Following the idea of choosing the most agnostic function, would you suggest to use the smoothed FIR instead of the standard FIR?

Both options are possible, sFIR allows for more spatial smoothing.

Based on the function selected, could the default parameter (32) for HRF's length be ok as a standard value or does it need to be optimised?

This is measured in seconds, so it's automatically adjusted to the TR. If for some reason you expert your HRF to be longer than 32 seconds (or much shorter), you can change this.

I still have some doubt about the meaning of k. Could be the default value (k=2) a good choice?

It's the range over which we search for the optimal lag (in TRs) between a pseudo-event and the HRF onset. The default is ok, unless you have a very short TR (0.2s or so).

we preprocessed rs-fMRI data using CONN toolbox. Which nuisance covariates in your opinion should be insered in the batch?

 If CONN is used for denoising, simply skip the denoising step in rsHRF (again, the default opinion is to do nothing). The order of processing might have an influence https://royalsocietypublishing.org/doi/1...