What we learned trying to copy Meta's RADAR: auto-approving low-risk PRs without breaking prod
retained text
On most teams, every PR waits in one queue for the same 2-3 senior devs. A README typo and a payment flow rewrite get the same attention, seniors start skimming, and risky changes slip through. Meta published how they handle this internally, a system called RADAR (paper ([https://arxiv.org/abs/2605.30208](https://arxiv.org/abs/2605.30208))). It scores every diff for risk, auto-approves the low-risk ones, and sends the risky ones to people who know that code. Auto-approved diffs had 50x fewer incidents and 3x fewer reverts, and reviews got 35% faster. We built a source-available version of this. A few things we learned: \- One signal isn't enough. We combine an LLM reading the diff, a code graph that measures blast radius (callers, DB writes, cross-package impact), and an ML model trained on the team's own incidents and reverts. \- Per-PR scores are too coarse. One PR can mix a 1-line auth change with 300 lines of CSS, so we score each hunk. \- Trust has to be earned. You replay a policy on past PRs before turning it on, and every false approval feeds back into the model. It ends up auto-approving around 30-50% of PRs. It's called LiveReview, and it's free and self-hostable: [producthunt.com/products/livereview](https://www.producthunt.com/products/livereview) Would your team trust a tool to auto-approve low-risk PRs? What would you need to see first? PS: I’m also behind [git-lrc](https://github.com/HexmosTech/git-lrc), #3 on PH, 1.5k+ GitHub stars
document fetched 2026-11-08 00:22 · canonical https://reddit.com/r/developersIndia/comments/1xq31o6/what_we_learned_trying_to_copy_metas_radar