Handle completion of % and # completion correctly.

[iteratee]

The current behavior if you try to complete :e %:p:h or something similar replaces only the h with the path, which isn't the desired behavior, because then nvim tries to open the wrong filename. Detect this case and produce edits that will replace the entire % chain with modifiers.

[iteratee]

My solution may not be ideal. I would be happy to re-write it given some pointers. I'm new to lua and vim plugins.

[aldevv]

any plans on merging this? this problem is been a pain for a while @hrsh7th

[fent]

please :(

[iteratee]

I'll try and rebase my current solution and push it up.

[0e4ef622]

I am installing from this PR using the commit hash, but it would be nice to have this merged.

[FoamScience]

In visual mode, completion doesn't keep track of typed chars, so generates wierdy items.
Fixed and rebased in iteratee#1 in case that takes some time to get merged, you can install from that PR instead

[liskin]

I tried this "fix" and it's not working well for me, it replaces the space before the % as well. Inspired by #110 (which fixes wildcards and %/# and breaks almost everything else), I believe this can be improved by vastly simplifying the code to just:

        if is_magic_file then
          item.insertText = item.label
          item.label = arglead .. ' → ' .. item.label
        end

All the complicated range stuff can be dropped.

---

Still, this is just a workaround, not an actual fix. Expansion of file wildcards is still broken, as are many other usecases where the fixed_input heuristics don't find the same offset as nvim's completion does internally.

The root cause of why this can't be fixed properly here is that getcompletion() doesn't expose the column where the returned matches are meant to be inserted. The person who added the ability to get cmdline completion from getcompletion() deals with this in their own funny way: https://github.com/Shougo/ddc-source-cmdline/blob/c03488e9239cba6c3bfbc000b17677dd47278e5a/denops/%40ddc-sources/cmdline.ts#L72-L86
But this can never work reliably enough. The correct way to deal with this is to extend getcompletion() to have a similar interface to omnifunc: https://github.com/vim/vim/blob/3d5065fc7553c8de3d0555c3085165bdd724663b/runtime/doc/insert.txt#L1105-L1122, so we can get the actual column where the matches are meant to be inserted without guessing.

Personally, I gave up and went back to nvim's builtin cmdline completion. nvim-cmp + cmp-cmdline didn't provide enough added value for me to sink additional hours into submitting a PR to neovim and then running a patched version until it makes it into a release and that release makes it to Debian. If someone else wants to pursue fixing this proper, feel free to get in touch for guidance.