OKF Harness synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From source registration to retained knowledge
OKF Harness is a local, agent-maintained wiki system, not an enclosing model runtime. This account covers its shipped source-registration, wiki-maintenance, retrieval, guidance and recovery paths at repository commit 0dd4f806c454a68088c5a3e64b4cbebbe2c216be, with an applicability cutoff of 2026-10-15. The sole source, SRC-1, supplies inspected implementation and declared workflows. Tests were read as executable expectations, not run. No host-agent session, target execution or causal experiment was supplied. The whole-system boundary includes shipped callers and agent instructions, but inspection is selective rather than a security or distribution audit. See boundary and its Source register.
The architecture divides semantic work from deterministic machinery. A host agent reads sources, writes wiki claims, judges conflicts and answers questions. TypeScript running in Node registers inputs, plans from metadata, reloads workspace state, checks structure and integrity, selects evidence and invokes Git recovery. The target does not own the host's model calls, scheduling, permissions or context lifecycle. RT-CMP-core, RT-CMP-host-agent.
On the workspace-local route, a skill tells the agent to sequence requested workflows, inspect command JSON and load only the reference needed for the current step. Native host guidance instead uses a setup launcher that dispatches the workspace's pinned CLI package through npm/npx. The native skill does not install the detailed workspace-local references, so these instruction payloads cannot be assumed equivalent. A runtime version selects software, not model weights or authenticated registry bytes. RT-BAP-workflow-guidance, RT-RTE-guidance-install, RT-RTE-runtime-launch.
Ingest first preserves an input, then asks the agent to interpret it. File registration hashes bytes, reuses a same-hash entry or creates a raw copy exclusively and appends its manifest record. This protects creation through those APIs, not against subsequent shell edits or restoration. URL registration creates a pointer document, not a webpage snapshot; reuse of a URL does not acquire its changed body. The ingest plan proposes paths and at most five concepts from metadata. It neither reads source bodies nor generates wiki claims. The agent must perform the semantic reading and bounded edits, then check and report them. Saved edits precede checking and are not automatically undone by a failed check. RT-RTE-source-ingest, MEM-OBJ-source-bytes, MEM-OBJ-url-metadata.
The retained knowledge is Markdown prose, with separately consumed symbolic paths, frontmatter and citation fields. Source manifests and reconciliation records are local JSONL; checkpoints retain Git trees and caller judgments. A hash proves recorded byte identity, not source truth. Neither a citation nor structural validity proves that a page faithfully represents its source. Actual source-to-claim transformations remain indeterminate: no produced input/output claims establish whether the agent preserved, derived or conjectured their content. MEM-OBJ-wiki-content, MEM-OBJ-wiki-metadata, MEM-OBJ-manifest, EPI-OBJ-wiki-page.
Maintaining a changed source
The representative multi-step workflow reconciles a registered file revision with the wiki. It is a connected shipped procedure, not an observed successful run. A check identifies a suspected prior→revision pair from same-filename registrations with differing hashes and exposes promoted-reference paths. That diagnostic hands identities to the host agent, not a semantic diff. The agent is directed to read both versions and every affected wiki document, revise the prose, run check, repair edit-caused findings and finally request an edge-specific acknowledgment. Omission alone is not evidence of removal; the procedure requires a positive textual signal. There is no separate reviewer or isolated worker handoff. See workflow and RT-RTE-source-reconcile.
Two completion products must remain distinct. Edited prose supplies later answers. The acknowledgment appends exact endpoint IDs and a nonempty note, then removes that pair from later pending-edge calculations. Core can reject invalid IDs, an unknown pair, an empty note or malformed ledger state, but does not inspect whether the agent corrected the wiki. The returned acknowledgment is therefore a protocol receipt, not verified semantic completion. Pending revision currency does not itself withhold the old prose from evidence; the maintenance-loop test explicitly expects unchanged evidence eligibility across acknowledgment. MEM-OBJ-reconciliation-edges, MEM-OBJ-reconciliation-notes, MEM-CLM-currency-evidence.
Recovery is caller-selected, not automatic rollback after check or acknowledgment failure. A checkpoint stages all Git-visible workspace changes, potentially including sources, configuration and instructions. Restore requires clean pending Git state, selects a prior tree and records the restoration as another commit. History exposes judgment bodies to callers; ledger notes have no established later semantic reader. Retaining reasons consequently does not establish an iterative criticism loop. RT-RTE-workspace-recovery, MEM-OBJ-checkpoint-tree, MEM-OBJ-checkpoint-judgment.
Bringing knowledge back into an answer
A question leads the host to request an Evidence Brief. Core reloads the workspace, runs the shared check pipeline, blocks on conformance failures, applies source-risk exclusions and ranks eligible wiki concepts lexically. It returns at most three excerpts, provenance fields, warnings, thin candidate cards and continuation cues. Candidates are not supporting evidence bodies or an exhaustive match list. Excerpt budgeting defaults to 400,000 characters, with presets and overrides; metadata and the host conversation are outside that budget. The agent is asked to perform at most one automatic cue-following read and then judge sufficiency and conflicts. This read limit is policy, not a call counter. RT-RTE-evidence-answer and its memory annotation.
Withholding has a precise enforcement boundary. Missing or drifted registered sources can seal directly linked concepts and incoming citation dependents within the evidence planner. These computed hops are not semantic contamination closure. Standalone search uses an empty exclusion set; direct read does not apply the planner's seals. Shell access is another counterpath. The agent must widen withholding when prose shows further dependence. These controls establish planner-local eligibility, not universal access denial or rejection of false propositions. RT-CLM-evidence-boundary, RT-RTE-direct-retrieval.
Selection is requested read-back, even though code chooses excerpts automatically. Wiki prose supplies knowledge at the answering agent and ranking input at the lexical scorer; symbolic fields govern identity, provenance and eligibility. Just-in-time reference loading and excerpt selection keep material out of a particular return. They do not remove accumulated turns, implement host compaction or establish a separate model frame whose context is discarded on return. Legacy log.md is excluded from normal search/evidence/graph and direct read despite instructions permitting some log edits. RT-BAP-evidence-context, MEM-OBJ-wiki-metadata, MEM-CLM-log-exclusion.
Retained contribution and assessment
The strongest supported contribution is an available persistence-and-consumption path: agent-authored wiki changes can reach later answering through executable retrieval. A separate, narrowly conditioned protocol permits keeping a conversation-derived answer. Only the no-match permission branch allows the offer; withheld/truncated results or a continuation read do not. Explicit user yes permits one concept and its index link, while silence/no permits no persistence. Yes means the answer is worth keeping, not certified true. The conversation is not captured as a raw source. This positive knowledge-retention route must not be mistaken for either demonstrated improvement or automatic trace learning. RT-RTE-evidence-answer, MEM-OBJ-wiki-content.
Citation-empty write-back also exposes a provenance limit. Workflow text tells the agent to disclose conversation origin when the parsed provenance is empty, but a manually authored unsourced page yields the same payload. Empty fields show the end of the parsed trail, not authenticated authorship. MEM-CLM-provenance-signal, EPI-CLM-provenance-origin.
For a source-revision maintenance scenario, the system provides concrete identities, persistent inputs, structural diagnostics and an exact completion receipt. Whether that procedure produces correct revised knowledge remains unobserved. For answering amid source damage, it supplies bounded, risk-filtered excerpts, conditional on represented dependencies and the agent following the intended route. For establishing accepted novel knowledge, its deterministic checks are insufficient: they evaluate structure, byte integrity and bookkeeping, while graph derivations establish parsed connectivity. No evidenced semantic acceptance decision or post-acceptance integration of an ampliative claim is supplied. MEM-RTE-workspace-check, MEM-RTE-graph-report, epistemic conclusion.
Learning, theory building, reflection, autonomy and self-improvement are separate questions, not a combined grade. The runtime records establish retained outputs and later delivery, but contain no formulated-theory assessment establishing theory-builder conditions 1–4. They likewise do not establish that criticism of a consumed theory improved future capacity, or a reflective or autonomous theory builder. The positive maintenance and retrieval paths remain supported. For self-improvement, changed wiki content has an available later behavioral-authority path, but no objective-specific update trace and subsequent operation demonstrate causal dependence on the change. Exercised self-improvement and successful improvement are therefore unestablished, not evidenced absent. Arbitrary input origins and caller control also prevent a complete negative about automatic trace-fed learning. RT-BAP-evidence-context, RT-RTE-source-reconcile, profile verification.
A provenance-linked agent run showing source reading, changed claims, checks, acknowledgment and later use would change the assessment of exercised maintenance. An evidence-consuming semantic decision would permit assessing accepted knowledge. An objective-defined update and controlled later comparison would permit assessing improvement at that comparison's granularity. Authenticated authorship, a semantic acknowledgment gate or broader enforced read controls would change the corresponding architectural limits. These are resolving evidence or system changes, not adoption recommendations.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No target execution, semantic session or causal comparison | SRC-1; RT-RTE-source-ingest, RT-RTE-source-reconcile, RT-BAP-evidence-context | Frozen code, workflow text and test definitions; no observed-run layer | Successful semantic maintenance, activation, answer benefit, improved future capacity or exercised self-improvement | Provenance-linked input/update/later-consumer traces; objective-specific comparisons for benefit and causal attribution |
| External host/model behavior and instruction payload differences | SRC-1; RT-CMP-host-agent, RT-RTE-guidance-install, RT-RTE-runtime-launch | Shipped adapters, launcher and workspace/native guidance; host/provider internals and registry artifacts excluded | Actual prompt loading, model fixity or updates, host isolation/compaction, instruction parity, authenticated deployed package identity | Host-specific loading and permission records, provider/version evidence and authenticated deployed artifacts |
| Partial instruction-withdrawal inventory, carried from report verification | SRC-1; RT-RTE-guidance-install, memory inventory, report-verification limit | Installation/synchronization account omits separate characterization of bootstrap uninstall; verification identifies ownership/status-checked host-entrypoint removal and dry-run | Exhaustive instruction-withdrawal or cross-host recovery characterization; installation and wiki-memory findings remain valid | A canonical route account of uninstall admission, effects and host recovery boundaries |
| Structural checks and ledger acknowledgment do not certify semantic changes | SRC-1; MEM-RTE-workspace-check, RT-RTE-source-reconcile, MEM-CLM-currency-evidence | Snapshot predicates, exact-pair append/filtering and maintenance-loop assertions | Ready/currency-sealed as truthful knowledge, proof of source fidelity or acknowledgment as verified reconciliation | Claim-level evidence-consuming evaluation; wiring that makes semantic results consequential at acknowledgment/admission |
| Planner-local, bounded withholding and excerpt-only budgets | SRC-1; RT-CLM-evidence-boundary, RT-RTE-direct-retrieval, RT-RTE-evidence-answer | Evidence selection and direct search/read counterpaths; source/citation graph and selected text | Universal read denial, complete semantic contamination closure, complete answer context or host-token/window bound | Dependency and alternate-path coverage, enforced shared gates if intended, and host-context accounting |
| Unresolved origin, encoding and admission of included retained material | SRC-1; MEM-OBJ-source-bytes, MEM-OBJ-wiki-content, MEM-OBJ-checkpoint-tree, RT-RTE-workspace-recovery | Arbitrary imported bytes, historical prose, broad checkpoint payloads and generic callers | Complete form/lineage/control classification, specific ingest curation transformation, or complete trace-learning/source negative | Original-input content and provenance, authorship/admission records and qualifying automatic-write/later-consumer traces |
| Retained reasons have incomplete semantic read-back | SRC-1; MEM-OBJ-reconciliation-notes, MEM-OBJ-checkpoint-judgment, MEM-OBJ-wiki-content | Ledger filtering consumes IDs; history returns judgments; wiki reasons are optional and excerpt-dependent | Retention as iterative criticism, complete note authority/read-back classification or theory-builder conditions | Named rationale consumers and traces of content-directed criticism, changed reliance/revision and reuse in a next round |
| Empty provenance does not authenticate conversation origin | SRC-1; MEM-CLM-provenance-signal, EPI-CLM-provenance-origin | Parsed citations/source pointers and disclosure convention | Unique origin attribution or factual certification of an unsourced answer | Authorship/provenance capture distinguishing compliant write-back from other unsourced pages |
| Selective implementation coverage, not universal security or evaluator coverage | SRC-1; RT-RTE-guidance-install, EPI-RTE-doctor-check, MEM-CLM-log-exclusion | Selected mutation/retrieval/integration branches; individual probes, filesystem races and distribution parity not exhaustively audited; logs excluded from normal answer retrieval | Complete security, reliability, integration/probe or trace-to-answer inventory | Targeted branch/probe audits and host evidence; an actual consumer route for any claimed log-based memory |