feat: 完善交互式部署与 MCP/LLM 配置能力

- 新增 MCP client 配置加载，支持 CLI/chat 通过配置文件接入 MCP - 完善 chat 交互命令，支持参数查看、事件查看、checkpoint 列表与加载 - 增加 LLM action 后诊断能力，支持真实 LLM 和本地规则兜底 - 将 chat 人工确认点接入 LangGraph interrupt/checkpointer - 更新 README、流程图、待办文档和打包说明 - 补充相关单元测试
2026-06-01 16:45:52 +08:00 · 2026-06-01 16:45:52 +08:00 · d01c4d3d06
commit d01c4d3d06
parent ba85e61379
24 changed files with 1189 additions and 27 deletions
--- a/README.md
+++ b/README.md
@ -28,6 +28,7 @@ pam_deploy_graph/
  params_loader.py     # 读取 JSON 或 config.txt 风格参数文件
  llm/                 # LLM structured output 接口、真实 HTTP client、提示词、规则 fallback 和 guardrails
  graph.py             # LangGraph StateGraph 集成入口
+  langgraph_runtime.py # chat 人工确认点的 LangGraph interrupt 运行器
  mcp_client.py        # MCP session/callable adapter 与 client 配置读取
  interactive.py       # 常驻式 CLI 对话框，会话命令、确认和续跑
  cli.py               # CLI 入口
@ -42,10 +43,12 @@ tests/

 docs/
  current_logic_flow.md      # 当前整体逻辑结构流程图
+  todo.md                    # chat 优化和 LLM action 后分析待办

 packaging/
  build_linux_self_contained.sh  # Linux 解压即用包构建脚本
  README_linux_package.md        # Linux 打包说明和包大小评估
+  mcp_client.example.json        # MCP stdio 配置示例
 ```

 ## 当前进度
@ -70,11 +73,15 @@ packaging/
 - 增加规则 fallback `RuleBasedLlmClient`，用于本地开发和测试。
 - 增加 LLM 输出 guardrails，禁止计划中出现可执行脚本命令和非法 action。
 - 引入 `langgraph` 依赖，并提供 `build_langgraph()` 图工厂。
+- chat 人工确认点已接入 LangGraph interrupt/checkpointer：`run` 到待回滚确认时暂停，`approve/reject` 通过 `Command(resume=...)` 恢复。
 - 引入 MCP client adapter，可包装 SDK session 或普通 callable，并提供 JSON client 配置读取。
+- CLI/chat 支持 `--mcp-config` 直接加载 stdio MCP 配置并构造 MCP runner。
 - 本地已安装 `langgraph` 和 `mcp`，并完成 LangGraph fake 全局流程 smoke。
 - CLI `analyze` 输出已做敏感字段脱敏。
- 增加 `chat` 常驻式 CLI 对话框，支持自然语言分析、参数设置、执行确认、回滚确认、状态查看和续跑。
- 添加基础测试，当前本地结果为 `31 passed, 1 skipped`。
+- 增加 `chat` 常驻式 CLI 对话框，支持自然语言分析、参数设置、执行确认、回滚确认、状态查看、事件查看、checkpoint 选择和续跑。
+- chat 可选启用 `rich` / `prompt_toolkit`，支持更清晰输出、命令补全和输入历史。
+- 增加 action 后 LLM/规则诊断，可通过 `--analyze-actions` 或 `llm action-analysis on` 显式开启。
+- 添加基础测试，当前本地结果为 `37 passed, 1 skipped`。

 未完成：

@ -108,16 +115,25 @@ python -m pam_deploy_graph.cli analyze \

 ## MCP Client 配置

-真实 MCP session 由外部接入，Agent 只依赖同步 `call_tool(name, arguments)` 接口。接入方式：
+CLI/chat 已支持通过 `--mcp-config` 直接加载 MCP 配置。当前内置支持 stdio transport；配置文件里提供 MCP server 启动命令后，Agent 会在调用 PAM_NODE action 时创建 MCP stdio session。
+
+CLI 示例：
+
+```bash
+python -m pam_deploy_graph.cli chat \
+  --config doc_scripts/config.txt.example \
+  --strategy hybrid_node_mcp \
+  --mcp-config mcp_client.json \
+  --checkpoint runtime/checkpoints/demo.json
+```
+
+代码内嵌方式：

 ```python
 from pam_deploy_graph.agent import PamDeployAgent
-from pam_deploy_graph.mcp_client import SessionMcpToolClient, load_mcp_client_config
-from pam_deploy_graph.mcp_runner import McpActionRunner
+from pam_deploy_graph.mcp_factory import build_mcp_runner_from_config

-config = load_mcp_client_config("mcp_client.json")
-client = SessionMcpToolClient(session)  # session 是你接入真实 MCP 后得到的 SDK session
-runner = McpActionRunner(client=client, tool_names=config.tool_names or None)
+runner = build_mcp_runner_from_config("mcp_client.json")
 agent = PamDeployAgent(mcp_runner=runner)
 ```

@ -126,6 +142,14 @@ agent = PamDeployAgent(mcp_runner=runner)
 ```json
 {
  "server_name": "pam-node-prod",
+  "transport": "stdio",
+  "command": "/opt/pam-node-mcp/server",
+  "args": ["--stdio"],
+  "cwd": "/opt/pam-node-mcp",
+  "env": {
+    "PAM_NODE_ENV": "prod"
+  },
+  "timeout_seconds": 60,
  "tool_names": {
    "get-online-ips": "pam_get_online_ips",
    "create-download-task": "pam_create_download_task",
@ -164,7 +188,7 @@ dist/linux_self_contained/pam-deploy-agent-linux-x86_64/
 dist/linux_self_contained/pam-deploy-agent-linux-x86_64.tar.gz
 ```

-发布包内的 `doc_scripts` 只包含运行必需文件：`deploy.sh`、`config.txt.example`、`PAM_AUTO_DEPLY_SKILL.md`。发布包内的 `README.md` 使用 `packaging/README_packaged_agent.md`，只介绍打包后 Agent 的使用方式。
+发布包内的 `doc_scripts` 只包含运行必需文件：`deploy.sh`、`config.txt.example`、`PAM_AUTO_DEPLY_SKILL.md`。发布包内的 `README.md` 使用 `packaging/README_packaged_agent.md`，只介绍打包后 Agent 的使用方式；同时会带上 `mcp_client.example.json` 作为 MCP 配置示例。

 目标机器解压后运行：

@ -192,12 +216,18 @@ PAM> set VERSION_NUMBER=2.0.6
 PAM> run
 即将执行真实 action；确认执行请输入 yes: yes
 PAM> status
+PAM> params
+PAM> events 5
+PAM> llm action-analysis on
+PAM> mcp config mcp_client.example.json
+PAM> list checkpoints
+PAM> load checkpoint runtime/checkpoints/chat-demo.json
 PAM> approve
 PAM> resume
 PAM> exit
 ```

-`chat` 默认仍要求在会话内显式输入 `run` 和 `yes` 才会执行 action；如果某个 IP 失败，会提示输入 `approve` 或 `reject [原因]`。`chat` 也支持 `--llm-base-url` / `--llm-api-key` / `--llm-model`，配置方式和 `analyze` 一致。
+`chat` 默认仍要求在会话内显式输入 `run`，并确认参数、目标 IP 范围和最终执行后才会执行 action；如果某个 IP 失败，会通过 LangGraph interrupt 暂停并提示输入 `approve` 或 `reject [原因]`，确认后恢复同一个图线程继续执行。`chat` 也支持 `--llm-base-url` / `--llm-api-key` / `--llm-model`、`--mcp-config` 和 `--analyze-actions`。

 预演：

@ -248,5 +278,5 @@ pytest -q

 1. 接入真实 PAM_NODE MCP session，并用 `SessionMcpToolClient` 包装。
 2. 在测试环境中做 smoke：HOME 脚本 `get-token/get-node-url` + NODE MCP `get-online-ips`。
-3. 把当前 checkpoint/confirmation 语义继续接入 LangGraph interrupt/checkpointer。
+3. 在测试环境验证真实脚本 action 的失败、回滚确认和续跑链路。
 4. 继续细化参数确认、IP 范围确认的交互式 UI 或上层编排。
--- a/docs/current_logic_flow.md
+++ b/docs/current_logic_flow.md
@ -20,20 +20,25 @@ flowchart TD

    CLI --> AGENT[PamDeployAgent]
    CHAT --> AGENT
+    CHAT --> LGR[langgraph_runtime.py chat interrupt 运行器]
    PARAMS --> AGENT
    RULE --> AGENT
    REAL --> AGENT

+    LGR --> AGENT
+    LGR --> LGCHECK[LangGraph InMemorySaver checkpointer]
    AGENT --> ROUTER[ActionRouter]
    ROUTER --> SCRIPT[ScriptActionRunner]
    ROUTER --> MCP[McpActionRunner]
    ROUTER --> FAKE[FakeActionRunner]

    SCRIPT --> DEPLOY[doc_scripts/deploy.sh 或 deploy.ps1]
-    MCP --> MCPCLIENT[mcp_client.py: Session/Function adapter]
+    MCP --> MCPFACTORY[mcp_factory.py 读取 --mcp-config]
+    MCPFACTORY --> MCPCLIENT[mcp_client.py: stdio/Session/Function adapter]
    FAKE --> FIXTURE[测试 fixture 或默认 fake 返回值]

    AGENT --> CHECKPOINT[checkpoint_store.py]
+    AGENT --> ACTIONLLM[action 后 LLM/规则诊断]
    AGENT --> REPORT[render_report 部署报告]
 ```

@ -99,6 +104,22 @@ flowchart LR
    C -- PAM_NODE action --> NM[MCP tool 执行]
 ```

+## action 后诊断
+
+```mermaid
+flowchart TD
+    A[action 执行完成] --> B{是否开启 analyze-actions}
+    B -- 否 --> X[只记录 ACTION_DONE/ACTION_FAIL]
+    B -- 是 --> C[整理 ActionResult 和 AgentState 摘要]
+    C --> D[敏感字段脱敏并截断长日志]
+    D --> E{真实 LLM 是否配置}
+    E -- 是 --> F[OpenAICompatibleLlmClient 输出结构化诊断]
+    E -- 否 --> G[RuleBasedLlmClient 本地规则诊断]
+    F --> H[追加 ACTION_ANALYSIS 事件]
+    G --> H
+    H --> I[诊断只作建议，不自动继续/回滚/改参数]
+```
+
 ## 失败、人工确认和续跑

 ```mermaid
@ -110,7 +131,11 @@ flowchart TD
    E --> F[设置 pending_confirmation=rollback-ip:IP]
    F --> G[保存 checkpoint 并暂停]

-    G --> H{用户决定}
+    G --> LG{是否来自 chat}
+    LG -- 是 --> LGI[LangGraph interrupt 输出确认请求]
+    LGI --> LGRS[approve/reject 通过 Command resume 恢复]
+    LGRS --> H{用户决定}
+    LG -- 否 --> H{用户决定}
    H -- approve --> I[confirm_pending 执行 rollback-ip]
    I --> J{rollback 是否成功}
    J -- 是 --> K[清空 pending_confirmation]
@ -128,10 +153,11 @@ flowchart TD
 - `ip_states[ip].status == SUCCESS`：成功 IP 会跳过。
 - `ip_states[ip].completed_steps`：同一个 IP 已完成的 action 会跳过。
 - `pending_confirmation`：存在待确认事项时，部署流程不继续执行，必须先 `approve` 或 `reject`。
+- chat 会话内的确认点由 `langgraph_runtime.py` 通过 LangGraph interrupt 和 InMemorySaver 托管；命令行一次性 `confirm/resume` 仍读取业务 checkpoint JSON。
 - checkpoint 为了真实续跑会保存完整参数，请放在受控目录中。

 ## 真实外部能力接入点

 - 真实 LLM：`llm.openai_compatible.OpenAICompatibleLlmClient`，通过 `PAM_LLM_BASE_URL`、`PAM_LLM_API_KEY`、`PAM_LLM_MODEL` 或 CLI 参数配置。
- 真实 MCP：外部建立 MCP session 后，用 `SessionMcpToolClient` 包装，再传给 `McpActionRunner`。
+- 真实 MCP：CLI/chat 可通过 `--mcp-config` 加载 stdio MCP 配置，内部由 `mcp_factory.py` 构造 `McpActionRunner`。
 - 真实脚本：PAM_HOME action 通过 `doc_scripts/deploy.sh` 或 `deploy.ps1` 调用。
--- a/docs/todo.md
+++ b/docs/todo.md
@ -0,0 +1,21 @@
+# 待办事项
+
+## chat 交互优化
+
+- [x] 使用 `rich` 输出表格、状态、错误和报告；未安装时自动降级为普通输出。
+- [x] 使用 `prompt_toolkit` 支持命令补全和历史记录；未安装时自动降级为 `input()`。
+- [x] 增加 `params` 命令，脱敏展示当前会话参数。
+- [x] 增加 `events` 命令，查看最近 action 执行记录。
+- [x] 增加 `load checkpoint` 和 `list checkpoints`，方便选择历史任务续跑。
+- [x] 增加参数确认和目标 IP 范围确认，不只在回滚阶段确认。
+- [x] 增加 LLM/MCP 配置热加载，例如 `llm config`、`mcp config`。
+- [x] 将 chat 的人工确认点接入 LangGraph interrupt/checkpointer；`run` 执行到回滚确认点后由 interrupt 暂停，`approve/reject` 通过 `Command(resume=...)` 恢复同一图线程。跨进程续跑仍保留业务 checkpoint JSON。
+
+## LLM action 后分析
+
+- [x] 每次 action 完成后，可把 `action`、`backend`、`ok`、`values`、`stderr`、`error_summary` 和当前 `AgentState` 摘要交给 LLM 分析。
+- [x] LLM 输出结构化结果：是否异常、异常等级、可能原因、建议动作、是否需要人工确认。
+- [x] LLM 分析只作为辅助建议，不直接决定继续执行、回滚或修改参数。
+- [x] 本地保留规则兜底：exit code、`verify-ip SUCCESS=false`、pending confirmation 等硬规则优先于 LLM。
+- [x] 对 LLM 输入做脱敏，禁止把 `CLIENT_SECRET`、token、Authorization、完整日志原文发送给模型。
+- [x] 通过 `--analyze-actions` 或 `llm action-analysis on` 显式开启，真实部署默认不启用。
--- a/packaging/README_linux_package.md
+++ b/packaging/README_linux_package.md
@ -19,7 +19,7 @@
 bash packaging/build_linux_self_contained.sh
 ```

-默认会安装 `.[mcp]`，即包含 MCP 可选依赖。如果只想打最小包：
+默认会安装 `.[mcp,chat]`，即包含 MCP 可选依赖和 chat 交互增强依赖。如果只想打最小包：

 ```bash
 PACKAGE_EXTRAS= bash packaging/build_linux_self_contained.sh
@ -42,6 +42,7 @@ pam-deploy-agent-linux-x86_64/
    deploy.sh
    config.txt.example
    PAM_AUTO_DEPLY_SKILL.md
+  mcp_client.example.json
  README.md
  LICENSE
 ```
@ -50,6 +51,7 @@ pam-deploy-agent-linux-x86_64/

 - `doc_scripts` 不会打入项目设计文档、测试脚本、Windows bat/PowerShell 脚本。
 - 发布包内的 `README.md` 来自 `packaging/README_packaged_agent.md`，只说明打包后 Agent 的使用方式。
+- 发布包内的 `mcp_client.example.json` 是 MCP stdio 配置示例，需要按真实 MCP server 修改。
 - 项目开发用 README 不会复制到发布包内。

 ## 解压后运行
--- a/packaging/README_packaged_agent.md
+++ b/packaging/README_packaged_agent.md
@ -12,6 +12,7 @@ pam-deploy-agent-linux-x86_64/
    deploy.sh            # Linux 脚本 action 入口
    config.txt.example   # 参数配置示例
    PAM_AUTO_DEPLY_SKILL.md
+  mcp_client.example.json
  README.md              # 当前说明
  LICENSE
 ```
@ -31,6 +32,9 @@ pam-deploy-agent-linux-x86_64/
 ./run.sh run-deploy --help
 ```

+发布包默认包含 `rich` 和 `prompt_toolkit`。如果终端支持，chat 会自动启用更清晰的输出、命令补全和输入历史；不可用时会自动降级为普通文本输入输出。
+chat 内的失败回滚确认由 LangGraph interrupt 托管；执行停在确认点后，输入 `approve` 或 `reject [原因]` 会恢复同一个图线程继续处理。
+
 ## 交互式使用

 推荐先用 fake 策略验证流程：
@ -39,6 +43,16 @@ pam-deploy-agent-linux-x86_64/
 ./run.sh chat --config doc_scripts/config.txt.example --strategy fake --checkpoint runtime/checkpoints/demo.json
 ```

+如果要启用 MCP，先按真实 MCP server 修改 `mcp_client.example.json`，再使用 `hybrid_node_mcp`：
+
+```bash
+./run.sh chat \
+  --config doc_scripts/config.txt.example \
+  --strategy hybrid_node_mcp \
+  --mcp-config mcp_client.example.json \
+  --checkpoint runtime/checkpoints/demo.json
+```
+
 进入对话框后可输入：

 ```text
@ -48,6 +62,12 @@ PAM> set VERSION_NUMBER=2.0.6
 PAM> run
 即将执行真实 action；确认执行请输入 yes: yes
 PAM> status
+PAM> params
+PAM> events 5
+PAM> llm action-analysis on
+PAM> mcp config mcp_client.example.json
+PAM> list checkpoints
+PAM> load checkpoint runtime/checkpoints/demo.json
 PAM> approve
 PAM> resume
 PAM> exit
@ -73,6 +93,28 @@ PAM> exit
 ./run.sh run-deploy --config doc_scripts/config.txt.example --strategy fake --checkpoint runtime/checkpoints/demo.json --confirm
 ```

+执行时开启 action 后诊断：
+
+```bash
+./run.sh run-deploy \
+  --config doc_scripts/config.txt.example \
+  --strategy fake \
+  --checkpoint runtime/checkpoints/demo.json \
+  --analyze-actions \
+  --confirm
+```
+
+使用 MCP 的完整部署：
+
+```bash
+./run.sh run-deploy \
+  --config doc_scripts/config.txt.example \
+  --strategy hybrid_node_mcp \
+  --mcp-config mcp_client.example.json \
+  --checkpoint runtime/checkpoints/demo.json \
+  --confirm
+```
+
 处理失败后的回滚确认：

 ```bash
@ -109,14 +151,54 @@ export PAM_LLM_MODEL="your-model-name"
  --llm-model your-model-name
 ```

+chat 内也可以热加载 LLM：
+
+```text
+PAM> llm config base_url=https://your-llm.example.com/v1 api_key=your-api-key model=your-model-name
+PAM> llm action-analysis on
+PAM> llm fallback
+```
+
 ## 策略说明

 - `fake`：全部使用 fake runner，不访问真实环境。
 - `script_only`：全部 action 走脚本。
 - `hybrid_node_mcp`：PAM_HOME 走脚本，PAM_NODE 走 MCP。

+## MCP 配置
+
+`--mcp-config` 指向 MCP client JSON 配置文件。当前支持 stdio transport：
+
+```json
+{
+  "server_name": "pam-node-prod",
+  "transport": "stdio",
+  "command": "/opt/pam-node-mcp/server",
+  "args": ["--stdio"],
+  "cwd": "/opt/pam-node-mcp",
+  "env": {
+    "PAM_NODE_ENV": "prod"
+  },
+  "timeout_seconds": 60,
+  "tool_names": {
+    "get-online-ips": "pam_get_online_ips",
+    "verify-ip": "pam_verify_ip",
+    "rollback-ip": "pam_rollback_ip"
+  }
+}
+```
+
+字段说明：
+
+- `command`：MCP server 启动命令。
+- `args`：MCP server 启动参数。
+- `cwd`：MCP server 工作目录，可为空。
+- `env`：传给 MCP server 的环境变量，可为空。
+- `timeout_seconds`：单次 tool 调用超时时间。
+- `tool_names`：Agent action 到 MCP tool name 的映射。
+
 ## 注意事项

 - 执行真实 action 前请确认配置文件中的 `HOME_BASE_URL`、`CLIENT_ID`、`CLIENT_SECRET`、`AIRPORT_CODE`、`APP_NAME`、`MODULE_NAME`、`VERSION_NUMBER`、`ZIP_FILE_PATH`。
 - `checkpoint` 会保存完整运行参数，请放在受控目录。
- 真实 MCP session 需要你在外部接入；当前包包含 MCP client adapter 和 action 映射能力。
+- `hybrid_node_mcp`、`resume`、`confirm` 如果需要执行 MCP action，请同时传入 `--mcp-config`。
--- a/packaging/build_linux_self_contained.sh
+++ b/packaging/build_linux_self_contained.sh
@ -9,7 +9,7 @@ cd "$ROOT_DIR"
 PYTHON_BIN="${PYTHON_BIN:-python3}"
 APP_NAME="pam-deploy-agent"
 RELEASE_NAME="${APP_NAME}-linux-x86_64"
-PACKAGE_EXTRAS="${PACKAGE_EXTRAS:-mcp}"
+PACKAGE_EXTRAS="${PACKAGE_EXTRAS:-mcp,chat}"
 BUILD_DIR="${BUILD_DIR:-$ROOT_DIR/build/linux_self_contained}"
 DIST_DIR="${DIST_DIR:-$ROOT_DIR/dist/linux_self_contained}"
 RELEASE_DIR="$DIST_DIR/$RELEASE_NAME"
@ -70,6 +70,7 @@ cp -a doc_scripts/PAM_AUTO_DEPLY_SKILL.md "$RELEASE_DIR/doc_scripts/PAM_AUTO_DEP
 chmod +x "$RELEASE_DIR/doc_scripts/deploy.sh"

 cp -a packaging/README_packaged_agent.md "$RELEASE_DIR/README.md"
+cp -a packaging/mcp_client.example.json "$RELEASE_DIR/mcp_client.example.json"
 cp -a LICENSE "$RELEASE_DIR/LICENSE"

 cat > "$RELEASE_DIR/run.sh" <<'RUN_SCRIPT'
@ -112,10 +113,19 @@ PAM 部署 Agent 解压即用包
  --target-ip <IP>
      指定目标工作站 IP。可重复传入多次。

+  --mcp-config <路径>
+      MCP client JSON 配置文件。hybrid_node_mcp 策略、resume 或 confirm
+      需要执行 MCP action 时使用。
+      示例：mcp_client.example.json
+
  --confirm
      非交互命令执行真实 action 前必须显式传入。
      chat 模式会在会话中要求输入 run 和 yes。

+  --analyze-actions
+      每个 action 完成后追加 LLM/规则诊断建议。诊断只作为辅助建议，
+      不会自动决定继续、回滚或修改参数。
+
 LLM 参数：
  --llm-base-url <URL>
      OpenAI-compatible LLM 服务地址，例如 https://example.com/v1
@ -134,6 +144,8 @@ LLM 环境变量：
 示例：
  ./run.sh chat --config doc_scripts/config.txt.example --strategy fake --checkpoint runtime/checkpoints/demo.json

+  ./run.sh chat --config doc_scripts/config.txt.example --strategy hybrid_node_mcp --mcp-config mcp_client.example.json --checkpoint runtime/checkpoints/demo.json
+
  ./run.sh analyze --config doc_scripts/config.txt.example --text "请用 MCP 预演部署 HET PAM Node 版本 2.0.5，不要动环境"

  ./run.sh run-deploy --config doc_scripts/config.txt.example --strategy fake --checkpoint runtime/checkpoints/demo.json --confirm
@ -148,7 +160,9 @@ LLM 环境变量：
 说明：
  1. 本包已包含 Python 运行时和 Python 依赖，目标机器不需要安装 Python 包。
  2. doc_scripts 只包含运行必需文件：deploy.sh、config.txt.example、PAM_AUTO_DEPLY_SKILL.md。
-  3. checkpoint 会保存完整运行参数，请放在受控目录。
+  3. mcp_client.example.json 是 MCP stdio 配置示例，需要按真实 MCP server 修改。
+  4. chat 内可使用 params、events、list checkpoints、load checkpoint、llm config、mcp config 等命令。
+  5. checkpoint 会保存完整运行参数，请放在受控目录。
 HELP_TEXT
 }

--- a/packaging/mcp_client.example.json
+++ b/packaging/mcp_client.example.json
@ -0,0 +1,23 @@
+{
+  "server_name": "pam-node-prod",
+  "transport": "stdio",
+  "command": "/opt/pam-node-mcp/server",
+  "args": ["--stdio"],
+  "cwd": "/opt/pam-node-mcp",
+  "env": {
+    "PAM_NODE_ENV": "prod"
+  },
+  "timeout_seconds": 60,
+  "tool_names": {
+    "get-online-ips": "pam_get_online_ips",
+    "create-download-task": "pam_create_download_task",
+    "poll-download-progress": "pam_poll_download_progress",
+    "upgrade-ip": "pam_upgrade_ip",
+    "poll-upgrade-progress": "pam_poll_upgrade_progress",
+    "start-ip": "pam_start_ip",
+    "stop-ip": "pam_stop_ip",
+    "verify-ip": "pam_verify_ip",
+    "download-log": "pam_download_log",
+    "rollback-ip": "pam_rollback_ip"
+  }
+}
--- a/pam_deploy_graph/agent.py
+++ b/pam_deploy_graph/agent.py
@ -7,6 +7,7 @@
 from __future__ import annotations

 import time
+from dataclasses import asdict
 from pathlib import Path
 from typing import Any

@ -33,6 +34,7 @@ class PamDeployAgent:
        mcp_runner: McpActionRunner | None = None,
        fake_runner: FakeActionRunner | None = None,
        llm_client: LlmClient | None = None,
+        action_analysis_enabled: bool = False,
    ) -> None:
        """初始化策略、脚本 runner、MCP runner、fake runner 和 LLM client。"""
        self.skill_policy = load_skill_policy(skill_path)
@ -41,6 +43,7 @@ class PamDeployAgent:
        self.fake_runner = fake_runner or FakeActionRunner()
        self.mcp_runner = mcp_runner
        self.llm_client = llm_client or RuleBasedLlmClient()
+        self.action_analysis_enabled = action_analysis_enabled
        self.router = ActionRouter(
            script_runner=self.script_runner,
            mcp_runner=mcp_runner,
@ -180,6 +183,7 @@ class PamDeployAgent:
                    "message": result.error_summary or "ok",
                }
            )
+            self._append_action_analysis(state, action, result)
            if not result.ok:
                state.last_failed_step = action
                self._save_checkpoint(state)
@ -243,6 +247,7 @@ class PamDeployAgent:
                        "message": result.error_summary or result.values.get("MESSAGE", "ok"),
                    }
                )
+                self._append_action_analysis(state, action, result, ip=ip)

                if failed:
                    self._record_ip_failure(state, ip, action, result.error_summary or str(result.values))
@ -322,6 +327,7 @@ class PamDeployAgent:
                "message": result.error_summary or result.values.get("MESSAGE", "ok"),
            }
        )
+        self._append_action_analysis(state, "rollback-ip", result, ip=ip)
        if result.ok:
            state.pending_confirmation = ""
            state.last_success_step = "rollback-ip"
@ -427,12 +433,61 @@ class PamDeployAgent:
                    "message": result.error_summary or "尽力下载日志失败",
                }
            )
+        self._append_action_analysis(state, "download-log", result, ip=ip)

    def _save_checkpoint(self, state: AgentState) -> None:
        """如果配置了 checkpoint 路径，则保存完整运行状态。"""
        if state.checkpoint_path:
            save_checkpoint(state, state.checkpoint_path, redact=False)

+    def _append_action_analysis(
+        self,
+        state: AgentState,
+        action: str,
+        result,
+        *,
+        ip: str | None = None,
+    ) -> None:
+        """启用 action 后分析时，把诊断结果追加到 events。"""
+        if not self.action_analysis_enabled:
+            return
+        try:
+            analysis = self.llm_client.analyze_action_result(
+                action=action,
+                result=result,
+                state_summary=self._state_summary_for_llm(state, ip=ip),
+            )
+        except Exception as exc:  # pragma: no cover - 诊断失败不应影响部署主流程
+            state.events.append(
+                {
+                    "type": "ACTION_ANALYSIS_FAIL",
+                    "stage": action,
+                    "ip": ip or "",
+                    "message": str(exc),
+                }
+            )
+            return
+        payload = asdict(analysis)
+        payload.update({"type": "ACTION_ANALYSIS", "stage": action})
+        if ip:
+            payload["ip"] = ip
+        state.events.append(payload)
+
+    def _state_summary_for_llm(self, state: AgentState, *, ip: str | None = None) -> dict[str, Any]:
+        """生成给 LLM action 分析使用的脱敏状态摘要。"""
+        return {
+            "run_id": state.run_id,
+            "execution_strategy": state.execution_strategy,
+            "completed_global_steps": state.completed_global_steps,
+            "online_ip_count": len(state.online_ips),
+            "target_ips": state.target_ips,
+            "current_ip": ip or "",
+            "current_ip_state": state.ip_states.get(ip, {}) if ip else {},
+            "pending_confirmation": state.pending_confirmation,
+            "last_success_step": state.last_success_step,
+            "last_failed_step": state.last_failed_step,
+        }
+
    def render_report(self, state: AgentState) -> str:
        """渲染当前部署状态报告。"""
        success = sum(1 for item in state.ip_states.values() if item.get("status") == "SUCCESS")
--- a/pam_deploy_graph/cli.py
+++ b/pam_deploy_graph/cli.py
@ -10,6 +10,7 @@ from .agent import PamDeployAgent
 from .checkpoint_store import load_agent_state, redact_mapping
 from .interactive import run_interactive_chat
 from .llm import build_llm_client
+from .mcp_factory import build_mcp_runner_from_config
 from .params_loader import load_params_file


@ -20,6 +21,16 @@ def add_llm_args(parser: argparse.ArgumentParser) -> None:
    parser.add_argument("--llm-model")


+def add_mcp_args(parser: argparse.ArgumentParser) -> None:
+    """为需要执行 MCP action 的子命令追加 MCP 配置参数。"""
+    parser.add_argument("--mcp-config", help="MCP client JSON 配置文件路径")
+
+
+def add_action_analysis_arg(parser: argparse.ArgumentParser) -> None:
+    """为执行类子命令追加 action 后诊断开关。"""
+    parser.add_argument("--analyze-actions", action="store_true", help="每个 action 后追加 LLM/规则诊断建议")
+
+
 def require_confirm(args: argparse.Namespace) -> None:
    """真实执行前强制要求命令行显式传入 --confirm。"""
    if not getattr(args, "confirm", False):
@ -54,12 +65,16 @@ def main() -> None:
    chat.add_argument("--target-ip", action="append", default=[])
    chat.add_argument("--checkpoint")
    add_llm_args(chat)
+    add_mcp_args(chat)
+    add_action_analysis_arg(chat)

    run = sub.add_parser("run-global")
    run.add_argument("--config", required=True)
    run.add_argument("--strategy", default="fake", choices=["hybrid_node_mcp", "script_only", "fake"])
    run.add_argument("--checkpoint")
    run.add_argument("--confirm", action="store_true")
+    add_mcp_args(run)
+    add_action_analysis_arg(run)

    deploy = sub.add_parser("run-deploy")
    deploy.add_argument("--config", required=True)
@ -67,16 +82,22 @@ def main() -> None:
    deploy.add_argument("--target-ip", action="append", default=[])
    deploy.add_argument("--checkpoint")
    deploy.add_argument("--confirm", action="store_true")
+    add_mcp_args(deploy)
+    add_action_analysis_arg(deploy)

    resume = sub.add_parser("resume")
    resume.add_argument("--checkpoint", required=True)
    resume.add_argument("--confirm", action="store_true")
+    add_mcp_args(resume)
+    add_action_analysis_arg(resume)

    confirm = sub.add_parser("confirm")
    confirm.add_argument("--checkpoint", required=True)
    confirm.add_argument("--decision", required=True, choices=["approve", "reject"])
    confirm.add_argument("--note", default="")
    confirm.add_argument("--confirm", action="store_true")
+    add_mcp_args(confirm)
+    add_action_analysis_arg(confirm)

    args = parser.parse_args()
    params = load_params_file(args.config) if getattr(args, "config", None) else {}
@ -87,7 +108,14 @@ def main() -> None:
            api_key=args.llm_api_key,
            model=args.llm_model,
        )
-    agent = PamDeployAgent(llm_client=llm_client)
+    mcp_runner = None
+    if getattr(args, "mcp_config", None):
+        mcp_runner = build_mcp_runner_from_config(args.mcp_config)
+    agent = PamDeployAgent(
+        llm_client=llm_client,
+        mcp_runner=mcp_runner,
+        action_analysis_enabled=bool(getattr(args, "analyze_actions", False)),
+    )

    if args.command == "analyze":
        result = agent.analyze_request(args.text, params)
--- a/pam_deploy_graph/interactive.py
+++ b/pam_deploy_graph/interactive.py
@ -3,12 +3,19 @@
 from __future__ import annotations

 import time
+import json
+import shlex
+import builtins
 from dataclasses import asdict
 from pathlib import Path
 from typing import Any, Callable

 from .agent import PamDeployAgent
 from .checkpoint_store import load_agent_state, redact_mapping
+from .langgraph_runtime import LangGraphDeploymentRuntime, LangGraphRunResult
+from .llm import build_llm_client
+from .llm.rule_based import RuleBasedLlmClient
+from .mcp_factory import build_mcp_runner_from_config
 from .models import AgentState, ExecutionStrategy

 InputFunc = Callable[[str], str]
@ -18,12 +25,20 @@ COMMAND_HELP = """可用命令：
  help                 显示帮助
  preview              查看当前参数和执行策略
  analyze <需求>       只做理解和计划，不执行
+  params               脱敏展示当前会话参数
+  events [数量]        查看最近 action 事件，默认 10 条
  set KEY=VALUE        修改当前会话参数
+  llm config KEY=VALUE 配置真实 LLM，支持 base_url/api_key/model
+  llm fallback         切回本地规则 fallback
+  llm action-analysis on|off  开关 action 后诊断
+  mcp config <路径>    加载 MCP client JSON 配置
  run                  创建部署任务并执行
  status               查看当前运行状态
  approve              确认待处理回滚
  reject [原因]        拒绝待处理回滚
  resume               从当前 checkpoint 续跑
+  list checkpoints     列出 checkpoint 目录下的 JSON 文件
+  load checkpoint <路径> 加载指定 checkpoint
  checkpoint           显示 checkpoint 路径
  exit                 退出

@ -51,10 +66,13 @@ class InteractiveCliSession:
        self.strategy = strategy
        self.checkpoint_path = checkpoint_path or _default_checkpoint_path()
        self.target_ips = list(target_ips or [])
-        self.input = input_func
-        self.output = output_func
+        self.input = _build_prompt_input(input_func)
+        self.output = _build_output_func(output_func)
        self.state: AgentState | None = None
        self.last_analysis: dict[str, Any] | None = None
+        self.llm_config: dict[str, str] = {}
+        self.mcp_config_path: str = ""
+        self.graph_runtime: LangGraphDeploymentRuntime | None = None

    def run(self) -> None:
        """启动 REPL 循环，直到用户 exit 或输入流结束。"""
@ -88,12 +106,24 @@ class InteractiveCliSession:
        if normalized == "preview":
            self.output(self.agent.preview(self.params, self.strategy))
            return True
+        if normalized == "params":
+            self._show_params()
+            return True
+        if normalized == "events":
+            self._show_events(rest.strip())
+            return True
        if normalized == "analyze":
            self._analyze(rest.strip())
            return True
        if normalized == "set":
            self._set_param(rest.strip())
            return True
+        if normalized == "llm":
+            self._configure_llm(rest.strip())
+            return True
+        if normalized == "mcp":
+            self._configure_mcp(rest.strip())
+            return True
        if normalized in ("run", "deploy", "execute"):
            self._run_deploy()
            return True
@ -112,6 +142,12 @@ class InteractiveCliSession:
        if normalized == "checkpoint":
            self.output(f"checkpoint: {self.checkpoint_path}")
            return True
+        if normalized == "list" and rest.strip().lower() == "checkpoints":
+            self._list_checkpoints()
+            return True
+        if normalized == "load" and rest.strip().lower().startswith("checkpoint"):
+            self._load_checkpoint(rest.strip()[len("checkpoint") :].strip())
+            return True

        self._analyze(text)
        return True
@ -159,12 +195,122 @@ class InteractiveCliSession:
        self.params[key] = value.strip()
        self.output(f"已设置 {key}")

+    def _show_params(self) -> None:
+        """脱敏展示当前会话参数。"""
+        self.output(_format_redacted_params(redact_mapping(self.params)))
+
+    def _show_events(self, count_text: str) -> None:
+        """展示最近若干条事件。"""
+        if self.state is None or not self.state.events:
+            self.output("当前没有事件。")
+            return
+        try:
+            count = int(count_text) if count_text else 10
+        except ValueError:
+            self.output("格式：events [数量]")
+            return
+        events = self.state.events[-max(count, 1) :]
+        self.output(json.dumps(redact_mapping(events), ensure_ascii=False, indent=2, default=str))
+
+    def _configure_llm(self, text: str) -> None:
+        """热加载 LLM 配置，或开关 action 后诊断。"""
+        if not text:
+            self.output("格式：llm config base_url=... api_key=... model=... | llm fallback | llm action-analysis on|off")
+            return
+        parts = shlex.split(text)
+        if parts[0] == "fallback":
+            self.agent.llm_client = RuleBasedLlmClient()
+            self.llm_config = {}
+            self.output("已切回本地规则 LLM fallback。")
+            return
+        if parts[0] == "action-analysis":
+            if len(parts) < 2 or parts[1] not in ("on", "off"):
+                self.output("格式：llm action-analysis on|off")
+                return
+            self.agent.action_analysis_enabled = parts[1] == "on"
+            self.output(f"action 后诊断已{'开启' if self.agent.action_analysis_enabled else '关闭'}。")
+            return
+        if parts[0] != "config":
+            self.output("未知 llm 命令。")
+            return
+        updates = _parse_key_values(parts[1:])
+        self.llm_config.update(updates)
+        try:
+            self.agent.llm_client = build_llm_client(
+                base_url=self.llm_config.get("base_url"),
+                api_key=self.llm_config.get("api_key"),
+                model=self.llm_config.get("model"),
+            )
+        except Exception as exc:
+            self.output(f"LLM 配置失败: {exc}")
+            return
+        safe = {**self.llm_config}
+        if safe.get("api_key"):
+            safe["api_key"] = "***"
+        self.output("LLM 配置已加载: " + json.dumps(safe, ensure_ascii=False))
+
+    def _configure_mcp(self, text: str) -> None:
+        """热加载 MCP client 配置。"""
+        command, _, path = text.partition(" ")
+        if command != "config" or not path.strip():
+            self.output("格式：mcp config <mcp_client.json>")
+            return
+        path = path.strip().strip('"')
+        try:
+            runner = build_mcp_runner_from_config(path)
+        except Exception as exc:
+            self.output(f"MCP 配置失败: {exc}")
+            return
+        self.agent.mcp_runner = runner
+        self.agent.router.mcp_runner = runner
+        self.mcp_config_path = path
+        self.output(f"MCP 配置已加载: {path}")
+
+    def _list_checkpoints(self) -> None:
+        """列出当前 checkpoint 目录下的 JSON 文件。"""
+        checkpoint_dir = Path(self.checkpoint_path).parent
+        if not checkpoint_dir.exists():
+            self.output(f"checkpoint 目录不存在: {checkpoint_dir}")
+            return
+        files = sorted(checkpoint_dir.glob("*.json"), key=lambda item: item.stat().st_mtime, reverse=True)
+        if not files:
+            self.output(f"checkpoint 目录没有 JSON 文件: {checkpoint_dir}")
+            return
+        lines = ["checkpoint 列表："]
+        for file in files[:20]:
+            lines.append(f"- {file}")
+        self.output("\n".join(lines))
+
+    def _load_checkpoint(self, path_text: str) -> None:
+        """加载指定 checkpoint 文件。"""
+        if not path_text:
+            self.output("格式：load checkpoint <路径>")
+            return
+        checkpoint = Path(path_text)
+        if not checkpoint.exists():
+            self.output(f"checkpoint 不存在: {checkpoint}")
+            return
+        self.state = load_agent_state(checkpoint)
+        self.state.checkpoint_path = str(checkpoint)
+        self.checkpoint_path = str(checkpoint)
+        self.params = dict(self.state.params)
+        self.strategy = self.state.execution_strategy
+        self.target_ips = list(self.state.target_ips)
+        self.graph_runtime = None
+        self.output(f"已加载 checkpoint: {checkpoint}")
+        if self.state.pending_confirmation:
+            self._print_confirmation()
+
    def _run_deploy(self) -> None:
        """在用户确认后创建状态并执行完整部署流程。"""
        if self.state and self.state.pending_confirmation:
            self._print_confirmation()
            return

+        if not self._confirm_params_and_scope():
+            self.output("已取消执行。")
+            return
+
        if not self._ask_yes_no("即将执行真实 action；确认执行请输入 yes: "):
            self.output("已取消执行。")
            return
@ -175,8 +321,20 @@ class InteractiveCliSession:
            checkpoint_path=self.checkpoint_path,
            target_ips=self.target_ips,
        )
+        self.graph_runtime = None
        self._execute_current_state()

+    def _confirm_params_and_scope(self) -> bool:
+        """执行前确认参数和目标 IP 范围。"""
+        self.output(_format_redacted_params(redact_mapping(self.params)))
+        if not self._ask_yes_no("确认以上参数请输入 yes: "):
+            return False
+        if self.target_ips:
+            self.output("目标 IP: " + ", ".join(self.target_ips))
+        else:
+            self.output("目标 IP: 未指定，将在 get-online-ips 后使用全部在线 IP。")
+        return self._ask_yes_no("确认目标范围请输入 yes: ")
+
    def _resume(self) -> None:
        """从内存状态或 checkpoint 文件继续执行部署流程。"""
        if self.state is None:
@ -186,6 +344,9 @@ class InteractiveCliSession:
                return
            self.state = load_agent_state(checkpoint)
            self.state.checkpoint_path = self.state.checkpoint_path or str(checkpoint)
+        if self.graph_runtime and self.graph_runtime.waiting_confirmation:
+            self._print_confirmation()
+            return
        self._execute_current_state()

    def _execute_current_state(self) -> None:
@ -193,7 +354,36 @@ class InteractiveCliSession:
        if self.state is None:
            self.output("当前没有运行状态。")
            return
-        self.state = self.agent.run_deploy_flow(self.state)
+        try:
+            if self.graph_runtime is None or not self.graph_runtime.waiting_confirmation:
+                self.graph_runtime = LangGraphDeploymentRuntime(agent=self.agent)
+            result = self.graph_runtime.start(self.state)
+        except RuntimeError as exc:
+            self.output(f"LangGraph 确认运行器不可用，降级为本地执行: {exc}")
+            self.graph_runtime = None
+            self.state = self.agent.run_deploy_flow(self.state)
+            self._print_state_report_and_checkpoint()
+            return
+        self._apply_graph_result(result)
+
+    def _apply_graph_result(self, result: LangGraphRunResult) -> None:
+        """把 LangGraph 运行结果同步回 chat 会话并输出用户可见状态。"""
+        if result.state is not None:
+            self.state = result.state
+        if self.state is None:
+            self.output("当前没有运行状态。")
+            return
+        self.output(result.report or self.agent.render_report(self.state))
+        if result.interrupted and result.confirmation:
+            self._print_confirmation_request(result.confirmation)
+        elif self.state.pending_confirmation:
+            self._print_confirmation()
+        self.output(f"checkpoint: {self.state.checkpoint_path or self.checkpoint_path}")
+
+    def _print_state_report_and_checkpoint(self) -> None:
+        """输出本地执行路径的状态报告和 checkpoint。"""
+        if self.state is None:
+            return
        self.output(self.agent.render_report(self.state))
        if self.state.pending_confirmation:
            self._print_confirmation()
@ -223,6 +413,15 @@ class InteractiveCliSession:
            self.output("当前没有待确认任务。")
            return

+        if self.graph_runtime and self.graph_runtime.waiting_confirmation:
+            try:
+                result = self.graph_runtime.resume(approved=approved, note=note)
+            except RuntimeError as exc:
+                self.output(f"LangGraph 确认恢复失败，降级为本地确认: {exc}")
+            else:
+                self._apply_graph_result(result)
+                return
+
        self.state = self.agent.confirm_pending(self.state, approved=approved, operator_note=note)
        self.output(self.agent.render_report(self.state))
        if self.state.pending_confirmation:
@ -235,6 +434,10 @@ class InteractiveCliSession:
        request = self.agent.build_confirmation_request(self.state)
        if not request:
            return
+        self._print_confirmation_request(request)
+
+    def _print_confirmation_request(self, request: dict[str, Any]) -> None:
+        """输出指定的人工确认请求。"""
        self.output("需要人工确认：")
        self.output(f"- type: {request.get('type')}")
        if request.get("ip"):
@ -307,3 +510,83 @@ def _format_redacted_params(params: dict[str, Any]) -> str:
    for key in sorted(params):
        lines.append(f"- {key}: {params[key]}")
    return "\n".join(lines)
+
+
+def _parse_key_values(parts: list[str]) -> dict[str, str]:
+    """解析 KEY=VALUE 参数列表。"""
+    values: dict[str, str] = {}
+    for part in parts:
+        if "=" not in part:
+            continue
+        key, value = part.split("=", 1)
+        if key:
+            values[key] = value
+    return values
+
+
+def _build_prompt_input(input_func: InputFunc) -> InputFunc:
+    """如果安装了 prompt_toolkit，则启用历史记录和命令补全。"""
+    if input_func is not builtins.input:
+        return input_func
+    try:
+        from prompt_toolkit import PromptSession
+        from prompt_toolkit.completion import WordCompleter
+        from prompt_toolkit.history import FileHistory
+    except ImportError:
+        return input_func
+
+    commands = [
+        "help",
+        "preview",
+        "analyze",
+        "params",
+        "events",
+        "set",
+        "llm config",
+        "llm fallback",
+        "llm action-analysis on",
+        "llm action-analysis off",
+        "mcp config",
+        "run",
+        "status",
+        "approve",
+        "reject",
+        "resume",
+        "list checkpoints",
+        "load checkpoint",
+        "checkpoint",
+        "exit",
+    ]
+    session = PromptSession(
+        history=FileHistory(str(Path("runtime") / "chat_history.txt")),
+        completer=WordCompleter(commands, ignore_case=True, sentence=True),
+    )
+    return session.prompt
+
+
+def _build_output_func(output_func: OutputFunc) -> OutputFunc:
+    """如果安装了 rich，则使用 rich 输出；否则保持原输出函数。"""
+    if output_func is not builtins.print:
+        return output_func
+    try:
+        from rich.console import Console
+        from rich.markdown import Markdown
+    except ImportError:
+        return output_func
+    console = Console()
+
+    def rich_print(value: str) -> None:
+        text = str(value)
+        stripped = text.lstrip()
+        if stripped.startswith("{") or stripped.startswith("["):
+            try:
+                console.print_json(text)
+                return
+            except Exception:
+                pass
+        if text.startswith("## ") or "\n| ---" in text:
+            console.print(Markdown(text))
+            return
+        console.print(text)
+
+    return rich_print
--- a/pam_deploy_graph/langgraph_runtime.py
+++ b/pam_deploy_graph/langgraph_runtime.py
@ -0,0 +1,148 @@
+"""chat 人工确认点的 LangGraph interrupt 运行器。"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any
+from uuid import uuid4
+
+from .agent import PamDeployAgent
+from .models import AgentState
+
+
+@dataclass(slots=True)
+class LangGraphRunResult:
+    """一次 LangGraph 执行或恢复后的结果摘要。"""
+
+    state: AgentState | None = None
+    report: str = ""
+    confirmation: dict[str, Any] = field(default_factory=dict)
+    interrupted: bool = False
+    chunks: list[dict[str, Any]] = field(default_factory=list)
+
+
+class LangGraphDeploymentRuntime:
+    """用 LangGraph interrupt/checkpointer 托管 chat 中的人工确认流程。"""
+
+    def __init__(self, *, agent: PamDeployAgent, thread_id: str | None = None) -> None:
+        """初始化图实例和会话线程 ID。"""
+        self.agent = agent
+        self.thread_id = thread_id or str(uuid4())
+        self._waiting_confirmation = False
+        self._graph = self._build_graph()
+
+    @property
+    def waiting_confirmation(self) -> bool:
+        """返回当前 LangGraph 会话是否停在 interrupt 确认点。"""
+        return self._waiting_confirmation
+
+    def start(self, state: AgentState) -> LangGraphRunResult:
+        """从给定 AgentState 开始执行，直到结束或遇到人工确认点。"""
+        self._waiting_confirmation = False
+        return self._consume(self._graph.stream({"agent_state": state}, self._config()))
+
+    def resume(self, *, approved: bool, note: str = "") -> LangGraphRunResult:
+        """把人工确认结果交回 LangGraph，并继续执行。"""
+        try:
+            from langgraph.types import Command
+        except ImportError as exc:  # pragma: no cover - 依赖缺失时由调用方降级
+            raise RuntimeError("未安装 langgraph，无法恢复 interrupt。") from exc
+
+        decision = {"approved": approved, "note": note}
+        return self._consume(self._graph.stream(Command(resume=decision), self._config()))
+
+    def _build_graph(self):
+        """构建 deploy -> confirm interrupt -> deploy 的循环图。"""
+        try:
+            from langgraph.checkpoint.memory import InMemorySaver
+            from langgraph.graph import END, START, StateGraph
+            from langgraph.types import interrupt
+        except ImportError as exc:  # pragma: no cover - 依赖缺失时由调用方降级
+            raise RuntimeError("未安装 langgraph，无法启用 chat interrupt。") from exc
+
+        def deploy_node(state: dict[str, Any]) -> dict[str, Any]:
+            """执行部署流，遇到 pending_confirmation 时由路由转入确认节点。"""
+            agent_state = self.agent.run_deploy_flow(state["agent_state"])
+            return {"agent_state": agent_state}
+
+        def confirm_node(state: dict[str, Any]) -> dict[str, Any]:
+            """把确认请求交给 LangGraph interrupt，并在恢复后执行确认动作。"""
+            agent_state = state["agent_state"]
+            request = self.agent.build_confirmation_request(agent_state)
+            decision = interrupt(request)
+            approved, note = _parse_confirmation_decision(decision)
+            agent_state = self.agent.confirm_pending(
+                agent_state,
+                approved=approved,
+                operator_note=note,
+            )
+            return {"agent_state": agent_state}
+
+        def report_node(state: dict[str, Any]) -> dict[str, Any]:
+            """渲染当前状态报告。"""
+            return {"report": self.agent.render_report(state["agent_state"])}
+
+        def route_after_deploy(state: dict[str, Any]) -> str:
+            """根据是否存在 pending_confirmation 决定下一步。"""
+            agent_state = state["agent_state"]
+            return "confirm" if agent_state.pending_confirmation else "report"
+
+        graph = StateGraph(dict)
+        graph.add_node("deploy", deploy_node)
+        graph.add_node("confirm", confirm_node)
+        graph.add_node("report", report_node)
+        graph.add_edge(START, "deploy")
+        graph.add_conditional_edges(
+            "deploy",
+            route_after_deploy,
+            {"confirm": "confirm", "report": "report"},
+        )
+        graph.add_edge("confirm", "deploy")
+        graph.add_edge("report", END)
+        return graph.compile(checkpointer=InMemorySaver())
+
+    def _config(self) -> dict[str, Any]:
+        """生成 LangGraph checkpointer 使用的线程配置。"""
+        return {"configurable": {"thread_id": self.thread_id}}
+
+    def _consume(self, chunks: Any) -> LangGraphRunResult:
+        """消费 LangGraph stream 输出，提取状态、报告和 interrupt 请求。"""
+        result = LangGraphRunResult()
+        for chunk in chunks:
+            result.chunks.append(chunk)
+            if "__interrupt__" in chunk:
+                result.interrupted = True
+                result.confirmation = _extract_interrupt_value(chunk["__interrupt__"])
+                continue
+
+            for value in chunk.values():
+                if not isinstance(value, dict):
+                    continue
+                if isinstance(value.get("agent_state"), AgentState):
+                    result.state = value["agent_state"]
+                if isinstance(value.get("report"), str):
+                    result.report = value["report"]
+
+        self._waiting_confirmation = result.interrupted
+        return result
+
+
+def _extract_interrupt_value(interrupts: Any) -> dict[str, Any]:
+    """从 LangGraph interrupt 对象中提取确认请求字典。"""
+    if not interrupts:
+        return {}
+    first = interrupts[0]
+    value = getattr(first, "value", first)
+    return value if isinstance(value, dict) else {"value": value}
+
+
+def _parse_confirmation_decision(value: Any) -> tuple[bool, str]:
+    """把 interrupt resume 值解析为 approved/note。"""
+    if isinstance(value, dict):
+        return bool(value.get("approved", False)), str(value.get("note", ""))
+    if isinstance(value, bool):
+        return value, ""
+    if isinstance(value, str):
+        normalized = value.strip().lower()
+        return normalized in ("approve", "approved", "yes", "y", "true"), value
+    return False, str(value)
--- a/pam_deploy_graph/llm/base.py
+++ b/pam_deploy_graph/llm/base.py
@ -4,7 +4,14 @@ from __future__ import annotations

 from typing import Any, Protocol

-from pam_deploy_graph.models import ExecutionStrategy, LlmDeployPlan, LlmIntentResult, LlmParamResult
+from pam_deploy_graph.models import (
+    ActionResult,
+    ExecutionStrategy,
+    LlmActionAnalysis,
+    LlmDeployPlan,
+    LlmIntentResult,
+    LlmParamResult,
+)


 class LlmClient(Protocol):
@ -27,3 +34,13 @@ class LlmClient(Protocol):
    ) -> LlmDeployPlan:
        """根据参数和意图生成部署计划。"""
        ...
+
+    def analyze_action_result(
+        self,
+        *,
+        action: str,
+        result: ActionResult,
+        state_summary: dict[str, Any],
+    ) -> LlmActionAnalysis:
+        """分析 action 执行结果，并给出辅助诊断建议。"""
+        ...
--- a/pam_deploy_graph/llm/openai_compatible.py
+++ b/pam_deploy_graph/llm/openai_compatible.py
@ -20,8 +20,9 @@ from pam_deploy_graph.constants import (
    SENSITIVE_KEYS,
 )
 from pam_deploy_graph.models import ExecutionStrategy, LlmDeployPlan, LlmIntentResult, LlmParamResult
+from pam_deploy_graph.models import ActionResult, LlmActionAnalysis

-from .prompts import INTENT_PROMPT, PARAM_PROMPT, PLAN_PROMPT, SYSTEM_PROMPT
+from .prompts import ACTION_ANALYSIS_PROMPT, INTENT_PROMPT, PARAM_PROMPT, PLAN_PROMPT, SYSTEM_PROMPT

 JsonTransport = Callable[[str, dict[str, str], dict[str, Any], float], dict[str, Any]]

@ -127,6 +128,40 @@ class OpenAICompatibleLlmClient:
            execution_strategy=_string(payload, "execution_strategy", strategy),  # type: ignore[arg-type]
        )

+    def analyze_action_result(
+        self,
+        *,
+        action: str,
+        result: ActionResult,
+        state_summary: dict[str, Any],
+    ) -> LlmActionAnalysis:
+        """调用 LLM 分析 action 结果，返回结构化诊断建议。"""
+        payload = self._complete_json(
+            ACTION_ANALYSIS_PROMPT,
+            {
+                "action": action,
+                "result": {
+                    "backend": result.backend,
+                    "ok": result.ok,
+                    "exit_code": result.exit_code,
+                    "tool_name": result.tool_name,
+                    "values": _redact_sensitive(result.values),
+                    "stderr": _truncate_text(result.stderr),
+                    "error_summary": result.error_summary,
+                },
+                "state_summary": _redact_sensitive(state_summary),
+            },
+        )
+        return LlmActionAnalysis(
+            action=_string(payload, "action", action),
+            has_anomaly=bool(payload.get("has_anomaly", False)),
+            severity=_string(payload, "severity", "info"),  # type: ignore[arg-type]
+            possible_reason=_string(payload, "possible_reason", ""),
+            suggested_action=_string(payload, "suggested_action", ""),
+            requires_confirmation=bool(payload.get("requires_confirmation", False)),
+            notes=_string_list(payload.get("notes")),
+        )
+
    def _complete_json(self, instruction: str, input_payload: dict[str, Any]) -> dict[str, Any]:
        """发送 chat/completions 请求，并解析 JSON 对象响应。"""
        request_payload = {
@ -229,6 +264,13 @@ def _redact_sensitive(value: Any) -> Any:
    return value


+def _truncate_text(value: str, limit: int = 1000) -> str:
+    """截断发送给 LLM 的长文本，避免传入完整日志。"""
+    if len(value) <= limit:
+        return value
+    return value[:limit] + "...[已截断]"
+
+
 def _string(payload: dict[str, Any], key: str, default: str) -> str:
    """安全读取字符串字段。"""
    value = payload.get(key, default)
--- a/pam_deploy_graph/llm/prompts.py
+++ b/pam_deploy_graph/llm/prompts.py
@ -65,3 +65,22 @@ PLAN_PROMPT = """生成 PAM 部署计划。
 计划只能使用允许 action；不要包含可执行脚本命令、命令行参数或密钥。
 PAM_HOME action 仍由脚本 action 执行；PAM_NODE action 在 hybrid_node_mcp 策略下走 MCP。
 """
+
+ACTION_ANALYSIS_PROMPT = """分析一次 PAM action 执行结果。
+
+输出 JSON schema：
+{
+  "action": "...",
+  "has_anomaly": false,
+  "severity": "info|low|medium|high",
+  "possible_reason": "...",
+  "suggested_action": "...",
+  "requires_confirmation": false,
+  "notes": ["..."]
+}
+
+要求：
+- 只给诊断建议，不决定继续执行、回滚或修改参数。
+- 如果 exit_code 非 0、ok=false、verify-ip SUCCESS=false、出现 pending_confirmation，应标记异常。
+- 不要输出密钥、token、Authorization 或完整日志原文。
+"""
--- a/pam_deploy_graph/llm/rule_based.py
+++ b/pam_deploy_graph/llm/rule_based.py
@ -11,7 +11,9 @@ from typing import Any

 from pam_deploy_graph.constants import GLOBAL_ACTION_SEQUENCE, REQUIRED_PARAMS
 from pam_deploy_graph.models import (
+    ActionResult,
    ExecutionStrategy,
+    LlmActionAnalysis,
    LlmDeployPlan,
    LlmIntentResult,
    LlmParamResult,
@ -145,6 +147,61 @@ class RuleBasedLlmClient:
            execution_strategy=strategy,
        )

+    def analyze_action_result(
+        self,
+        *,
+        action: str,
+        result: ActionResult,
+        state_summary: dict[str, Any],
+    ) -> LlmActionAnalysis:
+        """用本地规则分析 action 结果，作为真实 LLM 不可用时的兜底。"""
+        notes: list[str] = []
+        has_anomaly = not result.ok
+        severity = "info"
+        possible_reason = ""
+        suggested_action = "继续观察。"
+        requires_confirmation = False
+
+        if not result.ok:
+            severity = "medium"
+            possible_reason = result.error_summary or "action 返回失败状态。"
+            suggested_action = "查看 action stderr/raw_output，确认参数、网络和目标服务状态。"
+            notes.append("硬规则检测到 action 执行失败。")
+
+        if action == "verify-ip":
+            success = result.values.get("SUCCESS")
+            if success is not None and str(success).lower() not in ("true", "1", "yes"):
+                has_anomaly = True
+                severity = "high"
+                possible_reason = result.values.get("MESSAGE", "") or "工作站健康检查未通过。"
+                suggested_action = "先下载日志并人工确认是否执行回滚。"
+                requires_confirmation = True
+                notes.append("verify-ip SUCCESS 非成功值。")
+
+        if action == "rollback-ip" and not result.ok:
+            severity = "high"
+            suggested_action = "保持待确认状态，人工排查回滚失败原因后重试或转人工处理。"
+            requires_confirmation = True
+            notes.append("rollback-ip 失败需要人工处理。")
+
+        if result.values.get("PENDING_AGENT_CONFIRMATION"):
+            has_anomaly = True
+            severity = "high"
+            possible_reason = str(result.values["PENDING_AGENT_CONFIRMATION"])
+            suggested_action = "暂停自动流程，等待人工确认。"
+            requires_confirmation = True
+            notes.append("action 返回待人工确认标记。")
+
+        return LlmActionAnalysis(
+            action=action,
+            has_anomaly=has_anomaly,
+            severity=severity,  # type: ignore[arg-type]
+            possible_reason=possible_reason,
+            suggested_action=suggested_action,
+            requires_confirmation=requires_confirmation,
+            notes=notes,
+        )
+
    def _extract_key_values(self, text: str) -> dict[str, str]:
        """抽取 KEY=VALUE 形式的参数。"""
        params: dict[str, str] = {}
--- a/pam_deploy_graph/mcp_client.py
+++ b/pam_deploy_graph/mcp_client.py
@ -7,6 +7,7 @@ callable 或 SDK session 适配成这个接口，避免业务代码绑定具体
 from __future__ import annotations

 import json
+from datetime import timedelta
 from collections.abc import Callable
 from dataclasses import dataclass, field
 from pathlib import Path
@ -18,6 +19,12 @@ class McpClientConfig:
    """真实 MCP session 建立后需要传给 runner 的配置。"""

    server_name: str = "pam-node"
+    transport: str = "stdio"
+    command: str = ""
+    args: list[str] = field(default_factory=list)
+    env: dict[str, str] | None = None
+    cwd: str = ""
+    timeout_seconds: float = 60
    tool_names: dict[str, str] = field(default_factory=dict)

    @classmethod
@ -26,8 +33,20 @@ class McpClientConfig:
        tool_names = payload.get("tool_names") or payload.get("tools") or {}
        if not isinstance(tool_names, dict):
            raise ValueError("MCP tool_names 必须是 JSON object")
+        args = payload.get("args") or []
+        if not isinstance(args, list):
+            raise ValueError("MCP args 必须是数组")
+        env = payload.get("env")
+        if env is not None and not isinstance(env, dict):
+            raise ValueError("MCP env 必须是 JSON object")
        return cls(
            server_name=str(payload.get("server_name", "pam-node")),
+            transport=str(payload.get("transport", "stdio")),
+            command=str(payload.get("command", "")),
+            args=[str(item) for item in args],
+            env={str(key): str(value) for key, value in env.items()} if env else None,
+            cwd=str(payload.get("cwd", "")),
+            timeout_seconds=float(payload.get("timeout_seconds", 60)),
            tool_names={str(key): str(value) for key, value in tool_names.items()},
        )

@ -74,6 +93,56 @@ class SessionMcpToolClient:
        return normalize_mcp_sdk_result(result)


+class StdioMcpToolClient:
+    """通过 MCP Python SDK 启动 stdio server 并调用 tool。"""
+
+    def __init__(
+        self,
+        *,
+        command: str,
+        args: list[str] | None = None,
+        env: dict[str, str] | None = None,
+        cwd: str | None = None,
+        timeout_seconds: float = 60,
+    ) -> None:
+        """保存 stdio server 启动参数。"""
+        if not command:
+            raise ValueError("stdio MCP 配置必须提供 command")
+        self.command = command
+        self.args = list(args or [])
+        self.env = env
+        self.cwd = cwd or None
+        self.timeout_seconds = timeout_seconds
+
+    def call_tool(self, tool_name: str, arguments: dict[str, Any]) -> Any:
+        """创建一次 MCP stdio session，调用 tool 后关闭 session。"""
+        try:
+            import anyio
+            from mcp import ClientSession
+            from mcp.client.stdio import StdioServerParameters, stdio_client
+        except ImportError as exc:  # pragma: no cover - 依赖安装状态
+            raise RuntimeError("未安装 MCP Python SDK，请安装项目的 mcp 可选依赖") from exc
+
+        async def call_once() -> Any:
+            server = StdioServerParameters(
+                command=self.command,
+                args=self.args,
+                env=self.env,
+                cwd=self.cwd,
+            )
+            async with stdio_client(server) as (read_stream, write_stream):
+                async with ClientSession(read_stream, write_stream) as session:
+                    await session.initialize()
+                    result = await session.call_tool(
+                        tool_name,
+                        arguments,
+                        read_timeout_seconds=timedelta(seconds=self.timeout_seconds),
+                    )
+                    return normalize_mcp_sdk_result(result)
+
+        return anyio.run(call_once)
+
+
 def normalize_mcp_sdk_result(result: Any) -> Any:
    """把常见 MCP SDK 返回结构归一化成 dict/list/string。"""
    if hasattr(result, "structuredContent"):
--- a/pam_deploy_graph/mcp_factory.py
+++ b/pam_deploy_graph/mcp_factory.py
@ -0,0 +1,28 @@
+"""根据配置文件构造 MCP runner。"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from .mcp_client import McpClientConfig, StdioMcpToolClient, load_mcp_client_config
+from .mcp_runner import McpActionRunner
+
+
+def build_mcp_runner_from_config(path: str | Path) -> McpActionRunner:
+    """读取 MCP 配置文件，并构造可直接给 Agent 使用的 runner。"""
+    config = load_mcp_client_config(path)
+    client = build_mcp_client(config)
+    return McpActionRunner(client=client, tool_names=config.tool_names or None)
+
+
+def build_mcp_client(config: McpClientConfig):
+    """根据 transport 类型创建 MCP client。"""
+    if config.transport == "stdio":
+        return StdioMcpToolClient(
+            command=config.command,
+            args=config.args,
+            env=config.env,
+            cwd=config.cwd or None,
+            timeout_seconds=config.timeout_seconds,
+        )
+    raise ValueError(f"不支持的 MCP transport: {config.transport}")
--- a/pam_deploy_graph/models.py
+++ b/pam_deploy_graph/models.py
@ -10,6 +10,7 @@ ExecutionStrategy = Literal["hybrid_node_mcp", "script_only", "fake"]
 IntentName = Literal["deploy", "show_usage", "preview", "query_node_ips", "rollback"]
 ModePreference = Literal["MCP", "API脚本", "未指定"]
 StrategyPreference = Literal["hybrid_node_mcp", "script_only", "fake", "未指定"]
+ActionAnalysisSeverity = Literal["info", "low", "medium", "high"]


@dataclass(slots=True)
@ -88,6 +89,19 @@ class LlmDeployPlan:
    execution_strategy: StrategyPreference = "未指定"


+@dataclass(slots=True)
+class LlmActionAnalysis:
+    """LLM 或规则对单次 action 结果的诊断建议。"""
+
+    action: str
+    has_anomaly: bool = False
+    severity: ActionAnalysisSeverity = "info"
+    possible_reason: str = ""
+    suggested_action: str = ""
+    requires_confirmation: bool = False
+    notes: list[str] = field(default_factory=list)
+
+
@dataclass(slots=True)
 class AgentState:
    """一次部署运行的完整状态，可序列化到 checkpoint。"""
--- a/pyproject.toml
+++ b/pyproject.toml
@ -9,6 +9,7 @@ dependencies = [

 [project.optional-dependencies]
 mcp = ["mcp>=1"]
+chat = ["rich>=13", "prompt_toolkit>=3"]
 test = ["pytest"]

 [tool.pytest.ini_options]
--- a/tests/test_agent_flow.py
+++ b/tests/test_agent_flow.py
@ -60,6 +60,33 @@ def test_run_deploy_flow_stops_on_verify_failure(tmp_path: Path):
    assert any(event["type"] == "CONFIRMATION_REQUIRED" for event in state.events)


+def test_action_analysis_event_is_recorded_when_enabled(tmp_path: Path):
+    fake = FakeActionRunner(
+        {
+            "verify-ip:192.168.1.10": {
+                "ACTION": "verify-ip",
+                "IP": "192.168.1.10",
+                "SUCCESS": "false",
+                "MESSAGE": "health check failed",
+            }
+        }
+    )
+    agent = PamDeployAgent(fake_runner=fake, action_analysis_enabled=True)
+    state = agent.create_state(
+        params=PARAMS,
+        execution_strategy="fake",
+        config_path=str(tmp_path / "config.txt"),
+    )
+
+    agent.run_deploy_flow(state)
+
+    analyses = [event for event in state.events if event["type"] == "ACTION_ANALYSIS"]
+    verify_analysis = [event for event in analyses if event["stage"] == "verify-ip"][0]
+    assert verify_analysis["has_anomaly"] is True
+    assert verify_analysis["severity"] == "high"
+    assert verify_analysis["requires_confirmation"] is True
+
+
 def test_confirm_pending_rollback_runs_rollback_and_resume_continues(tmp_path: Path):
    fake = FakeActionRunner(
        {
--- a/tests/test_interactive_cli.py
+++ b/tests/test_interactive_cli.py
@ -50,7 +50,7 @@ def test_chat_run_executes_fake_deploy_and_writes_checkpoint(tmp_path: Path):
        checkpoint_path=str(checkpoint),
    )

-    run_session(session, ["run", "yes", "exit"])
+    run_session(session, ["run", "yes", "yes", "yes", "exit"])

    assert checkpoint.exists()
    assert session.state is not None
@ -76,9 +76,57 @@ def test_chat_approve_then_resume_continues_after_failed_ip(tmp_path: Path):
        checkpoint_path=str(tmp_path / "checkpoint.json"),
    )

-    run_session(session, ["run", "yes", "approve", "resume", "exit"])
+    run_session(session, ["run", "yes", "yes", "yes", "approve", "resume", "exit"])

    assert session.state is not None
    assert session.state.pending_confirmation == ""
    assert session.state.ip_states["192.168.1.10"]["rollback_status"] == "ROLLBACK_DONE"
    assert session.state.ip_states["192.168.1.11"]["status"] == "SUCCESS"
+
+
+def test_chat_params_events_and_checkpoint_commands(tmp_path: Path):
+    checkpoint = tmp_path / "checkpoint.json"
+    session = InteractiveCliSession(
+        agent=PamDeployAgent(fake_runner=FakeActionRunner(), action_analysis_enabled=True),
+        params=PARAMS,
+        strategy="fake",
+        checkpoint_path=str(checkpoint),
+    )
+
+    output = run_session(
+        session,
+        [
+            "params",
+            "llm action-analysis on",
+            "run",
+            "yes",
+            "yes",
+            "yes",
+            "events 2",
+            "list checkpoints",
+            "load checkpoint " + str(checkpoint),
+            "exit",
+        ],
+    )
+
+    assert session.state is not None
+    assert any("CLIENT_SECRET: ***" in item for item in output)
+    assert any("ACTION_ANALYSIS" in item for item in output)
+    assert any("checkpoint 列表" in item for item in output)
+
+
+def test_chat_can_hot_load_mcp_config(tmp_path: Path):
+    mcp_config = tmp_path / "mcp.json"
+    mcp_config.write_text('{"transport": "stdio", "command": "python"}', encoding="utf-8")
+    session = InteractiveCliSession(
+        agent=PamDeployAgent(),
+        params=PARAMS,
+        strategy="hybrid_node_mcp",
+        checkpoint_path=str(tmp_path / "checkpoint.json"),
+    )
+
+    output = run_session(session, ["mcp config " + mcp_config.as_posix(), "exit"])
+
+    assert session.agent.mcp_runner is not None
+    assert session.agent.router.mcp_runner is not None
+    assert any("MCP 配置已加载" in item for item in output)
--- a/tests/test_langgraph_runtime.py
+++ b/tests/test_langgraph_runtime.py
@ -0,0 +1,54 @@
+from pathlib import Path
+
+from pam_deploy_graph.agent import PamDeployAgent
+from pam_deploy_graph.fake_runner import FakeActionRunner
+from pam_deploy_graph.langgraph_runtime import LangGraphDeploymentRuntime
+
+
+PARAMS = {
+    "HOME_BASE_URL": "https://pam.home.example.com",
+    "CLIENT_ID": "client",
+    "CLIENT_SECRET": "secret",
+    "AIRPORT_CODE": "HET",
+    "APP_NAME": "PAM",
+    "MODULE_NAME": "Node",
+    "VERSION_NUMBER": "2.0.5",
+    "ZIP_FILE_PATH": "C:/pkg.zip",
+}
+
+
+def test_langgraph_runtime_interrupts_and_resumes_confirmation(tmp_path: Path):
+    fake = FakeActionRunner(
+        {
+            "verify-ip:192.168.1.10": {
+                "ACTION": "verify-ip",
+                "IP": "192.168.1.10",
+                "SUCCESS": "false",
+                "MESSAGE": "health check failed",
+            }
+        }
+    )
+    agent = PamDeployAgent(fake_runner=fake)
+    state = agent.create_state(
+        params=PARAMS,
+        execution_strategy="fake",
+        config_path=str(tmp_path / "config.txt"),
+        checkpoint_path=str(tmp_path / "checkpoint.json"),
+    )
+    runtime = LangGraphDeploymentRuntime(agent=agent)
+
+    first = runtime.start(state)
+
+    assert first.interrupted is True
+    assert runtime.waiting_confirmation is True
+    assert first.confirmation["type"] == "rollback-ip"
+    assert first.confirmation["ip"] == "192.168.1.10"
+
+    second = runtime.resume(approved=True)
+
+    assert second.interrupted is False
+    assert runtime.waiting_confirmation is False
+    assert second.state is not None
+    assert second.state.pending_confirmation == ""
+    assert second.state.ip_states["192.168.1.10"]["rollback_status"] == "ROLLBACK_DONE"
+    assert second.state.ip_states["192.168.1.11"]["status"] == "SUCCESS"
--- a/tests/test_llm_structured.py
+++ b/tests/test_llm_structured.py
@ -6,6 +6,7 @@ from pam_deploy_graph.llm.openai_compatible import OpenAICompatibleLlmClient
 from pam_deploy_graph.llm.rule_based import RuleBasedLlmClient
 from pam_deploy_graph.llm.validators import validate_deploy_plan
 from pam_deploy_graph.models import LlmDeployPlan
+from pam_deploy_graph.models import ActionResult


 def test_understand_request_prefers_hybrid_for_mcp():
@ -141,3 +142,50 @@ def test_openai_compatible_client_does_not_send_base_secret():
    serialized_prompt = str(calls[0])
    assert "real-secret" not in serialized_prompt
    assert result.extracted_params["CLIENT_SECRET"] == "real-secret"
+
+
+def test_openai_compatible_client_analyzes_action_result_with_redaction():
+    calls = []
+
+    def transport(url, headers, payload, timeout_sec):
+        calls.append(payload)
+        return {
+            "choices": [
+                {
+                    "message": {
+                        "content": (
+                            '{"action":"verify-ip","has_anomaly":true,"severity":"high",'
+                            '"possible_reason":"health check failed",'
+                            '"suggested_action":"download logs","requires_confirmation":true,'
+                            '"notes":["verify failed"]}'
+                        )
+                    }
+                }
+            ]
+        }
+
+    client = OpenAICompatibleLlmClient(
+        base_url="https://llm.example/v1",
+        api_key="secret-key",
+        model="model-a",
+        transport=transport,
+    )
+
+    analysis = client.analyze_action_result(
+        action="verify-ip",
+        result=ActionResult(
+            action="verify-ip",
+            backend="fake",
+            ok=False,
+            values={"CLIENT_SECRET": "real-secret", "SUCCESS": "false"},
+            stderr="x" * 1200,
+            error_summary="failed",
+        ),
+        state_summary={"params": {"CLIENT_SECRET": "real-secret"}},
+    )
+
+    serialized_prompt = str(calls[0])
+    assert analysis.has_anomaly is True
+    assert analysis.severity == "high"
+    assert "real-secret" not in serialized_prompt
+    assert "[已截断]" in serialized_prompt
--- a/tests/test_mcp_client.py
+++ b/tests/test_mcp_client.py
@ -2,8 +2,10 @@ from pam_deploy_graph.mcp_client import (
    FunctionMcpToolClient,
    load_mcp_client_config,
    SessionMcpToolClient,
+    StdioMcpToolClient,
    normalize_mcp_sdk_result,
 )
+from pam_deploy_graph.mcp_factory import build_mcp_runner_from_config


 def test_function_mcp_client_wraps_callable():
@ -31,11 +33,35 @@ def test_session_mcp_client_normalizes_text_json_content():
 def test_load_mcp_client_config(tmp_path):
    path = tmp_path / "mcp.json"
    path.write_text(
-        '{"server_name": "pam-node-prod", "tool_names": {"get-online-ips": "custom_ips"}}',
+        (
+            '{"server_name": "pam-node-prod", "transport": "stdio", '
+            '"command": "python", "args": ["-m", "server"], '
+            '"env": {"PAM_ENV": "test"}, "cwd": "/tmp", "timeout_seconds": 3, '
+            '"tool_names": {"get-online-ips": "custom_ips"}}'
+        ),
        encoding="utf-8",
    )

    config = load_mcp_client_config(path)

    assert config.server_name == "pam-node-prod"
+    assert config.transport == "stdio"
+    assert config.command == "python"
+    assert config.args == ["-m", "server"]
+    assert config.env == {"PAM_ENV": "test"}
+    assert config.cwd == "/tmp"
+    assert config.timeout_seconds == 3
    assert config.tool_names["get-online-ips"] == "custom_ips"
+
+
+def test_build_mcp_runner_from_stdio_config(tmp_path):
+    path = tmp_path / "mcp.json"
+    path.write_text(
+        '{"transport": "stdio", "command": "python", "tool_names": {"verify-ip": "custom_verify"}}',
+        encoding="utf-8",
+    )
+
+    runner = build_mcp_runner_from_config(path)
+
+    assert isinstance(runner.client, StdioMcpToolClient)
+    assert runner.tool_names["verify-ip"] == "custom_verify"