SoL-Pi is a standalone extension for the Pi agent framework that packages efficiency mechanisms for reducing token usage and inference work in long-running coding agents.
Long-running agents accumulate inefficiencies: file edits followed by predictable validation commands, large tool results replayed repeatedly, completed subtasks remaining in active context, and frontier models spending full requests reading logs when only a few lines matter for the next decision. SoL-Pi addresses these problems through four composable mechanisms that operate at different parts of the agent harness. Action Fusion allows edit or write operations to run their follow-up validation command in the same tool call. ObservationPack converts repeated large text results into stable handles with exact paged recall. The Evidence-Preserving Reducer compacts long diagnostic logs into compact receipts only when every retained quotation matches the archived source. Online Context Compact identifies completed plan steps as candidates for Pi's native compaction, subject to economic and window-pressure checks, allowing the agent to continue in a new turn. These mechanisms reduce repeated model turns, context replay, oversized observations, and unnecessary log reading while preserving the work and evidence agents need to complete tasks.
The tool suits teams running long-horizon coding agents where token efficiency and inference cost matter. Each mechanism is opt-in and disabled by default, so adoption can be incremental. The extension installs on top of an unmodified Pi release and uses Pi's public extension APIs, making it compatible with existing Pi deployments without requiring framework modifications.
The project maintains focused development on the four core mechanisms discovered through scaled auto-research loops. Documentation covers configuration details and the mechanisms are designed as composable, independent features rather than a monolithic system. The work is grounded in a research question about making agent harnesses more efficient before scaling agent loops themselves, with emphasis on constrained efficiency that avoids stopping early, skipping verification, or hiding evidence from the agent.
