>Unfortunately, much does depend on the particular task and chemical intuition; experimentation with the active space may be quite advisable (such as its gradual expansion). Bond breaking reactions would definitely require inclusion of the respective bonding and antibonding orbitals but more frontier orbitals may be relevant as well. To reduce the amount of "mess", one may request orbital localization and assemble the initial approximation to the active space from the localized orbitals. Have you seen this review: http://www.annualreviews.org/doi/abs/10.1146/annurev.physchem.49.1.233 ?
>I guess you are interested in the ground state. However, as Pavlo has pointed out, several states may come close, so you may need to select some reasonable value of NSTATE (not exceedingly large to spare computational time) and, if you really have a closely spaced ensemble of states to test the effects of state averaging via WSTATE. If the two states are of different symmetry and do not exhibit interaction via symmetry breaking, one may prefer to consider them separately; however, an averaged calculation would provide a more balanced description of their relative energy. One may also try to optimize more than one state using state tracking: the states can reorder depending on the geometry. To verify the adequacy of CASSCF, single point XMCQDPT2 tests can be helpful.