Continue to talk to everyone about my technical ideas and implementation of designing Skill in Waza. This time, we will discuss how I designed/check as a code review skill. First of all, we need to know that the model itself tells you that it has been completed, but it may not necessarily be completed. It is very likely that there will be some non-existent or problematic things left behind, often hiding some problems. In Agent design, my engineering experience is to give the model a checklist to verify whether it has really been completed. This is often much better than asking it to check the effect. /When Check was first designed, it was not a large and comprehensive reviewer, but a system for organizing and dividing tasks. SKILLmd was the main reviewer responsible for grading and controlling the review process. Agents had independent security and architecture reviewers, each managing their own tasks without interfering with each other. When to bring in someone, it would be determined by an activation rule, rather than the traditional keyword matching used by everyone. Hierarchical logic is also quite interesting. For code with less than 100 lines, we can quickly review it. For code with 100-500 lines, we can add experts as needed. For code with more than 500 lines, we can fully load it. Then, we can add another round of adversarial testing. Adversarial testing will look for vulnerabilities from these four perspectives, including violating assumptions, potential problems under combination failure, problems with cascading errors between superiors and subordinates, and abuse scenarios Then the problems discovered will be dealt with in four levels. Those that can be safely fixed will be fixed directly. Those that are roughly correct will be packaged for you to manually confirm whether they are okay. Those that need to be judged will be asked, and the parts for reference only will be overestimated. However, we will not ask you every small problem, nor will we overstep our authority to help you change the behavior logic. There is also a hard requirement that validation is not considered complete until it is completed. It comes with a detection script that can recognize project types such as Cargo, TypeScript, Python, etc. to run tests. If it cannot be detected, it will directly report an error without pretending to pass. This makes me more like an experienced technical expert, reviewing some of my experience in different situations, and I happen to have easily incorporated this experience into Waza's/check skills. Finally, if you have a better understanding of the code review process, I would greatly welcome you to write an MD for Waza, haha https://(GitHub)/tw93/waza