DoorDash发现单次AI代码审查仅捕获约30.7%的PR问题，为此构建了Sonnet 4.6+Opus 4.8多模型审查代理，召回率提升至53.6%，成本$3.91/PR。最新实验将审查拆分为侦察与验证两步：Kimi K2.6作为侦察模型扫描差异，Claude Fable 5作为评审验证。该方案在内部基准DashBench上达65.2%加权召回率、75.3% F1，每PR成本$3.81。Composer 2.5+GPT 5.5 medium方案精确率达92.2%，但召回率显著偏低。DoorDash同步开源了DashBench。
DoorDash showed open-source models can lift AI code review when used together.
And that harnesses now matter.
Single-pass AI reviewers caught only 30.7% of weighted PR issues.
The company’s production reviewer raised that to 53.6%, while costing $3.91 per PR.
That jump came from splitting review into two jobs, scouting first, then verifying.
A scout model scans the diff for risky areas before deeper reviewers test claims.
The strongest run used Kimi K2.6 as scout and Claude Fable 5 as reviewer.
That setup hit 65.2% weighted recall, 75.3% F1, and $3.81 per PR.
Composer 2.5 with GPT 5.5 medium reached 92.2% precision, but much lower recall.
Models out of the box have trouble reviewing DoorDash’s codebase, even when feeding them appropriate context. Single-pass AI reviewers only caught 30% of the real issues in our PRs. Instead, we built a multi-model (Sonnet 4.6 + Opus 4.8) code review agent that caught 53.6% issues at $3.91 a PR (most enterprise code review solutions perform significantly below 50% recall on DoorDash PRs). Today, we’re proud to share DashBench, our internal benchmark to compare multi-model/harness code review systems against each other. DashBench has given us confidence in maximizing the intelligence per dollar we’re spending on AI tokens to ship code safely.在 X 查看被引用的帖子
来源：Rohan Paul · x.com