If one has a finite dimension simple Lie algebra, one can easily calculate that taking the centralizer of a torus (or toral subalgebra), that is, summing the weight spaces that lie in some proper subspace of the dual Cartan, always gives a finite dimensional reductive Lie algebra; actually almost semi-simple, except that there is some central toral subalgebra (maybe bigger than the original torus).
For an affine Lie algebra, the picture is the same if you pick a subspace on which $(-,-)$ is positive definite, you get again something finite dimensional and reductive. If you pick a subspace containing $\delta$, however, things are a bit messier. Now you have an affine Lie algebra plus a central piece in every single one of the imaginary weight spaces; thus you have an infinite dimensional center. I would kind of like to think of this as a Borcherds algebra, where I add infinitely many rows and columns of zeros to the affine Cartan matrix.
For a hyperbolic Kac-Moody algebra, the picture is even worse; I still get a hyperbolic KM algebra attached to the root subsystem living in the subspace, but the "extra stuff" in imaginary weight spaces is much more complicated and not central anymore. This seems very complicated, but I hope to get some kind of handle on it:
Is it true that the centralizer of torus in a hyperbolic Kac-Moody algebra is a Borcherds algebra (necessarily infinite rank in most interesting cases)? Is there some nice description of its Cartan matrix?