Skip to content- Sources: primary
- Summary: The cacheWarming setting now defaults to idle, so pi keeps eligible prompt caches warm for up to 30 minutes between agent runs when a refresh is expected to pay for itself. Anthropic sessions use message threads, sending only new messages and falling back to the full request after compaction or a tool change. The release adds pi auth login and logout commands, makes Claude Opus 5.5 the default Anthropic model, and stops sending an empty tools array to OpenAI-compatible endpoints, which fixes errors from vLLM 0.20 and later.
- Why it matters: pi users on Anthropic will see cache-warming charges between runs unless they set cacheWarming back to streaming, and proxies that require tools with tool history now need a compat flag.
- Background: Anthropic prompt caching bills cached prefix reads far below fresh input, but cache entries expire after a short idle period. Warming sends a cheap request to keep the prefix alive, which pays off only if the next turn arrives before expiry.