A way to get "corfu-preview-current 'insert" without completion

[Haxxflaxx]

I generally like the way the 'insert' option confirms the preview on further input, and it works really well for cases where the language or setup doesn't involve template expansions.
The issue I find is when it is paired with eglot and yasnippet. From time to time I find myself typing something, cycle through the list, then realize that I want something else and start hitting backspace.
The result is something like the following:

SomeTypedText.
== hit M-DEL ==>
SomeTypedText.Previewed, arg2, arg3)

When doing any input it will complete the currently selected item triggering the yasnippet expansion. But what I really want it that it just inserts the previewed text.

Is there any way to get this behavior?
For now now I've switched to corfu-preview-current t but that has other jarring consequences where what I see isn't what is actually there.

[minad]

When doing any input it will complete the currently selected item triggering the yasnippet expansion. But what I really want it that it just inserts the previewed text.

It sounds as if the :exit-function of the respective Capf is not behaving correctly. When template expansion is not desired, Corfu exits with the state exact. You can check this via cape-capf-debug.

[Haxxflaxx]

What I'm looking for is a way to only have explicit completions and consistent behavior.

With corfu-preselect 'prompt I can expand the common parts of the string, but if there is only a single candidate it will trigger completion and exit, which is a behavior that can't be turned off.
I do find it a bit weird that we do completions from the expand function, but I can see how it is useful in certain cases. (For instance when previewing as this is my main way of triggering completions.)

For the preview I've noticed that if corfu--prepare calls corfu--insert with status nil, I get the behavior I would expect. But I can see how that would only be true for my usecase and might brake something else, say "tab and go".

Would it be possible to put this behind a corfu-on-exact-match option? So that insert would have the behavior we have today and show and quit would just expand the text without calling the exit-function?

[Haxxflaxx]

After spending some time digging into the implementations of corfu and eglot, I think this is both more clear and confusing at the same time.
For a start eglot handles finished and exact the same way. At least to my understanding. And I guess it doesn't handle the sole case at all.

 ‘:exit-function’: Function to run after completion is performed.

 ‘finished’ - text is now complete
 ‘sole’     - text cannot be further completed but
              completion is not finished
 ‘exact’    - text is a valid completion but may be further
              completed.

It sounds to me that based on the completion-extra-properties documentation that exact should indeed just insert the text and wait for further input, and only complete if finished is sent. So this should probably be reported to that project.

Another, similar thought I had was that with my newfound (possibly erroneous) understanding of how completions work in emacs; I find it a bit strange that corfu-expand exits with finished. I would assume it would use exact and sole and then call corfu-complete as it does when something is selected.

[Haxxflaxx]

After slightly more thought about it; maybe sole doesn't make much sense as corfu should just continue as normal in the case where it can't be completed without further input.

[jdtsmith]

I admit I'm fuzzy on the states used for :exit-function, but eglot only knows what the server tells it, and whether completion on the server side is "finished" or "exact" is not one of those things. You can see the full specs here.

[Haxxflaxx]

I have a hard time mapping the different options for the exit-function to what eglot should do, and maybe this is the reason why "finished" and "exact" is handled the same.
There seems like there should exist some partial complete and continue option, but maybe that's not on eglot to provide?

[jdtsmith]

Some servers provide incomplete completions, and can return new never-before-seen completions to the list after you add to the in-buffer text. That's probably why eglot is conservative on the matter. It's also why it tries to fully override the completion style and even UI, and use company configured in a way that every edit in the buffer calls out to the server anew (though there is caching).