Pi Continuity cleans context around signed thinking

Pi Continuity has a new route for clearing old tool results when Anthropic’s signed thinking makes local editing difficult. Version 10.6.0 of the community pi-smart-compact extension adds opt-in server-side context editing for direct Anthropic sessions. The following 10.6.1 patch addresses another long-session problem: a context-pressure figure that could drop misleadingly after a trim or rewind.

This is an extension for Pi Coding Agent, not a new model or a standalone coding assistant. Its workflow archives eligible tool output behind recoverable references, lets users rewind a research detour and reviews compacted summaries before applying them.

Cleanup that respects the provider boundary

The new providerContextEditing setting defaults to off. In a direct anthropic / anthropic-messages session, tool-results clears old shell and read-only research results on the server. The additional tool-results-and-thinking option also clears thinking older than the last four turns. The release says local session history stays unchanged and signed thinking remains valid.

That route has a cost-related caveat: clearing rebuilds the prompt cache from the cleared point. It should not be read as a blanket token-bill reduction. The developer also explicitly says this Anthropic integration was not exercised against the live API in the release evidence.

Context accounting after a rewind

Version 10.6.1 keeps the last measured provider usage and applies the estimated change from edits, with Pi’s estimate as a floor. In the author’s replay of seven real rewinds, this anchored figure was 0.3% to 2% from the next measured input, compared with Pi estimates 14% to 33% below it. These are developer replay results, not AITrending measurements or a guarantee for future sessions. Pi’s own footer is unchanged.

Installation is pi install npm:pi-smart-compact, requiring Pi 0.87.1 or newer and Node.js 22.19 or newer. Existing users reload or restart Pi after updating. The code is MIT licensed; model-provider access and billing remain separate. Optional memory engines are not installed automatically.

Summary approval is required by default. Rewind does not undo file changes, processes or external actions, and deterministic verification can miss information. The practical news is more explicit control over a long-running agent’s working context, with the provider-specific and replay-only evidence limits kept visible.