v0.1.2

• Full prompt caching for Claude in Pi and OpenCode. Claude models now go
  through Concentrate's Messages endpoint, where the cache grows with the
  conversation. On the Responses endpoint the cache fell behind during long
  tool loops, so earlier tool results were billed again at the cache-write
  price. On a long agent task this took the cache hit rate from about 40% to
  over 80%. Other models stay on Responses, where they already cache well.
• Pi and OpenCode cost estimates use each model's standard price. 0.1.1 took
  the first price listed, which for some models (such as Claude Opus 5.5) is
  the fast tier, so estimates showed double. Pi also gets the higher price for
  very long inputs where a model has one. Billing was never affected.
• cn changelog and the notes shown after cn update work in installed copies.