I agree with Sander's reply, but want to add that you have to find such a way to circumvent your issue. Standard mechanism design applies to Bayesian games. In a Bayesian game, the action space is type-independent, and since the designer does not know the types and can only decide on outcomes contingent on the agents' actions, we need to specify when the designer chooses an outcome that is "not allowed."
For example, such a problem shows up in mechanism design with evidence. It is discussed in this working paper by Roland Strausz (see footnote 1, this paper is superseded by another paper called "Principled Mechanism Design with Evidence" that I couldn't find online). The paper I linked seems to take Sander's approach (see the utility function on page 8), but also discusses what the other papers taking other approaches do.