展示HN：案例研究：一个编码代理重构了一个75万行代码的应用，无需代码审查
在以下案例研究中，AI编码代理在仅仅三天内重建了一个核心系统的不变性，几乎没有人类干预，进行了31次验证，并修正了201个错误，最终交付的代码没有任何漏洞、回归或技术债务。
免责声明：
本案例研究并不是：
- 从头开始编写的项目
- 又一个Rust转译
- 在大型语言模型训练数据中找到的开源项目的克隆
- HTML中的超级马里奥克隆
它是：
- 一项复杂的应用维护任务，涉及一个独特的问题，尽我所知没有开源等效物
- 一项真正的工程壮举，90%由ChatGPT 5.6 Sol Reasoning MAX驱动的代理完成
- 由AI代理自动修复的201个错误、模糊性和架构缺陷
- 证明在复杂的遗留软件中可以完美编程，而无需人类代码审查
欢迎反馈。
查看原文
In the following case study, the AI coding agent rebuilds a core system invariant in just three days, with minimal human input, running 31 verification passes and correcting 201 errors, before shipping code with zero bugs, zero regressions, and zero technical debt.<p><a href="https://aisovereignlabs.ai/docs/case-study/liveSession/case-study-live-session-en.pdf" rel="nofollow">https://aisovereignlabs.ai/docs/case-study/liveSession/case-...</a><p>Disclaimers:<p>This case study is NOT:
- a project written from scratch
- yet another Rust transpilation
- a clone of an open-source project found in the LLM's training data
- a Super Mario clone in HTML<p>It IS:
- a complex application maintenance task, about a unique problem with no open-source equivalent, to the best of my knowledge
- a true engineering feat, 90% driven by the agent powered by ChatGPT 5.6 Sol Reasoning MAX
- 201 bugs, ambiguities, and architectural flaws automatically fixed by the AI agent
- proof that one can program perfectly within complex legacy software without any need for human code review<p>Feedback welcome.