`#file:…` or `#buffer:…` doesn't seems to work

[tchernomax]

Hello

context

For example on this repository: https://gitlab.gnome.org/GNOME/gnome-shell

 % ls -l README.md NEWS fix
-rw-r--r-- 1 max users  892  9 févr.  2021 fix
-rw-r--r-- 1 max users 234K  9 févr.  2021 NEWS
-rw-r--r-- 1 max users 1,2K  9 févr.  2021 README.md

prompt examples

Here are some prompts, taken secondes/minutes apart, without any modification to any file or conf.
Each prompt was taken with a fresh neovim process (quit and reopen neovim between each prompt).

file and buffer doesn't work

## User

##file://README.md
what's the last line of the README ?

## Copilot

The last line of the README is empty (a blank line).

## User

##file:///home/max/Documents/repositories/gitlab.gnome.org/gnome-shell/README.md

## Copilot

The last line of the README is blank (an empty line).

## User

##buffer:///home/max/Documents/repositories/gitlab.gnome.org/gnome-shell/README.md

## Copilot

The last line of the README is blank (an empty line).

buffer works

## User

##buffer:///home/max/Documents/repositories/gitlab.gnome.org/gnome-shell/README.md

what's the last line of the README ?

## Copilot

The last line of the README is:
[license]: COPYING

file doesn't work

## User

##file://README.md

what's the last line of the README ?

## Copilot

The last line of the README is empty (a blank line).

file doesn't work, buffer works

## User

##buffer:///home/max/Documents/repositories/gitlab.gnome.org/gnome-shell/README.md

what's the last line of the README ?

## Copilot

The last line of the README is:
[license]: COPYING

## User

File not found: fix
(redacted: ↑ I wrote `#file:fix`)

│what's the last line of the "fix" file ?

## Copilot

There is no file named "fix" in the current directory, so I cannot provide its last line.

## User

##buffer:///home/max/Documents/repositories/gitlab.gnome.org/gnome-shell/fix

## Copilot

The last line of the "fix" file is:
23:10:56 gnome-shell[1304]: _onSessionCompleted end

file doesn't work

## User

File not found: /home/max/Documents/repositories/gitlab.gnome.org/gnome-shell/fix
(redacted: ↑ I wrote `#file:/home/max/Documents/repositories/gitlab.gnome.org/gnome-shell/fix`)

what's the last line of the "fix" file ?

## Copilot

There is no file named "fix" in /home/max/Documents/repositories/gitlab.gnome.org/gnome-shell, so I cannot provide its last line. If you create the file or specify a different file, I can help further.

buffer doesn't work

## User

##buffer:///home/max/Documents/repositories/gitlab.gnome.org/gnome-shell/NEWS

what's the first line of NEWS ?

## Copilot

The first line of NEWS is:

GNOME Shell News
(redacted: ↑ it's not… the real first line is "40.alpha.1.1")

notes

As you can see it's somewhat strange.
Sometimes it works, sometimes it doesn't.

I thinks the size of the file might explain why copilot never displayed the right answer for the NEWS file.
(it's responding too fast, I think, it doesn't get the full 234K, so he is hallucinating)

There is no error in the CopilotChat.log (in trace log level):

% grep -i error /home/max/.local/state/nvim/CopilotChat.log | grep -v 'errors or search result'
%

Here is my LazyVim conf (I disabled copilot.lua to be sure it wasn't the culprit):

==> /home/max/.config/nvim/lua/plugins/CopilotChat.lua <==
return {
	{
		"CopilotC-Nvim/CopilotChat.nvim",
		dependencies = {
			{ "nvim-lua/plenary.nvim", branch = "master" },
		},
		build = "make tiktoken",
		opts = {
			window = {
			},
			log_level = 'trace',
		},
	},
	{
		"ibhagwan/fzf-lua",
		dependencies = { "nvim-tree/nvim-web-devicons" },
		opts = {}
	},
}

==> /home/max/.config/nvim/lua/plugins/copilot.lua <==
return {
}

I can give you some logs if you want.

GitHub Copilot Business is active on my account.

note: I am a total noob on CopilotChat and somewhat neovim plugins and such (I installed & configured LazyVim only last week, to install CopilotChat… and it took me some times).

Do you have any idea on what's going on ?

Regards

[deathbeam]

Well fix does not work because neovim cannot detect its filetype. News is indeed massive so probably do not even fits into context so the whole file is skipped.

Try renaming README to README.md, maybe that will fix the filetype detection (neovim can detect the filetype regardless usually, but for that it needs to read the file partially, which is why buffer would work as then the file is already loaded).

But the limitation for only known filetypes in neovim is mostly so people do not accidentally embed binaries. Its probably not super necessary anymore as support for just loading all the files at once was removed so now its mostly filenames and 1by1 loading

[tchernomax]

Try renaming README to README.md

README is already README.md ( https://gitlab.gnome.org/GNOME/gnome-shell )

NEWS is indeed massive so probably do not even fits into context so the whole file is skipped.

Is there a possibility to be notified if there is a problem with "filetype" or if the file doesn't fit into the context.
(I looked at the logs but didn't see anything).

[deathbeam]

Try renaming README to README.md

README is already README.md ( https://gitlab.gnome.org/GNOME/gnome-shell )

NEWS is indeed massive so probably do not even fits into context so the whole file is skipped.

Is there a possibility to be notified if there is a problem with "filetype" or if the file doesn't fit into the context. (I looked at the logs but didn't see anything).

ah my bad. well with log level debug you can at least see what data is being sent so you can try to see the outgoing request and see if it contains README or not.

For the file detection, im considering just removing that and letting people embed whatever, as the biggest reason before for not doing that is because plugin had bulk file operations, but it doesnt anymore

[tchernomax]

ah my bad. well with log level debug you can at least see what data is being sent so you can try to see the outgoing request and see if it contains README or not.

Ok, I think I understand what's going on now.

When the code see a #file:… expression in the prompt, it transform it and add the content of the file to it.

For example:

• prompt: #file:LICENSE
• send to the API: "# file://LICENSE\n```text path=LICENSE start_line=1 end_line=675\n 1: GNU GENERAL PUBLIC L…

But, if the file is too big, it doesn't add the content, for example:

[DEBUG jeu. 04 sept. 2025 10:12:41] …/CopilotChat/client.lua:365: Prompt tokens: 15
[DEBUG jeu. 04 sept. 2025 10:12:41] …/CopilotChat/client.lua:366: System tokens: 1134
[DEBUG jeu. 04 sept. 2025 10:12:41] …/CopilotChat/client.lua:367: Resource tokens: 118681
…
[DEBUG jeu. 04 sept. 2025 10:12:42] …/CopilotChat/utils/curl.lua:78: POST request: https:>
    max_tokens = 16384,
…
        content = "\n##file://NEWS\n\nwhat's the first line of the NEWS file ?\n",

We see that "Resource tokens" > "max_tokens".

→ I think CopilotChat should warn the user when it happens !

Idem if the file type isn't detected (or isn't allowed ?).

For the file detection, im considering just removing that and letting people embed whatever, as the biggest reason before for not doing that is because plugin had bulk file operations, but it doesnt anymore

it's this part of the code, write ?

--- Get data for a file
---@param filename string
---@return string?, string?
function M.get_file(filename)
  local filetype = files.filetype(filename)
  if not filetype then
    return nil
  end

(just for me to try)

In any case, thanks for the help and for the tool !

[tchernomax]

Another idea (do you want me to create a new issue for it):

It would be great if CopilotChat could split it's prompt/request in multiple requests if it doesn't fit into one (because of max_input_token).

[deathbeam]

Another idea (do you want me to create a new issue for it):

It would be great if CopilotChat could split it's prompt/request in multiple requests if it doesn't fit into one (because of max_input_token).

That is not how it works. You need to always send the whole context, the remote api has no memory, the memory is in request. So if it do not fits it will never fit.

[deathbeam]

ah my bad. well with log level debug you can at least see what data is being sent so you can try to see the outgoing request and see if it contains README or not.

Ok, I think I understand what's going on now.

When the code see a #file:… expression in the prompt, it transform it and add the content of the file to it.

For example:

• prompt: #file:LICENSE
• send to the API: "# file://LICENSE\n```text path=LICENSE start_line=1 end_line=675\n 1: GNU GENERAL PUBLIC L…

But, if the file is too big, it doesn't add the content, for example:

[DEBUG jeu. 04 sept. 2025 10:12:41] …/CopilotChat/client.lua:365: Prompt tokens: 15
[DEBUG jeu. 04 sept. 2025 10:12:41] …/CopilotChat/client.lua:366: System tokens: 1134
[DEBUG jeu. 04 sept. 2025 10:12:41] …/CopilotChat/client.lua:367: Resource tokens: 118681
…
[DEBUG jeu. 04 sept. 2025 10:12:42] …/CopilotChat/utils/curl.lua:78: POST request: https:>
    max_tokens = 16384,
…
        content = "\n##file://NEWS\n\nwhat's the first line of the NEWS file ?\n",

We see that "Resource tokens" > "max_tokens".

→ I think CopilotChat should warn the user when it happens !

Idem if the file type isn't detected (or isn't allowed ?).

For the file detection, im considering just removing that and letting people embed whatever, as the biggest reason before for not doing that is because plugin had bulk file operations, but it doesnt anymore

it's this part of the code, write ?

--- Get data for a file
---@param filename string
---@return string?, string?
function M.get_file(filename)
  local filetype = files.filetype(filename)
  if not filetype then
    return nil
  end

(just for me to try)

In any case, thanks for the help and for the tool !

Issue with notifying the users about the request size is that it is getting filtered way too late so its pretty hard to do and there is also no dedicated notification area in the chat window other than in progress message. I guess I can technically imporove gc so it also filters stuff out but for simplicity for now its only in the client.

And for filtering, yes, you can just do if not filetype then filetype = '' or something

[deathbeam]

Moving to discussions as its not really an issue