Next Position Cursor Prediction (Inspired from Cursor Obvioslly...)

[PLAZMAMA]

I would like to take a crack at creating a cursor position prediction feature kind off like what the cursor team is working on.
Would this fit the projects goals/direction @olimorris?

[olimorris]

It's a very interesting feature.

I've shied away from implementing a suggestions strategy for a couple of reasons. Firstly, I don't think I could improve upon copilot.vim. Secondly, the amount of context that's needed to do suggestions properly, means that it's really only Ollama models that are viable for most users from a cost perspective. I expect that the next cursor position would need to know the history of where the user's cursor has been, the buffer's code and probably the overall structure of the cwd.

If you think this could be accomplished differently, let me know because I do think this is hugely compelling as a feature.

[PLAZMAMA]

Specific Responses to Your Comment (Thanks for responding btw)

1. Firstly, I don't think I could improve upon copilot.vim

Good point, copilot, supermaven and codeium do it so good already so I found myself just using them. But I think we should state this clearly in the readme to avoid confusion (which I had while trying to figure out where the inline suggestion feature was and how to activate it)

3. Secondly, the amount of context that's needed to do suggestions properly, means that it's really only Ollama models that are viable for most users from a cost perspective.

I think this can be addressed in the two following ways:

• Make context window size adjustable to the users need via a context_window_size option
• Limit context only be file specific.

My Current Implementation Strategy

Taking into account everything you laid out and my vision for this feature, I will formulate an implementation strategy below. I would like more feedback on this feature, especially stuff like the last comment.

Implementation overview (will be written in a bit as I have to switch to my laptop to commute)

I think the implementation is composed of four parts

• Context
• Prompt
• Response parsing and prediction display
• Event/prediction trigger

Context

I think the best way to get the context efficiently will be the following:

• Keep a list of the history of edited line numbers for each file. This list/table will look like the following:

local edits_history = {
    {
        ["name"] = "file_name_1.lua"
        ["edits_line_num_history"] = {
            33,
            22,
            33,
            9,
            102,
            ...
        }
    },
     {
        ["name"] = "file_name_2.lua"
        ["edits_line_num_history"] = {
            43,
            24,
            43,
            111,
            12,
            ...
        }
    },
    ...
}

• I'm going to flush everything below next week as i have school deadlines and a busy weekend
• I will tag you (Ollimorris) once I'm finished

Prompt

Response parsing and prediction display

Event/prediction trigger

[olimorris]

I'm not sure if by extension you mean plugin but I would strongly urge you to consider turning this into its own plugin at some point. I think it's potentially too good to merge into mine as so many other users could benefit from it. It really could be Harpoon but on steroids. From a CodeCompanion.nvim perspective, if there's any integration that we'd need to do, I'd make it a top priority.

BitNet looks like a great suggestion too.

Good luck with grad school and hope you nail it!

[PLAZMAMA]

I'm not sure if by extension you mean plugin.

Yes, I meant plugin, my bad. VScode and Chrome are too ingrained in my head, unfortunately.

I would strongly urge you to consider turning this into its own plugin at some point. I think it's potentially too good to merge into mine as so many other users could benefit from it.

Exactly!!!, I already created the empty repo for it https://github.com/PLAZMAMA/bunnyhop.nvim.

It really could be Harpoon but on steroids.

Yes, that is one of my favorite Nvim extensions and what initially made me switch from VSCode to Neovim, it also inspired this approach to the idea of cursor prediction.

Good luck with grad school and hope you nail it!

Thanks! But I think there was a little misunderstanding I'm finishing my Bachelor's in CS rn, I wish I was already doing my Master's lol.
I went with a less conventional route of learning code in High School -> 2yrs coding work for $ -> use $ to do a Bachelor's.

From a CodeCompanion.nvim perspective, if there's any integration that we'd need to do, I'd make it a top priority.

I have some ideas for that as well... but I'm trying not to jump too far ahead and focus on getting the base functionality in a good state first before I bring these ideas up and get too excited.

[olimorris]

Hey @PLAZMAMA, hope you're good. Did you make any progress on this btw?

[PLAZMAMA]

Hi @olimorris, since we last spoke I've completed my last 2 classes and am now finishing my capstone project. I hope to start working on this by the end of the month.

[olimorris]

Hi @olimorris, since we last spoke I've completed my last 2 classes and am now finishing my capstone project. I hope to start working on this by the end of the month.

Good luck on the project!

[PLAZMAMA]

Hi @olimorris, I've been working on the extension for a few days now, and so far, it's going alright. I've been able to create something that when a keybinding is pressed TRIES to predict your next cursor location and jump to it. I've still got a long way before this is usable though and there are a lot of improvements to be made.

I've run into some problems that I haven't been able to figure out yet so if you have any tips it would be appreciated.

• I've been using vim.fn.system for running curl which works but isn't async. When I try vim.system for curl commands I get a nested braces in the URL error. I've tried to debug it by stepping through Neovim's API code and wasn't able to figure out what is the raw command it creates. Is there a way to "dry run" this function or be able to see what is the final command it formulates?
• I've tried figuring out how to send requests to copilot chat, I haven't seen any documentation on the internet and wasn't able to figure it out from reading your and CopilotChat.nvim. Do you know of any examples or contained code I can use to accomplish the auth, token handling, and requests?

Additionally, after trying to debug all of these things for the last couple of days, I think it would be really helpful for people to have a library they can use to query all of the different kinds of AI API's. This will prevent the need to rewrite the same adapter logic over and over again for every plugin the wants to utilize these API's. Since Code Companion has the biggest adapter library, I think it might be worth for someone to try to take a look at the adapters implementation you got here and try to convert it to a single independent library for people to use.

Lastly, I wanted to say thank you again for nudging me to make this plugin happen and I hope soon I'll be able to say that its at a usable state.

[olimorris]

Hey @PLAZMAMA great to hear from you and this sounds awesome!

Is there a way to "dry run" this function or be able to see what is the final command it formulates?

I can't say I use vim.fn.system that much so unsure of its nuances. I'm using Plenary.Curl for all of CodeCompanion because of its asynchronous properties and because it's so well battle tested at this point. It can handle streaming too. Also, a great tip, if you use CodeCompanion's /fetch slash command and send the raw file to an LLM, you can ask it in detail how it works. That's helped me out loads when building the plugin.

Do you know of any examples or contained code I can use to accomplish the auth, token handling, and requests?

Check out my implementation, here. We've solved all of that. To see the full payload that's sent to the LLM, you may need to set the debug level to "TRACE" in your CodeCompanion settings and examine the JSON file that the plugin creates

Really excited to try this out.

[PLAZMAMA]

@olimorris I've been working on the plugin(bunnyhop.nvim) since we last talked and it is now at a point where I feel it can be of some use. Its still in a very alpha stage but you can start using it if you want. I have a lot of improvements (such as the context creation, tests, auto release, etc..) which I will continue making of course and if you have any other suggestions it would be appreciated.

[olimorris]

Amazing work!!! I can't wait to try it out. Reading the code base, I need to supply a hugging face API key in my config?

[PLAZMAMA]

Good question, you need to create a hugging face environment variable with your API key and supply the name of that environment variable in the config. I'll clarify it in the readme.