bug#69056: 30.0.50; history-add-new-input and recursive minibuffers

[Eshel Yaron via "Bug reports for GNU Emacs, the Swiss army knife of text editors" <bug-gnu-emacs@gnu.org>]

Eli Zaretskii <eliz@gnu.org> writes:

>> Date: Sun, 11 Feb 2024 16:54:43 +0100
>> From:  Eshel Yaron via "Bug reports for GNU Emacs,
>>  the Swiss army knife of text editors" <bug-gnu-emacs@gnu.org>
>>
>>
>> 1. emacs -Q
>> 2. (setq enable-recursive-minibuffers t)
>> 3. M-y
>> 4. In the minibuffer (with the prompt "Yank from kill-ring: "),
>>    type M-x calendar RET (or any other command).
>> 5. M-x M-p
>>    Expected: "calendar" is inserted in the minibuffer.
>>    Observed: error saying "Beginning of history; no preceding item".
>>
>> The problem is that the minibuffer history of M-x isn't recorded when
>> you invoke M-x from within the minibuffer of read-from-kill-ring (M-y).
>> The reason is that read-from-kill-ring let binds history-add-new-input,
>> and that affects all recursive minibuffers as well, so no minibuffer
>> history is recorded until you exit the first (non-recursive) minibuffer.
>>
>> AFAICT This issue affects all uses history-add-new-input, unfortunately,
>> not only read-from-kill-ring, since it's always used via let-bindings.
>
> I'm not sure we should be interested in fixing this.  Recursive
> minibuffers are not supposed to start a completely new command loop
> unaffected by whatever was before it, so we shouldn't try.

I see that, but the problem, IMO, is that there's nothing telling you
that you're in this state of not recording minibuffer history.  You
likely won't know that you're using a command that let-binds
history-add-new-input when you enter a recursive minibuffer, and losing
all minibuffer history from commands you invoked in the recursive edit
may come as an unpleasant surprise.

> Even if this particular case is solved (which I'm not sure we can),
> there are a legion of other similar situations, where something
> let-bound by a command entering the minibuffer affects all the
> recursive minibuffers.  Let-binding in commands that prompt users is
> ubiquitous in Emacs.

Indeed, this issue is possibly broader.  Often the solution is to use
minibuffer-setup-hook to bind a variable buffer-locally in a minibuffer,
rather than let-binding it (affecting all recursive minibuffers).  For
history-add-new-input this is slightly trickier since read_minibuf
checks the value of this variable only after the minibuffer is exited.
I'm experimenting with a possible solution where we change read_minibuf
to grab the buffer-local value of this variable from the minibuffer, and
change all users of history-add-new-input to set it buffer-locally
instead of let-binding it.  Works pretty well, but it doesn't cover
third party code that uses this variable, naturally.

> It's easy enough to work around the problem: C-g (perhaps more than
> once), then start afresh.
>
> So I tend to close this as wontfix.

All right, fair enough.