@buxianggongzuole 在 codex食用方式:白天new feature,晚上debug,日烧4亿token 中发帖
自从用AI写了个项目,代码已经到10w了,后端还行,前端一堆bug
于是写了下面几个instruction:
bug_finder.md: 使用openai新出的playwright-interactive从前端找bug
bug_identifier.md: 确认bug确实是个bug
reflextor.md: 反思找到的bug
bug_fixer.md:总结以上3个agent生成的report,修bug,要求对应bug写基于playwright的integration test直到通过test
搭配脚本:
start_agent.sh:起一个codex并输入其中一个agent的instruction,如start_agent.sh bug_finder(所以每个agent间是不共享上下文的,只通过各自生成的report沟通)
auto_agent.sh:新建个修bug的g...