Context: in my model a latent variable L is described by predictors x_i, nonlinear function f, and parameters \theta: L=f(x_1, x_2,\dots;\theta). L in turn affects two observables A and B, e.g., A \sim \mathcal{N}(L, \sigma) and B=\text{ordered_logstic}(L, \text{cut}). I don’t think brms supports this kind of model through formula (does it?), so I simply write the model for A in brms and add stan code related to B using stanvars, only that now after fitting I need extract params such as \text{cut}.

Indeed, brms does not support such a model yet. I don’t think that methods such a posterior_epred and friends are easy to enable for models that include manual stanvars parameters affecting predictions.