+1

I seriously thought I was going crazy when gpt-4o-mini was making up example code after knowingly reading the file contents and I explicitly instructed it to reference code from the repo. Turns out this was the problem. Made a fix locally and it works as expected now.

For reference, I changed the following:

to:

var result interface{}
if fileContent != nil {
    result, err = fileContent.GetContent()
    if err != nil {
        return nil, fmt.Errorf("failed to get file content: %w", err)
    }
} else {
    result = dirContent
}

Does having it in base64 also increase the token usage?

Does having it in base64 also increase the token usage?

yes. The model has to write code in which it reproduces the base64 file. This code will then produce the original file. This is 3x token usage.

Binary content is meant to be base64 encoded, and text content should not be. I have fixed this by using the raw API properly. It's been a TODO for me for ages, finally got round to it: #505

Things should be faster, and less brittle too.