I reviewed the underlying issue "addressed" in that LLM PR, turns out it was completely unnecessary and down to a misunderstanding. Who knows how many tokens were used to address a problem that doesn't exist if you use the API as intended. They tried to use the wrong tool for the job, when a better one was available, and when it didn't work they asked Claude to write hundreds of lines of code to 'fix' it. If they'd actually asked me (a human), I could have given them the correct answer 🤦‍♂️