德文评测:AI 阻止低质量内容
Devin Review: AI to Stop Slop

原始链接: https://cognition.ai/blog/devin-review

## Devin Review:AI 时代的代码审查扩展 随着 AI 编码代理增加代码输出和 PR 大小,代码审查正成为软件开发中的主要瓶颈。Devin Review 是一款新的免费工具,旨在利用 AI 增强人们对复杂代码变化的理解——无论这些代码是由人还是代理编写的。 Devin Review 通过三种关键方式改进审查流程:**智能 diff 组织**(分组逻辑更改并解释代码块)、**交互式聊天**(允许在完整代码库上下文中提问)和 **AI 漏洞检测**(标记潜在问题并带有严重程度级别)。 Devin Review 目前可通过 app.devin.ai/review 为所有公共和私有 GitHub PR 提供服务,通过 URL 交换(使用 devinreview.com 代替 github.com),或命令行工具 (`npx devin-review {pr-link}`) 使用。Devin Review 旨在克服传统代码审查的局限性,并帮助团队更快地交付更高质量的软件。它旨在解决“懒惰 LGTM”问题,并弥合 AI 辅助代码生成与实际生产力之间的差距。

## Devin AI:褒贬不一 最近的 Hacker News 讨论集中在 Cognition AI 的 Devin 上,这是一款旨在改进代码审查的 AI 工具。一些用户对 Devin 生成功能和协助识别潜在问题能力印象深刻,但对其定价模式存在担忧——特别是模糊且复杂的“ACU”系统。 一些评论员质疑 Devin 的用户基础以及在拥挤的 AI 编码领域的相关性,指出要与 Anthropic 和 OpenAI 等大型企业竞争所需的成功程度。另一些人则对依赖 AI 进行代码审查表示怀疑,担心这可能导致 PR 数量增加以及人工监督的彻底性下降。 Cognition 员工积极参与讨论,澄清 Devin Review 的设计目的是*辅助*人工审查员,而非取代他们,通过提高代码理解和组织能力来实现。他们还解决了关于 AI 生成的“垃圾”的担忧,强调 Devin 的作用是 UI 增强,而非自动评判者。尽管有这些澄清,但仍有人不信服,一位用户甚至开玩笑地抱怨该产品使用了他的名字。
相关文章

原文

As coding agents proliferate, the # of PRs increases, but the quality of that code can be mixed, and the size of each PR is increasing past maintainers' ability to understand. We're hearing from our customers that code review—not code generation—is now the bottleneck to shipping great products.

That's why we built Devin Review, a code review tool that uses state-of-the-art AI + UX to scale human understanding of ever-more-complex code diffs—whether authored by a human, or an agent.

Devin Review is currently free, while in early release, and works on any public or private GitHub PR. You can use it in three ways:

  1. Devin users: head to app.devin.ai/review to see all your open PRs.
  2. Everyone: swap github for devinreview in any PR URL (e.g. https://github.com/org/repo/pull/123 ⇒ https://devinreview.com/org/repo/pull/123). No login needed for public PRs.
  3. Everyone: npx devin-review {pr-link} - run this command inside the PR’s parent repo.

Check out the docs for more details.

The Birth and Stagnation of Code Review

15 years ago, GitHub set the standard for PR review... and then stopped there.

The first order problem with standard code review is well known - when PRs are small, they're easy to read and argue about. But this breaks down quickly for large reviews. We call this the "Lazy LGTM problem".

More broadly, the feeling that people have of extreme productivity with coding agents in their vibecoded prototypes, vs the disappointing feeling that most people actually see in the useful output of their engineers, organizations, and experience as end users, is the great mystery of our time.

To paraphrase Winston Churchill: "Never in the field of software engineering has so much code been created by so many, yet shipped to so few."

The Modern Code Review Workflow

Devin Review adds AI tooling to help you scale your PR understanding, so you quickly and fully understand the code you are about to merge.

As you review the PR, Devin helps you in every step:

  • Reading better:
    • GitHub shows you diffs by alphabetical order.
    • Solution: intelligent diff organization. PR Review analyzes your code, groups together changes that are logically connected, orders the hunks of code, and explains each hunk, so you can review from top to bottom. It's as if a smart colleague was walking you through the PR.
    • A quality-of-life improvement: when code is moved or renamed, GitHub shows the changes as full deletes and full writes. We detect what was copied/moved and don’t make a fuss.
  • Asking for more info:
    • When more context is needed on a diff, including context on code outside of the PR, GitHub doesn’t offer any solutions beyond token search.
    • Solution: Interactive chat. Devin Review pipe your diffs into an inline Ask Devin session with full codebase understanding, so you can chat about the changes, without leaving the review.
  • Catching bugs and issues
    • GitHub doesn’t catch bugs in PRs; it relies on your CI/linting. Other bugcatchers in the market are often seen as spammy and low signal.
    • Solution: AI Bug Detection. Devin Review scans the diffs and generates a list of issues categorized by seriousness: red for probable bugs, and yellow for warnings, and gray for FYI/commentary.
    • You can copy/paste or dismiss the AI flags, or otherwise just work with fellow humans in normal comment bubbles.

This is a starting point, and we’ll be pushing frequent updates as our users keep telling us what they want. We hope you'll try out Devin Review and give us feedback!

联系我们 contact @ memedata.com