pi v0.85.1: GPT-6 Astra, faster Alt scrolling, SDK import fix

v0.85.1 landed on September 5, one day after v0.85.0, and it is a focused follow-up rather than a feature drop. The headline is model access: GPT-6 Astra is now available through OpenAI API keys and OpenAI Codex subscriptions, with the provider setup documented in the API Keys and OpenAI Codex sections of the providers reference.

The fullscreen TUI gets three small quality-of-life improvements. Holding Alt while scrolling the mouse wheel now moves through the transcript five times faster, which matters on the long sessions the fullscreen view is built for. In the model and thinking selectors, saving a choice is no longer hard-coded to Ctrl+S — the save actions map to the configurable app.models.save and app.thinking.save bindings instead. The third is a hover fix: moving the pointer no longer changes the current selection or re-centers the autocomplete and settings lists, which could make a click land on a different item than intended.

The most consequential change is a repair. v0.85.0 unintentionally published internal experimental code and dependencies, breaking SDK imports downstream. v0.85.1 pulls those surfaces back: the experimental client and experimental/plugin subpaths and the server/client commands are now source-only through pi-test.sh, while the supported local SDK and the stdio RPC API are unchanged (#9132).

One provider-side fix closes the batch: long prompt-cache requests to GPT-5.6+ Responses models now send prompt_cache_options.ttl: "30m" instead of prompt_cache_retention: "24h".

The release notes, with PR and issue links where available, are on the GitHub release page.