A push-down automaton with two stacks which is equivalent to a linear-bounded automaton

Question

It is known that a PDA with two stacks is equivalent to a TM.

On the other hand a PDA with one stack is capable to recognise only context-free languages.

Hence there is a kind of a gap between the class of PDA with one stack and the class of PDA with two stacks which should be capable to recognise only context-sensitive languages.

I feel like it should be already an examined question, but I couldn't find an answer: what restrictions should we apply to a PDA with two stacks in order to make it equivalent to a Linear-bounded automaton?