terminusakivili/codev

CO-DEV is a lightweight system that keeps a human in control when AI is doing software development work. The idea is that an AI agent can keep producing code that looks reasonable step by step while slowly drifting away from what the person actually wanted, and by the time that drift is noticed it is no longer a small fix but a fundamental shape problem. CO-DEV stops the AI at chosen checkpoints and forces it to wait for a human to review the work before continuing. The author built this after trying to migrate an Android app to Windows with AI assistance and watching the result quietly go off track over the course of a day, even though documentation, architecture notes, and a roadmap were all in place. Their conclusion was that good instructions alone are not enough. What is needed is a review process that is cheap enough to actually use every time, but strict enough to catch drift before it compounds. CO-DEV positions itself as a governance layer rather than a tool that writes code itself. It tracks what the human wants, which piece of work is currently active, and when the AI must pause for approval, while separate tools handle the actual coding, testing, and building. A single state file at the root of a project records the current phase, the active checkpoint, and whether a human has approved, redirected, or rejected the latest batch of work. A set of rule files define behavior such as loading that state before doing anything, enforcing that no approval means no further work, and correcting several kinds of drift. A command line tool, built for PowerShell and working on Windows, macOS, and Linux, checks and updates this state file safely, using careful file handling so that approvals are not lost even if something goes wrong mid write. The project includes its own automated test suite covering these edge cases. CO-DEV ships as plugins for a few different AI coding tools, so it can be added to an existing project as a lightweight workflow rather than a full rewrite. It is aimed at developers who already use AI coding assistants and want a simple, repeatable way to keep those assistants honest to the original intent of a project.