Line content gets deleted after quitting Insert mode

[clangd]

Line content gets deleted after quitting Insert mode

[fannheyward]

What's the output :verbose imap <Esc>? Can't reproduce, maybe caused by other vim plugin.

[ythomop]

I checked and there is no mapping. The problem, as I wrote on the OP, only happens on files that make coc-clangd kick in, except for .cpp ones. It does not happen on other file types. So, I am pretty sure it is a coc-clangd bug.

[ythomop]

I tried every possible combination of settings and I managed to stop this behavior by removing the first line of my coc-settings.json:
"coc.preferences.formatOnInsertLeave": true

From the coc.nvim documentation, this setting does the following:

"Trigger format on type when insert leave by send \n to the server."

Maybe this has to do with clang specifically?

[fannheyward]

Reproduced.

[sam-mccall]

Minimal:

coc-settings.json (I needed all of these)

  "coc.preferences.formatOnInsertLeave": true,
  "coc.preferences.formatOnType": true,
  "coc.preferences.formatOnTypeFiletypes": ["cpp"],

Create file containing int;, open it.
Type ox<esc>
The new line containing x will be erased

Relevant messages:

• textDocument/didChange, insert "x" at line 1 char 0
• textDocument/onTypeFormatting, character is "\n"at line 1 char 1 (after 'x')
• formatting response: delete from line 1 char 0 to line 1 char 1

Where's the bug: unclear.

onTypeFormatting is underspecified, clangd expects the client to do what vscode does: edit the document first (including any client-side autoformatting) and then send the onTypeFormatting to the server.
Here, coc is sending a character of "\n" but not editing the document to include one, which doesn't seem to be a well-defined use of the protocol. In fact, formatOnType really does claim it needs to be triggered by a character that was typed, so binding exit-insert-mode to it is always going to be a stretch.
Clangd assumes that the previous "\n" is the one that was inserted, and anything between it and the point is automatic indentation. It formats the previous line.
We could hack around this by trying to detect the coc behavior on the server side, or make coc send something a bit less ambiguous, and teach clangd to do the right thing. What if we make "ch" the empty string?

[fannheyward]

@sam-mccall can you test the master branch? Just a quick fix to hack the ch to empty.

[fannheyward]

Here is an issue about onTypeFormatting: neoclide/coc.nvim#1363 (comment)

[sam-mccall]

@fannheyward Better with that patch indeed, it doesn't destroy text :-)

And it does some formatting, not sure if it's ever going to be great as we don't know how well-formed the code is after exiting insert mode. Depends on editing flow I guess.

That will stop "regular" format-on-type from working though: if you press enter while in insert mode it'd be nice if format-on-type activated (though I can't get that to work in coc today...)

[fannheyward]

Yes, this will stop format-on-type feature. I've revert this patch from master, and create a PR to trace it. #37 (I should do in this way from the beginning...)

[sam-mccall]

Yes, this will stop format-on-type feature

It actually doesn't stop format-on-type from working, clangd will reformat immediatly around the cursor if ch is sent as ''.
It's specifically ch='\n' that causes problems, because it tries to do line continuations and reindenting and stuff.