omnius 1.0.752 → 1.0.753

package/dist/index.js

@@ -1,4 +1,4 @@
 [diffend] Oversized file quarantined before diffing.
 name: package/dist/index.js
-size: 38514804 bytes
-sha256: 3ab14289e97823326a1c2463d6329a9ff83ccecf63536ad50d744c318704699f
+size: 38630756 bytes
+sha256: 2185bc5ae42eeb54157503bc2d6961c4d1b27879bd4dd8677bcf96230d239433

package/docs/DISCOVERY.json

@@ -26111,6 +26111,108 @@
         "Allowing an autonomous agent to invoke it without an explicit user action"
       ]
     },
+    {
+      "id": "command.selfimprove",
+      "kind": "command",
+      "title": "/selfimprove",
+      "summary": "Show durable review/proposal backlog, leases, retries, and last error",
+      "aliases": [
+        "/selfimprove",
+        "selfimprove",
+        "/self-improve",
+        "self-improve"
+      ],
+      "keywords": [
+        "command",
+        "runtime",
+        "selfimprove",
+        "status",
+        "run",
+        "review",
+        "autonomous",
+        "pause",
+        "rollback"
+      ],
+      "maturity": "stable",
+      "layer": "interface",
+      "audiences": [
+        "user",
+        "coding-agent",
+        "integrator"
+      ],
+      "interfaces": [
+        {
+          "type": "tui",
+          "target": "/selfimprove"
+        },
+        {
+          "type": "rest",
+          "target": "/v1/commands/selfimprove"
+        },
+        {
+          "type": "gateway-command",
+          "target": "/selfimprove"
+        }
+      ],
+      "references": [
+        {
+          "type": "source",
+          "target": "packages/cli/src/tui/command-registry.ts",
+          "relation": "registry"
+        }
+      ],
+      "signatures": [
+        {
+          "signature": "/selfimprove status",
+          "description": "Show durable review/proposal backlog, leases, retries, and last error"
+        },
+        {
+          "signature": "/selfimprove run",
+          "description": "Immediately drain eligible self-improvement work with the parent model"
+        },
+        {
+          "signature": "/selfimprove review",
+          "description": "Run evidence review and candidate generation without active promotion"
+        },
+        {
+          "signature": "/selfimprove autonomous",
+          "description": "Opt in to downstream candidate experiment and promotion orchestration"
+        },
+        {
+          "signature": "/selfimprove pause",
+          "description": "Pause processing while retaining the durable backlog"
+        },
+        {
+          "signature": "/selfimprove rollback <promotion-id> [reason]",
+          "description": "Restore the exact pre-promotion baseline and verify the rollback"
+        }
+      ],
+      "subcommands": [
+        "status",
+        "run",
+        "review",
+        "autonomous",
+        "pause",
+        "rollback"
+      ],
+      "implementation_status": "implemented",
+      "surfaces": {
+        "tui": true,
+        "rest": true,
+        "gateway": true,
+        "agentTool": false
+      },
+      "safety": {
+        "userOnly": true,
+        "destructive": false,
+        "networked": false,
+        "secretBearing": false,
+        "profileGated": false
+      },
+      "avoid_when": [
+        "Allowing an autonomous agent to invoke it without an explicit user action"
+      ]
+    },
     {
       "id": "command.selfmodify",
       "kind": "command",

package/docs/DISCOVERY.md

@@ -364,6 +364,7 @@ Daemon equivalents are `GET /v1/discovery/bootstrap`, `GET /v1/discovery?q=<inte
 | `command.scheduler` | /scheduler | Scheduled tasks control panel (list/kill/toggle) |
 | `command.score` | /score | Show inference capability scorecard (memory, compute, speed, models) |
 | `command.secrets` | /secrets | List registered secrets |
+| `command.selfimprove` | /selfimprove | Show durable review/proposal backlog, leases, retries, and last error |
 | `command.selfmodify` | /selfmodify | Toggle agent self-modify mode for natural-language runtime commands |
 | `command.sessions` | /sessions | Browse saved sessions |
 | `command.sethome` | /sethome | Set the project home for gateway/session use |

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

@@ -24,7 +24,7 @@ Completed implementation:
 
 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
+## WO-SI-02 — Inference review and candidate lineage — complete end to end
 
 Completed implementation:
 
@@ -33,7 +33,23 @@ Completed implementation:
 - 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.
+The TUI now connects review/proposal to the currently selected parent-model
+backend through a durable workspace controller. Every recorded outcome enters an
+oldest-first backlog; processing uses low-priority/yielding inference and no
+substitute model. Queue stage, attempts, retry horizon, last error, parent model,
+cross-process lease, downstream stage, and downstream receipt IDs survive
+restart in `.omnius/self-improvement/controller-state.json`. Stale leases are
+recovered, proposal failures resume after their completed review, and long
+downstream work renews its lease. `/selfimprove status|run|review|autonomous|pause`
+provides operator-visible state and control. Review-only is the default;
+autonomous downstream execution is connected to the materialization,
+evaluation, adjudication, training, and promotion runtime. Review-only remains
+the default. `/selfimprove autonomous` is the explicit opt-in that authorizes
+scoped production mutation; an unavailable dependency remains visibly
+retryable rather than being falsely reported as complete.
+
+Acceptance evidence: `packages/cli/tests/self-improvement-controller.test.ts`
+and the CLI command-registry/coherence regression suites.
 
 ## WO-SI-03 — Honest training export — complete
 
@@ -69,42 +85,87 @@ Completed implementation:
 
 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
+## WO-SI-06 — Private evaluation and Pareto archive — complete
 
 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.
+- Materialized candidates receive three independent private semantic reviews
+  (held-out, transfer, and adversarial) from the selected parent model.
+- Adapter candidates run isolated baseline, train, and private evaluation jobs
+  over at least three seeds, including tail and reward-hacking evidence.
+- Compatible historical candidates are compared through multi-objective Pareto
+  nondominance. Pareto output carries `authorizesPromotion: false` and is only
+  evidence for a separate inference adjudicator.
+- Candidate worktrees never receive private evaluation fixtures or prior
+  private judgments.
+
+## WO-SI-07 — Prompt, harness, tool, and code evolution — complete
+
+Implemented:
+
+- The selected parent model first locates a minimal path set from immutable
+  evidence, then compiles complete typed file artifacts against exact current
+  file contents and preimage hashes.
+- Prompt, harness, tool, and code candidates materialize in detached worktrees
+  (with an isolated candidate-directory fallback), never directly in the live
+  tree.
+- Direct-argv verification records bounded, hashed stdout/stderr receipts.
+- Private multi-perspective evaluation and Pareto evidence feed an independent
+  inference adjudicator; no scalar threshold or deterministic claim gate
+  decides semantic quality.
+- Only an inference `promote` verdict can enter the scoped transaction path.
+
+## WO-SI-08 — Adapter and preference experiments — complete orchestration
+
+Implemented:
+
+- Real transcript examples require an inference `include` review and immutable
+  provenance overlap before bundle export.
+- Fine-tuning suite intake freezes nonempty provenance-isolated
+  train/validation/test splits.
+- Each candidate executes baseline → train → private adapter evaluation for at
+  least three seeds through exact broker GPU UUID/VRAM/justification/horizon
+  leases.
+- Dataset IDs and job IDs are durably checkpointed; restart resumes polling
+  instead of submitting duplicate jobs.
+- Tail, reward-hacking, capability, and additional inference-reviewed metrics
+  remain separate objectives. SHA-256-addressed adapters on the seed frontier
+  are marked promotion-ready only after inference adjudication.
+- Training and evaluation are automatic when configured, while external model
+  publication and replacement of a selected model tag remain separate release
+  operations rather than an implicit side effect.
+
+## WO-SI-09 — Promotion, autonomous rewrite, and rollback — complete
+
+Implemented:
+
+- Review-only is the safe default. `/selfimprove autonomous` is the explicit,
+  durable opt-in for scoped autonomous source promotion.
+- Promotion requires an archived inference `promote` decision, an intact
+  materialization, explicit allowed paths, an enabled policy, and verifier
+  commands. Mechanical checks protect integrity but do not substitute for the
+  semantic adjudicator.
+- The transaction captures exact baselines, applies only candidate-owned paths,
+  verifies the live target, records a hash-addressed promotion receipt, and
+  automatically restores the baseline after apply, verification, or receipt
+  failure.
+- `/selfimprove rollback <promotion-id> [reason]` restores the exact recorded
+  baseline, refuses to overwrite later human edits, reruns verification, and
+  appends a rollback receipt.
+
+## Runtime configuration
+
+- `OMNIUS_SELF_IMPROVE_ALLOWED_PATHS`: comma-separated autonomous source scope.
+- `OMNIUS_SELF_IMPROVE_GPU_UUIDS`: exact broker UUIDs for adapter experiments;
+  the current `/gpu` preference is used when this is unset.
+- `OMNIUS_SELF_IMPROVE_VRAM_MIB`: requested VRAM per experiment (default 24576).
+- `OMNIUS_SELF_IMPROVE_DURATION_SECONDS`: lease horizon (default 14400).
+- `OMNIUS_SELF_IMPROVE_JUSTIFICATION`: visible broker justification.
+- `OMNIUS_SELF_IMPROVE_MODEL_ID`: fine-tuning suite base-model record.
+- `OMNIUS_FINE_TUNING_SUITE_URL`: suite endpoint.
+
+The complete path is restart-resumable and inherits the selected parent model;
+it contains no Nemotron or alternate-model fallback.

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.752",
+  "version": "1.0.753",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.752",
+      "version": "1.0.753",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.752",
+  "version": "1.0.753",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/library.js",