Re: add-zle-hook-widget and multiple hooks

[Daniel Shahaf <d.s@xxxxxxxxxxxxxxxxxx>]

Mikael Magnusson wrote on Mon, 08 Jun 2020 19:52 +0200:
> On 6/8/20, Daniel Shahaf <d.s@xxxxxxxxxxxxxxxxxx> wrote:
> > [moving Eric to bcc]
> >
> > Mikael Magnusson wrote on Sat, 06 Jun 2020 13:58 +0200:  
> >> On 6/6/20, Daniel Shahaf <d.s@xxxxxxxxxxxxxxxxxx> wrote:  
> >> > When two or more zle-line-pre-redraw hooks are registered using
> >> > add-zle-hook-widget, the value of $LASTWIDGET when the each hook is
> >> > called is the name of the former hook:
> >> >
> >> >     [Eric Freese wrote in
> >> > https://github.com/zsh-users/zsh-autosuggestions/issues/529#issuecomment-632113840]
> >> >     $ zsh -df
> >> >     % autoload add-zle-hook-widget
> >> >     % f() {}
> >> >     % g() { zle -M "$(typeset -p LASTWIDGET)" }
> >> >     % add-zle-hook-widget line-pre-redraw f
> >> >     % add-zle-hook-widget line-pre-redraw g
> >> >     % x<CURSOR>
> >> >     typeset -r LASTWIDGET=f
> >> >
> >> > The issue here is that g would like to to know what widget was invoked
> >> > immediately before the redraw.  In the example, that'd be self-insert.
> >> >
> >> > I've attached two proofs of concept.  WDYT?
> >> >
> >> > I'll add docs, etc, once an approach is chosen.
> >> >
> >> > Cheers,
> >> >
> >> > Daniel
> >> >
> >> > P.S.  For the latter patch, note that «zle $widget -f» is distinct from
> >> > «zle -f».  
> >>
> >> I think the warning message when -f is not followed by nolast should
> >> be phrased in the same way it would if there were other valid flags,
> >> since it would have to change when more are added anyway (we will
> >> probably not want to enumerate all possible flags in this warning
> >> message).  
> >
> > Thanks for the review.
> >
> > If we add more flags that are too many to list, then we should change
> > the error message, yes; nevertheless, _right now_ only one flag is
> > supported, so the error message might as well say that.  That'd be
> > a feature, not a bug.  Compare:
> >
> > % ssh $foo svn info --show-item=dept
> > svn: E205000: 'dept' is not a valid value for --show-item; did you mean
> > 'depth'?
> > % ssh $bar svn info --show-item=dept
> > svn: E205000: 'dept' is not a valid value for --show-item
> >
> > When I'm on $bar and get that error message, I know not to bother
> > trying the correct spelling because it won't work.  Same here:  If 5.9
> > supports -f nolast, 5.10 supports -f somethingelse as well, and then
> > somebody tries -f somethingelse in 5.9, a generic error message will
> > be less helpful to them than a specific one.
> >  
> >> Also, I think rather than reusing the concept of the -f option, it
> >> would be better to use another flag (maybe -l for LASTWIDGET) which is
> >> analogous to the existing -w option:  
> >
> > _Why_ would that be better?  
> 
> (sorry if the following is a bit rambly)
> 
> Just seems more consistent to me to have -l and -w, rather than -w and
> -f nolast/unrelated/flags.

I can't parse the last word.  In any case, one of the reason I used the
same option letter is that both -f flags set bits of the same bitfield
type.  Even if right now there's no bit that both -f flags set,
I suspect one might be added in the future.  (For example, a use-case for
«zle -f nolast» doesn't seem inconceivable, even if I don't have one on
me.)

> The point of the generic -f is that there
> had to be some option to initiate that mode of modifying global(ish)
> state without actually calling a widget, and adding 4 options instead
> of just one option that did 4 very related things felt weird. In a
> sense -f is like -N there, it is its own mode of operation. Eg, you
> can say zle -f yank kill and both words are arguments for -f (in a
> sense).

Yes, «zle -f» is a subcommand (like «git foo»).  All zle subcommands are
named as option flags, except for the "invoke a widget" subcommand which
is anonymous.

> In this case you're only adding a single flag and that flag is related
> to an existing option, and it is (for now) just a very long way to
> spell -l. It doesn't start a separate mode of operation, just modifies
> the current operation slightly.

That's what option flags usually do.  For example, that's what the -l
flag to ls(1) does.

> It's not like we're short on letters for options in the zle widgetname
> -opt namespace either (only 4 are used, 2 of which are already capital
> letters).

The difference will come once we support both «-f nolast» and «-f foo».
Then each such foo we add will save an option letter, and the API will
be easier to remember due to using more meaningful names.

> I think if we added more flags to your -f scheme, the
> convention would by necessity be -f nolast -f unrelated, which is
> inconsistent with the zle -f flag.

They could also be «-f nolast:unrelated», just like the $PATH envvar in
relation to the $path array.  We could even teach the toplevel -f to
support this syntax alongside the existing one for consistency, or teach
it to ignore flags called "-f" in its positional arguments, or add a «-f
foo» value that's an alias of -w, etc..

Over here the jury's still out.

Cheers,

Daniel