I have been developing a custom estimator class that follows scikit-learn's estimator API. Up to scikit-learn version 1.5.2, this was straightforward because the estimator checks did not enforce strict type checks on tags, so I didn't need to import scikit-learn at all, my custom estimator would simply implement the required interface.

However, starting with scikit-learn version 1.6, a new requirement was introduced: the estimator must implement a __sklearn_tags__() method that returns an instance of scikit-learn’s internal Tags class. The built-in estimator checks now assert that the returned object is exactly an instance of this Tags class, which forces me to subclass or at least import this class from scikit-learn. This creates a problem for projects that want to be compatible with the scikit-learn estimator API but do not want to include scikit-learn itself as a direct dependency, either to reduce install size or avoid version conflicts.

Is there an easy or recommended way to satisfy this new scikit-learn 1.6+ tagging API requirement without having to add scikit-learn itself as a dependency? For example, can I somehow provide the Tags class or a compatible substitute so my custom estimator passes the checks, but without fully depending on the scikit-learn package?