Hi Jim,
That's actually what I originally did (using -Iresp with 3dDeconvolve), but it was later suggested to me that maybe this wasn't such a good idea, in that the dumped out responses spill over into other modeled task events, although their timing is jittered. Of course, this would be a problem with mapfile averaging too.
When I flirted around in 3dDeconvolve with different values (i.e. durations) for the -stim_maxlag, the shape and peak of the -iresp-derived HRF dumps was very sensitive to the stim_maxlag setting- instead of just giving me more data at later time points, so it leaves me wondering what the "truth" is.
Maybe I can't handle the truth!