Anthropic公司Claude Code团队近日分享了其代码审查流程的演进经验。Claude AI编写了Claude Code自身约90%的代码1，随着项目规模扩大，团队面临着前所未有的代码审查压力。每日变更数量从5月的约2500个增至9月超过30000个1，促使团队将大部分审查工作从人工转向自动化检查。这一转变取得显著成效，从6月至9月期间，合并的拉取请求数量翻倍增长，同时需要还原的变更比例下降超过40%1。
在公开版本的Claude Code代码审查功能发布后，接收到实质性审查意见的内部拉取请求比例从16%上升至54%1。团队成员Boris Cherny认为，赋予Claude验证其工作的能力使最终结果提升了两到三倍1。审查过程中发现，约59%被审查人员反复标记的代码自其首次检查以来就一直存在1。在一个命令行界面拉取请求的优化过程中，该请求在一夜之间经历了超过50轮的审查与修复，直至作者决定停止1。
Anthropic's Claude AI system has authored approximately 90% of the code for Claude Code itself, according to the company's account of its evolving code review process 1. This substantial AI contribution has coincided with a dramatic acceleration in development velocity—daily code changes surged from roughly 2,500 in May to exceeding 30,000 by September 1.
To manage this explosive growth, the Claude Code team shifted much of its code review burden from manual inspection to automated checks 1. Despite the tenfold increase in merged pull requests between June and September, the proportion of changes that required reverting fell by more than 40 percent 1. When the public code review tool for Claude Code launched, the share of internal pull requests receiving substantive review comments jumped from 16% to 54% 1. Boris Cherny noted that providing Claude with a mechanism to verify its own work improved final results by a factor of two to three 1.
The review process has revealed persistent patterns: approximately 59% of issues flagged by reviewers involved code that had remained unchanged since the reviewer's initial pass 1. In one instance, a command-line interface pull request underwent more than fifty rounds of review and fixes overnight before its author halted the process 1.
评论
还没有评论，欢迎留下第一条。