SoL-Pi synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Entering a host-owned task loop
SoL-Pi is a Pi extension that changes execution and model-visible context, not an enclosing agent runtime. This account covers the standalone extension and its shipped Skill and Visual integrations at commit 41fa7e0e230bedc5e639aed2e7b9cc5698726db1, with cutoff 2026-10-14. Its evidence is inspected implementation and attributed design/documentation in SRC-1. No target execution, provider call, successful test run or causal comparison was supplied or performed. Pi owns foreground decisions, base tools, session persistence and native summarization; upstream implementations and the external research process are not frozen here (boundary, Source register). This synthesis uses the accepted members without additional source reads.
On session start, the extension registers enabled mechanisms once. Missing configuration disables all four. Project trust governs whether one project configuration file may be loaded; it is not a general permission boundary, and programmatic factories have caller-owned admission. No deployed grants were supplied. The extension runs with Pi's permissions (RT-RTE-config-admission).
Acting and returning narrower evidence
Action Fusion lets the foreground model propose a file mutation and optional then_run command in one tool call. Code executes the base mutation, samples content stability, then invokes bash without another planning-model decision. This supports the narrow avoided-round-trip claim for that route. It does not make the pair transactional: command failure leaves the mutation in place. Its canonical-path queue serializes cooperating fused wrappers, not all filesystem writers (RT-RTE-action-fusion, RT-CLM-fusion-roundtrip).
A diagnostic result can then enter a separate reduction route. Code selects eligible plain bash output or a marked fused-command suffix, preferring a safely loaded full-output file when available. It archives the selected body and gives one log plus metadata to an isolated reducer model, not the whole conversation or a tool loop. The model proposes quotations and labels; software admits a smaller receipt only after schema, source-hash, host-status and exact quote-membership checks. Ordinary provider or receipt failures retain the original observation. The foreground model retains diagnosis, repair and rerun decisions (RT-RTE-reducer-receipt, RT-CMP-reducer-model).
This is genuine evidence-consuming admission, but only for smaller evidence transport. It is not acceptance of a diagnosis or task correctness. A receipt can contain one failure-labelled matching quote while omitting a second failure. Label types are checked; their semantic appropriateness is not. Raw archives preserve an optional range-read route, not automatic use of omitted context. Thus the prompt's “lossless” description is supported for selected quotations, not for complete delivered evidence (EPI-OBJ-validated-receipt, EPI-OBJ-receipt-labels, EPI-CLM-lossless-reducer, MEM-OBJ-archive-object).
Limiting replay and continuing after compaction
ObservationPack alters provider replay without rewriting persisted conversation history. It archives large successful pure-text results and, after two locally counted full sends, projects a pointer with literal edge excerpts. Errors, image-bearing results and reducer receipts bypass packing. obs_recall supplies bounded pages on model request; it does not search archives semantically. Recall depends on surviving files, roots and identifiers. Exact decoded-text claims additionally require suitable offsets: arbitrary offsets inside UTF-8 characters can corrupt decoding even when storage is intact (RT-RTE-observation-replay, MEM-OBJ-observation, report verification limit). A pointer in context is not automatically dereferenced or recovered after it is forgotten.
Online Context Compact moves context maintenance to nominated plan boundaries. Completing a previously registered unfinished step through update_plan can nominate a boundary; successful tool completion and economic or window-pressure checks select whether to compact. Window pressure can override economic cooldown. Code aborts the loop, waits for idle settlement and requests Pi's native summary. Successful compaction and recognized safe refusals trigger a generic hidden continuation; cancellation and unknown errors have different paths. Manual/native compaction need not pass this selection policy (RT-RTE-online-compaction).
Two retained products must remain separate. Code restores plan/control state from the latest valid active-branch entry and uses trace-derived estimates for later selection. The later model receives Pi's summary/recent tail. Persisted progress assertions are neither independently verified work nor a dedicated task checkpoint automatically injected by SoL-Pi. A preservation instruction does not establish summary fidelity (RT-OBJ-online-state, MEM-OBJ-progress-summary, RT-RTE-online-compaction).
Delegating work and inspecting images
The representative connected workflow is the shipped Skill integration's Codex $sol-pi handoff. Software parses the user request, optionally exports a trimmed host transcript, starts detached Pi work in the shared working directory, records events and a result, then emits a correlated parent notification. Pi/model owns inner task actions; the launcher owns lifetime and process classification; the parent is asked to review changes and report to the user. This is example behavior, not implicit behavior of the default extension (RT-RTE-skill-handoff, workflow account).
The child has a separate conversation and configuration, not an OS sandbox. Workspace effects escape before return. The exported prefix is lossy reference context; its newest block may exceed the target budget, and the Codex adapter degrades error-label provenance. The result selects the last nonblank assistant text and aggregates metadata. Process done, result persistence and notification emission do not verify the product or establish host delivery. Execution timeout terminates the process group without rolling back edits; parent nonconcurrency and review are instructions, not enforced gates (MEM-OBJ-context-export, EPI-OBJ-exported-context, MEM-OBJ-run-result, workflow annotation).
Visual adds a different path: static workflow guidance and model-requested crops of existing images. The helper computes crop images and provenance, checks same-input cached artifacts, and the wrapper returns actual image bytes. This enables targeted inspection without generating new visual detail. It neither grades scene correctness nor enforces local edits. A static receipt flag is not an evaluator result (MEM-OBJ-magnify-artifact, RT-RTE-visual-inspection, RT-BAP-visual-policy).
Supported contribution and unresolved improvement
The strongest supported adaptive contribution is automatic trace-fed construction of retained receipts, continuation context, handoff artifacts and control estimates with later consumers. In the memory comparison's route sense, these support wired trace learning, not demonstrated improvement in capacity. In particular, OCC measurements influence later context-selection decisions. Raw archive retention alone does not establish such an update, and static policies are not inferred to have learned from outcomes (RT-OBJ-online-state, MEM-OBJ-validated-receipt, MEM-OBJ-context-export, MEM-OBJ-run-result, profile verification). The unresolved question is whether these changes improve later action, evidence use or cost at a specified task horizon.
The accepted runtime records do not supply a separate theory-builder conditions 1–4 assessment. They establish inspectable static guidance and model-facing delivery, but not reliable content-dependent model decisions, content-directed criticism and retained criticism-driven iteration over that guidance. No criticism of a consumed theory is shown to improve future capacity. OCC's retained measurements and later selector establish a bounded feedback path; they do not alone establish a reflective theory builder. Computational dispatch and admission are wired, but an autonomous theory builder is not established. Exercised self-improvement is likewise unestablished: subsequent operation causally depending on an improvement-directed organizational change is not observed. These are independent limits, not a bundled negative or an inference that opaque models cannot learn (RT-BAP-visual-policy, RT-BAP-reducer-instructions, RT-CMP-frontier-model, RT-OBJ-online-state).
For repeated large-text observations, the supported advantage is narrower replay with explicit reload. For verbose diagnostics, it is constrained evidence selection with a surviving source path. For long planned work, it is conditional compact/continue coordination. For delegated tasks, it is a separate context with durable handback artifacts. None establishes measured savings, successful repair, complete evidence use or reconstruction quality. Inspectable traces of recall, compaction settlement, parent delivery and task verification would establish activation at those routes. Controlled task comparisons with provider/configuration and workload conditions fixed would change benefit assessments; independently varying mechanisms would be needed for component attribution. Stronger completeness or safety guarantees would require changes such as semantic coverage checks, offset-safe recall, independently checked summaries or pre-effect admission/rollback, rather than stronger names for existing receipts and process statuses (EPI-CLM-preserve-evidence, RT-RTE-reducer-receipt, RT-RTE-online-compaction, RT-RTE-skill-handoff).
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No execution or causal comparison; selective evaluators are source, not test results | SRC-1; RT-CLM-fusion-roundtrip, EPI-CLM-preserve-evidence, RT-RTE-skill-handoff | Frozen extension and shipped examples; external benchmark trajectories not supplied | Exercised reliability, actual savings, task quality and component effects | Inspectable route/task traces and controlled comparisons with attribution at the treatment's level |
| Arbitrary byte offsets can start inside UTF-8 characters; recall does not recheck the whole hash per page | SRC-1; RT-RTE-observation-replay, MEM-OBJ-observation, EPI-OBJ-observation-text; report verification limit | Observation helper and recall path at the pin | Exact decoded text for arbitrary offsets, or tamper-proof readback; intact assembled storage and offset-zero/returned-offset paging remain supported | Character-boundary-safe start handling and integrity revalidation, with bounded paging checks |
| Reducer checks narrower than semantic completeness; archive input can be an inline preview | SRC-1; RT-RTE-reducer-receipt, EPI-OBJ-validated-receipt, EPI-OBJ-receipt-labels | Candidate, validator and replacement paths | All-failure coverage, correct labels, complete original log or accepted causal diagnosis | Full-output provenance and domain-specific semantic coverage/diagnosis evaluation |
| Archive/root lifetime and exceptional failures bound preservation | SRC-1; MEM-OBJ-observation, MEM-OBJ-archive-object, RT-RTE-reducer-receipt | Persistent and process-local temporary roots; provider fallback excludes some storage/journal exceptions | Indefinite recall, universal fallback or automatic cross-process temporary-root restoration | Lifecycle/recovery contracts and traces covering deletion, restoration and exceptional writes |
| Native summary internals and persistence remain external; progress is not a dedicated prompt checkpoint | SRC-1; RT-RTE-online-compaction, MEM-OBJ-progress-summary | SoL-Pi request/restore/continuation interfaces | Summary fidelity, complete memory substrate/selection inventory, verified work or automatic progress replay | Frozen Pi implementation plus committed summary/input/continuation traces and fidelity checks |
| Host transcript provenance and final delivery are not verified | SRC-1; MEM-OBJ-context-export, EPI-OBJ-exported-context, MEM-OBJ-run-result, RT-RTE-skill-handoff | Export adapters, launcher and hook emission | Faithful error provenance, hard total export budget, task acceptance or successful user receipt | Provenance-preserving adapters and bounded export checks; host/parent traces with independent product verification |
| Configuration/context separation does not confine effects | SRC-1; RT-RTE-config-admission, RT-RTE-action-fusion, RT-RTE-skill-handoff | Default loader, factory alternatives, fused queue and shared-cwd launcher | Deployment isolation, exclusive mutation access, transactional rollback or pre-change parent approval | Actual grants and enforced isolation/admission/rollback contracts covering alternate writers and entry paths |
| Image provenance guards are not visual evaluation; original image admission remains unresolved | SRC-1; MEM-OBJ-magnify-artifact, RT-RTE-visual-inspection, RT-BAP-visual-policy | Crop helper, cache, image delivery and static guidance | Scene correctness, hidden geometry, policy compliance or complete image-origin classification | Original image provenance, defined visual evaluator and crop-consumption/outcome traces |
| Provider parameters, criticism-driven revision and improvement outcomes are unestablished | SRC-1; RT-CMP-frontier-model, RT-CMP-reducer-model, RT-BAP-visual-policy, RT-OBJ-online-state | Extension call/feedback interfaces, not model internals or original research builder | Pinned-weight or training claims, theory-builder completion, criticism-improved capacity and exercised self-improvement | Pinned provider evidence and retained criticism/update/consumer records with subsequent operation and objective-relative comparisons |
The reconciliation reports no unresolved conflict. Profile verification declares no additional limit; its partial inventory supports existence comparisons, not complete-value statistics (profile verification).