MCMC and the inverse-Wishart distribution

I'm playing with a hierarchical Bayesian setup in which theta is a vector of model parameters with a Gaussian prior distribution. S is the covariance of that Gaussian (let the mean be zero for now) and has an inverse-Wishart prior. I cannot directly sample from theta and am using the metropolis algorithm to do so. How would I obtain samples of S that are conditioned on the data while I only have a single vector theta available after each metropolis step. Does one normally use all the vectors that have been sampled already to obtain samples for S? Is it possible at all?

Implementations that assume a Gaussian prior with spherical covariance I*sigma^2 seem to use the variance within the vector theta itself to compute the conditional inverse-gamma distribution. In my situation this is obviously not possible as I cannot compute the covariance matrix over a single vector.

EDIT: I realize that my question was a bit vague and I think I understand things a bit better after some thinking so I will give a bit more detail. I'm using metropolis sampling to sample vectors of parameters theta conditioned on the data (p(theta|X)) using a Gaussian prior. So the likelihood I evaluate at every sampling step is p(X|theta)p(theta|mu, S) (actually I'm not 100% sure I'm doing the right thing by including the prior in the sampling like that...). Now I want to place a prior on S as well in the form of an inverse-Wishart distribution. Based on other setups I've seen, it should be possible to use Gibbs sampling for this where given some value of theta I sample a value of S from the inverse-Wishart and use this to find a new value of theta using metropolis. The problem is of course that the covariance matrix cannot be determined from just theta but requires p(theta|X) itself.

These are the things I'm not sure about (In order of personal importance...):

-Does it make sense to run metropolis until convergence, compute the sample covariance, sample S from the IW distribution and run metropolis again using this new value of S? This would be some sort of metropolis within Gibbs with extremely sparse Gibbs updates.

-Is there a more efficient alternative to the above?

This one is more general:

-I know S can be integrated out replacing the Gaussian/inverse-Wishart combination by a multivariate Students T distribution but I didn't consider doing this yet because I'm actually interested in the distribution over S as well. Is this the reason why Gibbs sampling on priors is used in general or is it only done when it is not possible to integrate their parameters out?