Compute t/f representation with variable-length epochs?

In many studies we conduct, people solve simple arithmetic problems. I am interested in EEG activity during the problem solving phase, which differs in length (1) from epoch to epoch and (2) from person to person.

These variable-length epochs are problematic, because I cannot simply compute a t/f representation with standard MNE functionality. Since I am mainly looking at oscillatory activity in specific frequency bands, my current workaround is to average band power over time within each epoch so that only one value per epoch remains. Of course, this means that I completely lose any time information, which is why this is a really crude approach.

EEGLAB has an option to "timewarp" individual epochs (https://github.com/eeglabdevelopers/eeglab/blob/develop/functions/timefreqfunc/newtimef.m#L144). I haven't really looked at the code so I'm not exactly sure how this works, but it seems to adequately address my problem. Does MNE have something similar? What do others do with such kind of data? I haven't seen any studies using the same variable-length epochs structure out there, and it seems like people just use some (arbitrary) fixed length cutoff. However, this is not ideal because I would like to make sure that only task-relevant brain activity gets into the calculation, so I don't want to cut off epochs at a fixed latency. Instead, I want my epochs to only contain EEG activity related to the problem solving process (i.e. from problem presentation up to a response).

Any ideas or sugggestions?

BTW, this is not about creating an object to hold variable-length epochs (this has been discussed in #3533).

Start by reviewing MNE's standard time-frequency functionality and the issue's cited EEGLAB newtimef.m timewarp implementation. Compare how each handles epochs spanning from problem presentation to response, then define the expected time-frequency output and validation cases before considering implementation.