Really Cool Tool. Wish it was more like Aider though!

[roshaans]

I really love all the customizability and editabilty of this plugin. I just come from having significantly used Aider and it makes editing your codebase a breeze.

Some things that would make me switch over:

1. Seems like no matter what model I use, it only seems to make 1 change at a time. I have to currently fight it to make more general changes. Where as, Aider reads my mind and knows exactly what I am trying to do.

What I would like to see here is that it keeps prompting itself naturally (Aider style). I think by default the chat should assume that a user is trying to make a project wide change using the @editor, @files included. At the moment it seems like editing files is secondary and primary function is simply chatting with your files.

I have not tried to make good use of workflows yet, but if that would solve any of my problems do let me know!

2. Buffer management, Aider makes it very easy to do buffer management where as I am always confused what is in context and what is not. I have tried to use VectorCode and it may work, but still its extra things to setup. If VectorCode could be plugged in by default that'd be nice. Still though, even with VectorCode, It's still very hard to know what CodeCompanion is using as context.

Besides this, the entire expereince of having the entire chat history be a whole buffer is super neat and I have not seen that done before. Also MCP server integration with MCPHub was a breeze! Great work!

NOTES

Buffer Management seems to be fine for me now!

[roshaans]

I tried to use the Aider-mcp-server integrated into MCPHub.nvim and these are the results :/. It did not end up changing anything.

---

model: gpt-4o
reasoning_effort: medium
temperature: 0
max_tokens: 15000
top_p: 1
n: 1

Me

Context:

• access_mcp_resource
• use_mcp_tool

@mcp use aider to create a new controller called users

CodeCompanion (Copilot)

Tool Output

use_mcp_tool Tool: Returned the following:

success

Me

Context:

• access_mcp_resource
• use_mcp_tool
• src/api/controllers/leaderboard.ts
• cmd_runner
• editor
• files

@full_stack_dev use aider to create a new controller called users

CodeCompanion (Copilot)

Tool Output

use_mcp_tool Tool: Returned the following:

success

Me

Great. Now let's consider your code. I'd like you to check it carefully for correctness, style, and efficiency, and give constructive criticism for how to improve it.

CodeCompanion (Copilot)

Please provide the code you'd like me to review, or let me know where I can access it. If it's already in a Neovim buffer, I can analyze it directly.

Me

Thanks. Now let's revise the code based on the feedback, without additional explanations.

CodeCompanion (Copilot)

Please share the code or specify the buffer where it resides so I can revise it based on the feedback.

Me

I wnat you to add new new files...use the AIder tool it knows what to do!

CodeCompanion (Copilot)

Understood. I'll directly use the Aider tool to create the new files as per your request. Let me proceed.

Me

[cjoke]

Why build a second Aider?

[roshaans]

Well, learn from the best of Aider and bring it here. You can't deny that there is very much so that Aider has gotten right.

[olimorris]

Why build a second Aider?

No one has suggested this.

[ravitemer]

@roshaans Can you link the Aider MCP Server so that I can test it

[roshaans]

https://glama.ai/mcp/servers/@sengokudaikon/aider-mcp-server

Maybe there is a better mcp server. But I think something closer to a native experience would be better. I often find that CodeCompanion also masks important MCP server errors sometimes.

There does not seem to be any concept of retries. I think that would help a ton!

[ravitemer]

Maybe there is a better mcp server. But I think something closer to a native experience would be better. I often find that CodeCompanion also masks important MCP server errors sometimes.

I think this is not the case anymore. Please update both the plugins. The mcphub extension responds with the error. Can you give me an example if it's still the case?

[olimorris]

Seems like no matter what model I use, it only seems to make 1 change at a time

This seems to be very LLM dependent. It's quite the frustration. I'm not sure if we're missing a parameter in the request object which tells the LLM to call multiple function calls or if it's a system prompt fix. But it's annoying having to do that with certain LLMs.

It's still very hard to know what CodeCompanion is using as context.

You should see a big context block in the chat buffer which indicates what's been shared with the LLM. VectorCode is the exception.

[roshaans]

This seems to be very LLM dependent. It's quite the frustration. I'm not sure if we're missing a parameter in the request object which tells the LLM to call multiple function calls or if it's a system prompt fix. But it's annoying having to do that with certain LLMs.

I think studying the Aider codebase would be helpful here. Seems like they have a prompt or so after every response to figure out next steps and steps to resolve any issues. I believe they call it reflections and there are at max 3 of these allowed by the LLM

[olimorris]

If you investigate, please report back. Would be keen to learn any interesting techniques they're doing. I do have a plan to improve multi model editing in a buffer sometime soon.

[pratyushmittal]

The new @files tool does call multiple function tools together when updating multiple files.

@olimorris The larger issue here, I guess, is that we are not using recursion.

Once the LLMs make the tool calls, we don't submit back the responses automatically.

This is the pseudo code that we use in our other project.

function get_answer(messages)
    tool_calls = []
    for chunk in openai.get_completion(messages, stream=true)
        if text then
            yield chunk
        else
            tool_calls.append(...)

    if tool_calls:
        for tool in tool_calls:
            reply = call_tool(tool)
            messages.append(reply)
        get_answer(messages)

The last line here is most important. Calling get_answer(messages) once again if we have had tool calls. This creates a recursion.

Benefits of this approach:

• The errors are automatically submitted. LLM is able to fix them and auto-heal.
• The LLMs make multiple calls based on new outputs.

Do you think this can work?

[olimorris]

This has been a feature for a while now but I think it went undocumented so I've added this in, here.

My issue is that previously, with the XML tools approach, all LLMs seemed to have the ability to call multiple different tools within the same response. Now, it seems only a few can do that.

[pratyushmittal]

Mind blown 🤯. Awesome!!!