Jim,
I'm not so sure how you run your model wtih 3dDeconvolve. If you modeled the signal with regression analysis of totally four regressors (two events for each trial type), then a contrast of contrast ((RESPONSEyellow - CUEyellow) - (RESPONSEgreen - CUEgreen)) would simply assign two 1's and two -1's corresponding to the four events in your glt vector, just as Lukas suggested. If you are running deconvolution instead of regression, the same strategy should apply although you would have to consider the number of lags.
Let me know if I understand the situation correctly.
Gang