Hyalo synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Orient and retrieve
Hyalo is a Markdown knowledgebase tooling system, not the owner of an enclosing agent's continuing task. This account covers its CLI, shipped maintenance instructions, npm API, Pi integration, configuration and generators, and optional Jev classification helper at Git revision 697f215fe8a90a5c760fceb686f2a321c9db7221, with an analysis cutoff of 2026-10-14. The functional boundary is whole-system, but detailed inspection is selective. SRC-1 supplies implementation and doctrine/design only: no target, evaluator or provider was executed, and no observed-run or causal-experiment evidence was supplied. Host runtimes, deployed vaults and provider internals remain outside the frozen source. See boundary and evidence and the Source register.
The claimed work is persistent wiki retrieval and structural maintenance without requiring the agent to manipulate YAML unaided. An ordinary command loads configuration, prepares selection and index state, dispatches a deterministic handler, and returns results, diagnostics and an exit status. The user or agent owns the next query, read or repair. Shell, npm and Pi callers reach the CLI through different transports; none makes the CLI an autonomous scheduler. SRC-1, README.md, crates/hyalo-cli/src/run.rs, npm/hyalo/src/api.ts; EPI-CLM-tooling-layer, RT-RTE-query-read.
The durable corpus separates natural-language Markdown bodies from symbolic frontmatter. Filters and schemas consume metadata; humans and models interpret returned prose. Optional binary snapshots retain extracted entries, links, tasks and outlines plus BM25 lexical postings. These are derived access products, not learned weights or explanations of why assertions were accepted. Queries select by explicit paths, metadata, sections or lexical ranking and result limits. Retrieval supplies candidate evidence, not endorsement of source assertions. SRC-1, crates/hyalo-core/src/index.rs, crates/hyalo-cli/src/commands/find/run.rs; MEM-OBJ-body, MEM-OBJ-frontmatter, MEM-OBJ-index-entry, MEM-OBJ-bm25.
Snapshot compatibility is not a freshness proof. Reads can repair detected drift in process or fall back to disk without saving the repaired snapshot; explicit index creation and loaded-index mutation reconciliation can persist updates. Metadata-based drift detection does not authenticate unchanged bytes or coordinate every external writer. Read fidelity also has a concrete boundary: oversized or invalid-UTF-8 body lines become visible placeholders. Line positions can survive while supporting evidence does not. SRC-1, crates/hyalo-cli/src/prepared.rs, crates/hyalo-cli/src/commands/read.rs; MEM-RTE-snapshot-lifecycle, EPI-RTE-body-decoding.
Pi adds opt-in session-start orientation: extension code supplies a hidden aggregate summary without a foreground request. Its once-per-process flag is set before fetching, so a failed fetch is not retried by that handler; delivered context has no automatic refresh. Explicit find/read remains the reload route. The summary is not a Hyalo-retained continuation checkpoint. The long tidy instruction requests forked context, but the host owns isolation and return semantics. SRC-1, pi-package/extensions/hyalo.ts, crates/hyalo-cli/templates/skill-hyalo-tidy.md; RT-RTE-summary-injection, RT-RTE-tidy-review.
Diagnose and delegate selectively
The representative workflow is an explicitly requested Jev-assisted tidy. It is connected by shipped instructions rather than assembled from unrelated capabilities. The agent establishes authorization and vault scope, gathers a baseline summary, schemas, task/status selections, Git activity and structural diagnostics, and distinguishes defects from legitimate exceptions. Audits prescribe proposals only; repair requests prescribe relevant edits without deletion or unrelated changes. These are instruction-level boundaries, not filesystem isolation. The Jev branch uses live reads and skips index creation/removal even in repair mode. SRC-1, crates/hyalo-cli/templates/skill-hyalo-tidy.md, crates/hyalo-cli/templates/jev/jev.md; workflow account, RT-RTE-tidy-review.
Jev must be explicitly requested for the session, and its executable network gate requires --allow-network. The agent supplies local category descriptions, conventions and at most 25 exact Markdown selections. Read-only preparation builds a manifest; each body or section is nonempty and at most 18,000 bytes. Oversized evidence defers rather than being automatically summarized. The service receives selected text and questions under model identifier jev-1.13.0, not separate local-path/current-frontmatter fields; secrets embedded in text can still travel. This separates classification from the foreground context before intended-suggestion review, but establishes neither measured context savings nor provider deletion. SRC-1, npm/jev/src/prepare.ts, npm/jev/src/protocol.ts; RT-RTE-jev-suggestions, RT-OBJ-document-content, RT-CMP-jev.
Semantic evaluation, response coherence and advisory admission are distinct steps. Code checks model and answer identity, labels and numerical consistency. Choice suggestions require confidence at least .85 and winning probability at least .9; tag probabilities at least .9 suggest addition and at most .1 yield no-change. These expressly uncalibrated thresholds grant access to the suggestion channel, not acceptance of correctness. No actual proposition-level output settles whether semantic classification is entailed interpretation or non-entailed conjecture. SRC-1, npm/jev/src/decisions.ts, crates/hyalo-cli/templates/jev/jev.md; EPI-RTE-jev-evaluation, EPI-RTE-response-check, EPI-RTE-score-admission.
The helper never repairs the vault. Per-document failures produce invalid/unavailable rows and deferrals while other documents continue. An existing type can still produce a folder suggestion through an explicit local mapping after an independent provider failure; a failed classification cannot supply a missing type. The reviewer therefore needs the decision ledger, not just the process exit status. SRC-1, npm/jev/src/provider.ts; EPI-RTE-folder-mapping, RT-RTE-jev-suggestions.
Review, publish and report
Substantive review remains prescribed agent behavior. Before using a suggestion, the agent must inspect full current evidence, re-prepare and compare fingerprints, and check the complete effective schema and destination. Unknown required status/date/section values or ambiguous rewrites force deferral even after a high score. Intended use is scoped type/tag/location repair, not invention of completion evidence or taxonomy. Fingerprints establish supplied-input consistency, not authenticated origin or an atomic live read/write guard; section hashes cannot certify the rest of a document. SRC-1, crates/hyalo-cli/templates/jev/jev.md, npm/jev/src/protocol.ts; EPI-RTE-suggestion-review, RT-OBJ-manifest.
Ordinary mutations provide mechanical publication controls. For set, assigned-property constraints reject covered writes only when --validate or configured validation is active; this is not complete-document validation. Captured-source identity/byte checks protect prepared publication, but serial writes are not a batch rollback transaction. Earlier effects can survive a later failure. Loaded-index reconciliation refreshes effects or attempts invalidation, whose failure is reported. Removal has a distinct admission path and does not inherit set-specific validation. SRC-1, crates/hyalo-cli/src/commands/set.rs, crates/hyalo-cli/src/commands/apply.rs, crates/hyalo-cli/src/commands/remove.rs; RT-RTE-property-write, MEM-RTE-metadata-withdrawal.
Structural checks have narrower licenses than semantic judgments. Schema diagnostics establish rule-relative conformity; date normalization can preserve an encoded date, whereas defaults, type inference and nearest-enum repair do not establish factual values. Link destination guards check mechanical resolution; fuzzy/basename guesses require opt-in and a floor but can still select an unrelated note. Separate fragment-repair checks are acknowledged by report verification, not completely inventoried by the destination records. SRC-1, crates/hyalo-mdlint/src/schema.rs, crates/hyalo-core/src/link_fix.rs, crates/hyalo-core/src/anchor_fix.rs; EPI-RTE-schema-check, EPI-RTE-metadata-repair, EPI-RTE-link-guard, EPI-RTE-link-admission, report verification limits.
Pi appends lint feedback after covered host write/edit and typed set/task effects. It does not veto publication, enforce correction, or cover generic Hyalo and shell mutations. The maintenance procedure asks for lint, link/original-finding rechecks and a fresh non-indexed summary comparison. Delivery is a report of actual changes with reasons, unresolved issues and deltas; audit delivery contains proposals. This is not an enforced all-green barrier, and report retention or user receipt is unobserved. SRC-1, pi-package/extensions/hyalo.ts, crates/hyalo-cli/templates/skill-hyalo-tidy.md; RT-RTE-pi-feedback, workflow account.
Assessment at this boundary
For a scoped maintenance request, the strongest supported contribution is executable retrieval, structural diagnosis and mechanically guarded publication, coupled with a conservative instructed review procedure. For the Jev-assisted scenario, helper code implements bounded classification and failure isolation while leaving substantive adoption with the agent. For concurrent or partially failing repairs, effect reporting matters more than assuming transactional success. For establishing the truth of retained assertions, neither lint, target resolution, summary deltas nor provider scores supplies a general acceptance oracle. These conclusions are route-relative, not a product ranking or system-wide epistemic grade.
The learning-related positive is a prescribed trace-fed durable update: Git/project activity informs agent-selected metadata/link repairs that later queries and reads can consume. It remains claimed, not exercised or shown beneficial. Snapshot and navigation generation instead derive access products from documents. Automatic rule-generated metadata and navigation revisions are wired, but revision alone does not establish improved capacity for future action. SRC-1, crates/hyalo-cli/templates/skill-hyalo-tidy.md, crates/hyalo-cli/src/commands/madr.rs; RT-RTE-tidy-review, MEM-RTE-navigation-generation, profile verification.
Criticism of a consumed theory improving future capacity is independently unestablished: the records supply maintenance criteria, not a demonstrated content-directed method-criticism and retained-next-round chain. They do not provide separate assessments establishing theory-builder conditions 1–4, a reflective-system capacity, or autonomous theory building. Those properties remain unassessed here rather than absent; host control and model opacity do not settle them. Configuration publication and installed guidance can change later behavior, but are not evidence-responsive self-improvement merely because they persist. For the selected tidy pathway, exercised self-improvement is unestablished: no trace closes the links from objective-bearing evidence through a change to Hyalo's own organization to causally dependent subsequent operation. Successful improvement additionally lacks outcome comparison. SRC-1, crates/hyalo-cli/templates/skill-hyalo-tidy.md, crates/hyalo-cli/src/commands/types.rs; RT-RTE-tidy-review, RT-RTE-config-install.
A retained end-to-end run could alter the operation assessment by connecting request, evidence, rejected/adopted suggestions, published effects, rechecks and later consumption. Calibrated rubric-relative comparisons would change the semantic-accuracy assessment; independently varying Jev or review would be needed for component attribution. A retained criticism that changes method reliance and shapes another round could support theory-building conclusions. Mandatory current-evidence/whole-document admission and broader write-hook coverage would change the present enforcement boundary. These are resolving evidence or system changes, not inferred features. Source reads for this synthesis: none.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Incomplete cross-command inventory, carried from report verification | SRC-1; MEM-RTE-navigation-generation, MEM-RTE-metadata-withdrawal, RT-RTE-config-install, memory coverage, report limits | Traced MADR, set/remove, snapshot and configuration; detailed OKF selection, other mutations, lint-rule/deinit publication and specialized log/changelog consumers remain incomplete | Complete cross-command admission/recovery comparison or a verdict on every possible trace-fed update | Accepted route records tracing the remaining producers, admission controls, publication/recovery and later consumers |
| Omitted fragment-specific inventory, carried from report verification | SRC-1; EPI-RTE-link-guard, EPI-RTE-link-admission, report limits | Destination repair is characterized; separate anchor publication checks captured source bytes and current target headings, deferring conflicts | Complete link/anchor checking and recovery comparison; omission does not invalidate destination-versus-intent findings | Analyst records for the anchor route and its checks, admission, conflicts and consumers |
| No observed operation or causal comparison | SRC-1; Source register, RT-RTE-tidy-review, RT-RTE-jev-suggestions, workflow account | Frozen implementation and instructions only; no executed target, tests or provider calls | Successful maintenance, compliance, user receipt, accuracy, learning benefit or component effect | Inspectable provenance-linked runs; outcome and intervention comparisons at the claimed scope |
| External runtime/provider and initial-content unknowns | SRC-1; MEM-OBJ-body, MEM-OBJ-frontmatter, RT-CMP-jev, RT-CMP-foreground-agent, RT-OBJ-manifest | Caller-owned corpus and repository-side integrations; host/provider internals and deployed contents excluded | Authenticated source warrant, per-content human approval, exact foreground model, provider weight fixity or actual optional-artifact reuse | Provenance/admission evidence and separately bounded host/provider/deployment records |
| Context freshness and isolation are conditional | SRC-1; RT-RTE-summary-injection, RT-RTE-tidy-review, RT-RTE-jev-suggestions | Once-per-process aggregate supply, requested host fork and separate bounded helper calls | Guaranteed fresh orientation, host eviction/compaction, fork disposal or measured context savings | Host traces of context supply, isolation, retention, refresh and reload under named settings |
| Structural admission does not establish semantic acceptance | SRC-1; EPI-RTE-score-admission, EPI-RTE-suggestion-review, EPI-RTE-schema-check, EPI-RTE-link-admission | Protocol/score/schema/resolution checks and doctrine-only full-evidence review | Correct classification, true metadata, intended referent or lifecycle integration of accepted claims | Recorded evidence-consuming acceptance against scoped criteria and independent checks of semantic outcomes |
| Publication and feedback guarantees are route-local | SRC-1; RT-RTE-property-write, RT-RTE-pi-feedback, MEM-RTE-snapshot-lifecycle | Serial captured publication, loaded-index reconciliation and selected post-write hooks | Universal rollback, all-writer coherence, complete-document validity or mandatory repair | Broader verified controls or route-specific concurrent/failure evidence covering alternate writers and surviving effects |
| Learning, theory-building and self-improvement remain separate unresolved assessments | SRC-1; RT-RTE-tidy-review, RT-RTE-config-install, profile verification | Claimed activity-to-repair chain and wired caller-driven configuration; no retained criticism/next-round or improvement comparison | Improved future capacity from theory criticism, established conditions 1–4, reflection, autonomous theory building or exercised self-improvement | Property-specific records and traces of consumed theories, criticism, retained iteration, self-representation, organizational update and causally dependent later operation |