dark/agent_deply
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

完善 chat/runtime 的 LLM 审核、断点续跑与热更新，并同步打包文档

Browse Source 
调整 workflow 执行逻辑：每个 action 完成后统一进入 LLM/规则审核，审核开始/结果可播报，审核阻断时自动暂停并给出建议
增强 chat 交互：支持执行中 Ctrl+C 中断并保存 checkpoint，后续可 resume 继续
增加运行时热更新能力：支持 set KEY=VALUE 和 load params <路径> 同步更新当前 state、config.txt 和 checkpoint
支持自定义 action 审核提示词：新增 --llm-action-analysis-prompt-file / PAM_LLM_ACTION_ANALYSIS_PROMPT_FILE
新增 prompts/action_review.txt，落地保存当前默认审核提示词，便于后续按基线调整
更新 Linux 打包脚本，将 prompts/action_review.txt 一并带入发布包
同步更新 README、流程图、todo 和打包文档，修正 --analyze-actions 语义说明与 chat 最新行为说明
This commit is contained in:
dark 2026-06-03 17:02:17 +08:00
parent 5914e96693
commit 8d390aa416
19 changed files with 876 additions and 53 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

29
README.md
View File

@ -82,8 +82,13 @@ packaging/
- chat 在开发环境可选启用 `rich` / `prompt_toolkit`；PyInstaller 打包环境默认使用普通文本输入，避免交互兼容问题。
- chat 执行前会归一化参数并展示实际写入脚本配置的值；`script_only` / `hybrid_node_mcp` 会提前检查 `ZIP_FILE_PATH` 是否存在。
- chat 执行中会播报每个 action 的开始、完成或失败；action 执行失败会停在当前 checkpoint，不再误报 LangGraph 不可用。
- 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 `51 passed, 2 skipped`。
- 每个 action 完成后都会进入一次 LLM/规则审核；如果审核建议停止，流程会暂停并给出建议，等待用户 `resume`。
- `--analyze-actions` 和 `llm action-analysis on` 改为只控制是否把详细审核结果写入 `events`，不再控制审核是否执行。
- chat 会播报 action 审核开始、审核完成和审核失败，避免黑盒执行。
- chat 支持执行中按 `Ctrl+C` 中断，保存 checkpoint 后再 `resume`。
- chat 支持 `set KEY=VALUE` 和 `load params <路径>` 热更新当前运行参数，并同步回写运行中的 `config.txt` 与 checkpoint。
- 支持通过 `--llm-action-analysis-prompt-file`、`PAM_LLM_ACTION_ANALYSIS_PROMPT_FILE` 或 chat 内 `llm config action_analysis_prompt_file=...` 自定义 action 审核提示词。
- 添加基础测试，当前本地结果为 `57 passed, 2 skipped`。
未完成：
@ -113,6 +118,19 @@ python -m pam_deploy_graph.cli analyze \
--llm-model your-model-name
```
如需自定义 action 审核提示词，可再补充：
```bash
python -m pam_deploy_graph.cli analyze \
--config doc_scripts/config.txt.example \
--text "请分析这次部署" \
--llm-base-url https://your-llm.example.com/v1 \
--llm-model your-model-name \
--llm-action-analysis-prompt-file prompts/action_review.txt
```
仓库内已提供 [prompts/action_review.txt](/e:/AIcoding/agent_deply/prompts/action_review.txt) 作为“当前默认 action 审核提示词”的落地副本，后续自定义时可以先复制它再改，便于和内置默认行为对照。
真实 LLM 调用位置在 `pam_deploy_graph/llm/openai_compatible.py`，提示词在 `pam_deploy_graph/llm/prompts.py`。发送给 LLM 的 `base_params` 会脱敏，`CLIENT_SECRET` 不会进入 prompt；本地生成计划后仍会执行 guardrails 校验。
如果服务需要鉴权，再补充：
@ -253,14 +271,17 @@ python -m pam_deploy_graph.cli chat --config doc_scripts/config.txt.example --st
PAM> 请用 MCP 预演部署 HET PAM Node 版本 2.0.5，不要动环境
PAM> preview
PAM> set VERSION_NUMBER=2.0.6
PAM> load params runtime/override.txt
PAM> run
即将执行真实 action；确认执行请输入 yes: yes
开始执行 action: get-token [backend=fake]
开始分析 action 结果: get-token [backend=fake]
完成 action: get-token [backend=fake]
PAM> status
PAM> params
PAM> events 5
PAM> llm action-analysis on
PAM> llm config action_analysis_prompt_file=prompts/action_review.txt
PAM> mcp config mcp_client.example.json
PAM> list checkpoints
PAM> load checkpoint runtime/checkpoints/chat-demo.json
@ -269,7 +290,7 @@ PAM> resume
PAM> exit
```
`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。输入 `你好`、`hello` 这类问候不会触发 LLM/结构化分析；需要分析部署需求时可直接描述部署任务，或显式使用 `analyze <需求>`。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`。
`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。输入 `你好`、`hello` 这类问候不会触发 LLM/结构化分析；需要分析部署需求时可直接描述部署任务，或显式使用 `analyze <需求>`。每个 action 完成后都会自动进入一次 LLM/规则审核，并播报审核开始/结束；如果审核建议停止或审核本身失败，流程会暂停并输出建议，等待用户决定是否 `resume`。`--analyze-actions` 仅控制详细审核结果是否写入 `events`。执行中可按 `Ctrl+C` 中断，chat 会保存当前 checkpoint 并把流程标记为 `user_interrupted`。`set KEY=VALUE` 和 `load params <路径>` 会把更新同步到当前运行 state、`config.txt` 和 checkpoint。`chat` 也支持 `--llm-base-url` / `--llm-api-key` / `--llm-model` / `--llm-action-analysis-prompt-file`、`--mcp-config` 和 `--analyze-actions`。
预演：
@ -295,7 +316,7 @@ python -m pam_deploy_graph.cli run-deploy --config doc_scripts/config.txt.exampl
python -m pam_deploy_graph.cli confirm --checkpoint runtime/checkpoints/demo.json --decision approve --confirm
```
`confirm` 会通过 LangGraph interrupt resume 处理确认，并在确认后继续执行后续图节点；如果进程中断或需要再次续跑，再执行 `resume` 即可。
`confirm` 会通过 LangGraph interrupt resume 处理确认，并在确认后继续执行后续图节点；如果流程此前处于 `paused` 状态，`resume` 会先清理暂停标记，再从 checkpoint 继续执行。
拒绝回滚：

54
docs/current_logic_flow.md
View File

@ -1,6 +1,6 @@
# 当前整体逻辑结构流程图
本文描述当前 PAM 部署 Agent 的主要模块、运行路径、人工确认点和断点续跑逻辑。
本文描述当前 PAM 部署 Agent 的主要模块、运行路径、LLM 审核、人工确认点、热更新和断点续跑逻辑。
## 模块结构
@ -105,28 +105,41 @@ flowchart LR
C -- PAM_NODE action --> NM[MCP tool 执行]
```
## action 后诊断
## action 后审核
```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]
A[action 执行完成] --> 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]
E -- 是 --> F[OpenAICompatibleLlmClient 输出结构化审核]
E -- 否 --> G[RuleBasedLlmClient 本地规则审核]
F --> H{should_continue}
G --> H
H --> I[诊断只作建议，不自动继续/回滚/改参数]
H -- true --> I[继续后续 action]
H -- false --> J[暂停流程并写入 review_context]
J --> K[chat/CLI 播报审核建议并等待 resume]
F --> L{是否开启 analyze-actions}
G --> L
L -- 是 --> M[追加 ACTION_ANALYSIS 事件]
L -- 否 --> N[不写详细事件，仅播报审核过程]
```
说明：
- 每个 action 完成后都会进入一次审核，不再依赖 `--analyze-actions` 开关。
- `--analyze-actions` 或 `llm action-analysis on` 只控制是否把详细审核结果写入 `events`。
- 如果审核本身失败，也会生成“停止继续”的审核结果并暂停流程，避免黑盒继续执行。
## 失败、人工确认和续跑
```mermaid
flowchart TD
A[逐 IP action 执行] --> B{action 失败或业务校验失败}
B -- 否 --> C[记录 completed_steps 并保存 checkpoint]
C --> C1{LLM 审核是否允许继续}
C1 -- 是 --> C2[继续后续 action]
C1 -- 否 --> G[保存 checkpoint 并暂停]
B -- 是 --> D[记录 ip_state 为 FAILED]
D --> E[download-log 尽力下载日志]
E --> F[设置 pending_confirmation=rollback-ip:IP]
@ -148,18 +161,39 @@ flowchart TD
N --> O[跳过已完成全局步骤、成功 IP 和单 IP 已完成 action]
```
## 用户中断与热更新
```mermaid
flowchart TD
A[chat 执行中] --> B{用户是否按 Ctrl+C}
B -- 是 --> C[pause_state 标记 paused=user_interrupted]
C --> D[保存 checkpoint]
D --> E[chat 播报可 resume]
B -- 否 --> F[继续执行]
G[用户输入 set KEY=VALUE] --> H[normalize_params]
I[用户输入 load params <路径>] --> J[读取参数文件]
J --> H
H --> K[update_state_params]
K --> L[回写 state.params]
L --> M[回写运行中的 config.txt]
M --> N[保存 checkpoint]
```
## checkpoint 续跑语义
- `completed_global_steps`：全局阶段已经完成的 action 会跳过。
- `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`。
- `paused` / `pause_reason`：流程可能因 LLM 审核阻断、用户中断、回滚失败等原因暂停；`resume` 会先清理暂停标记，再继续执行。
- `review_context`：保存最近一次暂停时的审核建议、失败原因、IP 和阶段，供 chat/CLI 输出给用户。
- CLI/chat 的运行调度由 `langgraph_runtime.py` 通过 action 级 LangGraph 节点执行；chat 和 CLI confirm 的确认点使用 LangGraph interrupt 和 InMemorySaver。
- 跨进程续跑仍读取业务 checkpoint JSON；LangGraph checkpointer 负责单进程图恢复和 interrupt resume。
- 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 LLM：`llm.openai_compatible.OpenAICompatibleLlmClient`，通过 `PAM_LLM_BASE_URL`、`PAM_LLM_API_KEY`、`PAM_LLM_MODEL`、`PAM_LLM_ACTION_ANALYSIS_PROMPT_FILE` 或 CLI 参数配置。
- 真实 MCP：CLI/chat 可通过 `--mcp-config` 加载 streamable_http、sse 或 stdio MCP 配置，HTTP/SSE 支持独立 token 鉴权，并通过 `list_tools` 自动发现 server tools。
- 真实脚本：PAM_HOME action 通过 `doc_scripts/deploy.sh` 或 `deploy.ps1` 调用。

7
docs/todo.md
View File

@ -7,15 +7,18 @@
- [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 `load params <路径>`，允许从参数文件热更新当前会话和当前运行任务。
- [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] 增加执行中 `Ctrl+C` 中断处理：保存 checkpoint、标记 `user_interrupted`，再由 `resume` 继续。
- [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] LLM 分析结果会影响流程是否继续：`should_continue=false` 时自动暂停，并把建议输出给用户。
- [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。
- [x] 每个 action 都会执行审核；`--analyze-actions` 或 `llm action-analysis on` 只控制是否把详细审核结果写入 `events`。
- [x] 支持通过 `--llm-action-analysis-prompt-file`、环境变量或 chat 命令热加载自定义 action 审核提示词。

11
packaging/README_linux_package.md
View File

@ -42,6 +42,8 @@ pam-deploy-agent-linux-x86_64/
deploy.sh
config.txt.example
PAM_AUTO_DEPLY_SKILL.md
prompts/
action_review.txt
mcp_client.example.json
README.md
LICENSE
@ -50,6 +52,7 @@ pam-deploy-agent-linux-x86_64/
说明：
- `doc_scripts` 不会打入项目设计文档、测试脚本、Windows bat/PowerShell 脚本。
- `prompts/action_review.txt` 会随发布包一起带上，作为当前默认 action 审核提示词的参照版本。
- 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 server URL + 独立鉴权配置示例，需要按真实 MCP server 和 token 地址修改。
- 项目开发用 README 不会复制到发布包内。
@ -65,6 +68,14 @@ cd pam-deploy-agent-linux-x86_64
`run.sh --help` 是发布包专用的中文帮助，会解释命令、参数、环境变量和常见示例。`run.sh` 会切换到发布目录再启动可执行程序，因此默认的 `doc_scripts/...` 相对路径可以正常工作。
本次发布包对应的运行时行为也已同步到包内 `README.md`：
- 每个 action 完成后都会自动执行一次 LLM/规则审核。
- `--analyze-actions` 只控制是否把详细审核结果写入 `events`。
- chat 支持执行中 `Ctrl+C` 中断后保存 checkpoint，再通过 `resume` 继续。
- chat 支持 `set KEY=VALUE` 和 `load params <路径>` 热更新当前运行任务参数。
- 支持通过 `--llm-action-analysis-prompt-file` 或 chat 内 `llm config action_analysis_prompt_file=...` 自定义 action 审核提示词。
## 包大小评估
最终大小以脚本末尾打印的 `du` 结果为准。按当前依赖结构预估：

30
packaging/README_packaged_agent.md
View File

@ -12,12 +12,14 @@ 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
prompts/
action_review.txt # 当前默认 action 审核提示词基线
mcp_client.example.json
README.md # 当前说明
LICENSE
```
`doc_scripts` 只保留运行必需文件，不包含项目设计文档、测试脚本或 Windows 脚本。
`doc_scripts` 只保留运行必需文件，不包含项目设计文档、测试脚本或 Windows 脚本。`prompts/action_review.txt` 是当前默认 action 审核提示词的落地副本，便于复制后按需修改。
## 查看帮助
@ -34,7 +36,7 @@ pam-deploy-agent-linux-x86_64/
发布包默认使用普通文本输入，避免 PyInstaller 环境下 `prompt_toolkit` 兼容性问题；输出仍会在可用时使用 `rich` 做更清晰的文本展示。
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。
chat 会在执行前归一化并展示实际写入脚本配置的参数；`script_only` / `hybrid_node_mcp` 会先检查 `ZIP_FILE_PATH` 是否存在，避免脚本运行后才用默认路径失败。执行过程中每个 action 都会输出开始、完成或失败状态。
chat 会在执行前归一化并展示实际写入脚本配置的参数；`script_only` / `hybrid_node_mcp` 会先检查 `ZIP_FILE_PATH` 是否存在，避免脚本运行后才用默认路径失败。执行过程中每个 action 都会输出开始、完成或失败状态；每个 action 完成后还会自动进入一次 LLM/规则审核，并播报审核开始和审核结果。
## 交互式使用
@ -60,14 +62,17 @@ chat 会在执行前归一化并展示实际写入脚本配置的参数；`scrip
PAM> 请用 MCP 预演部署 HET PAM Node 版本 2.0.5，不要动环境
PAM> preview
PAM> set VERSION_NUMBER=2.0.6
PAM> load params runtime/override.txt
PAM> run
即将执行真实 action；确认执行请输入 yes: yes
开始执行 action: get-token [backend=fake]
开始分析 action 结果: get-token [backend=fake]
完成 action: get-token [backend=fake]
PAM> status
PAM> params
PAM> events 5
PAM> llm action-analysis on
PAM> llm config action_analysis_prompt_file=prompts/action_review.txt
PAM> mcp config mcp_client.example.json
PAM> list checkpoints
PAM> load checkpoint runtime/checkpoints/demo.json
@ -96,7 +101,7 @@ 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：
执行时把详细 action 审核结果写入 `events`：
```bash
./run.sh run-deploy \
@ -138,12 +143,13 @@ PAM> exit
```bash
export PAM_LLM_BASE_URL="https://your-llm.example.com/v1"
export PAM_LLM_API_KEY="your-api-key"
export PAM_LLM_MODEL="your-model-name"
./run.sh analyze --config doc_scripts/config.txt.example --text "请分析这次部署"
```
如果服务需要鉴权，再补 `PAM_LLM_API_KEY`；如果不需要鉴权，可以不配置，程序不会发送 `Authorization` 请求头。
也可以用 CLI 参数：
```bash
@ -151,14 +157,25 @@ export PAM_LLM_MODEL="your-model-name"
--config doc_scripts/config.txt.example \
--text "请分析这次部署" \
--llm-base-url https://your-llm.example.com/v1 \
--llm-api-key your-api-key \
--llm-model your-model-name
```
如需自定义 action 审核提示词：
```bash
./run.sh analyze \
--config doc_scripts/config.txt.example \
--text "请分析这次部署" \
--llm-base-url https://your-llm.example.com/v1 \
--llm-model your-model-name \
--llm-action-analysis-prompt-file prompts/action_review.txt
```
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 config action_analysis_prompt_file=prompts/action_review.txt
PAM> llm action-analysis on
PAM> llm fallback
```
@ -203,5 +220,8 @@ MCP token 获取方式与 HOME 一致，默认按 `client_credentials` POST 到
- 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`。
- `chat` 中输入 `你好`、`hello` 这类问候不会触发 LLM/结构化分析；需要分析部署需求时请直接描述部署任务，或显式使用 `analyze <需求>`。
- 每个 action 完成后都会自动执行一次 LLM/规则审核；`--analyze-actions` 和 `llm action-analysis on` 只控制是否把详细审核结果写入 `events`。
- 如果审核建议停止、审核本身失败，或用户在执行中按下 `Ctrl+C`，流程都会保存 checkpoint 并进入暂停状态；后续可使用 `resume` 继续。
- `set KEY=VALUE` 和 `load params <路径>` 会热更新当前运行任务的参数，并回写运行中的 `config.txt` 和 checkpoint。
- `checkpoint` 会保存完整运行参数，请放在受控目录。
- `hybrid_node_mcp`、`resume`、`confirm` 如果需要执行 MCP action，请同时传入 `--mcp-config`。

16
packaging/build_linux_self_contained.sh
View File

@ -69,6 +69,9 @@ cp -a doc_scripts/config.txt.example "$RELEASE_DIR/doc_scripts/config.txt.exampl
cp -a doc_scripts/PAM_AUTO_DEPLY_SKILL.md "$RELEASE_DIR/doc_scripts/PAM_AUTO_DEPLY_SKILL.md"
chmod +x "$RELEASE_DIR/doc_scripts/deploy.sh"
mkdir -p "$RELEASE_DIR/prompts"
cp -a prompts/action_review.txt "$RELEASE_DIR/prompts/action_review.txt"
cp -a packaging/README_packaged_agent.md "$RELEASE_DIR/README.md"
cp -a packaging/mcp_client.example.json "$RELEASE_DIR/mcp_client.example.json"
cp -a LICENSE "$RELEASE_DIR/LICENSE"
@ -162,12 +165,13 @@ 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. mcp_client.example.json 是 MCP server URL + 独立鉴权配置示例，需要按真实 MCP server 修改。
4. confirm 会通过 LangGraph interrupt resume 处理确认，并继续后续图节点；进程中断时再使用 resume。
5. chat 会在执行前归一化并展示实际写入脚本配置的参数；script_only / hybrid_node_mcp 会先检查 ZIP_FILE_PATH 是否存在。
6. chat 执行过程中会播报每个 action 的开始、完成或失败；普通问候不会触发 LLM/结构化分析。
7. chat 内可使用 params、events、list checkpoints、load checkpoint、llm config、mcp config 等命令。
8. checkpoint 会保存完整运行参数，请放在受控目录。
3. prompts/action_review.txt 是当前默认 action 审核提示词基线，可复制后自行修改。
4. mcp_client.example.json 是 MCP server URL + 独立鉴权配置示例，需要按真实 MCP server 修改。
5. confirm 会通过 LangGraph interrupt resume 处理确认，并继续后续图节点；进程中断时再使用 resume。
6. chat 会在执行前归一化并展示实际写入脚本配置的参数；script_only / hybrid_node_mcp 会先检查 ZIP_FILE_PATH 是否存在。
7. chat 执行过程中会播报每个 action 的开始、完成或失败；普通问候不会触发 LLM/结构化分析。
8. chat 内可使用 params、events、list checkpoints、load checkpoint、load params、llm config、mcp config 等命令。
9. checkpoint 会保存完整运行参数，请放在受控目录。
HELP_TEXT
}

173
pam_deploy_graph/agent.py
View File

@ -19,7 +19,7 @@ from .constants import DEFAULT_PARAMS, GLOBAL_ACTION_SEQUENCE, IP_ACTION_SEQUENC
from .fake_runner import FakeActionRunner
from .llm import LlmClient, RuleBasedLlmClient, validate_deploy_plan, validate_intent_result
from .mcp_runner import McpActionRunner
from .models import ActionResult, AgentState, ExecutionStrategy, LlmDeployPlan, LlmIntentResult, LlmParamResult
from .models import ActionResult, AgentState, ExecutionStrategy, LlmActionAnalysis, LlmDeployPlan, LlmIntentResult, LlmParamResult
from .script_runner import ScriptActionRunner, select_script_entry
from .skill_policy import load_skill_policy
@ -144,6 +144,38 @@ class PamDeployAgent:
target_ips=target_ips or [],
)
def pause_state(
self,
state: AgentState,
*,
reason: str,
review_context: dict[str, Any] | None = None,
) -> AgentState:
"""将当前 state 标记为暂停，并持久化 checkpoint。"""
state.paused = True
state.pause_reason = reason
state.review_context = dict(review_context or {})
self._save_checkpoint(state)
return state
def resume_state(self, state: AgentState) -> AgentState:
"""清理暂停标记，允许后续继续执行。"""
state.paused = False
state.pause_reason = ""
state.review_context = {}
self._save_checkpoint(state)
return state
def update_state_params(self, state: AgentState, updates: dict[str, Any]) -> AgentState:
"""热更新 state 中的参数，并回写 config 文件。"""
merged = {**state.params, **updates}
normalized = self.normalize_params(merged)
state.params = normalized
if state.config_path:
write_config(normalized, state.config_path)
self._save_checkpoint(state)
return state
def preview(self, params: dict[str, Any], strategy: ExecutionStrategy = "hybrid_node_mcp") -> str:
"""渲染部署预览，展示参数和 action 路由。"""
normalized = self.normalize_params(params)
@ -177,6 +209,9 @@ class PamDeployAgent:
def run_global_flow(self, state: AgentState) -> AgentState:
"""执行全局部署阶段，并跳过 checkpoint 中已完成的步骤。"""
if state.paused:
self._save_checkpoint(state)
return state
while True:
action = self.next_global_action(state)
if action is None:
@ -185,6 +220,8 @@ class PamDeployAgent:
def next_global_action(self, state: AgentState) -> str | None:
"""返回下一个未完成的全局 action。"""
if state.paused:
return None
for action in GLOBAL_ACTION_SEQUENCE:
if action in state.completed_global_steps:
continue
@ -221,7 +258,7 @@ class PamDeployAgent:
"message": result.error_summary or "ok",
}
)
self._append_action_analysis(state, action, result)
analysis = self._append_action_analysis(state, action, result)
if not result.ok:
self._emit_progress(
{
@ -232,6 +269,11 @@ class PamDeployAgent:
}
)
state.last_failed_step = action
self.pause_state(
state,
reason="action_failed",
review_context=self._review_context(action=action, analysis=analysis, result=result),
)
self._save_checkpoint(state)
raise RuntimeError(f"{action} 执行失败: {result.error_summary}")
missing_values = self._missing_required_values(action, result.values)
@ -246,6 +288,16 @@ class PamDeployAgent:
}
)
state.last_failed_step = action
self.pause_state(
state,
reason="action_missing_required_values",
review_context={
"type": "action_review",
"stage": action,
"message": message,
"missing_values": missing_values,
},
)
self._save_checkpoint(state)
raise RuntimeError(message)
self._apply_result(state, action, result.values)
@ -259,6 +311,14 @@ class PamDeployAgent:
"message": result.values.get("MESSAGE", "ok"),
}
)
if analysis is not None and not analysis.should_continue:
state.last_failed_step = action
self.pause_state(
state,
reason="llm_review_blocked",
review_context=self._review_context(action=action, analysis=analysis, result=result),
)
return state
self._save_checkpoint(state)
return state
@ -269,7 +329,7 @@ class PamDeployAgent:
def run_deploy_flow(self, state: AgentState) -> AgentState:
"""执行完整部署流程：全局阶段后进入逐 IP 阶段。"""
if state.pending_confirmation:
if state.pending_confirmation or state.paused:
self._save_checkpoint(state)
return state
self.run_global_flow(state)
@ -278,6 +338,9 @@ class PamDeployAgent:
def run_ip_flow(self, state: AgentState) -> AgentState:
"""执行逐 IP 部署流程，失败时停在人工确认点。"""
if state.paused:
self._save_checkpoint(state)
return state
while True:
work = self.next_ip_action(state)
if work is None:
@ -287,7 +350,7 @@ class PamDeployAgent:
def next_ip_action(self, state: AgentState) -> tuple[str, str] | None:
"""返回下一个待执行的单 IP action，并按需初始化 IP 状态。"""
if state.pending_confirmation:
if state.pending_confirmation or state.paused:
self._save_checkpoint(state)
return None
self._resolve_target_ips(state)
@ -358,7 +421,7 @@ class PamDeployAgent:
"message": result.error_summary or result.values.get("MESSAGE", "ok"),
}
)
self._append_action_analysis(state, action, result, ip=ip)
analysis = self._append_action_analysis(state, action, result, ip=ip)
if failed:
self._emit_progress(
@ -370,6 +433,11 @@ class PamDeployAgent:
"message": result.error_summary or result.values.get("MESSAGE", "action 执行失败"),
}
)
self.pause_state(
state,
reason="action_failed",
review_context=self._review_context(action=action, analysis=analysis, result=result, ip=ip),
)
self._record_ip_failure(state, ip, action, result.error_summary or str(result.values))
if action != "download-log":
self._download_log_best_effort(state, ip)
@ -388,6 +456,13 @@ class PamDeployAgent:
"message": result.values.get("MESSAGE", "ok"),
}
)
if analysis is not None and not analysis.should_continue:
self.pause_state(
state,
reason="llm_review_blocked",
review_context=self._review_context(action=action, analysis=analysis, result=result, ip=ip),
)
return state
self._save_checkpoint(state)
return state
@ -433,6 +508,9 @@ class PamDeployAgent:
}
)
state.pending_confirmation = ""
state.paused = False
state.pause_reason = ""
state.review_context = {}
self._save_checkpoint(state)
return state
@ -474,6 +552,9 @@ class PamDeployAgent:
state.pending_confirmation = ""
state.last_success_step = "rollback-ip"
state.last_failed_step = ""
state.paused = False
state.pause_reason = ""
state.review_context = {}
self._emit_progress(
{
"type": "ACTION_DONE",
@ -486,6 +567,8 @@ class PamDeployAgent:
else:
state.pending_confirmation = f"rollback-ip:{ip}"
state.last_failed_step = "rollback-ip"
state.paused = True
state.pause_reason = "rollback_failed"
self._emit_progress(
{
"type": "ACTION_FAIL",
@ -652,17 +735,23 @@ class PamDeployAgent:
result,
*,
ip: str | None = None,
) -> None:
) -> Any:
"""启用 action 后分析时，把诊断结果追加到 events。"""
if not self.action_analysis_enabled:
return
self._emit_progress(
{
"type": "ACTION_REVIEW_START",
"stage": action,
"ip": ip or "",
"message": "LLM 开始分析 action 结果",
}
)
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
except Exception as exc: # pragma: no cover - 审核失败时也要显式暂停，避免黑盒继续执行
state.events.append(
{
"type": "ACTION_ANALYSIS_FAIL",
@ -671,12 +760,42 @@ class PamDeployAgent:
"message": str(exc),
}
)
return
self._emit_progress(
{
"type": "ACTION_REVIEW_FAIL",
"stage": action,
"ip": ip or "",
"message": str(exc),
}
)
return LlmActionAnalysis(
action=action,
has_anomaly=True,
severity="high",
possible_reason=f"LLM 审核失败: {exc}",
suggested_action="请检查 LLM 配置、网络或 action 审核提示词文件后再继续。",
requires_confirmation=True,
should_continue=False,
notes=["action 结果未完成 LLM 审核，流程已自动暂停。"],
)
payload = asdict(analysis)
payload.update({"type": "ACTION_ANALYSIS", "stage": action})
if ip:
payload["ip"] = ip
if self.action_analysis_enabled:
state.events.append(payload)
self._emit_progress(
{
"type": "ACTION_REVIEW_DONE",
"stage": action,
"ip": ip or "",
"message": analysis.suggested_action or analysis.possible_reason or "LLM 审核完成",
"has_anomaly": analysis.has_anomaly,
"severity": analysis.severity,
"should_continue": analysis.should_continue,
}
)
return analysis
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。"""
@ -689,10 +808,42 @@ class PamDeployAgent:
"current_ip": ip or "",
"current_ip_state": state.ip_states.get(ip, {}) if ip else {},
"pending_confirmation": state.pending_confirmation,
"paused": state.paused,
"pause_reason": state.pause_reason,
"last_success_step": state.last_success_step,
"last_failed_step": state.last_failed_step,
}
def _review_context(
self,
*,
action: str,
analysis,
result,
ip: str | None = None,
) -> dict[str, Any]:
"""构造面向用户展示的审核暂停上下文。"""
context = {
"type": "action_review",
"stage": action,
"ip": ip or "",
"backend": result.backend,
"ok": result.ok,
"error_summary": result.error_summary,
}
if analysis is not None:
context.update(
{
"severity": analysis.severity,
"has_anomaly": analysis.has_anomaly,
"possible_reason": analysis.possible_reason,
"suggested_action": analysis.suggested_action,
"should_continue": analysis.should_continue,
"notes": list(analysis.notes),
}
)
return context
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")
@ -710,6 +861,8 @@ class PamDeployAgent:
f"- 成功: {success}",
f"- 失败: {failed}",
f"- 待确认: {state.pending_confirmation or '-'}",
f"- 暂停状态: {'是' if state.paused else '否'}",
f"- 暂停原因: {state.pause_reason or '-'}",
"",
"| IP | 状态 | 失败阶段 | 回滚状态 | 日志 |",
"| --- | --- | --- | --- | --- |",

16
pam_deploy_graph/cli.py
View File

@ -20,6 +20,7 @@ def add_llm_args(parser: argparse.ArgumentParser) -> None:
parser.add_argument("--llm-base-url")
parser.add_argument("--llm-api-key")
parser.add_argument("--llm-model")
parser.add_argument("--llm-action-analysis-prompt-file")
def add_mcp_args(parser: argparse.ArgumentParser) -> None:
@ -93,6 +94,7 @@ def main() -> None:
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_llm_args(run)
add_mcp_args(run)
add_action_analysis_arg(run)
@ -102,12 +104,14 @@ def main() -> None:
deploy.add_argument("--target-ip", action="append", default=[])
deploy.add_argument("--checkpoint")
deploy.add_argument("--confirm", action="store_true")
add_llm_args(deploy)
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_llm_args(resume)
add_mcp_args(resume)
add_action_analysis_arg(resume)
@ -116,17 +120,19 @@ def main() -> None:
confirm.add_argument("--decision", required=True, choices=["approve", "reject"])
confirm.add_argument("--note", default="")
confirm.add_argument("--confirm", action="store_true")
add_llm_args(confirm)
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 {}
llm_client = None
if args.command in ("analyze", "chat"):
if args.command != "preview":
llm_client = build_llm_client(
base_url=args.llm_base_url,
api_key=args.llm_api_key,
model=args.llm_model,
base_url=getattr(args, "llm_base_url", None),
api_key=getattr(args, "llm_api_key", None),
model=getattr(args, "llm_model", None),
action_analysis_prompt_path=getattr(args, "llm_action_analysis_prompt_file", None),
)
mcp_runner = None
if getattr(args, "mcp_config", None):
@ -173,6 +179,8 @@ def main() -> None:
if args.command == "resume":
state = load_agent_state(args.checkpoint)
state.checkpoint_path = state.checkpoint_path or args.checkpoint
if state.paused:
state = agent.resume_state(state)
result = run_graph_once(agent, state, flow="deploy")
print_graph_result(agent, result)
return

122
pam_deploy_graph/interactive.py
View File

@ -19,6 +19,7 @@ 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
from .params_loader import load_params_file
InputFunc = Callable[[str], str]
OutputFunc = Callable[[str], None]
@ -30,9 +31,9 @@ COMMAND_HELP = """可用命令：
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 config KEY=VALUE 配置真实 LLM，支持 base_url/api_key/model/action_analysis_prompt_file
llm fallback 切回本地规则 fallback
llm action-analysis on|off 开关 action 后诊断
llm action-analysis on|off 开关 action 审核详情写入 events
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
@ -40,11 +41,13 @@ COMMAND_HELP = """可用命令：
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 params <路径> 加载并热更新参数文件
load checkpoint <路径> 加载指定 checkpoint
checkpoint 显示 checkpoint 路径
exit 退出
也可以直接输入自然语言需求，Agent 会先分析并更新会话参数；执行仍需输入 run。
执行中可按 Ctrl+C 中断，保存 checkpoint 后再用 resume 继续。
"""
@ -85,6 +88,9 @@ class InteractiveCliSession:
while True:
try:
line = self.input("pam-deploy-agent> ")
except KeyboardInterrupt:
self.output("已取消当前输入。输入 exit 退出，或继续输入命令。")
continue
except EOFError:
self.output("bye")
return
@ -148,6 +154,9 @@ class InteractiveCliSession:
if normalized == "list" and rest.strip().lower() == "checkpoints":
self._list_checkpoints()
return True
if normalized == "load" and rest.strip().lower().startswith("params"):
self._load_params(rest.strip()[len("params") :].strip())
return True
if normalized == "load" and rest.strip().lower().startswith("checkpoint"):
self._load_checkpoint(rest.strip()[len("checkpoint") :].strip())
return True
@ -184,6 +193,7 @@ class InteractiveCliSession:
user_ips = param_result.extracted_control.get("user_specified_ips")
if isinstance(user_ips, list):
self.target_ips = [str(item) for item in user_ips]
self._sync_params_to_state()
safe_payload = redact_mapping({key: asdict(value) for key, value in result.items()})
self.output("已生成结构化理解：")
@ -208,6 +218,7 @@ class InteractiveCliSession:
self.output("参数名不能为空。")
return
self.params[key] = value.strip()
self._sync_params_to_state()
self.output(f"已设置 {key}")
def _show_params(self) -> None:
@ -230,7 +241,7 @@ class InteractiveCliSession:
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")
self.output("格式：llm config base_url=... api_key=... model=... action_analysis_prompt_file=... | llm fallback | llm action-analysis on|off")
return
parts = shlex.split(text)
if parts[0] == "fallback":
@ -243,7 +254,7 @@ class InteractiveCliSession:
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'}。")
self.output(f"action 审核详情写入 events 已{'开启' if self.agent.action_analysis_enabled else '关闭'}。")
return
if parts[0] != "config":
self.output("未知 llm 命令。")
@ -255,6 +266,7 @@ class InteractiveCliSession:
base_url=self.llm_config.get("base_url"),
api_key=self.llm_config.get("api_key"),
model=self.llm_config.get("model"),
action_analysis_prompt_path=self.llm_config.get("action_analysis_prompt_file"),
)
except Exception as exc:
self.output(f"LLM 配置失败: {exc}")
@ -315,6 +327,31 @@ class InteractiveCliSession:
self.output(f"已加载 checkpoint: {checkpoint}")
if self.state.pending_confirmation:
self._print_confirmation()
self._print_pause_context()
def _load_params(self, path_text: str) -> None:
"""从参数文件热更新当前会话参数，并同步到已暂停 state。"""
if not path_text:
self.output("格式：load params <路径>")
return
path = Path(path_text)
if not path.exists():
self.output(f"参数文件不存在: {path}")
return
try:
updates = load_params_file(path)
except Exception as exc:
self.output(f"参数文件加载失败: {exc}")
return
self.params.update(updates)
try:
self.params = self.agent.normalize_params(self.params)
except ValueError as exc:
self.output(f"参数热更新失败: {exc}")
return
self._sync_params_to_state()
self.output(f"已加载参数文件: {path}")
self.output(_format_redacted_params(redact_mapping(self.params)))
def _run_deploy(self) -> None:
"""在用户确认后创建状态并执行完整部署流程。"""
@ -370,6 +407,8 @@ class InteractiveCliSession:
return
self.state = load_agent_state(checkpoint)
self.state.checkpoint_path = self.state.checkpoint_path or str(checkpoint)
if self.state.paused:
self.state = self.agent.resume_state(self.state)
if self.graph_runtime and self.graph_runtime.waiting_confirmation:
self._print_confirmation()
return
@ -388,6 +427,9 @@ class InteractiveCliSession:
self.graph_runtime = None
try:
self.state = self.agent.run_deploy_flow(self.state)
except KeyboardInterrupt:
self._handle_execution_interrupt()
return
except Exception as fallback_exc:
self._handle_execution_error(fallback_exc)
return
@ -395,6 +437,9 @@ class InteractiveCliSession:
return
try:
result = self.graph_runtime.start(self.state)
except KeyboardInterrupt:
self._handle_execution_interrupt()
return
except Exception as exc:
self._handle_execution_error(exc)
return
@ -432,11 +477,27 @@ class InteractiveCliSession:
return
if self.state.last_failed_step:
self.output(f"最后失败步骤: {self.state.last_failed_step}")
self._print_pause_context()
if self.state.pending_confirmation:
self._print_confirmation()
self.output(f"checkpoint: {self.state.checkpoint_path or self.checkpoint_path}")
self.output("请修正参数或外部环境后，使用 load checkpoint <路径> / resume 继续，或重新 run。")
def _handle_execution_interrupt(self) -> None:
"""处理执行中的用户中断，并保留断点。"""
if self.state is None:
self.output("执行已中断。")
return
self.graph_runtime = None
self.state = self.agent.pause_state(
self.state,
reason="user_interrupted",
review_context={"type": "user_interrupt", "message": "用户手动中断执行"},
)
self.output("执行已由用户中断，当前 checkpoint 已保存。")
self._print_pause_context()
self.output(f"checkpoint: {self.state.checkpoint_path or self.checkpoint_path}")
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:
@ -449,6 +510,7 @@ class InteractiveCliSession:
self._print_confirmation_request(result.confirmation)
elif self.state.pending_confirmation:
self._print_confirmation()
self._print_pause_context()
self.output(f"checkpoint: {self.state.checkpoint_path or self.checkpoint_path}")
def _print_state_report_and_checkpoint(self) -> None:
@ -456,6 +518,7 @@ class InteractiveCliSession:
if self.state is None:
return
self.output(self.agent.render_report(self.state))
self._print_pause_context()
if self.state.pending_confirmation:
self._print_confirmation()
self.output(f"checkpoint: {self.state.checkpoint_path or self.checkpoint_path}")
@ -467,6 +530,7 @@ class InteractiveCliSession:
self.output(f"checkpoint: {self.checkpoint_path}")
return
self.output(self.agent.render_report(self.state))
self._print_pause_context()
if self.state.pending_confirmation:
self._print_confirmation()
@ -495,9 +559,49 @@ class InteractiveCliSession:
self.state = self.agent.confirm_pending(self.state, approved=approved, operator_note=note)
self.output(self.agent.render_report(self.state))
self._print_pause_context()
if self.state.pending_confirmation:
self._print_confirmation()
def _sync_params_to_state(self) -> None:
"""若当前已有 state，则把热更新参数同步到 checkpoint/config。"""
if self.state is None:
return
try:
self.state = self.agent.update_state_params(self.state, self.params)
except ValueError as exc:
self.output(f"参数同步到当前任务失败: {exc}")
return
self.params = dict(self.state.params)
if self.target_ips:
self.state.target_ips = list(self.target_ips)
def _print_pause_context(self) -> None:
"""输出暂停原因和审核建议，避免黑盒暂停。"""
if self.state is None or not self.state.paused:
return
context = self.state.review_context or {}
reason = self.state.pause_reason or "unknown"
self.output(f"当前流程已暂停: {reason}")
if context.get("stage"):
self.output(f"- stage: {context.get('stage')}")
if context.get("ip"):
self.output(f"- ip: {context.get('ip')}")
if context.get("possible_reason"):
self.output(f"- reason: {context.get('possible_reason')}")
elif context.get("error_summary"):
self.output(f"- reason: {context.get('error_summary')}")
if context.get("suggested_action"):
self.output(f"- suggestion: {context.get('suggested_action')}")
if context.get("severity"):
self.output(f"- severity: {context.get('severity')}")
if context.get("notes"):
self.output("- notes: " + "; ".join(str(item) for item in context.get("notes", [])))
if reason == "user_interrupted":
self.output("输入 resume 可从当前 checkpoint 继续。")
elif reason == "llm_review_blocked":
self.output("请根据以上建议判断后续；如需继续，输入 resume。")
def _on_progress(self, payload: dict[str, Any]) -> None:
"""把 Agent action 进度转成 chat 可见输出。"""
event_type = str(payload.get("type", ""))
@ -519,6 +623,14 @@ class InteractiveCliSession:
elif event_type == "ACTION_FAIL":
detail = f": {message}" if message else ""
self.output(f"失败 action: {stage}{suffix}{detail}")
elif event_type == "ACTION_REVIEW_START":
self.output(f"开始分析 action 结果: {stage}{suffix}")
elif event_type == "ACTION_REVIEW_DONE":
detail = f": {message}" if message else ""
self.output(f"分析完成: {stage}{suffix}{detail}")
elif event_type == "ACTION_REVIEW_FAIL":
detail = f": {message}" if message else ""
self.output(f"分析失败: {stage}{suffix}{detail}")
def _print_confirmation(self) -> None:
"""输出当前待人工确认事项。"""
@ -559,6 +671,7 @@ class InteractiveCliSession:
self.output(f"已加载 checkpoint: {checkpoint}")
if self.state.pending_confirmation:
self._print_confirmation()
self._print_pause_context()
def run_interactive_chat(
@ -704,6 +817,7 @@ def _build_prompt_input(input_func: InputFunc) -> InputFunc:
"reject",
"resume",
"list checkpoints",
"load params",
"load checkpoint",
"checkpoint",
"exit",

2
pam_deploy_graph/llm/base.py
View File

@ -42,5 +42,5 @@ class LlmClient(Protocol):
result: ActionResult,
state_summary: dict[str, Any],
) -> LlmActionAnalysis:
"""分析 action 执行结果，并给出辅助诊断建议。"""
"""分析 action 执行结果，并给出是否允许继续执行的建议。"""
...

15
pam_deploy_graph/llm/factory.py
View File

@ -5,7 +5,7 @@ from __future__ import annotations
import os
from .base import LlmClient
from .openai_compatible import OpenAICompatibleLlmClient
from .openai_compatible import OpenAICompatibleLlmClient, load_prompt_text
from .rule_based import RuleBasedLlmClient
@ -14,11 +14,17 @@ def build_llm_client(
base_url: str | None = None,
api_key: str | None = None,
model: str | None = None,
action_analysis_prompt_path: str | None = None,
) -> LlmClient:
"""根据显式参数或环境变量构造 LLM client。"""
actual_base_url = base_url or os.getenv("PAM_LLM_BASE_URL", "")
actual_api_key = api_key or os.getenv("PAM_LLM_API_KEY", "")
actual_model = model or os.getenv("PAM_LLM_MODEL", "")
actual_base_url = base_url if base_url is not None else os.getenv("PAM_LLM_BASE_URL", "")
actual_api_key = api_key if api_key is not None else os.getenv("PAM_LLM_API_KEY", "")
actual_model = model if model is not None else os.getenv("PAM_LLM_MODEL", "")
actual_action_prompt_path = (
action_analysis_prompt_path
if action_analysis_prompt_path is not None
else os.getenv("PAM_LLM_ACTION_ANALYSIS_PROMPT_FILE", "")
)
if not actual_base_url and not actual_api_key and not actual_model:
return RuleBasedLlmClient()
@ -35,4 +41,5 @@ def build_llm_client(
base_url=actual_base_url,
api_key=actual_api_key,
model=actual_model,
action_analysis_prompt=load_prompt_text(actual_action_prompt_path),
)

14
pam_deploy_graph/llm/openai_compatible.py
View File

@ -7,6 +7,7 @@
from __future__ import annotations
import json
from pathlib import Path
import urllib.request
from collections.abc import Callable
from typing import Any
@ -36,6 +37,7 @@ class OpenAICompatibleLlmClient:
base_url: str,
api_key: str,
model: str,
action_analysis_prompt: str | None = None,
timeout_sec: float = 30,
temperature: float = 0,
transport: JsonTransport | None = None,
@ -48,6 +50,7 @@ class OpenAICompatibleLlmClient:
self.base_url = base_url.rstrip("/")
self.api_key = api_key
self.model = model
self.action_analysis_prompt = action_analysis_prompt or ACTION_ANALYSIS_PROMPT
self.timeout_sec = timeout_sec
self.temperature = temperature
self.transport = transport or _default_transport
@ -135,7 +138,7 @@ class OpenAICompatibleLlmClient:
) -> 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,
self.action_analysis_prompt,
{
"action": action,
"result": {
@ -157,6 +160,7 @@ class OpenAICompatibleLlmClient:
possible_reason=_string(payload, "possible_reason", ""),
suggested_action=_string(payload, "suggested_action", ""),
requires_confirmation=bool(payload.get("requires_confirmation", False)),
should_continue=bool(payload.get("should_continue", True)),
notes=_string_list(payload.get("notes")),
)
@ -213,6 +217,14 @@ def _default_transport(
return decoded
def load_prompt_text(path: str | None) -> str:
"""读取自定义提示词文件。"""
if not path:
return ACTION_ANALYSIS_PROMPT
prompt_path = Path(path)
return prompt_path.read_text(encoding="utf-8").strip() or ACTION_ANALYSIS_PROMPT
def _chat_completions_url(base_url: str) -> str:
"""把 base_url 规范化为 chat/completions endpoint。"""
clean = base_url.rstrip("/")

3
pam_deploy_graph/llm/prompts.py
View File

@ -76,11 +76,12 @@ ACTION_ANALYSIS_PROMPT = """分析一次 PAM action 执行结果。
"possible_reason": "...",
"suggested_action": "...",
"requires_confirmation": false,
"should_continue": true,
"notes": ["..."]
}
要求：
- 只给诊断建议，不决定继续执行、回滚或修改参数。
- 必须明确给出 `should_continue`：没有问题时为 true；存在需要人工判断的问题时为 false。
- 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。
"""

6
pam_deploy_graph/llm/rule_based.py
View File

@ -161,12 +161,14 @@ class RuleBasedLlmClient:
possible_reason = ""
suggested_action = "继续观察。"
requires_confirmation = False
should_continue = True
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。")
should_continue = False
if action == "verify-ip":
success = result.values.get("SUCCESS")
@ -177,12 +179,14 @@ class RuleBasedLlmClient:
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。")
should_continue = False
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。")
should_continue = False
if result.values.get("PENDING_AGENT_CONFIRMATION"):
has_anomaly = True
@ -191,6 +195,7 @@ class RuleBasedLlmClient:
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。")
should_continue = False
return LlmActionAnalysis(
action=action,
@ -199,6 +204,7 @@ class RuleBasedLlmClient:
possible_reason=possible_reason,
suggested_action=suggested_action,
requires_confirmation=requires_confirmation,
should_continue=should_continue,
notes=notes,
)

4
pam_deploy_graph/models.py
View File

@ -99,6 +99,7 @@ class LlmActionAnalysis:
possible_reason: str = ""
suggested_action: str = ""
requires_confirmation: bool = False
should_continue: bool = True
notes: list[str] = field(default_factory=list)
@ -126,4 +127,7 @@ class AgentState:
last_success_step: str = ""
last_failed_step: str = ""
checkpoint_path: str = ""
paused: bool = False
pause_reason: str = ""
review_context: dict[str, Any] = field(default_factory=dict)
events: list[dict[str, Any]] = field(default_factory=list)

214
pam_deploy_graph/tool_catalog.py Normal file
View File

@ -0,0 +1,214 @@
"""统一维护面向 LLM 和 runtime 的 PAM action tool schema。"""
from __future__ import annotations
from pam_deploy_graph.action_router import build_action_backends
from pam_deploy_graph.constants import GLOBAL_ACTION_SEQUENCE, IP_ACTION_SEQUENCE
from pam_deploy_graph.models import AgentExecutionMode, ExecutionStrategy
from pam_deploy_graph.models import ActionToolSpec, SkillPolicy
ACTION_TOOL_SPECS: dict[str, ActionToolSpec] = {
"get-token": ActionToolSpec(
name="get_token",
action="get-token",
scope="global",
description="获取 PAM HOME OAuth token。",
risk_level="low",
required_param_fields=("HOME_BASE_URL", "CLIENT_ID", "CLIENT_SECRET"),
preferred_backend="script",
),
"create-version": ActionToolSpec(
name="create_version",
action="create-version",
scope="global",
description="创建版本记录。",
risk_level="medium",
preferred_backend="script",
),
"upload-package": ActionToolSpec(
name="upload_package",
action="upload-package",
scope="global",
description="上传软件包并返回 HASH_CODE。",
risk_level="high",
preferred_backend="script",
),
"publish-version": ActionToolSpec(
name="publish_version",
action="publish-version",
scope="global",
description="发布版本，需要已有 HASH_CODE。",
risk_level="high",
requires_confirmation=True,
required_runtime_fields=("hash_code",),
preferred_backend="script",
),
"get-node-url": ActionToolSpec(
name="get_node_url",
action="get-node-url",
scope="global",
description="获取目标 PAM NODE 地址。",
risk_level="low",
preferred_backend="script",
),
"get-online-ips": ActionToolSpec(
name="get_online_ips",
action="get-online-ips",
scope="global",
description="获取当前在线工作站 IP 列表。",
risk_level="low",
),
"create-download-task": ActionToolSpec(
name="create_download_task",
action="create-download-task",
scope="global",
description="创建云下载任务。",
risk_level="high",
),
"poll-download-progress": ActionToolSpec(
name="poll_download_progress",
action="poll-download-progress",
scope="global",
description="轮询云下载任务进度。",
risk_level="medium",
),
"upgrade-ip": ActionToolSpec(
name="upgrade_ip",
action="upgrade-ip",
scope="ip",
description="对单个工作站创建升级任务。",
risk_level="high",
requires_confirmation=True,
),
"poll-upgrade-progress": ActionToolSpec(
name="poll_upgrade_progress",
action="poll-upgrade-progress",
scope="ip",
description="轮询单个工作站升级进度。",
risk_level="medium",
),
"start-ip": ActionToolSpec(
name="start_ip",
action="start-ip",
scope="ip",
description="启动单个工作站应用。",
risk_level="high",
requires_confirmation=True,
),
"stop-ip": ActionToolSpec(
name="stop_ip",
action="stop-ip",
scope="ip",
description="停止单个工作站应用。",
risk_level="high",
requires_confirmation=True,
),
"verify-ip": ActionToolSpec(
name="verify_ip",
action="verify-ip",
scope="ip",
description="校验单个工作站版本和健康状态。",
risk_level="medium",
),
"download-log": ActionToolSpec(
name="download_log",
action="download-log",
scope="ip",
description="下载单个工作站日志。",
risk_level="low",
),
"rollback-ip": ActionToolSpec(
name="rollback_ip",
action="rollback-ip",
scope="ip",
description="对单个工作站执行回滚。",
risk_level="high",
requires_confirmation=True,
),
}
def ordered_actions_for_skill(policy: SkillPolicy) -> list[str]:
"""根据 skill 策略返回默认 action 顺序。"""
global_actions = list(policy.action_sequence or GLOBAL_ACTION_SEQUENCE)
ip_actions = list(policy.ip_action_sequence or IP_ACTION_SEQUENCE)
return [*global_actions, *ip_actions]
ACTION_DEPENDENCIES: dict[str, tuple[str, ...]] = {
"create-version": ("get-token",),
"upload-package": ("get-token", "create-version"),
"publish-version": ("get-token", "create-version", "upload-package"),
"get-node-url": ("get-token",),
"get-online-ips": ("get-token", "get-node-url"),
"create-download-task": ("get-token", "get-node-url", "get-online-ips"),
"poll-download-progress": ("get-token", "get-node-url", "get-online-ips", "create-download-task"),
}
for _ip_action in IP_ACTION_SEQUENCE:
ACTION_DEPENDENCIES[_ip_action] = tuple(GLOBAL_ACTION_SEQUENCE)
def allowed_tool_specs(policy: SkillPolicy) -> list[ActionToolSpec]:
"""按 skill 限制过滤并排序 tool specs。"""
ordered_actions = ordered_actions_for_skill(policy)
specs: list[ActionToolSpec] = []
for action in ordered_actions:
if action not in policy.allowed_actions:
continue
if action in policy.forbidden_actions:
continue
spec = ACTION_TOOL_SPECS.get(action)
if spec is not None:
specs.append(spec)
return specs
def tool_summaries(policy: SkillPolicy, strategy: ExecutionStrategy) -> list[dict[str, str]]:
"""生成给 LLM 使用的受控 tool 摘要。"""
routes = build_action_backends(strategy)
summaries: list[dict[str, str]] = []
for spec in allowed_tool_specs(policy):
summaries.append(
{
"name": spec.name,
"action": spec.action,
"scope": spec.scope,
"description": spec.description,
"risk_level": spec.risk_level,
"backend": routes.get(spec.action, spec.preferred_backend or ""),
"requires_confirmation": "true" if spec.requires_confirmation else "false",
}
)
return summaries
def normalize_planned_actions(
planned_actions: list[str],
*,
policy: SkillPolicy,
mode: AgentExecutionMode,
) -> list[str]:
"""按 skill 限制和依赖关系归一化 planned actions。"""
allowed = set(policy.allowed_actions)
forbidden = set(policy.forbidden_actions)
ordered = ordered_actions_for_skill(policy)
if not planned_actions:
return [action for action in ordered if action in allowed and action not in forbidden]
normalized: list[str] = []
for action in planned_actions:
if action in allowed and action not in forbidden and action not in normalized:
normalized.append(action)
expanded: list[str] = []
for action in normalized:
for dependency in ACTION_DEPENDENCIES.get(action, ()):
if dependency in allowed and dependency not in forbidden and dependency not in expanded:
expanded.append(dependency)
if action not in expanded:
expanded.append(action)
global_order = [action for action in ordered if action in GLOBAL_ACTION_SEQUENCE and action in expanded]
ip_order = [action for action in ordered if action in IP_ACTION_SEQUENCE and action in expanded]
return [*global_order, *ip_order]

18
prompts/action_review.txt Normal file
View File

@ -0,0 +1,18 @@
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,
"should_continue": true,
"notes": ["..."]
}
要求：
- 必须明确给出 `should_continue`：没有问题时为 true；存在需要人工判断的问题时为 false。
- 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。

114
tests/test_agent_flow.py
View File

@ -6,6 +6,7 @@ from pam_deploy_graph.agent import PamDeployAgent
from pam_deploy_graph.checkpoint_store import load_agent_state
from pam_deploy_graph.constants import GLOBAL_ACTION_SEQUENCE
from pam_deploy_graph.fake_runner import FakeActionRunner
from pam_deploy_graph.models import LlmActionAnalysis
PARAMS = {
@ -20,6 +21,25 @@ PARAMS = {
}
class BlockingReviewLlmClient:
def analyze_action_result(self, *, action, result, state_summary):
return LlmActionAnalysis(
action=action,
has_anomaly=True,
severity="high",
possible_reason="review blocked",
suggested_action="stop and inspect",
requires_confirmation=True,
should_continue=False,
notes=["blocked by test llm"],
)
class BrokenReviewLlmClient:
def analyze_action_result(self, *, action, result, state_summary):
raise RuntimeError("review transport failed")
def test_run_deploy_flow_success(tmp_path: Path):
agent = PamDeployAgent(fake_runner=FakeActionRunner())
state = agent.create_state(
@ -124,6 +144,49 @@ def test_action_analysis_event_is_recorded_when_enabled(tmp_path: Path):
assert verify_analysis["requires_confirmation"] is True
def test_successful_action_can_be_blocked_by_llm_review(tmp_path: Path):
agent = PamDeployAgent(
fake_runner=FakeActionRunner(),
llm_client=BlockingReviewLlmClient(),
)
state = agent.create_state(
params=PARAMS,
execution_strategy="fake",
config_path=str(tmp_path / "config.txt"),
checkpoint_path=str(tmp_path / "checkpoint.json"),
)
agent.run_deploy_flow(state)
assert state.paused is True
assert state.pause_reason == "llm_review_blocked"
assert state.last_failed_step == "get-token"
assert state.completed_global_steps == ["get-token"]
assert state.review_context["stage"] == "get-token"
assert state.review_context["suggested_action"] == "stop and inspect"
def test_action_review_failure_pauses_flow(tmp_path: Path):
agent = PamDeployAgent(
fake_runner=FakeActionRunner(),
llm_client=BrokenReviewLlmClient(),
)
state = agent.create_state(
params=PARAMS,
execution_strategy="fake",
config_path=str(tmp_path / "config.txt"),
checkpoint_path=str(tmp_path / "checkpoint.json"),
)
agent.run_deploy_flow(state)
assert state.paused is True
assert state.pause_reason == "llm_review_blocked"
assert state.review_context["stage"] == "get-token"
assert "LLM 审核失败" in state.review_context["possible_reason"]
assert any(event["type"] == "ACTION_ANALYSIS_FAIL" for event in state.events)
def test_confirm_pending_rollback_runs_rollback_and_resume_continues(tmp_path: Path):
fake = FakeActionRunner(
{
@ -215,3 +278,54 @@ def test_checkpoint_resume_skips_completed_global_and_success_ip(tmp_path: Path)
assert "get-token" not in called_actions
assert all(call[1].get("ip") != "192.168.1.10" for call in fake.calls)
assert loaded.ip_states["192.168.1.11"]["status"] == "SUCCESS"
def test_update_state_params_rewrites_config_and_checkpoint(tmp_path: Path):
initial_package = tmp_path / "pkg-a.zip"
updated_package = tmp_path / "pkg-b.zip"
checkpoint = tmp_path / "checkpoint.json"
config_path = tmp_path / "config.txt"
agent = PamDeployAgent(fake_runner=FakeActionRunner())
state = agent.create_state(
params={**PARAMS, "ZIP_FILE_PATH": str(initial_package)},
execution_strategy="fake",
config_path=str(config_path),
checkpoint_path=str(checkpoint),
)
agent.update_state_params(
state,
{
"APP_NAME": "PAM-NEW",
"ZIP_FILE_PATH": str(updated_package),
},
)
loaded = load_agent_state(checkpoint)
config_text = config_path.read_text(encoding="utf-8")
assert state.params["APP_NAME"] == "PAM-NEW"
assert state.params["ZIP_FILE_PATH"] == str(updated_package.resolve())
assert loaded.params["APP_NAME"] == "PAM-NEW"
assert loaded.params["ZIP_FILE_PATH"] == str(updated_package.resolve())
assert "APP_NAME=PAM-NEW" in config_text
assert f"ZIP_FILE_PATH={updated_package.resolve()}" in config_text
def test_resume_state_clears_pause_fields(tmp_path: Path):
checkpoint = tmp_path / "checkpoint.json"
agent = PamDeployAgent(fake_runner=FakeActionRunner())
state = agent.create_state(
params=PARAMS,
execution_strategy="fake",
checkpoint_path=str(checkpoint),
)
agent.pause_state(state, reason="manual_test", review_context={"stage": "get-token"})
resumed = agent.resume_state(state)
loaded = load_agent_state(checkpoint)
assert resumed.paused is False
assert resumed.pause_reason == ""
assert resumed.review_context == {}
assert loaded.paused is False
assert loaded.pause_reason == ""

79
tests/test_interactive_cli.py
View File

@ -6,6 +6,7 @@ import pytest
from pam_deploy_graph.agent import PamDeployAgent
from pam_deploy_graph.fake_runner import FakeActionRunner
from pam_deploy_graph.interactive import InteractiveCliSession, _build_prompt_input
from pam_deploy_graph.models import LlmActionAnalysis
PARAMS = {
@ -20,6 +21,20 @@ PARAMS = {
}
class BlockingReviewLlmClient:
def analyze_action_result(self, *, action, result, state_summary):
return LlmActionAnalysis(
action=action,
has_anomaly=True,
severity="high",
possible_reason="review blocked",
suggested_action="stop and inspect",
requires_confirmation=True,
should_continue=False,
notes=["blocked by test llm"],
)
def run_session(session: InteractiveCliSession, inputs: list[str]) -> list[str]:
output: list[str] = []
iterator = iter(inputs)
@ -74,6 +89,8 @@ def test_chat_run_prints_action_progress(tmp_path: Path):
assert any("开始执行 action: get-token" in item for item in output)
assert any("完成 action: verify-ip" in item for item in output)
assert any("开始分析 action 结果: get-token" in item for item in output)
assert any("分析完成: verify-ip" in item for item in output)
def test_chat_greeting_does_not_trigger_structured_analysis(tmp_path: Path):
@ -181,6 +198,68 @@ def test_chat_params_events_and_checkpoint_commands(tmp_path: Path):
assert any("checkpoint 列表" in item for item in output)
def test_chat_load_params_hot_updates_running_state_and_config(tmp_path: Path):
checkpoint = tmp_path / "checkpoint.json"
params_file = tmp_path / "params.txt"
params_file.write_text(
"\n".join(
[
"APP_NAME=PAM-HOT",
f"ZIP_FILE_PATH={tmp_path / 'updated.zip'}",
]
)
+ "\n",
encoding="utf-8",
)
session = InteractiveCliSession(
agent=PamDeployAgent(fake_runner=FakeActionRunner()),
params=PARAMS,
strategy="fake",
checkpoint_path=str(checkpoint),
)
run_session(
session,
[
"run",
"yes",
"yes",
"yes",
"load params " + str(params_file),
"exit",
],
)
assert session.state is not None
assert session.state.params["APP_NAME"] == "PAM-HOT"
assert session.state.params["ZIP_FILE_PATH"] == str((tmp_path / "updated.zip").resolve())
config_text = Path(session.state.config_path).read_text(encoding="utf-8")
assert "APP_NAME=PAM-HOT" in config_text
assert f"ZIP_FILE_PATH={(tmp_path / 'updated.zip').resolve()}" in config_text
def test_chat_llm_review_block_message_is_visible(tmp_path: Path):
checkpoint = tmp_path / "checkpoint.json"
session = InteractiveCliSession(
agent=PamDeployAgent(
fake_runner=FakeActionRunner(),
llm_client=BlockingReviewLlmClient(),
),
params=PARAMS,
strategy="fake",
checkpoint_path=str(checkpoint),
)
output = run_session(session, ["run", "yes", "yes", "yes", "exit"])
assert session.state is not None
assert session.state.paused is True
assert session.state.pause_reason == "llm_review_blocked"
assert any("当前流程已暂停: llm_review_blocked" in item for item in output)
assert any("- suggestion: stop and inspect" in item for item in output)
assert any("如需继续，输入 resume" 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")