Decapod reconciliation

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

Reconciliation

Boundary and connection basis

The reports share the identity and evidence boundary in boundary. Its Source register supplies static implementation and doctrine/design, not observed target execution or causal experiments. The workflow declares no records and adds annotations to runtime routes. No analyst answers or refusal were supplied to this attempt. The previous reconciliation is not current evidence: all connections below were reassessed from the pinned reports. No additional source reads were made.

Duplicate supersessions

These amendments establish referent identity, not replacement of an analyst's fields or a stronger status. Both declarations and their local evidence limits remain.

Amendment: EPI-OBJ-knowledge-entry is superseded by MEM-OBJ-knowledge-entry. Both identify KnowledgeEntry rows in the local database, containing title/content and provenance/lifecycle metadata, admitted and consumed by the same knowledge functions at SRC-1, src/decapod/plugins/knowledge.rs. The epistemic inventory and knowledge ledger concern this memory object, not another store; unknown source warrant remains unchanged.

Amendment: EPI-OBJ-internalization-manifest is superseded by MEM-OBJ-internalization-manifest. Both identify InternalizationManifest and associated mount/lease metadata with source/model/hash/TTL/replay fields, consumed by inspect and attach at SRC-1, src/decapod/plugins/internalize.rs. Their path descriptions do not establish different artifact families. The epistemic inventory and attachment ledger retain their metadata-only license.

Amendment: EPI-OBJ-adapter-payload is superseded by MEM-OBJ-adapter-payload. Both identify adapter.bin from noop or a configured executable, subsequently hashed during integrity checks at SRC-1, src/decapod/plugins/internalize.rs. The epistemic transformation assessment and memory lease account therefore concern the same bytes. Empty default output and unknown custom encoding/loading remain separate findings about alternatives.

Amendment: EPI-OBJ-feedback-entry is superseded by MEM-OBJ-feedback-entry. Both identify FeedbackEntry rows containing caller text/source/links, reloaded by the proposal formatter at SRC-1, src/decapod/plugins/feedback.rs. Unknown source truth and memory's canonical-database qualification apply to the same retained input.

Amendment: EPI-OBJ-lcm-summary is superseded by MEM-OBJ-summary-entry. Both identify stored summary text with original hashes, scope and time, consumed by summary lookup at SRC-1, src/decapod/plugins/lcm.rs. The epistemic inventory and transformation ledger concern SummaryEntry, not the separate original events. Memory's additional ingestion/rebuild inspection does not contradict the epistemic analyst's narrower integrity coverage.

Amendment: EPI-RTE-internalization-flow is superseded by MEM-RTE-internalization-lease. Both trace create/cache/profile execution, inspect, attach and detach through payload, manifest and mount metadata at SRC-1, src/decapod/plugins/internalize.rs. Memory additionally follows validation's lease consumer; epistemic adds cache-key and best-effort-source details. These additions do not change the lifecycle referent. The ledger's transformation, retention, checking, admission and recovery rows remain separate assessments, not a single acceptance event.

Amendment: EPI-RTE-feedback-proposal is superseded by MEM-RTE-feedback-proposal. Both identify add_feedback followed by propose_prefs: retain caller feedback, select the latest ten and return quotations plus a fixed placeholder diff without installing a rule. Their common evidence and passages are at SRC-1, src/decapod/plugins/feedback.rs. Memory additionally establishes CLI definition/dispatch at src/decapod/cli.rs and src/decapod/lib.rs. The transformation ledger and requested read-back finding concern one route; both now assign implementation wired and operation uninspected.

Overlaps without whole-record identity

MEM-RTE-knowledge-store spans ordinary admission and requested retrieval. EPI-RTE-knowledge-admit includes library conflict handling, while EPI-RTE-knowledge-select isolates selection with optional temporal/recency behavior. These scopes intersect but are not interchangeable whole routes. Neither epistemic route supersedes the valid combined memory account. Their shared object is amended above.

MEM-RTE-promotion-lifecycle groups promotion recording, procedural admission, conflict handling, decay and retrieval-feedback APIs. EPI-RTE-promotion-record and EPI-RTE-knowledge-withdraw isolate recording and expiry withdrawal. EPI-OBJ-promotion-event is an event rather than another KnowledgeEntry. No whole-record or split supersession is warranted: the broader record remains a valid grouping and its narrower operations retain different admission and invocation conditions. Both reports distinguish string/structure checks from authenticated approval and semantic procedural derivation. The memory aggregate afforded status does not erase its explicitly wired promotion branch; the epistemic withdrawal route now independently assigns afforded.

MEM-RTE-lcm-summary covers ingestion, derivation, summary/original retrieval and index rebuild. EPI-RTE-lcm-summary covers derivation, retention and summary display. Both cite chronological formatting, the 64 KiB prefix and retained original hashes at SRC-1, src/decapod/plugins/lcm.rs. They are not whole-route duplicates. Original retention and summary fidelity remain different conclusions, and both note unsafe UTF-8 slicing. MEM-OBJ-lcm-event remains distinct from the reconciled summary object.

MEM-OBJ-archive-entry and MEM-RTE-archive-context concern copied files and caller-authored summary/pointer replacement, not LCM's event-derived prefix. MEM-OBJ-portable-context-bundle and MEM-RTE-bundle-transfer concern persisted selected snippets and checked import. None is identical to RT-RTE-context-resolve's immediate context delivery. Shared context purpose does not justify supersession. Archive recovery retains the memory report's same-second overwrite and noncolliding-retention qualifications; no other report asserts universal archive durability.

MEM-RTE-governance-reload follows trajectory updates, later assurance reads and Jev ledger append/reset/validation. RT-RTE-assurance-observe owns provider invocation and immediate advisory. Persistence overlaps without whole-route identity. MEM-OBJ-trajectory-artifact is provider input; MEM-OBJ-jev-ledger is retained output; neither is RT-CMP-jev. The memory annotation supplies downstream consumers without asserting that prior probabilities feed the next assessment.

MEM-RTE-handoff-memory derives knowledge and an aptitude observation after successful handoff status. It is not RT-RTE-workspace-ensure's custody/filesystem preparation. MEM-OBJ-observation, MEM-OBJ-preference and MEM-OBJ-agent-prompt have different content, writes and selectors. MEM-RTE-aptitude-reminders connects known reads without establishing observation-to-preference adoption. None is the Jev route or feedback formatter merely because its label names observation or preference.

MEM-OBJ-federation-node and MEM-RTE-decision-memory concern valid federation writes, rejected RPC input and independent decision queries. They are not RT-OBJ-accepted-executor, whose authority is accepted-base Git identity and controller closure. RT-OBJ-agent-session is custody state rather than knowledge, a decision node or internalization lease.

EPI-OBJ-eval-artifacts remains a valid container alongside EPI-OBJ-eval-verdict and EPI-OBJ-eval-aggregate. Imported judgments and derived statistics have different consumers; the declared parts do not require container supersession. EPI-RTE-eval-judge-import and EPI-RTE-eval-statistics are neither Jev invocation nor exact-candidate readiness. EPI-RTE-eval-promotion-gate declares itself part of RT-RTE-workspace-publish. Its optional aggregate threshold and the runtime's current-candidate readiness have different inputs and licenses. Lack of current-code binding in that helper does not deny binding at the separate readiness check. No supersession follows.

EPI-OBJ-preference-proposal is returned formatter output, not stored feedback or an adopted aptitude preference. EPI-CLM-checked-work, EPI-CLM-eval-promotion and EPI-CLM-adapter-replay assert different operations from RT-CLM-session-auth. No claim supersession is warranted.

Reassessed disagreements and integration dispositions

The previous feedback status conflict no longer applies. MEM-RTE-feedback-proposal now supplies the shipped CLI dispatch passage and assigns wired, matching EPI-RTE-feedback-proposal. The affected read-back and integration passages consistently retain command wiring while withholding adoption. This is convergence on the corrected current finding, not selection of the stronger previous status.

The previous decay status conflict also no longer applies. EPI-RTE-knowledge-withdraw now assigns afforded and records a caller search in src/, assets/tools/ and .github/ finding only the declaration, matching the decay branch of MEM-RTE-promotion-lifecycle. Its ledger's architectural implemented value describes the inspected algorithm, not shipped invocation. The bounded conclusion likewise makes withdrawal conditional on invocation. No common scheduled-maintenance conclusion follows.

The corrected MEM-RTE-decision-memory rejects RPC decision upsert before persistence: its fixed rpc:store.upsert provenance fails the cited regex at SRC-1, src/decapod/plugins/federation.rs, through the caller at src/decapod/lib.rs. Valid federation library writes and independent RPC listing remain supported. The memory integration issues, object, write-side and read-back passages withdraw the former RPC producer-to-read-back chain. The epistemic coverage table mentions bounded RPC decision-write inspection but establishes no successful retained-node route; runtime suggests querying prior decisions without asserting this rejected producer's success. No dependent cross-report finding requires the withdrawn chain. Thus the correction does not leave an anchored disagreement in these pinned reports.

RT-RTE-context-resolve, its memory annotation and the epistemic ledger converge on selected text delivery without total model-window control or semantic obedience. First-content-line spec projection and configured identity precedence refine payload reachability rather than contradict runtime selection. MEM-BAP-retained-content and RT-BAP-context-delivery share a consumer class but differ in accumulated versus embedded/current material and selectors. Neither is RT-BAP-readiness-gates's enforcing publication path.

The workflow account and annotations extend RT-RTE-validate-candidate and RT-RTE-workspace-publish with sequencing and recovery, without contrary dependent findings. The host owns semantic repair; Decapod owns covered checks. Post-push failure can leave durable remote effects. Verified branch/PR publication does not imply merge, deployment or semantic correctness. Filesystem custody does not establish a separate reasoning frame.

The remaining memory integration issues retain compatible limits: library lifecycle options do not imply shipped scheduling; approver strings do not authenticate approval; metadata-only attach does not establish custom encoding or model loading; unknown prompt initialization does not erase known later selection; retained originals/pointers do not establish rationale completeness in summaries; and decision status changes do not automatically suppress old RPC-list text or expand lineage. The explicitly identified feedback duplicate is amended above; governance persistence and external-consumer overlaps remain distinct rather than superseding runtime routes.

Memory's additional positive witnesses do not complete epistemic assessment of decide, federation graph paths, aptitude, worker lessons, health/reflex remediation, schedulers/evaluators, trace consumers, spec initialization, full governance transformations or remote drivers. Both reports retain available-source inspection gaps beside these alternatives. Their different inspection depths and faithfully bounded uncertainty are not disagreements or evidence of missing behavior. RT-RTE-release-update remains version-driven executable replacement, not evidence-responsive self-improvement by association with retained experience.

No unresolved cross-report conflict is established by the supplied current findings. No analyst answers declining an objection were supplied, so no decline is treated as a defect or repair. Learning, reflection, theory-building and self-improvement remain independent bounded conclusions, not a bundled negative classification. This reconciliation allocates no records, rewrites no analyst fields and judges neither individual support nor complete-loop adequacy.