diff --git a/frontend-web/docs/multi-agent-backend-dev.md b/frontend-web/docs/multi-agent-backend-dev.md
index 4492f3e5..3e5bf522 100644
--- a/frontend-web/docs/multi-agent-backend-dev.md
+++ b/frontend-web/docs/multi-agent-backend-dev.md
@@ -20,6 +20,7 @@
 - **leader 共识 = `status: []` 空数组**；澄清 = `status: "clarification"`（字符串）；派活 = `status: [[name, task], ...]`。这三者**互斥不混用**。
 - **后端 picks clamp 2–6，前端用户可手动加到 8**。
 - **不写死任何具体模型名**：`intent / recommend / multi_agent` 三个 router 的 `model_name = str(payload.get("model") or "")`，空串会被下游 `lead_agent._resolve_model_name()` 自动回退到 `config.yaml` 的 `models[0]`（与 `/api/ai-writing` 行为一致）。内网部署只需配 `config.yaml`，前后端均无需改硬编码。
+- **草稿/推荐历史持久化（2026-06）**：圆桌页的「历史草稿」从浏览器 `localStorage` 迁到**后端 MySQL**（`/api/roundtable-drafts`，按登录用户隔离）；新增**会话级推荐历史**让推荐弹窗回显上次结果。两张表 `roundtable_drafts` / `roundtable_recommend_history`，启动 `create_all` 自动建表，详见 §7A。
 
 ---
 
@@ -73,7 +74,12 @@ offline-backend-20260512/backend/
 │       ├── multi_agent.py              # Step 2:init(SSE) + run/stream + 共享的 loopback 工具
 │       ├── intent.py                   # Step 1:init + stream(import multi_agent 工具)
 │       ├── recommend.py                # Step 1→2:推荐 stream(import multi_agent 工具)
-│       └── artifacts.py                # Step 3:列出/下载/编辑 thread 产出(通用,非圆桌专用)
+│       ├── artifacts.py                # Step 3:列出/下载/编辑 thread 产出(通用,非圆桌专用)
+│       └── roundtable_drafts.py        # 持久化:草稿 CRUD + 推荐历史(§7A,2026-06 新增)
+├── packages/harness/deerflow/persistence/roundtable_drafts/
+│   ├── model.py                        # ORM:RoundtableDraftRow + RoundtableRecommendHistoryRow
+│   ├── sql.py                          # RoundtableDraftRepository(CRUD + 历史)
+│   └── __init__.py                     # make_roundtable_draft_store(sf)
 ├── packages/harness/deerflow/tools/builtins/
 │   └── clarification_utils.py          # resolve_allow_multiple + coerce_bool(Step1/Step2 共享)
 └── .deer-flow/agents/
@@ -110,8 +116,15 @@ offline-backend-20260512/backend/
 | GET | `/api/threads/{thread_id}/artifacts` | 3 | 列出该 thread 的 outputs 文件 |
 | GET | `/api/threads/{thread_id}/artifacts/{path}` | 3 | 预览/下载单个产出（含 `.skill/` 内文件提取） |
 | PATCH | `/api/threads/{thread_id}/artifacts/{path}` | 3 | 编辑文本类产物（白名单后缀） |
+| GET | `/api/roundtable-drafts` | 持久化 | 列出当前用户草稿（轻量元数据） |
+| POST | `/api/roundtable-drafts` | 持久化 | 新建草稿 |
+| GET | `/api/roundtable-drafts/{id}` | 持久化 | 取草稿全量（含 step1/step2 快照） |
+| PUT | `/api/roundtable-drafts/{id}` | 持久化 | 局部更新（title / furthest_step / step1 / step2） |
+| DELETE | `/api/roundtable-drafts/{id}` | 持久化 | 删除草稿 |
+| GET | `/api/roundtable-drafts/{id}/recommendations` | 持久化 | 列出该草稿的推荐历史（倒序） |
+| POST | `/api/roundtable-drafts/{id}/recommendations` | 持久化 | 追加一次推荐记录 |
 
-字段级请求/响应见各 step 的 API 文档；下文只写**后端实现语义**。
+字段级请求/响应见各 step 的 API 文档；下文只写**后端实现语义**。圆桌草稿/推荐历史持久化见 §7A。
 
 ---
 
@@ -589,6 +602,53 @@ Step 3 **没有** 单独的 `multi_agent` 路由；使用既有 artifacts 能力
 
 ---
 
+## 7A. 圆桌草稿与推荐历史持久化（`roundtable_drafts.py`，2026-06）
+
+圆桌规划页的「历史记录」此前只存浏览器 `localStorage`（前端 `utils/drafts.ts`），换设备/清缓存即丢、大对话还会撞 5MB 配额。2026-06 迁到**后端 MySQL**，按登录用户隔离、跨设备保留；并新增**会话级推荐历史**，让推荐弹窗能回显上次推荐结果。
+
+### 7A.1 数据模型（`deerflow.persistence.roundtable_drafts`）
+
+仿 `user_prompts`（最简 CRUD）+ `ai_writing_sessions`（大 JSON 用 `PortableLongText` 存）两个现有模式。两张表：
+
+- **`roundtable_drafts`**（会话主表，一条 = 一次完整圆桌会话）
+  - `id`(PK, str64) / `user_id`(idx) / `title`(512) / `furthest_step`(int, 1|2|3)
+  - `step1` / `step2`：前端 `DraftStep{1,2}Snapshot` 的 JSON，列类型 **`PortableLongText`**（MySQL `LONGTEXT`），避免 Step 2 长对话 + `seatStubMessages` 撞 `TEXT` 64KB 上限
+  - `created_at` / `updated_at`：**`BeijingDateTime`**（项目规范，存北京时间且 tz-aware）
+- **`roundtable_recommend_history`**（推荐历史，按 `draft_id` 归属）
+  - `id`(PK) / `user_id`(idx) / `draft_id`(idx) / `objective`(512) / `status`(32: done|fallback|asking|error) / `model`(128)
+  - `rationale` / `picks`(JSON) / `candidates`(JSON 候选池快照)：均 `PortableLongText`
+  - `created_at`：`BeijingDateTime`
+
+JSON 列在 repo（`sql.py`）里用 `json.dumps/loads` 手动序列化（与 `ai_writing_sessions.transcript` 同法）。两个 Row 已登记进 `persistence/models/__init__.py`，**启动时 `Base.metadata.create_all` 自动建表**——新部署无需手动建表（需 mysql 账号有建库/表权限）。
+
+### 7A.2 路由（`app/gateway/routers/roundtable_drafts.py`，前缀 `/api/roundtable-drafts`）
+
+| 方法 | 路径 | 说明 |
+|------|------|------|
+| GET | `` | `list_drafts` → 轻量元数据列表（不含 step 大 blob），按 `updated_at` 倒序 |
+| POST | `` | `create_draft`，可带客户端 `id`（稳定 id 跨 create+update） |
+| GET | `/{id}` | `get_draft` 全量；非属主 → 404 |
+| PUT | `/{id}` | `update_draft`，`model_dump(exclude_unset=True)` → 只更新客户端实际发的字段 |
+| DELETE | `/{id}` | 204；非属主 → 404 |
+| GET | `/{id}/recommendations` | 推荐历史倒序 |
+| POST | `/{id}/recommendations` | 追加一次推荐（先校验草稿属主，404 拦截越权） |
+
+鉴权/属主：用 `_current_user_id(request)`（登录态取 `request.state.user.id`，无登录回退 `get_effective_user_id()` 的 `"default"`），与 `user_prompts` 一致；所有 repo 读写按 `user_id` 过滤，跨用户访问一律 404。store 在 `deps.py` 接为 `app.state.roundtable_draft_store`（`make_roundtable_draft_store(sf)`），`app.py` `include_router`。
+
+### 7A.3 与前端的契约要点
+
+- **PUT 不带 `title` 时不改标题**：前端自动保存只发 `furthest_step`/`step1`/`step2`，不发 `title`，因此**不会覆盖用户手动重命名**；标题只在 create（派生）与显式重命名（只发 title）时改。
+- **推荐时机**：前端在打开推荐弹窗前先 `ensureDraft` 落地草稿拿到 `draft_id`，推荐完成后 POST 一条历史；重开弹窗 GET 历史回显最近一次，可「重新分析」再 POST。
+- **字段命名**：后端 snake_case（`furthest_step`/`created_at`），前端 `api/drafts.ts` 在边界映射成 camelCase。
+
+### 7A.4 设计取舍
+
+- **整存 JSON 而非拆表**：step1/step2 形状宽且常变（前端 hook 里是 loose 类型），整存 JSON 避免每次改前端快照都要迁库；代价是后端不校验嵌套字段（与 `ai_writing_sessions.transcript` 同取舍）。
+- **推荐历史绑定草稿而非全局**：语义是「这个会话上次推荐了什么」，故按 `draft_id` 归属；这也要求推荐前必有 `draft_id`（前端 `ensureDraft` 保证）。
+- **本地 dev 走 sqlite**：`config.yaml` `database.backend` 本地默认 sqlite，走同一套 SQLAlchemy 代码，`PortableLongText/JSON` 各自降级；生产切 mysql 用同一份代码。
+
+---
+
 ## 8. 内置席位 SOUL 设计要点
 
 每个 `roundtable-*`（除 intent / recommender）遵循统一结构：角色定位、核心职责、输出风格（emoji + 简短结论）、边界（只产出本席位、等总控派活）。
@@ -1195,6 +1255,9 @@ client = httpx.AsyncClient(transport=transport, base_url=...)
 | 推荐路由 | `offline-backend-20260512/backend/app/gateway/routers/recommend.py` |
 | Step 2 路由 | `offline-backend-20260512/backend/app/gateway/routers/multi_agent.py` |
 | Step 3 列出/下载/编辑 | `offline-backend-20260512/backend/app/gateway/routers/artifacts.py` |
+| 草稿/推荐历史路由（§7A） | `offline-backend-20260512/backend/app/gateway/routers/roundtable_drafts.py` |
+| 草稿/推荐历史持久化 | `offline-backend-20260512/backend/packages/harness/deerflow/persistence/roundtable_drafts/` |
+| store 接线 | `offline-backend-20260512/backend/app/gateway/deps.py`（`app.state.roundtable_draft_store`） |
 | 路由注册 | `offline-backend-20260512/backend/app/gateway/app.py` |
 | 澄清单选/多选 | `offline-backend-20260512/backend/packages/harness/deerflow/tools/builtins/clarification_utils.py` |
 | 任务理解 Agent | `offline-backend-20260512/backend/.deer-flow/agents/roundtable-intent/` |
diff --git a/frontend-web/docs/multi-agent-frontend-dev.md b/frontend-web/docs/multi-agent-frontend-dev.md
index 89ee2963..ea5235f7 100644
--- a/frontend-web/docs/multi-agent-frontend-dev.md
+++ b/frontend-web/docs/multi-agent-frontend-dev.md
@@ -14,6 +14,7 @@
 - **步骤折叠卡与主聊天对齐**：思考过程 / 工具调用展示走 `MessageStepsCard`,**always-show-last-tool + 折叠 "查看其他 N 个步骤"** 的视觉与主聊天 `chats/new` 的 `MessageGroup` 一致。工具图标 / 文案映射在 `lib/step-display.tsx`,新增工具时与 `src/core/i18n/locales/zh-CN.ts` 的 `toolCalls` 字段对齐。
 - **`<think>` 标签自动提取为「思考过程」**：`lib/reasoning.ts::splitInlineReasoning(text)` 把 `<think>...</think>`(含未闭合的流式中段)抽离成 `reasoning`,正文走 markdown;由 `MessageBubble.tsx::ReasoningBlock` 渲染一个折叠卡。Step 1 / Step 2 气泡和推荐弹窗的 rationale 都走这一套,与主聊天 `Reasoning` 元素的视觉对齐。
 - **不使用 LangGraph SDK**：所有 API 都是自研的 `apiFetch` + `ReadableStream` SSE 解析（见 `api/intent.ts`、`api/recommend.ts`、`api/multi-agent.ts`）。
+- **草稿/历史持久化在后端（2026-06）**：草稿不再存浏览器 `localStorage`，改走后端 MySQL（`/api/roundtable-drafts`，按登录用户隔离、跨设备保留）。`api/drafts.ts` 封装 CRUD，`useDraftPersistence` 异步对接、对外接口不变。**推荐弹窗带「会话级推荐历史」**：每个草稿一串历史，重开弹窗先显示上次推荐结果，可「重新分析」重跑（详见 §5.7 / §6.2）。后端实现见 backend doc 的「圆桌草稿持久化」一节。
 - **不写死模型名**：与 `/page/canvas/ai-writing` 同样的策略 —— 全局选中模型由 `selectedModel` state 维护，首次 `useModels()` 完成后初始化为 `availableModels[0].name`，顶栏「模型」下拉可切换。所有 API 调用统一传 `model: selectedModel || undefined`；后端缺省由 `lead_agent._resolve_model_name()` 回退到 `config.yaml` 的 `models[0]`。**内网部署只需在 `config.yaml` 配自家模型即可，前后端均无需改动**。
 - **席位区间硬约束**：推荐弹窗要求 **2 ≤ 选中数 ≤ 8**（`MIN_SELECTED` / `MAX_SELECTED`）。
 - **协调智能体单例（2026-05）**：`COORDINATOR_AGENT_ID = "roundtable-coordinator"` 是固定常量，不再每会话生成新 id。后端 seeder 自动落地，「本次席位列表」改由后端通过 thread 历史 human 消息注入，前端无感。`newCoordinatorName()` 仅作为兼容性壳保留。
@@ -54,7 +55,8 @@ frontend-web/src/roundtable-planning/
 │   └── RoundtablePlanningPage.tsx      # 主页面：组合 hook + 顶栏/侧栏 + 三步路由（≈1 150 行）
 ├── api/
 │   ├── intent.ts                       # Step 1：initIntent、streamIntent
-│   ├── recommend.ts                    # 推荐：streamRecommend、filterRecommendCandidates、parseRecommendStreamContent
+│   ├── recommend.ts                    # 推荐：streamRecommend、filterRecommendCandidates、parseRecommendStreamContent；**+ 推荐历史 listRecommendHistory / saveRecommendHistory**
+│   ├── drafts.ts                       # **草稿后端 CRUD**：listDrafts / getDraft / createDraft / updateDraft / deleteDraft / deriveDraftName（替代旧 localStorage utils/drafts.ts）
 │   └── multi-agent.ts                  # Step 2/3：initMultiAgent、streamMultiAgent、listSessionArtifacts、buildArtifactDownloadUrl
 ├── components/
 │   ├── Avatars.tsx                     # 所有 SVG 头像 + getAvatar 选择器
@@ -73,14 +75,13 @@ frontend-web/src/roundtable-planning/
 │   ├── useAutoScroll.ts                # 滚动跟随 + 用户上滑暂停跟随
 │   ├── useStep1Intent.ts               # Step 1 状态机:init、send、stream、abort、INTENT_READY 同步
 │   ├── useStep2Orchestration.ts        # Step 2 状态机:initMultiAgent + 编排循环 + 对话流管理 + Persona
-│   └── useDraftPersistence.ts          # 草稿:自动保存(前进切换)/手动保存/加载/重命名/删除
+│   └── useDraftPersistence.ts          # 草稿(后端版):异步走 api/drafts;自动保存(前进切换)/手动保存/加载/重命名/删除/ensureDraft
 ├── lib/
 │   ├── clarification.ts                # resolveAllowMultiple（单选/多选解析）,与主聊天 ClarificationCard 对齐
 │   ├── reasoning.ts                    # splitInlineReasoning:从模型流式文本里提取 `<think>...</think>` 思考过程
 │   ├── roundtable-constants.ts         # 类型(RoleEntry / SelectedAgent) + 常量 + 纯工具函数(buildRoleTemplates / splitIntentReadyBlock 等)
 │   └── step-display.tsx                # Step 2 ChainOfThought 折叠卡图标 + 文案 + 子节点渲染
-├── utils/
-│   └── drafts.ts                       # localStorage 草稿存档（list / upsert / get / delete / rename / deriveName）
+├── utils/                              # （旧 drafts.ts 已删除 —— 草稿改走后端 api/drafts.ts）
 └── styles/
     └── roundtable-planning.css         # 页面主题变量（浅红品牌色 + 深色覆盖）
 ```
@@ -430,19 +431,28 @@ isOrchestrating = false
 - **禁用态**：`availableModels.length === 0` 或 `modelsLoading` → `disabled`，placeholder 显示「加载中…」或「未配置模型」
 - **可改换为 Radix Select**：当前用原生 `<select>` 与 Persona 抽屉保持一致；如需统一样式可改为 `@/components/ui/select`，行为不变
 
-### 5.7 草稿持久化（`utils/drafts.ts`）
+### 5.7 草稿持久化（`api/drafts.ts` + 后端 MySQL，2026-06 改造）
 
-- Key：`roundtable-planning:drafts:v1`（`localStorage`）
-- 草稿快照含：`step1.{intentThreadId, intentReady, interactiveMessages, timelineEntries}`、`step2.{threadIds, coordinatorName, step2RoundtableDialogues, lastLeaderContent, hasConsensus, consensusPercentage, budgetLimit}`、`furthestStep`
-- `deriveDraftName({intentReady, interactiveMessages})`：优先用 `intentReady.objective`，否则取第一个 user 消息，否则用 `未命名草稿 <时间>`，统一截断到 40 字
-- **自动保存策略**：仅在「前进」步骤切换时保存（1→2、2→3），后退不触发，避免覆盖 `furthestStep`
-- **手动保存**：顶栏 + Step 1 底部「保存草稿」按钮均调用 `handleManualSave`
-- **加载草稿**：`handleLoadDraft`
-  - 中断所有 in-flight（`abortAllInFlight + stopIntentStream`）
-  - **先**置 `loadedFromDraftRef.current = true`，再 `setCurrentStep(draft.furthestStep)`，避免 step1/step2 init effect 重新调用 init 把恢复的状态打掉
-  - `step1InitGenRef` 自增，淘汰可能晚归的旧 init
-  - `step2InitRef.current = draft.furthestStep >= 2`：曾经走到 Step 2 才标记已 init
-  - 角色面板基线状态：被恢复时全部置「在线」，否则 init 流程逐个置「在线」
+> **重大变更**：草稿从浏览器 `localStorage` 迁到**后端 MySQL**，按登录用户隔离、跨设备/会话保留。旧 `utils/drafts.ts` 已删除；`useDraftPersistence` 仍保持对外接口不变（列表项含 `name`，由后端 `title` 映射而来），主页面无需改动。后端表/路由见 backend doc「圆桌草稿持久化」。
+
+- **API 客户端 `api/drafts.ts`**：`listDrafts()`（轻量元数据）/ `getDraft(id)`（全量 step1+step2）/ `createDraft` / `updateDraft`（partial）/ `deleteDraft` / `deriveDraftName`。snake_case↔camelCase 映射在这一层收口（后端 `furthest_step/created_at/updated_at` ↔ 前端 `furthestStep/...`）。
+- **快照形状不变**：`DraftStep1Snapshot.{intentThreadId, intentReady, interactiveMessages, timelineEntries}`、`DraftStep2Snapshot.{threadIds, coordinatorName, step2RoundtableDialogues, lastLeaderContent, hasConsensus, consensusPercentage, budgetLimit, seatStubMessages}`；以 JSON 整存进后端 `step1`/`step2`（`PortableLongText`/LONGTEXT，避免大对话超 64KB）。
+- `deriveDraftName({intentReady, interactiveMessages})`：优先 `intentReady.objective`，否则第一个 user 消息，否则 `未命名草稿 <时间>`，截断 40 字。**仅在 create 时写一次**（见下方 bug 修复）。
+- **列表加载**：挂载时 `useEffect` 调 `listDrafts()`；增删改后 `refreshList()` 重拉。
+- **自动保存策略**：仅在「前进」步骤切换时保存（1→2、2→3），后退不触发。
+- **手动保存 / `ensureDraft`**：`handleManualSave` 走同一条保存队列；`ensureDraft(furthestStep?)` 幂等保证后端有一条对应草稿并返回 id（**推荐弹窗打开前调用**，让推荐历史有 `draft_id` 归属）。
+- **加载草稿 `handleLoadDraft`**：
+  - 走 `getDraft(id)` **实时拉服务端最新全量**（不再读内存里可能过期的列表快照）；404 → toast「草稿已不存在」并刷新列表。
+  - **先**置 `loadedFromDraftRef.current = true`，再 `setCurrentStep(draft.furthestStep)`（顺序不能反）。
+  - `setStep2InitGate(draft.furthestStep >= 2)`：曾走到 Step 2 才跳过自动 init。
+  - 角色面板基线：被恢复时全部「在线」，否则 init 流程逐个「在线」。
+
+**本次随手修复的 4 个旧 localStorage 版 bug**（对接后端时一并解决）：
+
+1. **自动保存覆盖手动重命名** —— 现在所有 `updateDraft` 一律**不带 title**；标题只由 create（派生）与显式「重命名」改写。
+2. **快速连续前进（1→2→3）产生重复草稿** —— 所有保存经一条**串行队列**（`saveQueueRef`）+ 同步镜像 `currentDraftIdRef`，create 完成即写 id，后续保存读到 id 走 update 而非再 create。
+3. **加载读到内存旧列表快照** —— 改为 `getDraft(id)` 实时拉。
+4. **localStorage 配额静默写失败 / 数据丢失** —— 迁后端后不复存在。
 
 ---
 
@@ -460,15 +470,22 @@ isOrchestrating = false
 
 ### 6.2 `RecommendAgentsDialog`
 
-- 打开时管线：`listAgents() → filterRecommendCandidates() → streamRecommend()`
+- **新增 `draftId` prop（2026-06）**：本次推荐归属的草稿/会话 id。由主页面在打开弹窗前 `await drafts.ensureDraft(1)` 拿到并传入（见 §6.6）。`null` 时退化为「直接推荐、不存历史」的旧行为。
+- 打开时管线（`start(forceFresh=false)`）：`listAgents() → filterRecommendCandidates()` → **若 `draftId` 有历史 → `listRecommendHistory()` 取最近一次直接展示（不重跑模型）**；否则 `streamRecommend()` 跑新推荐。
 - 阶段 UI（`InternalState.phase`）：
-  - `idle` → `loading_agents` → `recommending`（顶部「模型生成内容」区流式；下方推荐卡 streaming=true）→ `ready` / `fallback` / `error`
+  - `idle` → `loading_agents` → **`loading_history`** → **`history`**（展示上次推荐，可直接进入研讨或「重新分析」）
+  - 或 `loading_agents` → `recommending`（顶部「模型生成内容」流式；推荐卡 streaming=true）→ `ready` / `fallback` / `error`
+- **推荐历史（会话级）**：
+  - 打开时若该 `draftId` 有历史，进入 `history` 态：蓝色提示条标注上次推荐时间，picks 默认勾选、rationale 回显，**不**自动重跑模型。
+  - 「**重新分析**」按钮（`history` 态显示该文案，其它态为「重新推荐」）→ `start(true)` 强制跑新推荐，跳过历史。
+  - 每次**新推荐**完成（`ready`/`fallback`）后 `saveRecommendHistory(draftId, {objective,status,model,rationale,picks,candidates})` best-effort 追加一条（失败仅告警不阻断）。
+  - `history`/`fallback`/`ready` 三态都允许 `onConfirm`（用户可直接采纳历史 picks）。
 - **布局分两段**：
   - **顶部「模型生成内容」**：`RecommenderOutput` —— 状态行（`spinner + 调度推荐 / 已连接 / 正在生成`）+ `MessageStepsCard` 步骤卡 + `ReasoningBlock`（从 `<think>` 拆出来）+ markdown rationale + 末尾 `StreamingCursor`。视觉与 Step 1/Step 2 气泡的内容区一致，参考主聊天 `chats/new`。
   - **底部「推荐参会智能体」**：候选卡片(`AgentCard`),分「推荐参会 / 其他候选」两组。
 - 选择约束：`2 ≤ selected ≤ 8`；底部实时显示「至少再选 N 个」「已选 X / 8」
 - 失败兜底：候选为空 → error；status==="asking" 或 picks 不足 2 → `fallback`，自动用 `FALLBACK_AGENT_IDS` 与候选池交集勾选
-- 反竞态：内部 `generationRef`，每次 `runRecommendation` 自增，所有 setState 都校验 `myGen === generationRef.current`，避免「重新推荐」时旧 stream 写脏 state
+- 反竞态：内部 `generationRef`，每次 `start()` 自增，所有 setState（含历史读取/流式回调）都校验 `myGen === generationRef.current`，避免「重新分析」时旧 stream / 旧历史写脏 state
 - `onConfirm(agentIds)` → 父组件 `handleConfirmAgents`：复位 `step2InitRef` 并 `setCurrentStep(2)`，触发 Step 2 init
 
 ### 6.3 `HighFidelityReport`（Step 3）
@@ -613,7 +630,7 @@ interface StreamMultiAgentResult {
 | `FALLBACK_AGENT_IDS` | 6 内置 roundtable-* | RecommendAgentsDialog |
 | `BUILTIN_AGENT_META` | 6 内置头像/中文名 | RoundtablePlanningPage |
 | `FALLBACK_AVATAR_TYPES` | `[blue, purple, emerald, amber, pink]` | RoundtablePlanningPage |
-| `STORAGE_KEY` | `"roundtable-planning:drafts:v1"` | drafts.ts |
+| ~~`STORAGE_KEY`~~ | 已废弃 —— 草稿改存后端 `/api/roundtable-drafts`，不再用 localStorage | （旧 utils/drafts.ts，已删除） |
 | `INTENT_READY_RE` | `/\[INTENT_READY\]\s*```json\s*(\{[\s\S]*?\})\s*```/` | RoundtablePlanningPage |
 | `RECOMMEND_READY_MARKER` | `"[RECOMMEND_READY]"` | recommend.ts |
 | `EXCLUDED_AGENT_IDS` / 前缀 | `roundtable-intent`、`roundtable-recommender`、`roundtable-coordinator-*` | recommend.ts |
@@ -699,8 +716,8 @@ window.__MULTI_AGENT_DEBUG__       = true;  // stream 帧计数（默认开）
 
 1. **不要在 Panel 组件里写 fetch / setState 业务流程** —— Panel 只渲染,事件透传到 hook。
 2. **hook 之间不要互相 import** —— 跨 hook 的数据通过主页面的 ref / callback 桥接(参考 `buildInitialIntent` 与 `loadedFromDraftRef` 的实现)。
-3. **草稿 snapshot 形状**(Step 1 的 `Step1Snapshot` / Step 2 的 `Step2Snapshot`)与 `utils/drafts.ts` 的 `DraftStep{1,2}Snapshot` 是同一份语义,改字段时两边都要改。
-4. **lint/typecheck**:重构后 `pnpm typecheck` 与 `pnpm build` 必须保持与重构前同样的 0 个新增错误(预存量 50 个 pre-existing 错误来自其它模块如 `next/link` shim,与本特性无关)。
+3. **草稿 snapshot 形状**(Step 1 的 `Step1Snapshot` / Step 2 的 `Step2Snapshot`)与 `api/drafts.ts` 的 `DraftStep{1,2}Snapshot` 是同一份语义,改字段时两边都要改;后端以 JSON 整存,不强约束嵌套字段。
+4. **lint/typecheck**:重构后 `pnpm typecheck` 与 `pnpm build` 必须保持 0 个新增错误(预存量错误来自其它无关模块如 `strategy-components` / `next/link` shim,与本特性无关)。
 
 ### 12.7 文档协同
 
@@ -731,7 +748,8 @@ window.__MULTI_AGENT_DEBUG__       = true;  // stream 帧计数（默认开）
 | 常量/工具函数 | `frontend-web/src/roundtable-planning/lib/roundtable-constants.ts` |
 | ChainOfThought 渲染 | `frontend-web/src/roundtable-planning/lib/step-display.tsx` |
 | Step 1 API | `frontend-web/src/roundtable-planning/api/intent.ts` |
-| 推荐 API | `frontend-web/src/roundtable-planning/api/recommend.ts` |
+| 推荐 API（含推荐历史） | `frontend-web/src/roundtable-planning/api/recommend.ts` |
+| 草稿 API（后端 CRUD） | `frontend-web/src/roundtable-planning/api/drafts.ts` |
 | Step 2/3 API | `frontend-web/src/roundtable-planning/api/multi-agent.ts` |
 | 澄清卡片 | `frontend-web/src/roundtable-planning/components/IntentClarificationCard.tsx` |
 | 推荐弹窗 | `frontend-web/src/roundtable-planning/components/RecommendAgentsDialog.tsx` |
@@ -740,7 +758,7 @@ window.__MULTI_AGENT_DEBUG__       = true;  // stream 帧计数（默认开）
 | 回底按钮 | `frontend-web/src/roundtable-planning/components/ScrollToBottomButton.tsx` |
 | 滚动跟随 | `frontend-web/src/roundtable-planning/hooks/useAutoScroll.ts` |
 | 澄清单选逻辑 | `frontend-web/src/roundtable-planning/lib/clarification.ts` |
-| 草稿存档 | `frontend-web/src/roundtable-planning/utils/drafts.ts` |
+| 后端草稿/历史路由 | `offline-backend-20260512/backend/app/gateway/routers/roundtable_drafts.py` |
 | 路由 | `frontend-web/src/pages/WorkspaceRoutes.tsx` |
 | 样式 | `frontend-web/src/roundtable-planning/styles/roundtable-planning.css` |
 | 主聊天澄清参考 | `frontend-web/src/components/workspace/messages/message-list.tsx` |
diff --git a/frontend-web/src/roundtable-planning/api/drafts.ts b/frontend-web/src/roundtable-planning/api/drafts.ts
new file mode 100644
index 00000000..008305a5
--- /dev/null
+++ b/frontend-web/src/roundtable-planning/api/drafts.ts
@@ -0,0 +1,211 @@
+/**
+ * Backend-persisted draft storage for /page/roundtable/planning.
+ *
+ * One draft = one full roundtable session (Step 1 intent + Step 2 multi-agent
+ * roundtable). Persisted server-side (MySQL) bound to the logged-in user, so it
+ * follows the account across devices/sessions — replacing the old browser-only
+ * `localStorage` storage (see git history of `utils/drafts.ts`).
+ *
+ * The list endpoint returns lightweight metadata only; the full step1/step2
+ * snapshots are fetched on demand via {@link getDraft}. This keeps the history
+ * dropdown cheap and guarantees a load always reads the freshest server copy
+ * (the old localStorage path loaded from a possibly-stale in-memory array).
+ */
+
+import { apiFetch } from "@/core/api/fetch-client";
+import { getBackendBaseURL } from "@/core/config";
+
+import type { IntentSummary } from "@/roundtable-planning/api/intent";
+import type { ThreadIdMap } from "@/roundtable-planning/api/multi-agent";
+
+/**
+ * Step 1 snapshot. Shapes for the chat / timeline lists stay loose (`unknown[]`)
+ * because their inline types in the hooks are wide; the hook casts on hydrate.
+ */
+export interface DraftStep1Snapshot {
+  intentThreadId: string | null;
+  intentReady: IntentSummary | null;
+  interactiveMessages: unknown[];
+  timelineEntries: unknown[];
+}
+
+export interface DraftStep2Snapshot {
+  threadIds: ThreadIdMap | null;
+  coordinatorName: string | null;
+  step2RoundtableDialogues: unknown[];
+  lastLeaderContent: string;
+  hasConsensus: boolean;
+  consensusPercentage: number;
+  budgetLimit: number;
+  /** Per-seat stub AI messages for the sandbox `write-file:` virtual URL loader. */
+  seatStubMessages?: Record<string, unknown[]>;
+}
+
+/** Lightweight metadata row returned by the list endpoint. */
+export interface DraftMeta {
+  id: string;
+  /** Auto-derived task summary (objective → first user message → timestamp). */
+  title: string;
+  furthestStep: 1 | 2 | 3;
+  createdAt: string;
+  updatedAt: string;
+}
+
+/** Full record including the step snapshots. */
+export interface DraftRecord extends DraftMeta {
+  step1: DraftStep1Snapshot | null;
+  step2: DraftStep2Snapshot | null;
+}
+
+/** Fields accepted when creating a draft. */
+export interface DraftCreatePayload {
+  /** Optional stable id so the client can keep one id across create + updates. */
+  id?: string;
+  title: string;
+  furthestStep: 1 | 2 | 3;
+  step1: DraftStep1Snapshot;
+  step2: DraftStep2Snapshot;
+}
+
+/** Partial update — send only what changed. Omitting `title` preserves a rename. */
+export interface DraftUpdatePayload {
+  title?: string;
+  furthestStep?: 1 | 2 | 3;
+  step1?: DraftStep1Snapshot;
+  step2?: DraftStep2Snapshot;
+}
+
+const BASE = () => `${getBackendBaseURL()}/api/roundtable-drafts`;
+
+async function readJson<T>(res: Response, ctx: string): Promise<T> {
+  if (!res.ok) {
+    const text = await res.text().catch(() => "");
+    throw new Error(`${ctx} failed (${res.status}): ${text}`);
+  }
+  return (await res.json()) as T;
+}
+
+// ── server ⇄ client field mapping ──────────────────────────────────────────
+// Backend uses snake_case (furthest_step / created_at / updated_at); the rest
+// of the frontend reads camelCase. Map at this boundary so callers stay clean.
+
+interface ServerDraftMeta {
+  id: string;
+  title: string;
+  furthest_step: number;
+  created_at: string | null;
+  updated_at: string | null;
+}
+
+interface ServerDraftFull extends ServerDraftMeta {
+  step1: DraftStep1Snapshot | null;
+  step2: DraftStep2Snapshot | null;
+}
+
+function toMeta(s: ServerDraftMeta): DraftMeta {
+  const step = Math.min(3, Math.max(1, s.furthest_step || 1)) as 1 | 2 | 3;
+  return {
+    id: s.id,
+    title: s.title ?? "",
+    furthestStep: step,
+    createdAt: s.created_at ?? "",
+    updatedAt: s.updated_at ?? "",
+  };
+}
+
+function toRecord(s: ServerDraftFull): DraftRecord {
+  return { ...toMeta(s), step1: s.step1 ?? null, step2: s.step2 ?? null };
+}
+
+function fromCreate(p: DraftCreatePayload) {
+  return {
+    ...(p.id ? { id: p.id } : {}),
+    title: p.title,
+    furthest_step: p.furthestStep,
+    step1: p.step1,
+    step2: p.step2,
+  };
+}
+
+function fromUpdate(p: DraftUpdatePayload) {
+  const out: Record<string, unknown> = {};
+  if (p.title !== undefined) out.title = p.title;
+  if (p.furthestStep !== undefined) out.furthest_step = p.furthestStep;
+  if (p.step1 !== undefined) out.step1 = p.step1;
+  if (p.step2 !== undefined) out.step2 = p.step2;
+  return out;
+}
+
+// ── CRUD ────────────────────────────────────────────────────────────────────
+
+/** List the current user's drafts, newest first (lightweight metadata only). */
+export async function listDrafts(): Promise<DraftMeta[]> {
+  const res = await apiFetch(BASE(), { method: "GET" });
+  const data = await readJson<{ drafts: ServerDraftMeta[] }>(res, "list drafts");
+  return (data.drafts ?? []).map(toMeta);
+}
+
+/** Fetch the full draft (incl. step1/step2). Returns null on 404. */
+export async function getDraft(id: string): Promise<DraftRecord | null> {
+  const res = await apiFetch(`${BASE()}/${encodeURIComponent(id)}`, { method: "GET" });
+  if (res.status === 404) return null;
+  const data = await readJson<ServerDraftFull>(res, "get draft");
+  return toRecord(data);
+}
+
+export async function createDraft(payload: DraftCreatePayload): Promise<DraftRecord> {
+  const res = await apiFetch(BASE(), {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify(fromCreate(payload)),
+  });
+  const data = await readJson<ServerDraftFull>(res, "create draft");
+  return toRecord(data);
+}
+
+export async function updateDraft(id: string, patch: DraftUpdatePayload): Promise<DraftRecord | null> {
+  const res = await apiFetch(`${BASE()}/${encodeURIComponent(id)}`, {
+    method: "PUT",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify(fromUpdate(patch)),
+  });
+  if (res.status === 404) return null;
+  const data = await readJson<ServerDraftFull>(res, "update draft");
+  return toRecord(data);
+}
+
+export async function deleteDraft(id: string): Promise<void> {
+  const res = await apiFetch(`${BASE()}/${encodeURIComponent(id)}`, { method: "DELETE" });
+  if (!res.ok && res.status !== 404) {
+    const text = await res.text().catch(() => "");
+    throw new Error(`delete draft failed (${res.status}): ${text}`);
+  }
+}
+
+/**
+ * Derive a short task-summary name from whatever the user supplied so far.
+ * Preference: intent objective → first user message → timestamp. Trimmed to
+ * ~40 chars so dropdown rows stay scannable. (Ported from the old utils/drafts.)
+ */
+export function deriveDraftName(args: {
+  intentReady?: IntentSummary | null;
+  interactiveMessages?: unknown[];
+}): string {
+  const MAX = 40;
+  const truncate = (s: string) => (s.length > MAX ? s.slice(0, MAX) + "…" : s);
+
+  const objective = args.intentReady?.objective?.trim();
+  if (objective) return truncate(objective);
+
+  if (Array.isArray(args.interactiveMessages)) {
+    for (const m of args.interactiveMessages) {
+      if (!m || typeof m !== "object") continue;
+      const obj = m as Record<string, unknown>;
+      if (obj.sender === "user" && typeof obj.content === "string" && obj.content.trim()) {
+        return truncate(obj.content.trim());
+      }
+    }
+  }
+
+  return `未命名草稿 ${new Date().toLocaleString("zh-CN", { hour12: false })}`;
+}
diff --git a/frontend-web/src/roundtable-planning/api/recommend.ts b/frontend-web/src/roundtable-planning/api/recommend.ts
index 01fb53bb..c1975bb1 100644
--- a/frontend-web/src/roundtable-planning/api/recommend.ts
+++ b/frontend-web/src/roundtable-planning/api/recommend.ts
@@ -198,6 +198,104 @@ function extractAiTextDelta(parsed: unknown): { text: string; id: string } | nul
   return { text, id };
 }
 
+/**
+ * One persisted recommendation run, bound to a draft/session. Returned by
+ * {@link listRecommendHistory}; lets the dialog show "last recommendation for
+ * this session" before deciding whether to re-analyze.
+ */
+export interface RecommendHistoryItem {
+  id: string;
+  draftId: string;
+  objective: string;
+  /** How the run ended on the backend: "done" | "fallback" | "asking" | "error". */
+  status: string;
+  model: string | null;
+  rationale: string | null;
+  picks: RecommendPick[];
+  /** Candidate-pool snapshot captured at recommend time. */
+  candidates: RecommendCandidate[];
+  createdAt: string;
+}
+
+/** Payload persisted after a recommendation run completes. */
+export interface SaveRecommendHistoryPayload {
+  objective: string;
+  status: string;
+  model?: string;
+  rationale?: string;
+  picks: RecommendPick[];
+  candidates: RecommendCandidate[];
+}
+
+interface ServerHistoryItem {
+  id: string;
+  draft_id: string;
+  objective: string;
+  status: string;
+  model: string | null;
+  rationale: string | null;
+  picks: RecommendPick[];
+  candidates: RecommendCandidate[];
+  created_at: string | null;
+}
+
+function toHistoryItem(s: ServerHistoryItem): RecommendHistoryItem {
+  return {
+    id: s.id,
+    draftId: s.draft_id,
+    objective: s.objective ?? "",
+    status: s.status ?? "done",
+    model: s.model ?? null,
+    rationale: s.rationale ?? null,
+    picks: Array.isArray(s.picks) ? s.picks : [],
+    candidates: Array.isArray(s.candidates) ? s.candidates : [],
+    createdAt: s.created_at ?? "",
+  };
+}
+
+/** List a draft's recommendation history, newest first. */
+export async function listRecommendHistory(draftId: string): Promise<RecommendHistoryItem[]> {
+  const res = await apiFetch(
+    `${getBackendBaseURL()}/api/roundtable-drafts/${encodeURIComponent(draftId)}/recommendations`,
+    { method: "GET" },
+  );
+  if (res.status === 404) return [];
+  if (!res.ok) {
+    const text = await res.text().catch(() => "");
+    throw new Error(`list recommend history failed (${res.status}): ${text}`);
+  }
+  const data = (await res.json()) as { history: ServerHistoryItem[] };
+  return (data.history ?? []).map(toHistoryItem);
+}
+
+/** Append one recommendation run to a draft's history. */
+export async function saveRecommendHistory(
+  draftId: string,
+  payload: SaveRecommendHistoryPayload,
+): Promise<RecommendHistoryItem | null> {
+  const res = await apiFetch(
+    `${getBackendBaseURL()}/api/roundtable-drafts/${encodeURIComponent(draftId)}/recommendations`,
+    {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({
+        objective: payload.objective,
+        status: payload.status,
+        model: payload.model,
+        rationale: payload.rationale,
+        picks: payload.picks,
+        candidates: payload.candidates,
+      }),
+    },
+  );
+  if (res.status === 404) return null; // draft gone — skip silently
+  if (!res.ok) {
+    const text = await res.text().catch(() => "");
+    throw new Error(`save recommend history failed (${res.status}): ${text}`);
+  }
+  return toHistoryItem((await res.json()) as ServerHistoryItem);
+}
+
 /**
  * Open an SSE connection to `/api/recommend/stream` and resolve once the
  * upstream stream finishes. Each call is one recommendation round — the
diff --git a/frontend-web/src/roundtable-planning/components/RecommendAgentsDialog.tsx b/frontend-web/src/roundtable-planning/components/RecommendAgentsDialog.tsx
index 9bb89f75..6c8d09fb 100644
--- a/frontend-web/src/roundtable-planning/components/RecommendAgentsDialog.tsx
+++ b/frontend-web/src/roundtable-planning/components/RecommendAgentsDialog.tsx
@@ -31,9 +31,12 @@ import { MessageStepsCard } from "./MessageStepsCard";
 import type { IntentSummary } from "../api/intent";
 import {
   filterRecommendCandidates,
+  listRecommendHistory,
   parseRecommendStreamContent,
+  saveRecommendHistory,
   streamRecommend,
   type RecommendCandidate,
+  type RecommendHistoryItem,
   type RecommendStreamPhase,
 } from "../api/recommend";
 import { useAutoScroll } from "../hooks/useAutoScroll";
@@ -56,6 +59,13 @@ export interface RecommendAgentsDialogProps {
   intent: IntentSummary | null;
   /** Falls back to the recommender's own default if omitted. */
   model?: string;
+  /**
+   * Draft/session id this recommendation belongs to. When present, the dialog
+   * shows the session's last recommendation on open (with a「重新分析」button)
+   * and persists each fresh run to history. Null → behaves statelessly (always
+   * runs a fresh recommendation, no history).
+   */
+  draftId?: string | null;
   onCancel: () => void;
   /** Called with the user's final selection of agent ids (in display order). */
   onConfirm: (agents: Array<{ agent_id: string; name: string }>) => void;
@@ -81,6 +91,8 @@ const FALLBACK_AGENT_IDS = [
 type DialogPhase =
   | "idle"
   | "loading_agents"
+  | "loading_history"
+  | "history"
   | "recommending"
   | "ready"
   | "fallback"
@@ -111,6 +123,8 @@ interface InternalState {
   warning: string | null;
   /** Latest error message; only shown when `phase === "error"`. */
   errorMessage: string | null;
+  /** ISO time of the historical record being shown (phase === "history"). */
+  historyCreatedAt: string | null;
 }
 
 const INITIAL_STATE: InternalState = {
@@ -123,6 +137,7 @@ const INITIAL_STATE: InternalState = {
   recommendedIds: [],
   warning: null,
   errorMessage: null,
+  historyCreatedAt: null,
 };
 
 /** 推荐流程的 phase → 中文标签;与 RoundtablePlanningPage::phaseLabel 风格保持一致。 */
@@ -164,6 +179,7 @@ export function RecommendAgentsDialog({
   open,
   intent,
   model,
+  draftId,
   onCancel,
   onConfirm,
 }: RecommendAgentsDialogProps) {
@@ -185,172 +201,240 @@ export function RecommendAgentsDialog({
   }, []);
 
   /**
-   * The full pipeline: fetch agents → run recommendation → seed selection.
-   * Wrapped so the user can re-trigger it via the "重新推荐" button.
+   * Render a stored history record as the current view (no streaming). Used on
+   * open when the session already has a prior recommendation. The user can
+   * confirm these picks directly or click「重新分析」to run a fresh one.
    */
-  const runRecommendation = useCallback(async () => {
-    if (!intent || !intent.objective) {
-      setState({
-        ...INITIAL_STATE,
-        phase: "error",
-        errorMessage: "缺少任务目标,无法启动推荐",
-      });
-      return;
+  const applyHistory = useCallback((item: RecommendHistoryItem, candidates: RecommendCandidate[]) => {
+    const reasons: Record<string, string> = {};
+    const recommendedIds: string[] = [];
+    for (const p of item.picks) {
+      reasons[p.agent_id] = p.reason;
+      recommendedIds.push(p.agent_id);
     }
-
-    cancelInflight();
-    const myGen = ++generationRef.current;
-    const controller = new AbortController();
-    abortRef.current = controller;
-
-    setState({ ...INITIAL_STATE, phase: "loading_agents" });
-    setSelectedIds(new Set());
-
-    let agents: Agent[];
-    try {
-      agents = await listAgents();
-    } catch (err) {
-      if (myGen !== generationRef.current) return;
-      setState({
-        ...INITIAL_STATE,
-        phase: "error",
-        errorMessage: `加载智能体列表失败:${err instanceof Error ? err.message : String(err)}`,
-      });
-      return;
-    }
-    if (myGen !== generationRef.current) return;
-
-    const candidates = filterRecommendCandidates(agents);
-    if (candidates.length === 0) {
-      setState({
-        ...INITIAL_STATE,
-        phase: "error",
-        errorMessage: "候选智能体池为空,请先在系统里创建至少 2 个智能体",
-      });
-      return;
-    }
-
+    const seed = intersectInOrder(recommendedIds, candidates);
+    setSelectedIds(new Set(seed));
+    const { rationale } = parseRecommendStreamContent(item.rationale || "");
     setState((prev) => ({
       ...prev,
-      phase: "recommending",
+      phase: "history",
       candidates,
-      rationale: "",
-      streamPhase: "connecting",
-      // 推荐开始,清空上一轮的步骤时间线并打第一条点。
-      steps: [
-        { id: "0", key: "connecting", label: recommendPhaseLabel("connecting"), ts: nowTimeStr() },
-      ],
+      rationale: rationale || item.rationale || "",
+      recommendedIds: seed,
+      reasons,
+      streamPhase: null,
+      steps: [],
+      warning: null,
+      historyCreatedAt: item.createdAt || null,
     }));
+  }, []);
 
-    try {
-      const result = await streamRecommend({
-        intent,
+  /**
+   * Stream a fresh recommendation over the given candidate pool, seed the
+   * selection, and (when bound to a draft) persist the run to history.
+   */
+  const analyze = useCallback(
+    async (candidates: RecommendCandidate[], myGen: number, signal: AbortSignal) => {
+      if (!intent) return;
+      setState((prev) => ({
+        ...prev,
+        phase: "recommending",
         candidates,
-        model,
-        signal: controller.signal,
-        onRationaleUpdate: (rationale) => {
-          if (myGen !== generationRef.current) return;
-          setState((prev) => ({ ...prev, rationale }));
-        },
-        onPicksUpdate: (picks) => {
-          if (myGen !== generationRef.current) return;
-          const reasons: Record<string, string> = {};
-          const recommendedIds: string[] = [];
-          for (const p of picks) {
+        rationale: "",
+        reasons: {},
+        recommendedIds: [],
+        warning: null,
+        historyCreatedAt: null,
+        streamPhase: "connecting",
+        // 推荐开始,清空上一轮的步骤时间线并打第一条点。
+        steps: [
+          { id: "0", key: "connecting", label: recommendPhaseLabel("connecting"), ts: nowTimeStr() },
+        ],
+      }));
+      setSelectedIds(new Set());
+
+      try {
+        const result = await streamRecommend({
+          intent,
+          candidates,
+          model,
+          signal,
+          onRationaleUpdate: (rationale) => {
+            if (myGen !== generationRef.current) return;
+            setState((prev) => ({ ...prev, rationale }));
+          },
+          onPicksUpdate: (picks) => {
+            if (myGen !== generationRef.current) return;
+            const reasons: Record<string, string> = {};
+            const recommendedIds: string[] = [];
+            for (const p of picks) {
+              reasons[p.agent_id] = p.reason;
+              recommendedIds.push(p.agent_id);
+            }
+            const seed = intersectInOrder(recommendedIds, candidates);
+            setSelectedIds((prev) => {
+              const next = new Set(prev);
+              for (const id of seed) next.add(id);
+              return next;
+            });
+            setState((prev) => ({ ...prev, recommendedIds, reasons }));
+          },
+          onPhaseChange: (phase) => {
+            if (myGen !== generationRef.current) return;
+            setState((prev) => {
+              const key = phase;
+              const last = prev.steps[prev.steps.length - 1];
+              const steps =
+                last && last.key === key
+                  ? prev.steps
+                  : [
+                      ...prev.steps,
+                      {
+                        id: String(prev.steps.length),
+                        key,
+                        label: recommendPhaseLabel(phase),
+                        ts: nowTimeStr(),
+                      },
+                    ];
+              return { ...prev, streamPhase: phase, steps };
+            });
+          },
+        });
+        if (myGen !== generationRef.current) return;
+
+        const reasons: Record<string, string> = {};
+        const recommendedIds: string[] = [];
+        if (result.picks) {
+          for (const p of result.picks) {
             reasons[p.agent_id] = p.reason;
             recommendedIds.push(p.agent_id);
           }
-          const seed = intersectInOrder(recommendedIds, candidates);
-          setSelectedIds((prev) => {
-            const next = new Set(prev);
-            for (const id of seed) next.add(id);
-            return next;
-          });
-          setState((prev) => ({
-            ...prev,
-            recommendedIds,
-            reasons,
-          }));
-        },
-        onPhaseChange: (phase) => {
-          if (myGen !== generationRef.current) return;
-          setState((prev) => {
-            const key = phase;
-            const last = prev.steps[prev.steps.length - 1];
-            // 连续重复阶段不重复 push,只在阶段切换时打点。
-            const steps =
-              last && last.key === key
-                ? prev.steps
-                : [
-                    ...prev.steps,
-                    {
-                      id: String(prev.steps.length),
-                      key,
-                      label: recommendPhaseLabel(phase),
-                      ts: nowTimeStr(),
-                    },
-                  ];
-            return { ...prev, streamPhase: phase, steps };
-          });
-        },
-      });
-      if (myGen !== generationRef.current) return;
-
-      const reasons: Record<string, string> = {};
-      const recommendedIds: string[] = [];
-      if (result.picks) {
-        for (const p of result.picks) {
-          reasons[p.agent_id] = p.reason;
-          recommendedIds.push(p.agent_id);
         }
-      }
-
-      if (result.status === "done" && recommendedIds.length >= MIN_SELECTED) {
-        const seed = intersectInOrder(recommendedIds, candidates);
-        setSelectedIds(new Set(seed));
         const { rationale } = parseRecommendStreamContent(result.content || "");
+
+        let finalPhase: DialogPhase;
+        let finalIds: string[];
+        let warning: string | null;
+        if (result.status === "done" && recommendedIds.length >= MIN_SELECTED) {
+          finalPhase = "ready";
+          finalIds = intersectInOrder(recommendedIds, candidates);
+          warning = null;
+        } else {
+          // Fallback: status === "asking" or picks too few. Use the 6 built-in
+          // roundtable seats, intersected against the live pool.
+          finalPhase = "fallback";
+          const fallback = intersectInOrder(FALLBACK_AGENT_IDS, candidates);
+          finalIds =
+            fallback.length >= MIN_SELECTED
+              ? fallback
+              : candidates.slice(0, MIN_SELECTED).map((c) => c.agent_id);
+          warning = "智能推荐未生成有效结果,已为你勾选默认的圆桌席位,可手动调整。";
+        }
+
+        setSelectedIds(new Set(finalIds));
         setState((prev) => ({
           ...prev,
-          phase: "ready",
+          phase: finalPhase,
           rationale: rationale || prev.rationale,
-          recommendedIds,
+          recommendedIds: finalIds,
           reasons,
           streamPhase: null,
-          warning: null,
+          warning,
         }));
+
+        // Persist this run to the session's history (best-effort).
+        if (draftId) {
+          const picks = finalIds.map((id) => ({ agent_id: id, reason: reasons[id] ?? "" }));
+          void saveRecommendHistory(draftId, {
+            objective: intent.objective,
+            status: finalPhase === "ready" ? "done" : "fallback",
+            model,
+            rationale: rationale || result.content || "",
+            picks,
+            candidates,
+          }).catch((err) => console.warn("[recommend] 保存推荐历史失败:", err));
+        }
+      } catch (err) {
+        if (myGen !== generationRef.current) return;
+        // AbortError fires when the user closes the dialog mid-stream — no
+        // state update needed because the dialog is going away.
+        if (err instanceof DOMException && err.name === "AbortError") return;
+        setState((prev) => ({
+          ...prev,
+          phase: "error",
+          streamPhase: null,
+          errorMessage: `推荐过程出错:${err instanceof Error ? err.message : String(err)}`,
+        }));
+      }
+    },
+    [intent, model, draftId],
+  );
+
+  /**
+   * Open pipeline: load candidates → (unless forceFresh) show the session's
+   * last recommendation from history → otherwise run a fresh analysis.
+   * `forceFresh=true` is the「重新分析」/「重新推荐」path.
+   */
+  const start = useCallback(
+    async (forceFresh: boolean) => {
+      if (!intent || !intent.objective) {
+        setState({ ...INITIAL_STATE, phase: "error", errorMessage: "缺少任务目标,无法启动推荐" });
         return;
       }
 
-      // Fallback: status === "asking" or picks too few. Use the 6 built-in
-      // roundtable seats, intersected against the live pool.
-      const fallback = intersectInOrder(FALLBACK_AGENT_IDS, candidates);
-      const seed = fallback.length >= MIN_SELECTED
-        ? fallback
-        : candidates.slice(0, MIN_SELECTED).map((c) => c.agent_id);
-      setSelectedIds(new Set(seed));
-      const { rationale } = parseRecommendStreamContent(result.content || "");
-      setState((prev) => ({
-        ...prev,
-        phase: "fallback",
-        rationale: rationale || prev.rationale,
-        recommendedIds: seed,
-        reasons,
-        streamPhase: null,
-        warning: "智能推荐未生成有效结果,已为你勾选默认的圆桌席位,可手动调整。",
-      }));
-    } catch (err) {
+      cancelInflight();
+      const myGen = ++generationRef.current;
+      const controller = new AbortController();
+      abortRef.current = controller;
+
+      setState({ ...INITIAL_STATE, phase: "loading_agents" });
+      setSelectedIds(new Set());
+
+      let agents: Agent[];
+      try {
+        agents = await listAgents();
+      } catch (err) {
+        if (myGen !== generationRef.current) return;
+        setState({
+          ...INITIAL_STATE,
+          phase: "error",
+          errorMessage: `加载智能体列表失败:${err instanceof Error ? err.message : String(err)}`,
+        });
+        return;
+      }
       if (myGen !== generationRef.current) return;
-      // AbortError fires when the user closes the dialog mid-stream — no
-      // state update needed because the dialog is going away.
-      if (err instanceof DOMException && err.name === "AbortError") return;
-      setState((prev) => ({
-        ...prev,
-        phase: "error",
-        streamPhase: null,
-        errorMessage: `推荐过程出错:${err instanceof Error ? err.message : String(err)}`,
-      }));
-    }
-  }, [intent, model, cancelInflight]);
+
+      const candidates = filterRecommendCandidates(agents);
+      if (candidates.length === 0) {
+        setState({
+          ...INITIAL_STATE,
+          phase: "error",
+          errorMessage: "候选智能体池为空,请先在系统里创建至少 2 个智能体",
+        });
+        return;
+      }
+
+      // 若该会话已有推荐历史,默认展示最近一次(可重新分析),不自动重跑模型。
+      if (!forceFresh && draftId) {
+        setState((prev) => ({ ...prev, phase: "loading_history", candidates }));
+        try {
+          const history = await listRecommendHistory(draftId);
+          if (myGen !== generationRef.current) return;
+          if (history.length > 0) {
+            applyHistory(history[0], candidates);
+            return;
+          }
+        } catch (err) {
+          if (myGen !== generationRef.current) return;
+          // 历史拉取失败不阻断:退化为直接跑一次新推荐。
+          console.warn("[recommend] 拉取推荐历史失败,回退到新推荐:", err);
+        }
+      }
+
+      await analyze(candidates, myGen, controller.signal);
+    },
+    [intent, draftId, cancelInflight, analyze, applyHistory],
+  );
 
   // Trigger the pipeline whenever the dialog opens. Closing it aborts any
   // in-flight stream so the next open starts clean.
@@ -359,11 +443,11 @@ export function RecommendAgentsDialog({
       cancelInflight();
       return;
     }
-    void runRecommendation();
+    void start(false);
     return () => {
       cancelInflight();
     };
-  }, [open, runRecommendation, cancelInflight]);
+  }, [open, start, cancelInflight]);
 
   const handleToggle = useCallback((agentId: string) => {
     setSelectedIds((prev) => {
@@ -391,7 +475,7 @@ export function RecommendAgentsDialog({
 
   const selectedCount = selectedIds.size;
   const canConfirm =
-    (state.phase === "ready" || state.phase === "fallback") &&
+    (state.phase === "ready" || state.phase === "fallback" || state.phase === "history") &&
     selectedCount >= MIN_SELECTED &&
     selectedCount <= MAX_SELECTED;
 
@@ -452,7 +536,9 @@ export function RecommendAgentsDialog({
            ============================================================ */}
         <CollapsibleModelOutput
           isLoading={
-            state.phase === "loading_agents" || state.phase === "recommending"
+            state.phase === "loading_agents" ||
+            state.phase === "loading_history" ||
+            state.phase === "recommending"
           }
           phase={state.phase}
           scrollDeps={[state.rationale, state.steps, state.streamPhase]}
@@ -497,8 +583,20 @@ export function RecommendAgentsDialog({
                 </div>
               )}
             </>
-          ) : (state.phase === "ready" || state.phase === "fallback") ? (
+          ) : (state.phase === "ready" || state.phase === "fallback" || state.phase === "history") ? (
             <>
+              {state.phase === "history" && (
+                <div className="text-xs text-sky-700 dark:text-sky-300 p-2 border border-sky-300/50 dark:border-sky-700/50 rounded-md bg-sky-50 dark:bg-sky-950/30 flex items-center justify-between gap-2">
+                  <span>
+                    📌 这是本会话上次的推荐结果
+                    {state.historyCreatedAt
+                      ? `（${new Date(state.historyCreatedAt).toLocaleString("zh-CN", { hour12: false })}）`
+                      : ""}
+                    ，可直接进入研讨,或点「重新分析」让模型重新推荐。
+                  </span>
+                </div>
+              )}
+
               {state.warning && (
                 <div className="text-xs text-amber-700 dark:text-amber-300 p-2 border border-amber-300/50 dark:border-amber-700/50 rounded-md bg-amber-50 dark:bg-amber-950/30">
                   {state.warning}
@@ -551,15 +649,18 @@ export function RecommendAgentsDialog({
             >
               {selectionHint}
             </span>
-            {(state.phase === "ready" || state.phase === "fallback" || state.phase === "error") && (
+            {(state.phase === "ready" ||
+              state.phase === "fallback" ||
+              state.phase === "history" ||
+              state.phase === "error") && (
               <Button
-                variant="ghost"
+                variant={state.phase === "history" ? "secondary" : "ghost"}
                 size="sm"
-                onClick={() => void runRecommendation()}
+                onClick={() => void start(true)}
                 className="h-7 px-2 text-xs"
               >
                 <RefreshCwIcon className="size-3 mr-1" />
-                重新推荐
+                {state.phase === "history" ? "重新分析" : "重新推荐"}
               </Button>
             )}
           </div>
@@ -624,7 +725,8 @@ function CollapsibleModelOutput({
       hasAutoClosedRef.current = false;
       return;
     }
-    const isCompleted = phase === "ready" || phase === "fallback" || phase === "error";
+    const isCompleted =
+      phase === "ready" || phase === "fallback" || phase === "error" || phase === "history";
     if (isCompleted && !hasAutoClosedRef.current && open) {
       const timer = setTimeout(() => {
         setOpen(false);
@@ -639,6 +741,10 @@ function CollapsibleModelOutput({
     switch (phase) {
       case "loading_agents":
         return { label: "加载候选", className: "bg-slate-100 text-slate-600" };
+      case "loading_history":
+        return { label: "读取历史", className: "bg-slate-100 text-slate-600" };
+      case "history":
+        return { label: "历史记录", className: "bg-sky-50 text-sky-700 border border-sky-200" };
       case "recommending":
         return { label: "正在生成", className: "bg-purple-50 text-purple-700 border border-purple-200" };
       case "ready":
@@ -713,11 +819,11 @@ function CollapsibleModelOutput({
  *  - 流式中正文末尾跟一个闪烁竖线
  */
 function RecommenderOutput({ state }: { state: InternalState }) {
-  if (state.phase === "loading_agents") {
+  if (state.phase === "loading_agents" || state.phase === "loading_history") {
     return (
       <div className="flex items-center gap-2 text-muted-foreground text-sm">
         <Loader2Icon className="size-4 animate-spin" />
-        正在加载候选智能体...
+        {state.phase === "loading_history" ? "正在读取历史推荐..." : "正在加载候选智能体..."}
       </div>
     );
   }
@@ -733,7 +839,8 @@ function RecommenderOutput({ state }: { state: InternalState }) {
   const showRationaleStage =
     state.phase === "recommending" ||
     state.phase === "ready" ||
-    state.phase === "fallback";
+    state.phase === "fallback" ||
+    state.phase === "history";
   if (!showRationaleStage) return null;
 
   const isStreaming = state.phase === "recommending";
diff --git a/frontend-web/src/roundtable-planning/hooks/useDraftPersistence.ts b/frontend-web/src/roundtable-planning/hooks/useDraftPersistence.ts
index 524a2b01..c9b61a34 100644
--- a/frontend-web/src/roundtable-planning/hooks/useDraftPersistence.ts
+++ b/frontend-web/src/roundtable-planning/hooks/useDraftPersistence.ts
@@ -1,26 +1,42 @@
 /**
- * 圆桌规划草稿持久化 hook。
+ * 圆桌规划草稿持久化 hook（后端版）。
  *
- * 维护 localStorage 草稿列表与 currentDraftId、自动保存(前进切换 1→2、2→3)、
- * 手动保存、加载 / 重命名 / 删除。Step 1 / Step 2 的状态由各自的 hook 提供
- * 快照函数与 hydrateFromDraft;本 hook 只负责编排「什么时机存 / 加载到哪里」。
+ * 2026-06 改造：草稿从浏览器 localStorage 迁移到后端 MySQL（`/api/roundtable-drafts`），
+ * 按登录用户隔离、跨设备/会话保留。本 hook 仍只负责编排「什么时机存 / 加载到哪里」，
+ * Step 1 / Step 2 的状态由各自的 hook 提供快照函数与 hydrate。
+ *
+ * 对外返回的接口（`drafts` 列表项含 `name` 等字段、各 handler 签名）与旧版保持一致，
+ * 主页面无需改动。
+ *
+ * 同时修复了旧 localStorage 版的几个存储 bug：
+ *  1. **自动保存会覆盖手动重命名** —— 现在 create 时写一次派生标题，之后所有 update
+ *     一律不带 title；标题只由显式「重命名」改写。
+ *  2. **快速连续前进（1→2→3）产生重复草稿** —— 所有保存经一条串行队列，create 完成
+ *     即写入 `currentDraftIdRef`，后续保存读到 id 走 update 而非再 create。
+ *  3. **加载读到 state 里的旧列表快照** —— load 走 `getDraft(id)` 实时拉服务端最新全量。
+ *  4. localStorage 配额导致的静默写失败 / 数据丢失 —— 改后端后不复存在。
  */
 import { useCallback, useEffect, useRef, useState } from "react";
 
-import type {
-  Step1Snapshot,
-} from "@/roundtable-planning/hooks/useStep1Intent";
-import type {
-  Step2Snapshot,
-} from "@/roundtable-planning/hooks/useStep2Orchestration";
+import type { Step1Snapshot } from "@/roundtable-planning/hooks/useStep1Intent";
+import type { Step2Snapshot } from "@/roundtable-planning/hooks/useStep2Orchestration";
 import {
+  createDraft,
   deleteDraft,
   deriveDraftName,
+  getDraft,
   listDrafts,
-  renameDraft,
-  upsertDraft,
-  type DraftRecord,
-} from "@/roundtable-planning/utils/drafts";
+  updateDraft,
+} from "@/roundtable-planning/api/drafts";
+
+/** 历史下拉渲染用的轻量列表项（`name` 映射自后端 `title`）。 */
+export interface DraftListItem {
+  id: string;
+  name: string;
+  furthestStep: 1 | 2 | 3;
+  createdAt: string;
+  updatedAt: string;
+}
 
 export interface UseDraftPersistenceOptions {
   currentStep: number;
@@ -44,16 +60,12 @@ export interface UseDraftPersistenceOptions {
 }
 
 export interface UseDraftPersistenceReturn {
-  drafts: DraftRecord[];
+  drafts: DraftListItem[];
   currentDraftId: string | null;
   isHistoryOpen: boolean;
   renamingDraftId: string | null;
   renameDraftValue: string;
   historyDropdownRef: React.MutableRefObject<HTMLDivElement | null>;
-  /**
-   * Shared latch: 加载草稿后置 true,影响 LeftSidebar 在 Step 2 时角色面板的
-   * 初始 status(在线 vs 正在上线)。由主页面在加载完后 read,使用完可手动复位。
-   */
   loadedFromDraftRef: React.MutableRefObject<boolean>;
   setIsHistoryOpen: React.Dispatch<React.SetStateAction<boolean>>;
   setRenameDraftValue: React.Dispatch<React.SetStateAction<string>>;
@@ -61,9 +73,14 @@ export interface UseDraftPersistenceReturn {
   handleManualSave: () => void;
   handleLoadDraft: (id: string) => void;
   handleDeleteDraft: (id: string) => void;
-  handleStartRename: (draft: DraftRecord) => void;
+  handleStartRename: (draft: DraftListItem) => void;
   handleCommitRename: (id: string) => void;
   handleCancelRename: () => void;
+  /**
+   * 幂等地保证后端存在一条对应当前会话的草稿，返回其 id（无可保存内容时返回 null）。
+   * 推荐弹窗打开前调用，确保推荐历史有 draft_id 可归属。
+   */
+  ensureDraft: (furthestStep?: 1 | 2 | 3) => Promise<string | null>;
 }
 
 export function useDraftPersistence(
@@ -81,8 +98,8 @@ export function useDraftPersistence(
     setStep2InitGate,
   } = options;
 
-  const [drafts, setDrafts] = useState<DraftRecord[]>(() => listDrafts());
-  const [currentDraftId, setCurrentDraftId] = useState<string | null>(null);
+  const [drafts, setDrafts] = useState<DraftListItem[]>([]);
+  const [currentDraftId, setCurrentDraftIdState] = useState<string | null>(null);
   const [isHistoryOpen, setIsHistoryOpen] = useState<boolean>(false);
   const [renamingDraftId, setRenamingDraftId] = useState<string | null>(null);
   const [renameDraftValue, setRenameDraftValue] = useState<string>("");
@@ -91,95 +108,164 @@ export function useDraftPersistence(
   const prevStepRef = useRef<number>(currentStep);
   const historyDropdownRef = useRef<HTMLDivElement | null>(null);
 
-  const buildDraftPayload = useCallback(
-    (
-      furthestStep: 1 | 2 | 3,
-      idOverride?: string | null,
-    ): Parameters<typeof upsertDraft>[0] => {
-      const step1 = getStep1Snapshot();
-      const step2 = getStep2Snapshot();
-      const resolvedId = idOverride ?? currentDraftId;
-      return {
-        ...(resolvedId ? { id: resolvedId } : {}),
-        name: deriveDraftName({
-          intentReady: step1.intentReady,
-          interactiveMessages: step1.interactiveMessages,
-        }),
-        furthestStep,
-        step1,
-        step2,
-      };
+  // 同步镜像，供保存队列在 render 之前读到最新 id（修复重复建草稿竞态）。
+  const currentDraftIdRef = useRef<string | null>(null);
+  // 串行化所有写操作，避免并发 create。
+  const saveQueueRef = useRef<Promise<unknown>>(Promise.resolve());
+
+  // 包一层 setter：同时更新 state 与同步 ref。主页面通过它 setCurrentDraftId(null)。
+  const setCurrentDraftId = useCallback<React.Dispatch<React.SetStateAction<string | null>>>(
+    (action) => {
+      setCurrentDraftIdState((prev) => {
+        const next = typeof action === "function" ? (action as (p: string | null) => string | null)(prev) : action;
+        currentDraftIdRef.current = next;
+        return next;
+      });
     },
-    [currentDraftId, getStep1Snapshot, getStep2Snapshot],
+    [],
   );
 
-  const saveDraft = useCallback(
-    (furthestStep: 1 | 2 | 3): DraftRecord | null => {
-      const step1 = getStep1Snapshot();
-      const step2 = getStep2Snapshot();
-      const hasUserMessage = step1.interactiveMessages.some(
-        (m) => m.sender === "user",
+  const refreshList = useCallback(async () => {
+    try {
+      const metas = await listDrafts();
+      setDrafts(
+        metas.map((m) => ({
+          id: m.id,
+          name: m.title,
+          furthestStep: m.furthestStep,
+          createdAt: m.createdAt,
+          updatedAt: m.updatedAt,
+        })),
       );
-      if (!hasUserMessage && step2.step2RoundtableDialogues.length === 0) {
-        return null;
-      }
-      const saved = upsertDraft(buildDraftPayload(furthestStep));
-      setCurrentDraftId(saved.id);
-      setDrafts(listDrafts());
-      return saved;
+    } catch (err) {
+      console.warn("[drafts] 刷新草稿列表失败:", err);
+    }
+  }, []);
+
+  // 首次挂载拉一次草稿列表。
+  useEffect(() => {
+    void refreshList();
+  }, [refreshList]);
+
+  /** 当前会话是否有值得保存的内容。 */
+  const hasSavableContent = useCallback((step1: Step1Snapshot, step2: Step2Snapshot): boolean => {
+    const hasUserMessage = step1.interactiveMessages.some((m) => m.sender === "user");
+    return hasUserMessage || step2.step2RoundtableDialogues.length > 0;
+  }, []);
+
+  /**
+   * 串行执行一次保存：有 id 走 update（不带 title，保护重命名）；无 id 走 create
+   * （写一次派生标题）。create/update 成功后同步 currentDraftIdRef。
+   * 返回 draft id（无可保存内容返回 null）。
+   */
+  const enqueueSave = useCallback(
+    (furthestStep: 1 | 2 | 3): Promise<string | null> => {
+      const run = async (): Promise<string | null> => {
+        const step1 = getStep1Snapshot();
+        const step2 = getStep2Snapshot();
+        if (!hasSavableContent(step1, step2)) return null;
+
+        const id = currentDraftIdRef.current;
+        if (id) {
+          const updated = await updateDraft(id, { furthestStep, step1, step2 });
+          if (updated) return updated.id;
+          // 服务端已不存在（被删/换账号）→ 落回新建。
+        }
+        const created = await createDraft({
+          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>;
     },
-    [buildDraftPayload, getStep1Snapshot, getStep2Snapshot],
+    [getStep1Snapshot, getStep2Snapshot, hasSavableContent, setCurrentDraftId],
+  );
+
+  const ensureDraft = useCallback(
+    async (furthestStep: 1 | 2 | 3 = (currentStep as 1 | 2 | 3) || 1): Promise<string | null> => {
+      if (currentDraftIdRef.current) return currentDraftIdRef.current;
+      const id = await enqueueSave(furthestStep);
+      if (id) await refreshList();
+      return id;
+    },
+    [currentStep, enqueueSave, refreshList],
   );
 
   const handleManualSave = useCallback(() => {
-    const furthest = (currentStep as 1 | 2 | 3) ?? 1;
-    const saved = saveDraft(furthest);
-    if (saved) {
-      triggerToast(`💾 已保存为草稿:${saved.name}`);
-    } else {
-      triggerToast("当前会话还没有可保存的内容");
-    }
-  }, [currentStep, saveDraft, triggerToast]);
+    const furthest = (currentStep as 1 | 2 | 3) || 1;
+    void (async () => {
+      const id = await enqueueSave(furthest);
+      if (id) {
+        await refreshList();
+        triggerToast("💾 已保存为草稿");
+      } else {
+        triggerToast("当前会话还没有可保存的内容");
+      }
+    })();
+  }, [currentStep, enqueueSave, refreshList, triggerToast]);
 
   // 前进切换时自动保存;后退不保存以避免覆盖 furthestStep。
   useEffect(() => {
     const prev = prevStepRef.current;
     prevStepRef.current = currentStep;
     if (currentStep > prev) {
-      saveDraft(currentStep as 1 | 2 | 3);
+      void (async () => {
+        const id = await enqueueSave(currentStep as 1 | 2 | 3);
+        if (id) await refreshList();
+      })();
     }
     // eslint-disable-next-line react-hooks/exhaustive-deps
   }, [currentStep]);
 
   const handleLoadDraft = useCallback(
     (id: string) => {
-      const draft = drafts.find((d) => d.id === id) ?? null;
-      if (!draft) {
-        triggerToast("⚠️ 该草稿已不存在");
-        setDrafts(listDrafts());
-        return;
-      }
+      void (async () => {
+        let draft;
+        try {
+          draft = await getDraft(id);
+        } catch (err) {
+          console.warn("[drafts] 加载草稿失败:", err);
+          triggerToast("⚠️ 加载草稿失败,请重试");
+          return;
+        }
+        if (!draft) {
+          triggerToast("⚠️ 该草稿已不存在");
+          await refreshList();
+          return;
+        }
 
-      cancelAllStreams();
-      // 先置 latch,再 setCurrentStep —— 顺序反过来会让 step1/step2 init effect
-      // 把 loadedFromDraftRef.current 当成 false 处理,导致角色面板卡在「正在上线」。
-      loadedFromDraftRef.current = true;
+        cancelAllStreams();
+        // 先置 latch,再 setCurrentStep —— 顺序反过来会让 step1/step2 init effect
+        // 把 loadedFromDraftRef.current 当成 false 处理,导致角色面板卡在「正在上线」。
+        loadedFromDraftRef.current = true;
 
-      hydrateStep1(draft.step1 as Step1Snapshot);
-      hydrateStep2(draft.step2 as Step2Snapshot);
-      setStep2InitGate(draft.furthestStep >= 2);
+        if (draft.step1) hydrateStep1(draft.step1 as Step1Snapshot);
+        if (draft.step2) hydrateStep2(draft.step2 as Step2Snapshot);
+        setStep2InitGate(draft.furthestStep >= 2);
 
-      setCurrentDraftId(draft.id);
-      setIsHistoryOpen(false);
-      prevStepRef.current = draft.furthestStep;
-      setCurrentStep(draft.furthestStep);
-      triggerToast(`📂 已加载草稿:${draft.name}`);
+        setCurrentDraftId(draft.id);
+        setIsHistoryOpen(false);
+        prevStepRef.current = draft.furthestStep;
+        setCurrentStep(draft.furthestStep);
+        triggerToast(`📂 已加载草稿:${draft.title}`);
+      })();
     },
     [
       cancelAllStreams,
-      drafts,
       hydrateStep1,
       hydrateStep2,
+      refreshList,
+      setCurrentDraftId,
       setCurrentStep,
       setStep2InitGate,
       triggerToast,
@@ -191,31 +277,54 @@ export function useDraftPersistence(
       if (typeof window !== "undefined" && !window.confirm("确定要删除这条历史记录吗?此操作不可恢复。")) {
         return;
       }
-      deleteDraft(id);
-      setDrafts(listDrafts());
-      if (currentDraftId === id) setCurrentDraftId(null);
-      if (renamingDraftId === id) setRenamingDraftId(null);
-      triggerToast("🗑 已删除历史记录");
+      void (async () => {
+        try {
+          await deleteDraft(id);
+        } catch (err) {
+          console.warn("[drafts] 删除草稿失败:", err);
+          triggerToast("⚠️ 删除失败,请重试");
+          return;
+        }
+        await refreshList();
+        if (currentDraftIdRef.current === id) setCurrentDraftId(null);
+        if (renamingDraftId === id) setRenamingDraftId(null);
+        triggerToast("🗑 已删除历史记录");
+      })();
     },
-    [currentDraftId, renamingDraftId, triggerToast],
+    [refreshList, renamingDraftId, setCurrentDraftId, triggerToast],
   );
 
-  const handleStartRename = useCallback((draft: DraftRecord) => {
+  const handleStartRename = useCallback((draft: DraftListItem) => {
     setRenamingDraftId(draft.id);
     setRenameDraftValue(draft.name);
   }, []);
 
   const handleCommitRename = useCallback(
     (id: string) => {
-      const updated = renameDraft(id, renameDraftValue);
-      if (updated) {
-        setDrafts(listDrafts());
-        triggerToast("✏️ 已重命名");
+      const trimmed = renameDraftValue.trim();
+      if (!trimmed) {
+        setRenamingDraftId(null);
+        setRenameDraftValue("");
+        return;
       }
-      setRenamingDraftId(null);
-      setRenameDraftValue("");
+      void (async () => {
+        try {
+          // 只发 title —— 与自动保存的「不带 title」互补,重命名是唯一改标题的入口。
+          const updated = await updateDraft(id, { title: trimmed });
+          if (updated) {
+            await refreshList();
+            triggerToast("✏️ 已重命名");
+          }
+        } catch (err) {
+          console.warn("[drafts] 重命名失败:", err);
+          triggerToast("⚠️ 重命名失败,请重试");
+        } finally {
+          setRenamingDraftId(null);
+          setRenameDraftValue("");
+        }
+      })();
     },
-    [renameDraftValue, triggerToast],
+    [refreshList, renameDraftValue, triggerToast],
   );
 
   const handleCancelRename = useCallback(() => {
@@ -253,5 +362,6 @@ export function useDraftPersistence(
     handleStartRename,
     handleCommitRename,
     handleCancelRename,
+    ensureDraft,
   };
 }
diff --git a/frontend-web/src/roundtable-planning/pages/RoundtablePlanningPage.tsx b/frontend-web/src/roundtable-planning/pages/RoundtablePlanningPage.tsx
index 5195efce..1aa88409 100644
--- a/frontend-web/src/roundtable-planning/pages/RoundtablePlanningPage.tsx
+++ b/frontend-web/src/roundtable-planning/pages/RoundtablePlanningPage.tsx
@@ -320,6 +320,7 @@ export default function RoundtablePlanningPage() {
         open={isRecommendDialogOpen}
         intent={step1.intentReady}
         model={selectedModel || undefined}
+        draftId={drafts.currentDraftId}
         onCancel={() => setIsRecommendDialogOpen(false)}
         onConfirm={handleConfirmAgents}
       />
@@ -467,7 +468,9 @@ 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;
                 }
-                setIsRecommendDialogOpen(true);
+                // 先把草稿落地拿到 draft_id,推荐历史才有归属;失败也照常打开弹窗
+                // (弹窗在无 draftId 时退化为「直接推荐、不存历史」)。
+                void drafts.ensureDraft(1).finally(() => setIsRecommendDialogOpen(true));
               }}
               onManualSave={drafts.handleManualSave}
               onTriggerToast={triggerToast}
diff --git a/frontend-web/src/roundtable-planning/utils/drafts.ts b/frontend-web/src/roundtable-planning/utils/drafts.ts
deleted file mode 100644
index bcb4a6f3..00000000
--- a/frontend-web/src/roundtable-planning/utils/drafts.ts
+++ /dev/null
@@ -1,200 +0,0 @@
-/**
- * Local-only draft storage for /page/roundtable/planning.
- *
- * One draft = one full roundtable session (Step 1 intent + Step 2 multi-agent
- * roundtable). Persisted in `localStorage` so refresh / browser-restart keeps
- * it; cross-device sync is out of scope (would need a backend). The Step 2
- * back-end LangGraph threads are reused on load, so list/preview/download of
- * artifacts in Step 3 continue to work as long as the user hasn't deleted the
- * threads server-side.
- */
-
-import type { IntentSummary } from "@/roundtable-planning/api/intent";
-import type { ThreadIdMap } from "@/roundtable-planning/api/multi-agent";
-
-const STORAGE_KEY = "roundtable-planning:drafts:v1";
-
-/**
- * Snapshot of every piece of UI state needed to resume a session. We capture
- * shapes loosely (`unknown[]`) for the chat / role lists because their inline
- * types in the page component are wide and would force a refactor we don't
- * want as part of the persistence feature.
- */
-export interface DraftStep1Snapshot {
-  intentThreadId: string | null;
-  intentReady: IntentSummary | null;
-  interactiveMessages: unknown[];
-  timelineEntries: unknown[];
-}
-
-export interface DraftStep2Snapshot {
-  threadIds: ThreadIdMap | null;
-  coordinatorName: string | null;
-  step2RoundtableDialogues: unknown[];
-  lastLeaderContent: string;
-  hasConsensus: boolean;
-  consensusPercentage: number;
-  budgetLimit: number;
-  /**
-   * Per-seat stub AI messages used by the sandbox `write-file:` virtual URL
-   * loader. Older drafts won't have this — callers should default to `{}`.
-   */
-  seatStubMessages?: Record<string, unknown[]>;
-}
-
-export interface DraftRecord {
-  id: string;
-  /** Auto-derived task summary (objective → first user message → timestamp). */
-  name: string;
-  createdAt: string;
-  updatedAt: string;
-  /** Furthest step the user had reached when this snapshot was taken. */
-  furthestStep: 1 | 2 | 3;
-  step1: DraftStep1Snapshot;
-  step2: DraftStep2Snapshot;
-}
-
-function isBrowser(): boolean {
-  return typeof window !== "undefined" && typeof window.localStorage !== "undefined";
-}
-
-function readAll(): DraftRecord[] {
-  if (!isBrowser()) return [];
-  try {
-    const raw = window.localStorage.getItem(STORAGE_KEY);
-    if (!raw) return [];
-    const parsed = JSON.parse(raw);
-    if (!Array.isArray(parsed)) return [];
-    return parsed as DraftRecord[];
-  } catch (err) {
-    console.warn("[drafts] failed to parse localStorage payload:", err);
-    return [];
-  }
-}
-
-function writeAll(records: DraftRecord[]): void {
-  if (!isBrowser()) return;
-  try {
-    window.localStorage.setItem(STORAGE_KEY, JSON.stringify(records));
-  } catch (err) {
-    // Quota errors are the realistic failure mode — surface but don't throw,
-    // since the user can keep working with their in-memory state.
-    console.warn("[drafts] failed to write localStorage payload:", err);
-  }
-}
-
-function newId(): string {
-  if (isBrowser() && typeof window.crypto !== "undefined" && typeof window.crypto.randomUUID === "function") {
-    return window.crypto.randomUUID();
-  }
-  return `draft-${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 8)}`;
-}
-
-/** List drafts newest first. */
-export function listDrafts(): DraftRecord[] {
-  return readAll()
-    .slice()
-    .sort((a, b) => (a.updatedAt < b.updatedAt ? 1 : -1));
-}
-
-export function getDraft(id: string): DraftRecord | null {
-  return readAll().find((d) => d.id === id) ?? null;
-}
-
-/**
- * Create-or-update a draft. If `id` is given and exists, that record is
- * overwritten (updatedAt refreshed). Otherwise a new record is inserted.
- * Returns the (re-)saved record so callers can read back the new id.
- */
-export function upsertDraft(
-  payload: Omit<DraftRecord, "createdAt" | "updatedAt" | "id"> & { id?: string },
-): DraftRecord {
-  const records = readAll();
-  const now = new Date().toISOString();
-
-  if (payload.id) {
-    const idx = records.findIndex((d) => d.id === payload.id);
-    if (idx >= 0) {
-      const existing = records[idx];
-      const updated: DraftRecord = {
-        ...existing,
-        name: payload.name,
-        furthestStep: payload.furthestStep,
-        step1: payload.step1,
-        step2: payload.step2,
-        updatedAt: now,
-      };
-      records[idx] = updated;
-      writeAll(records);
-      return updated;
-    }
-  }
-
-  const created: DraftRecord = {
-    id: payload.id ?? newId(),
-    name: payload.name,
-    furthestStep: payload.furthestStep,
-    step1: payload.step1,
-    step2: payload.step2,
-    createdAt: now,
-    updatedAt: now,
-  };
-  records.push(created);
-  writeAll(records);
-  return created;
-}
-
-export function deleteDraft(id: string): void {
-  const records = readAll();
-  const next = records.filter((d) => d.id !== id);
-  if (next.length !== records.length) writeAll(next);
-}
-
-/**
- * Rename a draft. No-op when `id` isn't found or `name` is empty. Returns the
- * updated record so callers can refresh local UI state.
- */
-export function renameDraft(id: string, name: string): DraftRecord | null {
-  const trimmed = name.trim();
-  if (!trimmed) return null;
-  const records = readAll();
-  const idx = records.findIndex((d) => d.id === id);
-  if (idx < 0) return null;
-  const updated: DraftRecord = {
-    ...records[idx],
-    name: trimmed,
-    updatedAt: new Date().toISOString(),
-  };
-  records[idx] = updated;
-  writeAll(records);
-  return updated;
-}
-
-/**
- * Derive a short task-summary name from whatever the user has supplied so far.
- * Preference order: intent agent's structured objective → first user message
- * in Step 1 → ISO timestamp fallback. The result is trimmed to ~40 chars so
- * the dropdown rows stay scannable.
- */
-export function deriveDraftName(args: {
-  intentReady?: IntentSummary | null;
-  interactiveMessages?: unknown[];
-}): string {
-  const MAX = 40;
-  const truncate = (s: string) => (s.length > MAX ? s.slice(0, MAX) + "…" : s);
-
-  const objective = args.intentReady?.objective?.trim();
-  if (objective) return truncate(objective);
-
-  if (Array.isArray(args.interactiveMessages)) {
-    for (const m of args.interactiveMessages) {
-      if (!m || typeof m !== "object") continue;
-      const obj = m as Record<string, unknown>;
-      if (obj.sender === "user" && typeof obj.content === "string" && obj.content.trim()) {
-        return truncate(obj.content.trim());
-      }
-    }
-  }
-
-  return `未命名草稿 ${new Date().toLocaleString("zh-CN", { hour12: false })}`;
-}
diff --git a/offline-backend-20260512/backend/CLAUDE.md b/offline-backend-20260512/backend/CLAUDE.md
index c269059c..832ded19 100644
--- a/offline-backend-20260512/backend/CLAUDE.md
+++ b/offline-backend-20260512/backend/CLAUDE.md
@@ -244,6 +244,7 @@ FastAPI application on port 8001 with health check at `GET /health`. Set `GATEWA
 | **Runs** (`/api/runs`) | `POST /stream` - stateless run + SSE; `POST /wait` - stateless run + block; `GET /{rid}/messages` - paginated messages by run_id `{data, has_more}` (cursor: `after_seq`/`before_seq`); `GET /{rid}/feedback` - list feedback by run_id |
 | **Thread Shares** (`/api/thread-shares`) | Per-conversation sharing with two modes (`thread_shares.mode`). **`import`**: `POST /` `{thread_id, mode}` - create (or reuse) a revocable share code; `GET /{code}/preview` + `POST /{code}/import` - copy that conversation into the caller's account (fresh `thread_id` via `app/gateway/thread_copy.py` + new `threads_meta` row + `user-data` copy; dedup on `metadata.shared_origin`). **`view`**: a public, read-only static page — see Public Shares below. Both modes: `GET /` - list my codes; `DELETE /{code}` - revoke. |
 | **Public Shares** (`/api/public/shares`) | **Unauthenticated** (the `/api/public/` prefix is whitelisted in `auth_middleware._PUBLIC_PATH_PREFIXES`). `GET /{code}` - read-only conversation snapshot (title + serialized messages + artifacts) for a `view`-mode share; messages are read straight from the latest checkpoint. `GET /{code}/files/{path}` - serves a file referenced by that conversation (images, Markdown, text, PDF, …) so the shared page's links are openable; path resolution is confined to the owner's thread directory. Active content (`text/html`, `application/xhtml+xml`, `image/svg+xml`) is forced to download (`_public_share_disposition`); everything else is served inline. Missing/revoked/non-`view` codes all return 404. |
+| **Roundtable Drafts** (`/api/roundtable-drafts`) | Per-user persistence for the roundtable-planning page (replaces the old browser `localStorage`). `GET /` - list drafts (lightweight meta); `POST /` - create; `GET/PUT/DELETE /{id}` - full read / partial update / delete. A draft = one full session (Step 1 intent + Step 2 roundtable), step snapshots stored as JSON in `PortableLongText`. Recommendation history bound to a draft: `GET /{id}/recommendations` - list newest-first; `POST /{id}/recommendations` - append one run (objective/status/rationale/picks/candidates). Backed by `deerflow.persistence.roundtable_drafts` (tables `roundtable_drafts`, `roundtable_recommend_history`); store wired in `deps.py` as `app.state.roundtable_draft_store`. |
 | **Scheduled Tasks** (`/api/scheduled-tasks`) | Per-user cron-like agent tasks. `GET/POST /` - list/create owned tasks; `PATCH/DELETE /{id}` - update/delete (owner); `POST /{id}/{pause,resume,trigger}`; `GET /{id}/runs`, `GET /runs/{rid}`, run-file preview/edit/download; `GET /runs/{rid}/messages` - agent messages for the run's underlying agent run (used to inspect a stuck/running run; authorized via `_load_run`); `DELETE /runs/{rid}` - owner-scoped, with an admin fallback so admins can delete any user's run; publish/subscribe under `/{id}/{publish,unpublish,subscribe}` and `GET /published`. **Admin (cross-user):** `GET /admin/all` - every task; `PATCH/DELETE /admin/{id}` - edit/delete any task; `POST /admin/{id}/trigger`; `GET /admin/{id}/runs`. Admin routes require `system_role == "admin"`. |
 
 Proxied through nginx: `/api/langgraph/*` → LangGraph, all other `/api/*` → Gateway.
diff --git a/offline-backend-20260512/backend/app/gateway/app.py b/offline-backend-20260512/backend/app/gateway/app.py
index 50319e12..b720496d 100644
--- a/offline-backend-20260512/backend/app/gateway/app.py
+++ b/offline-backend-20260512/backend/app/gateway/app.py
@@ -42,6 +42,7 @@ from app.gateway.routers import (
     public_scheduled_tasks,
     recommend,
     recommended_questions,
+    roundtable_drafts,
     runs,
     scheduled_tasks,
     skills,
@@ -833,6 +834,9 @@ This gateway provides custom endpoints for models, MCP configuration, skills, an
 
     app.include_router(user_prompts.router)
 
+    # Roundtable-planning drafts + recommendation history at /api/roundtable-drafts
+    app.include_router(roundtable_drafts.router)
+
     # Suggestions API is mounted at /api/threads/{thread_id}/suggestions
     app.include_router(suggestions.router)
 
diff --git a/offline-backend-20260512/backend/app/gateway/deps.py b/offline-backend-20260512/backend/app/gateway/deps.py
index 741c345f..27fb52ac 100644
--- a/offline-backend-20260512/backend/app/gateway/deps.py
+++ b/offline-backend-20260512/backend/app/gateway/deps.py
@@ -152,6 +152,10 @@ async def langgraph_runtime(app: FastAPI) -> AsyncGenerator[None, None]:
 
         app.state.user_prompt_store = make_user_prompt_store(sf)
 
+        from deerflow.persistence.roundtable_drafts import make_roundtable_draft_store
+
+        app.state.roundtable_draft_store = make_roundtable_draft_store(sf)
+
         # Run event store (has its own factory with config-driven backend selection)
         run_events_config = getattr(config, "run_events", None)
         app.state.run_event_store = make_run_event_store(run_events_config)
diff --git a/offline-backend-20260512/backend/app/gateway/routers/roundtable_drafts.py b/offline-backend-20260512/backend/app/gateway/routers/roundtable_drafts.py
new file mode 100644
index 00000000..5834b9b2
--- /dev/null
+++ b/offline-backend-20260512/backend/app/gateway/routers/roundtable_drafts.py
@@ -0,0 +1,223 @@
+"""CRUD API for roundtable-planning drafts + recommendation history.
+
+A draft is one full roundtable session (Step 1 intent + Step 2 multi-agent
+roundtable). Bound to the requesting user (not the browser), so it follows the
+account across devices/sessions — replacing the old ``localStorage`` storage.
+
+Recommendation history is bound to a draft: the recommend dialog shows "last
+recommendation for this session" on open, while still letting the user
+re-analyze (which appends a new history row).
+
+Routes (prefix ``/api/roundtable-drafts``):
+    GET    ""                         list current user's drafts (lightweight)
+    POST   ""                         create a draft
+    GET    "/{draft_id}"              full draft (incl. step1 / step2 blobs)
+    PUT    "/{draft_id}"              partial update (title / step / snapshots)
+    DELETE "/{draft_id}"              delete
+    GET    "/{draft_id}/recommendations"   list recommendation history
+    POST   "/{draft_id}/recommendations"   append one recommendation run
+"""
+
+from __future__ import annotations
+
+import logging
+from datetime import datetime
+from typing import Any
+from uuid import uuid4
+
+from fastapi import APIRouter, HTTPException, Request
+from pydantic import BaseModel, Field
+
+from deerflow.runtime.user_context import get_effective_user_id
+
+logger = logging.getLogger(__name__)
+router = APIRouter(prefix="/api/roundtable-drafts", tags=["roundtable-drafts"])
+
+
+# ── schemas ────────────────────────────────────────────────────────────────
+
+
+class DraftMetaResponse(BaseModel):
+    id: str
+    title: str = ""
+    furthest_step: int = 1
+    created_at: datetime | str | None = None
+    updated_at: datetime | str | None = None
+
+
+class DraftListResponse(BaseModel):
+    drafts: list[DraftMetaResponse]
+
+
+class DraftResponse(DraftMetaResponse):
+    step1: dict[str, Any] | None = None
+    step2: dict[str, Any] | None = None
+
+
+class DraftCreateRequest(BaseModel):
+    # Optional client-provided id so the browser can keep one stable id across
+    # the create + subsequent updates without an extra round-trip.
+    id: str | None = Field(default=None, max_length=64)
+    title: str = Field(default="", max_length=512)
+    furthest_step: int = Field(default=1, ge=1, le=3)
+    step1: dict[str, Any] | None = None
+    step2: dict[str, Any] | None = None
+
+
+class DraftUpdateRequest(BaseModel):
+    # All optional → callers send only what changed. Auto-save omits ``title``
+    # so it never clobbers a manual rename.
+    title: str | None = Field(default=None, max_length=512)
+    furthest_step: int | None = Field(default=None, ge=1, le=3)
+    step1: dict[str, Any] | None = None
+    step2: dict[str, Any] | None = None
+
+
+class RecommendPickSchema(BaseModel):
+    agent_id: str
+    reason: str = ""
+
+
+class RecommendCandidateSchema(BaseModel):
+    agent_id: str
+    name: str = ""
+    description: str = ""
+
+
+class RecommendHistoryResponse(BaseModel):
+    id: str
+    draft_id: str
+    objective: str = ""
+    status: str = "done"
+    model: str | None = None
+    rationale: str | None = None
+    picks: list[RecommendPickSchema] = []
+    candidates: list[RecommendCandidateSchema] = []
+    created_at: datetime | str | None = None
+
+
+class RecommendHistoryListResponse(BaseModel):
+    history: list[RecommendHistoryResponse]
+
+
+class RecommendHistoryCreateRequest(BaseModel):
+    objective: str = Field(default="", max_length=512)
+    status: str = Field(default="done", max_length=32)
+    model: str | None = None
+    rationale: str | None = None
+    picks: list[RecommendPickSchema] = []
+    candidates: list[RecommendCandidateSchema] = []
+
+
+# ── helpers ──────────────────────────────────────────────────────────────
+
+
+def _current_user_id(request: Request) -> str:
+    user = getattr(request.state, "user", None)
+    if user is not None:
+        return str(user.id)
+    return get_effective_user_id()
+
+
+def _get_store(request: Request):
+    store = getattr(request.app.state, "roundtable_draft_store", None)
+    if store is None:
+        raise HTTPException(status_code=503, detail="Roundtable draft store not available")
+    return store
+
+
+# ── draft routes ───────────────────────────────────────────────────────────
+
+
+@router.get("", response_model=DraftListResponse)
+async def list_drafts(request: Request) -> DraftListResponse:
+    store = _get_store(request)
+    user_id = _current_user_id(request)
+    rows = await store.list_drafts(user_id)
+    return DraftListResponse(drafts=[DraftMetaResponse(**r) for r in rows])
+
+
+@router.post("", response_model=DraftResponse, status_code=201)
+async def create_draft(request: Request, body: DraftCreateRequest) -> DraftResponse:
+    store = _get_store(request)
+    user_id = _current_user_id(request)
+    row = await store.create_draft(
+        user_id,
+        {
+            "id": (body.id or uuid4().hex),
+            "title": body.title.strip(),
+            "furthest_step": body.furthest_step,
+            "step1": body.step1,
+            "step2": body.step2,
+        },
+    )
+    return DraftResponse(**row)
+
+
+@router.get("/{draft_id}", response_model=DraftResponse)
+async def get_draft(request: Request, draft_id: str) -> DraftResponse:
+    store = _get_store(request)
+    user_id = _current_user_id(request)
+    row = await store.get_draft(draft_id, user_id)
+    if row is None:
+        raise HTTPException(status_code=404, detail="Draft not found")
+    return DraftResponse(**row)
+
+
+@router.put("/{draft_id}", response_model=DraftResponse)
+async def update_draft(request: Request, draft_id: str, body: DraftUpdateRequest) -> DraftResponse:
+    store = _get_store(request)
+    user_id = _current_user_id(request)
+    # Build kwargs that distinguish "omitted" from "explicitly None". The store
+    # uses a sentinel; here we pass only the fields the client actually sent.
+    fields = body.model_dump(exclude_unset=True)
+    row = await store.update_draft(draft_id, user_id, **fields)
+    if row is None:
+        raise HTTPException(status_code=404, detail="Draft not found")
+    return DraftResponse(**row)
+
+
+@router.delete("/{draft_id}", status_code=204)
+async def delete_draft(request: Request, draft_id: str) -> None:
+    store = _get_store(request)
+    user_id = _current_user_id(request)
+    deleted = await store.delete_draft(draft_id, user_id)
+    if not deleted:
+        raise HTTPException(status_code=404, detail="Draft not found")
+
+
+# ── recommendation history routes ────────────────────────────────────────
+
+
+@router.get("/{draft_id}/recommendations", response_model=RecommendHistoryListResponse)
+async def list_recommendations(request: Request, draft_id: str) -> RecommendHistoryListResponse:
+    store = _get_store(request)
+    user_id = _current_user_id(request)
+    rows = await store.list_recommendations(draft_id, user_id)
+    return RecommendHistoryListResponse(history=[RecommendHistoryResponse(**r) for r in rows])
+
+
+@router.post("/{draft_id}/recommendations", response_model=RecommendHistoryResponse, status_code=201)
+async def add_recommendation(
+    request: Request, draft_id: str, body: RecommendHistoryCreateRequest
+) -> RecommendHistoryResponse:
+    store = _get_store(request)
+    user_id = _current_user_id(request)
+    # Ownership: the draft must exist and belong to the caller.
+    draft = await store.get_draft(draft_id, user_id)
+    if draft is None:
+        raise HTTPException(status_code=404, detail="Draft not found")
+    row = await store.add_recommendation(
+        user_id,
+        draft_id,
+        {
+            "id": uuid4().hex,
+            "objective": body.objective,
+            "status": body.status,
+            "model": body.model,
+            "rationale": body.rationale,
+            "picks": [p.model_dump() for p in body.picks],
+            "candidates": [c.model_dump() for c in body.candidates],
+        },
+    )
+    return RecommendHistoryResponse(**row)
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/persistence/models/__init__.py b/offline-backend-20260512/backend/packages/harness/deerflow/persistence/models/__init__.py
index c5ca237c..9766c380 100644
--- a/offline-backend-20260512/backend/packages/harness/deerflow/persistence/models/__init__.py
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/persistence/models/__init__.py
@@ -22,6 +22,10 @@ from deerflow.persistence.llm_metrics.model import LlmCallMetricRow
 from deerflow.persistence.models.run_event import RunEventRow
 from deerflow.persistence.notifications.model import NotificationRow
 from deerflow.persistence.recommended_questions.model import RecommendedQuestionRow
+from deerflow.persistence.roundtable_drafts.model import (
+    RoundtableDraftRow,
+    RoundtableRecommendHistoryRow,
+)
 from deerflow.persistence.run.model import RunRow
 from deerflow.persistence.scheduled_tasks.model import (
     ScheduledTaskDeliveryProfileRow,
@@ -45,6 +49,8 @@ __all__ = [
     "LlmCallMetricRow",
     "NotificationRow",
     "RecommendedQuestionRow",
+    "RoundtableDraftRow",
+    "RoundtableRecommendHistoryRow",
     "RunEventRow",
     "RunRow",
     "ScheduledTaskDeliveryProfileRow",
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/persistence/roundtable_drafts/__init__.py b/offline-backend-20260512/backend/packages/harness/deerflow/persistence/roundtable_drafts/__init__.py
new file mode 100644
index 00000000..0e17516b
--- /dev/null
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/persistence/roundtable_drafts/__init__.py
@@ -0,0 +1,22 @@
+"""Roundtable-planning draft + recommendation-history persistence."""
+
+from __future__ import annotations
+
+from deerflow.persistence.roundtable_drafts.model import (
+    RoundtableDraftRow,
+    RoundtableRecommendHistoryRow,
+)
+from deerflow.persistence.roundtable_drafts.sql import RoundtableDraftRepository
+
+__all__ = [
+    "RoundtableDraftRepository",
+    "RoundtableDraftRow",
+    "RoundtableRecommendHistoryRow",
+    "make_roundtable_draft_store",
+]
+
+
+def make_roundtable_draft_store(session_factory):
+    if session_factory is None:
+        return None
+    return RoundtableDraftRepository(session_factory)
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/persistence/roundtable_drafts/model.py b/offline-backend-20260512/backend/packages/harness/deerflow/persistence/roundtable_drafts/model.py
new file mode 100644
index 00000000..94ee13c9
--- /dev/null
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/persistence/roundtable_drafts/model.py
@@ -0,0 +1,69 @@
+"""ORM models for roundtable-planning drafts and recommendation history.
+
+A "draft" is one full roundtable-planning session (Step 1 intent + Step 2
+multi-agent roundtable), persisted per user so it follows the account across
+devices/sessions — replacing the old browser-only ``localStorage`` storage.
+
+``step1`` / ``step2`` are stored as JSON serialized into ``PortableLongText``
+(MySQL ``LONGTEXT``) rather than ``PortableJSON`` (MySQL ``TEXT``, 64 KB cap),
+because the Step 2 snapshot can carry a long roundtable transcript plus per-seat
+stub messages that easily exceed 64 KB — mirroring ``ai_writing_sessions``.
+"""
+
+from __future__ import annotations
+
+from datetime import UTC, datetime
+
+from sqlalchemy import Integer, String
+from sqlalchemy.orm import Mapped, mapped_column
+
+from deerflow.persistence.base import Base
+from deerflow.persistence.types import BeijingDateTime, PortableLongText
+
+
+class RoundtableDraftRow(Base):
+    __tablename__ = "roundtable_drafts"
+
+    id: Mapped[str] = mapped_column(String(64), primary_key=True)
+    user_id: Mapped[str] = mapped_column(String(64), nullable=False, index=True)
+    # Auto-derived task summary (objective → first user message → timestamp).
+    title: Mapped[str] = mapped_column(String(512), nullable=False, default="")
+    # Furthest step reached when this snapshot was taken (1 | 2 | 3).
+    furthest_step: Mapped[int] = mapped_column(Integer, nullable=False, default=1)
+    # JSON blobs matching the frontend Step1Snapshot / Step2Snapshot shapes.
+    step1: Mapped[str | None] = mapped_column(PortableLongText, nullable=True)  # JSON object
+    step2: Mapped[str | None] = mapped_column(PortableLongText, nullable=True)  # JSON object
+    created_at: Mapped[datetime] = mapped_column(BeijingDateTime(), nullable=False, default=lambda: datetime.now(UTC))
+    updated_at: Mapped[datetime] = mapped_column(
+        BeijingDateTime(),
+        nullable=False,
+        default=lambda: datetime.now(UTC),
+        onupdate=lambda: datetime.now(UTC),
+    )
+
+
+class RoundtableRecommendHistoryRow(Base):
+    """One agent-recommendation run, bound to a draft/session.
+
+    Lets the recommend dialog show "last recommendation for this session" on
+    open, while still letting the user re-analyze (which appends a new row).
+    """
+
+    __tablename__ = "roundtable_recommend_history"
+
+    id: Mapped[str] = mapped_column(String(64), primary_key=True)
+    user_id: Mapped[str] = mapped_column(String(64), nullable=False, index=True)
+    draft_id: Mapped[str] = mapped_column(String(64), nullable=False, index=True)
+    # Denormalized intent objective at recommend time (list display / context).
+    objective: Mapped[str] = mapped_column(String(512), nullable=False, default="")
+    # How the run ended on the backend: "done" | "fallback" | "asking" | "error".
+    status: Mapped[str] = mapped_column(String(32), nullable=False, default="done")
+    model: Mapped[str | None] = mapped_column(String(128), nullable=True)
+    # Natural-language rationale (may be long → LONGTEXT).
+    rationale: Mapped[str | None] = mapped_column(PortableLongText, nullable=True)
+    # JSON: [{"agent_id": str, "reason": str}, ...]
+    picks: Mapped[str | None] = mapped_column(PortableLongText, nullable=True)
+    # JSON snapshot of the candidate pool at recommend time:
+    # [{"agent_id": str, "name": str, "description": str}, ...]
+    candidates: Mapped[str | None] = mapped_column(PortableLongText, nullable=True)
+    created_at: Mapped[datetime] = mapped_column(BeijingDateTime(), nullable=False, default=lambda: datetime.now(UTC))
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/persistence/roundtable_drafts/sql.py b/offline-backend-20260512/backend/packages/harness/deerflow/persistence/roundtable_drafts/sql.py
new file mode 100644
index 00000000..bf668b42
--- /dev/null
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/persistence/roundtable_drafts/sql.py
@@ -0,0 +1,191 @@
+"""SQLAlchemy-backed roundtable-planning draft + recommendation-history store.
+
+JSON columns (``step1`` / ``step2`` / ``picks`` / ``candidates``) are stored as
+serialized strings in ``PortableLongText`` and round-tripped here with
+``json.dumps`` / ``json.loads`` — same approach as ``ai_writing_sessions`` — so
+large Step 2 transcripts never hit MySQL's 64 KB ``TEXT`` cap.
+
+All write/read paths are scoped by ``user_id`` for ownership isolation.
+"""
+
+from __future__ import annotations
+
+import json
+from datetime import UTC, datetime
+from typing import Any
+
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker
+
+from deerflow.persistence.roundtable_drafts.model import (
+    RoundtableDraftRow,
+    RoundtableRecommendHistoryRow,
+)
+
+_UNSET = object()
+
+
+def _loads(value: Any) -> Any:
+    if not isinstance(value, str):
+        return value
+    if not value:
+        return None
+    try:
+        return json.loads(value)
+    except Exception:
+        return None
+
+
+def _dumps(value: Any) -> str | None:
+    if value is None:
+        return None
+    return json.dumps(value, ensure_ascii=False)
+
+
+class RoundtableDraftRepository:
+    def __init__(self, session_factory: async_sessionmaker[AsyncSession]) -> None:
+        self._sf = session_factory
+
+    # ── drafts ────────────────────────────────────────────────────────────
+
+    @staticmethod
+    def _draft_to_meta(row: RoundtableDraftRow) -> dict[str, Any]:
+        """Lightweight projection for the history dropdown (no big blobs)."""
+        return {
+            "id": row.id,
+            "title": row.title,
+            "furthest_step": row.furthest_step,
+            "created_at": row.created_at.isoformat() if isinstance(row.created_at, datetime) else row.created_at,
+            "updated_at": row.updated_at.isoformat() if isinstance(row.updated_at, datetime) else row.updated_at,
+        }
+
+    @classmethod
+    def _draft_to_dict(cls, row: RoundtableDraftRow) -> dict[str, Any]:
+        d = cls._draft_to_meta(row)
+        d["step1"] = _loads(row.step1)
+        d["step2"] = _loads(row.step2)
+        return d
+
+    async def list_drafts(self, user_id: str, *, limit: int = 100) -> list[dict[str, Any]]:
+        stmt = (
+            select(RoundtableDraftRow)
+            .where(RoundtableDraftRow.user_id == user_id)
+            .order_by(RoundtableDraftRow.updated_at.desc())
+            .limit(limit)
+        )
+        async with self._sf() as session:
+            result = await session.execute(stmt)
+            return [self._draft_to_meta(row) for row in result.scalars()]
+
+    async def get_draft(self, draft_id: str, user_id: str) -> dict[str, Any] | None:
+        async with self._sf() as session:
+            row = await session.get(RoundtableDraftRow, draft_id)
+            if row is None or row.user_id != user_id:
+                return None
+            return self._draft_to_dict(row)
+
+    async def create_draft(self, user_id: str, data: dict[str, Any]) -> dict[str, Any]:
+        now = datetime.now(UTC)
+        row = RoundtableDraftRow(
+            id=data["id"],
+            user_id=user_id,
+            title=(data.get("title") or "").strip(),
+            furthest_step=int(data.get("furthest_step") or 1),
+            step1=_dumps(data.get("step1")),
+            step2=_dumps(data.get("step2")),
+            created_at=now,
+            updated_at=now,
+        )
+        async with self._sf() as session:
+            session.add(row)
+            await session.commit()
+            await session.refresh(row)
+            return self._draft_to_dict(row)
+
+    async def update_draft(
+        self,
+        draft_id: str,
+        user_id: str,
+        *,
+        title=_UNSET,
+        furthest_step=_UNSET,
+        step1=_UNSET,
+        step2=_UNSET,
+    ) -> dict[str, Any] | None:
+        async with self._sf() as session:
+            row = await session.get(RoundtableDraftRow, draft_id)
+            if row is None or row.user_id != user_id:
+                return None
+            if title is not _UNSET and title is not None:
+                row.title = str(title).strip()
+            if furthest_step is not _UNSET and furthest_step is not None:
+                row.furthest_step = int(furthest_step)
+            if step1 is not _UNSET:
+                row.step1 = _dumps(step1)
+            if step2 is not _UNSET:
+                row.step2 = _dumps(step2)
+            row.updated_at = datetime.now(UTC)
+            await session.commit()
+            await session.refresh(row)
+            return self._draft_to_dict(row)
+
+    async def delete_draft(self, draft_id: str, user_id: str) -> bool:
+        async with self._sf() as session:
+            row = await session.get(RoundtableDraftRow, draft_id)
+            if row is None or row.user_id != user_id:
+                return False
+            await session.delete(row)
+            await session.commit()
+            return True
+
+    # ── recommendation history ────────────────────────────────────────────
+
+    @staticmethod
+    def _history_to_dict(row: RoundtableRecommendHistoryRow) -> dict[str, Any]:
+        return {
+            "id": row.id,
+            "draft_id": row.draft_id,
+            "objective": row.objective,
+            "status": row.status,
+            "model": row.model,
+            "rationale": row.rationale,
+            "picks": _loads(row.picks) or [],
+            "candidates": _loads(row.candidates) or [],
+            "created_at": row.created_at.isoformat() if isinstance(row.created_at, datetime) else row.created_at,
+        }
+
+    async def list_recommendations(
+        self, draft_id: str, user_id: str, *, limit: int = 20
+    ) -> list[dict[str, Any]]:
+        stmt = (
+            select(RoundtableRecommendHistoryRow)
+            .where(
+                RoundtableRecommendHistoryRow.draft_id == draft_id,
+                RoundtableRecommendHistoryRow.user_id == user_id,
+            )
+            .order_by(RoundtableRecommendHistoryRow.created_at.desc())
+            .limit(limit)
+        )
+        async with self._sf() as session:
+            result = await session.execute(stmt)
+            return [self._history_to_dict(row) for row in result.scalars()]
+
+    async def add_recommendation(self, user_id: str, draft_id: str, data: dict[str, Any]) -> dict[str, Any]:
+        now = datetime.now(UTC)
+        row = RoundtableRecommendHistoryRow(
+            id=data["id"],
+            user_id=user_id,
+            draft_id=draft_id,
+            objective=(data.get("objective") or "").strip()[:512],
+            status=(data.get("status") or "done")[:32],
+            model=(data.get("model") or None),
+            rationale=data.get("rationale") or None,
+            picks=_dumps(data.get("picks") or []),
+            candidates=_dumps(data.get("candidates") or []),
+            created_at=now,
+        )
+        async with self._sf() as session:
+            session.add(row)
+            await session.commit()
+            await session.refresh(row)
+            return self._history_to_dict(row)