OpenViking reconciliation

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

Reconciliation

Evidence and status boundary

The reports share the revision and complete-artifact, partial-loop scope of boundary. Its Source register supplies implementation, doctrine/design and attributed operation, but no observed run or causal experiment. The epistemic ledger's implemented architectural status and the analysts' wired implementation and uninspected operation conclusions are different fields. They do not disagree merely because their vocabularies differ.

This reconciliation reassesses the supplied members. No source ambiguity required another source read. No correction-answer file or declined objection was supplied; the memory member's own correction and integration statements are assessed below, without inferring an analyst's answer from them.

Identity and containment dispositions

No exact duplicate referent is established, so no supersession is proposed. The consequential overlaps are ruled out as identities on the following grounds.

Archive and derived context. RT-OBJ-session-commit is the archive/queue container, with raw messages and scheduling metadata. MEM-OBJ-archive-overview and MEM-OBJ-checkpoint are declared derived parts with model-context consumers. Their distinct content and admission checks do not invalidate the container. MEM-OBJ-tool-result is separately stored tool output with a reduced display representation, not another archive declaration. MEM-RTE-session-reconstruction is the later-context branch of RT-RTE-session-commit, not the archive/enqueue protocol itself. MEM-RTE-tool-externalization also runs during message admission and has offload/hydration endpoints; it is not identical to commit.

Memory content and its representations. MEM-OBJ-memory-file individuates the persisted ordinary user-memory file, including separately consumed body and metadata. EPI-OBJ-memory-fields individuates schema-defined assertion fields for epistemic assessment. Their scopes overlap at retained assertion content, but the fields are not the complete file with its storage and metadata consumers. MEM-OBJ-memory-index is a derived vector/URI/abstract representation, not the file or RT-CMP-account-embedder that produces its numerical representation. Shared substrate or producer does not establish identity.

Application and constituent checks. MEM-RTE-memory-application covers streaming/grouped submission, merge, application, deletion safeguards and indexing handoff across several producers. EPI-RTE-memory-write isolates upsert admission and retention; EPI-RTE-memory-withdrawal isolates dependency-guarded deletes. Shared updater passages establish overlap, not equality of these complete routes. Their declared containment in RT-RTE-session-commit identifies commit's application branches without making all other uses of the updater commit operations. EPI-RTE-memory-guards ends at finalized extraction operations before physical application. EPI-RTE-memory-consolidation uses the same loop/updater with selected existing-memory directories and a compile trigger, rather than the session archive protocol. None supersedes the commit container or the broader application route.

Experience objects and feedback. MEM-OBJ-case, MEM-OBJ-trajectory and MEM-OBJ-experience distinguish case routing/training input, diagnostic content and persisted ExperienceSet policy. EPI-OBJ-experience-instructions identifies the experiences schema's instruction-content contract, not the entire policy object, links and snapshots. EPI-OBJ-trajectory-evaluation is the evaluation result, not the extracted trajectory. These are different operative scopes, not duplicate records.

MEM-RTE-experience-maintenance follows the compressor's trace-to-experience chain. EPI-RTE-rollout-feedback isolates evaluator selection, fallback and feedback consumption. EPI-RTE-policy-training traces generic batch engine/pipeline successor flow. RT-RTE-skill-admission traces conditional session-skill production with a different persisted product and updater. Common analyzer, gradient and optimizer roles do not equate these triggers, products and endpoints.

Recall parts and consumers. EPI-OBJ-recall-digest and EPI-RTE-digest-contract are declared output and checking parts of RT-RTE-context-assembly, not duplicate assembly routes. MEM-OBJ-recall-ledger is retained access metadata; MEM-BAP-ledger-selection affects server candidate exclusion. RT-BAP-recall-context affects host input after selection. Their consumers, channels and force differ.

MEM-RTE-bot-recall and MEM-BAP-experience-reminder cover VikingBot's independent formatting and reminder delivery, not OpenCode's RT-RTE-plugin-recall. WF-RTE-session-capture covers OpenCode event capture, local state and transport before server commit. It is neither that recall route nor the server archive/extraction phase. Workflow annotations reuse the runtime declarations rather than duplicate them.

Withdrawal, settings and claims. MEM-RTE-resource-withdrawal starts with service resource removal and linked-memory cleanup; EPI-RTE-memory-withdrawal starts with proposed memory deletes and replacement dependencies. Low-level filesystem removal is a shared executor, not route identity. Account-template admission described in the memory member's Write side controls extraction-template overrides; RT-RTE-config-patch controls runtime settings publication and model-resource selection. These mechanisms are not identical. RT-CLM-scoped-retrieval concerns query-scope construction. EPI-CLM-agent-knowledge, EPI-CLM-reusable-experience and EPI-CLM-memory-benefit concern different public/design claims; none duplicates that scope claim.

Integration issues and convergence

Corrected template admission does not conflict with retained-template consumption. The memory member's integration issues and Write side now distinguish managed edits from loading persisted overrides. Description, eligible content-template and selected field-description edits are admitted; changed embedding-template values are locked against deployment defaults. MEM-OBJ-memory-index still identifies an index consumer of schema embedding guidance, including possible older persisted configurations. That consumer finding does not require a managed embedding-template editing capability. Neither RT-RTE-config-patch nor the epistemic member's source-and-claim boundary, which leaves account overrides uninspected, asserts the contrary. The prior admission assertion has been corrected in its declaring member; no current dependent finding requires it. No alternate embedding-template editing route is established by joining these reports.

Ledger details extend a local inspection boundary. MEM-OBJ-recall-ledger and its annotation on context assembly supply file persistence, pruning, URI-only exemption and failure fallback where RT-RTE-context-assembly leaves internals uninspected. This is added implementation evidence, not a disagreement about operation or a universal recall-lifetime guarantee. The positive metadata path coexists with both reports' host-use limits.

Archive acceptance and content admission are separate transitions. RT-RTE-session-commit distinguishes accepted archive/queue work from completed extraction. MEM-RTE-memory-application, EPI-RTE-memory-write and EPI-RTE-memory-withdrawal converge on mechanical application controls and partial failures. Computational curation in the memory report does not claim truth acceptance; the epistemic report's separation of operational admission from epistemic acceptance therefore does not contradict it. Caller initiation is not per-item human approval in any of these accounts.

Withdrawal ordering adds coverage without contradicting a supplied positive. MEM-RTE-resource-withdrawal records physical resource deletion before the function named before_resource_delete, plus low-level bypass of service cleanup. The runtime left this maintenance unresolved, and the epistemic record concerns memory-operation deletes instead. Their connection establishes no universal derivative revocation, semantic-index convergence or retraction of already-delivered text.

Fallback evaluation and oracle uncertainty concern different facts. MEM-OBJ-trajectory and EPI-RTE-rollout-feedback converge that fallback passed/score measures extracted-trajectory existence, not task success. The epistemic record adds evaluator dispatch, feedback timing and confidence consumption. The memory and runtime accounts' uninspected expected-answer oracle remains compatible with that fallback and the optional evaluator branch: its concrete target, criterion and reference outcome are not supplied. No single evaluation domain follows from the shared result type.

Requested injection is not a delivered-channel finding. EPI-OBJ-experience-instructions quotes a schema requesting system-prompt injection and explicitly does not establish actual injection. MEM-BAP-experience-reminder identifies VikingBot's concrete user-role reminder. Those statements do not disagree. Likewise, EPI-RTE-digest-contract establishes URI-membership and presentation checks, not entailment by source text. RT-RTE-context-assembly and RT-RTE-plugin-recall provide its consequential consumers: no_relevant can suppress visible recall, and digest preference can omit the original fragments. A hit, checked URI or response entry is not interchangeable with delivered evidence or compliance.

Local capture recovery and server recovery have separate owners. WF-RTE-session-capture extends the memory annotation's local persistence findings: queued messages can count as captured, and commit can proceed without complete transmission. RT-RTE-session-commit describes durable processing of the session state that reached the server, not lossless capture of the host transcript. Their handoff supports the selected workflow account without establishing end-to-end exactly-once effects, eventual delivery or an actual item-level extraction-to-recall trace.

Successor wiring does not erase independent unknowns. RT-RTE-skill-admission, MEM-RTE-experience-maintenance and EPI-RTE-policy-training retain persistent, editable instruction products and successor-consumer wiring. Their theory-builder conditions, learning, reflection, autonomy and exercised self-improvement remain separately uninspected. The epistemic batch-flow evidence is additional architectural coverage, not evidence of actual candidate criticism, changed later behavior or benefit. The reported gains in EPI-CLM-memory-benefit are not run provenance for those routes.

Conflicts and remaining limits

No disagreement between the current supplied findings is established. No duplicate supersession or missing operative part requiring an unresolved-conflict paragraph is identified within their declared, bounded accounts. The template correction removes an admission assertion rather than a supported consumer path; the other integration issues add coverage or retain local uncertainty.

The memory member's limitations explicitly leave server auto-commit and resource-addition producer triggers/provenance incomplete. Its integration issues, the epistemic coverage boundary and the workflow account also retain alternate admission, parser/privacy/watch, evaluator/optimizer variant, provider, bot write/compaction and other-host limits. The reports do not jointly establish exhaustive producers, uniform policy controls, complete rationale preservation, host compliance or benefit. These limits remain limits, not absence findings or conflicts between their positive records.