代码提交

frontend-web/docs/recommend-agents-progress.md

@@ -0,0 +1,268 @@
+# 智能体推荐弹窗功能 — 开发进度
+
+> 分支: `dhk-1`  
+> 最后更新: 2026-05-25  
+> 状态: **全部完成 ✅**
+
+---
+
+## 功能概述
+
+在圆桌规划流程的 Step 1（任务意图理解）和 Step 2（多智能体圆桌研讨）之间，插入一个**推荐弹窗**：
+
+1. 用户完成任务澄清后，点击"进入多智能体研讨"→ 弹出推荐弹窗
+2. 弹窗拉取 `GET /api/agents` 候选池，调用 `roundtable-recommender` 智能体流式推荐 2-6 个最合适的智能体
+3. 推荐结果默认勾选，用户可取消勾选 / 从候选池增选（总选 2-8 个）
+4. 确认后 Step 2 使用用户最终选定的智能体列表，不再硬编码 6 个圆桌席位
+
+---
+
+## 已完成的所有步骤
+
+### Step 1 — 后端：新建 `roundtable-recommender` 智能体 ✅
+
+**文件：**
+- `offline-backend-20260512/backend/.deer-flow/agents/roundtable-recommender/config.yaml`
+- `offline-backend-20260512/backend/.deer-flow/agents/roundtable-recommender/SOUL.md`
+
+**要点：**
+- `id: roundtable-recommender`，`name: 智能体推荐`
+- SOUL.md 规定输出协议：自然语言推理 + `[RECOMMEND_READY]` 标记 + ` ```json {"picks": [...]} ``` ` 块
+- 每个 pick: `{"agent_id": "...", "reason": "..."}`，agent_id 必须来自候选池
+- 约束：2 ≤ picks ≤ 6，禁止使用任何工具
+
+---
+
+### Step 2 — 后端：新建 `/api/recommend/stream` 接口 ✅
+
+**文件：**
+- `offline-backend-20260512/backend/app/gateway/routers/recommend.py`（新建）
+- `offline-backend-20260512/backend/app/gateway/app.py`（注册路由）
+
+**接口规格：**
+```
+POST /api/recommend/stream
+Content-Type: application/json
+
+Body:
+{
+  "intent": { "objective": "...", "constraints": [...], "assumptions": [...] },
+  "candidates": [{ "agent_id": "...", "name": "...", "description": "..." }, ...],
+  "model": "deepseek-chat"   // 可选
+}
+
+SSE 流：透传上游 LangGraph frames + 最终状态帧
+最终帧: { "status": "done"|"asking"|"error", "content": "...", "picks": [...] | null }
+```
+
+**关键实现：**
+- 无 `/init`，每次请求内部创建新 thread（stateless）
+- `_validate_picks()`：对比候选池 id、去重、夹断到 2-6，< 2 返回 `[]`
+- `_build_prompt()`：将 intent + 候选池渲染为单条 user message
+- `_RECOMMEND_READY_PATTERN`：正则提取 `[RECOMMEND_READY]` 后的 JSON 块
+- `httpx.AsyncClient(trust_env=False)`：防止 loopback 被代理拦截
+- 复用 `multi_agent.py` 的 `_auth_headers`, `_create_thread`, `_stream_upstream` 等
+
+**app.py 注册：**
+```python
+from app.gateway.routers import recommend
+app.include_router(recommend.router)
+```
+
+---
+
+### Step 3 — 前端：`recommend.ts` API 客户端 ✅
+
+**文件：** `src/roundtable-planning/api/recommend.ts`
+
+**导出：**
+```typescript
+// 类型
+export interface RecommendPick { agent_id: string; reason: string }
+export interface StreamRecommendResult { status: "done"|"asking"|"error"; content: string; picks: RecommendPick[]|null }
+export type RecommendStreamPhase = "connecting"|"connected"|"streaming"
+export interface RecommendCandidate { agent_id: string; name: string; description: string }
+
+// 候选池过滤（排除规则集中管理）
+export function filterRecommendCandidates(agents: Agent[]): RecommendCandidate[]
+
+// 流式推荐
+export async function streamRecommend(req: StreamRecommendRequest): Promise<StreamRecommendResult>
+```
+
+**排除规则（`filterRecommendCandidates`）：**
+- `roundtable-intent` — Step 1 澄清专员
+- `roundtable-recommender` — 推荐者自身
+- 前缀 `roundtable-coordinator-` — 每 session 生成的一次性协调器
+
+---
+
+### Step 4 — 前端：`RecommendAgentsDialog.tsx` 弹窗组件 ✅
+
+**文件：** `src/roundtable-planning/components/RecommendAgentsDialog.tsx`
+
+**Props：**
+```typescript
+export interface RecommendAgentsDialogProps {
+  open: boolean;
+  intent: IntentSummary | null;
+  model?: string;
+  onCancel: () => void;
+  onConfirm: (agentIds: string[]) => void;
+}
+```
+
+**内部阶段机器：**
+```
+"idle" → "loading_agents" → "recommending" → "ready"
+                                           ↘ "fallback"  (推荐失败，兜底)
+                                           ↘ "error"
+```
+
+**UI 分区：**
+- **推荐参会**（默认全选）：推荐智能体卡片，带 reason 说明
+- **其他候选**（默认未选）：可增选，补足 8 个上限
+
+**选择约束：** min 2，max 8；确认按钮在范围外 disabled
+
+**安全机制：**
+- `generationRef` 计数器防止 stale async 污染状态
+- `AbortController` 取消在途推荐请求
+- `intersectInOrder()`：前端二次验证推荐 id 在候选池中存在
+- 兜底列表 `FALLBACK_AGENT_IDS`：6 个内置 roundtable-* 智能体
+
+**确认顺序：** 推荐且已选 → 用户增选（按交互顺序）
+
+---
+
+### Step 5 — 前端：`RoundtablePlanningPage.tsx` 动态 Agent 角色 ✅
+
+**删除：** `ROUNDTABLE_ROLE_TEMPLATES`, `makeRoles`, `ROUNDTABLE_AGENT_IDS`, `AGENT_NAME_TO_ROLE`（全部硬编码常量）
+
+**新增模块级：**
+```typescript
+const COMMANDER_SEAT_ID = 1
+type SelectedAgent = { agent_id: string; name: string }
+const BUILTIN_AGENT_META: Record<string, { avatarType: string; displayName: string }>
+const FALLBACK_AVATAR_TYPES = ["blue","purple","emerald","amber","pink"]
+function getAvatarTypeFor(agentId, index): string
+function getDisplayNameFor(agent: SelectedAgent): string
+function buildRoleTemplates(agents): ReadonlyArray<Omit<RoleEntry,"status">>
+function makeRolesFromTemplates(templates, defaultStatus, overrides?): RoleEntry[]
+function buildAgentNameToRole(agents): Record<string, {speakerId,avatarType,displayName}>
+const DEFAULT_SELECTED_AGENTS: SelectedAgent[]  // 原 6 个内置席位
+```
+
+**新增 state + useMemo：**
+```typescript
+const [selectedAgents, setSelectedAgents] = useState<SelectedAgent[]>(DEFAULT_SELECTED_AGENTS);
+const roleTemplates    = useMemo(() => buildRoleTemplates(selectedAgents),    [selectedAgents]);
+const agentNameToRole  = useMemo(() => buildAgentNameToRole(selectedAgents),  [selectedAgents]);
+const roundtableAgentIds = useMemo(() => selectedAgents.map(a => a.agent_id), [selectedAgents]);
+```
+
+**替换调用点：**
+| 替换前 | 替换后 |
+|---|---|
+| `AGENT_NAME_TO_ROLE[name]` | `agentNameToRole[name]` |
+| `ROUNDTABLE_AGENT_IDS` | `roundtableAgentIds` |
+| `makeRoles(status, ...)` | `makeRolesFromTemplates(roleTemplates, status, ...)` |
+| 角色面板 `role.id === 2 ? <RobotAvatar ...>` 链 | `getAvatar(agentNameToRole[role.agentId]?.avatarType ?? "blue", "h-7 w-7")` |
+| `HighFidelityReport seatNameMap` 用 `AGENT_NAME_TO_ROLE` | 用 `agentNameToRole` |
+| useEffect 依赖 | 加入 `roleTemplates` |
+| dispatch useCallback 依赖 | 加入 `agentNameToRole` |
+
+---
+
+### Step 6 — 前端：弹窗接入页面 ✅
+
+**新增 import：**
+```typescript
+import { RecommendAgentsDialog } from "@/roundtable-planning/components/RecommendAgentsDialog";
+```
+
+**新增 state：**
+```typescript
+const [isRecommendDialogOpen, setIsRecommendDialogOpen] = useState<boolean>(false);
+```
+
+**新增 callback `handleConfirmAgents`：**
+```typescript
+const handleConfirmAgents = useCallback((agentIds: string[]) => {
+  setSelectedAgents(
+    agentIds.map(id => ({
+      agent_id: id,
+      name: BUILTIN_AGENT_META[id]?.displayName ?? id,
+    })),
+  );
+  step2InitRef.current = false;   // 重置 Step 2，让新 coordinator 按新名单初始化
+  setIsRecommendDialogOpen(false);
+  setCurrentStep(2);
+  setShowToast("成功确认！已切换至第二阶段圆桌研讨系统");
+}, []);
+```
+
+**两个"进入多智能体研讨"按钮改为：**
+```typescript
+onClick={() => setIsRecommendDialogOpen(true)}
+```
+
+**JSX 渲染（在 showToast 之前）：**
+```tsx
+<RecommendAgentsDialog
+  open={isRecommendDialogOpen}
+  intent={intentReady}
+  model={DEFAULT_MODEL}
+  onCancel={() => setIsRecommendDialogOpen(false)}
+  onConfirm={handleConfirmAgents}
+/>
+```
+
+---
+
+## 文件清单（本次改动）
+
+### 后端（新建）
+| 文件 | 说明 |
+|---|---|
+| `.deer-flow/agents/roundtable-recommender/config.yaml` | 推荐智能体元数据 |
+| `.deer-flow/agents/roundtable-recommender/SOUL.md` | 推荐智能体系统提示 |
+| `app/gateway/routers/recommend.py` | `/api/recommend/stream` 接口 |
+
+### 后端（修改）
+| 文件 | 说明 |
+|---|---|
+| `app/gateway/app.py` | 注册 `recommend.router` |
+
+### 前端（新建）
+| 文件 | 说明 |
+|---|---|
+| `src/roundtable-planning/api/recommend.ts` | 前端 SSE 客户端 |
+| `src/roundtable-planning/components/RecommendAgentsDialog.tsx` | 推荐弹窗组件 |
+
+### 前端（修改）
+| 文件 | 说明 |
+|---|---|
+| `src/roundtable-planning/pages/RoundtablePlanningPage.tsx` | 动态 Agent 角色 + 弹窗接入 |
+
+---
+
+## 下次开始前检查项
+
+1. **后端是否已重启** — `recommend.router` 注册需重启 FastAPI（`make dev` 或 `./scripts/restart-all.sh`）
+2. **`roundtable-recommender` 智能体已注册** — 访问 `GET /api/agents` 确认列表里有 `roundtable-recommender`
+3. **typecheck 干净** — `pnpm typecheck 2>&1 | grep roundtable-planning` 无输出
+4. **弹窗流程端到端测试**：
+   - Step 1 完成意图澄清 → `intentReady` 不为 null
+   - 点击"进入多智能体研讨"→ 弹窗出现，开始加载智能体列表
+   - 弹窗展示推荐结果（流式文字 + 卡片列表）
+   - 调整勾选 → 确认 → Step 2 用所选智能体初始化
+
+---
+
+## 潜在后续工作（未列入本次范围）
+
+- 弹窗内显示每个候选智能体的完整 description（目前从 `/api/agents` 取，已有字段）
+- 推荐失败时的 Toast 提示（目前静默兜底）
+- 推荐理由（reason）在 Step 2 角色面板中展示
+- `HighFidelityReport` 的 `avatarType` 字段按新选角色动态渲染

frontend-web/src/roundtable-planning/pages/RoundtablePlanningPage.tsx

@@ -1,4 +1,4 @@
-import React, { useState, useEffect, useRef, useCallback } from "react";
+import React, { useState, useEffect, useRef, useCallback, useMemo } from "react";
 import ReactMarkdown from "react-markdown";
 import remarkGfm from "remark-gfm";
 import { HighFidelityReport } from "@/roundtable-planning/components/HighFidelityReport";
@@ -21,6 +21,7 @@ import {
   type IntentStreamPhase,
   type IntentSummary,
 } from "@/roundtable-planning/api/intent";
+import { RecommendAgentsDialog } from "@/roundtable-planning/components/RecommendAgentsDialog";
 import {
   deleteDraft,
   deriveDraftName,
@@ -159,42 +160,111 @@ type RoleEntry = {
   status: string;
 };
 
-const ROUNDTABLE_ROLE_TEMPLATES: ReadonlyArray<Omit<RoleEntry, "status">> = [
+/** Stable seat id reserved for the human commander row. Agent seats start at 2. */
-  { id: 1, name: "指挥员 / 主持人", type: "human" },
+const COMMANDER_SEAT_ID = 1;
-  { id: 2, name: "总结陈述", agentId: "roundtable-summary", type: "agent" },
-  { id: 3, name: "情报收集", agentId: "roundtable-intelligence", type: "agent" },
-  { id: 4, name: "环境评估", agentId: "roundtable-environment", type: "agent" },
-  { id: 5, name: "方案设计", agentId: "roundtable-solution-design", type: "agent" },
-  { id: 6, name: "风险审查", agentId: "roundtable-risk-review", type: "agent" },
-  { id: 7, name: "执行规划", agentId: "roundtable-execution-plan", type: "agent" },
-  { id: 8, name: "总结陈述", agentId: "roundtable-summary", type: "agent" },
-];
 
-function makeRoles(defaultStatus: string, overrides: Record<number, string> = {}): RoleEntry[] {
+/** One entry per agent the user picked for this round of Step 2. */
-  return ROUNDTABLE_ROLE_TEMPLATES.map((t) => ({
+type SelectedAgent = { agent_id: string; name: string };
+
+/**
+ * Display metadata for the 6 built-in roundtable seats — kept so the
+ * familiar emoji/color/display name survives when the user keeps the
+ * defaults. Custom agents fall through to {@link FALLBACK_AVATAR_TYPES}.
+ */
+const BUILTIN_AGENT_META: Record<string, { avatarType: string; displayName: string }> = {
+  "roundtable-intelligence": { avatarType: "purple", displayName: "情报收集" },
+  "roundtable-environment": { avatarType: "emerald", displayName: "环境评估" },
+  "roundtable-solution-design": { avatarType: "human", displayName: "方案设计" },
+  "roundtable-risk-review": { avatarType: "amber", displayName: "风险审查" },
+  "roundtable-execution-plan": { avatarType: "pink", displayName: "执行规划" },
+  "roundtable-summary": { avatarType: "blue", displayName: "总结陈述" },
+};
+
+/**
+ * Visual rotation for non-built-in agents. Each entry must be a key
+ * recognized by {@link getAvatar} — anything else falls back to the plain
+ * `RobotAvatar`.
+ */
+const FALLBACK_AVATAR_TYPES = ["blue", "purple", "emerald", "amber", "pink"] as const;
+
+function getAvatarTypeFor(agentId: string, index: number): string {
+  return (
+    BUILTIN_AGENT_META[agentId]?.avatarType ??
+    FALLBACK_AVATAR_TYPES[index % FALLBACK_AVATAR_TYPES.length]
+  );
+}
+
+function getDisplayNameFor(agent: SelectedAgent): string {
+  return BUILTIN_AGENT_META[agent.agent_id]?.displayName ?? agent.name;
+}
+
+/**
+ * Build the seat-template list shown in the right-rail role panel.
+ * The commander always sits at id=1; agents take id=2..N+1, in the order
+ * the user picked them so the panel mirrors the recommend dialog.
+ */
+function buildRoleTemplates(
+  agents: SelectedAgent[],
+): ReadonlyArray<Omit<RoleEntry, "status">> {
+  const out: Omit<RoleEntry, "status">[] = [
+    { id: COMMANDER_SEAT_ID, name: "指挥员 / 主持人", type: "human" },
+  ];
+  agents.forEach((a, i) => {
+    out.push({
+      id: i + 2,
+      name: getDisplayNameFor(a),
+      agentId: a.agent_id,
+      type: "agent",
+    });
+  });
+  return out;
+}
+
+function makeRolesFromTemplates(
+  templates: ReadonlyArray<Omit<RoleEntry, "status">>,
+  defaultStatus: string,
+  overrides: Record<number, string> = {},
+): RoleEntry[] {
+  return templates.map((t) => ({
     ...t,
     status: overrides[t.id] ?? defaultStatus,
   }));
 }
 
-// Subset of role templates that map to real backend agents (excludes the
+/**
-// human commander seat) — used as the `agent_names` payload on init.
+ * Map `agent_name` (lowercase) → bubble meta. Used while streaming output so
-const ROUNDTABLE_AGENT_IDS = Array.from(
+ * the right speaker bubble lights up and the dialogue card picks the right
-  new Set(
+ * avatar. Speaker ids stay in sync with {@link buildRoleTemplates} (agents
-    ROUNDTABLE_ROLE_TEMPLATES.filter((r) => r.agentId).map((r) => r.agentId as string),
+ * are id 2..N+1, indexed by selection order).
-  ),
+ */
-);
+function buildAgentNameToRole(
+  agents: SelectedAgent[],
+): Record<string, { speakerId: number; avatarType: string; displayName: string }> {
+  const out: Record<string, { speakerId: number; avatarType: string; displayName: string }> = {};
+  agents.forEach((a, i) => {
+    out[a.agent_id] = {
+      speakerId: i + 2,
+      avatarType: getAvatarTypeFor(a.agent_id, i),
+      displayName: getDisplayNameFor(a),
+    };
+  });
+  return out;
+}
 
-// agent_name (lowercase) → bubble meta. Used when streaming output, so we can
+/**
-// light up the right speaker bubble and pick an avatar for the dialogue card.
+ * Default Step-2 roster when the user hasn't picked anything yet (e.g. the
-const AGENT_NAME_TO_ROLE: Record<string, { speakerId: number; avatarType: string; displayName: string }> = {
+ * recommend dialog was skipped or hasn't run). Matches the original
-  "roundtable-summary": { speakerId: 2, avatarType: "blue", displayName: "总结陈述" },
+ * hand-coded 6 roundtable seats so the page renders identically until the
-  "roundtable-intelligence": { speakerId: 3, avatarType: "purple", displayName: "情报收集" },
+ * recommend flow overrides it.
-  "roundtable-environment": { speakerId: 4, avatarType: "emerald", displayName: "环境评估" },
+ */
-  "roundtable-solution-design": { speakerId: 5, avatarType: "human", displayName: "方案设计" },
+const DEFAULT_SELECTED_AGENTS: SelectedAgent[] = [
-  "roundtable-risk-review": { speakerId: 6, avatarType: "amber", displayName: "风险审查" },
+  { agent_id: "roundtable-summary", name: "总结陈述" },
-  "roundtable-execution-plan": { speakerId: 7, avatarType: "pink", displayName: "执行规划" },
+  { agent_id: "roundtable-intelligence", name: "情报收集" },
-};
+  { agent_id: "roundtable-environment", name: "环境评估" },
+  { agent_id: "roundtable-solution-design", name: "方案设计" },
+  { agent_id: "roundtable-risk-review", name: "风险审查" },
+  { agent_id: "roundtable-execution-plan", name: "执行规划" },
+];
 
 function nowTime() {
   return new Date().toLocaleTimeString("zh-CN", { hour12: false });
@@ -334,17 +404,27 @@ export default function RoundtablePlanningPage() {
     }
   ]);
 
-  // Roles in roundtable state (updates as we switch steps!)
+  // The user's chosen Step-2 participants. Defaults to the 6 built-in
-  const [roles, setRoles] = useState<RoleEntry[]>([
+  // roundtable seats so behavior is unchanged before the recommend dialog
-    { id: 1, name: "指挥员 / 主持人", type: "human", status: "等待激活" },
+  // overrides it. Step 6 will replace this via the dialog's onConfirm.
-    { id: 2, name: "总结陈述", agentId: "roundtable-summary", type: "agent", status: "等待激活" },
+  const [selectedAgents, setSelectedAgents] = useState<SelectedAgent[]>(DEFAULT_SELECTED_AGENTS);
-    { id: 3, name: "情报收集", agentId: "roundtable-intelligence", type: "agent", status: "等待激活" },
+
-    { id: 4, name: "环境评估", agentId: "roundtable-environment", type: "agent", status: "等待激活" },
+  // Derived: template list for the seat panel + dispatch-time name lookup +
-    { id: 5, name: "方案设计", agentId: "roundtable-solution-design", type: "agent", status: "等待激活" },
+  // the `agent_names` payload for /api/multi-agent/init. All three regenerate
-    { id: 6, name: "风险审查", agentId: "roundtable-risk-review", type: "agent", status: "等待激活" },
+  // together whenever the selection changes, so seat ids stay in sync.
-    { id: 7, name: "执行规划", agentId: "roundtable-execution-plan", type: "agent", status: "等待激活" },
+  const roleTemplates = useMemo(() => buildRoleTemplates(selectedAgents), [selectedAgents]);
-    { id: 8, name: "总结陈述", agentId: "roundtable-summary", type: "agent", status: "等待陈述" },
+  const agentNameToRole = useMemo(() => buildAgentNameToRole(selectedAgents), [selectedAgents]);
-  ]);
+  const roundtableAgentIds = useMemo(() => selectedAgents.map((a) => a.agent_id), [selectedAgents]);
+
+  // Roles in roundtable state (updates as we switch steps!). Seeded from
+  // `roleTemplates`'s initial value (selectedAgents starts at the default
+  // roster), so the panel renders identically to the previous hand-coded list.
+  const [roles, setRoles] = useState<RoleEntry[]>(() =>
+    makeRolesFromTemplates(buildRoleTemplates(DEFAULT_SELECTED_AGENTS), "等待激活"),
+  );
+
+  // Dialog: recommend agents before entering Step 2.
+  const [isRecommendDialogOpen, setIsRecommendDialogOpen] = useState<boolean>(false);
 
   // Step 2 dialogue list — populated by the real orchestration loop once the
   // user enters the roundtable. Starts empty so we never flash mock content.
@@ -739,7 +819,7 @@ export default function RoundtablePlanningPage() {
         // when the leader 派活给多个席位。改为顺序后,同一时刻只有一个 SSE 在跑。
         for (const [subName, task] of leaderRes.dispatched) {
           if (orchestrationCancelRef.current) break;
-          const meta = AGENT_NAME_TO_ROLE[subName.toLowerCase()];
+          const meta = agentNameToRole[subName.toLowerCase()];
           if (meta) markSpeakerStreaming(meta.speakerId);
           const perSeatModel = meta ? agentConfigs[meta.speakerId]?.model : undefined;
 
@@ -806,7 +886,23 @@ export default function RoundtablePlanningPage() {
       clearAllStreamingSpeakers();
       setIsOrchestrating(false);
     }
-  }, [appendStreamingDialogue, appendToDialogue, finalizeDialogue, highlightSpeaker, markSpeakerStreaming, markSpeakerIdle, clearAllStreamingSpeakers, isOrchestrating, agentConfigs]);
+  }, [appendStreamingDialogue, appendToDialogue, finalizeDialogue, highlightSpeaker, markSpeakerStreaming, markSpeakerIdle, clearAllStreamingSpeakers, isOrchestrating, agentConfigs, agentNameToRole]);
+
+  // Called by RecommendAgentsDialog onConfirm: update the participant roster,
+  // reset step-2 init so the new coordinator is created for the new roster,
+  // then navigate to Step 2.
+  const handleConfirmAgents = useCallback((agentIds: string[]) => {
+    setSelectedAgents(
+      agentIds.map((id) => ({
+        agent_id: id,
+        name: BUILTIN_AGENT_META[id]?.displayName ?? id,
+      })),
+    );
+    step2InitRef.current = false;
+    setIsRecommendDialogOpen(false);
+    setCurrentStep(2);
+    setShowToast("成功确认！已切换至第二阶段圆桌研讨系统");
+  }, []);
 
   // Public stop handler used by the pause/stop button and the step-2 teardown.
   // Sets the cancel flag and aborts every in-flight stream so the loop can
@@ -849,7 +945,7 @@ export default function RoundtablePlanningPage() {
       const coord = newCoordinatorName();
       try {
         const initRes = await initMultiAgent({
-          agentNames: ROUNDTABLE_AGENT_IDS,
+          agentNames: roundtableAgentIds,
           mainAgentName: coord,
           model: DEFAULT_MODEL,
           // As each sub-agent's thread comes online, flip every role that
@@ -1473,11 +1569,11 @@ export default function RoundtablePlanningPage() {
       // seat to 在线 — assume all seats are online so the panel doesn't get
       // stuck at "正在上线".
       const baseStatus = loadedFromDraftRef.current ? "在线" : "正在上线";
-      setRoles(makeRoles(baseStatus, { 1: "在线" }));
+      setRoles(makeRolesFromTemplates(roleTemplates, baseStatus, { 1: "在线" }));
     } else {
-      setRoles(makeRoles("已完成"));
+      setRoles(makeRolesFromTemplates(roleTemplates, "已完成"));
     }
-  }, [currentStep, intentThreadId, intentInitError]);
+  }, [currentStep, intentThreadId, intentInitError, roleTemplates]);
 
   // Inject user input to the roundtable room — sends it as a new leader call.
   const handleInjectToRoundtable = () => {
@@ -1544,6 +1640,14 @@ export default function RoundtablePlanningPage() {
     >
       
       {/* Toast Notice Banner */}
+      <RecommendAgentsDialog
+        open={isRecommendDialogOpen}
+        intent={intentReady}
+        model={DEFAULT_MODEL}
+        onCancel={() => setIsRecommendDialogOpen(false)}
+        onConfirm={handleConfirmAgents}
+      />
+
       {showToast && (
         <div className="fixed top-4 left-1/2 transform -translate-x-1/2 z-50 bg-primary text-white text-xs px-4 py-2.5 rounded-xl shadow-lg border border-primary/40 font-medium flex items-center gap-2 animate-bounce">
           <Sparkles className="h-4 w-4 shrink-0" />
@@ -2009,16 +2113,10 @@ export default function RoundtablePlanningPage() {
                         }`}
                       >
                         <div className="flex items-center gap-2.5">
-                          {/* Avatar chooser based on role.id */}
                           {role.type === "human" ? (
                             <Man1Avatar size="h-7 w-7" />
                           ) : (
-                            role.id === 2 ? <RobotAvatar headColor="#f5222d" size="h-7 w-7" /> :
+                            getAvatar(role.agentId ? (agentNameToRole[role.agentId]?.avatarType ?? "blue") : "blue", "h-7 w-7")
-                            role.id === 3 ? <PurpleRobotAvatar size="h-7 w-7" /> :
-                            role.id === 4 ? <EmeraldRobotAvatar size="h-7 w-7" /> :
-                            role.id === 6 ? <AmberRobotAvatar size="h-7 w-7" /> :
-                            role.id === 7 ? <PinkRobotAvatar size="h-7 w-7" /> :
-                            <BlueRobotAvatar size="h-7 w-7" />
                           )}
                           
                           <div className="flex flex-col">
@@ -2146,10 +2244,7 @@ export default function RoundtablePlanningPage() {
                       </div>
                       <button
                         id="confirm-roundtable-btn"
-                        onClick={() => {
+                        onClick={() => setIsRecommendDialogOpen(true)}
-                          setCurrentStep(2);
-                          triggerToast("成功确认！已切换至第二阶段圆桌研讨系统");
-                        }}
                         className="bg-primary hover:bg-[var(--brand-900)] text-white font-bold text-xs px-4 py-2 rounded-xl transition shadow-[0_2px_8px_rgba(245,34,45,0.2)] active:scale-95 shrink-0"
                       >
                         用户确认，进入多智能体研讨
@@ -2435,7 +2530,7 @@ export default function RoundtablePlanningPage() {
                           triggerToast("⚠️ 请先与任务理解智能体完成澄清,生成任务摘要后再进入");
                           return;
                         }
-                        setCurrentStep(2);
+                        setIsRecommendDialogOpen(true);
                       }}
                       disabled={!intentReady}
                       title={intentReady ? undefined : "请先完成任务意图澄清"}
@@ -3123,7 +3218,7 @@ export default function RoundtablePlanningPage() {
                   content: String(d.content ?? ""),
                 }))}
               seatNameMap={Object.fromEntries(
-                Object.entries(AGENT_NAME_TO_ROLE).map(([k, v]) => [k, v.displayName]),
+                Object.entries(agentNameToRole).map(([k, v]) => [k, v.displayName]),
               )}
               onReset={handleStartNewTask}
             />