llm-wiki-cli synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From agent work to retained knowledge
LWC is an agent-operated memory and context system, not an enclosing inference runtime. External agents reason, select evidence and author knowledge; LWC preserves sources, replaceable Wiki pages, citations, history and selection state. This account covers the shipped CLI, agent instructions, host adapters, optional learning stores, replication and readers at Git commit e1be6ab30e83947e4b0dba97ba910acfe3bd9821, with an applicability cutoff of 2026-11-08. The evidence is source code and doctrine from SRC-1, with limited attributed reporting, not a supplied execution of the write-and-recall loop or a causal comparison. Whole-system inclusion does not mean every included transformation was deeply inspected. See the boundary and evidence, Source register and RT-CMP-host-model.
The ordinary progression begins with an external agent recalling relevant knowledge for a user task, checking mutable claims against live evidence, doing the work and capturing a verified milestone. Canonical Wiki content resides in SQLite; materialized Markdown and graph/index structures are projections. Immutable source snapshots remain separately addressable from agent-authored pages. Citation linkage preserves a route to evidence, not an endorsement of the page's propositions. Explicit provenance can also admit durable observations without immutable source IDs. See RT-CLM-agent-memory, MEM-OBJ-sources and MEM-OBJ-pages.
Source ingestion is the representative multi-step workflow because shipped instructions connect its steps. After reviewed source admission, the agent claims an ingest job and receives purpose/schema, bounded page summaries and source windows. The agent submits analysis, writes a cited source summary and writes a cited integration page, or supplies the permitted specific no-derived-pages reason. Claim and analysis transitions persist state; they do not launch a reasoning worker. Completion counts citation joins and page kinds. It cannot attest that every window was read or that the resulting synthesis is faithful. See MEM-RTE-source-admit, WF-RTE-ingest-preparation and RT-RTE-ingest-complete.
Admission, publication and recovery
Page replacement validates references and concurrency, commits canonical content and then reconciles derived projections. The caller must resubmit still-valid citations: replacement does not infer which old references to preserve. For dependent edits the skill prescribes a sparse draft changeset, validation and publication. Changeset commit rejects conflicts and blocking lint on a fresh publication without an override; code also supports a reasoned lint override and selector-free direct-live page writes. These are supported alternate paths, not evidence that the stronger agent policy was followed. No universal prepublication lint or human-approval guarantee follows. See RT-RTE-page-write, RT-RTE-changeset-commit and EPI-RTE-lint-check.
The policy's semantic acceptance criterion is source-relative verification, explicit provenance, lint and fixed original/paraphrase retrieval checks in the top five, repeated against live state after commit. That is a consequential host procedure, not an independent Rust truth evaluator. An ingest-complete flag is therefore weaker than the prescribed delivered result. The agent owns repair and final delivery to the user. A queued Work receipt likewise is not maintenance success: terminal state and result must be inspected. Because canonical commit precedes projection work, an error or interrupted response does not establish that nothing was saved. See EPI-RTE-milestone-review, the workflow account and RT-RTE-work-runner.
Later consumption and context control
Explicit CLI reads and MCP exploration provide requested recall. MCP ranks candidates and bounds delivered page bodies to 15,000 characters each and 60,000 in aggregate, with truncation flags; a source hit is not automatically full source delivery. Lifecycle hooks instead push tag-selected pages and tracked task cues. Tag loading can omit whole pages under tag/global/render budgets and fail open to readiness-only or empty output. Pi caches one-shot context around lifecycle events and appends it to the system prompt. The intended authority remains untrusted reference data, not instructions allowed to override host policy. Neither delivery position nor source code demonstrates model compliance. Full explicit reads remain reload routes. See RT-RTE-mcp-recall, RT-RTE-hook-context, RT-BAP-wiki-context and MEM-RTE-task-continuity.
These mechanisms mainly select bounded external state and recover exact identities after context loss; they are not a general LWC summarizer of discarded host conversation. Host compaction remains external. Discussion adds a different path: hooks conditionally capture exact raw visible replies, while agents separately submit derived summaries. Revising referenced items marks dependent summaries stale and unconfirmed. Raw capture alone is not a derived behavior update, and a generated summary is not user authorization. Temporal capsules separately retain selected changes, reasons and outcomes; recall explicitly reports lack of live verification. See MEM-RTE-discussion-capture, MEM-RTE-discussion-revise, MEM-BAP-discussion-summary and MEM-RTE-temporal-cycle.
Read-only exploration does not make the entire tool surface read-only. MCP core commands and Discussion support mutation; core commands prohibit the changeset selector. Workspace path validation and command allowlisting are scoped controls, not a general operating-system sandbox. Replication and archive import add publication and conflict-recovery gates, but authorized transfer does not establish semantic preservation. See RT-RTE-core-command, RT-RTE-team-sync and MEM-RTE-archive-admit.
Retained adaptation and its evidential limits
The strongest directly traced automatic adaptation is Practice review rating/time → durable scheduling card and debt → later budgeted review queue. Code consumes the retained result to change future selection. This is a positive trace-fed update capacity, not merely raw-log storage; no comparison establishes improved capacity or learning outcomes. Discussion summaries and Tutor checkpoints separately afford trace-derived continuation through an external agent, rather than an internal summarizer. Their production and recovery were not observed. See MEM-RTE-practice-review, MEM-BAP-review-schedule, MEM-BAP-discussion-summary and RT-RTE-tutor-turn.
Other adaptation has different force. Explicit retrieval feedback changes numeric ranks, not the truth of stored text. Tutor's versioned Soul supplies standing teaching guidance through cold/recovery status and a prescribed hot-turn cache. Local approval fields are not independently authenticated learner consent, and turn commit cannot enforce the host's commit-before-display ordering. Book separately leases ordered source blocks and gates synthesis on committed coverage and anomaly disposition; coverage proves accepted report submission, not comprehension. Its token-labelled budget uses character arithmetic, not exact model token counting. See MEM-RTE-retrieval-adjust, RT-RTE-soul-publish, MEM-RTE-book-window and MEM-RTE-book-synthesis.
Practice also separates item eligibility from answer evaluation. Verification checks repeated fields and source identity/hash/revision; source-relative semantic verification remains agent policy. Objective grading derives exact JSON equality against a frozen answer, not source truth or semantic equivalence. Subjective accepted/pending state depends on submitted judgment and confidence, not independent rubric verification. See EPI-RTE-practice-verify and EPI-RTE-practice-grade.
For reuse of task knowledge, the supported contribution is durable evidence-linked content plus explicit and lifecycle recall, with mechanical recovery boundaries. For continued discussion and teaching, the contribution is separately retained exchanges, checkpoints, guidance and review state. Neither scenario has supplied evidence of useful compounding. The accepted records do not establish a formulated-theory pathway sufficient to decide theory-builder conditions 1–4 separately: localized theory content, content-dependent decisions, content-directed criticism and retained iterative dependence remain unestablished, not evidenced absent system-wide. Nor do they establish that criticism of a consumed theory improved future capacity. Reflection's two causal directions and role-complete autonomy remain independently unestablished. Standing guidance revisions and queue updates do not by themselves establish exercised self-improvement: subsequent operation causally depending on an evidence-responsive change to the system's own organization was not supplied. These limits do not negate the supported persistence and adaptation capacities. See the epistemic bounded conclusion, RT-BAP-soul-guidance and MEM-RTE-practice-review.
A linked source-to-candidate-to-review-to-publication-to-later-use trace would change the assessment of semantic acceptance and exercised reliance. Retained criticisms tied to changed claims and a next round would bear on theory-building; comparisons of subsequent capacity would bear on improvement. Closing architectural gaps instead requires tracing the named opaque admission and delivery alternatives, not merely executing another search. No additional source reads were made for this synthesis.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No supplied end-to-end or causal evidence for semantic write-back and reuse | SRC-1; RT-CLM-agent-memory, EPI-RTE-milestone-review, MEM-RTE-practice-review | Source implementation and agent policies; no target execution | Exercised semantic acceptance, model compliance, useful compounding, improved capacity or exercised self-improvement | Linked inputs, candidate/check/disposition records and later operation; objective-relative comparisons with adequate attribution |
| Host inference, provider parameters, deployment state and external executable bytes are outside the frozen implementation evidence | SRC-1; RT-CMP-host-model, RT-RTE-runtime-install, RT-RTE-team-sync | Inspected caller and installation/publication contracts | Weight fixity or learning, release/source equivalence, deployed isolation and service reliability | Pinned provider/runtime identities, release evidence and scoped deployment/run evidence |
| Semantic transformations lack concrete source/candidate and criticism histories | SRC-1; MEM-RTE-book-synthesis, MEM-RTE-discussion-revise, EPI-RTE-milestone-review | Report/synthesis submission and instructed host review | Faithful reshaping, accepted ampliative knowledge, theory-building iteration, reflection or autonomous criticism; no whole-system absence follows | Concrete paired content, stated challenged claims and blame, dispositions and retained next-round dependence; role and self-representation traces where relevant |
| Included learning-update and curation alternatives remain incompletely inspected | SRC-1; MEM-OBJ-soul, MEM-OBJ-practice-items, MEM-OBJ-derived-indexes, MEM-RTE-task-continuity | Selected Tutor, Practice, Plan/Todo and graph/span paths | Complete adaptation, admission, curation or assessment-quality inventory; positive witnesses do not classify opaque alternatives | Trace learner-fact/goal/plan, item retirement/blueprint/override, graph/retraction and span-fusion consumers with their admission and lineage |
| Delivery and recovery alternatives are partial; scoped controls are not universal guarantees | SRC-1; RT-RTE-hook-context, MEM-RTE-task-continuity, MEM-RTE-book-synthesis, RT-RTE-team-sync, RT-RTE-work-runner | Budgeted pi/hook/MCP paths, selected Book, archive/team and maintenance paths | Host parity, universal context isolation, complete byte delivery, private-unit Sync preservation or restore success | Trace non-pi and legacy fallback selectors, frontend/cloud rendering, Book display/projection, private Sync/merge and restore/recovery branches |
| Withdrawal is not uniform erasure | SRC-1; MEM-RTE-content-withdraw, MEM-RTE-temporal-cycle, MEM-OBJ-tutor-turns | Guarded Wiki removal, temporal eviction and private learning retention doctrine | All-memory purge across history, archives, replicas, private stores or prior model contexts | Scope-specific erasure contracts and traced propagation through all retained copies and consumers |
| Benchmark caller's automatic raw-message import and evidence read-back lack separate memory-route characterization | SRC-1; report-verification limits, MEM-RTE-source-admit, boundary | Verifier inspected benchmarks/agent_memory/lwc_backend.py: add-many prepares source files/add-manifest; search requests passages then rereads full sources |
Complete trace-origin, automatic raw-trace admission and evaluator-delivery inventory; this is not evidence of model-led Wiki synthesis or benefit | Canonical records tracing this caller's inputs, automatic admission, source selection and final delivered payload |
| Learning authority of conversational inputs is not separately classified at every update consumer | SRC-1; profile-verification limits, MEM-RTE-discussion-revise, MEM-BAP-discussion-summary, RT-RTE-tutor-turn, MEM-RTE-practice-review | Partial authority inventory covers scheduler learning authority and downstream conversational continuation, but not each conversational update consumer | Complete consumer-by-consumer learning-authority inventory or counts; scheduler's independent positive remains supported | Classify retained conversational inputs at Discussion-summary and Tutor-checkpoint update consumers, separately from output authority |