alibaba/open-code-review tops today’s ranking. It sends code changes through deterministic checks and an LLM agent, then places each review comment on the exact line it refers to, with built-in rules for common bug classes like null pointer errors and SQL injection. If your team’s reviews tend to end in vague summaries, this is the opposite style: every suggestion carries an address you can jump to and check.
Two entries attack the cost of running models from different angles. JustVugg/colibri is a dependency-free inference engine written in pure C that streams MoE expert weights from disk, so a frontier mixture-of-experts model can run on hardware you already own. bojieli/ai-infra-book takes the teaching route: an open manuscript that derives LLM inference and training design from hardware capacity, bandwidth, and model architecture, with a freshly typeset PDF and companion experiments in the repo.
If you work with Tailwind, shadcn-ui/lint turns your design system conventions into rules that coding agents verify as they edit, and its errors suggest the accepted pattern instead of just saying no. For a lighter find, ayghri/i-have-adhd is a plain-text skill file that makes assistants put the answer first, number their steps, and close with one concrete next step.