llm-wiki synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Install a maintainer, not an enclosing runtime
llm-wiki supplies a Claude Code skill for maintaining a personal or team wiki in Logseq or Obsidian. It combines a model-interpreted procedure, page templates and configuration with executable installation and initialization scripts. The whole-system analysis covers the shipped paths that create, maintain and consume retained state at commit 11538c1195666db24e25e23cc64bb1d70587aef9, with an analysis cutoff of 2026-10-17. The evidence is SRC-1, the frozen repository: principally doctrine/design, with a narrower implementation layer for setup. No inspected execution trace or causal experiment establishes compliance or benefit. Claude Code's model, permissions, session management and automatic memory loading are external; so are the two wiki applications and real user vaults. See the boundary and Source register.
The plugin loads the root skill with a configured wiki directory. An alternate installer copies it into a project's command file; that copy does not update with plugin releases. The foreground host agent owns command progression, interpretation, selection and tool requests. The target supplies no separate worker handoff in the representative ingest procedure. The initializer receives flags and templates, writes missing configuration and skeleton files, and optionally initializes Git and attempts a commit. Its existing-file guard is not a system-wide non-overwrite guarantee: the interactive installer can first remove an existing configuration after confirmation. Neither path supplies transactional rollback. RT-BAP-skill-dispatch, RT-RTE-wiki-setup, RT-CMP-claude-code.
Turn sources into current state and discoverable pages
The representative workflow is /llm-wiki ingest <source>, the skill's default multi-step operation. The agent resolves the root, reads configuration and Schema, acquires a URL/file/text, extracts facts and entities, inspects existing pages, and decides what to create or revise. A superseding fact replaces current wording; the old wording moves verbatim into dated History. Ingest also maintains each affected page's hub routing description and tags. This is the consequential handoff to a later query: a saved page without a usable routing entry can be absent from preferred retrieval. Bulk import is a distinct route and does not explicitly require the same description production. RT-RTE-source-ingest, MEM-OBJ-page-state, MEM-OBJ-page-history, MEM-OBJ-hub-index, workflow account.
The same agent then checks properties, cross-references, routing, credential exclusion and page-touch counts. This is a post-write structural self-check, not an independent factual reviewer or atomic admission barrier. Ingest does not require human approval for each extracted fact. Its prescribed delivery is persisted page/routing changes, a change-and-warning report and a structural Git commit; successful execution is unobserved. History and Git afford manual recovery, not guaranteed rollback. Acquiring and retaining source claims does not establish their truth or complete fact-to-source provenance. RT-RTE-source-ingest, workflow account, epistemic ledger.
Retrieve selectively, then retain evidence of use
A later question requests retained knowledge. The foreground model reads candidate namespace hubs, ranks short descriptions and tags, and selects three child pages, at most five under the operative skill. It reads selected pages and produces an attributed answer with confidence and staleness qualifications. If routing is ineffective, grep across all wiki pages supplies an alternate path, including archived pages. This is request-driven retrieval, not unsolicited context injection. Page descriptions influence ranking; page text supplies informational material for synthesis. Neither role establishes factual endorsement. RT-RTE-wiki-query, MEM-BAP-wiki-answer, MEM-OBJ-hub-index.
The skill requests at most three pages loaded simultaneously and JIT batching, but supplies no context deletion, isolated child frame or checkpoint mechanism. It limits selected objects, not tokens, and does not demonstrate removal of previously read text. Folded History is a UI convention, not a proven filter on model-visible full-page results. The supported mechanism is selective admission before detailed reads, not context reclamation. Supporting specifications contain a seven-page scenario, so the skill's limit must not be generalized to every design document. RT-RTE-wiki-query, MEM-OBJ-page-history, report verification limits.
Query appends actual full-page reads and short selection reasons to Access-Log before synthesis. That write is separate from optional knowledge write-back, which requires user confirmation. Confirmation authorizes retaining a proposed summary or gap fill; it is not a named evidence-consuming factual acceptance test. Without actual synthesis inputs and outputs, the answer edge cannot be classified as faithful reshaping, warranted derivation or non-entailing conjecture. MEM-OBJ-access-log, RT-RTE-wiki-query, epistemic ledger.
Use feedback to change selection and review standing rules
Access traces can shape future preferred retrieval. Requested or externally scheduled prune uses latest logged access, creation date when access is missing, and exemptions to propose cold pages. Only confirmed candidates receive archive markers and have routing lines moved verbatim from live Index to Archive. Files remain available through links and grep. This is a prescribed, human-gated change in selection salience, not factual rejection or an enforced access restriction. The skill offers re-promotion on an archived re-hit; README's automatic wording conflicts with that procedure. No comparison establishes the claimed precision benefit. RT-RTE-index-prune, MEM-OBJ-hub-archive, EPI-CLM-automatic-repromotion, RT-CLM-cache-precision.
Lint supplies another model-mediated feedback route: diagnostics, optional --fix repairs and a retained Dashboard update. Status instead requests no modifications. Repairs can backfill descriptions and downgrade old confidence, but age and structure do not establish factual reliability. --fix authorizes repairs without per-fix confirmation and never authorizes L1 edits. RT-RTE-health-maintenance, MEM-OBJ-dashboard, EPI-CLM-confidence-lifecycle.
The strongest supported criticism-and-revision contribution is prune --l1, which reviews standing memory rules intended to steer future sessions. It selects due current-behavior claims, gathers local read-only evidence, reports supports/contradicts/inconclusive per claim, and asks the user to re-verify, correct, demote, delete or skip. Contradiction permits re-verification only together with corrected text; inconclusive cases require user knowledge. Approved corrections synchronize the rule body, frontmatter description and index claim. Demotion must successfully write L2 carry-over history before deleting L1; credentials cannot be demoted. This is a bounded design-level evidence and acceptance protocol for continued local-rule reliance, not merely structural lint. Actual acceptance and later use remain unobserved. RT-RTE-l1-verification, MEM-OBJ-l1-rules, MEM-OBJ-l1-index, EPI-CLM-l1-honesty.
L1 denotes standing host memory; L2 denotes the on-demand wiki. Corrected L1 content has an afforded later host consumer, but exact body/description/index delivery and budgets are unknown. Review dates select checks, not permission to act. llm-wiki explicitly has no action-time freshness gate. Credential separation and local read-only checks are agent policies, not target-enforced isolation; arbitrary memory-directory placement is not guaranteed to exclude L1 from Git. MEM-BAP-l1-loading, RT-ABS-action-gate, memory integration issues.
What the revision pathways establish
For L1 verification, the runtime separately assesses theory-builder conditions: 1, localized content: afforded, because files can hold whole behavior claims; 2, content-dependent consumption: claimed, because meaning determines checks and recommendations; 3, content-directed criticism: claimed, through per-claim contradiction and correction; 4, iteration: afforded, because corrected rules and due metadata can shape another round. Actual theories, criticism episodes and iteration were not supplied. These design findings are not a demonstrated theory-building run. RT-RTE-l1-verification.
Criticism can produce durable corrected guidance, but improved capacity for future action is unestablished: there is no outcome or attribution evidence. Reflection is likewise only partially characterized: the route represents and edits behavior-steering claims, while correspondence with actual operation and both causal directions remain uninspected. A reflective theory-builder conclusion about llm-wiki's own method is unestablished. The verification route is not autonomous because users make required internal admission decisions. These are independent findings, not a ladder. RT-RTE-l1-verification, epistemic bounded conclusion.
For self-improvement at the shipped knowledge-system boundary, the positive contribution is an improvement-directed design: use evidence can alter future ranking, and local criticism can alter durable guidance. The unresolved link is operative dependence: no supplied subsequent operation shows that a changed index or rule entered actual behavior and causally affected it. Exercised self-improvement, and successful improvement separately, are therefore unestablished. Automatic trace learning also remains not determinable, not absent: raw logging and human-approved durable changes do not establish a qualifying automatic chain, while initial L1 production and the conflicting automatic re-promotion claim remain unresolved. RT-RTE-index-prune, RT-RTE-l1-verification, profile verification.
For a hub-first wiki-query scenario, the source supports an explicit selection procedure, fallback and reversible salience maintenance, not measured accuracy or context savings. For a stale local-rule scenario, it supports evidence-responsive, user-controlled correction, not action-time protection. A linked ingest/query/prune trace would establish exercised transitions; a later session showing dependence on a corrected rule would strengthen the activation account. Controlled outcome comparisons would be needed for precision or future-action benefit, with independent variation required to attribute effects to particular mechanisms. A target-owned context-removal mechanism, verified host loading contract or action interceptor would materially change the corresponding architectural assessment. None is established by the present records. RT-CLM-cache-precision, RT-RTE-wiki-query, RT-RTE-l1-verification, RT-ABS-action-gate.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No inspected execution or comparison | SRC-1; RT-RTE-source-ingest, RT-RTE-wiki-query, RT-CLM-cache-precision | Frozen shipped procedures and setup implementation; no user vault or execution trace | Actual compliance, delivery, retrieval benefit, improved future-action capacity and exercised self-improvement | Provenance-linked runs through later consumption; objective-specific outcome comparisons and suitable causal controls |
| Supporting-doctrine inventory is incomplete and page limits vary | SRC-1; report verification limits, RT-RTE-wiki-query, RT-RTE-health-maintenance, epistemic ledger | Operative skill plus verifier's scoped inspection of openspec/specs/query.md and openspec/specs/l1-l2-routing.md |
Exhaustive comparison of evidence-ranking and L1 policies: confidence weighting, explicit contradiction presentation and promotion/merge recommendations are not separately retained; no uniform page limit across all doctrine | Accepted records covering these policies and reconciled specification limits; these prescriptions alone would still not prove factual acceptance or automatic admission |
| Re-promotion and Dashboard-writer wording conflict | SRC-1; report verification limits, EPI-CLM-automatic-repromotion, MEM-OBJ-dashboard, RT-RTE-health-maintenance | README, skill and Dashboard templates | One uniform contract, automatic re-promotion, status-driven Dashboard persistence or observed resolution; skill-level offer and lint writer remain supported | Harmonized source contracts and traces showing the actual disposition/writer |
Obsidian logical hub links versus _index.md paths |
SRC-1; report verification limits, RT-RTE-wiki-setup, MEM-OBJ-hub-index | Initializer, generated naming and Dashboard links; external UI uninspected | Demonstrated end-to-end link resolution or usable preferred routing in a real Obsidian vault | Inspected application resolution contract or a provenance-linked vault integration trace |
| Initial L1 admission and host payload are unknown | SRC-1; MEM-OBJ-l1-rules, MEM-OBJ-l1-index, MEM-BAP-l1-loading, RT-CMP-claude-code | Target's file-maintenance and attributed external-consumer roles; no host implementation, actual rules or provider state | Complete lineage/admission classification, universal body loading, exact context budget, fixed-model or parametric-learning conclusions | Initial-content provenance, host selection/loading evidence and separately scoped provider evidence |
| Safety and freshness are policy-scoped | SRC-1; RT-ABS-action-gate, RT-RTE-l1-verification, memory integration issues | Skill, manifests and shell setup; deployed permissions and arbitrary host routes excluded | Target-enforced credential isolation, universal L1 Git exclusion or stale-rule action veto | Enforced path/permission controls, deployment configuration and a separately evidenced action-time gate |
| Selection bounds do not establish window reclamation | SRC-1; RT-RTE-wiki-query, MEM-OBJ-page-history | Skill's selective reads/JIT instructions; host context processing excluded | Token bound, eviction, isolated task frame or omission of folded History from model input | Host-visible payload/context traces or an inspected isolation/removal implementation |
| Fact warrant, semantic transformations and retained reasons remain partial | SRC-1; MEM-OBJ-page-state, MEM-OBJ-page-history, MEM-OBJ-access-log, EPI-CLM-confidence-lifecycle, RT-RTE-l1-verification | Prescribed extraction, synthesis, logging and L1 correction without actual inputs/outputs or a required full checking ledger | Universal provenance, verified high-confidence content, warranted derivation/novel conjecture, or reconstruction of every later decision's reasons | Linked source/output propositions, evidence-consuming fact checks and retained checking rationales with later consumers |
| Preferred routing and recovery are conditional | SRC-1; RT-RTE-source-ingest, RT-RTE-wiki-setup, RT-RTE-health-maintenance | Import, generated illustrative hubs, lint repair, installer and initializer | Complete usable indexes after setup/import, atomic rejection or guaranteed recovery; L1 deletion has no Git undo | Validated routing outputs, explicit transactional/recovery controls and observed failure-path recovery |
No source files were read for this synthesis; it uses the supplied accepted members. The profile verification declares no additional limits. No unresolved inter-report conflict was supplied.