SoL-Pi boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

This run examines how SoL-Pi changes a model-driven agent's tool execution, observation replay, diagnostic-log consumption and context maintenance. The target is the standalone extension plus the integration paths shipped in this repository, not the upstream Pi runtime or the research process that originally discovered the mechanisms. The target class is extension or tool mechanism. Its operation serves Pi's model loop and, for diagnostic reduction, issues an additional model call through Pi's registry. The boundary kind is whole-system relative to this repository's shipped SoL-Pi responsibilities, not a claim to cover the enclosing hosts.

The frozen evidence is SRC-1 at 41fa7e0e230bedc5e639aed2e7b9cc5698726db1, with applicability cutoff 2026-10-14. The checkout's HEAD matches that pin, and the tracked-worktree status was clean when inspected. The evidence tier is code-grounded: entry points, state producers, admission checks, persistence, reload and consumption paths were inspected as source. No target code, test, example, model call or service was executed. Test source is implementation evidence for the evaluator's design, not an observed successful run. Compatibility and benchmark statements in documentation remain attributed reports.

Entry points and functional inclusions

The package declares src/sol-pi/index.ts as its Pi extension. That entry reads the effective configuration at session_start and conditionally registers the four mechanisms. This is a host-loaded extension, rather than an enclosing agent runtime (SRC-1, package.json, src/sol-pi/index.ts, src/sol-pi/config.ts).

"pi": { "extensions": [ "./src/sol-pi/index.ts" ] }, --- package.json

export function registerConfiguredFeatures(pi: ExtensionAPI, config: SolPiConfig): void { if (config.actionFusion) registerActionFusion(pi); if (config.observationPack) registerObservationPack(pi); if (config.evidencePreservingReducer) { --- src/sol-pi/index.ts

The included core paths are:

• Action Fusion: replacement edit and write definitions, optional follow-up shell execution, queue/path handling and result presentation. Its caller is the host's model-selected tool invocation; the actual base tools and shell are supplied by Pi (SRC-1, src/sol-pi/extensions/action-fusion/index.ts, src/sol-pi/extensions/action-fusion/then-run.ts, src/sol-pi/extensions/action-fusion/file-queue.ts).
• ObservationPack: eligibility and object identity, local archive creation, projected placeholders, replay accounting, paged obs_recall and JSONL ledger. The context event consumes stored tool-result messages without replacing their persisted history; recall later reads the archived object (SRC-1, src/sol-pi/extensions/observation-pack/index.ts, src/sol-pi/extensions/observation-pack/observation.ts, src/sol-pi/extensions/observation-pack/ledger.ts).
• Evidence-Preserving Reducer: diagnostic candidate selection, raw-log archives, reducer prompt and provider call, receipt admission, replacement tool results and decision journals. The resulting receipt points the agent to an archive path for exact readback; that readback uses the host's ordinary shell rather than a separate SoL-Pi recall tool (SRC-1, src/sol-pi/extensions/evidence-preserving-reducer/index.ts, src/sol-pi/extensions/evidence-preserving-reducer/candidate.ts, src/sol-pi/extensions/evidence-preserving-reducer/archive.ts, src/sol-pi/extensions/evidence-preserving-reducer/provider.ts, src/sol-pi/extensions/evidence-preserving-reducer/receipt.ts, src/sol-pi/extensions/evidence-preserving-reducer/journal.ts).
• Online Context Compact: the model-facing plan tool and its guidance, plan-transition detection, economic/window-pressure selection, session-state append/restore, native compaction callbacks and automatic continuation. State restoration reads the active session branch. Compaction and the subsequent task turn belong to the same inspected extension path, rather than to an assumed external resumer (SRC-1, src/sol-pi/extensions/online-context-compact/tools.ts, src/sol-pi/extensions/online-context-compact/extension.ts, src/sol-pi/extensions/online-context-compact/state.ts, src/sol-pi/extensions/online-context-compact/economics.ts).

export function appendOnlineState(pi: ExtensionAPI, state: OnlineState): void { pi.appendEntry(ONLINE_STATE_ENTRY, state); } --- src/sol-pi/extensions/online-context-compact/state.ts

pi.sendMessage( { customType: "sol-pi-online-context-compact", content: compacted ? POST_COMPACTION_PLAN_REMINDER : SKIPPED_COMPACTION_REMINDER, display: false, }, { triggerTurn: true }, ); --- src/sol-pi/extensions/online-context-compact/extension.ts

Shared configuration, trust-dependent admission, runtime archive roots and TUI reporting are included. Session-directory and in-memory/no-session archive modes are both within scope (SRC-1, src/sol-pi/config.ts, src/sol-pi/runtime-paths.ts, docs/configuration.md, docs/compatibility.md).

The shipped examples are also included, despite not being in the package's default extension entry:

• SoL-Pi Skill: the Python launcher configures an isolated Pi agent directory, starts persistent or detached runs, records prompts/events/session paths, rereads those records for status and cost summaries, and reports the result back through Claude Code or Codex hooks. The transcript exporter selects and trims host conversation content; bundled skills describe a model-mediated fallback when hooks do not run. These are shipped producer/consumer and host-integration responsibilities, not excluded merely because they launch Pi as a subprocess (SRC-1, examples/sol-pi-skill/bin/sol-pi-run, examples/sol-pi-skill/bin/sol-pi-context, examples/sol-pi-skill/probe/sol-pi-probe.ts, examples/sol-pi-skill/codex-hooks/hooks.json, examples/sol-pi-skill/claude-plugin/sol-pi/hooks/hooks.json, examples/sol-pi-skill/codex-skill/sol-pi/SKILL.md, examples/sol-pi-skill/claude-plugin/sol-pi/skills/sol-pi/SKILL.md).
• SoL-Pi Visual: explicit extension loading, trusted-project configuration, model prompt guidance, magnify tool registration, image/receipt production and image readback are included. Blender's scene creation/rendering and visual grading are not implemented by this extension (SRC-1, examples/solpi-visual/README.md, examples/solpi-visual/src/index.ts, examples/solpi-visual/src/scene/index.ts, examples/solpi-visual/src/scene/magnify.py, examples/solpi-visual/src/scene/policies.ts).

write_json(run_dir / 'result.json', result) return result --- examples/sol-pi-skill/bin/sol-pi-run

if (run_dir / 'result.json').exists(): return json.loads((run_dir / 'result.json').read_text()) --- examples/sol-pi-skill/bin/sol-pi-run

External dependencies and excluded conclusions

Pi owns the enclosing agent/session loop, tool definitions, native summarization, provider registry/authentication, extension dispatch and trust API. Node.js, TypeBox and the Pi TUI support the extension; the skill launcher additionally uses Python and host transcript/hook conventions; Visual uses Python/Pillow and external scene/render tools (SRC-1, package.json, src/sol-pi/extensions/evidence-preserving-reducer/provider.ts, examples/sol-pi-skill/README.md, examples/solpi-visual/README.md). Only SoL-Pi's calls and declared expectations of those dependencies are covered. Their implementations are not frozen here, so this run cannot establish upstream runtime correctness, actual host hook delivery, live authentication behavior or cross-version compatibility independently.

The linked paper, blog, pricing pages and benchmark dataset are not part of the frozen source. No raw benchmark trajectories, live-provider results or intervention/comparison runs were supplied as separate evidence. This prevents independently establishing claimed savings, reconstruction quality or the causal effectiveness of the original auto-research process. Repository statements about those results may be cited only as reported operation, not as observed runs or causal experiments (SRC-1, README.md, examples/solpi-visual/README.md, docs/compatibility.md). Missing execution evidence does not make the source inspection incomplete.

Top-level coverage

The table accounts for every top-level tracked file or directory at the pin. “Included” fixes functional scope; it does not assert that every leaf or test assertion has already been inspected. Named uninspected leaves are available evidence, not access gaps. All paths below belong to SRC-1.

Source register