Cursor's Planning mode: Ultimate tokens saver

[abuquack]

Would love to hear your thoughts on this.

So far, the only open-source project that embraced Cursor's (and probably Windsurf's) method of applying diffs is avante.nvim, as detailed in their blog post.

Simply put, in Cursor, the resultant diffs look something like this:

You can notice the comments starting and ending with three dots. Cursor isn’t applying these diffs using some magical algorithm—it’s using a distilled, private LLM model that’s very small, fast, and has a large context window, then comparing for accuracy.

They originally relied on the LLaMa 70B model, which was slower and had more errors. But later, they trained their own model (as they mentioned in their now-deleted article, which seems to be the secret sauce 😆).

I think this could save us a lot of headaches and enable a faster workflow by letting the LLM output smaller windows of code, then using Groq’s insanely fast models to apply the diffs.

What do you guys think? Critiques on this approach are much appreciated. It really looks like a magical solution to the extra expenses caused by tools like Cline, Roo-code, and Aider, where the agent struggles to apply edits efficiently.

[jakob-gross]

isnt there diff mode in roo? Aider has it i think