Ollama disabled Claude Code's token countdown over cache misses

Ollama disabled Claude Code's live tokens-left countdown in its v0.33.0-rc2 release candidate, published August 21. A cache miss isn't free, it means reprocessing tokens you already paid to process once, and this one was happening on every single request.

The countdown was a system message Ollama kept moving to the front of the prompt. Prompt caching only works when the front of the prompt stays identical, so that one line was invalidating the KV cache every time.

The same release fixes a second bug: cancelled prefills were discarding restore points. On recurrent-layer models, a request that matched 46,000 of 47,000 tokens against the previous turn still reprocessed the whole thing from zero, per the release notes.

Both bugs share a cause, something dynamic sitting where the cache expects something static. The release also adds a front-end feature alongside the fix: a menu-bar toggle for turning individual local Ollama models on or off for use inside Claude.