detecting drift

ghost began with an idea about decentralized UI libraries.

shadcn gives teams the code and lets them own what happens next. A product can solve a local problem without waiting for a central design systems team to accept it and release a new version. This also worked well for model-generated code because the relevant context stayed local.

The cost is that every copy evolves. One team changes the spacing. Another rebuilds a component around its workflow. A third finds a better pattern with no way to show the rest of the system. The copies stop moving together. By design, but that leads to the formation of the next problem to solve. Some movements are the library evolving organically.

ghost is meant to add that missing view.

drift as a signal

ghost currently scans implementation details and generates a fingerprint of a design system. The fingerprint measures palette, spacing, typography, surfaces, and architecture. A fleet view can compare systems, find clusters and outliers, and show how they move over time.

The point is not to force every copy back into alignment. A central library cannot anticipate every workflow, audience, platform, or constraint its consumers will meet. Local divergence gives the system room to learn. The gap is letting it happen invisibly.

ghost treats drift as a neutral signal that needs a recorded stance. The CLI can acknowledge a difference, adopt a new fingerprint as a baseline, or mark one dimension as diverging. Projects can declare a parent, and history files record how their fingerprints move.

This puts authority closer to the people making products. A product can move without permission, but other teams can see what changed and respond with intent. Agents should inspect these systems continuously and help people interpret their differences, rather than serve as features added to a human dashboard.

Much of that observatory exists. The social mechanism around it does not.