SoL-Pi

Type: types/note.md

Evidence basis: source code and repository documentation at NVlabs/SoL-Pi commit 1559b5c, inspected 2026-09-24; no live execution or causal efficiency experiment.

SoL-Pi is an extension for Pi that packages four optional efficiency mechanisms. All are disabled by default. It owns tool wrappers, context transforms and compaction scheduling; Pi owns the main agent loop, permission envelope, model transport and native summary generation. Its README attributes the mechanisms to historical auto-research, but that research process is outside this release boundary.

Action Fusion lets a model request an edit or write followed by a shell command without an intervening model call. A per-file queue serializes this extension's operations, and two hashes around a yield detect some intervening changes. Mutation failure skips the command; command failure leaves the mutation in place. This is neither transactional rollback nor a global filesystem lock. Implementation.

ObservationPack archives eligible large successful text results, keeps their first two full context sends, and then projects a handle with a bounded excerpt. obs_recall retrieves bounded exact slices from the current session archive. The stored history and outgoing projection are different objects: shortening one does not erase the other. Recall availability still depends on the local archive and host session contract. Projection implementation.

The Evidence-Preserving Reducer archives eligible diagnostic logs and asks a configured model to select quotations. It checks source identity, status, schema, quotation occurrence and size before replacing the tool result with a shorter receipt. These checks establish provenance of selected text, not completeness or correct diagnostic labels. The receipt leaves diagnosis and repair with the main agent and affords raw-log readback through Bash. Handled model/validation failures preserve the original result; archive errors occur outside that fallback and depend on Pi's hook handling. Reducer, receipt checks.

Online Context Compact retains plan status and request/context statistics. Completed plan steps can trigger a cost/window calculation, followed by native compaction after the current run settles. Success triggers a new turn with a generic instruction to rebuild the remaining plan. Plan completion is model-declared, not independently verified. Retained progress wording has no established later model reader on the custom-state route; native summary contents and replay selection remain host-owned. Compaction control.

The strongest supported result is the wiring of recoverable evidence and adaptive context scheduling. Diagnostic receipts, native compaction continuation and persistent scheduling statistics meet the analysis's trace-learning criterion, without establishing improved capacity. The economic formula is an operative, inspectable prediction; runtime updates change its inputs, not its content. Criticism-driven improvement and self-improvement of the released extension remain uninspected. Narrow operational reflection is wired: prior compaction count and cache debt represent its own operations and change later scheduling. This does not establish a reflective theory builder.

Compatibility documentation reports deterministic fake-provider checks and explicitly excludes live token-savings claims. Measured savings, preserved task quality and downstream dependence on recalled evidence require additional execution evidence. The extension also supplies no sandbox: it runs with Pi's permissions.