SkillNote reconciliation

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

Reconciliation

The four reports share the frozen source and implementation/doctrine boundary in the Source register. None introduces observed operation or causal evidence. This reconciliation uses the supplied members only; no source reads were needed. No correction answers or declined objections were supplied. No cross-report disagreement was found; the connections and remaining limits below do not settle the support of an individual report.

Duplicate identity

Amendment: EPI-RTE-install-maintenance is superseded by MEM-RTE-install-maintenance. Both identify the same CLI update/remove invocations in SRC-1, cli/src/commands/update.ts and cli/src/commands/remove.ts: installed-manifest selection, active-release/current checksum checks, destination replacement, shared-global retention and manifest changes. Both retain the no-active-version skip, nontransactional replacement limit, wired implementation and uninspected operation. The shared global-removal passage and corresponding route progression establish identity, not merely similar labels. The epistemic ledger's check and operational-consumption entries and its withdrawal conclusion therefore refer to the memory-declared route without gaining a stronger status or a broader cleanup guarantee. Both declarations remain. Neither is identical to RT-RTE-bundle-install, whose publication and initial-add route is supplemented by these maintenance invocations.

Containers, parts and nonidentical overlaps

MEM-RTE-feedback-catalog, EPI-RTE-bundle-metadata and EPI-RTE-bundle-selection all concern SRC-1, backend/app/api/openclaw.py, and explicitly belong to RT-RTE-openclaw-use. They are not interchangeable identities. The memory record covers the candidate endpoint's delivery, aggregation and ordering; the epistemic records separately assess arithmetic/status production and feedback-dependent selection. The endpoint container remains valid alongside these functional parts, so no split supersession is warranted. The reports converge on the alphabetic preliminary window, usage/rating ordering, final truncation and lack of established resolver invocation. Implemented preprocessing does not acquire the host-selection policy's status, or establish its execution.

MEM-RTE-inbound-resolution is a declared part, not a duplicate, of RT-RTE-browser-sync. Runtime leaves full server collision admission uninspected; memory adds staging, policy resolution, promotion, deletion and audit behavior from SRC-1, backend/app/api/claude_ai.py. The epistemic report's import ledger entry inherits the runtime's bounded acquisition finding and retains its own parsing limit. These are different inspection extents, not opposing findings. Staged availability, current-row adoption and remote delivery remain distinct milestones; memory's added detail does not establish semantic acceptance or remote operation.

EPI-OBJ-skill-content spans instruction text carried in rows, snapshots and installed files. MEM-OBJ-skill describes the current skill including access/lineage fields; MEM-OBJ-content-version describes historical and staged snapshots. Neither whole object is identical to the cross-substrate text account. Their overlap does not justify supersession: the text account remains a valid container, while the memory objects distinguish consumers and restoration semantics. MEM-OBJ-skill-version is separately a published archive reference, not another name for a content snapshot. Its opaque archive payload cannot inherit the natural-language classification of a displayed body.

EPI-OBJ-feedback-text groups attributed assessment text and scores across Comment and SkillRating. MEM-OBJ-comment and MEM-OBJ-usage-event identify separate retained stores and fields. These are not duplicate whole-object identities. Likewise, EPI-OBJ-context-bundle is a recomputed response, not the persistent source objects or their delivery route. EPI-OBJ-skill-push is shipped policy text, not the registry-authoring route that consumes it.

Integration dispositions and independent convergence

RT-RTE-feedback-report, EPI-RTE-rating-admission, EPI-RTE-comment-check and EPI-RTE-comment-retention remain distinct: MCP completion/current-version attribution, REST rating submission, comment validation and comment persistence have different inputs or functions. Memory's comment-to-SkillRating fan-out connects stores without merging them. Runtime explicitly leaves their identity unestablished; memory and epistemic agree that bundle means read Comment and comment deletion does not retract the emitted rating in the inspected handler. No correction of runtime uncertainty is needed, and no shared invalidation policy follows.

MEM-RTE-trace-reporting is neither rating admission nor the context-bundle usage route. Its watcher writes analytics read events and cursor state, not the SkillUsageEvent store counted by the bundle. MEM-OBJ-watcher-state is recovery/deduplication state, not learned procedure text. The epistemic report's uninspected watcher attribution is a local coverage limit, not a denial of memory's parser finding. Session-evaluation generation and consequential evaluation-driven content revision remain unestablished.

The workflow's authoring annotation adds optional specialist responsibility, user choices and rejection recovery to RT-RTE-registry-write, without declaring another route identity. Its later installation milestone uses RT-RTE-project-sync. Runtime, memory and epistemic distinguish prompt approval from structural API admission; workflow's additional schema and collection checks do not contradict that distinction. Epistemic's doctrine-only authoring/review functions and runtime's wired prompt/API route describe different layers, not conflicting statuses. No universal approval gate or expected-answer oracle follows.

RT-RTE-context-hooks and its memory annotation converge on pointers/reminders rather than retained task-reasoning summaries; PostCompact model visibility and host frame behavior remain unresolved. RT-BAP-host-instructions and RT-CMP-host-model bound all later instruction use by external-host interpretation. Distribution, discovery and compliance remain separate.

The epistemic claim comparison connects EPI-CLM-pattern-learning to persistence/distribution and EPI-CLM-registry-improves to feedback-conditioned delivery. These claim records are not duplicate route declarations or conflicts with analysts' narrower findings. RT-RTE-adapter-update similarly establishes an update path, not an improvement experiment. Across the reports, implementation capacity, unobserved host behavior and unestablished benefit stay separate. Unknown corpus rationale, archive contents, arbitrary metadata semantics and global-install ownership internals remain local limits rather than system-wide negatives.