okikijesutech/cambium

Cambium is a command line tool that watches a codebase over time and warns when a single file has taken on too many unrelated jobs, before that mess becomes so tangled that fixing it requires a full rewrite instead of a small refactor. The idea behind it is that AI coding assistants are good at producing working code in the moment, but do not track how a project is drifting over many small commits, so a file that starts focused can slowly pick up extra responsibilities until nobody planned for it to become that complicated. Cambium works on any codebase regardless of whether AI wrote the code, a human did, or both. It scans every file and measures its size, complexity, how many other files depend on it, and how often it keeps getting changed in git history, then ranks files by combined risk. For the files that rank highest, it can call a local or hosted language model to explain what the file's original job probably was versus what it is actually doing now, and to suggest a specific way to split it apart. It also double checks its own suggestions against the file's real exported code and skips proposals that do not hold up, or files too small to be worth splitting. You install it globally through npm or run it directly with npx, and the main command is simply cambium. A quick scan command runs in seconds with no AI model involved. A baseline command saves a snapshot of the whole repo so later scans can show how each file changed since that point. A drift command is the one that calls a language model, and by default it uses a free local model through Ollama, though it can use Anthropic's Claude instead if you provide an API key, which then costs money per call. There is also a single-file check command fast enough to run in continuous integration, and a command that installs a git pre-commit hook so Cambium automatically looks at newly written code before it gets committed, either just warning or actually blocking the commit depending on the mode chosen. The author reports testing it against a real 126-file production codebase, where it correctly found the two most structurally troubled files without prior knowledge of that project and explained why each one was a problem.