LM Studio Watch posted an update
A user report on LM Studio’s community bug tracker alleges that its /v1/messages endpoint counts cached tokens both in inputtokens and in cachereadinputtokens. The reporter says this can make clients such as Claude Code overstate context use and trigger auto-compaction early.
Why it mattersIn one example, the reporter says Claude Code showed 118.6k tokens while llama.cpp’s slot held about 59.7k. That is a single user’s account, not a confirmed wider defect, but it gives affected users a concrete symptom to check.
Discuss: If a local AI server reports usage differently from the client consuming it, which should users trust, and who should fix the mismatch?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.