code-review

Ars Technica reports on a study showing AI coding agents boost code output but slow overall delivery, with pull request review times rising 49%, changes-requested rates nearly doubling, and no significant employment changes attributable to AI so far.

GitGlow 发布于 Product Hunt，提供精确暂存、清晰历史记录和 checkout 感知终端的一体化本地 Git 工作区，并支持 AI 代理变更审查；个人本地克隆免费使用，无需注册账号。

Diffle 是一个键盘优先的浏览器端 Git diff 查看工具，支持按行、块或文件添加评论，并可将评论作为 prompt 交给 AI agent 或提交为 GitHub 待处理评审，内置 LSP 支持且可通过 pixi、npm 或独立二进制安装。

GitHub announced ReviewBench, an open benchmark for AI code review agents built from analysis of 103.9M GitHub pull requests, featuring a 219-PR corpus across 19 languages with multi-source ground truth and precision/recall metrics so teams can evaluate and compare their own review agents.

The author introduces Hutch, an open-source Emacs/Magit-based interface that runs LLM-powered code reviews locally on staged, unpushed, or branch-diff changes, letting developers queue and bulk-apply suggestion patches inside a familiar keyboard-driven workflow.

Review is a free, open-source (MIT) desktop app for reviewing Git branches and GitHub PRs locally, letting you comment privately and add AI reviewers powered by your own models (OpenAI, Anthropic, Gemini, OpenRouter, or local via Ollama), splitting review tasks across a team of models with per-rule visibility and targeted retries.

Matt Pocock shares a daily prompt idea: retroactively analyze GitHub PR review comments from a team and generate suggestions for updating CODING_STANDARDS.md, turning past code review feedback into forward-looking coding guidelines.

Flirt, a local-first code review tool designed for the amend/rebase/force-push patch series workflow, has been open-sourced on Codeberg. It tracks previously reviewed state and diffs against the current changes, supporting multiple backends like GitHub and mailing lists, though the author warns it's not yet ready for general use.

DHH argues that reviewing every line of AI-agent-generated code is unsustainable and that the future lies in adversarial agent reviews plus automated testing, while @msimoni pushes back, expressing distrust of software written by AI that nobody actually read.

Reviu relaunches as a review workspace for code written by coding agents, letting users run Claude Code, Codex, Gemini and other ACP agents in durable sessions while inspecting diffs, adding inline comments, and managing Git branches, with Pro adding GitHub PR review and merge features.

A tweet criticizes veteran developers like Uncle Bob Martin and DHH for endorsing 'AI psychosis' around letting coding agents review code without human oversight, warning that skipping review risks vulnerabilities like shell injection.

A developer is building Alloyqa, a tool that lets an AI agent take a coding task, work on it in the cloud, and return a pull request for review, with preview support for iOS and macOS tasks and persistent project context. They are seeking early testers and feedback.

Baoyu's commentary on the podcast interview with Lauren Tan and Matt Pocock: Lauren builds the Grok Bot at SpaceXAI and merges 2,500 PRs a month without reviewing each one, relying instead on AI-driven automated verification, code conventions, and an open-source Skill system called pstack to guarantee quality. The article cautions readers not to rush to copy her approach — first understand what makes it possible for her: models strong enough, token costs, the limits of AI verification, and having your own domain expertise to steer things.

GitHub's blog walks beginners through the GitHub Copilot app's built-in diff, terminal, and browser panels, showing how developers can review, run, and preview agent-generated code in one place without tab hopping.

A developer at a SaaS company describes how his PM built a 3,000-line feature over the weekend using Claude and CodeRabbit, reflecting the growing friction between AI-assisted coding and traditional code review practices.

The author describes running an overnight bug-fixing experiment where a cheap model (DeepSeek) opened 14 PRs, but only 6 survived review by a stronger model (GPT-6.1 Sol) due to regressions, ineffective tests, and security bypasses. They then flipped the pipeline to have the strong model design and review while the cheap model only implements.

A practitioner with C/C++ experience at Microsoft and Google runs 12 AI code-review experiments on real PRs, finding that fresh reviewer agents with no shared context catch far more defects, that executing code (not just reading it) is essential for logic bugs, and that model choice affects cost-effectiveness more than raw capability.

Charlie Marsh argues that 'no one writes code anymore' is no longer a meaningful differentiator in AI-driven development; the key question is now whether teams still read and review code at all.

heal.dev 推出一款 AI 驱动的代码审查补充工具，通过构建带外部服务桩的生产级沙箱来运行应用，自动发现传统 code review 无法捕捉的 bug、性能回退和运行时漏洞，并在报告前复现每个缺陷。

Qodo 3.0 发布，作为面向 agentic software factory 的质量与治理层，支持导入 AGENTS.md 和 CLAUDE.md 文件作为代码审查标准，使现有的 agent 编码规则成为审查依据。