Apache Maka report verification

Type: agentic-system-analyses/types/agentic-system-verification.md

Verification

The reports support a bounded implementation account. No publication blocker was found. I checked the handed reports and reconciliation afresh, including every declaration, source attribution, evidence status, annotation, containment relation and supersession. The supplied structural check says none; the four supplied answers files are empty. There is no correction argument or structural failure requiring disposition.

The boundary and Source register remain authoritative. Independent Git inspection confirmed the reviewed HEAD and clean checkout, and the tracked top-level tree agreed with the boundary inventory. Additional searches and reads used only SRC-1 at that revision. I did not execute Maka, tests, models or examples. Implementation and prompt text establish wiring and requested behavior, not observed compliance, successful recovery, task success or causal benefit. The reports preserve that separation; none declares an evidenced absence.

Runtime and authority

I traced CLI prompt admission, tool settlement, child creation, backend alternatives, composition, scheduled admission and recovery against their implementation anchors. RT-RTE-cli-turn correctly separates Session, Turn and Run identities and client observation from Host execution. SRC-1 packages/cli/src/runtime-host-run-command.ts requires the returned Run identity. RT-RTE-tool-settlement is supported by SRC-1 packages/runtime/src/ai-sdk-turn.ts and packages/runtime/src/tool-runtime.ts: the main batch waits for consumption, reads durable current-run evidence before dispatch, resolves active names, rejects provider-executed calls and drops raw settled results before continuation. The nested Code Mode settlement seam does not imply that arbitrary plugin I/O is mediated.

The live client-capability guard admits preparation under Bypass or ask; the report correctly avoids treating its diagnostic wording as a Bypass-only invariant. Managed filesystem and inspected Bash missing-sandbox branches fail closed at SRC-1 packages/runtime/src/filesystem-executor.ts and packages/runtime/src/builtin-tools.ts. These support the scoped enforcement in RT-BAP-tool-boundary, not universal process isolation. RT-RTE-executor-plugin correctly preserves the different executor and direct-LLM seams: SRC-1 packages/runtime/src/plugin-executor-backend.ts excludes Maka boundary requests, while packages/runtime/src/plugin-llm-service.ts can dispatch directly to a visible adapter. Acknowledgement/abandonment is not rollback of remote effects.

RT-RTE-child-graph retains separate child contexts and durable lineage. SRC-1 packages/runtime/src/session-manager.ts verifies parent admission, creates child metadata and rereads the parent before starting an unstarted child; executor-backed children receive no local tool list. Graph wakes have their own identity and admission checks in packages/runtime-host/src/server/agent-graph-execution-coordinator.ts. RT-RTE-durable-wake correctly distinguishes scheduled-fire admission from final task success and from recovery validation, supported by packages/runtime-host/src/server/scheduled-task-coordinator.ts and packages/runtime-host/src/server/hosted-execution-recovery.ts.

RT-CMP-ai-model, RT-OBJ-run-composition and RT-BAP-composed-prompt distinguish provider-resolved parameters, symbolic baseline metadata and interpreted prompt content. SRC-1 packages/runtime-host/src/server/execution-model-composition.ts constructs the summary model from the selected target and conditionally records the immutable baseline; it does not certify exact weights. packages/runtime-host/src/server/interactive-run-composer.ts caches base composition by Session/Turn and distinguishes child construction. RT-RTE-plugin-admission preserves desired-authority commit, runtime convergence, rollback and unknown-outcome fencing as different phases; its cited plugin-platform and composition-loader paths do not establish semantic improvement. RT-CLM-task-runtime remains an attributed README claim, not a verified comparative result. The runtime guarantee table is bounded by its enforcement owners and alternate paths.

Memory, evidence and later consumers

I checked the local bundle, extracted items, checkpoints, indexes, Matters and generated-file/archive objects against their cited producers, stores and consumers. MEM-OBJ-local-entries and MEM-OBJ-bundle-metadata correctly separate interpreted bodies from software-consumed lifecycle fields. MEM-RTE-local-memory and EPI-RTE-local-memory-use agree with SRC-1 packages/runtime-host/src/server/memory-coordinator.ts and packages/core/src/local-memory.ts: proposal review is not the sole write path, replacement has separate admission, and subsequent prompt selection uses policy, active status, session scope and a cap rather than task relevance. Producer-written approval metadata does not authenticate human review.

MEM-OBJ-memory-item, MEM-OBJ-item-index and the broader EPI-OBJ-memory-item preserve retained text, lookup/control metadata and transient candidate scope. I independently followed MEM-RTE-fact-extraction through EPI-RTE-memory-evidence, EPI-RTE-memory-canonicalize, EPI-RTE-memory-disposition and EPI-RTE-memory-commit. SRC-1 packages/runtime/src/memory-extraction-proposal.ts checks lexical citation membership, not entailment. packages/runtime/src/memory-extraction.ts builds canonicalization input from evidence without original candidate wording, checks result identities, reattaches citations, repeats admission and rechecks eligibility before commit. This supports the reported fallible evidence-consuming acceptance criterion, not external factual verification or demonstrated accepted novelty.

MEM-RTE-recall-history correctly identifies the actual later consumer. SRC-1 packages/runtime-host/src/server/execution-composition.ts projects stored facts to content/kind/time and omits stored source references. packages/runtime/src/recall-tools.ts excludes the current turn, while archived sessions and archived fact items have different rules. The cited core and storage paths support literal transcript matching, key-prefix fact selection and active-by-default fact lookup. MEM-RTE-session-import correctly stops short of native-trace equivalence or proving all downstream import consumers.

RT-RTE-context-projection, MEM-OBJ-text-checkpoint, MEM-OBJ-provider-checkpoint and MEM-RTE-archive-read retain distinct effective-history, text, opaque-state and reload roles. SRC-1 packages/runtime/src/ai-sdk-compaction.ts selects over effective history and rereads transitions after writes; packages/runtime/src/history-compact-checkpoint.ts distinguishes text-event replay from provider state. The archive resource and ledger-reader anchors enforce session/source/digest/size checks, not semantic fidelity. MEM-BAP-local-guidance is supported by workspace and skill selection/loading anchors; static instruction delivery is not automatically trace-fed learning.

MEM-OBJ-index-documents, MEM-OBJ-index-cursors and MEM-RTE-index-maintenance correctly separate body edits, directory navigation and coverage checkpoints. SRC-1 packages/runtime-host/src/assistant/memory-network/host.ts, corpus.ts and controller.ts support revision/range rejection and continued work from saved notes; coverage completion is not a semantic completeness certificate. MEM-OBJ-extraction-files and MEM-RTE-history-extract are independently saved products, not implicit index adoption. The extractor's generated/needs-review/truncated status does not certify truth.

MEM-OBJ-matter-state, MEM-OBJ-matter-reviews and MEM-RTE-matter-state agree with the Matter store, tools, prompt and review anchors: intermediate checkpointing differs from reviewed settlement; rejection returns feedback; subsequent activations are directed to current state files. The review prompt bounds execution history and does not establish an external oracle. MEM-OBJ-review-archive correctly distinguishes model-generated daily-review sections from fallback publication and does not infer automatic agent-method revision.

Epistemic and workflow conclusions

EPI-OBJ-goal-evaluation, EPI-RTE-goal-judge and EPI-RTE-goal-selection are supported by SRC-1 goal-evaluator, goal-continuation and Host goal-coordinator paths. The judge receives recent projected conversation, not independently executed checks. Parsed flags can stop/wait/continue; evaluator failure is excluded from real progress accounting; pending steering can be retained. EPI-CLM-goal-judge remains doctrine, not demonstrated prevention of premature completion. EPI-OBJ-local-memory remains a valid bundle container, not an assertion that all its fields reach the model.

EPI-OBJ-verifier-result and EPI-RTE-eval-verification correctly separate imported reward from verification completion. SRC-1 packages/eval/src/harness-executor.ts permits a present zero score to receive completed status; packages/eval/src/runner.ts avoids verification of uncertain subject states and consumes verification in reporting. Neither route validates the external oracle or establishes causal benchmark advantage.

The selected workflow is coherent rather than assembled from unrelated capabilities. WF-RTE-patch-publication is supported by SRC-1 packages/runtime/src/session-manager.ts and packages/runtime-host/src/server/execution-artifacts.ts: latest-terminal/quiescent workspace capture publishes an artifact, not parent application or semantic approval. The workflow annotations preserve explicit child output and foreground client delivery. Structural result decoding, terminal state and artifact publication do not become task-quality verification.

Relations, omission questions and stopping decision

The reconciliation retains the one local-memory route supersession with matching identity evidence. I accept it. The epistemic bundle and candidate containers remain valid alongside finer memory parts; their findings do not require split supersession. All five declared Part of: relations are supported: the two checkpoints and reload branch belong to effective-history control; local guidance belongs to composed prompt authority; patch publication belongs to child work. Annotations do not replace their declaring records' statuses. No unresolved cross-report contradiction was found. Additional analyst inspection resolves earlier inspection gaps without making those earlier scoped uncertainties contrary findings.

Omission questions were selected from the frozen tree, alternate execution paths and producer/consumer dependencies. Could plugin/client execution invalidate a uniform sandbox claim? The alternate seams are already distinguished, so no such universal claim survives. Could non-CLI judging refute the workflow's verification account? Goal, Eval and Matter routes are consequential but distinct; the selected generic foreground-child path does not require their semantic approval. SRC-1 packages/runtime-host/src/assistant/proactive-matters/host.ts scopes its before-finish gate to an active Matter activation, rather than every child turn. Could retained state be counted without an actual later reader? Fact projection, checkpoint replay, index body reads and Matter file workflows provide the positive consumers; independent extraction files and daily archives are explicitly narrower.

A further search of assistant scheduling found an ordinary-conversation heartbeat alternative, recorded below as a local inventory limit. Further omission expansion stopped because these selected challenges have dispositions and no known contradiction or unsupported emitted comparison remains. Uninspected source adapters, platform internals and external verifier bodies still prevent universal coverage, isolation, withdrawal and reliability conclusions; they do not negate the positive paths traced here. This is not an exhaustive inventory or audit.

Blockers

none

Limits

• The assistant inventory remains incomplete beyond MEM-RTE-index-maintenance, MEM-RTE-matter-state and RT-RTE-durable-wake. Independent inspection found a distinct ordinary-session heartbeat path in SRC-1 packages/runtime-host/src/assistant/index-initiative/host.ts, controller.ts and prompt.ts: it resumes the configured chat, admits a follow-up and waits for its history/idle boundary rather than creating an independent index worker or requiring a Matter checkpoint. The reports already mark detailed heartbeat coverage uninspected and make no complete autonomy or read-back claim, so this does not invalidate their findings. Withhold a complete comparison of proactive triggers, retained-context reuse and assistant delivery modes; do not generalize Matter settlement or index-worker isolation to all assistant work.

  const agent = await this.asOwner(s, () => this.ctx.agents.resume({ sessionId: s.sessionId })); --- packages/runtime-host/src/assistant/index-initiative/controller.ts const admitted = await agent.followup(wake(state)); --- packages/runtime-host/src/assistant/index-initiative/controller.ts