This is something I’ve needed for my own research! I’m going to warn you that this part of the interface is going to be changing soon as part of a larger change to the driver logic and the optimizer.
Here’s what you can do for now:
Call psi4.optimize(your-function-here, dertype=1, molecule=your-molecule-here). The signature Psi expects of your function is shown here. gradient is Psi’s own gradient method, which is passed as the first argument for historical reasons. Your function must return a Psi4 Matrix object as the first argument and a second argument. Although Psi expects this to be a wavefunction, nothing bad will happen if you have something else as your second argument, as long as you don’t pass return_wfn=True to your psi4.optimize call.
In your case, your gradient function will probably be a regular Psi4 gradient call, a call to your dispersion function, and then returning the sum of the two.