bug#58839: [Patch] Re: bug#58839: 29.0.50; project-kill-buffer fails when Eglot is running

[Philip Kaludercic]

João Távora <joaotavora@gmail.com> writes:

> Dmitry Gutov <dgutov@yandex.ru> writes:
>
>> Anyway, if we do decide to flip the switch, it should be through
>> project-kill-buffer-conditions, so the user can make a different
>> choice through customization.
>
> project-kill-buffer-conditions doesn't work, I've tried it, it has this
> fundamental-mode thing there that makes it impossible.  Supposedly it is
> there to serve some purpose that no-one seems to be able to find a
> argumentative basis for.

The condition can be written entirely, if we are willing to accept a
breaking change.  In the case of `project-kill-buffer', this ought to be
acceptable if fewer buffers are killed.

> It's quite clear that _some_ non-file-visiting buffers can be considered
> as belonging to a project's working set.  But it's very very easy to
> come up with many that cannot be considered so.

I have to admit that I am more and more inclined to make the list a
opt-in thing, where  we explicitly mark those major modes that are tied
to a project.

> Because "killing buffers" is a destructive operation, being greedy here
> is a really bad design decision, as it catches an arbitrary number of
> unsuspecting extensions off-guard, which have been using earmuffed
> buffers for many years.  
>
> All in all, it's like you're making a gun that only backfires 5% of the
> time.
>
> In the little time I've used this feature since the start of this
> discussion I have discovered it backfires no small number of occasions:
> Eglot, CIDER, *scratch*, *ielm*, *sly-scratch*, *Completions*,...  Heck
> even *ibuffer* itself is targeted by this.
>
> Project-kill-buffers is off. Its intention pretty useful, but its
> implementation is a blunder.  The root cause is this overgreedy
> project-buffers.  When "killing a project" the echo area asks me if I
> want to kill a number of buffers that I didn't even know I had, because
> of hidden buffers.  This cannot be logical and the only way the
> "argument can be made both ways" is out of stubborness.
>
> JSONRPC's buffers are hidden implementation details: the argument that
> they are somehow under the responsibility of project.el just because it
> can see them through (buffer-list) is blind tiranny.
>
> The mini-languages invented in project-kill-buffers-conditions and
> project-ignore-buffer-conditions are abominations.  If project-buffers
> just been conservatively designed, we'd need nothing more than the
> existing hooks for the exceptions.  *earmuffed* buffers interested in
> opting in could declare if it belonged or not in one line.  

What existing hooks?

>                                                             Just like
>
> diff --git a/lisp/vc/vc-dispatcher.el b/lisp/vc/vc-dispatcher.el
> index dc3ed52650..718bebc7cd 100644
> --- a/lisp/vc/vc-dispatcher.el
> +++ b/lisp/vc/vc-dispatcher.el
> @@ -179,6 +179,7 @@ vc-setup-buffer
>    (let ((camefrom (current-buffer))
>       (olddir default-directory))
>      (set-buffer (get-buffer-create buf))
> +    (setq-local project-owned t)
>      (let ((oldproc (get-buffer-process (current-buffer))))
>        ;; If we wanted to wait for oldproc to finish before doing
>        ;; something, we'd have used vc-eval-after.
>
> To name one.  The above is just the converse of the solution proposed by
> Philip before.

I would be fine with this in principle, my only worry is backwards
compatibility for those who use project.el from ELPA.

> Anyway, I've now suggested and presented 2 actually tested, actually
> working patches to project.el.  I don't have anything more to add.
>
> João