更新了skill

langgraph-plan-review/SKILL.md

@@ -7,6 +7,32 @@ description: Use the LangGraph plan-review workflow from the current Claude Code
 
 Use this skill to run the LangGraph plan-review workflow. All scripts are bundled under the skill's own `scripts\` directory (`C:\Users\bowei\.claude\skills\langgraph-plan-review\scripts\`) — no dependency on `E:\LangGraph_Design\lab00`.
 
+## ⚠️ 强制执行规则（MANDATORY — read first）
+
+**plan-review 工作流必须、且只能通过 `langgraph dev` 服务器来驱动。**
+
+- ✅ **唯一允许的入口**：`run_langgraph.ps1` → `run_workflow.py`。后者是一个**纯 REST/SDK 客户端**
+  （`langgraph_sdk.get_sync_client(url=http://127.0.0.1:2024)`），每次 start/resume 都把 run
+  投递到 `langgraph dev` 服务器执行。先用 `manage_dev_server.ps1 start` 把服务器拉起。
+  这样 LangGraph Studio 才能实时追踪线程、节点、中断与状态。
+
+- 🚫 **绝对禁止**：直接 `import` / 在进程内运行本 skill 的图定义模块
+  （`scripts\graph.py`、`scripts\workflow.py`、`scripts\agents.py`）来跑工作流。包括但不限于：
+  - `python -c "from graph import ...; graph.invoke(...)"` 或 `.stream(...)`
+  - 写临时脚本 build 编译图后在进程内调用
+  - 任何绕过 `http://127.0.0.1:2024` 的「就地执行」方式
+
+  这类调用在 dev server 进程**之外**运行，Studio 追踪不到状态，**一律不允许**——
+  即使只是想「快速验证一个节点」也不行。
+
+- ℹ️ **例外说明**：`review_task.py`（Codex 评审 shim）会直接调 `codex`，这是**合规的**——
+  它是由运行在 dev server 内的图通过 `codex_review_command` 调起的评审步骤，不是绕过
+  dev server 去驱动工作流本身。区别在于：驱动图 = 必须走 dev server；图内部调用的工具 = 正常。
+
+- 🛠 **改了 `graph.py`/`workflow.py`/`agents.py` 后**：dev server 内存里仍是旧代码。先
+  `manage_dev_server.ps1 stop` 再 `start`（或让 `--reload` 生效），再经 `run_langgraph.ps1`
+  重跑，使改动在被追踪的运行中生效。
+
 ## Pre-flight: Environment Setup
 
 Before running the workflow for the first time (or when unsure the environment is ready), verify that the required packages are installed in the `langgraph` conda env:
@@ -32,6 +58,7 @@ conda run -n langgraph langgraph --version
 
 ## Invariants
 
+- **Drive the graph only through the dev server** (see the mandatory rule above): `run_langgraph.ps1` → `run_workflow.py` → `http://127.0.0.1:2024`. Never import or run `graph.py`/`workflow.py`/`agents.py` in-process to execute the workflow — that bypasses the dev server and LangGraph Studio cannot track it.
 - Use the current Claude Code session for all Claude Code work. Do not launch another `claude` process.
 - Assume the current Claude Code session already has `--dangerously-skip-permissions`.
 - Run Python through the recorded conda environment. `run_langgraph.ps1` resolves conda from `-CondaExe`, `CONDA_EXE`, `conda` on PATH, or the user-profile miniconda fallback: