diff --git a/api_doc/README.md b/api_doc/README.md
new file mode 100644
index 0000000..a0ef1d6
--- /dev/null
+++ b/api_doc/README.md
@@ -0,0 +1,127 @@
+# API v1 接入文档（后台任务 + 轮询进度 + 文件）
+
+本目录面向**第三方/多端**（Web、Flutter、脚本、其它服务）对接，目标是：
+
+- API 调用不依赖 WebSocket：网页刷新/断网/关闭不会影响后台任务运行；
+- 客户端使用 **HTTP 轮询**获取与网页端一致的细粒度进度（token 输出、工具块、工具结果等）；
+- API 账号与网页账号隔离：避免互相抢占任务/文件/对话状态。
+
+## 基本信息
+
+- API 版本：v1
+- 默认后端地址（本地）：`http://localhost:8091`
+- 鉴权方式：`Authorization: Bearer <token>`
+- API 前缀：`/api/v1`
+- 数据落盘：对话与文件落盘到 API 用户独立目录（见 `auth.md`）
+- 任务与事件：**仅内存**保存（进程重启后任务/事件不可恢复）
+- 默认最大迭代次数 `max_iterations`：**100**（可在调用 `/api/v1/messages` 时覆盖）
+
+## 只保留的接口（共 6 组能力）
+
+1. 创建对话：`POST /api/v1/conversations`
+2. 发送消息/启动后台任务：`POST /api/v1/messages`
+3. 轮询任务事件：`GET /api/v1/tasks/<task_id>?from=<offset>`
+4. 停止任务：`POST /api/v1/tasks/<task_id>/cancel`
+5. 文件上传（仅 user_upload）：`POST /api/v1/files/upload`
+6. 文件浏览/下载：`GET /api/v1/files`、`GET /api/v1/files/download`
+
+详细参数与返回请看：
+
+- `auth.md`：API 用户与 Token、目录结构、安全注意事项
+- `messages_tasks.md`：发送消息/轮询/停止
+- `events.md`：事件流格式与事件类型说明（与 WebSocket 同源）
+- `files.md`：上传/列目录/下载
+- `errors.md`：HTTP 错误码与常见排查
+- `examples.md`：curl/Python/JS/Flutter 示例
+- `openapi.yaml`：OpenAPI 3.0 规范（可导入 Postman/Swagger）
+
+## 快速开始（curl）
+
+> 将 `<TOKEN>` 替换为你的 Bearer Token。  
+> 默认后端：`http://localhost:8091`，如不同请修改。
+
+### 0）创建对话
+
+```bash
+curl -sS -X POST \
+  -H "Authorization: Bearer <TOKEN>" \
+  http://localhost:8091/api/v1/conversations
+```
+
+返回示例：
+
+```json
+{ "success": true, "conversation_id": "conv_20260123_234245_036" }
+```
+
+### 1）发送消息（创建后台任务）
+
+```bash
+curl -sS -X POST \
+  -H "Authorization: Bearer <TOKEN>" \
+  -H "Content-Type: application/json" \
+  -d '{
+    "conversation_id": "conv_20260123_234245_036",
+    "message": "请用中文简要介绍《明日方舟：终末地》",
+    "run_mode": "fast",            # 可选：fast / thinking / deep
+    "model_key": null,
+    "max_iterations": 100         # 默认 100，示例显式填写便于对齐
+  }' \
+  http://localhost:8091/api/v1/messages
+```
+
+返回示例（202）：
+
+```json
+{
+  "success": true,
+  "task_id": "60322db3-f884-4a1e-a9b3-6eeb07fbab47",
+  "conversation_id": "conv_20260123_234245_036",
+  "status": "running",
+  "created_at": 1769182965.30
+}
+```
+
+### 2）轮询进度（按 offset 增量拉取）
+
+首次从 `from=0` 开始，之后使用返回中的 `next_offset`。
+
+```bash
+curl -sS \
+  -H "Authorization: Bearer <TOKEN>" \
+  "http://localhost:8091/api/v1/tasks/60322db3-f884-4a1e-a9b3-6eeb07fbab47?from=0"
+```
+
+### 3）停止任务（可选）
+
+```bash
+curl -sS -X POST \
+  -H "Authorization: Bearer <TOKEN>" \
+  http://localhost:8091/api/v1/tasks/60322db3-f884-4a1e-a9b3-6eeb07fbab47/cancel
+```
+
+## 统一约定：响应结构与错误处理
+
+### 成功响应
+
+大部分接口返回：
+
+```json
+{ "success": true, ... }
+```
+
+### 失败响应
+
+大部分失败返回：
+
+```json
+{ "success": false, "error": "原因说明" }
+```
+
+并使用合适的 HTTP 状态码（如 400/401/404/409/503）。
+
+## 重要限制（务必阅读）
+
+- **单用户禁止并发任务**：同一个 API 用户同一时间只允许一个 `running/pending` 任务。重复发送消息会返回 `409`。
+- **事件缓冲为内存队列（maxlen=1000）**：任务特别长或轮询太慢会导致早期事件被丢弃；请按推荐频率轮询并在客户端持久化你需要的内容。
+- **进程重启不可恢复**：重启后任务/事件会消失，但对话/文件已落盘的不受影响。
diff --git a/api_doc/auth.md b/api_doc/auth.md
new file mode 100644
index 0000000..255280f
--- /dev/null
+++ b/api_doc/auth.md
@@ -0,0 +1,119 @@
+# 认证与 API 用户（JSON 手动维护）
+
+本项目的 API v1 使用 **Bearer Token** 鉴权，且 **API 用户与网页用户完全隔离**。
+
+## 1. 鉴权方式（每个请求都要带）
+
+请求头：
+
+```
+Authorization: Bearer <TOKEN>
+```
+
+服务端会对 `<TOKEN>` 做 `SHA256`，与 `data/api_users.json` 中保存的 `token_sha256` 进行匹配。
+
+## 2. API 用户配置文件
+
+路径：
+
+- `data/api_users.json`
+
+格式（示例）：
+
+```json
+{
+  "users": {
+    "api_jojo": {
+      "token_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "created_at": "2026-01-23",
+      "note": "for flutter app"
+    }
+  }
+}
+```
+
+字段说明：
+
+- `users`：对象，key 是 API 用户名（建议全小写、只包含字母数字下划线/连字符）。
+- `token_sha256`：必填，token 的 SHA256 十六进制字符串。
+- `created_at` / `note`：可选，仅用于记录与审计。
+
+注意：
+
+- 当前实现为**启动时加载**该 JSON；如果你修改了 `data/api_users.json`，请**重启服务**使其生效。
+
+## 3. 生成 token_sha256 的方法
+
+### 3.1 用 Python 生成
+
+```bash
+python3 - <<'PY'
+import hashlib
+token = "YOUR_TOKEN"
+print(hashlib.sha256(token.encode("utf-8")).hexdigest())
+PY
+```
+
+把输出粘到 `token_sha256`。
+
+### 3.2 token 建议
+
+- 长度建议：>= 32 字符（越长越好）
+- 生成建议：使用密码管理器或安全随机源生成
+- 存储建议：客户端侧使用系统安全存储（Keychain/Keystore），不要写死在前端代码里
+
+## 4. API 用户的独立落盘目录
+
+每个 API 用户会创建独立工作区，默认根目录：
+
+- `api/users/<api_username>/`
+
+典型结构：
+
+```
+api/users/<api_username>/
+  project/
+    user_upload/              # API 上传文件只能落在这里（以及其子目录）
+    ...                       # 智能体运行产生的项目文件也在 project 下
+  data/
+    conversations/            # 对话 JSON 落盘目录
+    memory.md
+    task_memory.md
+    ...
+  logs/
+```
+
+对话文件示例路径：
+
+- `api/users/api_jojo/data/conversations/conv_20260123_234245_036.json`
+
+对话 JSON 结构（简化示例）：
+
+```json
+{
+  "id": "conv_20260123_234245_036",
+  "title": "新对话",
+  "created_at": "2026-01-23T23:42:45.036Z",
+  "updated_at": "...",
+  "messages": [
+    { "role": "user", "content": "你好", "timestamp": "..." },
+    { "role": "assistant", "content": "…", "reasoning_content": "", "timestamp": "..." },
+    { "role": "tool", "name": "web_search", "content": "{...}", "timestamp": "..." }
+  ],
+  "metadata": { "run_mode": "fast", "model_key": "kimi", "...": "..." }
+}
+```
+
+说明：
+
+- API v1 **没有提供“读取历史对话”的 HTTP 接口**；如确实需要历史，请直接读落盘文件或后续再加只读接口。
+
+## 5. 安全注意事项
+
+- Bearer Token 等同密码：不要放到 Git、不要发到日志、不要在网页端暴露。
+- 如果怀疑泄露：替换 token（更新 JSON），重启服务；旧 token 会立即失效。
+- 生产环境建议：
+  - 通过反向代理限制来源 IP；
+  - 记录 API 调用审计日志（至少记录 user、endpoint、时间、返回码）；
+  - 加入速率限制（避免被刷）。
+
diff --git a/api_doc/errors.md b/api_doc/errors.md
new file mode 100644
index 0000000..609718d
--- /dev/null
+++ b/api_doc/errors.md
@@ -0,0 +1,70 @@
+# 错误码与常见问题（HTTP 层）
+
+本 API 的错误分两类：
+
+1) **HTTP 层错误**：接口直接返回 `success=false`，并用 HTTP 状态码表达错误类型。  
+2) **任务内错误**：接口本身 200，但轮询的 `events` 中出现 `type=error`，且任务状态变为 `failed`。
+
+## 1. HTTP 状态码约定
+
+| 状态码 | 何时出现 | 返回体示例 |
+|---:|---|---|
+| 200 | 成功（大多数 GET/POST） | `{ "success": true, ... }` |
+| 202 | 成功但已进入后台执行（发送消息） | `{ "success": true, "task_id": "...", "status": "running", ... }` |
+| 400 | 参数错误（缺字段/格式不对/路径非法） | `{ "success": false, "error": "..." }` |
+| 401 | 未授权（缺 token 或 token 无效） | `{ "success": false, "error": "..." }` |
+| 404 | 资源不存在（task/file/path 不存在） | `{ "success": false, "error": "..." }` |
+| 409 | 并发冲突（已有 running/pending 任务） | `{ "success": false, "error": "已有运行中的任务，请稍后再试。" }` |
+| 500 | 服务端内部错误（少见） | `{ "success": false, "error": "..." }` |
+| 503 | 系统未初始化/资源繁忙（容器不可用等） | `{ "success": false, "error": "..." }` |
+
+## 2. 常见错误与排查
+
+### 2.1 401 Unauthorized
+
+原因：
+
+- 忘记带 `Authorization: Bearer ...`
+- token 不存在或已被替换
+
+排查：
+
+- 检查 `data/api_users.json` 是否包含该 token 的 SHA256
+- 修改 JSON 后是否重启服务（当前为启动时加载）
+
+### 2.2 409 已有运行中的任务
+
+原因：
+
+- 同一个 API 用户在任务未结束时再次调用 `/api/v1/messages`
+
+处理建议：
+
+- 客户端 UI：禁用“发送”按钮，直到轮询显示 `status != running`
+- 或先调用 `/api/v1/tasks/<task_id>/cancel` 再发送新任务
+
+### 2.3 轮询一直 running，但没有 events
+
+可能原因：
+
+- 模型调用还没开始（排队/网络慢）
+- 事件被缓冲上限丢弃（极少见，通常是轮询太慢）
+
+建议：
+
+- 将轮询间隔调小（0.5s~1s）
+- 确保客户端从 `next_offset` 增量拉取，不要重复拉取导致 UI“看似卡住”
+
+### 2.4 任务失败（轮询 events 里出现 type=error）
+
+原因可能来自：
+
+- 模型调用异常（上游 API/Key/网络）
+- 工具执行异常（例如 web_search/extract_webpage 失败）
+- 运行时资源问题
+
+建议：
+
+- 读取 `GET /api/v1/tasks/<task_id>` 返回的 `data.error`
+- 同时遍历 `events` 中的 `type=error` 与 `system_message`，这些往往包含更具体原因
+
diff --git a/api_doc/events.md b/api_doc/events.md
new file mode 100644
index 0000000..163ac80
--- /dev/null
+++ b/api_doc/events.md
@@ -0,0 +1,255 @@
+# 事件流说明（轮询返回 events）
+
+`GET /api/v1/tasks/<task_id>` 返回的 `events` 字段，是一个按时间顺序（`idx` 递增）的**事件流**。
+
+本项目的目标是：**与网页端 WebSocket 事件保持同一粒度**。因此你会看到：
+
+- token 级别的 `text_chunk`/`thinking_chunk`
+- 工具调用生命周期事件（准备/意图/开始/状态更新）
+- 工具执行结果的状态更新与系统消息
+
+> 提示：客户端应只依赖 `type` 字符串与 `data` 字段，并对未知事件类型保持兼容（忽略或记录）。
+
+## 1. 事件统一包裹格式（envelope）
+
+每个事件的结构：
+
+```json
+{
+  "idx": 12,
+  "type": "text_chunk",
+  "data": { "content": "你好", "index": 5, "elapsed": 0.003 },
+  "ts": 1769182968.154797
+}
+```
+
+字段含义：
+
+- `idx`：事件序号（从 0 开始递增），用于轮询 offset 与去重
+- `type`：事件类型（字符串）
+- `data`：事件载荷（不同 type 不同结构）
+- `ts`：服务端记录的 UNIX 时间戳（秒，float）
+
+## 2. offset 与去重
+
+轮询接口支持 `from` 参数：
+
+- `GET /api/v1/tasks/<task_id>?from=0`：从头拉取
+- `GET /api/v1/tasks/<task_id>?from=next_offset`：增量拉取
+
+正确做法：
+
+1. 客户端本地保存 `offset`（初始 0）
+2. 轮询请求带上 `from=offset`
+3. 处理返回 `events`
+4. 将 `offset` 更新为返回的 `next_offset`
+
+## 3. 常见事件类型与载荷
+
+以下为**常见**事件类型（实际可能随版本增加更多 type）。
+
+### 3.1 AI 消息生命周期
+
+#### `ai_message_start`
+
+AI 新一轮回复开始（一次用户消息可能触发多次迭代，但 UI 通常以此作为“助手消息开始”的信号）。
+
+```json
+{ "type": "ai_message_start", "data": {} }
+```
+
+### 3.2 思考（reasoning）
+
+#### `thinking_start`
+
+```json
+{ "type": "thinking_start", "data": {} }
+```
+
+#### `thinking_chunk`
+
+逐段思考内容（细粒度流式）。
+
+```json
+{ "type": "thinking_chunk", "data": { "content": "..." } }
+```
+
+#### `thinking_end`
+
+思考结束（带完整思考文本）。
+
+```json
+{ "type": "thinking_end", "data": { "full_content": "..." } }
+```
+
+建议客户端处理方式：
+
+- 若你想还原“实时思考流”，就持续 append `thinking_chunk.data.content`；
+- 若你只想最终值，可忽略 chunk，只在 `thinking_end` 读取 `full_content`。
+
+### 3.3 正文输出（token 级）
+
+#### `text_start`
+
+```json
+{ "type": "text_start", "data": {} }
+```
+
+#### `text_chunk`
+
+逐 token/子串输出。字段：
+
+- `content`：本次增量内容
+- `index`：从 1 开始的 chunk 序号（仅用于调试/对齐）
+- `elapsed`：与上一 chunk 的时间间隔（秒）
+
+```json
+{
+  "type": "text_chunk",
+  "data": { "content": "你好", "index": 5, "elapsed": 0.003 }
+}
+```
+
+#### `text_end`
+
+正文结束，包含 `full_content`（完整文本）。
+
+```json
+{ "type": "text_end", "data": { "full_content": "..." } }
+```
+
+建议客户端处理方式：
+
+- 实时显示：append `text_chunk.data.content`
+- 最终落盘：以 `text_end.data.full_content` 为准（如果你做了实时 append，最终可用 full_content 校验/纠偏）
+
+> 注意：若模型在工具调用后没有再次输出总结文本，则可能出现“最后一轮事件是工具相关，没有 text_end”的情况；这属于模型行为/迭代次数限制导致，并不代表任务异常。
+
+### 3.4 工具调用（tool）相关
+
+工具链路通常会出现如下事件：
+
+1) `tool_hint`（可选，提前猜测意图）  
+2) `tool_preparing`（模型开始输出 tool_calls 时）  
+3) `tool_intent`（从增量 arguments 中抽取 intent 字段）  
+4) `tool_start`（真正执行工具时）  
+5) `update_action`（执行进度与结果状态更新）  
+
+#### `tool_hint`（可选）
+
+```json
+{
+  "type": "tool_hint",
+  "data": {
+    "id": "early_web_search_...",
+    "name": "web_search",
+    "message": "检测到可能需要调用 web_search...",
+    "confidence": "low",
+    "conversation_id": "conv_..."
+  }
+}
+```
+
+#### `tool_preparing`
+
+```json
+{
+  "type": "tool_preparing",
+  "data": {
+    "id": "web_search:0",
+    "name": "web_search",
+    "message": "准备调用 web_search...",
+    "intent": "搜索明日方舟终末地游戏信息",
+    "conversation_id": "conv_..."
+  }
+}
+```
+
+#### `tool_intent`
+
+```json
+{
+  "type": "tool_intent",
+  "data": {
+    "id": "web_search:0",
+    "name": "web_search",
+    "intent": "搜索明日方舟终末地游戏信息",
+    "conversation_id": "conv_..."
+  }
+}
+```
+
+#### `tool_start`
+
+```json
+{
+  "type": "tool_start",
+  "data": {
+    "id": "tool_0_web_search_...",
+    "name": "web_search",
+    "arguments": { "...": "..." },
+    "preparing_id": "web_search:0",
+    "monitor_snapshot": null,
+    "conversation_id": "conv_..."
+  }
+}
+```
+
+说明：
+
+- `preparing_id` 用于把“模型输出的 tool_call”与“真实执行的工具卡片”关联起来
+- `arguments` 为工具参数（通常是 JSON 对象）
+
+#### `update_action`
+
+该事件用于更新工具卡片（或其他 action）的状态。字段随工具不同可能略有差异，但通常包含：
+
+- `id`：与 `tool_start.data.id` 对应
+- `status`：`running/completed/failed/...`
+- `message`：可读描述
+- 可能还带 `result/summary/monitor_snapshot/...`
+
+客户端建议：
+
+- 把 `update_action` 当作通用的“状态更新”事件处理；
+- 对未知字段保持兼容（直接展示或忽略）。
+
+### 3.5 系统消息与错误
+
+#### `system_message`
+
+```json
+{ "type": "system_message", "data": { "content": "..." } }
+```
+
+#### `error`
+
+任务内错误（不同于 HTTP 层错误）。出现后通常任务会进入 `failed`。
+
+```json
+{ "type": "error", "data": { "message": "..." } }
+```
+
+### 3.6 任务结束
+
+#### `task_complete`
+
+```json
+{
+  "type": "task_complete",
+  "data": {
+    "total_iterations": 2,
+    "total_tool_calls": 1,
+    "auto_fix_attempts": 0
+  }
+}
+```
+
+> 结束条件以轮询返回的 `status != running` 为准；`task_complete` 是一个“完成事件”，但客户端仍应读取 `status`。
+
+## 4. 事件缓冲与性能建议
+
+- 服务端每个任务保存最近 **1000** 条事件（队列会丢弃更早的数据）
+- 建议轮询间隔 `0.5s ~ 2s`
+- 长任务建议客户端把 `text_chunk` 与关键工具事件持久化到本地，避免错过
+
diff --git a/api_doc/examples.md b/api_doc/examples.md
new file mode 100644
index 0000000..663acdc
--- /dev/null
+++ b/api_doc/examples.md
@@ -0,0 +1,264 @@
+# 示例（curl / Python / JS / Flutter）
+
+本文提供“最小可用”的端到端示例：创建对话 → 发送消息 → 轮询输出 →（可选）停止任务 → 文件上传/下载。
+
+请先阅读 `auth.md` 并准备好 token。
+
+## 0. 统一变量
+
+- `BASE_URL`：例如 `http://localhost:8091`
+- `TOKEN`：你的 Bearer Token（明文）
+
+---
+
+## 1) curl：完整对话流程
+
+### 1.1 创建对话
+
+```bash
+BASE_URL="http://localhost:8091"
+TOKEN="<TOKEN>"
+
+curl -sS -X POST \
+  -H "Authorization: Bearer $TOKEN" \
+  "$BASE_URL/api/v1/conversations"
+```
+
+假设返回：
+
+```json
+{ "success": true, "conversation_id": "conv_20260123_234245_036" }
+```
+
+### 1.2 发送消息（创建任务）
+
+```bash
+curl -sS -X POST \
+  -H "Authorization: Bearer $TOKEN" \
+  -H "Content-Type: application/json" \
+  -d '{
+    "conversation_id": "conv_20260123_234245_036",
+    "message": "请用中文简要介绍《明日方舟：终末地》",
+    "run_mode": "fast",
+    "max_iterations": 100
+  }' \
+  "$BASE_URL/api/v1/messages"
+```
+
+假设返回：
+
+```json
+{ "success": true, "task_id": "60322db3-...", "status": "running", "conversation_id": "conv_...", "created_at": 1769182965.30 }
+```
+
+### 1.3 轮询（建议脚本处理 next_offset）
+
+```bash
+TASK_ID="60322db3-..."
+OFFSET=0
+
+curl -sS -H "Authorization: Bearer $TOKEN" \
+  "$BASE_URL/api/v1/tasks/$TASK_ID?from=$OFFSET"
+```
+
+---
+
+## 2) Python：轮询并实时拼接文本
+
+依赖：`pip install requests`
+
+```python
+import time
+import requests
+
+BASE_URL = "http://localhost:8091"
+TOKEN = "<TOKEN>"
+H = {"Authorization": f"Bearer {TOKEN}"}
+
+def post_json(path, payload):
+    r = requests.post(BASE_URL + path, json=payload, headers={**H, "Content-Type":"application/json"}, timeout=30)
+    r.raise_for_status()
+    return r.json()
+
+def get_json(path, params=None):
+    r = requests.get(BASE_URL + path, params=params or {}, headers=H, timeout=30)
+    r.raise_for_status()
+    return r.json()
+
+# 1) create conversation
+conv = requests.post(BASE_URL + "/api/v1/conversations", headers=H, timeout=30).json()
+assert conv["success"]
+conversation_id = conv["conversation_id"]
+
+# 2) send message
+task = post_json("/api/v1/messages", {
+    "conversation_id": conversation_id,
+    "message": "请用中文简要介绍《明日方舟：终末地》",
+    "run_mode": "fast",
+    "max_iterations": 100
+})
+task_id = task["task_id"]
+
+# 3) poll events
+offset = 0
+text_buf = []
+think_buf = []
+
+while True:
+    data = get_json(f"/api/v1/tasks/{task_id}", params={"from": offset})
+    if not data["success"]:
+        raise RuntimeError(data.get("error"))
+    info = data["data"]
+    events = info["events"]
+    offset = info["next_offset"]
+
+    for ev in events:
+        t = ev["type"]
+        d = ev["data"] or {}
+        if t == "text_chunk":
+            text_buf.append(d.get("content",""))
+            print(d.get("content",""), end="", flush=True)
+        elif t == "text_end":
+            print("\\n--- text_end ---\\n")
+        elif t == "thinking_chunk":
+            think_buf.append(d.get("content",""))
+        elif t == "tool_start":
+            print(f\"\\n[tool_start] {d.get('name')}\\n\")
+        elif t == "update_action":
+            # 工具状态更新
+            st = d.get("status") or ""
+            msg = d.get("message") or ""
+            if st or msg:
+                print(f\"\\n[update_action] {st} {msg}\\n\")
+        elif t == "error":
+            print(f\"\\n[error] {d.get('message')}\\n\")
+
+    if info["status"] != "running":
+        break
+
+    time.sleep(1.0)
+
+final_text = "".join(text_buf)
+final_thinking = "".join(think_buf)
+print("final status:", info["status"])
+print("final text length:", len(final_text))
+print("final thinking length:", len(final_thinking))
+```
+
+---
+
+## 3) JavaScript（浏览器/Node）要点
+
+浏览器端直接跨域请求时，请确保服务端允许 CORS（当前服务端已启用 CORS）。请求示例：
+
+```js
+const BASE_URL = "http://localhost:8091";
+const TOKEN = "<TOKEN>";
+
+async function api(path, options = {}) {
+  const resp = await fetch(BASE_URL + path, {
+    ...options,
+    headers: {
+      "Authorization": `Bearer ${TOKEN}`,
+      ...(options.headers || {})
+    }
+  });
+  const data = await resp.json();
+  if (!resp.ok || !data.success) throw new Error(data.error || resp.statusText);
+  return data;
+}
+
+const conv = await api("/api/v1/conversations", { method: "POST" });
+const msg = await api("/api/v1/messages", {
+  method: "POST",
+  headers: { "Content-Type": "application/json" },
+  body: JSON.stringify({
+    conversation_id: conv.conversation_id,
+    message: "请用中文简要介绍《明日方舟：终末地》",
+    run_mode: "fast",
+    max_iterations: 100
+  })
+});
+
+let offset = 0;
+let text = "";
+while (true) {
+  const poll = await api(`/api/v1/tasks/${msg.task_id}?from=${offset}`);
+  const info = poll.data;
+  for (const ev of info.events) {
+    if (ev.type === "text_chunk") text += ev.data.content || "";
+    if (ev.type === "tool_start") console.log("tool:", ev.data.name);
+  }
+  offset = info.next_offset;
+  if (info.status !== "running") break;
+  await new Promise(r => setTimeout(r, 1000));
+}
+console.log("done:", text);
+```
+
+---
+
+## 4) Flutter（Dart）轮询示例（伪代码）
+
+依赖：`http` 包或 `dio` 包均可，这里用 `http` 表达逻辑。
+
+```dart
+import 'dart:convert';
+import 'dart:async';
+import 'package:http/http.dart' as http;
+
+const baseUrl = "http://localhost:8091";
+const token = "<TOKEN>";
+
+Map<String,String> headersJson() => {
+  "Authorization": "Bearer $token",
+  "Content-Type": "application/json",
+};
+
+Future<String> createConversation() async {
+  final resp = await http.post(Uri.parse("$baseUrl/api/v1/conversations"),
+      headers: {"Authorization":"Bearer $token"});
+  final data = jsonDecode(resp.body);
+  if (resp.statusCode != 200 || data["success"] != true) throw Exception(data["error"]);
+  return data["conversation_id"];
+}
+
+Future<String> sendMessage(String convId, String message) async {
+  final resp = await http.post(Uri.parse("$baseUrl/api/v1/messages"),
+      headers: headersJson(),
+      body: jsonEncode({
+        "conversation_id": convId,
+        "message": message,
+        "run_mode": "fast",
+        "max_iterations": 100,
+      }));
+  final data = jsonDecode(resp.body);
+  if (resp.statusCode != 202 || data["success"] != true) throw Exception(data["error"]);
+  return data["task_id"];
+}
+
+Stream<String> pollText(String taskId) async* {
+  int offset = 0;
+  while (true) {
+    final resp = await http.get(Uri.parse("$baseUrl/api/v1/tasks/$taskId?from=$offset"),
+        headers: {"Authorization":"Bearer $token"});
+    final data = jsonDecode(resp.body);
+    if (resp.statusCode != 200 || data["success"] != true) throw Exception(data["error"]);
+    final info = data["data"];
+    final events = (info["events"] as List);
+    for (final ev in events) {
+      if (ev["type"] == "text_chunk") {
+        yield (ev["data"]["content"] ?? "");
+      }
+    }
+    offset = info["next_offset"];
+    if (info["status"] != "running") break;
+    await Future.delayed(Duration(milliseconds: 800));
+  }
+}
+```
+
+提示：
+
+- Flutter UI 展示建议：把 `pollText()` 的输出 append 到一个 `StringBuffer`，并用 `setState()`/状态管理更新。
+- 同时可以订阅 tool_* 事件在 UI 中显示“正在搜索/正在执行工具”等状态。
diff --git a/api_doc/files.md b/api_doc/files.md
new file mode 100644
index 0000000..8a1be39
--- /dev/null
+++ b/api_doc/files.md
@@ -0,0 +1,166 @@
+# 文件接口（上传 / 列目录 / 下载）
+
+API v1 只提供最小化的文件能力，用于把外部输入文件放入工作区，并下载产物。
+
+核心限制：
+
+- **上传只能落在 `project/user_upload/` 目录及其子目录**；
+- 列目录/下载也只允许访问 `user_upload` 内部路径（路径穿越会被拒绝）。
+
+> 目录结构见 `auth.md`。
+
+## 1) 上传文件
+
+### POST `/api/v1/files/upload`
+
+#### 请求
+
+- Headers：`Authorization: Bearer <TOKEN>`
+- Content-Type：`multipart/form-data`
+- Form 字段：
+  - `file`：必填，上传的文件对象
+  - `filename`：可选，覆盖原文件名（会做安全清洗）
+  - `dir`：可选，`user_upload` 下的子目录（例如 `docs/`）；不传则落在 `user_upload/` 根目录
+
+> 如果你希望“完全不允许指定目录”，客户端请不要传 `dir`，统一上传到根目录即可。
+
+#### 响应
+
+成功（200）：
+
+```json
+{
+  "success": true,
+  "path": "docs/spec.pdf",
+  "filename": "spec.pdf",
+  "size": 12345,
+  "sha256": "..."
+}
+```
+
+字段说明：
+
+- `path`：相对 `user_upload` 的路径（用于 list/download）
+- `sha256`：服务端计算的文件哈希（便于客户端校验/去重）
+
+可能错误：
+
+- `400`：缺少 file / 文件名非法 / 目录非法
+- `401`：缺少或无效 token
+- `503`：系统未初始化
+- `500`：保存失败（例如写入异常、扫描/隔离异常等）
+
+#### curl 示例
+
+上传到根目录：
+
+```bash
+curl -sS -X POST \
+  -H "Authorization: Bearer <TOKEN>" \
+  -F "file=@./hello.txt" \
+  http://localhost:8091/api/v1/files/upload
+```
+
+上传到子目录：
+
+```bash
+curl -sS -X POST \
+  -H "Authorization: Bearer <TOKEN>" \
+  -F "dir=inputs" \
+  -F "file=@./hello.txt" \
+  http://localhost:8091/api/v1/files/upload
+```
+
+## 2) 列出目录内容
+
+### GET `/api/v1/files?path=<dir>`
+
+#### 请求
+
+- Headers：`Authorization: Bearer <TOKEN>`
+- Query：
+  - `path`：可选，相对 `user_upload` 的目录路径；不传表示根目录
+
+#### 响应
+
+成功（200）：
+
+```json
+{
+  "success": true,
+  "base": "inputs",
+  "items": [
+    {
+      "name": "hello.txt",
+      "is_dir": false,
+      "size": 16,
+      "modified_at": 1769182550.594278,
+      "path": "inputs/hello.txt"
+    }
+  ]
+}
+```
+
+字段说明：
+
+- `base`：你请求的目录（相对 `user_upload`），根目录时为 `.`
+- `items[].path`：相对 `user_upload` 的路径（用于下载或继续列目录）
+- `modified_at`：文件 mtime（UNIX 秒）
+
+可能错误：
+
+- `400`：path 不是目录或 path 非法
+- `401`：缺少或无效 token
+- `404`：目录不存在
+- `503`：系统未初始化
+
+#### curl 示例
+
+```bash
+curl -sS \
+  -H "Authorization: Bearer <TOKEN>" \
+  "http://localhost:8091/api/v1/files?path=inputs"
+```
+
+## 3) 下载文件/目录
+
+### GET `/api/v1/files/download?path=<file_or_dir>`
+
+#### 请求
+
+- Headers：`Authorization: Bearer <TOKEN>`
+- Query：
+  - `path`：必填，相对 `user_upload` 的文件或目录路径
+
+#### 响应
+
+1) 若 `path` 指向文件：返回文件二进制流（`Content-Disposition: attachment`）。  
+2) 若 `path` 指向目录：服务端会把目录打包成 zip 后返回（`application/zip`）。
+
+可能错误：
+
+- `400`：缺少 path 或 path 非法
+- `401`：缺少或无效 token
+- `404`：文件/目录不存在
+- `503`：系统未初始化
+
+#### curl 示例
+
+下载文件：
+
+```bash
+curl -L -o out.txt \
+  -H "Authorization: Bearer <TOKEN>" \
+  "http://localhost:8091/api/v1/files/download?path=hello.txt"
+```
+
+下载目录（zip）：
+
+```bash
+curl -L -o inputs.zip \
+  -H "Authorization: Bearer <TOKEN>" \
+  "http://localhost:8091/api/v1/files/download?path=inputs"
+```
+
+> zip 内的路径目前是相对 `project/` 的（因此通常会包含 `user_upload/` 前缀）。
+
diff --git a/api_doc/messages_tasks.md b/api_doc/messages_tasks.md
new file mode 100644
index 0000000..a4cfabb
--- /dev/null
+++ b/api_doc/messages_tasks.md
@@ -0,0 +1,181 @@
+# 发送消息与任务系统（后台运行 + 轮询）
+
+本节覆盖：
+
+- 创建对话（可选）：`POST /api/v1/conversations`
+- 发送消息（创建后台任务）：`POST /api/v1/messages`
+- 轮询事件：`GET /api/v1/tasks/<task_id>`
+- 停止任务：`POST /api/v1/tasks/<task_id>/cancel`
+
+## 1) 创建对话
+
+### POST `/api/v1/conversations`
+
+创建一个新的对话，并返回 `conversation_id`。
+
+请求：
+
+- Headers：`Authorization: Bearer <TOKEN>`
+- Body：无
+
+响应（200）：
+
+```json
+{
+  "success": true,
+  "conversation_id": "conv_20260123_234245_036"
+}
+```
+
+可能错误：
+
+- `401`：缺少或无效 token
+- `503`：系统未初始化（资源繁忙/容器不可用等）
+- `500`：创建失败
+
+## 2) 发送消息（创建后台任务）
+
+### POST `/api/v1/messages`
+
+创建一个后台任务执行智能体流程，并立即返回 `task_id`。
+
+#### 请求
+
+Headers：
+
+- `Authorization: Bearer <TOKEN>`
+- `Content-Type: application/json`
+
+Body(JSON)：
+
+| 字段 | 类型 | 必填 | 说明 |
+|---|---:|---:|---|
+| `message` | string | 是（或 images） | 用户消息文本；`message` 与 `images` 至少其一不为空 |
+| `conversation_id` | string | 否 | 不传则自动新建对话；建议显式传入以便客户端管理会话 |
+| `model_key` | string/null | 否 | 指定模型 key（可选）；具体可用值取决于服务端配置 |
+| `run_mode` | string/null | 否 | 运行模式：`"fast"` \| `"thinking"` \| `"deep"`；若传入则优先使用 |
+| `thinking_mode` | boolean/null | 否 | 兼容字段：true=thinking，false=fast；当 `run_mode` 为空时才使用 |
+| `max_iterations` | integer/null | 否 | 最大迭代次数，默认服务端配置为 **100**（`config.MAX_ITERATIONS_PER_TASK`）；传入可覆盖 | 
+| `images` | string[] | 否 | 图片路径列表（服务端可访问的路径）；一般配合特定模型使用 |
+
+优先级：`run_mode` > `thinking_mode` > 终端当前配置。`run_mode="deep"` 将启用深度思考模式（若模型与配置允许）。
+
+#### 响应
+
+成功（202）：
+
+```json
+{
+  "success": true,
+  "task_id": "60322db3-f884-4a1e-a9b3-6eeb07fbab47",
+  "conversation_id": "conv_20260123_234245_036",
+  "status": "running",
+  "created_at": 1769182965.3038778
+}
+```
+
+字段说明：
+
+- `task_id`：任务唯一 ID（用于轮询/停止）
+- `conversation_id`：本次任务归属对话
+- `status`：初始为 `running`
+- `created_at`：UNIX 时间戳（秒，float）
+
+可能错误：
+
+- `400`：`message/images` 都为空；或 `conversation_id` 无法加载
+- `401`：缺少或无效 token
+- `409`：该 API 用户已有运行中的任务（禁止并发）
+- `503`：系统未初始化/资源繁忙
+
+并发冲突示例（409）：
+
+```json
+{ "success": false, "error": "已有运行中的任务，请稍后再试。" }
+```
+
+## 3) 轮询任务事件（增量）
+
+### GET `/api/v1/tasks/<task_id>?from=<offset>`
+
+用于以 **HTTP 轮询**方式获取任务执行过程中的流式事件（与 WebSocket 同粒度）。
+
+#### 请求
+
+- Headers：`Authorization: Bearer <TOKEN>`
+- Query：
+  - `from`：起始 offset（默认 0）
+
+#### 响应（200）
+
+```json
+{
+  "success": true,
+  "data": {
+    "task_id": "60322db3-f884-4a1e-a9b3-6eeb07fbab47",
+    "status": "running",
+    "created_at": 1769182965.30,
+    "updated_at": 1769182968.15,
+    "error": null,
+    "events": [
+      { "idx": 0, "type": "ai_message_start", "data": {}, "ts": 1769182965.30 },
+      { "idx": 1, "type": "text_start", "data": {}, "ts": 1769182968.09 },
+      { "idx": 2, "type": "text_chunk", "data": { "content": "我来", "index": 1, "elapsed": 0.0 }, "ts": 1769182968.15 }
+    ],
+    "next_offset": 3
+  }
+}
+```
+
+字段说明：
+
+- `status`：任务状态，常见值：
+  - `running`：执行中
+  - `cancel_requested`：已请求停止，等待实际停下
+  - `canceled`：已停止
+  - `succeeded`：成功结束
+  - `failed`：失败结束（`error` 会给出原因）
+- `events`：从 `idx>=from` 的事件列表（按 idx 升序）
+- `next_offset`：建议下一次轮询的 `from` 值
+
+可能错误：
+
+- `401`：缺少或无效 token
+- `404`：任务不存在（或不属于该 API 用户）
+
+#### 推荐轮询策略
+
+- 轮询间隔：`0.5s ~ 2s`（任务越密集越推荐更快）
+- 客户端必须：
+  1) 保存 `next_offset`
+  2) 追加处理 events
+  3) 当 `status != running` 时停止轮询
+
+> 重要：服务端事件缓冲是 `deque(maxlen=1000)`，轮询过慢会丢失早期事件；客户端应自行落盘你需要的内容。
+
+## 4) 停止任务
+
+### POST `/api/v1/tasks/<task_id>/cancel`
+
+请求停止某个任务。
+
+请求：
+
+- Headers：`Authorization: Bearer <TOKEN>`
+- Body：无
+
+成功（200）：
+
+```json
+{ "success": true }
+```
+
+说明：
+
+- 该接口是“请求停止”，任务可能不会立刻停下；
+- 停止后的最终状态在轮询里体现：`status` 变为 `canceled` 或 `failed/succeeded`（极少数情况下已接近结束）。
+
+可能错误：
+
+- `401`：缺少或无效 token
+- `404`：任务不存在
diff --git a/api_doc/openapi.yaml b/api_doc/openapi.yaml
new file mode 100644
index 0000000..8b83926
--- /dev/null
+++ b/api_doc/openapi.yaml
@@ -0,0 +1,351 @@
+openapi: 3.0.3
+info:
+  title: Agents API v1
+  version: "1.0"
+  description: |
+    后台任务 + HTTP 轮询进度 + 文件上传/下载。
+
+    - 鉴权：Authorization: Bearer <token>（token 为自定义字符串，服务端存 SHA256）
+    - 事件流：与网页端 WebSocket 同粒度（text_chunk、tool_start 等）
+servers:
+  - url: http://localhost:8091
+
+components:
+  securitySchemes:
+    bearerAuth:
+      type: http
+      scheme: bearer
+      bearerFormat: token
+  schemas:
+    ErrorResponse:
+      type: object
+      properties:
+        success: { type: boolean, example: false }
+        error: { type: string, example: "无效的 Token" }
+      required: [success, error]
+
+    CreateConversationResponse:
+      type: object
+      properties:
+        success: { type: boolean, example: true }
+        conversation_id: { type: string, example: "conv_20260123_234245_036" }
+      required: [success, conversation_id]
+
+    SendMessageRequest:
+      type: object
+      properties:
+        message:
+          type: string
+          description: 用户消息文本（message 与 images 至少其一不为空）
+        images:
+          type: array
+          items: { type: string }
+          description: 图片路径列表（相对 project 根目录）
+        conversation_id:
+          type: string
+          description: 对话 ID；不传则自动创建
+        model_key:
+          type: string
+          nullable: true
+          description: 模型 key（可选，取决于服务端配置）
+        run_mode:
+          type: string
+          nullable: true
+          description: 运行模式：fast | thinking | deep；若传入则优先使用
+        thinking_mode:
+          type: boolean
+          nullable: true
+          description: 是否开启思考模式（true/false）
+        max_iterations:
+          type: integer
+          nullable: true
+          description: 最大迭代次数（最多允许多少轮模型调用/工具循环）
+      additionalProperties: false
+
+    SendMessageResponse:
+      type: object
+      properties:
+        success: { type: boolean, example: true }
+        task_id: { type: string, example: "60322db3-f884-4a1e-a9b3-6eeb07fbab47" }
+        conversation_id: { type: string, example: "conv_20260123_234245_036" }
+        status: { type: string, example: "running" }
+        created_at: { type: number, format: double, example: 1769182965.3038778 }
+      required: [success, task_id, conversation_id, status, created_at]
+
+    TaskEvent:
+      type: object
+      properties:
+        idx: { type: integer, example: 12 }
+        type: { type: string, example: "text_chunk" }
+        data:
+          type: object
+          additionalProperties: true
+        ts: { type: number, format: double, example: 1769182968.154797 }
+      required: [idx, type, data, ts]
+
+    TaskPollResponse:
+      type: object
+      properties:
+        success: { type: boolean, example: true }
+        data:
+          type: object
+          properties:
+            task_id: { type: string }
+            status: { type: string, example: "running" }
+            created_at: { type: number, format: double }
+            updated_at: { type: number, format: double }
+            error: { type: string, nullable: true }
+            events:
+              type: array
+              items: { $ref: "#/components/schemas/TaskEvent" }
+            next_offset: { type: integer, example: 3 }
+          required: [task_id, status, created_at, updated_at, error, events, next_offset]
+      required: [success, data]
+
+    CancelResponse:
+      type: object
+      properties:
+        success: { type: boolean, example: true }
+      required: [success]
+
+    UploadResponse:
+      type: object
+      properties:
+        success: { type: boolean, example: true }
+        path: { type: string, example: "inputs/hello.txt" }
+        filename: { type: string, example: "hello.txt" }
+        size: { type: integer, example: 16 }
+        sha256: { type: string, example: "0f5bd6..." }
+      required: [success, path, filename]
+
+    ListFilesItem:
+      type: object
+      properties:
+        name: { type: string, example: "hello.txt" }
+        is_dir: { type: boolean, example: false }
+        size: { type: integer, example: 16 }
+        modified_at: { type: number, format: double, example: 1769182550.594278 }
+        path: { type: string, example: "inputs/hello.txt" }
+      required: [name, is_dir, size, modified_at, path]
+
+    ListFilesResponse:
+      type: object
+      properties:
+        success: { type: boolean, example: true }
+        base: { type: string, example: "." }
+        items:
+          type: array
+          items: { $ref: "#/components/schemas/ListFilesItem" }
+      required: [success, base, items]
+
+security:
+  - bearerAuth: []
+
+paths:
+  /api/v1/conversations:
+    post:
+      summary: 创建对话
+      security: [{ bearerAuth: [] }]
+      responses:
+        "200":
+          description: OK
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/CreateConversationResponse" }
+        "401":
+          description: Unauthorized
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/ErrorResponse" }
+
+  /api/v1/messages:
+    post:
+      summary: 发送消息并创建后台任务
+      security: [{ bearerAuth: [] }]
+      requestBody:
+        required: true
+        content:
+          application/json:
+            schema: { $ref: "#/components/schemas/SendMessageRequest" }
+      responses:
+        "202":
+          description: Accepted
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/SendMessageResponse" }
+        "400":
+          description: Bad Request
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/ErrorResponse" }
+        "401":
+          description: Unauthorized
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/ErrorResponse" }
+        "409":
+          description: Conflict (已有运行任务)
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/ErrorResponse" }
+
+  /api/v1/tasks/{task_id}:
+    get:
+      summary: 轮询任务事件
+      security: [{ bearerAuth: [] }]
+      parameters:
+        - in: path
+          name: task_id
+          required: true
+          schema: { type: string }
+        - in: query
+          name: from
+          required: false
+          schema: { type: integer, default: 0 }
+      responses:
+        "200":
+          description: OK
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/TaskPollResponse" }
+        "401":
+          description: Unauthorized
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/ErrorResponse" }
+        "404":
+          description: Not Found
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/ErrorResponse" }
+
+  /api/v1/tasks/{task_id}/cancel:
+    post:
+      summary: 请求停止任务
+      security: [{ bearerAuth: [] }]
+      parameters:
+        - in: path
+          name: task_id
+          required: true
+          schema: { type: string }
+      responses:
+        "200":
+          description: OK
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/CancelResponse" }
+        "401":
+          description: Unauthorized
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/ErrorResponse" }
+        "404":
+          description: Not Found
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/ErrorResponse" }
+
+  /api/v1/files/upload:
+    post:
+      summary: 上传文件到 user_upload
+      security: [{ bearerAuth: [] }]
+      requestBody:
+        required: true
+        content:
+          multipart/form-data:
+            schema:
+              type: object
+              properties:
+                file:
+                  type: string
+                  format: binary
+                filename:
+                  type: string
+                dir:
+                  type: string
+                  description: user_upload 下子目录
+              required: [file]
+      responses:
+        "200":
+          description: OK
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/UploadResponse" }
+        "400":
+          description: Bad Request
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/ErrorResponse" }
+        "401":
+          description: Unauthorized
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/ErrorResponse" }
+
+  /api/v1/files:
+    get:
+      summary: 列出 user_upload 目录内容
+      security: [{ bearerAuth: [] }]
+      parameters:
+        - in: query
+          name: path
+          required: false
+          schema: { type: string, default: "" }
+      responses:
+        "200":
+          description: OK
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/ListFilesResponse" }
+        "400":
+          description: Bad Request
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/ErrorResponse" }
+        "401":
+          description: Unauthorized
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/ErrorResponse" }
+        "404":
+          description: Not Found
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/ErrorResponse" }
+
+  /api/v1/files/download:
+    get:
+      summary: 下载文件或目录（目录会打包成 zip）
+      security: [{ bearerAuth: [] }]
+      parameters:
+        - in: query
+          name: path
+          required: true
+          schema: { type: string }
+      responses:
+        "200":
+          description: File content
+          content:
+            application/octet-stream:
+              schema:
+                type: string
+                format: binary
+            application/zip:
+              schema:
+                type: string
+                format: binary
+        "400":
+          description: Bad Request
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/ErrorResponse" }
+        "401":
+          description: Unauthorized
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/ErrorResponse" }
+        "404":
+          description: Not Found
+          content:
+            application/json:
+              schema: { $ref: "#/components/schemas/ErrorResponse" }
diff --git a/config/paths.py b/config/paths.py
index c19de5b..1281650 100644
--- a/config/paths.py
+++ b/config/paths.py
@@ -11,6 +11,11 @@ USERS_DB_FILE = f"{DATA_DIR}/users.json"
 INVITE_CODES_FILE = f"{DATA_DIR}/invite_codes.json"
 ADMIN_POLICY_FILE = f"{DATA_DIR}/admin_policy.json"
 
+# API 专用用户与工作区（与网页用户隔离）
+API_USER_SPACE_DIR = "./api/users"
+API_USERS_DB_FILE = f"{DATA_DIR}/api_users.json"
+API_TOKENS_FILE = f"{DATA_DIR}/api_tokens.json"
+
 __all__ = [
     "DEFAULT_PROJECT_PATH",
     "PROMPTS_DIR",
@@ -20,4 +25,7 @@ __all__ = [
     "USERS_DB_FILE",
     "INVITE_CODES_FILE",
     "ADMIN_POLICY_FILE",
+    "API_USER_SPACE_DIR",
+    "API_USERS_DB_FILE",
+    "API_TOKENS_FILE",
 ]
diff --git a/modules/api_user_manager.py b/modules/api_user_manager.py
new file mode 100644
index 0000000..9f74f22
--- /dev/null
+++ b/modules/api_user_manager.py
@@ -0,0 +1,160 @@
+"""API 专用用户与工作区管理（JSON + Bearer Token 哈希）。
+
+仅支持手动维护：在 `API_USERS_DB_FILE` 中添加用户与 SHA256(token)。
+结构示例：
+{
+  "users": {
+    "api_jojo": {
+      "token_sha256": "abc123...",
+      "created_at": "2026-01-23",
+      "note": "for mobile app"
+    }
+  }
+}
+"""
+
+from __future__ import annotations
+import json
+import hashlib
+import threading
+from dataclasses import dataclass
+from datetime import datetime
+from pathlib import Path
+from typing import Dict, Optional, Tuple
+
+from config import (
+    API_USER_SPACE_DIR,
+    API_USERS_DB_FILE,
+    API_TOKENS_FILE,
+)
+from modules.personalization_manager import ensure_personalization_config
+
+
+@dataclass
+class ApiUserRecord:
+    username: str
+    token_sha256: str
+    created_at: str
+    note: str = ""
+
+
+@dataclass
+class ApiUserWorkspace:
+    username: str
+    root: Path
+    project_path: Path
+    data_dir: Path
+    logs_dir: Path
+    uploads_dir: Path
+    quarantine_dir: Path
+
+
+class ApiUserManager:
+    """最小化的 API 用户管理：只校验 token 哈希并准备隔离工作区。"""
+
+    def __init__(
+        self,
+        users_file: str = API_USERS_DB_FILE,
+        tokens_file: str = API_TOKENS_FILE,
+        workspace_root: str = API_USER_SPACE_DIR,
+    ):
+        self.users_file = Path(users_file)
+        self.tokens_file = Path(tokens_file)
+        self.workspace_root = Path(workspace_root).expanduser().resolve()
+        self.workspace_root.mkdir(parents=True, exist_ok=True)
+
+        self._users: Dict[str, ApiUserRecord] = {}
+        self._lock = threading.Lock()
+
+        self._load_users()
+
+    # ----------------------- public APIs -----------------------
+    def get_user_by_token(self, bearer_token: str) -> Optional[ApiUserRecord]:
+        if not bearer_token:
+            return None
+        token_sha = self._sha256(bearer_token)
+        with self._lock:
+            for user in self._users.values():
+                if user.token_sha256 == token_sha:
+                    return user
+        return None
+
+    def ensure_workspace(self, username: str) -> ApiUserWorkspace:
+        """为 API 用户创建隔离工作区。"""
+        root = (self.workspace_root / username).resolve()
+        project_path = root / "project"
+        data_dir = root / "data"
+        logs_dir = root / "logs"
+        uploads_dir = project_path / "user_upload"
+
+        for path in (project_path, data_dir, logs_dir, uploads_dir):
+            path.mkdir(parents=True, exist_ok=True)
+
+        # 数据子目录
+        (data_dir / "conversations").mkdir(parents=True, exist_ok=True)
+        (data_dir / "backups").mkdir(parents=True, exist_ok=True)
+        ensure_personalization_config(data_dir)
+
+        # 上传隔离区（沿用 uploads 配置）
+        from config import UPLOAD_QUARANTINE_SUBDIR
+        quarantine_root = Path(UPLOAD_QUARANTINE_SUBDIR).expanduser()
+        if not quarantine_root.is_absolute():
+            quarantine_root = (self.workspace_root.parent / UPLOAD_QUARANTINE_SUBDIR).resolve()
+        quarantine_dir = (quarantine_root / username).resolve()
+        quarantine_dir.mkdir(parents=True, exist_ok=True)
+
+        return ApiUserWorkspace(
+            username=username,
+            root=root,
+            project_path=project_path,
+            data_dir=data_dir,
+            logs_dir=logs_dir,
+            uploads_dir=uploads_dir,
+            quarantine_dir=quarantine_dir,
+        )
+
+    # ----------------------- internal helpers -----------------------
+    def _sha256(self, token: str) -> str:
+        return hashlib.sha256((token or "").encode("utf-8")).hexdigest()
+
+    def _load_users(self):
+        """加载用户列表，读取 token_sha256；不支持明文存储。"""
+        if not self.users_file.exists():
+            self._save_users()
+            return
+        try:
+            raw = json.loads(self.users_file.read_text(encoding="utf-8"))
+        except json.JSONDecodeError as exc:
+            raise RuntimeError(f"无法解析 API 用户文件: {self.users_file} ({exc})")
+
+        users = raw.get("users", {}) if isinstance(raw, dict) else {}
+        for username, payload in users.items():
+            if not isinstance(payload, dict):
+                continue
+            token_sha = (payload.get("token_sha256") or "").strip()
+            if not token_sha:
+                continue
+            record = ApiUserRecord(
+                username=username.strip().lower(),
+                token_sha256=token_sha,
+                created_at=payload.get("created_at") or "",
+                note=payload.get("note") or "",
+            )
+            self._users[record.username] = record
+
+    def _save_users(self):
+        payload = {
+            "users": {
+                username: {
+                    "token_sha256": record.token_sha256,
+                    "created_at": record.created_at or datetime.utcnow().isoformat(),
+                    "note": record.note,
+                }
+                for username, record in self._users.items()
+            }
+        }
+        self.users_file.parent.mkdir(parents=True, exist_ok=True)
+        self.users_file.write_text(json.dumps(payload, ensure_ascii=False, indent=2), encoding="utf-8")
+
+
+__all__ = ["ApiUserManager", "ApiUserRecord", "ApiUserWorkspace"]
diff --git a/server/api_auth.py b/server/api_auth.py
new file mode 100644
index 0000000..f26d761
--- /dev/null
+++ b/server/api_auth.py
@@ -0,0 +1,44 @@
+"""Bearer Token 认证：用于 API v1。
+
+策略：
+- 从 Authorization: Bearer <token> 读取明文 token。
+- 计算 SHA256，与 `api_user_manager` 里的 token_sha256 匹配。
+- 验证通过后，把 username 写入 Flask session 与 `g.api_username`，并标记 `is_api_user=True`。
+- 该装饰器跳过 CSRF（在 server/security.requires_csrf_protection 中已放行 Bearer）。
+"""
+from __future__ import annotations
+import functools
+from flask import request, jsonify, session, g
+
+from . import state
+
+
+def _extract_bearer_token() -> str:
+    auth_header = request.headers.get("Authorization") or ""
+    if not auth_header.lower().startswith("bearer "):
+        return ""
+    return auth_header.split(" ", 1)[1].strip()
+
+
+def api_token_required(view_func):
+    @functools.wraps(view_func)
+    def wrapped(*args, **kwargs):
+        token = _extract_bearer_token()
+        if not token:
+            return jsonify({"success": False, "error": "缺少 Bearer Token"}), 401
+
+        record = state.api_user_manager.get_user_by_token(token)
+        if not record:
+            return jsonify({"success": False, "error": "无效的 Token"}), 401
+
+        # 写入 session 以复用现有上下文/工作区逻辑
+        session["username"] = record.username
+        session["role"] = "api"
+        session["is_api_user"] = True
+        g.api_username = record.username
+        return view_func(*args, **kwargs)
+
+    return wrapped
+
+
+__all__ = ["api_token_required"]
diff --git a/server/api_v1.py b/server/api_v1.py
new file mode 100644
index 0000000..1643864
--- /dev/null
+++ b/server/api_v1.py
@@ -0,0 +1,238 @@
+"""API v1：Bearer Token 版轻量接口（后台任务 + 轮询 + 文件）。"""
+from __future__ import annotations
+import os
+import zipfile
+from io import BytesIO
+from pathlib import Path
+from typing import Dict, Any
+
+from flask import Blueprint, request, jsonify, send_file, session
+
+from .api_auth import api_token_required
+from .tasks import task_manager
+from .context import get_user_resources, ensure_conversation_loaded, get_upload_guard
+from .files import sanitize_filename_preserve_unicode
+from .utils_common import debug_log
+
+api_v1_bp = Blueprint("api_v1", __name__, url_prefix="/api/v1")
+
+
+def _within_uploads(workspace, rel_path: str) -> Path:
+    base = Path(workspace.uploads_dir).resolve()
+    rel = rel_path or ""
+    # 兼容传入包含 user_upload/ 前缀的路径
+    if rel.startswith("user_upload/"):
+        rel = rel.split("user_upload/", 1)[1]
+    rel = rel.lstrip("/").strip()
+    target = (base / rel).resolve()
+    if not str(target).startswith(str(base)):
+        raise ValueError("非法路径")
+    return target
+
+
+@api_v1_bp.route("/conversations", methods=["POST"])
+@api_token_required
+def create_conversation_api():
+    username = session.get("username")
+    terminal, workspace = get_user_resources(username)
+    if not terminal:
+        return jsonify({"success": False, "error": "系统未初始化"}), 503
+    result = terminal.create_new_conversation()
+    if not result.get("success"):
+        return jsonify({"success": False, "error": result.get("error") or "创建对话失败"}), 500
+    return jsonify({"success": True, "conversation_id": result.get("conversation_id")})
+
+
+@api_v1_bp.route("/messages", methods=["POST"])
+@api_token_required
+def send_message_api():
+    username = session.get("username")
+    payload = request.get_json() or {}
+    message = (payload.get("message") or "").strip()
+    images = payload.get("images") or []
+    conversation_id = payload.get("conversation_id")
+    model_key = payload.get("model_key")
+    thinking_mode = payload.get("thinking_mode")
+    run_mode = payload.get("run_mode")
+    max_iterations = payload.get("max_iterations")
+    if not message and not images:
+        return jsonify({"success": False, "error": "消息不能为空"}), 400
+
+    terminal, workspace = get_user_resources(username)
+    if not terminal or not workspace:
+        return jsonify({"success": False, "error": "系统未初始化"}), 503
+    try:
+        conversation_id, _ = ensure_conversation_loaded(terminal, conversation_id)
+    except Exception as exc:
+        return jsonify({"success": False, "error": f"对话加载失败: {exc}"}), 400
+
+    try:
+        rec = task_manager.create_chat_task(
+            username=username,
+            message=message,
+            images=images,
+            conversation_id=conversation_id,
+            model_key=model_key,
+            thinking_mode=thinking_mode,
+            run_mode=run_mode,
+            max_iterations=max_iterations,
+        )
+    except ValueError as exc:
+        return jsonify({"success": False, "error": str(exc)}), 400
+    except RuntimeError as exc:
+        # 并发等业务冲突仍用 409
+        return jsonify({"success": False, "error": str(exc)}), 409
+    return jsonify({
+        "success": True,
+        "task_id": rec.task_id,
+        "conversation_id": rec.conversation_id,
+        "status": rec.status,
+        "created_at": rec.created_at,
+    }), 202
+
+
+@api_v1_bp.route("/tasks/<task_id>", methods=["GET"])
+@api_token_required
+def get_task_events(task_id: str):
+    username = session.get("username")
+    rec = task_manager.get_task(username, task_id)
+    if not rec:
+        return jsonify({"success": False, "error": "任务不存在"}), 404
+    try:
+        offset = int(request.args.get("from", 0))
+    except Exception:
+        offset = 0
+    events = [e for e in rec.events if e["idx"] >= offset]
+    next_offset = events[-1]["idx"] + 1 if events else offset
+    return jsonify({
+        "success": True,
+        "data": {
+            "task_id": rec.task_id,
+            "status": rec.status,
+            "created_at": rec.created_at,
+            "updated_at": rec.updated_at,
+            "error": rec.error,
+            "events": events,
+            "next_offset": next_offset,
+        }
+    })
+
+
+@api_v1_bp.route("/tasks/<task_id>/cancel", methods=["POST"])
+@api_token_required
+def cancel_task_api_v1(task_id: str):
+    username = session.get("username")
+    ok = task_manager.cancel_task(username, task_id)
+    if not ok:
+        return jsonify({"success": False, "error": "任务不存在"}), 404
+    return jsonify({"success": True})
+
+
+@api_v1_bp.route("/files/upload", methods=["POST"])
+@api_token_required
+def upload_file_api():
+    username = session.get("username")
+    terminal, workspace = get_user_resources(username)
+    if not terminal or not workspace:
+        return jsonify({"success": False, "error": "系统未初始化"}), 503
+    if 'file' not in request.files:
+        return jsonify({"success": False, "error": "未找到文件"}), 400
+    file_obj = request.files['file']
+    raw_name = request.form.get('filename') or file_obj.filename
+    filename = sanitize_filename_preserve_unicode(raw_name)
+    if not filename:
+        return jsonify({"success": False, "error": "非法文件名"}), 400
+    subdir = request.form.get("dir") or ""
+    try:
+        target_dir = _within_uploads(workspace, subdir)
+        target_dir.mkdir(parents=True, exist_ok=True)
+        target_path = (target_dir / filename).resolve()
+    except Exception as exc:
+        return jsonify({"success": False, "error": str(exc)}), 400
+
+    guard = get_upload_guard(workspace)
+    rel_path = str(target_path.relative_to(workspace.uploads_dir))
+    try:
+        result = guard.process_upload(
+            file_obj,
+            target_path,
+            username=username,
+            source="api_v1",
+            original_name=raw_name,
+            relative_path=rel_path,
+        )
+    except Exception as exc:
+        return jsonify({"success": False, "error": f"保存文件失败: {exc}"}), 500
+
+    metadata = result.get("metadata", {})
+    return jsonify({
+        "success": True,
+        "path": rel_path,
+        "filename": target_path.name,
+        "size": metadata.get("size"),
+        "sha256": metadata.get("sha256"),
+    })
+
+
+@api_v1_bp.route("/files", methods=["GET"])
+@api_token_required
+def list_files_api():
+    username = session.get("username")
+    _, workspace = get_user_resources(username)
+    if not workspace:
+        return jsonify({"success": False, "error": "系统未初始化"}), 503
+    rel = request.args.get("path") or ""
+    try:
+        target = _within_uploads(workspace, rel)
+        if not target.exists():
+            return jsonify({"success": False, "error": "路径不存在"}), 404
+        if not target.is_dir():
+            return jsonify({"success": False, "error": "路径不是文件夹"}), 400
+        items = []
+        for entry in sorted(target.iterdir(), key=lambda p: p.name):
+            stat = entry.stat()
+            rel_entry = entry.relative_to(workspace.uploads_dir)
+            items.append({
+                "name": entry.name,
+                "is_dir": entry.is_dir(),
+                "size": stat.st_size,
+                "modified_at": stat.st_mtime,
+                "path": str(rel_entry),
+            })
+        return jsonify({"success": True, "items": items, "base": str(target.relative_to(workspace.uploads_dir))})
+    except Exception as exc:
+        return jsonify({"success": False, "error": str(exc)}), 400
+
+
+@api_v1_bp.route("/files/download", methods=["GET"])
+@api_token_required
+def download_file_api():
+    username = session.get("username")
+    _, workspace = get_user_resources(username)
+    if not workspace:
+        return jsonify({"success": False, "error": "系统未初始化"}), 503
+    rel = request.args.get("path")
+    if not rel:
+        return jsonify({"success": False, "error": "缺少 path"}), 400
+    try:
+        target = _within_uploads(workspace, rel)
+    except Exception as exc:
+        return jsonify({"success": False, "error": str(exc)}), 400
+    if not target.exists():
+        return jsonify({"success": False, "error": "文件不存在"}), 404
+
+    if target.is_dir():
+        memory_file = BytesIO()
+        with zipfile.ZipFile(memory_file, mode='w', compression=zipfile.ZIP_DEFLATED) as zf:
+            for root, _, files in os.walk(target):
+                for file in files:
+                    full_path = Path(root) / file
+                    arcname = full_path.relative_to(workspace.project_path)
+                    zf.write(full_path, arcname=str(arcname))
+        memory_file.seek(0)
+        download_name = f"{target.name}.zip"
+        return send_file(memory_file, as_attachment=True, download_name=download_name, mimetype='application/zip')
+    return send_file(target, as_attachment=True, download_name=target.name)
+
+
+__all__ = ["api_v1_bp"]
diff --git a/server/app_legacy.py b/server/app_legacy.py
index 5dba7a4..558e559 100644
--- a/server/app_legacy.py
+++ b/server/app_legacy.py
@@ -32,6 +32,7 @@ from server.chat import chat_bp
 from server.usage import usage_bp
 from server.status import status_bp
 from server.tasks import tasks_bp
+from server.api_v1 import api_v1_bp
 from server.socket_handlers import socketio
 from server.security import attach_security_hooks
 from werkzeug.utils import secure_filename
@@ -291,6 +292,7 @@ app.register_blueprint(chat_bp)
 app.register_blueprint(usage_bp)
 app.register_blueprint(status_bp)
 app.register_blueprint(tasks_bp)
+app.register_blueprint(api_v1_bp)
 
 # 安全钩子（CSRF 校验 + 响应头）
 attach_security_hooks(app)
diff --git a/server/chat_flow.py b/server/chat_flow.py
index 7f6a6c4..f9bb276 100644
--- a/server/chat_flow.py
+++ b/server/chat_flow.py
@@ -508,8 +508,8 @@ async def handle_task_with_sender(terminal: WebTerminal, workspace: UserWorkspac
     last_tool_call_time = 0
     detected_tool_intent: Dict[str, str] = {}
     
-    # 设置最大迭代次数
-    max_iterations = MAX_ITERATIONS_PER_TASK
+    # 设置最大迭代次数（API 可覆盖）
+    max_iterations = getattr(web_terminal, "max_iterations_override", None) or MAX_ITERATIONS_PER_TASK
     
     pending_append = None  # {"path": str, "tool_call_id": str, "buffer": str, ...}
     append_probe_buffer = ""
@@ -1291,37 +1291,6 @@ async def handle_task_with_sender(terminal: WebTerminal, workspace: UserWorkspac
                     content_chunks += 1
                     debug_log(f"  正式内容 #{content_chunks}: {repr(content[:100] if content else 'None')}")
                     
-                    # 通过文本内容提前检测工具调用意图
-                    if not detected_tools:
-                        # 检测常见的工具调用模式
-                        tool_patterns = [
-                            (r'(创建|新建|生成).*(文件|file)', 'create_file'),
-                            (r'(读取|查看|打开).*(文件|file)', 'read_file'),
-                            (r'(修改|编辑|更新).*(文件|file)', 'write_file_diff'),
-                            (r'(删除|移除).*(文件|file)', 'delete_file'),
-                            (r'(搜索|查找|search)', 'web_search'),
-                            (r'(执行|运行).*(Python|python|代码)', 'run_python'),
-                            (r'(执行|运行).*(命令|command)', 'run_command'),
-                            (r'(等待|sleep|延迟)', 'sleep'),
-                            (r'(聚焦|focus).*(文件|file)', 'focus_file'),
-                            (r'(终端|terminal|会话|session)', 'terminal_session'),
-                        ]
-                        
-                        for pattern, tool_name in tool_patterns:
-                            if re.search(pattern, content, re.IGNORECASE):
-                                early_tool_id = f"early_{tool_name}_{time.time()}"
-                                if early_tool_id not in detected_tools:
-                                    sender('tool_hint', {
-                                        'id': early_tool_id,
-                                        'name': tool_name,
-                                        'message': f'检测到可能需要调用 {tool_name}...',
-                                        'confidence': 'low',
-                                        'conversation_id': conversation_id
-                                    })
-                                    detected_tools[early_tool_id] = tool_name
-                                    debug_log(f"    ⚡ 提前检测到工具意图: {tool_name}")
-                                    break
-                    
                     if in_thinking and not thinking_ended:
                         in_thinking = False
                         thinking_ended = True
diff --git a/server/context.py b/server/context.py
index 5fd78e6..083966b 100644
--- a/server/context.py
+++ b/server/context.py
@@ -2,7 +2,7 @@
 from __future__ import annotations
 from functools import wraps
 from typing import Optional, Tuple, Dict, Any
-from flask import session, jsonify
+from flask import session, jsonify, has_request_context
 
 from core.web_terminal import WebTerminal
 from modules.gui_file_manager import GuiFileManager
@@ -39,14 +39,20 @@ def get_user_resources(username: Optional[str] = None) -> Tuple[Optional[WebTerm
     username = (username or get_current_username())
     if not username:
         return None, None
-    record = get_current_user_record()
-    workspace = state.user_manager.ensure_user_workspace(username)
+    is_api_user = bool(session.get("is_api_user")) if has_request_context() else False
+    # API 用户与网页用户使用不同的 manager
+    if is_api_user:
+        record = None
+        workspace = state.api_user_manager.ensure_workspace(username)
+    else:
+        record = get_current_user_record()
+        workspace = state.user_manager.ensure_user_workspace(username)
     container_handle = state.container_manager.ensure_container(username, str(workspace.project_path))
-    usage_tracker = get_or_create_usage_tracker(username, workspace)
+    usage_tracker = None if is_api_user else get_or_create_usage_tracker(username, workspace)
     terminal = state.user_terminals.get(username)
     if not terminal:
-        run_mode = session.get('run_mode')
-        thinking_mode_flag = session.get('thinking_mode')
+        run_mode = session.get('run_mode') if has_request_context() else None
+        thinking_mode_flag = session.get('thinking_mode') if has_request_context() else None
         if run_mode not in {"fast", "thinking", "deep"}:
             preferred_run_mode = None
             try:
@@ -78,51 +84,53 @@ def get_user_resources(username: Optional[str] = None) -> Tuple[Optional[WebTerm
             terminal.terminal_manager.broadcast = terminal.message_callback
         state.user_terminals[username] = terminal
         terminal.username = username
-        terminal.user_role = get_current_user_role(record)
-        terminal.quota_update_callback = lambda metric=None: emit_user_quota_update(username)
-        session['run_mode'] = terminal.run_mode
-        session['thinking_mode'] = terminal.thinking_mode
+        terminal.user_role = "api" if is_api_user else get_current_user_role(record)
+        terminal.quota_update_callback = (lambda metric=None: emit_user_quota_update(username)) if not is_api_user else None
+        if has_request_context():
+            session['run_mode'] = terminal.run_mode
+            session['thinking_mode'] = terminal.thinking_mode
     else:
         terminal.update_container_session(container_handle)
         attach_user_broadcast(terminal, username)
         terminal.username = username
-        terminal.user_role = get_current_user_role(record)
-        terminal.quota_update_callback = lambda metric=None: emit_user_quota_update(username)
+        terminal.user_role = "api" if is_api_user else get_current_user_role(record)
+        terminal.quota_update_callback = (lambda metric=None: emit_user_quota_update(username)) if not is_api_user else None
 
     # 应用管理员策略
-    try:
-        from core.tool_config import ToolCategory
-        from modules import admin_policy_manager
-        policy = admin_policy_manager.get_effective_policy(
-            record.username if record else None,
-            get_current_user_role(record),
-            getattr(record, "invite_code", None),
-        )
-        categories_map = {
-            cid: ToolCategory(
-                label=cat.get("label") or cid,
-                tools=list(cat.get("tools") or []),
-                default_enabled=bool(cat.get("default_enabled", True)),
-                silent_when_disabled=bool(cat.get("silent_when_disabled", False)),
+    if not is_api_user:
+        try:
+            from core.tool_config import ToolCategory
+            from modules import admin_policy_manager
+            policy = admin_policy_manager.get_effective_policy(
+                record.username if record else None,
+                get_current_user_role(record),
+                getattr(record, "invite_code", None),
             )
-            for cid, cat in policy.get("categories", {}).items()
-        }
-        forced_states = policy.get("forced_category_states") or {}
-        disabled_models = policy.get("disabled_models") or []
-        terminal.set_admin_policy(categories_map, forced_states, disabled_models)
-        terminal.admin_policy_ui_blocks = policy.get("ui_blocks") or {}
-        terminal.admin_policy_version = policy.get("updated_at")
-        if terminal.model_key in disabled_models:
-            for candidate in ["kimi", "deepseek", "qwen3-vl-plus", "qwen3-max"]:
-                if candidate not in disabled_models:
-                    try:
-                        terminal.set_model(candidate)
-                        session["model_key"] = terminal.model_key
-                        break
-                    except Exception:
-                        continue
-    except Exception as exc:
-        debug_log(f"[admin_policy] 应用失败: {exc}")
+            categories_map = {
+                cid: ToolCategory(
+                    label=cat.get("label") or cid,
+                    tools=list(cat.get("tools") or []),
+                    default_enabled=bool(cat.get("default_enabled", True)),
+                    silent_when_disabled=bool(cat.get("silent_when_disabled", False)),
+                )
+                for cid, cat in policy.get("categories", {}).items()
+            }
+            forced_states = policy.get("forced_category_states") or {}
+            disabled_models = policy.get("disabled_models") or []
+            terminal.set_admin_policy(categories_map, forced_states, disabled_models)
+            terminal.admin_policy_ui_blocks = policy.get("ui_blocks") or {}
+            terminal.admin_policy_version = policy.get("updated_at")
+            if terminal.model_key in disabled_models:
+                for candidate in ["kimi", "deepseek", "qwen3-vl-plus", "qwen3-max"]:
+                    if candidate not in disabled_models:
+                        try:
+                            terminal.set_model(candidate)
+                            session["model_key"] = terminal.model_key
+                            break
+                        except Exception:
+                            continue
+        except Exception as exc:
+            debug_log(f"[admin_policy] 应用失败: {exc}")
     return terminal, workspace
 
 
@@ -213,8 +221,9 @@ def ensure_conversation_loaded(terminal: WebTerminal, conversation_id: Optional[
         if not result.get("success"):
             raise RuntimeError(result.get("message", "创建对话失败"))
         conversation_id = result["conversation_id"]
-        session['run_mode'] = terminal.run_mode
-        session['thinking_mode'] = terminal.thinking_mode
+        if has_request_context():
+            session['run_mode'] = terminal.run_mode
+            session['thinking_mode'] = terminal.thinking_mode
         created_new = True
     else:
         conversation_id = conversation_id if conversation_id.startswith('conv_') else f"conv_{conversation_id}"
@@ -237,8 +246,9 @@ def ensure_conversation_loaded(terminal: WebTerminal, conversation_id: Optional[
                     terminal.api_client.start_new_task(force_deep=terminal.deep_thinking_mode)
                 else:
                     terminal.api_client.start_new_task()
-                session['run_mode'] = terminal.run_mode
-                session['thinking_mode'] = terminal.thinking_mode
+                if has_request_context():
+                    session['run_mode'] = terminal.run_mode
+                    session['thinking_mode'] = terminal.thinking_mode
             except Exception:
                 pass
     return conversation_id, created_new
diff --git a/server/security.py b/server/security.py
index 1373c6f..00d6439 100644
--- a/server/security.py
+++ b/server/security.py
@@ -111,6 +111,13 @@ def get_csrf_token(force_new: bool = False) -> str:
 
 
 def requires_csrf_protection(path: str) -> bool:
+    # Bearer Token 请求走无状态认证，跳过 CSRF
+    auth_header = (request.headers.get("Authorization") or "").lower()
+    if auth_header.startswith("bearer "):
+        return False
+    # API v1 统一跳过 CSRF；若未携带 Authorization，将由鉴权层返回 401
+    if path.startswith("/api/v1/"):
+        return False
     if path in state.CSRF_EXEMPT_PATHS:
         return False
     if path in state.CSRF_PROTECTED_PATHS:
diff --git a/server/state.py b/server/state.py
index 082c14c..85c0037 100644
--- a/server/state.py
+++ b/server/state.py
@@ -12,9 +12,11 @@ from modules.custom_tool_registry import CustomToolRegistry
 from modules.usage_tracker import UsageTracker
 from modules.user_container_manager import UserContainerManager
 from modules.user_manager import UserManager
+from modules.api_user_manager import ApiUserManager
 
 # 全局实例
 user_manager = UserManager()
+api_user_manager = ApiUserManager()
 custom_tool_registry = CustomToolRegistry()
 container_manager = UserContainerManager()
 user_terminals: Dict[str, WebTerminal] = {}
@@ -68,6 +70,7 @@ PROJECT_MAX_STORAGE_MB = PROJECT_MAX_STORAGE_MB
 
 __all__ = [
     "user_manager",
+    "api_user_manager",
     "custom_tool_registry",
     "container_manager",
     "user_terminals",
diff --git a/server/tasks.py b/server/tasks.py
index 700d085..47aea53 100644
--- a/server/tasks.py
+++ b/server/tasks.py
@@ -7,6 +7,7 @@ from collections import deque
 from typing import Dict, Any, Optional, List
 
 from flask import Blueprint, request, jsonify
+from flask import current_app, session
 
 from .auth_helpers import api_login_required, get_current_username
 from .context import get_user_resources, ensure_conversation_loaded
@@ -27,9 +28,25 @@ class TaskRecord:
         "events",
         "thread",
         "error",
+        "model_key",
+        "thinking_mode",
+        "run_mode",
+        "max_iterations",
+        "session_data",
+        "stop_requested",
     )
 
-    def __init__(self, task_id: str, username: str, message: str, conversation_id: Optional[str]):
+    def __init__(
+        self,
+        task_id: str,
+        username: str,
+        message: str,
+        conversation_id: Optional[str],
+        model_key: Optional[str],
+        thinking_mode: Optional[bool],
+        run_mode: Optional[str],
+        max_iterations: Optional[int],
+    ):
         self.task_id = task_id
         self.username = username
         self.status = "pending"
@@ -40,6 +57,12 @@ class TaskRecord:
         self.events: deque[Dict[str, Any]] = deque(maxlen=1000)
         self.thread: Optional[threading.Thread] = None
         self.error: Optional[str] = None
+        self.model_key = model_key
+        self.thinking_mode = thinking_mode
+        self.run_mode = run_mode
+        self.max_iterations = max_iterations
+        self.session_data: Dict[str, Any] = {}
+        self.stop_requested: bool = False
 
 
 class TaskManager:
@@ -50,9 +73,40 @@ class TaskManager:
         self._lock = threading.Lock()
 
     # ---- public APIs ----
-    def create_chat_task(self, username: str, message: str, images: List[Any], conversation_id: Optional[str]) -> TaskRecord:
+    def create_chat_task(
+        self,
+        username: str,
+        message: str,
+        images: List[Any],
+        conversation_id: Optional[str],
+        model_key: Optional[str] = None,
+        thinking_mode: Optional[bool] = None,
+        run_mode: Optional[str] = None,
+        max_iterations: Optional[int] = None,
+    ) -> TaskRecord:
+        if run_mode:
+            normalized = str(run_mode).lower()
+            if normalized not in {"fast", "thinking", "deep"}:
+                raise ValueError("run_mode 只支持 fast/thinking/deep")
+            run_mode = normalized
+        # 单用户互斥：禁止并发任务
+        existing = [t for t in self.list_tasks(username) if t.status in {"pending", "running"}]
+        if existing:
+            raise RuntimeError("已有运行中的任务，请稍后再试。")
         task_id = str(uuid.uuid4())
-        record = TaskRecord(task_id, username, message, conversation_id)
+        record = TaskRecord(task_id, username, message, conversation_id, model_key, thinking_mode, run_mode, max_iterations)
+        # 记录当前 session 快照，便于后台线程内使用
+        try:
+            record.session_data = {
+                "username": session.get("username"),
+                "role": session.get("role"),
+                "is_api_user": session.get("is_api_user"),
+                "run_mode": session.get("run_mode"),
+                "thinking_mode": session.get("thinking_mode"),
+                "model_key": session.get("model_key"),
+            }
+        except Exception:
+            record.session_data = {}
         with self._lock:
             self._tasks[task_id] = record
         thread = threading.Thread(target=self._run_chat_task, args=(record, images), daemon=True)
@@ -77,6 +131,7 @@ class TaskManager:
         rec = self.get_task(username, task_id)
         if not rec:
             return False
+        rec.stop_requested = True
         # 标记停止标志；chat_flow 会检测 stop_flags
         entry = stop_flags.get(task_id)
         if not isinstance(entry, dict):
@@ -107,10 +162,45 @@ class TaskManager:
 
     def _run_chat_task(self, rec: TaskRecord, images: List[Any]):
         username = rec.username
+        terminal = None
+        workspace = None
+        stop_hint = False
         try:
-            terminal, workspace = get_user_resources(username)
+            # 为后台线程构造最小请求上下文，填充 session
+            from server.app import app as flask_app
+            with flask_app.test_request_context():
+                try:
+                    for k, v in (rec.session_data or {}).items():
+                        if v is not None:
+                            session[k] = v
+                except Exception:
+                    pass
+                terminal, workspace = get_user_resources(username)
             if not terminal or not workspace:
                 raise RuntimeError("系统未初始化")
+            stop_hint = bool(stop_flags.get(rec.task_id, {}).get("stop"))
+
+            # API 传入的模型/模式配置
+            if rec.model_key:
+                try:
+                    terminal.set_model(rec.model_key)
+                except Exception as exc:
+                    debug_log(f"[Task] 设置模型失败 {rec.model_key}: {exc}")
+            if rec.run_mode:
+                try:
+                    terminal.set_run_mode(rec.run_mode)
+                except Exception as exc:
+                    debug_log(f"[Task] 设置运行模式失败 {rec.run_mode}: {exc}")
+            elif rec.thinking_mode is not None:
+                try:
+                    terminal.set_run_mode("thinking" if rec.thinking_mode else "fast")
+                except Exception as exc:
+                    debug_log(f"[Task] 设置思考模式失败: {exc}")
+            if rec.max_iterations:
+                try:
+                    terminal.max_iterations_override = int(rec.max_iterations)
+                except Exception:
+                    terminal.max_iterations_override = None
 
             # 确保会话加载
             conversation_id = rec.conversation_id
@@ -142,8 +232,9 @@ class TaskManager:
             )
 
             # 结束状态
+            canceled_flag = rec.stop_requested or stop_hint or bool(stop_flags.get(rec.task_id, {}).get("stop"))
             with self._lock:
-                rec.status = "canceled" if rec.task_id in stop_flags and stop_flags[rec.task_id].get('stop') else "succeeded"
+                rec.status = "canceled" if canceled_flag else "succeeded"
                 rec.updated_at = time.time()
         except Exception as exc:
             debug_log(f"[Task] 后台任务失败: {exc}")
@@ -155,6 +246,12 @@ class TaskManager:
         finally:
             # 清理 stop_flags
             stop_flags.pop(rec.task_id, None)
+            # 清理一次性配置
+            if terminal and hasattr(terminal, "max_iterations_override"):
+                try:
+                    delattr(terminal, "max_iterations_override")
+                except Exception:
+                    terminal.max_iterations_override = None
 
 
 task_manager = TaskManager()
@@ -192,7 +289,21 @@ def create_task_api():
     conversation_id = payload.get("conversation_id")
     if not message and not images:
         return jsonify({"success": False, "error": "消息不能为空"}), 400
-    rec = task_manager.create_chat_task(username, message, images, conversation_id)
+    model_key = payload.get("model_key")
+    thinking_mode = payload.get("thinking_mode")
+    max_iterations = payload.get("max_iterations")
+    try:
+        rec = task_manager.create_chat_task(
+            username,
+            message,
+            images,
+            conversation_id,
+            model_key=model_key,
+            thinking_mode=thinking_mode,
+            max_iterations=max_iterations,
+        )
+    except RuntimeError as exc:
+        return jsonify({"success": False, "error": str(exc)}), 409
     return jsonify({
         "success": True,
         "data": {
@@ -241,4 +352,3 @@ def cancel_task_api(task_id: str):
     if not ok:
         return jsonify({"success": False, "error": "任务不存在"}), 404
     return jsonify({"success": True})
-