LLM-WIKI-MCP reconciliation

Type: agentic-system-analyses/types/agentic-system-reconciliation-report.md

Reconciliation

Duplicate identities

Amendment: EPI-OBJ-wiki-pages is superseded by MEM-OBJ-wiki. Both denote the retained vault Markdown page corpus, not its writer or the current prompt. Their SRC-1 page-write passages identify the same filesystem substrate and subsequent indexing/retrieval consumers. The epistemic record's extraction-placeholder, sanitizer and alternate CLI-import findings remain qualifications on that corpus. The affected epistemic inventory and acquisition, admission, retention and retrieval ledger entries concern this same object; supersession does not replace their findings with a truth endorsement.

Amendment: EPI-OBJ-page-notes is superseded by MEM-OBJ-notes. Both denote generated Markdown sidecars under notes/. The memory record and the epistemic object's supporting EPI-RTE-note-template retain matching term-extraction and note-snippet-selection passages. The epistemic inventory and note/selection ledger rows therefore concern the same retained object. Its template-derived text, symbolic scores, indeterminate semantic relation and pre-promotion ranking influence remain separate findings, not stronger statuses inferred from identity.

Amendment: EPI-RTE-page-withdrawal is superseded by MEM-RTE-page-withdrawal. Both trace caller-selected rename/delete through file changes, optional link rewriting or tombstone preservation, reindex and later active-page reads. Their shared tombstone-copy/unlink/reindex passages establish route identity. The epistemic withdrawal ledger and bounded conclusion retain their local-removal and no-cascade qualifications. Memory places this route within RT-RTE-knowledge-write; epistemic distinguishes rename/delete from the parent's enumerated create/update/ingest operations. A specialized operation can belong to that broader write family without being identical to the parent, so these comparisons do not conflict.

Containers, parts and distinct referents

No further cross-report identity is established:

Disagreement reassessment

No unresolved cross-report disagreement is identified in the supplied versions. In particular, the previous reconciliation's context-pack conflict no longer holds against the pinned runtime record.

Two historical channels now agree. RT-OBJ-context-pack explicitly distinguishes reduced recent Q/A from serialized wiki_ask_history tool evidence. It says the latter preserves sanitized stored prompt fields, subject to a 6000-character serialization cap, source allocation and whole-pack trimming. RT-RTE-agentic-ask and RT-CLM-context-budget carry the same scope correction. This agrees with memory's annotation and MEM-OBJ-ask-history, whose supporting branch and sanitizer findings establish the second channel. Memory integration issue 1 still describes the earlier qualification, but the current runtime no longer makes the disputed object-wide exclusion. EPI-RTE-history-reshape restricts its verbatim-prompt exclusion to the recent-Q/A renderer; the workflow account makes no exhaustive historical-exposure claim. No record now warrants treating the dedicated history budget or its token estimate as a bound on all history in the first prompt.

Convergence and integration dispositions

Orchestration and answer availability. RT-RTE-agentic-ask, EPI-RTE-answer-check and the workflow annotation converge on code-owned evidence selection, generation requests and blank-answer recovery, not model-selected tools or factual acceptance. Workflow's truthy-answer display branch separates visible delivery from provider ok and persistence success. EPI-RTE-fallback-compose adds a concrete readiness assertion possible without supporting pages, distinct from generic extractive fallback. This does not contradict runtime's qualified fallback account or workflow's warning that nonempty text may not resolve the subject question. Neither report licenses correctness from availability.

Prompt scope and reconstruction. RT-CLM-context-budget, the workflow account and memory's read-back account agree that first-agentic-pack controls do not cover the separately built second-pass prompt, RT-RTE-plain-ask or RT-RTE-store-ask. They also retain the first stored/exposed prompt even when a second request supplies the answer. These limits coexist with the corrected two-channel history finding. No report establishes exact provider capacity or full replay provenance.

Archive consumption and compaction. Memory integration issue 2 connects MEM-RTE-history-compaction, MEM-RTE-recursive-context and EPI-RTE-history-reshape. The compressed archive has an explicit local reader, but the final pack's recent-history section does not read it. Registered context return does not establish subsequent host/model use. Memory's 500-row selection and destructive-rewrite limits add detail without contradicting epistemic's narrower account. Loss of older exact live rows is not total erasure of all independently retained answer copies.

Diagnostic retention correction. Memory integration issue 6 and MEM-OBJ-runtime-traces now explicitly establish that sanitized provider answer/raw-response metadata can persist in metrics independently of history append and compression. Complete recent metrics rows can return through metrics summaries, usage and self-access tools. CLI diagnostic displays expose narrower selected fields. These diagnostic consumers are not the normal answer-history reader or automatic final-pack delivery. This corrects memory's own earlier exclusion of full answer bodies, not runtime's positive metrics finding. It also does not conflict with the history records' claims about loss from live JSONL: those claims concern that particular substrate. Exact answer-producing prompt reconstruction remains unestablished. Journal-helper capability stays separate from final automatic journal production and from score-driven policy adaptation.

Admission, withdrawal and source refresh. Memory integration issues 3 and 5 are handled by the identities and part distinctions above. MEM-OBJ-provenance and the knowledge-write annotation explain that unchanged metadata may skip a separately deleted generated page; EPI-OBJ-wiki-pages distinguishes one-shot import from provenance ingestion. Neither finding establishes universal restoration. Both withdrawal records distinguish active-file removal from erasing sidecars, provenance, historic prompts, metrics or external copies. Cache and concurrency limits remain limits rather than a disagreement about guaranteed freshness. EPI-RTE-structure-check supplies structural check domains consistent with runtime's operational admission predicates. Memory and epistemic independently agree that sidecar templates can affect ranking/snippets before later editorial promotion, without establishing semantic acceptance or improvement.

Navigation and instruction-surface limits. Memory integration issue 7 explicitly adds direct inspection of TOC generation and separate editable/default schema guidance, but leaves full schema dispatch/read-back untraced. Epistemic's coverage likewise includes schema/TOC bodies without claiming exhaustive wrapper assessment. MEM-OBJ-wiki is therefore not an exhaustive inventory of editable instructions, and the current context-pack record establishes no automatic schema delivery. This is a retained coverage limit, not a conflicting positive or a basis for silently merging schema guidance into the page-corpus identity. Complete instruction-surface and maintenance-route comparisons remain prevented.

Authority and evidence limits. Memory integration issue 4 retains unknown deployment corpus, caller identity and downstream use. RT-RTE-runtime-config and its memory annotation distinguish append enablement from reads/compression of existing files. RT-BAP-config-control describes executable setting selection; RT-BAP-answer-guidance describes interpreted prompt guidance. They are different authority paths, not competing statuses. RT-CMP-ollama-model leaves provider parameters and exact deployed identity unresolved. The Source register supplies implementation and doctrine/design evidence, not observed operation or causal benefit. Each report's wired, implemented and afforded conclusions keeps its own scope; none is promoted to a strongest system-wide status or a conclusion about improved capacity.

No analyst answers or refusal were supplied. The memory report's internal corrections are assessed above from the current findings; no declined objection is treated as an established defect. All connections use supplied members and their retained evidence passages. No frozen-source rereads were needed.