SoL-Pi is NVIDIA Research’s MIT-licensed extension for the Pi coding harness that packages four context-efficiency mechanisms from scaled auto-research loops: Action Fusion, ObservationPack paged recall, an evidence-preserving reducer with verbatim-quote verification, and online context compaction.
Its argument is that harness efficiency is a scaling problem: the mechanisms exist because the team ran auto-research loops long enough to watch token cost grow with loop length, and each one attacks a measured waste site rather than the context window in general.
What it is #
A standalone TypeScript extension published by NV Labs (with the accompanying paper, arXiv 2609.20519) that installs beside an unmodified Pi harness. Action Fusion collapses repeated tool sequences into single replayable actions; ObservationPack pages old tool output out of the live window into recallable chunks; the reducer compresses history while preserving verbatim quotes that downstream verification can check; and online compaction rewrites the working context mid-loop instead of at overflow. The project page reports the efficiency results from the paper’s auto-research benchmarks, and the harness keeps running ordinary Pi sessions when the extension is off.
Status #
Young and active: 3,443 stars, 273 forks, and 78 open issues as of 2026-10-13, created 2026-09-02 and pushed the day of this check, MIT licensed, with work continuing toward a Pi v1 baseline. The attention record is thin for the star count: three small Hacker News posts between September and October 2026 (3, 2, and 2 points), so the stars ride the NV Labs name and the awesome-list circuit rather than debate. It extends unmodified Pi rather than forking it, which distinguishes it from the rejected third-Pi-derivative family in the harnesses category.
Strengths #
- The mechanisms target measured waste sites (repeated actions, cold tool output, lossy summaries, overflow-time compaction), so each has a reason to exist independent of the paper.
- Verbatim-quote verification on the reducer is the right candor mechanism: compression that claims to preserve evidence can be checked against the quotes it kept.
- It extends Pi without forking it, so Pi upstream improvements keep flowing to its users.
- NVIDIA Research provenance with a paper, a project page, and reproducible benchmark claims.
- MIT and standalone, so the mechanisms are readable as a design menu even for teams on other harnesses.
Cautions #
- Pi-only: every mechanism is implemented against one harness, so teams elsewhere get ideas, not code.
- Young (one month at verification) with 78 open issues already, so the extension surface is still settling.
- The efficiency claims come from the authors’ own auto-research benchmarks, reproduced by nobody independent yet, and the HN footprint is three small posts.
- Compaction and paging change what the model sees, so behavior differences are possible even when token counts improve.
- The paper’s loops are research workflows; ordinary feature work may never hit the loop lengths where the wins appear.
Pricing #
Free and open source under MIT. No hosted tier and no paid plan; the cost is the Pi harness you already run.
Compared to #
- LLMLingua: Microsoft’s prompt-compression line compresses text for any LLM call; SoL-Pi’s mechanisms are structural (fusion, paging, verified reduction) and live inside one harness’s loop.
- Headroom: the receipt-backed compression layer for tool output in general; SoL-Pi is narrower (Pi only) but adds the verified-reducer idea Headroom lacks.
- qmd: the session-search tool that answers from archives after the fact; SoL-Pi keeps the live window small during the loop instead.
Bottom line #
Recommended for Pi users running long autonomous loops who want token cost flat in loop length without losing verifiable evidence. Not for other harnesses (port the ideas instead), or for teams needing independently reproduced efficiency numbers before adopting a month-old extension.
Changes #
- 2026-10-13 - Created.
See also #
- Context Engines Feature Matrix - the category comparison this note joins
- LLMLingua - the harness-agnostic compression line
- Headroom - the receipt-backed tool-output compressor
- qmd - the session-archive search counterpart
- Pi - the harness this extension builds on
References #
https://github.com/NVlabs/SoL-Pi - repository, MIT license, stars, forks, push activity, and the four-mechanism structure as of 2026-10-13
https://arxiv.org/abs/2609.20519 - the paper grounding the four mechanisms and the auto-research benchmark claims
https://nvlabs.github.io/SoL-Pi/ - the NVIDIA project page with the reported efficiency results
https://news.ycombinator.com/ - Algolia search returned three small SoL-Pi posts (3, 2, and 2 points, September to October 2026), the recorded thin community footprint