Avante extension - v2.2.0

[ravitemer]

The Avante extension automatically updates your [mode].avanterules file (e.g. planning.avanterules) whenever MCP servers or their tools are updated. This ensures Avante always has up-to-date information about available MCP capabilities in its system prompt.

⚠️ Important Loading Behavior: Avante loads rules files only once when AvanteChat is opened after Neovim starts. Changes to enabled/disabled MCP servers will update the rules file, but Avante won't see these changes until you:

1. Exit Neovim completely
2. Start Neovim again
3. Open Avante fresh

Commands like /reset, /new, or /clear do not cause Avante to reload the rules file.

⚠️ Tool Conflicts: Disable any built-in Avante tools that might conflict with enabled MCP servers to prevent duplicate functionality or unexpected behavior.

Setup

Add MCP capabilities to Avante by including the MCP tool in your setup:

require("avante").setup({
    -- other config
    custom_tools = {
        -- optional: mode is "planning" that is when you open the chat using toggle or <leader>aa
        --optional: cwd can be a string or a function that should returns path
        require("mcphub.extensions.avante").mcp_tool("planning",function()
          return require("avante.utils").get_project_root()
        end)
    }
})

⚠️ Warning: File Override Behavior

This extension modifies your [mode].avanterules file. When creating a new file or replacing content, it uses this template:

{% block mcp_servers %}
# Active MCP Servers:
- server1: tool1, tool2
- server2: tool3, tool4

# Available Tools and Resources:
[Detailed list of capabilities...]
{% endblock %}

File handling works as follows:

1. If file has the MCP servers block:
   • Only content within the block is modified
   • Content outside the block remains untouched
2. If no block present:
   • ENTIRE FILE CONTENT WILL BE REPLACED
   • A new file with proper block structure will be created
   • All custom instructions will be lost

To safely use custom instructions:

1. Add the MCP servers block at the END of your [mode].avanterules file
2. Put your custom instructions BEFORE the block
3. Keep the block at the END to prevent MCP content from being overwritten

Example .avanterules file with custom instructions:

# Your Custom Instructions
You should always write tests for your code.
Handle edge cases carefully.

# IMPORTANT: Keep this block at the end of file
{% block mcp_servers %}
[MCP server capabilities will be automatically updated here]
{% endblock %}

@yingmanwumen please check it out, and let me know your thoughts.

[yingmanwumen]

This is a wired error... I'm looking it:

Error executing vim.schedule lua callback: Vim:E729: Using a Funcref as a String
stack traceback:
	[C]: in function 'expand'
	...e/nvim/lazy/mcphub.nvim/lua/mcphub/extensions/avante.lua:6: in function 'update_avanterules'
	...e/nvim/lazy/mcphub.nvim/lua/mcphub/extensions/avante.lua:70: in function 'cb'
``

[yingmanwumen]

Will try it out with mcp tool as with the default prompt it used to make even sequential calls in one response hallucinating the response of the first tool call result.

I've researched the way that cline does: it terminate the streaming immediately after it recieved a whole tool invocation even if the llm has something to say after the tool invocation. I think that's also a good method, too. But it is hard to implement it in codecompanion based on current architechture.

Also I have a lot of ideas, such as "pin" a reference:

olimorris/codecompanion.nvim#1099

Why we have to leave the historic file contents which are useless in messages? A smart solution is to only keep the lateset updated reference to the file, and remove the unneccessary history messages.

And the contents of the tool invocation inner xml codeblocks often break the xml codeblocks. For example, when writting documents with llm, it will produce a lot of code examples which are wrapped with backtics:

```xml
<tools>
<tool>
....
here is the example of the codes:
```rust     <-- The tool invocation is broken here
far(boo())
```      <-- or here
....
</tool>
</tools>
``` <-- And this tool will not be executed

And it is hard to tell LLM to avoid it...

Although codecompanion is flexible enough for other users, it is still lack of flexibility for me. So I think I may develop a new agent with features in my head. Or I may just use cline...

[ravitemer]

Agree on everything. I do like the cline way of parsing the response it kind of finds the tool as the output is being streamed using the starting tag (which looked strange and error prone on reading the source code but I guess it works flawlessly in real use) thereby restricting it to one tool use at a time.

Regarding pin, I've never used it because I have setup what exactly you are talking about just updating the current file contents everytime we chat rather than keeping the history of the whole file reads, writes, updates etc. See olimorris/codecompanion.nvim#420 how I setup auto context for llm without explicit references way back.

And it is hard to tell LLM to avoid it...

True

Or I may just use cline...

Currently cline is the only reason that I open vscode. So much so that atleast once a day I want to create cline.nvim , Cline but for neovim. It should be possible with integrating multiple adapters logic from codecompanion, some inbuilt tools and the prompt of cline. But I want to be sure that the UI and UX are super clean as we can't have css in neovim 😄 . One day!

[yingmanwumen]

So much so that atleast once a day I want to create cline.nvim

But I want to be sure that the UI and UX are super clean as we can't have css in neovim 😄 .

I think neovim might not be able to handle a complex ai agent, since it is only a text editor without UI support, and lua is not powerful enought to maintain a large project...

But I think a powerful and general-using AI agent alone with a neovim mcp server would be better (or why we have to open a text editor since AI will make everything right? Image that we open 4-5 agents at the background in the same time, doing different coding tasks automatically, and all you need is to review and append some suggestiones, then everything is done...

I don't know the detailed definition of AGI. In my point of view, AGI is when an AI Agent could do part of coding work like an intern 😄 And much more cheaper and obedient

[yingmanwumen]

Regarding pin, I've never used it because I have setup what exactly you are talking about just updating the current file contents everytime we chat rather than keeping the history of the whole file reads, writes, updates etc.

unfortunately, even if cline doesn't do it well....

[ravitemer]

Cline doesn't care about tokens 😂. Even for a single line change, the whole file is sent for review. But as long as it does the work I'm happy. I am using VSCodeLLM api thanks to copilot subscription. Otherwise I can't afford cline😂

[ravitemer]

@yingmanwumen avante integration is updated with 17a86c7 commit with some changes in the avante plugin itself that supports dynamic system prompt. It is working flawlessly. Make sure to update avante and mcphub.nvim and test it out. Delete any previous .avanterules files created by the extension. See the commit and readme to configure avante with this update.

[yingmanwumen]

I'll have a try when I am free. Thanks for your work! T_T Being busy at work now