fix: Outdated completion item with mini.snippets

[abeldekat]

Hello @saghen,

I recently fixed this issue in nvim-cmp.

The problem also occurs in blink, and the fix is very similar.

Are there any drawbacks to the following approach?

.../completion/init.lua:

  -- trigger -> sources: request completion items from the sources on show
  trigger.show_emitter:on(function(event)
    vim.schedule(function()
      -- schedule to avoid outdated completion items
      sources.request_completions(event.context)
    end)
  end)

[saghen]

This would add latency to all of the sources. We want to get the request out to the LSP ASAP. I'll read up on that issue you linked soon

[echasnovski]

The ASAP part might indeed be the culprit here, because after that there might be other TextChanged{I,P} callbacks that modify buffer text immediately which causes sent request to be for an outdated buffer state. Here are more details.

[abeldekat]

For nvim-cmp, we might have a solution fixing this concrete problem, without falling back to vim.schedule. I think that a similar approach for blink is more complex unfortunately.

So, for now, in my config I apply this hack before the setup:

-- HACK: See blink cmp completion init, line 20
-- Fix outdated completion items:
local sources_lib = require("blink.cmp.sources.lib")
local orig_request_completions = sources_lib.request_completions
sources_lib.request_completions = vim.schedule_wrap(orig_request_completions)

local blink = require("blink.cmp")
blink.setup(opts)

[abeldekat]

Hello @saghen,

I implemented the same approach for blink.cmp as I did for nvim-cmp.

There is no vim.schedule. This is the relevant fragment in blink.cmp.config.snippets.lua:

-- ... expand
      mini_snippets = function(snippet)
        if not _G.MiniSnippets then error('mini.snippets has not been setup') end
        local insert = MiniSnippets.config.expand.insert or MiniSnippets.default_insert
        insert({ body = snippet })
        require('blink.cmp').resubscribe()
      end,

-- ...

I profiled the solution. For each expand using mini.snippets, an extra 0.057723 ms is added. Measured on an old laptop... For nvim-cmp I measured 0.0475465 ms.

[saghen]

Brilliant, thank you! This does seem a bit hacky, albeit less so than vim.schedule so I'll add some notes around the code to try to solve the core issue eventually. It would be great if neovim had a way to control autocmd ordering

[abeldekat]

It would be great if neovim had a way to control autocmd ordering

Indeed!