Changing the regression models in lmtree terminal nodes

Question

When building a tree with lmtree, we get a tree with regression object in each terminal node. The problem is, that the model in each terminal node might not always have only significant predictors included (i.e. the p-value for some is quite large).

What I would like is to retrain the model or perform variable selection on each terminal node separately, as the variables that might be significant in one node might not be in another. Alternatively, I might just want to fit a different model, based on the data distribution in that node.

In the end, I would still like to have a party object tree, from which I could predict as normal, but with custom terminal models.

How can I do this?