户"出错了但已自动改用其他技能"
+            const isError = step.type === 'skill_error';
+            const isFallback = step.type === 'fallback';
             const isActive = idx === activeIdx;
             return (
               <li key={idx} className="flex items-start gap-2 text-[11px]">
@@ -79,11 +86,13 @@ export function SkillStepsCard({ steps, inProgress }: SkillStepsCardProps) {
                     'shrink-0 mt-0.5 inline-flex items-center justify-center size-4 rounded-full text-[10px] font-medium',
                     isActive
                       ? 'bg-violet-500 text-white'
-                      : isStopEarly
-                        ? 'bg-amber-100 dark:bg-amber-900/60 text-amber-600 dark:text-amber-300'
-                        : isResult
-                          ? 'bg-emerald-100 dark:bg-emerald-900/60 text-emerald-600 dark:text-emerald-300'
-                          : 'bg-violet-100 dark:bg-violet-900/60 text-violet-600 dark:text-violet-300',
+                      : isError
+                        ? 'bg-rose-100 dark:bg-rose-900/60 text-rose-600 dark:text-rose-300'
+                        : isStopEarly || isFallback
+                          ? 'bg-amber-100 dark:bg-amber-900/60 text-amber-600 dark:text-amber-300'
+                          : isResult
+                            ? 'bg-emerald-100 dark:bg-emerald-900/60 text-emerald-600 dark:text-emerald-300'
+                            : 'bg-violet-100 dark:bg-violet-900/60 text-violet-600 dark:text-violet-300',
                   )}
                 >
                   {isActive ? <Loader2 className="size-2.5 animate-spin" /> : idx + 1}
@@ -92,11 +101,13 @@ export function SkillStepsCard({ steps, inProgress }: SkillStepsCardProps) {
                   <div
                     className={cn(
                       'flex items-center gap-1 font-medium',
-                      isStopEarly
-                        ? 'text-amber-700 dark:text-amber-300'
-                        : isResult
-                          ? 'text-emerald-700 dark:text-emerald-300'
-                          : 'text-violet-700 dark:text-violet-300',
+                      isError
+                        ? 'text-rose-700 dark:text-rose-300'
+                        : isStopEarly || isFallback
+                          ? 'text-amber-700 dark:text-amber-300'
+                          : isResult
+                            ? 'text-emerald-700 dark:text-emerald-300'
+                            : 'text-violet-700 dark:text-violet-300',
                     )}
                   >
                     <Icon className="size-3 shrink-0" />
diff --git a/frontend-web/src/open-canvas/types/ai-writing.ts b/frontend-web/src/open-canvas/types/ai-writing.ts
index bb6b783e..361cb090 100644
--- a/frontend-web/src/open-canvas/types/ai-writing.ts
+++ b/frontend-web/src/open-canvas/types/ai-writing.ts
@@ -20,6 +20,8 @@ export interface AIMaterial {
 //   select_skills：LLM 按相关度排序的候选
 //   invoke_skill：开始调用某个技能去检索
 //   skill_result：某个技能检索返回了 N 条素材
+//   skill_error：某个技能调用失败（超时/报错），将改用其他技能（容错）
+//   fallback：配置技能全部失败，回退到其他技能继续搜集（容错）
 //   stop_early：累计素材达到上限，停止后续技能
 export interface SkillSearchStep {
   type:
@@ -28,6 +30,8 @@ export interface SkillSearchStep {
     | "select_skills"
     | "invoke_skill"
     | "skill_result"
+    | "skill_error"
+    | "fallback"
     | "stop_early";
   message: string;
   skillNames?: string[];
diff --git a/offline-backend-20260512/backend/CLAUDE.md b/offline-backend-20260512/backend/CLAUDE.md
index f6812132..0e623e64 100644
--- a/offline-backend-20260512/backend/CLAUDE.md
+++ b/offline-backend-20260512/backend/CLAUDE.md
@@ -265,7 +265,7 @@ The AI-writing graph (`packages/harness/deerflow/agents/ai_writing/`) has four s
 
 Seeding mirrors the roundtable pattern: verified `config.yaml` + `SOUL.md` ship in `app/gateway/routers/_ai_writing_seed_assets/<id>/`; `_ai_writing_seed.py::ensure_ai_writing_functional_agents()` copies them to `.deer-flow/agents/` when missing (never overwrites local edits), called from the startup lifespan (before `_sync_legacy_agents`, so the agents appear in `/api/agents` on first boot) and self-heals from `GET /api/ai-writing/sessions` + `GET /api/ai-writing/article-types`. Tests: `tests/test_ai_writing_seed.py`, `tests/test_ai_writing_agent_runner.py`.
 
-**Skill-driven retrieval (检索类型).** `material_source` has exactly two values: `general` (通用检索, default) and `notebook` (我的空间); legacy `knowledge_base`/`skill` are normalized to `general` by the researcher node. General retrieval is orchestrated from the **skills configured on the `ai-writing-researcher` agent** (`config.yaml → skills`, editable from the agent management page): the built-in `knowledge-base-search` skill (知识库检索) runs the native keyword fan-out web/intranet search (the old default path), while any other configured skill searches with a `description`-targeted query; skills run in configured order with early-stop at `max_materials`, and an empty/unresolvable skills config falls back to the built-in knowledge search so retrieval always works. The skill itself is seeded from `_ai_writing_seed_assets/skills/knowledge-base-search/SKILL.md` into `skills/public/` (never overwrites), and a pre-existing researcher `config.yaml` lacking the `skills` key is patched in place with the default (explicit `skills: []` is respected). **PAUSE-1 input append flow**: when the user types natural language into the material-confirm input (`re_search`/legacy `enable_skill_search` + `user_query`), the node matches the most relevant configured skill via `SkillPicker(candidates=…)` (skipped when only one skill is configured), searches with the user's words, and **appends** the new materials to the existing package (URL-dedup, original order preserved, no LLM re-rank/truncation, 100-item hard cap). Tests: `tests/test_ai_writing_material_source.py`.
+**Skill-driven retrieval (检索类型).** `material_source` has exactly two values: `general` (通用检索, default) and `notebook` (我的空间); legacy `knowledge_base`/`skill` are normalized to `general` by the researcher node. General retrieval is orchestrated from the **skills configured on the `ai-writing-researcher` agent** (`config.yaml → skills`, editable from the agent management page): the built-in `knowledge-base-search` skill (知识库检索) runs the native keyword fan-out web/intranet search (the old default path), while any other configured skill searches with a `description`-targeted query; skills run in configured order with early-stop at `max_materials`, and an empty/unresolvable skills config falls back to the built-in knowledge search so retrieval always works. The skill itself is seeded from `_ai_writing_seed_assets/skills/knowledge-base-search/SKILL.md` into `skills/public/` (never overwrites), and a pre-existing researcher `config.yaml` lacking the `skills` key is patched in place with the default (explicit `skills: []` is respected). **PAUSE-1 input append flow**: when the user types natural language into the material-confirm input (`re_search`/legacy `enable_skill_search` + `user_query`), the node matches the most relevant configured skill via `SkillPicker(candidates=…)` (skipped when only one skill is configured), searches with the user's words, and **appends** the new materials to the existing package (URL-dedup, original order preserved, no LLM re-rank/truncation, 100-item hard cap). **Skill-failure fault tolerance**: a configured skill that times out / raises / has all its keyword searches fail (vs. legitimately returning 0 results) no longer just silently moves on — the researcher node emits a `skill_search_step` of type `skill_error` (carrying the error `reason`), and if the configured skills collected **no** new materials at all it runs a `fallback` pass over other available skills (`_fallback_candidate_skills`: the built-in `knowledge-base-search` native path first, then other enabled catalog skills not yet tried) instead of erroring out. The frontend `SkillStepsCard` renders `skill_error` (rose) and `fallback` (amber) so the timeline shows "技能报错 → 改用其他技能搜集". Tests: `tests/test_ai_writing_material_source.py`.
 
 ### Admin Leaderboard Snapshots (`/api/admin/users/leaderboard*`)
 
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/researcher.py b/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/researcher.py
index c4ceb06f..8a4c58f0 100644
--- a/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/researcher.py
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/researcher.py
@@ -141,6 +141,30 @@ def _load_configured_skills() -> list:
     return resolved or [_fallback_knowledge_skill()]
 
 
+def _fallback_candidate_skills(tried: set[str]) -> list:
+    """容错兜底：从技能目录里挑出**尚未尝试过**的可用技能，内置「知识库检索」优先。
+
+    当配置技能调用全部失败、且没收集到任何新素材时，researcher 用这些技能再试一轮，
+    而不是直接报错。内置知识库检索（原生多关键词并发，最稳）排在最前，其后是技能目录里
+    其它已启用的技能。技能目录加载失败时只回退内置知识库检索，保证兜底链永不断。
+    """
+    candidates: list = []
+    seen: set[str] = set(tried)
+    if KNOWLEDGE_SEARCH_SKILL_NAME not in seen:
+        candidates.append(_fallback_knowledge_skill())
+        seen.add(KNOWLEDGE_SEARCH_SKILL_NAME)
+    try:
+        from deerflow.skills.storage import get_or_new_skill_storage
+
+        for s in get_or_new_skill_storage().load_skills(enabled_only=True):
+            if getattr(s, "name", None) and s.name not in seen:
+                candidates.append(s)
+                seen.add(s.name)
+    except Exception:
+        logger.warning("researcher: 兜底加载技能目录失败", exc_info=True)
+    return candidates
+
+
 def _resolve_keyword_count(raw: object) -> int:
     """把用户传入的检索关键词数量规整到 [1, _MAX_KEYWORD_COUNT]，非法值回退默认 6。"""
     try:
@@ -511,24 +535,35 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
                 added += 1
             return added
 
-        async def _timed_search(skill_name: str, query: str, limit: int) -> list[SearchResult]:
+        async def _timed_search(skill_name: str, query: str, limit: int) -> tuple[list[SearchResult], str]:
             # 加超时兜底：单个技能检索最多等 _SKILL_SEARCH_TIMEOUT_SECONDS 秒，
             # 超时不再阻塞后续技能，避免整段过程看起来"冻结"。
+            # 返回 (结果, 错误详情)：错误详情非空 = 该技能调用失败（触发兜底）。
             try:
-                return await asyncio.wait_for(
+                results = await asyncio.wait_for(
                     provider.search(query, max_results=limit),
                     timeout=_SKILL_SEARCH_TIMEOUT_SECONDS,
-                ) or []
+                )
+                return (results or []), ""
             except asyncio.TimeoutError:
+                msg = f"检索超时（>{_SKILL_SEARCH_TIMEOUT_SECONDS}s）"
                 logger.warning("skill[%s] search timed out after %ss", skill_name, _SKILL_SEARCH_TIMEOUT_SECONDS)
-                search_errors.append(f"skill[{skill_name}]: 检索超时（>{_SKILL_SEARCH_TIMEOUT_SECONDS}s）")
-                return []
+                search_errors.append(f"skill[{skill_name}]: {msg}")
+                return [], msg
             except Exception as e:
                 logger.warning("skill[%s] search failed: %s", skill_name, e)
                 search_errors.append(f"skill[{skill_name}]: {e}")
-                return []
+                return [], str(e)
 
-        for sk in ordered:
+        def _skill_display(sk) -> str:
+            return _KNOWLEDGE_SKILL_DISPLAY if sk.name == KNOWLEDGE_SEARCH_SKILL_NAME else sk.name
+
+        async def _execute_skill(sk) -> tuple[int, str]:
+            """运行单个技能检索，返回 (新增素材数, 错误详情)。
+
+            错误详情非空 = 该技能整体调用失败（超时 / 抛异常 / 全部检索词失败），
+            调用方据此 emit ``skill_error`` 步骤并触发"改用其它技能"的兜底逻辑。
+            """
             is_knowledge = sk.name == KNOWLEDGE_SEARCH_SKILL_NAME
             if is_knowledge and not is_user_query_search:
                 # 原生知识库检索：对全部检索词并发检索（旧版默认路径）。
@@ -557,13 +592,20 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
 
                 gathered = await asyncio.gather(*[_search_one(kw) for kw in search_keywords], return_exceptions=True)
                 added = 0
+                errs: list[str] = []
                 for results in gathered:
                     if isinstance(results, Exception):
                         logger.warning("search failed: %s", results)
-                        search_errors.append(f"{type(results).__name__}: {results}")
+                        errs.append(f"{type(results).__name__}: {results}")
                         continue
                     added += _collect(results)
                 search_errors.extend(getattr(provider, "errors", []))
+                search_errors.extend(errs)
+                # 全部检索词都抛异常且 0 新增 → 视为该技能整体调用失败，触发兜底
+                error_detail = ""
+                if added == 0 and errs and len(errs) == total:
+                    error_detail = "；".join(dict.fromkeys(errs))[:200]
+                return added, error_detail
             else:
                 # 定向技能检索：user_query 模式优先用用户原话；非知识库技能再
                 # 附上 description 让 query 贴合该技能的领域。
@@ -576,14 +618,35 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
                     "message": f"使用技能「{sk.name}」检索：{skill_query[:60]}",
                     "skill_name": sk.name,
                 })
-                added = _collect(await _timed_search(sk.name, skill_query, per_skill))
+                results, error_detail = await _timed_search(sk.name, skill_query, per_skill)
+                return _collect(results), error_detail
 
-            _emit_skill_step({
-                "type": "skill_result",
-                "message": f"技能「{_KNOWLEDGE_SKILL_DISPLAY if is_knowledge else sk.name}」检索到 {added} 条新素材",
-                "skill_name": sk.name,
-                "matched_count": added,
-            })
+        # 失败技能名（去重后用于兜底提示）：记录配置技能里调用失败的那些
+        failed_skill_names: list[str] = []
+        tried_names: set[str] = set()
+
+        def _emit_skill_outcome(sk, added: int, error_detail: str) -> None:
+            """单个技能跑完后，按成功 / 失败 emit 对应步骤事件。"""
+            if error_detail:
+                failed_skill_names.append(_skill_display(sk))
+                _emit_skill_step({
+                    "type": "skill_error",
+                    "message": f"技能「{_skill_display(sk)}」调用失败，尝试使用其他技能检索",
+                    "skill_name": sk.name,
+                    "reason": error_detail,
+                })
+            else:
+                _emit_skill_step({
+                    "type": "skill_result",
+                    "message": f"技能「{_skill_display(sk)}」检索到 {added} 条新素材",
+                    "skill_name": sk.name,
+                    "matched_count": added,
+                })
+
+        for sk in ordered:
+            added, error_detail = await _execute_skill(sk)
+            tried_names.add(sk.name)
+            _emit_skill_outcome(sk, added, error_detail)
 
             # 早停：累积素材已达上限
             if len(new_results) >= max_materials:
@@ -594,6 +657,27 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
                 })
                 break
 
+        # ── 容错兜底：配置技能调用失败且没收集到任何新素材 → 改用其它可用技能继续检索 ──
+        # （不直接报错；前端时间线靠 skill_error / fallback 步骤展示"技能报错，改用其它技能搜集"）
+        if not new_results and failed_skill_names:
+            fallback_skills = _fallback_candidate_skills(tried_names)
+            if fallback_skills:
+                _emit_skill_step({
+                    "type": "fallback",
+                    "message": (
+                        f"已配置技能（{'、'.join(dict.fromkeys(failed_skill_names))}）检索失败，"
+                        "改用其他技能继续搜集素材"
+                    ),
+                    "skill_names": [_skill_display(s) for s in fallback_skills],
+                })
+                for sk in fallback_skills:
+                    added, error_detail = await _execute_skill(sk)
+                    tried_names.add(sk.name)
+                    _emit_skill_outcome(sk, added, error_detail)
+                    if new_results:
+                        # 兜底已拿到素材即止，不再无谓尝试后续技能
+                        break
+
     # 合并去重
     # - 首次检索 / 普通 re_search（无输入，用户只想重出）：旧素材丢弃，覆盖式替换
     # - 素材不足求助 supplement_search：保留旧素材 + 追加新检索结果

From 8d1669adad4a8b758a988b01ead40c182528a5de Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E8=91=A3=E5=AE=8F=E5=87=AF?= <hkdong@kland.com.cn>
Date: Tue, 16 Jun 2026 14:17:30 +0800
Subject: [PATCH 8/9] =?UTF-8?q?feat(roundtable):=20=E5=90=8E=E5=8F=B0?=
 =?UTF-8?q?=E4=BD=9C=E4=B8=9A=E6=8C=89=E7=94=A8=E6=88=B7=E5=B9=B6=E5=8F=91?=
 =?UTF-8?q?=E9=99=90=E6=B5=81=20+=20=E8=8D=89=E7=A8=BF=E6=8C=81=E4=B9=85?=
 =?UTF-8?q?=E5=8C=96=E4=BC=98=E5=8C=96?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

后台挂起作业每个 agent 轮次（leader/seat/report/summary/dashboard）开跑前
先取该用户的后台并发额度（默认 2，ROUNDTABLE_BACKGROUND_SLOTS_PER_USER 可调），
跑完释放；同一用户多个后台作业共享额度，避免把单连接 SQLite checkpointer 和
离线模型端点打满拖垮前台实时会商。前台实时路径不经过此额度，不受限。

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---
 .../hooks/useDraftPersistence.ts              |  51 +++++++
 .../pages/RoundtablePlanningPage.tsx          | 125 +++++++++++++++---
 .../gateway/roundtable_inprocess_gateway.py   |  48 ++++++-
 3 files changed, 204 insertions(+), 20 deletions(-)

diff --git a/frontend-web/src/roundtable-planning/hooks/useDraftPersistence.ts b/frontend-web/src/roundtable-planning/hooks/useDraftPersistence.ts
index 56b4cce5..599b62cb 100644
--- a/frontend-web/src/roundtable-planning/hooks/useDraftPersistence.ts
+++ b/frontend-web/src/roundtable-planning/hooks/useDraftPersistence.ts
@@ -120,6 +120,11 @@ export interface UseDraftPersistenceReturn {
    * 发起前 / 点查看前各存一次，保住当时的聊天记录不丢。返回 draft id。
    */
   saveDraftNow: (furthestStep?: 1 | 2 | 3) => Promise<string | null>;
+  /**
+   * 用**显式传入的 step2 快照**强制存一次（不读 live 状态）。「后台挂起直接跳转」用：
+   * 跳转会清空 step2，调用方在跳转前冻结快照传进来，保证落库的是挂起那一刻的研讨记录。
+   */
+  saveStep2SnapshotNow: (step2: DraftStep2Snapshot, furthestStep?: 1 | 2 | 3) => Promise<string | null>;
   /** 强制刷新草稿列表。 */
   refresh: () => Promise<void>;
 }
@@ -287,6 +292,51 @@ export function useDraftPersistence(
     [currentStep, enqueueSave, refreshList],
   );
 
+  /**
+   * 用**显式传入的 step2 快照**强制存一次（不读 live 状态）。
+   *
+   * 「后台挂起改直接跳转」用：点挂起后页面立刻 resetToStep1 会清空 step2，而存稿是
+   * fire-and-forget 在跳转之后才真正落库——若仍走 `saveDraftNow`（其 `getStep2Snapshot`
+   * 是延迟读取）就会存到被重置后的空 step2。调用方在跳转前同步 `getStep2Snapshot()` 冻结
+   * 一份快照传进来，这里只认这份，确保存的是挂起那一刻的研讨记录。
+   *
+   * 有草稿 id → update（保护重命名，不带 title）；无 id → 用当前 step1 + 传入 step2 建档。
+   */
+  const saveStep2SnapshotNow = useCallback(
+    (step2: DraftStep2Snapshot, furthestStep: 1 | 2 | 3 = 2): Promise<string | null> => {
+      const run = async (): Promise<string | null> => {
+        const id = currentDraftIdRef.current;
+        if (id) {
+          const updated = await updateDraft(id, { furthestStep, step2 });
+          if (updated) {
+            currentDraftIdRef.current = updated.id;
+            return updated.id;
+          }
+          // 服务端已不存在 → 落回新建。
+        }
+        const step1 = getStep1Snapshot();
+        if (!hasSavableContent(step1, step2)) return null;
+        const created = await createDraft({
+          ...(taskIdRef.current ? { taskId: taskIdRef.current } : {}),
+          title: deriveDraftName({
+            intentReady: step1.intentReady,
+            interactiveMessages: step1.interactiveMessages,
+          }),
+          furthestStep,
+          step1,
+          step2,
+        });
+        currentDraftIdRef.current = created.id;
+        setCurrentDraftId(created.id);
+        return created.id;
+      };
+      const next = saveQueueRef.current.then(run, run);
+      saveQueueRef.current = next.catch(() => undefined);
+      return next as Promise<string | null>;
+    },
+    [getStep1Snapshot, hasSavableContent, setCurrentDraftId],
+  );
+
   const handleManualSave = useCallback(() => {
     const furthest = (currentStep as 1 | 2 | 3) || 1;
     void (async () => {
@@ -535,6 +585,7 @@ export function useDraftPersistence(
     handleCancelRename,
     ensureDraft,
     saveDraftNow,
+    saveStep2SnapshotNow,
     /** 强制刷新草稿列表（如后台挂起后让「进行中」状态立刻出现）。 */
     refresh: refreshList,
   };
diff --git a/frontend-web/src/roundtable-planning/pages/RoundtablePlanningPage.tsx b/frontend-web/src/roundtable-planning/pages/RoundtablePlanningPage.tsx
index c0b01eb6..7ad4ee81 100644
--- a/frontend-web/src/roundtable-planning/pages/RoundtablePlanningPage.tsx
+++ b/frontend-web/src/roundtable-planning/pages/RoundtablePlanningPage.tsx
@@ -1143,6 +1143,26 @@ export default function RoundtablePlanningPage() {
     });
   }, [resetToStep1]);
 
+  // 「挂起中」本地标记：点「后台挂起」后立刻跳转，建作业是 fire-and-forget——从点击到作业
+  // 真正建好的窗口里，用它让历史记录显示「挂起中」（作业建好后转 queued/running，由后端状态接管）。
+  // resetToStep1 不卸载本页，故该 state 跨「跳转」存活。
+  const [pendingSuspendIds, setPendingSuspendIds] = useState<Set<string>>(() => new Set());
+  const markPendingSuspend = useCallback((draftId: string) => {
+    setPendingSuspendIds((prev) => {
+      const next = new Set(prev);
+      next.add(draftId);
+      return next;
+    });
+  }, []);
+  const clearPendingSuspend = useCallback((draftId: string) => {
+    setPendingSuspendIds((prev) => {
+      if (!prev.has(draftId)) return prev;
+      const next = new Set(prev);
+      next.delete(draftId);
+      return next;
+    });
+  }, []);
+
   // 「后台挂起」核心（手动按钮 / 离开页面自动挂起共用）：幂等保证草稿存在 →
   // 同草稿已有在跑作业则跳过 → 启动后台作业。后端另有「同草稿未终态作业幂等返回」
   // 兜底（覆盖跨标签页 / beforeunload 与本函数并发触发），绝不重复建作业。
@@ -1191,27 +1211,87 @@ export default function RoundtablePlanningPage() {
     buildInitialIntent,
   ]);
 
-  // 「后台挂起」按钮：把当前研讨交给后端自动跑完 Step2 研讨 + Step3（总结报告 + 结果绘制）。
-  // 挂起后页面像「新建任务」一样重置回第一步（后台作业继续在历史记录里跑）。
-  const handleBackgroundSuspend = useCallback(async () => {
-    if (step2.selectedAgents.length === 0) {
+  // 「后台挂起」按钮（2026-06-16 改「直接跳转」）：点击后**立刻**跳回第一步并提示，挂起动作
+  // （存稿 → 去重 → 建后台作业）改为 fire-and-forget 在后台完成，不再等接口返回才跳转。
+  // 历史记录在作业真正建好前显示「挂起中」（pendingSuspendIds），建好后由后端状态（queued→
+  // running）接管。⚠️ resetToStep1 会清空 step2 等 live 状态，故挂起入参 + step2 快照必须在
+  // 跳转前同步冻结（getStep2SnapshotWithRuns 返回的是不可变快照对象，reset 替换 state 不影响它）。
+  const handleBackgroundSuspend = useCallback(() => {
+    const selected = step2.selectedAgents;
+    if (selected.length === 0) {
       triggerToast("⚠️ 请先选择研讨席位再后台挂起");
       return;
     }
-    let jobId: string | null = null;
-    try {
-      jobId = await startBackgroundJob();
-    } catch (err) {
-      triggerToast(`❌ 后台挂起失败：${(err as Error).message}`);
-      return;
-    }
-    // 注册到全工作区通知器，作业完成时**跨页**弹「达成共识同款」提示。
-    if (jobId) watchJob(jobId);
-    // 重置回第一步（同新建任务）；后台作业在历史记录里继续跑、可点开看进度。
+    // —— 跳转前同步冻结所有需要的输入（reset 之后这些 live 状态都会被清空）——
+    const step2Snapshot = getStep2SnapshotWithRuns();
+    const agents = selected.map((a) => ({ agentId: a.agent_id, name: a.name }));
+    const intent = buildInitialIntent();
+    const model = selectedModel || undefined;
+    const seatModeSnap = seatMode;
+    const mode = orchestrationMode;
+    const chainSnap = chainMeta;
+    const planSnap = mode === "dag" ? orchestrationPlan : null;
+    const draftId = drafts.currentDraftId;
+
+    // 立刻：标记「挂起中」+ 跳回第一步 + 提示（不等接口）。
+    if (draftId) markPendingSuspend(draftId);
     resetToStep1();
     triggerToast("🚀 已转入后台，任务会自动跑完研讨、总结报告与结果绘制；可在「历史记录」查看进度");
-    void drafts.refresh();
-  }, [step2, drafts, startBackgroundJob, resetToStep1, triggerToast, watchJob]);
+
+    // 后台 fire-and-forget：用冻结快照存稿 → 去重 → 建作业 → 注册完成通知。失败只提示、不挡跳转。
+    void (async () => {
+      let watchId: string | null = null;
+      try {
+        const savedId = (await drafts.saveStep2SnapshotNow(step2Snapshot, 2)) ?? draftId ?? null;
+        // 防重复：该草稿已有未终态作业就不再新建（同 startBackgroundJob 的客户端去重，
+        // 后端另有「同草稿未终态作业幂等返回」兜底）。
+        const existing = savedId ? await getJobByDraft(savedId).catch(() => null) : null;
+        const alreadyRunning =
+          existing &&
+          (existing.status === "running" ||
+            existing.status === "awaiting_input" ||
+            existing.status === "queued");
+        if (alreadyRunning) {
+          watchId = existing?.id ?? null;
+        } else {
+          const job = await startJob({
+            draftId: savedId ?? undefined,
+            intent,
+            agents,
+            model,
+            seatMode: seatModeSnap,
+            orchestrationMode: mode,
+            chain: chainSnap,
+            orchestrationPlan: planSnap,
+          });
+          watchId = job.id;
+        }
+      } catch (err) {
+        triggerToast(`❌ 后台挂起失败：${(err as Error).message}`);
+      } finally {
+        // 注册到全工作区通知器，作业完成时**跨页**弹「达成共识同款」提示。
+        if (watchId) watchJob(watchId);
+        // 清「挂起中」本地标记：此后历史记录的状态由后端作业（queued→running）权威接管。
+        if (draftId) clearPendingSuspend(draftId);
+        void drafts.refresh();
+      }
+    })();
+  }, [
+    step2,
+    drafts,
+    getStep2SnapshotWithRuns,
+    buildInitialIntent,
+    selectedModel,
+    seatMode,
+    orchestrationMode,
+    orchestrationPlan,
+    chainMeta,
+    resetToStep1,
+    triggerToast,
+    watchJob,
+    markPendingSuspend,
+    clearPendingSuspend,
+  ]);
 
   // ── 自动后台挂起：研讨进行中离开本页面（路由卸载）或刷新 / 关标签页时，任务自动转后台 ──
   // 最新状态进 ref：卸载 cleanup 与 beforeunload 都是「拍快照时刻」的同步回调，
@@ -1427,6 +1507,7 @@ export default function RoundtablePlanningPage() {
             setIsOpen={drafts.setIsHistoryOpen}
             drafts={drafts.drafts}
             currentDraftId={drafts.currentDraftId}
+            pendingSuspendIds={pendingSuspendIds}
             renamingDraftId={drafts.renamingDraftId}
             renameDraftValue={drafts.renameDraftValue}
             setRenameDraftValue={drafts.setRenameDraftValue}
@@ -1752,6 +1833,8 @@ interface HistoryDropdownProps {
   setIsOpen: React.Dispatch<React.SetStateAction<boolean>>;
   drafts: ReturnType<typeof useDraftPersistence>["drafts"];
   currentDraftId: string | null;
+  /** 正在「挂起中」（已点后台挂起、作业尚未建好）的草稿 id 集合 —— 历史记录显示「挂起中」。 */
+  pendingSuspendIds: Set<string>;
   renamingDraftId: string | null;
   renameDraftValue: string;
   setRenameDraftValue: React.Dispatch<React.SetStateAction<string>>;
@@ -1774,6 +1857,7 @@ function HistoryDropdown(props: HistoryDropdownProps) {
     setIsOpen,
     drafts,
     currentDraftId,
+    pendingSuspendIds,
     renamingDraftId,
     renameDraftValue,
     setRenameDraftValue,
@@ -1844,6 +1928,8 @@ function HistoryDropdown(props: HistoryDropdownProps) {
               {drafts.map((d) => {
                 const isCurrent = d.id === currentDraftId;
                 const isRenaming = d.id === renamingDraftId;
+                // 「挂起中」：已点后台挂起、作业还没建好（本地标记），或作业刚建好仍排队（queued）。
+                const isSuspending = pendingSuspendIds.has(d.id) || d.status === "queued";
                 // 「进行中 / 待澄清」且有 jobId → 点击打开流程图弹窗，而非加载草稿。
                 const isActiveJob =
                   (d.status === "running" || d.status === "awaiting_input") && !!d.jobId;
@@ -1881,7 +1967,12 @@ function HistoryDropdown(props: HistoryDropdownProps) {
                               {d.name}
                             </span>
                             <span className="text-[10px] text-slate-400 mt-0.5 flex items-center gap-1 font-mono">
-                              {d.status === "running" ? (
+                              {isSuspending ? (
+                                <span className="inline-flex items-center gap-1 text-sky-500 font-semibold">
+                                  <Loader2 className="w-3 h-3 animate-spin" />
+                                  挂起中
+                                </span>
+                              ) : d.status === "running" ? (
                                 <span className="inline-flex items-center gap-1 text-primary font-semibold">
                                   <Loader2 className="w-3 h-3 animate-spin" />
                                   进行中
diff --git a/offline-backend-20260512/backend/app/gateway/roundtable_inprocess_gateway.py b/offline-backend-20260512/backend/app/gateway/roundtable_inprocess_gateway.py
index b0c494d4..5eee3c38 100644
--- a/offline-backend-20260512/backend/app/gateway/roundtable_inprocess_gateway.py
+++ b/offline-backend-20260512/backend/app/gateway/roundtable_inprocess_gateway.py
@@ -14,8 +14,10 @@
 
 from __future__ import annotations
 
+import asyncio
 import json
 import logging
+import os
 import uuid
 from datetime import UTC, datetime
 from types import SimpleNamespace
@@ -51,6 +53,33 @@ _SUMMARY_FILE = "方案总结报告.md"
 # 这里漏禁了 web_search，导致后台席位 / 报告能联网搜索而前端不能，现由 policy 统一。
 
 
+# ── 后台作业「按用户限流」（2026-06-16）─────────────────────────────────────────
+# 痛点：进 2+ 个后台挂起后，整个网关变慢——建立智能体连接、等待模型响应都慢。
+# 根因：① checkpointer 用单连接 aiosqlite（AsyncSqliteSaver，内部一把 asyncio.Lock 串
+# 行全进程的 checkpoint 读写）；② 后台作业无并发上限——每个作业进程内驱动真实 leader/
+# seat/report run，DAG/recommend 的 stage 还用 asyncio.gather 把席位真并行铺开。多个后台
+# 作业同时把几十个 agent run 压到同一把 SQLite 锁 + 同一个离线模型端点上，前台实时会商
+# 的 checkpoint 读写 / 模型调用只能排队 → 「干啥都慢」。
+#
+# 解法：给**每个用户**一个并发额度（默认 2），后台每个 agent 轮次（leader/seat/report/
+# summary/dashboard）真正开跑前先取一个额度，跑完释放。同一用户的所有后台作业共享这 2 个
+# 额度——随便挂多少个作业都能立刻建好，但同一瞬间真正在打模型/写 checkpoint 的后台轮次
+# 至多 2 个，不会把共享资源打满。**前台实时路径（multi_agent._leader_run/_special_run、
+# 普通聊天）不经过本文件的 `_run_agent_turn`，永不取这个额度，故前台不受限。**
+_BACKGROUND_SLOTS_PER_USER = max(1, int(os.getenv("ROUNDTABLE_BACKGROUND_SLOTS_PER_USER", "2") or "2"))
+# 按 user_id 缓存 Semaphore。网关全程跑在同一个事件循环上，普通 dict 取/建即可（无跨线程）。
+_user_background_sems: dict[str, asyncio.Semaphore] = {}
+
+
+def _user_background_semaphore(user_id: str) -> asyncio.Semaphore:
+    """取该用户的后台并发额度信号量（懒建，每用户一把）。"""
+    sem = _user_background_sems.get(user_id)
+    if sem is None:
+        sem = asyncio.Semaphore(_BACKGROUND_SLOTS_PER_USER)
+        _user_background_sems[user_id] = sem
+    return sem
+
+
 # ── 进程内底层操作 ───────────────────────────────────────────────────────────
 
 
@@ -156,6 +185,7 @@ def _tool_calls_since_last_human(messages: list[dict[str, Any]]) -> list[dict[st
 async def _run_agent_turn(
     app,
     *,
+    user_id: str,
     thread_id: str,
     agent_name: str,
     message: str,
@@ -170,6 +200,10 @@ async def _run_agent_turn(
 
     镜像 multi_agent ``_run_payload`` 的 body 形状：agent 通过 ``context.agent_name`` 选，
     assistant_id 固定 lead_agent；关闭 memory recall / subagent / plan mode（临时研讨任务）。
+
+    并发受**该用户的后台额度**约束（``_user_background_semaphore``）：真正开跑（start_run +
+    await run task）前先取额度，跑完释放。同一用户的所有后台作业共享这几个额度，避免 2+
+    后台挂起把单连接 SQLite checkpointer / 离线模型打满、拖垮前台实时会商。
     """
     from app.gateway.routers.thread_runs import RunCreateRequest
     from app.gateway.services import start_run
@@ -201,9 +235,12 @@ async def _run_agent_turn(
     # 虚拟 request：带上空 ``state``（auth 中间件在真请求上会写 state.user / state.auth；
     # 进程内没有这层，下游凡是读 request.state.X 的代码必须能拿到 None 而不是 AttributeError）。
     request = SimpleNamespace(app=app, state=SimpleNamespace(user=None, auth=None))
-    record = await start_run(body, thread_id, request)  # type: ignore[arg-type]
-    if getattr(record, "task", None) is not None:
-        await record.task  # 阻塞到这轮 agent 跑完，checkpoint 已落盘
+    # 取该用户的后台额度后再开跑：start_run + await task 是最重的部分（建 agent、调模型、
+    # 落 checkpoint）。额度满时本轮在此 await 排队，等同用户某个在跑轮次释放额度后再开。
+    async with _user_background_semaphore(user_id):
+        record = await start_run(body, thread_id, request)  # type: ignore[arg-type]
+        if getattr(record, "task", None) is not None:
+            await record.task  # 阻塞到这轮 agent 跑完，checkpoint 已落盘
     return await _read_thread_messages(app, thread_id)
 
 
@@ -385,6 +422,7 @@ class InProcessRoundtableGateway:
         policy = roundtable_run_policy("leader")
         messages = await _run_agent_turn(
             self._app,
+            user_id=self._user_id,
             thread_id=thread_ids[coordinator_name],
             agent_name=coordinator_name,
             message=msg,
@@ -426,6 +464,7 @@ class InProcessRoundtableGateway:
         # 席位模式覆盖 policy 的 thinking/reasoning（None → policy 默认），与前端 _special_run 一致。
         messages = await _run_agent_turn(
             self._app,
+            user_id=self._user_id,
             thread_id=thread_ids[agent_id],
             agent_name=agent_id,
             message=task,
@@ -460,6 +499,7 @@ class InProcessRoundtableGateway:
         for attempt in range(1 + _ARTIFACT_RETRY_LIMIT):
             messages = await _run_agent_turn(
                 self._app,
+                user_id=self._user_id,
                 thread_id=report_thread_id,
                 agent_name=REPORT_AGENT_ID,
                 message=message,
@@ -503,6 +543,7 @@ class InProcessRoundtableGateway:
         for attempt in range(1 + _ARTIFACT_RETRY_LIMIT):
             messages = await _run_agent_turn(
                 self._app,
+                user_id=self._user_id,
                 thread_id=summary_thread_id,
                 agent_name=SUMMARY_AGENT_ID,
                 message=message,
@@ -538,6 +579,7 @@ class InProcessRoundtableGateway:
         for attempt in range(1 + _ARTIFACT_RETRY_LIMIT):
             messages = await _run_agent_turn(
                 self._app,
+                user_id=self._user_id,
                 thread_id=thread_id,
                 agent_name=REPORT_AGENT_ID,
                 message=message,

From 8794c51f88deee0a2dd897c9803408c0258f36a7 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E8=91=A3=E5=AE=8F=E5=87=AF?= <hkdong@kland.com.cn>
Date: Tue, 16 Jun 2026 14:18:15 +0800
Subject: [PATCH 9/9] =?UTF-8?q?=E6=8F=90=E4=BA=A4?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 offline-backend-20260512/backend/config.yaml | 24 ++++++++++----------
 1 file changed, 12 insertions(+), 12 deletions(-)

diff --git a/offline-backend-20260512/backend/config.yaml b/offline-backend-20260512/backend/config.yaml
index f2971f53..5c28a316 100644
--- a/offline-backend-20260512/backend/config.yaml
+++ b/offline-backend-20260512/backend/config.yaml
@@ -10,18 +10,18 @@ gateway:
   port: 8001
 
 models:
-  - name: deepseek-chat
-    display_name: DeepSeek Chat
-    use: langchain_openai:ChatOpenAI
-    model: deepseek-chat
-    api_key: sk-35345154983047aba9e12faa86c5f62c
-    base_url: https://api.deepseek.com
-    request_timeout: 600.0
-    max_retries: 2
-    max_tokens: 4096
-    temperature: 1.0
-    supports_vision: false
-    supports_thinking: true
+  # - name: deepseek-chat
+  #   display_name: DeepSeek Chat
+  #   use: langchain_openai:ChatOpenAI
+  #   model: deepseek-chat
+  #   api_key: sk-e1c22b7931dc446f8b67e57878edd105
+  #   base_url: https://api.deepseek.com
+  #   request_timeout: 600.0
+  #   max_retries: 2
+  #   max_tokens: 4096
+  #   temperature: 1.0
+  #   supports_vision: false
+  #   supports_thinking: true
 
   # - name: zai-org/GLM-5-FP8
   #   display_name: zai-org/GLM-5-FP8