refine autonomous_sop single-task enforcement + subagent input.txt length rule
This commit is contained in:
@@ -6,6 +6,7 @@
|
||||
- 启动:`python agentmain.py --task {task_name} [--llm_no N]`(cwd=代码根),其中agentmain.py位于代码根目录
|
||||
- 流程:写 input.txt → 启动 → 轮询 output.txt → 读回复 → 写 reply.txt 继续 → 不写则5min自动退出
|
||||
- input.txt原则:写目标+约束,可指定SOP名。禁写具体实现步骤——除非主agent已读过该SOP确认正确。凭印象猜的步骤会误导subagent
|
||||
- **input.txt长度红线**:input.txt必须精简。只写目标+约束+关键参数。大量数据给路径让subagent自己读,禁止入input.txt
|
||||
- output.txt:首轮对话的流式输出(持续append),用mtime/size判断更新
|
||||
- output1.txt, output2.txt...:reply后各轮的流式输出(递增编号),同样持续append
|
||||
|
||||
|
||||
Reference in New Issue
Block a user