fix: repair one-click installer

README.md

@@ -111,7 +111,7 @@ Tool Registry 首先确定操作风险：
 
 ## 🤖 当前 MCP 能力
 
-目前共 **11 个 MCP Tools**：
+目前共 **12 个 MCP Tools**：
 
 | Tool | Risk | 功能 |
 | :--- | :--- | :--- |
@@ -177,6 +177,28 @@ Agent 只主动连接 Hub，不需要公网入站端口。
 
 ---
 
+## 🚀 一键安装与升级
+
+### 一键安装
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/kyaring/vps-commander/main/deploy/install.sh | sudo bash
+```
+
+安装脚本自动识别 `amd64/arm64` 并进入 Hub / Agent 安装流程。
+
+### 一键升级 Agent
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/kyaring/vps-commander/main/deploy/upgrade-agent.sh | sudo bash
+```
+
+升级脚本下载目标 Release、校验文件大小、备份旧版本并在启动失败时自动回滚。
+
+生产环境的 Agent 凭据以独立设备凭据为准；不要将 Cluster Secret 作为长期生产凭据重新启用。
+
+---
+
 ## 📄 文档索引
 
 - [文档基线与全局索引](docs/DOCUMENTATION_BASELINE_20261002.md)

cmd/agent/main.go

@@ -152,6 +152,17 @@ func run(hub, name, token string, insecure bool) error {
 
 func handle(writeJSON func(cluster.Message) error, exec executor.Local, mcpRuntime *cluster.MCPRuntime, sessions *executor.SessionManager, msg cluster.Message) error {
 	switch msg.Action {
+	case "diagnostic_command":
+		var p executor.DiagnosticRequest
+		if err := decodePayload(msg.Payload, &p); err != nil {
+			return sendError(writeJSON, msg.ID, err)
+		}
+		v, err := executor.Diagnostic(context.Background(), p)
+		if err != nil {
+			return sendError(writeJSON, msg.ID, err)
+		}
+		return writeJSON(cluster.Message{ID: msg.ID, Status: "success", Payload: v})
+
 	case "mcp_list":
 		return writeJSON(cluster.Message{ID: msg.ID, Status: "success", Payload: mcpRuntime.List()})
 	case "mcp_test":

cmd/hub/main.go

@@ -70,6 +70,7 @@ func main() {
 		AdminToken:    adminToken,
 		ExecLimiter:   make(chan struct{}, 32),
 		SearchLimiter: make(chan struct{}, 8),
+		LocalSessions: executor.NewSessionManager(0),
 	}
 	if err := s.LoadSecurityPolicySnapshot(); err != nil {
 		log.Fatalf("load security policy snapshot: %v", err)
@@ -121,6 +122,13 @@ func main() {
 		IdleTimeout:       90 * time.Second,
 	}
 	log.Printf("vps-commander-hub listening on %s as %s", *addr, local)
+	go func() {
+		ticker := time.NewTicker(time.Minute)
+		defer ticker.Stop()
+		for range ticker.C {
+			s.LocalSessions.Cleanup()
+		}
+	}()
 	go func() {
 		if err := server.ListenAndServe(); err != nil && err != http.ErrServerClosed {
 			log.Fatal(err)

deploy/README.md

@@ -47,9 +47,53 @@ chmod 0600 /etc/vps-commander/agent.env
 
 ---
 
-## 三、Hub 服务端部署步骤
+## 三、一键安装与升级
 
-### 3.1 编译
+### 3.1 一键安装
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/kyaring/vps-commander/main/deploy/install.sh | sudo bash
+```
+
+脚本自动识别 `amd64/arm64`，并进入 Hub / Agent 安装流程。
+
+- Hub：自动生成 API Key 与 Web 面板密码。
+- Agent：输入 Hub WSS 地址、节点名称和独立 Agent Token。
+- 生产环境不再要求 Cluster Secret 作为长期凭据。
+
+### 3.2 一键升级 Agent
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/kyaring/vps-commander/main/deploy/upgrade-agent.sh | sudo bash
+```
+
+升级流程：
+
+```text
+下载 Release
+   ↓
+临时文件完整性检查
+   ↓
+备份当前 Agent
+   ↓
+原子替换
+   ↓
+systemd restart
+   ↓
+健康检查
+   ├─ 成功 → 删除备份
+   └─ 失败 → 自动恢复旧版本
+```
+
+### 3.3 手工安装
+
+如果需要固定版本、离线部署或灰度验证，再使用下面的手工安装流程。
+
+---
+
+## 四、Hub 服务端部署步骤
+
+### 4.1 编译
 ```bash
 cd /root/.openclaw/workspace/vps-commander
 go test ./...
@@ -65,13 +109,13 @@ vcs.modified=false
 vcs.revision=<release commit>
 ```
 
-### 3.2 安装
+### 4.2 安装
 ```bash
 install -d -m 0750 /opt/vps-commander/data /etc/vps-commander
 install -m 0755 bin/vps-commander-hub-linux-amd64 /opt/vps-commander/vps-commander-hub
 ```
 
-### 3.3 配置
+### 4.3 配置
 ```bash
 cat << 'ENV' > /etc/vps-commander/hub.env
 VPS_COMMANDER_API_KEY=your_secure_api_key
@@ -80,7 +124,7 @@ ENV
 chmod 0600 /etc/vps-commander/hub.env
 ```
 
-### 3.4 Systemd
+### 4.4 Systemd
 ```ini
 [Service]
 EnvironmentFile=-/etc/vps-commander/hub.env
@@ -89,7 +133,7 @@ Restart=always
 RestartSec=3s
 ```
 
-### 3.5 启动
+### 4.5 启动
 ```bash
 systemctl daemon-reload
 systemctl enable --now vps-commander-hub
@@ -98,15 +142,15 @@ systemctl status vps-commander-hub
 
 ---
 
-## 四、远程 Agent 节点部署步骤
+## 五、远程 Agent 节点部署步骤
 
-### 4.1 安装二进制
+### 5.1 安装二进制
 ```bash
 install -d -m 0750 /etc/vps-commander /opt/vps-commander
 install -m 0755 vps-commander-agent-linux-amd64 /opt/vps-commander/vps-commander-agent
 ```
 
-### 4.2 配置
+### 5.2 配置
 ```bash
 cat << 'ENV' > /etc/vps-commander/agent.env
 VPS_COMMANDER_HUB_WS_URL=wss://vpstool.kory.kdns.fr/agent/ws
@@ -116,7 +160,7 @@ ENV
 chmod 0600 /etc/vps-commander/agent.env
 ```
 
-### 4.3 Systemd
+### 5.3 Systemd
 ```ini
 [Service]
 EnvironmentFile=-/etc/vps-commander/agent.env
@@ -127,7 +171,7 @@ RestartSec=3s
 
 当前 v2.0 生产迁移阶段可兼容显式 token；v2.1 发布后必须改为不出现在 ExecStart 的方式。
 
-### 4.4 启动
+### 5.4 启动
 ```bash
 systemctl daemon-reload
 systemctl enable --now vps-commander-agent
@@ -138,24 +182,24 @@ Agent 启动后主动连接 Hub，并在设备列表中显示为 `online`。
 
 ---
 
-## 五、控制面板与 MCP 使用
+## 六、控制面板与 MCP 使用
 
-### 5.1 管理面板
+### 6.1 管理面板
 `https://vpstool.kory.kdns.fr/panel/login`
 
 面板包含设备状态、安全策略、操作控制、审计和相关管理功能。
 
-### 5.2 MCP
+### 6.2 MCP
 远程入口：
 `https://vpstool.kory.kdns.fr/mcp/sse`
 
 本地 stdio 使用 `vps-commander-mcp-stdio`。
 
-当前共 11 个 MCP tools，详见 `docs/MCP_INTEGRATION.md`。
+当前共 12 个 MCP tools，详见 `docs/MCP_INTEGRATION.md`。
 
 ---
 
-## 六、生产升级流程
+## 七、生产升级流程
 
 ```text
 Backup
@@ -177,7 +221,7 @@ Full Acceptance
 
 ---
 
-## 七、日常运维检查
+## 八、日常运维检查
 
 ```bash
 systemctl status vps-commander-hub
@@ -189,7 +233,7 @@ curl -fsS http://127.0.0.1:9521/healthz
 
 ---
 
-## 八、当前生产审计提示
+## 九、当前生产审计提示
 
 2026-10-02 审计发现生产 Hub / Agent 曾由 `vcs.modified=true` 的工作树构建。
 因此生产现场二进制不能直接视为当前 HEAD 的可复现产物。

deploy/install.sh

@@ -72,15 +72,15 @@ install_agent() {
     read -rp "请输入本节点名称 (默认: ${DEFAULT_NAME}): " AGENT_NAME
     AGENT_NAME=${AGENT_NAME:-${DEFAULT_NAME}}
 
-    read -rp "请输入 Cluster Secret (集群通信密钥): " CLUSTER_SECRET
-    while [ -z "$CLUSTER_SECRET" ]; do
-        read -rp "Cluster Secret 不能为空，请重新输入: " CLUSTER_SECRET
+    read -rp "请输入 Agent 独立凭据 Token: " AGENT_TOKEN
+    while [ -z "$AGENT_TOKEN" ]; do
+        read -rp "Agent Token 不能为空，请重新输入: " AGENT_TOKEN
     done
 
     cat << ENV > "${CONF_DIR}/agent.env"
 VPS_COMMANDER_HUB_WS_URL=${HUB_WS}
 VPS_COMMANDER_AGENT_NAME=${AGENT_NAME}
-VPS_COMMANDER_CLUSTER_SECRET=***
+VPS_COMMANDER_AGENT_TOKEN=${AGENT_TOKEN}
 ENV
     chmod 0600 "${CONF_DIR}/agent.env"
 
@@ -133,23 +133,18 @@ install_hub() {
     read -rp "请输入监听端口 (默认 9521): " HUB_PORT
     HUB_PORT=${HUB_PORT:-9521}
 
-    GEN_API_KEY=*** -c 24 /dev/urandom | xxd -p)
-    GEN_SECRET=*** -c 24 /dev/urandom | xxd -p)
-    GEN_PASS=$(head -c 16 /dev/urandom | xxd -p)
+    GEN_API_KEY=$(od -An -N24 -tx1 /dev/urandom | tr -d ' \n')
+    GEN_PASS=$(od -An -N16 -tx1 /dev/urandom | tr -d ' \n')
 
     read -rp "请输入 API Key (默认随机生成): " API_KEY
-    API_KEY=***
-
-    read -rp "请输入 Cluster Secret (默认随机生成): " CLUSTER_SECRET
-    CLUSTER_SECRET=${CLUS…RET}
+    API_KEY=${API_KEY:-$GEN_API_KEY}
 
     read -rp "请输入 Web 管理面板密码 (默认随机生成): " WEB_PASS
     WEB_PASS=${WEB_PASS:-$GEN_PASS}
 
     cat << ENV > "${CONF_DIR}/hub.env"
-VPS_COMMANDER_API_KEY=***
-VPS_COMMANDER_CLUSTER_SECRET=***
-VPS_COMMANDER_WEB_PASSWORD=***
+VPS_COMMANDER_API_KEY=${API_KEY}
+VPS_COMMANDER_WEB_PASSWORD=${WEB_PASS}
 ENV
     chmod 0600 "${CONF_DIR}/hub.env"
 
@@ -185,7 +180,6 @@ SVC
         echo -e "\n================== Hub 配置凭据 =================="
         echo -e "监听地址: 127.0.0.1:${HUB_PORT}"
         echo -e "API Key: ${API_KEY}"
-        echo -e "Cluster Secret: ${CLUSTER_SECRET}"
         echo -e "Web 面板密码: ${WEB_PASS}"
         echo -e "==================================================\n"
         warn "请妥善保管上述凭据，反代请将域名反代至 127.0.0.1:${HUB_PORT}"

docs/ARCHITECTURE.md

@@ -88,6 +88,7 @@ Agent 返回结果后，Hub 完成审计并将结果返回 HTTP/MCP 调用方。
 | :--- | :--- |
 | Low | read_file、read_multiple_files、list_directory、file_info、search、list_processes |
 | Medium | edit_block、write_file、create_directory、move_file、mcp_call |
+| Low | diagnostic_command：严格白名单只读诊断（Docker、systemd、journal、网络监听、磁盘、内存、uptime） |
 | High | exec_command、start_process、interact_with_process、kill_process、force_terminate |
 
 ### 4.3 Policy Snapshot
@@ -121,7 +122,7 @@ kill / terminate 使用 Session 内部进程句柄，不接受客户端任意 PI
 
 ## 六、MCP 架构
 
-当前 MCP 共 11 个工具：
+当前 MCP 共 12 个工具：
 1. `list_devices`
 2. `exec_command`
 3. `read_file`
@@ -129,6 +130,8 @@ kill / terminate 使用 Session 内部进程句柄，不接受客户端任意 PI
 5. `create_directory`
 6. `move_file`
 7. `list_processes`
+8. `diagnostic_command`：严格白名单的只读系统诊断；
+
 8. `edit_block`
 9. `write_file`
 10. `list_agent_mcp`

docs/DEV_SPEC.md

@@ -86,7 +86,7 @@ Hub 根据 Credential 建立 Canonical Device Identity。
 Server version：`1.2.0`。
 Protocol Version：`2024-11-05`。
 
-当前工具共 11 个：
+当前工具共 12 个：
 `list_devices`、`exec_command`、`read_file`、`read_multiple_files`、`create_directory`、`move_file`、`list_processes`、`edit_block`、`write_file`、`list_agent_mcp`、`call_agent_mcp`。
 
 远程 transport：legacy HTTP + SSE。

docs/DOCUMENTATION_BASELINE_20261002.md

@@ -40,7 +40,7 @@ VPS-Commander 是自托管、多节点、AI Agent 友好的 Linux 运维控制
 
 ## 三、当前 MCP 基线
 
-11 个工具：
+12 个工具：
 `list_devices`、`exec_command`、`read_file`、`read_multiple_files`、`create_directory`、`move_file`、`list_processes`、`edit_block`、`write_file`、`list_agent_mcp`、`call_agent_mcp`。
 
 远程：legacy SSE。

docs/KNOWN_ISSUE_STREAMABLE_HTTP_20261002.md

@@ -0,0 +1,49 @@
+# 已知问题：MCP Streamable HTTP 传输不受支持（待排期）
+
+- **发现日期**：2026-10-02
+- **发现人**：审核专家 B（生产问题分析）
+- **状态**：待排期，未修复
+
+---
+
+## 现象
+
+Android 客户端使用 MCP Kotlin SDK 的 `StreamableHttpClientTransport` 连接 vps-commander，
+发送第一条消息（`initialize`）时失败：
+
+```
+io.modelcontextprotocol.kotlin.sdk.types.McpException: Error while sending message: Streamable HTTP error:
+Caused by: io.modelcontextprotocol.kotlin.sdk.client.StreamableHttpError: Streamable HTTP error:
+    at ...StreamableHttpClientTransport.performSend(...)
+```
+
+## 根因
+
+传输协议代际不匹配：
+
+| | 服务端 (vps-commander) | 客户端 (Kotlin SDK) |
+|---|---|---|
+| 协议 | SSE 传输（MCP 2024-11-05） | Streamable HTTP（MCP 2025-03-26） |
+| 端点 | `GET /mcp/sse` + `POST /mcp/message` | `POST /mcp`（单端点） |
+| 行为 | `HandleSSE` 非 GET 直接 405（`internal/mcp/mcp.go:120`） | POST 收到非 2xx 即抛 `StreamableHttpError` |
+
+服务端没有 `POST /mcp` 路由。客户端 POST 打到 `/mcp/sse` 得 405，打到 `/mcp` 得 404，
+两种情况都会触发该异常。SDK 的错误信息为空（未携带 HTTP 状态码），需抓包/curl 确认。
+
+## 验证
+
+```bash
+curl -i -X POST https://vpstool.kory.kdns.fr/mcp/sse \
+  -H "Content-Type: application/json" -d '{}'
+# 预期：405 method not allowed（坐实协议不匹配）
+```
+
+## 修复选项（待决策）
+
+1. **客户端改（快）**：Android 端改用 `SseClientTransport`，URL 指向 `/mcp/sse`，与服务端现状匹配。
+2. **服务端加（彻底）**：实现 Streamable HTTP（`POST /mcp`，MCP 2025-03-26），属新功能开发。
+
+## 备注
+
+- 与 2026-10-02 bug hunt 的 4 个 bug 无关，是独立的协议兼容性问题。
+- 决策前不要动生产端点。

docs/MCP_INTEGRATION.md

@@ -20,7 +20,7 @@ VPS-Commander 当前提供两种 MCP 接入模式：
 
 ---
 
-## 二、当前支持的 11 大 MCP Tools
+## 二、当前支持的 12 大 MCP Tools
 
 | Tool 名称 | 风险等级 | 参数/功能说明 |
 | :--- | :--- | :--- |
@@ -31,6 +31,7 @@ VPS-Commander 当前提供两种 MCP 接入模式：
 | `create_directory` | Medium | 创建目录 |
 | `move_file` | Medium | 移动/重命名文件 |
 | `list_processes` | Low | 查看进程 |
+| `diagnostic_command` | Low | 严格白名单只读诊断，不接受任意 Shell |
 | `edit_block` | Medium | 唯一匹配 + 原子写回 |
 | `write_file` | Medium | 创建或覆写文件 |
 | `list_agent_mcp` | Low | 查询 Agent MCP 服务 |
@@ -145,6 +146,12 @@ API Key 不应提交到 Git 仓库。
 ### 6.6 list_processes
 - `device` 可选。
 
+### 6.7 diagnostic_command
+- `device`、`action` 必填。
+- `action` 仅允许：`docker_ps`、`docker_logs`、`docker_inspect`、`systemctl_status`、`journalctl`、`ss_listen`、`df`、`free`、`uptime`。
+- 这是 Low 风险只读诊断通道，不接受任意 Shell；Medium 模式可使用，但 `exec_command` 仍保持 High。
+
+
 ### 6.7 edit_block
 - `device`、`path`、`old_text`、`new_text`。
 - `old_text` 必须唯一匹配。
@@ -162,7 +169,31 @@ API Key 不应提交到 Git 仓库。
 
 ---
 
-## 七、Agent MCP
+## 七、一键安装与升级
+
+### 7.1 一键安装
+
+仓库已经提供统一安装脚本，支持 Linux `amd64` 与 `arm64`，安装时自动识别架构并从 GitHub Release 获取对应二进制。
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/kyaring/vps-commander/main/deploy/install.sh | sudo bash
+```
+
+脚本会进入交互式安装流程，可选择 Hub 或 Agent。生产环境使用时，Agent 应使用独立设备凭据，不应重新启用全局 Shared Secret。
+
+### 7.2 一键升级 Agent
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/kyaring/vps-commander/main/deploy/upgrade-agent.sh | sudo bash
+```
+
+升级采用临时文件下载 → 基础完整性检查 → 旧二进制备份 → 原子替换 → systemd 重启 → 健康检查；启动失败时自动回滚旧版本。
+
+> Hub 的生产升级建议继续采用滚动发布：先完成 Release 与构建校验，再替换 Hub，确认节点重新上线后再继续后续节点。避免把 Hub 升级做成无保护的“一键重启”。
+
+---
+
+## 八、Agent MCP
 
 Hub 可以发现远程 Agent 暴露的 MCP Service。
 

docs/PROGRESS_TRACKING.md

@@ -28,7 +28,7 @@
 
 ## 二、MCP 当前能力
 
-当前共 11 个 Tools：
+当前共 12 个 Tools：
 1. `list_devices`
 2. `exec_command`
 3. `read_file`

internal/api/api.go

@@ -49,6 +49,7 @@ type Server struct {
 	AdminToken    string
 	ExecLimiter   chan struct{}
 	SearchLimiter chan struct{}
+	LocalSessions *executor.SessionManager
 }
 
 const (
@@ -782,9 +783,13 @@ func (s *Server) processSession(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 	var req struct {
-		Device, SessionID, Command, Workdir, Data string
-		Force                                     bool
-		RingCapacity                              int `json:"ring_capacity"`
+		Device       string `json:"device"`
+		SessionID    string `json:"session_id"`
+		Command      string `json:"command"`
+		Workdir      string `json:"workdir"`
+		Data         string `json:"data"`
+		Force        bool   `json:"force"`
+		RingCapacity int    `json:"ring_capacity"`
 	}
 	if err := json.NewDecoder(http.MaxBytesReader(w, r.Body, 256*1024)).Decode(&req); err != nil {
 		http.Error(w, "bad json", 400)
@@ -796,38 +801,106 @@ func (s *Server) processSession(w http.ResponseWriter, r *http.Request) {
 	if !s.requireSecurity(w, r, req.Device, "exec", req.SessionID) {
 		return
 	}
-	if s.Cluster == nil {
-		http.Error(w, "cluster unavailable", 503)
-		return
-	}
 	ctx, cancel := context.WithTimeout(r.Context(), 30*time.Second)
 	defer cancel()
 	action := r.URL.Query().Get("action")
+	owner := sessionOwner(r)
 	var result any
 	var err error
-	switch action {
-	case "start":
-		if req.SessionID == "" {
-			req.SessionID, err = newSessionID()
-			if err != nil {
-				http.Error(w, "session id generation failed", 500)
-				return
+	if s.isRemote(req.Device) {
+		if s.Cluster == nil || !s.Cluster.Online(req.Device) {
+			http.Error(w, "device offline", 409)
+			return
+		}
+		switch action {
+		case "start":
+			if req.SessionID == "" {
+				req.SessionID, err = newSessionID()
+				if err != nil {
+					http.Error(w, "session id generation failed", 500)
+					return
+				}
 			}
+			result, err = s.Cluster.StartProcess(ctx, req.Device, cluster.StartProcessPayload{Command: req.Command, Workdir: req.Workdir, SessionID: req.SessionID, Owner: owner, RingCapacity: req.RingCapacity})
+		case "read":
+			result, err = s.Cluster.ReadProcessOutput(ctx, req.Device, req.SessionID, owner)
+		case "input":
+			err = s.Cluster.InteractProcess(ctx, req.Device, req.SessionID, req.Data, owner)
+			result = map[string]any{"ok": true}
+		case "stop":
+			err = s.Cluster.ForceTerminate(ctx, req.Device, req.SessionID, owner)
+			result = map[string]any{"ok": true}
+		case "list":
+			result, err = s.Cluster.ListSessions(ctx, req.Device)
+		default:
+			http.Error(w, "unknown action", 400)
+			return
+		}
+	} else {
+		if s.LocalSessions == nil {
+			http.Error(w, "local session manager unavailable", 503)
+			return
+		}
+		switch action {
+		case "start":
+			if req.SessionID == "" {
+				req.SessionID, err = newSessionID()
+				if err != nil {
+					http.Error(w, "session id generation failed", 500)
+					return
+				}
+			}
+			var sess *executor.Session
+			sess, err = s.LocalSessions.Start(context.Background(), req.SessionID, req.Command, req.Workdir, owner, req.RingCapacity)
+			if err == nil {
+				result = map[string]any{"session_id": sess.ID, "started_at": sess.StartedAt.Unix()}
+			}
+		case "read":
+			var sess *executor.Session
+			sess, ok := s.LocalSessions.Get(req.SessionID)
+			if !ok {
+				err = fmt.Errorf("session not found")
+				break
+			}
+			if !sess.Authorized(owner) {
+				err = fmt.Errorf("session ownership denied")
+				break
+			}
+			code, done, _ := sess.Status()
+			stdout, stderr := sess.Output()
+			result = cluster.ProcessOutputPayload{SessionID: req.SessionID, Stdout: stdout, Stderr: stderr, ExitCode: code, Running: !done}
+		case "input":
+			var sess *executor.Session
+			sess, ok := s.LocalSessions.Get(req.SessionID)
+			if !ok {
+				err = fmt.Errorf("session not found")
+				break
+			}
+			if !sess.Authorized(owner) {
+				err = fmt.Errorf("session ownership denied")
+				break
+			}
+			err = sess.Stdin([]byte(req.Data))
+			result = map[string]any{"ok": true}
+		case "stop":
+			var sess *executor.Session
+			sess, ok := s.LocalSessions.Get(req.SessionID)
+			if !ok {
+				err = fmt.Errorf("session not found")
+				break
+			}
+			if !sess.Authorized(owner) {
+				err = fmt.Errorf("session ownership denied")
+				break
+			}
+			err = s.LocalSessions.Stop(req.SessionID, true)
+			result = map[string]any{"ok": true}
+		case "list":
+			result = s.LocalSessions.List()
+		default:
+			http.Error(w, "unknown action", 400)
+			return
 		}
-		result, err = s.Cluster.StartProcess(ctx, req.Device, cluster.StartProcessPayload{Command: req.Command, Workdir: req.Workdir, SessionID: req.SessionID, Owner: sessionOwner(r), RingCapacity: req.RingCapacity})
-	case "read":
-		result, err = s.Cluster.ReadProcessOutput(ctx, req.Device, req.SessionID, sessionOwner(r))
-	case "input":
-		err = s.Cluster.InteractProcess(ctx, req.Device, req.SessionID, req.Data, sessionOwner(r))
-		result = map[string]any{"ok": true}
-	case "stop":
-		err = s.Cluster.ForceTerminate(ctx, req.Device, req.SessionID, sessionOwner(r))
-		result = map[string]any{"ok": true}
-	case "list":
-		result, err = s.Cluster.ListSessions(ctx, req.Device)
-	default:
-		http.Error(w, "unknown action", 400)
-		return
 	}
 	if err != nil {
 		http.Error(w, err.Error(), 409)

internal/api/api_test.go

@@ -9,6 +9,7 @@ import (
 	"strings"
 	"sync"
 	"testing"
+	"time"
 
 	"github.com/wjyhk/vps-commander/internal/executor"
 	"github.com/wjyhk/vps-commander/internal/storage"
@@ -92,6 +93,38 @@ func TestSecurityEffectiveModeAndExecGate(t *testing.T) {
 	}
 }
 
+func TestProcessSessionLocalSnakeCaseAndLifecycle(t *testing.T) {
+	dir := t.TempDir()
+	store, err := storage.Open(filepath.Join(dir, "db.sqlite"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer store.DB.Close()
+	if err := store.SetGlobalSecurityMode("high"); err != nil {
+		t.Fatal(err)
+	}
+	s := &Server{Exec: executor.Local{MaxOutput: 1024 * 1024}, Store: store, LocalName: "local", LocalSessions: executor.NewSessionManager(0)}
+	if err := s.LoadSecurityPolicySnapshot(); err != nil {
+		t.Fatal(err)
+	}
+	req := httptest.NewRequest(http.MethodPost, "/api/v1/process/session?action=start", strings.NewReader(`{"device":"local","session_id":"session-test-001","command":"sleep 0.2; printf local-ok"}`))
+	w := httptest.NewRecorder()
+	s.processSession(w, req)
+	if w.Code != http.StatusOK {
+		t.Fatalf("start status=%d body=%s", w.Code, w.Body.String())
+	}
+	time.Sleep(1 * time.Second)
+	req = httptest.NewRequest(http.MethodPost, "/api/v1/process/session?action=read", strings.NewReader(`{"device":"local","session_id":"session-test-001"}`))
+	w = httptest.NewRecorder()
+	s.processSession(w, req)
+	if w.Code != http.StatusOK {
+		t.Fatalf("read status=%d body=%s", w.Code, w.Body.String())
+	}
+	if !strings.Contains(w.Body.String(), "local-ok") {
+		t.Fatalf("missing session output: %s", w.Body.String())
+	}
+}
+
 func TestEditBlockSecurityAndReplacement(t *testing.T) {
 	dir := t.TempDir()
 	store, err := storage.Open(filepath.Join(dir, "db.sqlite"))

internal/api/mcp_adapter.go

@@ -98,6 +98,33 @@ func (b *MCPBackend) ExecCommand(ctx context.Context, clientIP, device, command,
 	}, nil
 }
 
+func (b *MCPBackend) DiagnosticCommand(ctx context.Context, clientIP, device string, req executor.DiagnosticRequest) (any, error) {
+	if device == "" {
+		device = b.server.LocalName
+	}
+	if err := b.authorize(device, "diagnostic", clientIP); err != nil {
+		return nil, err
+	}
+	if req.Action == "" {
+		return nil, errors.New("diagnostic action required")
+	}
+	if b.server.isRemote(device) {
+		if b.server.Cluster == nil || !b.server.Cluster.Online(device) {
+			return nil, fmt.Errorf("device %s is offline", device)
+		}
+		res, err := b.server.Cluster.Diagnostic(ctx, device, req)
+		if err != nil {
+			return nil, err
+		}
+		return res, nil
+	}
+	res, err := executor.Diagnostic(ctx, req)
+	if err != nil {
+		return nil, err
+	}
+	return res, nil
+}
+
 func (b *MCPBackend) ReadFile(ctx context.Context, clientIP, device, path string, offset, limit int64) (any, error) {
 	if device == "" {
 		device = b.server.LocalName

internal/cluster/manager.go

@@ -13,6 +13,7 @@ import (
 	"time"
 
 	"github.com/gorilla/websocket"
+	"github.com/wjyhk/vps-commander/internal/executor"
 	"github.com/wjyhk/vps-commander/internal/security"
 	"github.com/wjyhk/vps-commander/internal/storage"
 	"github.com/wjyhk/vps-commander/internal/sysinfo"
@@ -464,6 +465,19 @@ func (m *Manager) Exec(ctx context.Context, name string, p ExecPayload) (Result,
 	return Result{msg.ExitCode, msg.Stdout, msg.Stderr, msg.DurationMS}, nil
 }
 
+func (m *Manager) Diagnostic(ctx context.Context, name string, req executor.DiagnosticRequest) (executor.DiagnosticResult, error) {
+	msg, err := m.call(ctx, name, "diagnostic_command", req)
+	if err != nil {
+		return executor.DiagnosticResult{}, err
+	}
+	b, _ := json.Marshal(msg.Payload)
+	var out executor.DiagnosticResult
+	if err := json.Unmarshal(b, &out); err != nil {
+		return out, err
+	}
+	return out, nil
+}
+
 func (m *Manager) ReadFile(ctx context.Context, name string, p FileReadPayload) (FileResult, error) {
 	msg, err := m.call(ctx, name, "file_read", p)
 	if err != nil {

internal/executor/diagnostic.go

@@ -0,0 +1,117 @@
+package executor
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"os/exec"
+	"strconv"
+	"strings"
+	"time"
+)
+
+// DiagnosticRequest is a structured, read-only diagnostic operation. It is deliberately
+// not a shell command: every action maps to a fixed executable/argument set.
+type DiagnosticRequest struct {
+	Action string `json:"action"`
+	Target string `json:"target,omitempty"`
+	Lines  int    `json:"lines,omitempty"`
+}
+
+type DiagnosticResult struct {
+	Action     string `json:"action"`
+	Target     string `json:"target,omitempty"`
+	ExitCode   int    `json:"exit_code"`
+	Stdout     string `json:"stdout"`
+	Stderr     string `json:"stderr"`
+	DurationMS int64  `json:"duration_ms"`
+}
+
+func Diagnostic(ctx context.Context, req DiagnosticRequest) (DiagnosticResult, error) {
+	args, err := diagnosticArgs(req)
+	if err != nil {
+		return DiagnosticResult{}, err
+	}
+	if req.Lines <= 0 {
+		req.Lines = 100
+	}
+	if req.Lines > 500 {
+		req.Lines = 500
+	}
+	// docker logs is the only operation where a caller-controlled numeric limit is useful.
+	if req.Action == "docker_logs" {
+		args = []string{"logs", "--tail", strconv.Itoa(req.Lines), req.Target}
+	}
+
+	start := time.Now()
+	c, cancel := context.WithTimeout(ctx, 15*time.Second)
+	defer cancel()
+	cmd := exec.CommandContext(c, args[0], args[1:]...)
+	out, runErr := cmd.CombinedOutput()
+	const maxOutput = 256 * 1024
+	if len(out) > maxOutput {
+		out = append(out[:maxOutput], []byte("\n[output truncated at 256 KiB]\n")...)
+	}
+	code := 0
+	if cmd.ProcessState != nil {
+		code = cmd.ProcessState.ExitCode()
+	}
+	result := DiagnosticResult{Action: req.Action, Target: req.Target, ExitCode: code, Stdout: string(out), DurationMS: time.Since(start).Milliseconds()}
+	if runErr != nil && code == -1 {
+		return result, fmt.Errorf("diagnostic failed: %w", runErr)
+	}
+	return result, nil
+}
+
+func diagnosticArgs(req DiagnosticRequest) ([]string, error) {
+	target := strings.TrimSpace(req.Target)
+	switch req.Action {
+	case "docker_ps":
+		if target != "" {
+			return nil, errors.New("docker_ps does not accept target")
+		}
+		return []string{"docker", "ps", "--no-trunc"}, nil
+	case "docker_logs":
+		if target == "" || strings.ContainsAny(target, " \t\r\n;|&<>$`") {
+			return nil, errors.New("invalid docker container target")
+		}
+		return []string{"docker", "logs", "--tail", "100", target}, nil
+	case "docker_inspect":
+		if target == "" || strings.ContainsAny(target, " \t\r\n;|&<>$`") {
+			return nil, errors.New("invalid docker target")
+		}
+		return []string{"docker", "inspect", target}, nil
+	case "systemctl_status":
+		if target == "" || strings.ContainsAny(target, " \t\r\n;|&<>$`") {
+			return nil, errors.New("invalid systemd unit")
+		}
+		return []string{"systemctl", "status", "--no-pager", "--full", target}, nil
+	case "journalctl":
+		if target == "" || strings.ContainsAny(target, " \t\r\n;|&<>$`") {
+			return nil, errors.New("invalid journal unit")
+		}
+		return []string{"journalctl", "--no-pager", "-u", target, "-n", "100"}, nil
+	case "ss_listen":
+		if target != "" {
+			return nil, errors.New("ss_listen does not accept target")
+		}
+		return []string{"ss", "-lntp"}, nil
+	case "df":
+		if target != "" {
+			return nil, errors.New("df does not accept target")
+		}
+		return []string{"df", "-h"}, nil
+	case "free":
+		if target != "" {
+			return nil, errors.New("free does not accept target")
+		}
+		return []string{"free", "-h"}, nil
+	case "uptime":
+		if target != "" {
+			return nil, errors.New("uptime does not accept target")
+		}
+		return []string{"uptime"}, nil
+	default:
+		return nil, fmt.Errorf("unsupported diagnostic action: %s", req.Action)
+	}
+}

internal/executor/diagnostic_test.go

@@ -0,0 +1,24 @@
+package executor
+
+import (
+	"context"
+	"strings"
+	"testing"
+)
+
+func TestDiagnosticRejectsShellInjection(t *testing.T) {
+	_, err := Diagnostic(context.Background(), DiagnosticRequest{Action: "docker_logs", Target: "edgeever;rm -rf /"})
+	if err == nil || !strings.Contains(err.Error(), "invalid docker") {
+		t.Fatalf("expected target rejection, got %v", err)
+	}
+}
+
+func TestDiagnosticAllowsReadOnlyUptime(t *testing.T) {
+	got, err := Diagnostic(context.Background(), DiagnosticRequest{Action: "uptime"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got.ExitCode != 0 || strings.TrimSpace(got.Stdout) == "" {
+		t.Fatalf("unexpected uptime result: %#v", got)
+	}
+}

internal/mcp/mcp.go

@@ -12,6 +12,7 @@ import (
 	"time"
 
 	"github.com/wjyhk/vps-commander/internal/auth"
+	"github.com/wjyhk/vps-commander/internal/executor"
 )
 
 // MCP Protocol / JSON-RPC 2.0 Types
@@ -77,6 +78,10 @@ type ActionBackend interface {
 	EditBlock(ctx context.Context, clientIP, device, path, oldText, newText string) (any, error)
 }
 
+type DiagnosticBackend interface {
+	DiagnosticCommand(ctx context.Context, clientIP, device string, req executor.DiagnosticRequest) (any, error)
+}
+
 type ExtendedBackend interface {
 	ReadMultipleFiles(ctx context.Context, clientIP, device string, paths []string, limit int64) (any, error)
 	CreateDirectory(ctx context.Context, clientIP, device, path string) (any, error)
@@ -280,12 +285,14 @@ func (s *Server) handleRequest(ctx context.Context, r *http.Request, req JSONRPC
 			{Name: "create_directory", Description: "创建目录", InputSchema: InputSchema{Type: "object", Properties: map[string]PropertyDef{"device": {Type: "string", Description: "目标设备"}, "path": {Type: "string", Description: "目录路径"}}, Required: []string{"path"}}},
 			{Name: "move_file", Description: "移动或重命名文件", InputSchema: InputSchema{Type: "object", Properties: map[string]PropertyDef{"device": {Type: "string", Description: "目标设备"}, "source": {Type: "string", Description: "源路径"}, "destination": {Type: "string", Description: "目标路径"}}, Required: []string{"source", "destination"}}},
 			{Name: "list_processes", Description: "列出受控节点当前进程", InputSchema: InputSchema{Type: "object", Properties: map[string]PropertyDef{"device": {Type: "string", Description: "目标设备"}}}},
+			{Name: "diagnostic_command", Description: "执行严格白名单的只读系统诊断，不接受任意 Shell", InputSchema: InputSchema{Type: "object", Properties: map[string]PropertyDef{"device": {Type: "string", Description: "目标设备"}, "action": {Type: "string", Description: "docker_ps|docker_logs|docker_inspect|systemctl_status|journalctl|ss_listen|df|free|uptime"}, "target": {Type: "string", Description: "容器名或 systemd unit（按 action 使用）"}, "lines": {Type: "integer", Description: "docker_logs 最大行数，1-500"}}, Required: []string{"action"}}},
 			{
 				Name:        "edit_block",
 				Description: "对指定文件执行唯一文本块的原子替换；old_text 必须恰好匹配一次",
 				InputSchema: InputSchema{Type: "object", Properties: map[string]PropertyDef{"device": {Type: "string", Description: "目标节点名称"}, "path": {Type: "string", Description: "目标文件路径"}, "old_text": {Type: "string", Description: "必须唯一匹配的原文本块"}, "new_text": {Type: "string", Description: "替换后的文本块"}}, Required: []string{"path", "old_text", "new_text"}},
 			},
 			{
+				Name:        "write_file",
 				Description: "向指定受控节点创建或覆写文件",
 				InputSchema: InputSchema{
 					Type: "object",
@@ -429,6 +436,24 @@ func (s *Server) callTool(ctx context.Context, clientIP string, params CallToolP
 			return map[string]string{"error": err.Error()}, true
 		}
 		return res, false
+	case "diagnostic_command":
+		var a struct {
+			Device string `json:"device"`
+			Action string `json:"action"`
+			Target string `json:"target"`
+			Lines  int    `json:"lines"`
+		}
+		_ = json.Unmarshal(params.Arguments, &a)
+		b, ok := s.Backend.(DiagnosticBackend)
+		if !ok {
+			return map[string]string{"error": "diagnostic tools unavailable"}, true
+		}
+		res, err := b.DiagnosticCommand(ctx, clientIP, a.Device, executor.DiagnosticRequest{Action: a.Action, Target: a.Target, Lines: a.Lines})
+		if err != nil {
+			return map[string]string{"error": err.Error()}, true
+		}
+		return res, false
+
 	case "list_processes":
 		var a struct {
 			Device string `json:"device"`

internal/mcp/mcp_test.go

@@ -79,9 +79,26 @@ func TestMCPInitializeAndListTools(t *testing.T) {
 		t.Fatalf("expected map result")
 	}
 	tools, ok := resMap["tools"].([]any)
-	if !ok || len(tools) != 11 {
+	if !ok || len(tools) != 12 {
 		t.Fatalf("expected 11 tools, got %d", len(tools))
 	}
+	foundWriteFile := false
+	for _, raw := range tools {
+		tool, ok := raw.(map[string]any)
+		if !ok {
+			t.Fatalf("tool entry is not an object: %#v", raw)
+		}
+		name, _ := tool["name"].(string)
+		if name == "" {
+			t.Fatalf("tool entry has empty name: %#v", tool)
+		}
+		if name == "write_file" {
+			foundWriteFile = true
+		}
+	}
+	if !foundWriteFile {
+		t.Fatalf("write_file missing from tools/list")
+	}
 }
 
 func TestMCPCallTool(t *testing.T) {

internal/security/registry.go

@@ -15,6 +15,7 @@ var registry = map[string]ToolSpec{
 	"start_search":            {Name: "start_search", RequiredRisk: RiskLow, Category: "search", ReadOnly: true},
 	"get_more_search_results": {Name: "get_more_search_results", RequiredRisk: RiskLow, Category: "search", ReadOnly: true},
 	"list_processes":          {Name: "list_processes", RequiredRisk: RiskLow, Category: "process", ReadOnly: true},
+	"diagnostic_command":      {Name: "diagnostic_command", RequiredRisk: RiskLow, Category: "diagnostic", ReadOnly: true},
 	"list_sessions":           {Name: "list_sessions", RequiredRisk: RiskLow, Category: "process", ReadOnly: true},
 	"read_process_output":     {Name: "read_process_output", RequiredRisk: RiskLow, Category: "process", ReadOnly: true},
 	"edit_block":              {Name: "edit_block", RequiredRisk: RiskMedium, Category: "file", ReadOnly: false},
@@ -31,7 +32,7 @@ var registry = map[string]ToolSpec{
 
 var actionTools = map[string]string{
 	"read": "read_file", "probe": "read_file", "write": "write_file",
-	"mcp": "mcp_call", "exec": "exec_command",
+	"mcp": "mcp_call", "exec": "exec_command", "diagnostic": "diagnostic_command",
 }
 
 func Lookup(name string) (ToolSpec, bool) {