ivy-switch-buffer: filtering does not preserve virtual buffer ordering

[Wilfred]

I use ivy-switch-buffer with ivy-use-virtual-buffers and it generally does exactly what I want.

However, if I kill a buffer, then try to switch to buffer with the same name, I accidentally reopen the previous file:

In this picture, I've just killed a lib.rs buffer. I have two other files called lib.rs open. However, typing C-x b lib.rs RET reopens the lib.rs file I just closed.

Is it possible to reorder the results to prevent this?

[abo-abo]

It looks like the case here is the perfect match logic: when you type in the exact candidate name, it will be selected no matter where in the list it is.

I think if you type in only lib, the virtual one will be lower in the list.

[Wilfred]

This doesn't seem to be the case:

What would you expect to happen in this case?

[abo-abo]

What would you expect to happen in this case?

The virtual buffers are always below for me. What settings are you using?

[Wilfred]

I think my only relevant setting is this one:

(setq ivy-re-builders-alist '((t . ivy--regex-ignore-order)))

I have set four other ivy related settings: https://github.com/Wilfred/.emacs.d/blob/40cc6ebc474f04d14762f3898677589f17781373/user-lisp/minibuffer-completion-customisations.el#L47-L70

I'm not seeing virtual buffers at the bottom. In one case (note the case difference), I am even seeing the virtual buffers in the middle:

Is there code to ensure that virtual buffers are last? Could you point me at it?

[abo-abo]

See ivy--buffer-list. It has an append of current buffers and virtual buffers.

[Wilfred]

OK, so ivy--buffer-list is only called once, and ordering is correct at that point. It's only when I start typing characters to narrow down the options that the order is not preserved.

This seems odd because my understanding of ivy-sort-functions-alist is that we shouldn't change the ordering for buffers.

Edit: strike that, ivy-sort-functions-alist seems to be for the initial sort. I think the function I'm interested in is ivy--filter.

[Wilfred]

I think the problem might be this line in ivy--filter:

(setq ivy--old-cands (ivy--sort name cands))

This seems to be putting package.json first in the list of candidates, even though it's a virtual buffer. Changing this line to:

(setq ivy--old-cands cands)

Fixes the issue. Is this a reasonable patch? Assuming the data is sorted anyway, sorting when filtering seems redundant.

[abo-abo]

Likely the issue with ivy--regex-fuzzy and flx: flx doesn't care if a buffer is virtual or not when sorting.

[Wilfred]

I have a stupid question: why does ivy--filter sort at all?

Edit: oh, I see. ivy-sort-function-buffer prefers prefix matches first.

For what it's worth, I don't have flx installed, so I don't end up using ivy--flx-sort. It's ivy-sort-function-buffer that seems to not be preserving virtual buffer ordering.

[Wilfred]

Aha, I think I see the problem. I have the following in my init.el:

(setq uniquify-buffer-name-style 'post-forward-angle-brackets)

This means that the virtual buffer name is just package.json, whereas the existing buffers have prefixes: eslint/package.json and so on. Since ivy-sort-function-buffer prefers a prefix match, it is prioritising the virtual buffer name over the uniquified buffer name.

[abo-abo]

Can you make a generic fix for uniquify? Patches welcome.
Otherwise, you can adjust your ivy-sort-matches-functions-alist setting.

[Wilfred]

I'm not sure what the ideal behaviour is, really. Uniquify appending text seems to be in conflict with preferring prefix matches.

How would you feel about a patch to ivy-sort-function-buffer that ensures virtual buffers are always last?

Alternatively, I'm very happy just to set ivy-sort-matches-functions-alist if you think this is all too much of a corner case to be worth handling.

[abo-abo]

How would you feel about a patch to ivy-sort-function-buffer that ensures virtual buffers are always last?

Sounds good.