off-topic / non-blocker:: imo this is an uninteresting use of "!" (although common). Users can always do :wall to save everything before restart. And :confirm restart to opt-in to confirm "unsaved changes".

We might want to use "!" to mean "don't restore state/session" or something like that.

Don't need to figure this out right now.

FUTURE: we might want to give more control over how the server is stopped. E.g.:

:restart +qall

(similar to :help +cmd). This also frees up "!" to be used for some other purpose.

#34788

This presumably sends the event to all UIs. But we should only send to the "current UI". See ex_detach for how to get the current UI.

Also, should the server explicitly detach all other UIs?

At most one UI can restart the server

Can the UI event reach multiple UIs? If so, how will each UI know whether to attempt restart?

A UI knows whether it is the one that started the server in the first place. Only that UI can stop the server and then restart it. Other UIs can't stop the server at all.

Will ui_client_event_restart be invoked on each UI? If so, we're depending on channel_close(ui_client_channel_id,...) to fail I guess?

But a non-embedding client does not know what arguments the server was started with, so it still can't restart the server...

Yeah that'd need to be handled... We could do something similar to v:argv by setting another global and using that for both the embedded case and the remote case

Why can't we just only send the restart UI event to the "current UI", like how :detach works?

1. How can a client restart the server if it doesn't know how the server was started?
2. If one client restarts the server, shouldn't the other clients have to know that they need to reattach?

How can a client restart the server if it doesn't know how the server was started?

As long as client+server are on the same machine, v:argv gives the info needed to restart, no? (Maybe also need to think about environment variables and file handles, but that can be a "revisit later" topic.)

If the server is "remote" and not restartable, we should show an error and cancel the :restart.

If one client restarts the server, shouldn't the other clients have to know that they need to reattach?

The current spec for this feature is that all other clients will just detach, not reattach. Reattaching all the clients is a "nice to have" that in practice may not be a common use-case.

Tracked in #34296