In response to the above question, Aki writes:

I recommend as an easy default option
real nu;
nu ~ gamma(2,0.1);

This was proposed and anlysed by Juárez and Steel (2010) (Model-based clustering of non-Gaussian panel data based on skew-t distributions. Journal of Business & Economic Statistics 28, 52–66.). Juárez and Steel compere this to Jeffreys prior and report that the difference is small. Simpson et al (2014) (arXiv:1403.4630) propose a theoretically well justified “penalised complexity (PC) prior”, which they show to have a good behavior for the degrees of freedom, too. PC prior might be the best choice, but requires numerical computation of the prior (which could computed in a grid and interpolated etc.). It would be feasible to implement it in Stan, but it would require some work. Unfortunately no-one has compared PC prior and this gamma prior directly, but based on the discussion with Daniel Simpson, although PC prior would be better this gamma(2,0.1) prior is not a bad choice. Thus, I would use it until someone implements the PC prior for degrees of freedom of the Student’s t in Stan.