Interesting to read someone else ideas about that topic, which I though myself quite a lot about. The basic building block of a better desktop OS is IMHO – and as the OP wrote – a communication contract between capabilities and the glue (a.k.a apps). I don't think we would need that many capability-services to be able to build something useful (it doesn't even need to be efficient at first). For the start it might be enough to wrap existing tools and expose them and see if things work or not.

Maybe by starting to build command-line apps and see how good the idea works (cross-platform would be nice). I guess that the resulting system would have some similarities with RxJava, which allows to compose things together (get asynchronously A & B, then build C and send it to D if it contains not Foo).

If an app would talk to a data-service it would no longer have to know where the data is coming from or how it got there. This would allow to build a whole new kind of abstractions, e.g. data could be stored in the cloud and only downloaded to a local cache when frequently used, just to be later synced back to the cloud transparently (maybe even ahead of time because a local AI learned your usage patterns). I know that you can have such sync-things today, they are just complicated to setup, or cost a lot of money, or work only for specific things/applications, also they are often not accessible to normal users.

Knowing how to interact with the command-line gives advanced users superpowers. I think it is time to give those superpowers to normal users too. And no, learning how to use the command-line is not the way to go ;-)

A capability-services based OS could even come with a quite interesting monetization strategy by selling extra capabilities, like storage, async-computation or AI services, beside of selling applications.