Here are some features that I have discussed with Warren in the in the past:

1) Vector graphics output (preferably in Adobe Illustrator format, SVG, postscript, or PDF format).  I tend to make all of my posters and publication figures in Adobe illustrator (due to its scalability) and images generated from PyMOL seem to always slow me down especially when I need to go back and ray trace an image or sets of images over again since they were smaller than I had expected.  I am no expert but I feel like there should be an "easy" way to achieve this.  For example, the output for POV-RAY generated by PyMOL is essentially a list of primitive objects and how they should be drawn and I feel like it could be converted into some sort of postscript-like information which could be interpreted in programs like Adobe Illustrator or other vector-based programs. 

2) In line with the first request, I stumbled across the fact that VMD now has a really cool plug-in for PDF files where you can open a PDF file that contains a figure and actually MOVE the 3D-image (or a molecule) within the PDF file!  This could really change the way we interpret molecular structures in the future!  More information can be found here:

3) Being a molecular dynamics guy, I've spent countless hours writing python scripts to read in MD trajectory files.  Though successful, the hurdle lies within the fact that the process is extremely slow.  Essentially, the python script opens the trajectory file, reads one MD frame, duplicates a state in PyMOL, copies over the coordinates from the MD frame over to the new state via the alter command (or alter_state?), and repeat for as many states as needed.  The problem is really the last part of using the alter command which is extremely slow.  I understand that certain versions of PyMOL contains the ability to read trajectories (which I currently use) but I also understand that not everyone has access to the same resources so it would be nice to have a faster way to access/load/modify the coordinates for different states that can be handled efficiently via python scripting (and not C).  Alternatively, one may want to load several PDB files (or NMR models) into a single object if you close PyMOL (on purpose or accidentlally) then you will be forced to load the files again.  This could be tedious.  Instead, if you could save the states as an MD trajectory then all you would have to do next time is load the corresponding PDB file followed by the trajectory and that's it.  Now, I have written the python code for this (to save states to a trajectory file) but then I run into the same issue of not being able to read in a trajectory file via python (the hard-coded method using C to read trajectories is much much faster).  Thus, the ability to load trajectories is always a plus.  In addition, having the ability to load trajectories not only makes loading multiple files very fast and efficient but it is more favorable than opening a movie/video generated by PyMOL since with the trajectory, you can still move the molecule while the movie is played as is without further interaction.

4) Going along the same MD theme, PyMOL seems to not be able to handle certain representations (i.e. cartoon, spheres, surface) of large biomolecules (160,000+ atoms) when a large number of states is loaded (20,000+ states).  As much as I love PyMOL and as much as I've convinced other members of my lab to use it, it's hard to compete with VMD in this regard.  I would be more than willing to help provide examples for this (and I had conversed with Warren about this in the past and is also where I was made aware of the "defer_builds_mode" setting).  

5) It would be nice to have a GUI for drawing CGO objects.

Hopefully, this will get the ball rolling for more ideas!