OpenWiki report verification
Type: agentic-system-analyses/types/agentic-system-verification.md
Verification
The current reports support publication as a bounded source-level account. The workflow correction removes the remaining unsupported conversation-isolation inference. The pending-page projection limit remains local; per-submission durability and final whole-run checks still hold within their stated filesystem and dependency assumptions.
I read every supplied member, correction answer, type, contract and the previous verification completely. I assessed all declarations, annotations, source anchors, evidence statuses, containment relations and reconciliation dispositions in the runtime records, memory records, epistemic records and workflow account. The structural report says none; there is no structural failure to route. Runtime, memory and workflow supply corrected answers; epistemic's answer file is empty. Answers and the earlier verification were treated as arguments, not proof.
I independently checked checkout HEAD against the Source register, enumerated the top-level frozen tree, and traced consequential producer, admission, persistence and consumer behavior in SRC-1. Source reads and searches below add implementation and source-native doctrine/design coverage, not observed operation. The offload tests were read, not run. No target, example, test, model provider, connector service or evaluator was executed. Deterministic citation and quotation checks remain code's responsibility; occurrence alone was not treated as claim support. The boundary remains supported: OpenWiki includes shipped production, maintenance and consumption callers, while external host and provider internals are not attributed to it.
Runtime and workflow
RT-RTE-native-generation matches dispatch and runner construction in SRC-1 src/agent/index.ts and src/agent/repository-runner.ts. Direct reads of src/generation/repository-run.ts establish begin/resume, source-drift replanning, canonical plan admission, pending-job selection, snapshot restoration, submission and finish. The model proposes semantic scope and content; fixed code owns queue progression. RT-OBJ-repository-run-state and RT-OBJ-page-guidance correctly separate code-interpreted statuses from interpreted task guidance in src/generation/run-state.ts and src/agent/repository-prompts.ts. Only quickstart gets the complete planned page map. RT-BAP-page-instructions establishes delivery, not adherence.
The workflow's corrected stage 3, retry account and dependent annotations now distinguish selected initial prompts and fresh uncheckpointed graphs from possible conversation-derived retention. I rechecked both worker builders, retry restoration and stream forwarding in SRC-1 src/agent/repository-runner.ts, the shared mount in src/agent/agent-backend.ts, home resolution in src/config/openwiki-home.ts, and the complete test/agent/conversation-history-offload.test.ts. MEM-OBJ-conversation-history and MEM-RTE-history-offload retain the load-bearing mount and direct-write/read-back passages. Native planners and page workers share the configured home-backed history surface. Tool-write denial does not deny middleware backend writes. Page/Claim snapshots do not restore that surface. The current workflow no longer claims that all research conversations are erased or cannot cross stages. Actual payload, production trigger, automatic reload and cross-worker use remain unresolved, not denied. Runtime and memory's dependent findings preserve the same distinction. All three corrected answers are substantiated.
SRC-1 src/agent/repository-runner.ts also supports process-local claimed-job acquisition, quickstart's deferred single-slot pass when concurrent, two-attempt recovery, rate-limit retry bypass and pool reduction, bounded tool-event forwarding, and fatal-error propagation before finish. A successful submission remains successful if the subsequent stream fails. These controls do not prove cross-process reservation, successful recovery in operation or host compaction behavior.
RT-CLM-native-confinement is supported over model-facing interception paths in SRC-1 src/agent/docs-only-backend.ts, src/agent/agent-backend.ts and the worker builder. Assigned-page grants, personal shell refusal and the repository shell allowlist are not OS isolation or backend-wide history-write denial. RT-RTE-host-generation correctly assigns research and physical writes to the external host: src/integrations/core/session-manager.ts checks active run identity and guards adapter operations, while integrations/openwiki/SKILL.md requests sequential assigned-page work and host sandbox compliance. Neither source establishes compliance or per-content human approval. RT-CMP-chat-model distinguishes model-ID configuration from immutable served parameters; provider-side mutability remains uninspected.
RT-CLM-page-durability follows Claims finalization, page durability assertion, manifest completion and queue write ordering in SRC-1 src/generation/repository-run.ts. Finish refuses pending work, requires skipped-page snapshots, finalizes artifacts and Claims, replaces manifest coverage and removes resumable state last. Skipped restored pages and source drift can coexist with a complete return and interrupted metadata. src/generation/page-manifest.ts preserves prior source checkpoints for untouched covered pages. These are bounded protocol guarantees, not semantic truth or a whole-run transaction.
The alternate routes remain distinct. RT-RTE-personal-ingestion follows deterministic pulls and fresh-thread local-wiki update in SRC-1 src/ingestion/ingestion.ts; it does not use repository page submission. RT-RTE-mcp-connector and RT-CLM-mcp-policy follow allowlist/annotation/Notion eligibility and configured stdio startup in src/connectors/mcp-runtime.ts and src/connectors/mcp-client.ts. shell: false and client cleanup do not prove server nonmutation or process isolation. RT-RTE-page-translation follows direct model invocation, empty/token-limit refusal, rewrite and retry marking in src/agent/translation-middleware.ts; semantic fidelity is requested, not checked independently.
RT-RTE-maintenance-admission correctly separates local lifecycle admission from Git publication. SRC-1 .github/workflows/openwiki-update.yml can propose progress despite CLI failure and deletes transient queue state. src/ingestion/code-mode.ts supplies managed instructions and connector enrichment; src/agent/skills.ts refreshes bundled skills while preserving unrelated directories. RT-BAP-managed-guidance establishes a host-facing instruction surface, not observed loading or an actual merge decision.
Memory and epistemic warrant
MEM-OBJ-claim, MEM-OBJ-evidence, MEM-OBJ-page-claims and MEM-OBJ-page-manifest match the statement, resource/version, envelope and coverage-ledger distinctions in SRC-1 src/claims/brains/code/store.ts, src/claims/core/mutations.ts, src/claims/brains/code/session.ts and src/generation/page-manifest.ts. Strict sidecars contain current state, not a mandatory criticism or retraction-rationale history. Prose may separately retain reasons. EPI-OBJ-page-claims remains a valid mixed container alongside those parts.
MEM-RTE-claims-reconciliation, MEM-BAP-claims-maintenance and the five native-generation Part of: records in the epistemic declarations have supported progression and containment. SRC-1 src/claims/brains/code/preflight.ts, src/generation/page-jobs.ts and src/generation/repository-run.ts establish version issues, explicit decisions for issue-bearing Claims, automatic confirmation of omitted issue-free Claims, rejection of an empty final set, and optional compact inspection without opaque versions. Shared host use does not make these subfunctions duplicates of a complete invocation. Version resolution does not judge statement entailment.
The epistemic ledger appropriately separates checking, protocol disposition and retention from semantic acceptance and lifecycle integration. SRC-1 src/claims/brains/code/session.ts and src/claims/brains/code/runtime.ts recheck evidence, persist eligible state and reject warning-bearing finalization. EPI-OBJ-verification-events and EPI-RTE-verification-projection match src/okf/claims-verification.ts: foreign events survive; absent active events remove producer-family stamps without removing prose; unsafe active stamps can be rolled back after failed hash refresh. The pending-page propagation exception is retained below.
EPI-OBJ-page-sources and EPI-RTE-source-projection match whole-file mapping, deduplication and preservation of independent entries in SRC-1 src/okf/claim-sources.ts. The concrete examples are code deductions, not observed fixtures. EPI-CLM-grounded-claims accurately retains the README's repository-evidence scope and connector-derived-fact exclusion. Current evidence identity, a nonempty Claim set and durable verification do not establish independent semantic acceptance of a produced ampliative claim.
MEM-OBJ-wiki-page, RT-RTE-wiki-retrieval and RT-BAP-retrieved-sections preserve the distinction between stored prose, compact search display and full-section delivery. I traced SRC-1 src/retrieval/wiki.ts: ephemeral FTS5 matching, path/coverage/BM25 ordering, Markdown reload and explicit requested-section extraction. Search's deprecated filter is not repeated by direct read. No sidecar proof package or total section-token budget is supplied by those functions. integrations/openwiki/SKILL.md requests current-source verification and non-instruction treatment; src/agent/prompts/personal.ts requests wiki-first answers and narrow raw fallback. Prompt policy establishes neither compliance nor warrant.
MEM-OBJ-connector-data matches SRC-1 src/connectors/io.ts and src/connectors/tools.ts: twenty retained run-list entries do not imply raw-file deletion, and raw reads slice decoded strings. MEM-OBJ-trace and MEM-RTE-trace-documentation match anomaly-weighted selection and compact structure/metrics/errors in src/connectors/sources/langsmith/index.ts and src/connectors/sources/langsmith/runs.ts, followed by code-mode synthesis guidance. This is a supported trace-fed authoring capacity, not observed criticism, learning or causal improvement.
MEM-OBJ-chat-checkpoint and MEM-RTE-chat-continuation match chat disk persistence, same-session thread supply and SQL cleanup in SRC-1 src/agent/index.ts and src/cli/app/app.tsx. Removing older snapshots does not remove old turns from surviving state or erase separately mounted history files. Infrastructure wiring, afforded history access and uninspected payload/operation are appropriately separate statuses.
Other consumers, reconciliation and omission disposition
MEM-RTE-visualizer-delivery matches real-file collection, body projection and bounded node/edge construction in SRC-1 src/visualize/graph.ts; full browser fidelity is not claimed. MEM-RTE-evaluation-transfer matches archive validation, compatible-cache selection, generation/setup, wiki/AGENTS transfer and the downstream caller in evals/deepswe/openwiki_codex.py. Delivery does not establish uptake or benefit; compatible caching does not imply current-package regeneration.
MEM-RTE-ledger-evaluation, EPI-OBJ-ledger-verdicts, EPI-RTE-ledger-judgment and EPI-RTE-ledger-reporting retain supported distinct functions. SRC-1 evals/ledger/evaluator/precision.ts and evals/ledger/evaluator/prompts.ts establish bounded current/historical source judgments and nondegraded caching. evals/ledger/run/evaluate-checkpoint.ts retains verdicts, metrics and obsolete watch state. evals/ledger/run/runner.ts passes artifact/evidence/carry to evaluation, but not scores/carry to the next product update. A benchmark semantic evaluator thus coexists with no independent semantic answer oracle in the inspected production admission loop. Neither evaluator accuracy nor causal benefit follows.
The reconciliation correctly supersedes EPI-OBJ-wiki-markdown by MEM-OBJ-wiki-page on referent identity without transferring warrant. Valid Claim, maintenance and evaluator containers remain alongside their parts. Guidance, retrieved context, SQLite and filesystem history are not aliases. I found no further required supersession, invalid containment or undeclared material part undermining the emitted findings. No analyst declares an evidenced absence. The prior context conflict is resolved by current wording; the pending-page projection qualification is consistent across reports. Different retrieval/evaluator inspection coverage remains a refinement, not disagreement. No current unresolved conflict needs routing.
Selected omission questions asked whether fresh workers have another retained-context channel; whether verification or retrieval licenses truth or universal withdrawal; and whether shipped evaluator consumers feed production revision. Their search basis was the frozen tree, alternate native/host builders, backend/home dependencies and adjacent offload tests, Claim producer/projector consumers, direct section reads and evaluator caller dependencies. The history question supports the corrections without proving exercised reuse. Warrant and withdrawal remain bounded by their actual code consumers. Evaluator feedback is confined to benchmark reporting/carry in the inspected sequence. Further optional omission expansion stopped because each identified material challenge has a supported disposition and no unsupported consequential conclusion remains. Connector-specific synchronization, personal curation, dependency summarization, browser packaging and cross-platform scheduling remain partial inventories: no absence, equivalence or exhaustive coverage conclusion is licensed for them. Missing operation and experiment evidence continues to prevent empirical reliability, learning and benefit claims.
Blockers
none
Limits
- RT-CLM-page-durability, the workflow page-admission account and EPI-RTE-verification-projection do not establish complete pending-page noninterference. SRC-1
src/generation/repository-run.tssupplies other pending pages as exclusions;src/claims/brains/code/session.tshonors them for session persistence and omits their eligibility entries. However,src/claims/brains/code/runtime.tscalls the verification projector without that exclusion set, andsrc/okf/claims-verification.tsvisits all discovered pages. An omitted active event can remove an existing OpenWiki stamp on another pending page. Withhold the inference that pending Markdown stays untouched or that exclusions uniformly prevent transient metadata writes. Per-submission durability ordering and final whole-run checks remain supported. The corrected workflow and reconciliation acknowledge this qualification; no observed race, semantic corruption or failed delivery is established. The cited epistemic record retains the load-bearing traversal and missing-event passages.