Alibaba's Open Code Reviewer Mixes Rules With Agents on Purpose

Alibaba just open-sourced open-code-review, and it is trending because it refuses to pick a side in the dumbest debate in the space. Should code review be deterministic linting, or should it be an LLM that reads your diff? open-code-review says both, wired together. Deterministic pipelines handle the rules and static checks that a model would hallucinate around, and LLM agents handle the judgment calls a linter can't make. It is written in Go, it is self-hostable, and it talks to both OpenAI and Anthropic models.

The reason this matters is that everyone who has shipped a pure-LLM reviewer has hit the same wall. It is confidently wrong, it misses obvious rule violations, and it invents problems that aren't there. A pure static analyzer has the opposite failure, it is noisy and literal and can't tell you whether the change actually makes sense. Bolting them together, deterministic gates first and agent reasoning on top, is the pragmatic answer people kept building privately.

The signal here is not the tool, it is who shipped it. When a vendor the size of Alibaba open-sources the hybrid pattern instead of a pure-agent demo, that is the pattern quietly winning. The interesting agent architectures in 2026 are not the ones that hand everything to the model, they are the ones that know exactly which parts to keep dumb and deterministic. https://github.com/alibaba/open-code-review