New in V4.0.0 - ✨ Improved @editor tool to update your buffers

[olimorris]

Firstly, there's a potential tonne of breaking changes which I've explained in the specific commits. So please consult the CHANGELOG.

@editor tool

I've made a lot of changes to this tool to try and make it a viable means of updating a buffer from the chat. It's now capable of some pretty complex modifications:

Complex.modifications.mp4

Note: I'm using claude-3-5-sonnet and it took 454 tokens to make that modification

Previously, the tool could only handle single modifications. Now, it can stack multiple actions together such as add, delete or update. For full disclosure, the XML from the last video was:

<tools>
  <tool name="editor">
    <action type="delete">
      <buffer>7</buffer>
      <start_line>20</start_line>
      <end_line>28</end_line>
    </action>
    <action type="add">
      <buffer>7</buffer>
      <code>---Count the number of elements in a table (array or non-array)
---@param t table
---@return integer
M.count_improved = function(t)
  if M.is_array(t) then
    return #t
  else
    local count = 0
    for _ in pairs(t) do
      count = count + 1
    end
    return count
  end
end</code>
      <line>20</line>
    </action>
  </tool>
</tools>

I've also made sure that the prompt_library provides buffer numbers where possible so that you can leverage @editor after calling those.

Call to action

I'd ask the community to try this out and curate the system prompt that comes with the editor tool to make it even better. I've tried the example above with Copilot, Anthropic, OpenAI, Google and all work well. But I expect to hit some edge cases and I'd value your investigation on how to solve them.

:CodeCompanionActions can now use external providers

A valid challenge in recent days was making :CodeCompanionActions available for external providers such as Telescope, especially as we reference it as an optional dependency. That is now possible in v4.0.0 and I'd ask the community to look at that implementation and think about adding mini.pick or fzflua.

Streamlined commands

It's now possible to do :CodeCompanionChat why is Lua such an awesome language? to open up a chat buffer and automatically submit the prompt. Saves a few keystrokes and feels more inline with what I would expect if I was using the plugin for the first time.

[bassamsdata]

Thank you for all the hard work. I'm testing the new additions.

I'd ask the community to look at that implementation and consider adding mini.pick or fzf-lua.

I'll work on adding mini.pick in the coming days.

One question, though, it's a bit out of context: what color scheme are you using? I really like the diff colors.

[olimorris]

Thanks for testing. Some of the older models are notoriously bad with this type of request unfortunately. I've tried tweaking the system prompts for them but they seemingly struggle with basic XML. I had really good experience with Gemini Pro. Could you share the specific prompt you sent to it? I'm not actually sure how many people use Gemini as an API endpoint but I'd like to make sure I shake it down properly.

Regarding the code base, just search editor and it will come up. I think the repository is in a pretty clean state these days so easy to poke around in. Btw, what's your default LLM of choice?

[raine]

Thanks for yet another great release!

What would be the most streamlined way to request changes to the current buffer with minimal keystrokes?

[S1M0N38]

I believe that the ability to generate or use XML tags depends on how the model was fine-tuned during the final stages of training. Anthropic's models are among the first to make extensive use of XML tags for structuring prompts.

I'm not sure about other models; they might simply be mimicking the usage of XML tags because they are sophisticated enough to do so (just a speculative thought).

[olimorris]

I watched a video some weeks ago where a dev was sharing his experiences of working with LLMs in production. The use of tags was something he shared as being key to getting reliable performance from them.

I've made sure the tool prompts have few-shot inference to try and circumvent any lack of training they may have had. Certainly had no issues with the "main" models

[olimorris]

@raine I've streamline the system prompt so you can just tag the editor without saying anything