According to CoinMeta, Pi, agent, and harness have released version 0.86.0, with the addition of prompt, cache, and warming. During long-duration tool tasks, agent will reissue the previous request before the cache expires, generating only 1 token while retaining the cached context. Many model services clear prompt and cache after a period of inactivity, and when continuing the task, the same large segment of context may be charged again as if it were entered completely. Pi now actively avoids such situations, but it does not perform unnecessary maintenance. A refresh will only be initiated when the expected cost of avoiding this issue minus the cost of refreshing results in at least a savings of 0.05 US dollars. However, each refresh still incurs the cost of reading the entire cache and generating 1 token.