OKF synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From a task to selected project knowledge
OKF is an agent-operated knowledge ecosystem, not an enclosing model runtime. It gives project explanations a durable home in Markdown, supplies retrieval and structural checks, and asks an external agent to keep the content current. This account covers the four shipped gems, skills, plugin and Pro installation templates at commit cfb7ffac7fef37d56e97b5b7dd8fe97ffc648c7c, with an applicability cutoff of 2026-11-08. SRC-1 is the sole registered source. The evidence is inspected implementation and authored doctrine/design, not supplied session traces, passing test executions or causal experiments. The host conversation loop, model internals and optional backend internals are outside inspected implementation; see the boundary and Source register.
A user brings a task to an agent in a project with an OKF bundle. The skill routes intent to a playbook. Consume requests directory shape, a relevant branch index and recent history, then only the concept bodies the task needs; pointed questions transfer to search. The CLI parses visible Markdown and returns maps, matches or selected bodies. The external agent chooses subsequent steps, interprets the content and delivers the task answer. There is no OKF-owned answer oracle or terminal-answer evaluator in this route. The knowledge side and the host's task-answer side therefore have separate delivery boundaries: RT-RTE-skill-work, RT-CMP-agent, WF-RTE-maintenance-closeout.
The durable corpus separates concept prose from provenance and lifecycle fields, authored indexes, and update logs. Pro adds a work board, daily journal, retained counter snapshots and friction events. Natural-language explanations guide agents and humans; symbolic fields separately control filters, counts and admissibility checks. A source reference records an origin claim, not proof that the source supports the prose. The body and its metadata thus have different consumers and warrants: MEM-OBJ-concept-body, MEM-OBJ-concept-metadata, MEM-OBJ-index, MEM-OBJ-log.
MCP supplies shipped tools, resources and prompts for later consumption. Requested searches filter and project results; exact-id concept reads return original Markdown. A host can request a resource attachment without the model deciding to fetch it. Browser and TUI support human consumption and registry management. These are concrete retrieval interfaces, not proof that delivered content changed an agent's behavior: RT-RTE-mcp-retrieve, RT-RTE-registry-manage, MEM-BAP-retained-knowledge.
Selection keeps material out before it enters context; it does not establish control over the host's accumulated conversation. MCP row projections and log cuts, filtered hook feedback and computed opening state are distinct supply controls. The log byte bound is per file's content and scales with the requested limit; it does not bound the complete response, full concept reads, resource reads or host window. Cache eviction removes Ruby residency, not already delivered turns. Live body bytes also do not guarantee fresh selection metadata: path/mtime/size fingerprints can miss preserving edits, reserved-log reads can fall back to a loaded snapshot, and browser structural views are boot snapshots. SafeRead's containment checks do not cover hardlinks or the resolve/open race. These are implementation qualifications, not observed stale or escaping executions: RT-RTE-mcp-retrieve, RT-CLM-live-bodies.
From a discovery to maintained files
The representative multi-step workflow is connected by an explicit handoff: a durable discovery during consume triggers maintain. The same agent finds affected concepts, updates bodies and attribution, repairs enumerations and links, and appends a dated log entry. It then checks the index, runs validate and lint, and judges loose concepts and tags. This is procedural delegation to commands, not an evidenced handoff to another model worker. Validate can reject conformance with exit 1; lint is advisory by default unless a failure threshold is selected. Direct edits already persist before these checks. Recovery is agent correction and rechecking, not automatic rollback: RT-RTE-skill-work, WF-RTE-maintenance-closeout.
Acquisition and maintenance have several forms. Produce requests distillation from code, documents or human knowledge; document adoption requests preservation of existing bodies, with permitted relocation-link edits. That adoption branch must not stand for every migration mode: the report verification identifies a spelling-migration mode that moves citations into metadata and deletes their old body section. The transformation inventory is consequently incomplete. Correction and deprecation can preserve superseded reasoning for later readers, but distillation alone does not establish a novel claim, semantic preservation or evidence-consuming acceptance: MEM-OBJ-concept-body, EPI-RTE-semantic-disposition, report-verification limits.
There are several different write controls, not one universal approval gate:
- Library Writer validates a materialized temporary bundle before locked promotion, with attempted backup restoration on failed overwrite. This protects that publication route, not arbitrary host file edits: RT-RTE-publish-bundle.
- The base plugin supplies filtered post-edit findings and exits zero. It can be disabled and cannot undo the edit: RT-RTE-curation-feedback.
- Installed Pro hooks request asks or blocks for recognized host events; Stop, staged-tree pre-commit and CI are separate later doors. Enforcement depends on setup and the host honoring the protocol. Shell recognition is heuristic, external write tools are not enclosed by the named matchers, and an empty object event can supply no target. Whether real edit events can reach that last branch is unresolved, not a demonstrated bypass: RT-RTE-pro-gates.
- Pro's mechanical writers construct constrained board/journal/closure changes and check declared line-multiset deltas. Conservation does not establish prose truth, line placement, concurrent-update safety or multi-file transactionality. Interrupted writes can require repair: RT-RTE-pro-writes.
What checks license, and what later sessions receive
OKF implements useful deductions within parsed structural domains: field membership, links, declared expiry, trust classification and snapshot agreement. These are not semantic source checks. Keyed attribution joins labels; it does not show that a cited source supports a sentence. A human-prefixed verification mapping yields a human-reviewed tier without binding that declaration to current body bytes or freshness. The tier matters operationally: optional search filters select membership, and Pro pairing treats human review as discharging a reference-reading obligation. Full concept reads have no universal human-review prerequisite. The intended owner decision is actual review, but neither a permission ask nor recorded metadata authenticates that review: EPI-RTE-structural-check, EPI-RTE-trust-tier, EPI-RTE-read-pairing, EPI-RTE-semantic-review.
Pro's eligible Write feedback searches a few filename terms and presents bounded collision snippets, including deprecated hits. It asks the agent to compare claims, correct/deprecate a loser or retain a linked unresolved board obligation. It neither detects semantic contradiction nor automatically accepts a winner; Edit does not take this Write-only branch. The prescribed semantic disposition has meaningful intended consequences for later reliance, but no operated evidence-consuming acceptance and subsequent integration chain is supplied: MEM-RTE-reconcile-search, EPI-RTE-semantic-disposition.
Later context also arrives without an explicit request for those retained parts. SessionStart selects the latest dated Snapshot alongside cheap live state and selected friction cues. The historical snapshot is labeled not live; its command computes a line that the caller must append. Friction automatically records raw instrumented intent, not necessarily completed actions, and later supplies a count/report rather than a durable revised method. This supports cross-session state feedback, not demonstrated lesson extraction or improved prioritization: MEM-RTE-snapshot-context, MEM-RTE-friction-feedback, MEM-BAP-state-context. The profile verification preserves the distinction between these supported paths and unresolved trace-learning/source classifications; neither a positive learning value nor a system-wide negative follows.
Attested Computation is a separate offered protocol. A consumer binds permitted parameters to sanctioned code, an independent deterministic attester compares the executed artifact and displayed value against authoritative receipt data, and a display gate is supposed to refuse failure. The inspected CLI/MCP surfaces expose contracts and inventories, not this complete execution/check/display chain. Even a conforming attestation would license artifact/value fidelity for that call, not the sanctioned definition's real-world correctness. Arbitrary payload meaning and concrete executor/attester behavior remain unresolved: MEM-OBJ-references, EPI-OBJ-computation, EPI-RTE-run-attestation, EPI-RTE-display-gate.
Adaptation and scenario-relative assessment
The strongest supported adaptation contribution is an inspectable way to propose and retain changes to retrieval layout or instruction organization, supported by explicit principles and requested before/after checks. Refine reserves execution approval to the user; the separate instruction-restructuring skill does not impose that same approval rule. This is a policy-level pathway, not an observed improvement comparison. Governance installation can change instruction/hook files, and extension discovery can load capability code in-process; neither installation nor a naming check establishes trace-responsive improvement or sandboxing: RT-RTE-refine-proposal, RT-RTE-governance-install, RT-RTE-extension-load.
For the structural principles assessed in the runtime record, theory-builder condition 1 is wired: localized explanatory units exist. Condition 2 is claimed: the playbook requests decisions governed by their content, without observed consumption. Conditions 3 and 4 are uninspected: diagnosis under unchanged principles is not evidenced content-directed criticism of those principles, and no retained criticism-to-next-round chain is shown. Criticism of a consumed theory improving future capacity is therefore unestablished, independently of the useful revision pathway. Reflection is unestablished because both causal directions through a bounded self-representation are not shown. Autonomous theory building is unestablished; actual role execution and host internals are not frozen, and refine retains human approval. Exercised self-improvement is also unestablished: no evidence-shaped organizational update is traced into subsequent operation causally depending on it. These are separate findings, not a combined negative judgment about OKF or its external agent: RT-RTE-refine-proposal.
For a task needing navigable, editable project explanations, the evidence supports durable retention, selective later access and differentiated structural/work-state controls. For a task requiring authenticated source review, automatic semantic reconciliation or demonstrated cross-session improvement, those same routes are insufficient evidence. Paired source/concept and review records could establish semantic preservation or accepted claims; configured-host traces could establish actual gate reachability and delivery; retained-update/later-consumer traces plus an appropriate comparison could establish improvement at the level varied. Concrete executor/attester receipts would change the computation assessment, and backend implementation evidence would resolve the opaque alternative. None of these missing observations negates the implemented paths already described.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No supplied operation or outcome evidence | SRC-1; RT-RTE-skill-work, WF-RTE-maintenance-closeout, MEM-BAP-retained-knowledge | Frozen source implementation and instructions; no host session or causal experiment | Actual write-back, successful closeout, semantic correctness, reliable host compliance or measured retrieval/task benefit | Inspectable source-linked session and later-task traces; outcome comparisons appropriate to the claimed effect |
| Migration coverage across modes is incomplete; carried report-verification limit | SRC-1; MEM-OBJ-concept-body, memory annotation, epistemic coverage, report-verification limits | Adoption account plus verifier's identified v0.1-to-v0.2 citation reshaping in the same playbook | Every migration preserves body bytes, or every provenance-reshaping route is enumerated | Mode-by-mode transformation records including citation removal and later consumption |
| Host gates are conditional and channel-bounded | SRC-1; RT-RTE-pro-gates, RT-RTE-curation-feedback, RT-RTE-pro-writes | Named host events, heuristic shell patterns, distinct writer/commit/CI paths | Universal approved-write exclusion, authenticated attestation history, or end-to-end rollback; real-host reachability of incomplete events | Actual host event/enforcement contract and traces across relevant channels; separately evidenced transaction/recovery guarantees |
| Fresh bytes, bounded projections and containment have narrower scope than universal freshness or window control | SRC-1; RT-RTE-mcp-retrieve, RT-CLM-live-bodies | Fingerprinted MCP selection, live guarded reads, log cuts and browser snapshots | Universally fresh selection, complete delivery of retained reasons, whole-response/window bounds or universal filesystem containment | Stronger invalidation and containment contracts; measured complete delivery/window traces covering alternate reads |
| Semantic warrant and review authenticity remain unestablished | SRC-1; EPI-RTE-semantic-review, EPI-RTE-semantic-disposition, EPI-RTE-trust-tier, EPI-RTE-read-pairing | Requested source review and implemented metadata/check consumers | Source-supported distillation, current authenticated human review, or accepted ampliative knowledge followed by integration | Paired candidates/sources, evidence-consuming review/disposition records and body-bound provenance with later consumers |
| Optional backend, arbitrary payloads and initial caller control are unresolved | SRC-1; RT-RTE-mcp-retrieve, MEM-OBJ-references, MEM-OBJ-board, MEM-RTE-snapshot-context | Fallback/backend seam, payload inventory/contracts, known subsequent transforms | Complete storage/form/lineage/admission/consumer inventory; fallback findings generalized to opaque alternatives | Backend and per-payload implementation records; explicit initial-content and Snapshot admission decisions |
| Attestation is not established as an operated end-to-end route | SRC-1; EPI-RTE-run-attestation, EPI-RTE-display-gate, report verification | SPEC contract and inspected CLI/MCP dispatch; minimal fixture is not full protocol wiring | Implemented or observed sanctioned-artifact comparison and display refusal; repository-wide absence of all attester functions | Concrete executor, independent binding/authoritative reread, attester and display-consumer implementation plus receipts |
| Trace learning, future capacity and stronger self-change properties are unresolved independently | SRC-1; MEM-RTE-friction-feedback, MEM-RTE-snapshot-context, RT-RTE-refine-proposal, profile verification | Raw telemetry, caller-retained counters and requested refinement; external model internals uninspected | Positive or system-wide negative trace-learning/source classifications; demonstrated learning, reflection, autonomous theory building or exercised self-improvement | Qualifying automatic trace-to-durable-output records; criticism/revision/iteration and later-consumer evidence; property-specific causal links and outcome comparisons |
No source rereads were needed for this synthesis. No unresolved inter-report conflict was declared; profile verification declared no additional limits. Available but uninspected examples, UI details and evaluator leaves do not support system-complete negative conclusions.