llm-wiki synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
A knowledge service inside an external agent's work
llm-wiki makes a Markdown knowledge bundle available to later agent work. It supplies lexical retrieval, concept writes, navigation maintenance, directory rules and synchronization. It does not run the enclosing model loop: the external host decides when to consult the bundle, investigates domain questions, proposes semantic content and produces the final domain answer. The reviewed whole-system boundary covers the repository's knowledge-system responsibilities, including DSH, pi, MCP, Codex and skill/CLI adapters, maintenance scripts and evaluation material. It excludes external host/model internals and live service state. Evidence is source code and attributed design from SRC-1 at commit f70e680d852a78d48504425d87252531e68c5624, with applicability cutoff 2026-11-08. No execution trace, passing test run or causal experiment was supplied (boundary, Source register, RT-CMP-core).
The shipped forms expose a common core but not identical operational behavior. CLI returns stdout; DSH registers tools, context assembly and management routes; pi registers tools and a prompt hook; MCP supplies tools and initialization instructions; Codex combines a stdio tool server with context hooks. External hosts own scheduling and model interpretation. Frozen vendor comparisons support shared semantics for inspected core files, not installed-package parity or equivalence of wrappers. Rule refresh, formatting and online guards differ materially (runtime entry paths, RT-CLM-form-equivalence).
From request to selected knowledge
The representative procedure is skill-guided rules→list→search→get→create. The host first requests rules and an overview, chooses search terms, then retrieves a selected concept. This is a coherent source-prescribed workflow, not an executable sequence constraint or an observed session. Commands resolve the bundle independently; continuity depends on the caller retaining the intended selection. No internal model worker receives a delegated investigation (workflow account).
Files are the retrieval source of truth. Core rebuilds a graph from nonhidden, nonreserved Markdown, deriving links and backlinks. Search uses lexical matches, selected metadata, filters and a result limit; get resolves IDs, unique titles or unique prefixes. Search returns summaries rather than full passages. Get supplies the chosen body and backlink references, not the bodies or explanations of all linked concepts. A later get or separate file read is needed to obtain those reasons. Neither a failed search nor a high score establishes absence or truth (RT-RTE-bundle-read, MEM-OBJ-concept-body).
Stored provenance is richer than common model-facing responses. CLI/MCP get omit many structured sources, verification, freshness and computation-binding fields; MCP also caps output. A concept can therefore reach the host without the stored evidence needed to interpret its trust label or attester pointer. Verified, stale and deprecated states are not universal search/get exclusion criteria. Retained text acts as knowledge at the interpreting host and as ranking input at lexical search; those are distinct effects, neither an endorsement (MEM-OBJ-concept-metadata, EPI-OBJ-concept).
From investigation to retained content
When knowledge is missing, the skill asks the external agent to investigate and record facts or diagnosed pitfalls. This is the strongest supported experience-reuse contribution: caller-authored explanations can persist across invocations and become available to later work. The demo pitfall includes a mechanism, not merely a warning label. However, the records do not identify a qualifying recorded-trace input and automatic semantic producer, demonstrate later reliance on a particular new item, or show improved future capacity. The experience-recording affordance survives those uncertainties; neither trace learning nor its absence is established (EPI-CLM-knowledge-recording, MEM-OBJ-concept-body, memory write side, profile verification).
Creation has an executable but scoped confirmation protocol. The nearest AGENTS gate declaration determines which types require a caller-supplied confirmation flag. Covered unconfirmed creates are refused; a cooperating host can show the proposal, obtain consent and retry. The flag is not an authenticated approval record. Core then checks structural preconditions and writes the concept before log/index maintenance. A later maintenance failure can leave the concept durable. Updates can change body or type without the create gate; ingest, migrations, exported core calls and direct editing provide other admission routes. Thus the create veto is real, but it is neither universal human approval nor factual acceptance. Earlier verification entries can survive an unconfirmed body revision (RT-RTE-concept-create, RT-RTE-concept-update, EPI-OBJ-trust-fields, MEM-RTE-bundle-migration).
Source ingestion is registration, not semantic synthesis. It copies a new destination, then conditionally creates a small Reference page if stem-based concept lookup finds nothing. A readable copied Markdown file can itself satisfy that lookup, so ingest does not uniformly produce a raw/derived pair. An existing destination is not refreshed. The optional page takes a short first-line excerpt from the current external input; it can omit qualifications or differ from an older retained destination. New raw bytes can be preserved while source warrant remains unknown and the excerpt's semantic relation remains indeterminate (RT-RTE-source-ingest, MEM-OBJ-reference-source, EPI-OBJ-derived-reference).
Checks, maintenance and later behavioral influence
Validation and lint derive bounded facts about parsed structure, graph links and dates. They return diagnostics; they do not check domain truth or automatically accept and revise knowledge. CLI validation can fail its own command, but neither diagnostic is mandatory before ordinary writes. Index regeneration reorganizes navigation; standalone refresh does not imply a change-log entry. Deprecation retains bodies and marks reliance as discouraged without preventing ordinary retrieval. None of these operations establishes evidence-consuming acceptance of a new domain claim or post-acceptance integration (EPI-RTE-format-check, EPI-RTE-health-check, MEM-RTE-index-refresh, MEM-RTE-directory-deprecation).
The “Attested Computation” type does not close that gap. The SQL equality attester is a demo specification, with no implemented receipt-to-verdict consumer established in inspected paths. Even a realized equality check would warrant executed-versus-sanctioned SQL agreement, not business-definition or data truth. Smoke assertions are test expectations, not observed outcomes. Cloud migration's Markdown-only selector can omit SQL assets while retaining pointers to them (EPI-RTE-attester-check, EPI-OBJ-smoke-checks, EPI-RTE-cloud-migration).
Retained rules have a separate behavioral path. Lifecycle injection broadly collects APPEND text; CLI get/create/update independently append rules selected by the operated-on directory's ancestor chain. Requested concept retrieval is pull, while the unsolicited rule appendix is a separate push. Broad lifecycle selection and directory-targeted appendix selection must not be conflated. pi rereads per hook, DSH can use cached snapshots, MCP initialization text is static after construction, and Codex suppresses unchanged rendered text except on SessionStart. Caps, errors and compact AGENTS extraction can omit material. MCP get does not provide the rule appendix promised by its documentation; explicit rules calls remain available. These mechanisms deliver by-value text, not live references, host-turn compaction or isolated reasoning frames. Delivery supports an instruction channel, not demonstrated obedience (RT-RTE-context-delivery, RT-BAP-rule-delivery, profile verification).
Git/Feishu synchronization extends persistence and access, not semantic warrant. Local writes and remote publication have separate outcomes. Selected prewrite guards can fail open on exceptions, and pi update differs from guarded wrappers. Ordinary Feishu no-delete synchronization is not a guarantee about Git tracked deletions or explicit Feishu deletion. Explicit deletion can have partial irreversible effects; deprecation, deregistration and ledger forgetting have different, narrower effects. Backups and history afford recovery without proving successful restoration or recall of text already delivered to a host (RT-RTE-remote-sync, MEM-OBJ-wiki-cloud, MEM-OBJ-wiki-registry, report verification).
What the architecture supports, and what would change the assessment
For a host retrieving reusable domain advice, the discriminating capability is persistent, editable text with lexical access, backlinks and complementary rule delivery. For a scenario requiring every write to be human-approved, every retrieved proposition to carry current warrant, or every retired concept to be excluded, the inspected routes do not supply those guarantees. This is a scenario-relative distinction, not a product ranking.
No accepted internal semantic theory-building loop is established. The retained demo explanation supplies localized content, but storage and delivery alone do not establish theory-builder condition 2, dependence on its content; conditions 3 and 4 require content-directed criticism and retained criticism shaping another round, which the supplied records do not establish. The runtime does not supply a complete conditions 1–4 assessment, so no system-wide theory-builder verdict follows. Separately, criticism improving future capacity, reflection's bidirectional self-representation, autonomous internal theory-building and exercised self-improvement remain unestablished. Editable rules provide a potential later instruction path; they do not show evidence determining a self-update or subsequent behavior causally depending on it (MEM-OBJ-concept-body, EPI-CLM-knowledge-recording, RT-BAP-rule-delivery, epistemic bounded conclusion).
Linked proposal, evidence, approval, revision and later-consumption traces would permit stronger judgments about exercised recording and criticism. An independently specified improvement objective and a comparison tracing changed retained content into later behavior would be needed for capacity or self-improvement attribution. System changes such as consistent admission across mutation routes, version-bound approval, provenance-preserving responses or consumer-enforced withdrawal would change the corresponding guarantee assessment; their effectiveness would still require evidence. Neither source code alone nor a successful domain answer would isolate the contribution of any one mechanism.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No observed execution or causal comparison | SRC-1; RT-RTE-bundle-read, RT-RTE-concept-create, EPI-OBJ-smoke-checks | Frozen code, doctrine and synthetic examples; no target execution | Operational reliability, factual accuracy, retrieval effectiveness and downstream benefit | Inspectable workload traces and outcome comparisons with stated conditions and appropriate independent variation |
| Host behavior and actual approval remain external | SRC-1; RT-RTE-concept-create, RT-BAP-rule-delivery, EPI-CLM-knowledge-recording | Target-side flags, metadata and instruction channels, not consent or model traces | Authenticated approval, rule obedience, exercised experience reuse, autonomous criticism or self-improvement | Content-linked consent and investigation records, host input/output traces, retained revisions and later causal dependence |
| Stored provenance and delivered evidence differ | SRC-1; MEM-OBJ-concept-metadata, EPI-OBJ-derived-reference, RT-RTE-context-delivery | Common get renderers, conditional excerpts and capped/cached rule builders | Complete provenance or qualification delivery; current verification of revised bodies; universal rule fallback | Consumer-visible responses linked to exact stored versions, source/proposal comparisons and demonstrated alternate reads where needed |
| Raw origin and semantic transformation are unresolved | SRC-1; MEM-OBJ-reference-source, MEM-OBJ-concept-body, EPI-OBJ-derived-reference | Arbitrary copied payloads and caller-authored content | Complete encoding/lineage classification, warranted source propositions, trace-learning classification or accepted ampliative knowledge | Original inputs, authorship history, transformation traces and evidence-consuming decisions with criterion, intended use and scope |
| Remote services and exhaustive reconciliation options are not established | SRC-1; RT-RTE-remote-sync, MEM-OBJ-wiki-cloud | Inspected Git/Feishu branches and local ledger/backups; no live service evidence | Universal sync safety, race freedom, atomic recovery or successful remote publication | Scoped service traces, concurrent/conflict cases and recovery evidence for the relevant options |
| Attestation and external conformance remain bounded declarations | SRC-1; EPI-RTE-attester-check, EPI-RTE-format-check, EPI-RTE-cloud-migration | Demo attester, local parser predicates and Markdown-only migration | Running attestation, domain truth, independent OKF conformance or complete computation transfer | Implemented consequential receipt/verdict route, checked premises, pinned specification comparison and transferred-asset inventory |
| Whole-system scope does not certify every available surface | SRC-1; boundary coverage, RT-CLM-form-equivalence, report verification | Selected material paths across all forms; packaging/UI, translation parity, other tests/examples and remote options remain uncertified | Installed-form equivalence, exhaustive evaluator absence or complete behavior inventory | Targeted inspection of the unresolved surface needed for each stronger conclusion, plus execution evidence where operation is claimed |