diff --git a/.gitignore b/.gitignore
index 54b9972..70945cb 100644
--- a/.gitignore
+++ b/.gitignore
@@ -55,6 +55,9 @@ memory/*
 # Subagent SOP
 !memory/subagent_sop.md
 
+# Plan SOP
+!memory/plan_sop.md
+
 # ADB UI tool
 !memory/adb_ui.py
 
diff --git a/agent_loop.py b/agent_loop.py
index 70344aa..edfa89a 100644
--- a/agent_loop.py
+++ b/agent_loop.py
@@ -54,8 +54,12 @@ def agent_runner_loop(client, system_prompt, user_input, handler, tools_schema,
         yield f"**LLM Running (Turn {turn+1}) ...**\n\n"
         if (turn+1) % 10 == 0: client.last_tools = ''  # 每10轮重置一次工具描述，避免上下文过大导致的模型性能下降
         response_gen = client.chat(messages=messages, tools=tools_schema)
-        response = yield from response_gen
-        if verbose: yield '\n\n'
+        if verbose:
+            response = yield from response_gen
+            yield '\n\n'
+        else:
+            response = exhaust(response_gen)
+            yield response.content
 
         if not response.tool_calls:
             tool_name, args = 'no_tool', {}
diff --git a/memory/plan_sop.md b/memory/plan_sop.md
new file mode 100644
index 0000000..ca99c34
--- /dev/null
+++ b/memory/plan_sop.md
@@ -0,0 +1,17 @@
+# Plan Mode SOP
+
+## 拆分
+1. 分析任务，必要时先读相关SOP确定子步骤
+2. 拆成可验证的原子步骤，风险步前置
+
+## 写入 checkpoint
+将以下写入 working checkpoint（忽略长度限制）：
+
+=== PLAN ===
+[ ] 步骤1
+[ ] 步骤2
+...
+=== PLAN RULES ===
+- 每完成/跳过一步，重新 update working checkpoint
+- 任何 checkpoint update 必须保留 PLAN
+================
diff --git a/memory/subagent_sop.md b/memory/subagent_sop.md
index e33f021..1793323 100644
--- a/memory/subagent_sop.md
+++ b/memory/subagent_sop.md
@@ -22,8 +22,9 @@ proc = subprocess.Popen(
 
 - 必须 Popen，禁止 subprocess.run（会阻塞）
 - stdout.log/stderr.log 用于调试subagent卡死、LLM调用失败等问题
-- `--llm_no` 默认=sonnet 4.5，`--llm_no 1`=opus 4.6
 - 文件统一 UTF-8，subagent 无 reply 5min 自动退出无需清理
+- **禁止合并启动+轮询到同一个code_run**——会阻塞自己。启动Popen立即返回，下一轮再poll output.txt。这是并行的前提
+- 新建/复用任务目录时，先删除旧 output*.txt（否则会读到上次结果误判完成）
 
 ## 场景1：测试模式 - 行为验证
 **用途**：观察agent真实行为，修正RULES/L2/L3/SOP