nvlabs/sol-pi, published by NVIDIA Research, is an automated research harness aimed at cutting the repeated tool calls and context usage of agents. The four techniques the repository describes run on top of pi's public extension points and do not include or modify pi's own source. It is written in TypeScript and is MIT licensed. There is no public latest release.
It bundles an edit and its follow-up verification command into a single tool call, and replaces large results with a handle that can be read again. Long diagnostic logs are compressed into a receipt form only when the citations match the original text, while completed plan steps are promoted to native compression candidates. Nothing is enabled unless configured, so all four techniques are off by default. The repository explains that it contains the techniques that survived repeated automated research loops. The configuration therefore focuses on reducing repetitive work in the harness rather than changing the model itself.
The performance evidence comes from a Blender scene restoration example: compared with a code command-line tool, cost per scene dropped 37 percent and model calls dropped 13 percent. The comparison material is published as model-scored result images for fifteen scenes. No basis is therefore given for extending those numbers into overall harness-wide cost savings. Installation assumes global pi 1.3.0, and it notes that pi 1.0.x and 1.1.0 fail to start due to an internal package pinning problem. Judging effectiveness is a step that has to weigh the example's limited evidence together with the supported environment.