NVlabs/SoL-Pi

What it solves

SoL-Pi addresses the inefficiency of long-running coding agents that accumulate redundant work, oversized observations, and bloated context windows. It reduces token traffic, inference costs, and the number of agent turns without sacrificing the evidence or verification steps necessary to complete a task.

How it works

It functions as a standalone extension for the Pi coding agent, implementing four efficiency mechanisms:

• Action Fusion: Combines a file edit or write with its subsequent validation command into a single tool call.
• ObservationPack: Replaces repeated large text results with stable handles that allow for exact paged recall.
• Evidence-Preserving Reducer: Compresses long diagnostic logs into compact receipts, ensuring every retained quotation matches the original source.
• Online Context Compact: Identifies completed plan steps as candidates for context compaction to reduce window pressure, starting a new turn after successful compaction.

Who it’s for

Developers using the Pi coding agent who want to reduce LLM operational costs and improve performance by optimizing how the agent handles context and tool outputs.

Highlights

• Non-intrusive: Installs on top of an unmodified Pi release via public APIs without requiring patches to the Pi source code.
• Opt-in: All efficiency mechanisms are disabled by default and must be explicitly enabled in configuration.
• Evidence-focused: Ensures original observations remain available and prevents data loss during reduction.
• Auto-research derived: Based on mechanisms discovered through scaled auto-research loops to optimize agent harnesses.