Hacker News discussion: Sober: Local-first code reviewer for agentic PR (deterministic and model review)

A new open-source tool called Sober (v0.8.5-beta) has been released, designed to provide deterministic, local-first code review for agent-generated pull requests without cloud dependency. It runs entirely on the user's machine, uses their own AI models, and never auto-merges—keeping merge authority with humans. The latest version includes a 'Defense Module' that detects recently observed supply-chain attacks targeting AI coding agents through configuration files. Sober uses tree-sitter to parse over 320 grammars, catches secrets and unsafe patterns without AI, and stores all review evidence locally for inspection. It is offered as a free community or pay-what-you-want enterprise OCI image, with no subscription required.