Logic systems are a good place where the library first approach tends to work well: you keep a small, boring core (the inference engine), and let people extend it with predicates, rules/operators, and domain packages.

I’ve been experimenting with a small defeasible-logic core (argumentation semantics + priorities + conflict detection) that stays stable, while most of the real meaning lives in extension points. https://github.com/canto-lang/canto-lang