FDR aside, doing both cluster- and voxel-wise inference isn't "fishing" per se, but it will result in inflated false positive rates.  Hence, once has to commit to a single inference strategy before seeing the data to be sure you can trust your P-values.

To be clear, cluster inference in FSL always uses a familywise error (FWE) correction--a 0.05-corrected significant cluster controls the chance of 1 or more false positive clusters anywhere in the brain.  Similarly, the standard FSL voxel-wise corrected significance controls for the chance of 1 or more false postive *voxels* anywhere in the brain.  FDR is only available for voxel-wise inference (it has been proposed for cluster inference, but FDR works best with many many tests and sparse signals, and there just aren't that many clusters in a typical image).

Hence your question suggested a combination of FWE cluster-wise inference and voxel-wise FDR inference, which isn't so much apples and oranges, as apples and turnips.

Hope this helps!

-Tom