omnius 1.0.745 → 1.0.746

package/dist/index.js

@@ -1,4 +1,4 @@
 [diffend] Oversized file quarantined before diffing.
 name: package/dist/index.js
-size: 38450286 bytes
-sha256: e8298c4ffeab6235a49fe30663830527f1c938fbc9a745cd6f950ce81a750d73
+size: 38482354 bytes
+sha256: 01d1ef9f810ffdf50441681bb5f1426370e29b3d11481907c98ba18dc9faad18

package/docs/DISCOVERY.json

@@ -32725,6 +32725,40 @@
         }
       ]
     },
+    {
+      "id": "guide.work-orders-recurrent-self-improvement-readme-uppercase",
+      "kind": "guide",
+      "title": "Recurrent self-improvement work orders",
+      "summary": "Updated: 2026-09-25. Research basis and quality assessments live in .aiwg/research/self-improvement/.",
+      "keywords": [
+        "work",
+        "orders",
+        "recurrent",
+        "self",
+        "improvement",
+        "README",
+        "md"
+      ],
+      "maturity": "internal",
+      "audiences": [
+        "maintainer",
+        "large-context-agent"
+      ],
+      "layer": "documentation",
+      "interfaces": [
+        {
+          "type": "file",
+          "target": "docs/work-orders/recurrent-self-improvement/README.md"
+        }
+      ],
+      "references": [
+        {
+          "type": "documentation",
+          "target": "docs/work-orders/recurrent-self-improvement/README.md",
+          "relation": "canonical-artifact"
+        }
+      ]
+    },
     {
       "id": "guide.work-orders-runtime-health-remediation-audit-uppercase",
       "kind": "guide",

package/docs/DISCOVERY.md

@@ -529,6 +529,7 @@ Daemon equivalents are `GET /v1/discovery/bootstrap`, `GET /v1/discovery?q=<inte
 | `guide.work-orders-hermes-architecture-deltas-index-uppercase` | Hermes Architecture Deltas Workorder Index | Status: ready for implementation planning. These documents are workorders only; they do not claim the features are implemented. |
 | `guide.work-orders-long-horizon-feature-delivery-wo-01-native-research-feature-workflow-uppercase` | WO-01 — Native research-to-feature workflow | Status: complete — repository implementation, twelve-scenario acceptance and scoped Git delivery. Authorized by the user request to bring the AIWG research-team mechanism into Omnius. Publication and live acceptance remain with the user. |
 | `guide.work-orders-omnius-context-engineering-behavior-fixes` | Omnius Context Engineering Behavior Fixes | Source observation: current Omnius behavior while supervising a noclip/earth sidebar cleanup task. The observed run mixed internal cognitive-agent artifacts with user-task artifacts, repeated stale edit attempts after the file had changed, and overclaimed completion despite stale or missing verification. |
+| `guide.work-orders-recurrent-self-improvement-readme-uppercase` | Recurrent self-improvement work orders | Updated: 2026-09-25. Research basis and quality assessments live in .aiwg/research/self-improvement/. |
 | `guide.work-orders-runtime-health-remediation-audit-uppercase` | Runtime Health Remediation Audit Log | Session: RHR-2026-09-02 Host user: roko Program root: /home/roko/Documents/Projects/Adjacent/omnius/omnius |
 | `guide.work-orders-runtime-health-remediation-cli-current-main-suite-evidence-uppercase` | Current-main CLI suite reconciliation evidence | This tranche reconciles the complete CLI suite except the protected generated discovery outputs and packages/cli/tests/update-service.test.ts, as directed. It does not edit publish/, the shared runtime-health tracker or audit, or the protected discovery files. |
 | `guide.work-orders-runtime-health-remediation-extended-work-orders-uppercase` | Extended Omnius Work Orders | This register preserves the remaining operator requests that cross the core runtime-health orders. Each item is an implementation obligation. A checked item requires code, deterministic verification, and a named evidence entry in AUDIT.md. Prompt-specific keyword rules, output stripping, and case-specific reply heuristics do not satisfy an item. |

package/docs/work-orders/recurrent-self-improvement/README.md

@@ -0,0 +1,110 @@
+# Recurrent self-improvement work orders
+
+Updated: 2026-09-25. Research basis and quality assessments live in `.aiwg/research/self-improvement/`.
+
+## Non-negotiable contract
+
+- Improvement is evidence-driven inference, not a deterministic claim gate.
+- Tests, hashes, receipts, and metrics are evidence. They do not decide semantic quality by themselves.
+- Full source traces remain immutable and addressable; summaries and scalar rewards are never the only copy.
+- Search and held-out evidence are separated. The proposer never receives private selection results.
+- The system maintains branching candidate lineage. It does not silently overwrite the active implementation.
+- Generated data is labeled synthetic and linked to exact sources and an inference review.
+- Human/original data remains in every recurrent training generation.
+- Promotion is versioned, reversible, explicitly adjudicated, and auditable.
+
+## WO-SI-01 — Immutable observation spine — complete
+
+Completed implementation:
+
+- Hash-chained JSONL event archive under `.omnius/self-improvement/events.jsonl`.
+- Immutable, content-addressed trajectory and full-transcript evidence per completed run.
+- Evidence references retain identifiers, URIs, SHA-256, media types, and optional byte ranges.
+- All outcomes are recorded. A deterministic low-signal heuristic may emit an advisory but cannot exclude an observation.
+
+Acceptance evidence: `packages/orchestrator/tests/self-improvement.test.ts` and `packages/orchestrator/tests/agenticRunner.test.ts`.
+
+## WO-SI-02 — Inference review and candidate lineage — complete at framework level
+
+Completed implementation:
+
+- Typed inference callbacks for REVIEW, PROPOSE, and ADJUDICATE.
+- Candidate surfaces cover memory, prompt, harness, tool, training data, adapters, and weights.
+- Candidates retain parent IDs, review IDs, proposer identity, hypotheses, and artifact provenance.
+- Decisions cite only evaluation events available for that candidate and require a rollback plan.
+
+The TUI now connects review/proposal to the configured parent-model backend every five completed tasks as low-priority work. Remaining production work: expose lineage in `/memory` or a dedicated improvement view.
+
+## WO-SI-03 — Honest training export — complete
+
+Completed implementation:
+
+- Replaced fabricated rejection-fine-tuning examples.
+- Export now requires real assistant transcript content.
+- Every export requires an inference review whose disposition is `include` and whose evidence overlaps the example's sources.
+- Export manifests record review IDs, source hashes, and synthetic/original counts.
+
+Remaining production work: add privacy/license review status to review output before external dataset publication.
+
+## WO-SI-04 — Fine-tuning suite bridge — complete
+
+Completed implementation:
+
+- Omnius client covers models, datasets, actions, jobs, cancellation, evaluations, comparisons, Ollama inspection, and omni/audio/adapter contracts and execution.
+- Suite intake validates `omnius.self-improvement.bundle.v1`.
+- It creates immutable provenance-grouped train/validation/test files and a frozen hash manifest.
+- Dataset selection automatically binds the suite's `DISTILL_TRAIN_FILE` and `DISTILL_VAL_FILE` inputs.
+- Held-out/test split names are not bound into the training job.
+
+Acceptance evidence: Omnius fine-tuning client tests and `/srv/fine_tuning_suite/tests/test_omnius_intake.py`.
+
+## WO-SI-05 — Broker-safe training — complete for REST jobs
+
+Completed implementation:
+
+- CUDA actions require a lease owner, meaningful justification, expected duration, exact GPU UUID scope, VRAM request, and readiness command.
+- The runner performs `docker gpu discover` before starting and wraps the child with `docker gpu run`.
+- `CUDA_VISIBLE_DEVICES` and `DISTILL_GPUS` are both set to the exact reserved UUID scope.
+- Lease details, excluding tokens, are retained in job metadata.
+
+Direct baseline, training, evaluation, and export commands also refuse to run outside a broker lease or with static GPU ordinals.
+
+## WO-SI-06 — Private evaluation and Pareto archive — partial
+
+Completed implementation:
+
+- Search, held-out, transfer, and adversarial evaluation types.
+- Held-out/transfer/adversarial events are rejected unless marked private to the proposer.
+- Multiple metrics and directionality are retained without collapsing them to one reward.
+
+Remaining production work:
+
+1. Add a Pareto-front service over capability, regression, latency, inference tokens, VRAM, and reward-hacking metrics.
+2. Run at least three seeds for stochastic candidates.
+3. Add private cross-model transfer suites and a frozen canary set.
+4. Make evaluation fixtures unreadable from candidate sandboxes.
+
+## WO-SI-07 — Prompt and harness evolution — pending experiment
+
+1. Generate prompt candidates from exact traces and inferred diagnoses.
+2. Generate harness/code candidates in isolated worktrees.
+3. Validate interfaces before expensive evaluation.
+4. Compare greedy, beam, and Pareto selection.
+5. Explicitly probe for benchmark-specific shortcuts and trace contamination.
+
+Exit criterion: statistically credible held-out and transfer improvement with no material regression in safety, provenance, latency, or tool correctness.
+
+## WO-SI-08 — Adapter and preference experiments — pending execution
+
+1. Establish frozen base-model and current-production baselines.
+2. Run SFT/LoRA and QLoRA candidates over inference-reviewed real traces plus original-data anchors.
+3. Construct DPO pairs only when independent inference review explains why one real response is preferred.
+4. Compare environment-feedback curricula against static transcript SFT.
+5. Measure catastrophic forgetting, tail performance, calibration, tool integrity, and cross-model transfer.
+6. Never merge or publish weights automatically.
+
+Exit criterion: candidate beats the baseline on private held-out and transfer suites, passes model-collapse/tail checks, and retains a working rollback artifact.
+
+## WO-SI-09 — Explicit promotion transaction — pending
+
+Implement a separate operator-visible promotion command. It must verify the cited decision and artifacts, capture the previous version, apply one reversible target change, run smoke tests, record a promotion event, and support rollback. A running inference must never call this implicitly.

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.745",
+  "version": "1.0.746",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.745",
+      "version": "1.0.746",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.745",
+  "version": "1.0.746",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/library.js",