As I have it written up now, there are frames to exchange the PCI for a new preshared connection. But I'm wondering if we should leave this completely up to the application layer. This document could just describe the format of the data and how it's used. The primary scenario I'm interested right now (P2P between two clients of the same server) wouldn't even use this in-band method, since the server would have to proxy the request from one client to the other, using some application protocol.

Do you expect to document a protocol wherein both parties can confidentially establish a secret that the shared-server cannot interfere or eavesdrop on?

More globally, do you plan to cater to a scenario where the shared-server is not 100% trustworthy? Or is the only expectation here that if we do not trust the shared-server, that you'd have to use out-of-bound data to establish a secure-confidential channel?