bug#69584: 29.2.50; project-find-functions should have access to maybe-prompt

[Spencer Baugh]

Dmitry Gutov <dmitry@gutov.dev> writes:
>>> With regards to caching, for example, if some caller wanted to do so
>>> (some related discussion:
>>> https://github.com/joaotavora/breadcrumb/issues/18#issuecomment-1984615275),
>>> then would also need to take this parameter into account.
>> True, but it's already not correct to cache when maybe-prompt=t,
>> right?
>> Because the returned project may just be the one that the user choose
>> interactively at the prompt.
>
> It's... a good point, but so far the main exception was the
> "transient" project, for which one could make an exception somehow, or
> even cache without major downsides, as long as buffer stays the same
> and the cache interval is low.
>
> Disabling cache altogether with maybe-prompt=t might be a net
> negative, given that many users' interaction with project.el might be
> limited to commands that only do such invocations. But perhaps it's
> the price to pay for flexibility: as long as we're talking about
> external cache, it will be up to the callers to avoid caching where
> the results can be non-deterministic, such as after a prompt.

Hm, I'm slightly confused, isn't the problem more general than just the
transient project?  If I run (project-current t directory), and I get a
project back, I have no idea whether that project is actually for
DIRECTORY or not: if DIRECTORY is not in a project at all, the returned
project is instead some project selected by the user with
project-prompter.

>>> So the first thing I'd ask is whether you see a different way to
>>> implement the same features. I don't see the whole usage scenarios, so
>>> it's hard to judge.
>> Let me give some context about my concrete use case.
>> At Jane Street we have a code review system built on top of Emacs,
>> called FE.  A user opens a code review in a buffer in Emacs, and can see
>> that review's diff.  If they want to comment on the review, or build the
>> code or test it or anything like that, they need to also have a local
>> working copy of the code in the review.  The local working copy for each
>> review is kept separately, like git worktrees.
>> (There are some other details and screenshots in
>> https://blog.janestreet.com/putting-the-i-back-in-ide-towards-a-github-explorer/
>> but this should suffice)
>> So my use case then is this: when a user opens code review FE-123 in
>> a
>> buffer, they look at the diff and then decide they want to do something
>> in a working copy of the code.  Currently, to do that they run one of a
>> variety of internal commands which duplicate things like
>> project-find-file, but which are aware of whether or not there's a local
>> working copy, and operate the local working copy if any, and otherwise
>> prompt to create a local working copy and then error.
>> I'd like to replace those internal commands with just normal
>> project-find-file, and also allow other commands which use
>> project-current to determine the current project to just work.
>
> If you set up a project instance in a buffer-local way, would it even
> work correctly outside of that buffer?

Hm, I don't see why it wouldn't?  It's not really any different, again,
from project-prompter returning a project when DIRECTORY isn't a
project.  I'm intending for these functions to return a totally normal
vc project, to be clear - the only magic is in initially finding that vc
project, when default-directory isn't in that vc project.

> Would project history work fine? When you pick a project from recently
> visited, you basically just apply its last root directory and expect
> the project to be "found".

The project-root would be just be the normal directory that the project
actually is located in, in the filesystem.  And since it would be a
normal vc project, project-find-functions would return the same project
instance when run on its root.  So that would work fine too.

> I've read the article (thanks!), but I'm not sure yet of what would be
> the ergonomic savings in such scenario when instead of having a
> separate command to check out a branch and visit a file in it (perhaps
> bound to 'o f' inside the major mode map for the branches list's
> buffer), you call project-find-file right away. In the former scenario
> such command would make sure the branch is checked out, so its
> directory has proper contents, and then it could delegate to
> project-find-file inside said directory. And later visits (e.g. from
> project-switch-project) would work fine until the directory is
> deleted.

Consider project-vc-dir or project-dired.  The default-directory of
these directories is the project root, so if you want to operate on the
project, you can do that in these buffers.  And that's convenient and
good - you can do things like find-file or project-find-file or
whatever, because these buffers are conceptually "within" the project.

The branch overview is like project-vc-dir, but you can also open it
when there's no local working copy for a branch.

If there *is* a local working copy, the branch overview has a
default-directory in the project, so you can treat it like
project-vc-dir or project-dired.  This is the common case, this works
great.

If there isn't a local working copy, the branch overview has a
default-directory of "/" just because there's no sensible
default-directory for the buffer.  And if you open a branch overview and
you know there's no local working copy, you could run a command to
create a local working copy and only then start treating it like
project-vc-dir, running commands which operate on the project.

But, it's convenient to be able to ignore whether a given branch
overview has a local copy or not.  Indeed, there are heuristics which
pre-create local copies for branches you are likely to interact with,
e.g. branches you need to review code for.  So for normal development,
there will usually be a local working copy before you open the branch
overview, even without your intervention.  So you can get away with only
rarely explicitly creating one.

So when you open up a branch overview, you'll usually assume there's a
local copy, and so your first action will probably some command which
uses project-current.  But if there's no working copy, then you'll get
dropped to a prompt to choose a project, instead of (say) a
project-find-file prompt, which you might not immediately notice, which
is confusing, and you'll have to C-g out of it, and then run some other
command to create the working copy.  All that is a hassle.

A few other potential things I could do to solve that confusing
situation:

- My project-find-function could detect if it's running in a branch
  overview buffer without a local copy and immediately error, which
  stops project-current from running, so it can't prompt.

- I could make the branch overview buffer always have a
  default-directory of the location where the local copy *will* be
  created, even if it doesn't currently exist.  (All of the local
  working copies are created as subdirectories of one specific
  directory.)  Then my project-find-function could look at the
  default-directory string without touching the filesystem, detect that
  it's in the directory for projects managed by my package, and return a
  project instance with a project-root that doesn't actually exist, so
  then project-find-file will fail when it tries to list files for a
  nonexistent project.

I'm guessing both of those also have undesirable implications for the
project-current semantics, though?

>>>> Since adding a new argument to project-find-functions is hard, maybe we
>>>> could do this by introducing a new dynamic variable
>>>> project-find-functions-may-prompt which we let-bind?  Like:
>>>> diff --git a/lisp/progmodes/project.el b/lisp/progmodes/project.el
>>>> index c7c07c3d34c..3975182b88d 100644
>>>> --- a/lisp/progmodes/project.el
>>>> +++ b/lisp/progmodes/project.el
>>>> @@ -242,8 +242,9 @@ project-current
>>>>            (setq pr (cons 'transient directory))))
>>>>        pr))
>>>>    -(defun project--find-in-directory (dir)
>>>> -  (run-hook-with-args-until-success 'project-find-functions dir))
>>>> +(defun project--find-in-directory (dir &optional maybe-prompt)
>>>> +  (let ((project-find-functions-may-prompt maybe-prompt))
>>>> +    (run-hook-with-args-until-success 'project-find-functions dir)))
>>>>      (defvar project--within-roots-fallback nil)
>>>
>>> As far as the implementation goes, a dynamic variable can work. We
>>> could also try reusing an existing one: non-essential, and we'd set it
>>> to nil when maybe-prompt is non-nil.
>>>
>>> I wonder how it would interact with Tramp (ask for password in
>>> disconnected buffers?), but that seems to fall into the same general
>>> category as your other cases.
>> Nice idea, it does seem like we should probably already be binding
>> non-essential=t around project-find-functions when maybe-prompt is nil.
>> Since already TRAMP can cause prompting even when a programmer calls
>> project-current with maybe-prompt=nil.
>
> If this change will be enough to cover your scenario, let's go ahead
> and add the 'non-essential' binding. It does seem to make sense for
> Tramp, at least.

Yes, that completely covers my scenario.  (Putting aside whether my
scenario is a good idea :) )

So I would be happy with that.