Exo reconciliation

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

Reconciliation

Identity and overlap dispositions

Amendment: EPI-OBJ-memory-entry is superseded by MEM-OBJ-exo-memory. Both declarations identify the agent-scoped entries in memory/exo-memory.json, with text, ID and creation time in a versioned agent artifact. The epistemic declaration supplies no separate observed entry instance. Identity evidence is SRC-1 exo/tools/memory-tools.ts, including the artifact-path passage retained in the memory record. This connects the epistemic inventory and memory write/read/forget findings to the same retained entry collection; it changes neither the epistemic assessment nor the memory record's provenance limits.

Amendment: EPI-OBJ-todo is superseded by MEM-OBJ-exo-todos. Both identify conversation-scoped todos/exo-todos.json, natural-language tasks and symbolic statuses, rather than a separately supplied task instance. SRC-1 exo/tools/todo-tools.ts, with the path passage in the memory record, supplies identity evidence. The affected epistemic inventory and todo-state ledger remain assessments of the same retained plan. Neither declaration treats a completed status as verified work.

The combined EPI-OBJ-skill-record covers both skills/index.json and skills/<name>.json. MEM-OBJ-skills-index and MEM-OBJ-skill-record separately identify catalog metadata and payload. SRC-1 exoharness/typescript/harness/skill-tools.ts, with the path and lifecycle passages retained in memory, supports this container/parts relation. No identity with either single part is asserted, and no split supersession is needed: the combined record's findings remain compatible with the distinction. The epistemic install/read records and ledger also distinguish catalog, selected body and supporting-file access. Both required parts are declared; no missing-part conflict is retained.

The specialized routes are not duplicate identities of their broader containers:

EPI-RTE-todo-state and MEM-RTE-plan-replacement describe the same full-list proposal, structural validation, conversation-artifact write and conditional developer-message replay. Their different analytical emphases do not create different routes.

Amendment: EPI-RTE-todo-state is superseded by MEM-RTE-plan-replacement. Identity evidence is SRC-1 exo/tools/todo-tools.ts: the epistemic record retains the pending/in_progress counting passage and memory retains the no-open-items injection guard. Endpoints, owner, selection and withdrawal agree. The affected epistemic ledger functions remain separate assessments of checking, retention and consumption; supersession supplies no stronger verification of task completion.

EPI-RTE-harbor-task overlaps but does not replace WF-RTE-harbor-trial. The former traces instruction delivery, shared state and attempted export in SRC-1 eval/harbor/src/exo_harbor/agent.py and eval/harbor/src/exo_harbor/plugin.py; the latter additionally traces client configuration, container selection, subprocess timeout and export construction through eval/harbor/src/exo_harbor/exo.py and eval/harbor/src/exo_harbor/trajectory.py. The narrower caller account is not identical to that whole orchestration. Both remain alongside downstream RT-RTE-managed-turn and RT-RTE-responses-loop, neither of which owns Harbor delivery. WF-OBJ-trajectory is a separate event projection, not RT-OBJ-turn-work, an artifact-version store or a guardian outcome.

Other object comparisons do not establish identity. MEM-OBJ-artifact-versions describes the versioned storage mechanism, not each specific stored content set. MEM-OBJ-sandbox-snapshots and MEM-RTE-sandbox-rewind concern backend filesystem state, not native conversation recovery. Text skill records differ from executable RT-OBJ-tool-registry. EPI-OBJ-web-result and EPI-RTE-web-access add imported content and acquisition, not another declaration of the Responses loop. Their source anchors and operative consumers are retained in those records.

Independent convergence and integration issues

The native-compaction qualification in memory is compatible with runtime. RT-RTE-native-bridge leaves native algorithms and internal compaction uninspected; MEM-RTE-codex-compaction establishes the bridge's request between replay chunks and later consumption of opaque native state. SRC-1 exoharness/typescript/codex/replay.ts supplies the request passages. Known trigger/wiring does not expose the summary representation. Neither record depends on the other asserting complete context preservation.

Memory's corruption and withdrawal issues converge with the epistemic account. MEM-RTE-fact-revision and the epistemic memory routes agree on blocked corrupt writes and warning-based reads. The todo routes instead treat corrupt reads as empty. The skill lifecycle and epistemic skill-read route agree that uninstall removes catalog membership while direct supporting-file access can survive. These are different policies at different consumers, not contradictory findings about one universal policy. SRC-1 exo/tools/memory-tools.ts, exo/tools/todo-tools.ts and exoharness/typescript/harness/skill-tools.ts support the distinctions through the cited records.

MEM-RTE-result-offload and RT-RTE-responses-loop agree on the 8,000-character inline threshold and on offload not establishing universal automatic reload. The unresolved model-facing full-result reload question remains a limitation, not an absence finding. WF-OBJ-trajectory adds a further consumer transformation: export may use an already truncated preview and does not paginate beyond its requested event page. Full stored bytes, model-visible results and exported observations therefore cannot substitute for one another. The records retain their respective SRC-1 implementation passages; no full-history or full-evidence delivery conclusion follows.

RT-RTE-guardian-update, EPI-RTE-development-checks and EPI-CLM-recursive-improvement concern different mechanisms or claims. The runtime's lack of an independent improvement gate in the inspected guardian pipeline does not deny development hooks elsewhere. The epistemic record explicitly leaves ordinary guardian builds outside that hook gate. The public recursive-improvement rationale is not identical to the runtime's wired self-change route, and neither analyst reports observed improvement. SRC-1 exo/scripts/exo-service-guardian, scripts/precommit.mjs and docs/RSI.md are the respective anchors retained in those records.

Likewise, EPI-CLM-memory-authority records requested reliance, not an additional truth check or a duplicate of RT-BAP-harness-instructions. EPI-CLM-shared-trials records the learning rationale behind sequential trials, while the Harbor routes establish the configuration gate and state reuse. MEM-CLM-history-inspection is requested telemetry access, not the automatic delivery of RT-RTE-background-wakeup. These claim/route and requested/automatic distinctions rule out further identity supersessions.

Memory's remaining integration questions—opaque native payloads, alternate harness policy, hosted retention, UI/evaluator selection, adapter lifecycle, snapshot fidelity and local-profile initialization—remain bounded inspection or external-evidence limits. The other reports do not supply the missing facts or rely on them being established. In particular, workflow's detailed Harbor export wiring does not establish external grading, and epistemic's acquisition analysis does not authenticate scheduler-origin content. No system-complete negative or cross-backend parity conclusion follows.

Cross-report disagreements and correction dispositions

No unresolved cross-report disagreement was identified in these pinned members. All route implementation statuses remain wired, operation remains uninspected, and epistemic architectural status is a separate field rather than a stronger substitute. The reports share the static-evidence boundary in boundary; no supplied observed run or causal experiment changes those limits. Faithful uncertainty about opaque or uninspected alternatives is not treated as conflict.

No analyst correction-answer or declined-objection input was supplied to this attempt. The dispositions above reconcile the current reports afresh, including memory's six integration issues, without presuming a prior objection was established. No source files were read to resolve ambiguity; all connections use the supplied records and their retained passages. No analyst value or member is rewritten by this reconciliation.