Writing NCM2 sources

Category: vim

Tags: vim

NCM2 is a plugin for Neovim and Vim which provides users with a unified completion menu. By default it displays nothing, users need to add sources for completion. These plugins instruct NCM2 on how to get completion candidates for a particular use-case, such as words in the current buffer, or function names in a programming language. The NCM2 documentation explains how to write the individual pieces of a source, and in this blog post I am going to dive into how to connect those pieces into a working source.

We will start out with a simple synchronous source and finish with a case-study of my asynchronous ncm2-vlime source. Along the way I will be pointing out the relevant entries in the NCM2 manual for context.

Source for the day of the month

Let us start with a simple source which fills in the day of the month. Set up the Vim plugin as usual first, then create the directory ncm2-plugin. NCM2 will search all runtime directories for the ncm2-plugin directory and load all Vimscript and Python files it finds there.

We will be using Vimscript for our sources because it is the canonical and ubiquitous language for scripting Vim. Create a new file named ncm2-plugin/months.vim (the name of the file does not really matter). We can now define our source in this file.

let s:months_source = {
    \ 'name': 'Months',
    \ 'mark': 'Month',
    \ 'enable': 1,
    \ 'ready': 1,
    \ 'priority': 2,
    \ 'word_pattern': '[a-zA-Z]+',
    \ 'complete_length': 2}

A source definition (see ncm2#register_source()) is a dictionary with certain entries, some of which are optional. The values so far are self-explanatory. Do keep in mind that the regular expression patterns are in PCRE format, not in Vim's format. In particular this means you cannot use things like the \v magic flag.

So far we have only specified passive properties of the source, but we also need to instruct NCM2 on how to actively complete the text. This is done by specifying a callback function which will be called by NCM2.

" Completions results are always the same
let s:months = ['January', 'February', 'March', 'April', 'May', 'June', 'July',
    \ 'August', 'September', 'October', 'November', 'December']

" The callback function
function! s:complete_month(context)
    call ncm2#complete(a:context, a:context.startccol, s:months)

" Add the callback to the source defintion
let s:months_source['on_complete'] = function('s:complete_month')

Completion in NCM2 is a two-step process: NCM2 analyses our input and when there is something to complete it calls the callback function of a source. This in itself does not present any completion results to the user yet, it just gives the source power to decide what is to be done next. In the case of such a simple source we can tell NCM2 right away what the completion results are.

The callback function takes one argument (a dictionary describing the ncm2#context in which completion was triggered) and returns no result. The ncm2#complete() function takes three arguments: a context, the first column (counted in character) and the list of completion results.

Why have such a two-step process? In a more complicated case we might not have all the results yet, and computing them all might take a while. If we had to return the completion results right away this would block the entire editor. Instead we can delay the decision by calling some asynchronous function and return right away. The editor remains responsive and the popup-menu will show once completion results are available.

Finally we need to register the source to make it available in NCM2.

call ncm2#register_source(s:months_source)

We can now start the editor, start typing the name of a month and see its name being completed. Since our source has no scope entry the source will work for any file type.

Case-study: ncm2-vlime

Let us now consider how one might write an asynchronous completion source. The definition of the source is similar to above, so I will skip over it. The callback function is defined as follows:

function! ncm2#vlime#on_complete(context)
    let l:connection = vlime#connection#Get(v:true)
    call l:connection.SimpleCompletions(a:context.base, {c,r->ncm2#vlime#complete_simple(a:context, r)})

I have omitted some boilerplate code and the error checking. As it happens, Vlime already provides an asynchronous way of getting completion, so we can focus on just connecting the wires: we get a connection object (a Vimscript dictionary) and call its SimpleCompletions method. This method is asynchronous, it does not return any result, but it lets us specify a function to call when the results are available. The function takes two arguments: the connection (c) and a list of results (r).

This is where the two-step process comes in. The callback is a lambda (anonymous function) which calls another function. To recapitulate: NCM2 calls the callback of a source, the callback then tells Vlime to fetch completion results and what to do when the results are in, then returns. When Vlime is done fetching completion results its own callback calls the following function:

function! ncm2#vlime#complete_simple(context, result)
    let l:matches = a:result[0]
    let l:startccol = a:context.startccol
    call ncm2#complete(a:context, l:startccol, l:matches)

Just as in the example with the months, the function receives the completion context and a list of results. Depending on the format of the results we might need to do some transformation of them first, but other than that this is effectively all there is to it.


Defining a source requires us to ask two questions: what to do when completion is requested, and what to do when the results are in. In the case of simple and synchronous sources both questions are one, but more complex and asynchronous plugins must consider those two questions separately. I have tried to illustrate the process here:

  [ Completion gets requested ]
[ Callback of the source called ] ---> [ Completion callback gets set up ]
                 |                                      |
                 |                                      |
                 V                                      V
    [ Resume Editor control ]                  [ Compute results ]
                 |                                      |
                 |                                      |
                 V                                      V
      [ User keeps typing ]                  [ Present completions ]
                 |                                      |
      [ Apply completions ]

The left-hand column represents the main sequence of execution. In the second step a parallel sequence of execution is set up, which eventually merges back into the original sequence.