Completion list not accurate at first type?

[eric-hansen]

Contributing guidelines

• I have read CODE_OF_CONDUCT.md

Module(s)

mini.completion

Question

Not sure if this is a me-thing, so figured I would raise this as a discussion/Q&A first.

But I've noticed with mini.completion, the completion list is never accurate when I'm first typing for something. So I have to exit from insert to normal, go back in, and continue trying to complete what I'm looking for to get an accurate list.

It's sort of hard to explain, and not sure if this is an LSP server matter, but I struggle to lean towards that since it works accurately first-time around with nvim-cmp.

My mini.completion set up is pretty plain Jane:

require("mini.completion").setup({
    lsp_completion = {
        source_func = "omnifunc",
    },
})

I've also noticed this is only really happening for one LSP. But it's the same LSP I've used before when I was using NvChad. Which I've been migrating things away from because I got exhausted from dealing with the abstractions NvChad was doing.

[echasnovski]

The actual reproduction steps would be much help here. In particular: which LSP server, what file/project structure, what file content, which place to start typing and what to type.

But I've noticed with mini.completion, the completion list is never accurate when I'm first typing for something. So I have to exit from insert to normal, go back in, and continue trying to complete what I'm looking for to get an accurate list.

Does pressing <C-Space> help instead of switching modes?

[echasnovski]

Any and any. Not sure how familiar you are with PHP, but whether I'm trying to extend a previous class, use a trait or just try and reference a method on an object.

Yes, I am not really familiar with PHP and thus need a step-by-step guidance in order to reproduce the issue. Without actually reproducing it there are very limited ways to help. Here is one of the examples (the "Steps to reproduce" section) of what would help very much here.

Does pressing help instead of switching modes?

Yes, this does seem to resolve the issue from what I've tested thus far. Though I'm not on the computer currently with the bigger code base, I was able to reproduce the issue easily just by requiring Laravel as a depndency and try to create a new collection.

See https://asciinema.org/a/HahHnPNKqSZjbJ5HlILQzHOsy which I was able to capture. The initial completion doesn't offer Illuminate\Support\Collection, which I would expect to see. However, when I do <C-Space> as requested, it shows.

My guess would be that LSP server sets an isIncomplete field (meaning that it didn't return all possible candidates for some reason and request needs to be redone afterwards) which 'mini.completion' doesn't support at least currently. I plan to take a closer look in the upcoming 'mini.completion' updates, but making it work is somewhat problematic.

The reason here is because 'mini.completion' deliberately uses Neovim's built-in 'omnifunc'/'completefunc' mechanism to filter down candidates once it received response from LSP server(s). As far as I understand the LSP specification, respecting isIncomplete would require to somehow interject into that order of things, which doesn't look straightforward.

I'll take a closer look here to confirm this in case I'll be able to reproduce the behavior.

[eric-hansen]

Yes, I am not really familiar with PHP and thus need a step-by-step guidance in order to reproduce the issue. Without actually reproducing it there are very limited ways to help. Here is #1230 (comment) (the "Steps to reproduce" section) of what would help very much here.

I need to make some assumptions, namely you're using Ubuntu though the general premise applies to Mac or other Linux distros as well. But as that is what I use in WSL, I need to approach it from that:

• Install PHP & Composer: sudo apt install php-cli composer
• Install Node & Intelephense (free version should suffice): sudo apt install nodejs npm && npm install -g intelephense
• You should be able to run intelephense and get an error about input stream not set (expected): intelephense
• Create a folder (this will assume /tmp/phptest moving forward) and cd into it
• Run composer require illuminate/support
• Have mini.completion & lsp configured in Neovim, and configured to use intelephense (see lua config below)
• Create a new PHP file in Neovim:

<?php

class A {
  public function b()
  {
    // cursor here
  }
}

In the script above, remove the // line and type $var = new Collect, and you won't see the expected value of Illuminate\Support\Collection, but when you do <C-Space> the list will refresh and the expected item will exist.

Lua config (this can probably be modified but it's mimicking what I have with most fluff stripped out):

local config = {
    lsp = {
        ["intelephense"] = {},
    },
}

local lspconfig = require("lspconfig")

local on_attach = function(_, bufnr)
end

local common_capabilities = function(server)
    local capabilities = vim.lsp.protocol.make_client_capabilities()
    capabilities.textDocument.completion.completionItem.snippetSupport = false
    return capabilities
end

for server, opts in pairs(config.lsp) do
    local lsp_opts = {
        on_attach = on_attach,
        capabilities = common_capabilities(server),
    }

    lspconfig[server].setup(lsp_opts)
end

require("mini.completion").setup({
    lsp_completion = {
        source_func = "omnifunc",
    },
})

I'm not sure if there's anything else you would need, but this is what I can think of.

Follow up question

I was reading through the source but couldn't find an answer to this. Since the two-step result works for what I need, is there a way to force this to happen instead of whatever happens the first time, whether by language or globally?

[echasnovski]

Thanks, I'll give it a try in the coming days.

I was reading through the source but couldn't find an answer to this. Since the two-step result works for what I need, is there a way to force this to happen instead of whatever happens the first time, whether by language or globally?

If by "two-step" you mean the "get autocompletion, type something, press <C-Space>", then I don't think so. The whole premise of the isIncomplete (which is the most probable assumption at the moment) is that it is just the way LSP server responds at the particular context. So if autocompletion doesn't return all results, then forcing it in the same context with <C-Space> won't help.

[eric-hansen]

In that way though is there anything you can think of for me to try with the LSP?

I'll also see if I can just create a minimal repo for you to try as well to simplify most of this setup. Just can't guarantee when I'd be able to do so.

I'm fairly certain it does caching so I may also see if that's playing a part in this. Since as I mentioned it worked fine using nvim-cmp, so it's not like it shouldn't work as expected, there just may be something missing.

[echasnovski]

In that way though is there anything you can think of for me to try with the LSP?

Maybe if there is a limit on number of returned completion items as a configuration, try to increase it.

I'll also see if I can just create a minimal repo for you to try as well to simplify most of this setup. Just can't guarantee when I'd be able to do so.

Thanks! Installing intelephense through 'mason.nvim' is pretty straightforward, but getting an exact project/file/position to replicate needs domain knowledge.

[echasnovski]

@eric-hansen, I am afraid I can not reproduce with these steps. Here is the demo:

mini-completion-php-no-item.mp4

[echasnovski]

With the help of some logging, I am 99% sure that the problem here is that indeed LSP server returns isComplete = true. I can see that even on my machine. Just in my case the target Collection seems to be present initially, while in your demo it is not.

I am taking a look now into the best way to incorporate reaction to isIncomplete = true in 'mini.completion'.

[echasnovski]

@eric-hansen, would you mind checking the backlog-completion branch, please? It should:

• Automatically force re-trigger completion if the returned completion list is incomplete. Request is done immediately to reduce flickering.
• Force completion immediately if inserted key is a trigger character for LSP. This might change in the future, but currently it is the way reduce flickering in case of isIncomplete

Does this fix the issue you experience with not present items? I'd also appreciate hearing about any other "suspicious" new behavior, like extra flickering (of completion popup menu), etc.

[eric-hansen]

I can, but I don't have the whole mini.nvim suite installed currently. Since this is specific to completion (I think) can the branch be pushed also to the https://github.com/echasnovski/mini.completion repo? If not I understand.

I found it frustrating I couldn't reproduce the issue with a minimal PHP set up the following day as well, yet I could when I raised this ticket, even though the same laptop was used for both instances so nothing really changed. However, I still experience the issue on my other machine that does have hefty/heavy projects pulling from intelephense.

[echasnovski]

I can, but I don't have the whole mini.nvim suite installed currently. Since this is specific to completion (I think) can the branch be pushed also to the https://github.com/echasnovski/mini.completion repo? If not I understand.

Sure, I've pushed the isincomplete branch to a standalone repo.

I also opted out of triggering completion immediately for trigger characters; only retriggering after incomplete previous list is immediate. As per LSP spec, those are meant to have the same nature as regular "identifier characters" (usually [a-zA-Z], but in Neovim it is controlled by 'iskeyword' option), just specific to the language itself (like ., $, etc.).

[eric-hansen]

Thanks! I'll give this a try tomorrow when I'm back on the machine that I can really test this on and let you know how it goes.

[echasnovski]

Just a heads up:

• I was able to reproduce the original "missing items; need to " issue in a bigger project. It is indeed caused by isIncomplete and the pushed branches should indeed fix that.
• There is currently visible flickering (popup menu disappears and reappears in very quick succession) for when 'mini.completion' tries to force update the list after isIncomplete. I think I have a solution, though.

As these changes are probably going to be merged into main anyway, confirming the above findings is not a high priority. If you're busy, you can wait until those are merged with a cleaner implementation.

[echasnovski]

This now should be resolved on latest main (both 'mini.nvim' and standalone 'mini.completion') by respecting isIncomplete. I've tested in PHP files specifically :)
Thanks again for providing very useful reproduction steps!