sneakoscope 2.0.4 → 2.0.5

package/README.md

@@ -16,7 +16,7 @@ Set up this agent project with Sneakoscope Codex. Use [[mandarange/Sneakoscope-C
 
 ## Current Release
 
-SKS **2.0.4** is a P0 Codex App Fast UI and MAD Zellij worker-pane closure patch on top of the 2.0 execution layer. `sks --mad` now relies on launch-time Fast/high overrides instead of user-level Codex config rewrites, safe Fast UI repair runs through `sks doctor --fix`, provider badges read env/auth/config.toml consistently, and interactive MAD worker panes attach to the real Zellij session as scheduler slots spawn.
+SKS **2.0.5** is a P0 Codex App Fast UI and MAD Zellij worker-pane closure patch on top of the 2.0 execution layer. `sks --mad` now relies on launch-time Fast/high overrides instead of user-level Codex config rewrites, safe Fast UI repair runs through `sks doctor --fix`, provider badges read env/auth/config.toml consistently, and interactive MAD worker panes attach to the real Zellij session as scheduler slots spawn.
 
 What changed:
 
@@ -398,9 +398,9 @@ Manual fan-out syntax:
 
 Effort is assigned per agent. Simple read-only/docs slices can run low, ordinary tooling and lease mapping use medium, safety/DB/schema/release lanes use high, and frontier/forensic research can escalate to xhigh. If a lease conflict, schema failure, proof blocker, DB risk, or release risk appears, the parent can escalate that lane while keeping unrelated lanes cheaper and faster.
 
-### Naruto Shadow Clone Swarm (`$Naruto`)
+### Naruto Massive Parallel Work Swarm (`$Naruto`)
 
-`$Naruto` (影分身 / Kage Bunshin no Jutsu) is a high-scale mode of the native agent kernel for broad fan-out work — codebase-wide sweeps, parallel drafting, large audits. It lifts the standard 20-agent ceiling to **up to 100 clone sessions** (only for this route; every other route keeps the 20 cap).
+`$Naruto` (影分身 / Kage Bunshin no Jutsu) is the hardware-safe massive parallel work mode of the native agent kernel. It is not limited to validation. A Naruto run builds a mixed work graph, keeps a safe active worker pool full, and assigns clones to implementation, modification, test generation, verification, research, documentation, conflict resolution, rollback planning, integration support, and GPT final review input work. It lifts the standard 20-agent ceiling to **up to 100 total clone generations** for this route while keeping active workers under the live hardware, lease, memory, terminal UI, file descriptor, local LLM, and remote API caps.
 
 ```sh
 sks naruto run "sweep the codebase for TODO comments and summarize"
@@ -411,9 +411,11 @@ sks naruto status
 
 Aliases: `$ShadowClone`, `$Kagebunshin`, and the CLI flag `sks --naruto`.
 
-- **System-aware concurrency:** `--clones N` is the total work fan-out, but `$Naruto` never spawns the whole count at once. Live concurrency is throttled to a host-safe number derived from CPU cores and free memory (heavier cap for real `codex-sdk` workers, tighter packing for in-process `fake`). So `--clones 100` on a small host still processes all 100 work units while only running a safe handful at a time; the run reports when it throttles. Override with `SKS_NARUTO_MAX_CONCURRENCY=<n>`.
+- **Hardware-safe governor:** `--clones N` is the total work fan-out, but `$Naruto` never spawns the whole count at once. Live concurrency is throttled by current load, memory, file descriptors, Zellij pane budget, local LLM request budget, remote API budget, disk pressure, pending queue, and active lease conflicts.
+- **Dynamic active pool:** completed workers are drained and replaced while runnable work remains, so the active pool does not sit empty between generations.
 - **Dynamic per-clone effort (like Team):** truly simple / no-tool work runs at `low`, any tool use lifts a clone to `medium` (never high/xhigh), and every clone runs in fast service tier.
-- **Safe parallel writes:** clones coordinate through the same lease-based patch-swarm (merge coordinator + conflict rebase + transaction journal) as Team.
+- **Safe parallel writes:** write-capable clones produce patch envelopes for leased files. Non-overlapping envelopes can apply in parallel; overlapping envelopes serialize or route to conflict resolution. Local worker output remains a draft until the GPT final arbiter approves or modifies it.
+- **Massive UI without pane overload:** Zellij shows visible active worker panes up to the UI cap and tracks the remaining active headless workers in the Naruto dashboard.
 
 See [docs/naruto.md](docs/naruto.md) for the full reference.
 
@@ -513,18 +515,20 @@ $DB inspect this migration for destructive risk
 
 ### Optional Local LLM Workers
 
-Local Ollama workers are off by default, so SKS stays GPT-only unless you explicitly enable them. Use the Codex App prompt commands:
+Local model workers are off by default, so SKS stays GPT-only unless you explicitly enable them. Use the Codex App prompt commands:
 
 ```text
 $with-local-llm-on
 $with-local-llm-off
 ```
 
-When enabled, the local model can only help with policy-eligible simple code patch-envelope work or read-only collection. GPT/Codex still owns strategy, planning, design, review, verification, safety, and integration. Check or tune the machine-local setting from the terminal:
+When enabled, SKS auto-detects an installed local model from a running MLX LM server, OpenAI-compatible local server, or Ollama. If no local model is available, activation stays blocked and reports that no local model was found. The local model can only help with policy-eligible simple code patch-envelope work or read-only collection. GPT/Codex still owns strategy, planning, design, review, verification, safety, and integration. Check or tune the machine-local setting from the terminal:
 
 ```sh
 sks with-local-llm status --json
-sks with-local-llm on --model rafw007/qwen36-a3b-claude-coder:q4_K_M
+sks with-local-llm on
+sks with-local-llm on --provider mlx-lm --model mlx-community/Qwen3.6-35B-A3B-4bit --base-url http://127.0.0.1:8080
+sks with-local-llm on --provider ollama --model rafw007/qwen36-a3b-claude-coder:q4_K_M
 sks with-local-llm off
 ```
 

package/crates/sks-core/Cargo.lock

@@ -76,7 +76,7 @@ dependencies = [
 
 [[package]]
 name = "sks-core"
-version = "2.0.4"
+version = "2.0.5"
 dependencies = [
  "serde_json",
 ]

package/crates/sks-core/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "sks-core"
-version = "2.0.4"
+version = "2.0.5"
 edition = "2021"
 
 [dependencies]

package/crates/sks-core/src/main.rs

@@ -4,7 +4,7 @@ use std::io::{self, Read, Seek, SeekFrom};
 fn main() {
     let mut args = std::env::args().skip(1);
     match args.next().as_deref() {
-        Some("--version") => println!("sks-rs 2.0.4"),
+        Some("--version") => println!("sks-rs 2.0.5"),
         Some("compact-info") => {
             let mut input = String::new();
             let _ = io::stdin().read_to_string(&mut input);

package/dist/.sks-build-stamp.json

@@ -1,8 +1,8 @@
 {
   "schema": "sks.dist-build-stamp.v1",
   "package_name": "sneakoscope",
-  "package_version": "2.0.4",
-  "source_digest": "bd67ddf989e2e0b702453d422e2cf991e8ca397d3d4afc14fdfa06b4610452c9",
-  "source_file_count": 1879,
-  "built_at_source_time": 1780569940553
+  "package_version": "2.0.5",
+  "source_digest": "696706cb9d175c4200c967130fbc5e680f4b6c905ec6e7fad5b1671827bcb4cb",
+  "source_file_count": 1948,
+  "built_at_source_time": 1780613485190
 }

package/dist/bin/sks.js

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-const FAST_PACKAGE_VERSION = '2.0.4';
+const FAST_PACKAGE_VERSION = '2.0.5';
 const args = process.argv.slice(2);
 try {
     if (args[0] === '--agent' && args[1] === 'worker') {

package/dist/build-manifest.json

@@ -1,16 +1,16 @@
 {
   "schema": "sks.dist-build.v2",
-  "version": "2.0.4",
-  "package_version": "2.0.4",
+  "version": "2.0.5",
+  "package_version": "2.0.5",
   "typescript": true,
   "mjs_runtime_files": 0,
-  "compiled_file_count": 949,
-  "compiled_js_count": 949,
+  "compiled_file_count": 1014,
+  "compiled_js_count": 1014,
   "compiled_dts_count": 0,
-  "source_digest": "bd67ddf989e2e0b702453d422e2cf991e8ca397d3d4afc14fdfa06b4610452c9",
-  "source_file_count": 1879,
-  "source_files_hash": "c43d09b591370ad58937365d6bbfca2be0b60fcf04a43fdbb9a56d9288752814",
-  "source_list_hash": "c43d09b591370ad58937365d6bbfca2be0b60fcf04a43fdbb9a56d9288752814",
+  "source_digest": "696706cb9d175c4200c967130fbc5e680f4b6c905ec6e7fad5b1671827bcb4cb",
+  "source_file_count": 1948,
+  "source_files_hash": "50906865545a9076d7b2736ee379a428160a8869749b85ee79c4ddab8eab4b12",
+  "source_list_hash": "50906865545a9076d7b2736ee379a428160a8869749b85ee79c4ddab8eab4b12",
   "src_mjs_runtime_files": 0,
   "dist_stamp_schema": "sks.dist-build-stamp.v1",
   "files": [
@@ -216,6 +216,8 @@
     "core/codex-control/gpt-final-context-compressor.js",
     "core/codex-control/gpt-final-proof-pack.js",
     "core/codex-control/gpt-final-review-schema.js",
+    "core/codex-control/python-codex-sdk-adapter.js",
+    "core/codex-control/python-codex-sdk-event-translator.js",
     "core/codex-control/schemas/agent-worker-result.schema.js",
     "core/codex-exec-output-schema.js",
     "core/codex-hooks/codex-hook-actual-discovery.js",
@@ -237,6 +239,7 @@
     "core/codex/codex-0-134-compat.js",
     "core/codex/codex-0-135-compat.js",
     "core/codex/codex-0-136-compat.js",
+    "core/codex/codex-0-137-compat.js",
     "core/codex/codex-cli-syntax-builder.js",
     "core/codex/codex-config-eperm-repair.js",
     "core/codex/codex-config-readability.js",
@@ -362,7 +365,22 @@
     "core/json-schema-validator.js",
     "core/language-preference.js",
     "core/local-llm/local-collaboration-policy.js",
+    "core/local-llm/local-llm-backpressure.js",
+    "core/local-llm/local-llm-capability.js",
+    "core/local-llm/local-llm-client.js",
     "core/local-llm/local-llm-config.js",
+    "core/local-llm/local-llm-context-cache.js",
+    "core/local-llm/local-llm-control-adapter.js",
+    "core/local-llm/local-llm-json-repair.js",
+    "core/local-llm/local-llm-metrics.js",
+    "core/local-llm/local-llm-ollama-client.js",
+    "core/local-llm/local-llm-openai-compatible-client.js",
+    "core/local-llm/local-llm-prompt-cache.js",
+    "core/local-llm/local-llm-scheduler.js",
+    "core/local-llm/local-llm-schema-enforcer.js",
+    "core/local-llm/local-llm-smoke.js",
+    "core/local-llm/local-llm-warmup.js",
+    "core/local-llm/local-worker-eligibility.js",
     "core/loop-blocker.js",
     "core/mad-sks/audit-ledger.js",
     "core/mad-sks/authorization-manifest.js",
@@ -393,6 +411,22 @@
     "core/mission.js",
     "core/mistake-memory.js",
     "core/mistake-recall.js",
+    "core/naruto/hardware-capacity-probe.js",
+    "core/naruto/naruto-active-pool.js",
+    "core/naruto/naruto-backpressure.js",
+    "core/naruto/naruto-concurrency-governor.js",
+    "core/naruto/naruto-finalizer.js",
+    "core/naruto/naruto-generation-scheduler.js",
+    "core/naruto/naruto-gpt-final-pack.js",
+    "core/naruto/naruto-parallel-patch-apply.js",
+    "core/naruto/naruto-patch-transaction-batch.js",
+    "core/naruto/naruto-role-policy.js",
+    "core/naruto/naruto-verification-dag.js",
+    "core/naruto/naruto-verification-pool.js",
+    "core/naruto/naruto-work-graph.js",
+    "core/naruto/naruto-work-item.js",
+    "core/naruto/naruto-work-stealing.js",
+    "core/naruto/resource-pressure-monitor.js",
     "core/no-question-guard.js",
     "core/openclaw.js",
     "core/perf-bench.js",
@@ -406,6 +440,8 @@
     "core/pipeline/agent-stage-policy.js",
     "core/pipeline/final-gpt-patch-stage.js",
     "core/pipeline/final-gpt-review-stage.js",
+    "core/pipeline/finalize-pipeline-result.js",
+    "core/pipeline/gpt-final-required.js",
     "core/pipeline/pipeline-plan-writer.js",
     "core/pipeline/plan-schema.js",
     "core/pipeline/prompt-context-answer.js",
@@ -438,6 +474,7 @@
     "core/ppt.js",
     "core/preflight/parallel-preflight-engine.js",
     "core/prompt-context-builder.js",
+    "core/prompt/prompt-placeholder-guard.js",
     "core/proof-field.js",
     "core/proof/auto-finalize.js",
     "core/proof/claim-ledger.js",
@@ -568,6 +605,7 @@
     "core/zellij/zellij-lane-runtime.js",
     "core/zellij/zellij-launcher.js",
     "core/zellij/zellij-layout-builder.js",
+    "core/zellij/zellij-naruto-dashboard.js",
     "core/zellij/zellij-pane-proof.js",
     "core/zellij/zellij-screen-proof.js",
     "core/zellij/zellij-worker-pane-manager.js",
@@ -679,6 +717,7 @@
     "scripts/codex-0-134-runner-truth-check.js",
     "scripts/codex-0-135-compat-check.js",
     "scripts/codex-0-136-compat-check.js",
+    "scripts/codex-0-137-compat-check.js",
     "scripts/codex-app-fast-ui-preservation-check.js",
     "scripts/codex-app-provider-badge-check.js",
     "scripts/codex-app-ui-clobber-guard-check.js",
@@ -696,6 +735,7 @@
     "scripts/codex-control-structured-output-check.js",
     "scripts/codex-control-thread-registry-check.js",
     "scripts/codex-control-tool-call-sequence-repair-check.js",
+    "scripts/codex-environment-scoped-approvals-check.js",
     "scripts/codex-exec-output-schema-actual-syntax-check.js",
     "scripts/codex-fast-mode-profile-propagation-check.js",
     "scripts/codex-history-search-check.js",
@@ -710,6 +750,7 @@
     "scripts/codex-managed-proxy-env-check.js",
     "scripts/codex-output-schema-fixture-check.js",
     "scripts/codex-permission-profiles-check.js",
+    "scripts/codex-plugin-list-json-check.js",
     "scripts/codex-profile-primary-check.js",
     "scripts/codex-project-config-policy-merge-regression.js",
     "scripts/codex-project-config-policy-splitter-check.js",
@@ -731,6 +772,7 @@
     "scripts/codex-sdk-thread-registry-check.js",
     "scripts/codex-sdk-ux-ppt-review-pipeline-check.js",
     "scripts/codex-sdk-zellij-pane-binding-check.js",
+    "scripts/codex-thread-runtime-choice-check.js",
     "scripts/codex-web-adapter-check.js",
     "scripts/computer-use-live-evidence-check.js",
     "scripts/computer-use-live-optional-check.js",
@@ -802,9 +844,18 @@
     "scripts/lib/real-codex-parallel-gate.js",
     "scripts/lib/real-codex-parallel-proof-fixture.js",
     "scripts/lib/valid-png-fixture.js",
+    "scripts/local-collab-all-pipelines-final-gpt-check.js",
     "scripts/local-collab-gpt-final-availability-check.js",
     "scripts/local-collab-no-local-only-final-check.js",
     "scripts/local-collab-policy-check.js",
+    "scripts/local-llm-all-pipelines-check.js",
+    "scripts/local-llm-cache-performance-check.js",
+    "scripts/local-llm-capability-check.js",
+    "scripts/local-llm-smoke-check.js",
+    "scripts/local-llm-structured-output-check.js",
+    "scripts/local-llm-throughput-check.js",
+    "scripts/local-llm-tool-call-repair-check.js",
+    "scripts/local-llm-warmup-check.js",
     "scripts/loop-blocker-check.js",
     "scripts/mad-preflight-blocks-unreadable-config-check.js",
     "scripts/mad-sks-actual-executor-blackbox.js",
@@ -831,7 +882,16 @@
     "scripts/mcp-tool-naming-parity-check.js",
     "scripts/memory-summary-rebuild-check.js",
     "scripts/mutation-callsite-coverage-check.js",
+    "scripts/naruto-active-pool-check.js",
+    "scripts/naruto-concurrency-governor-check.js",
+    "scripts/naruto-gpt-final-pack-check.js",
+    "scripts/naruto-parallel-patch-apply-check.js",
+    "scripts/naruto-real-local-gpt-final-smoke.js",
+    "scripts/naruto-role-distribution-check.js",
     "scripts/naruto-shadow-clone-swarm-check.js",
+    "scripts/naruto-verification-pool-check.js",
+    "scripts/naruto-work-graph-check.js",
+    "scripts/naruto-zellij-massive-ui-check.js",
     "scripts/non-recursive-pipeline-check.js",
     "scripts/npm-publish-performance-check.js",
     "scripts/official-docs-compat-report.js",
@@ -856,8 +916,13 @@
     "scripts/prepublish-fast-check.js",
     "scripts/prepublish-release-check-or-fast.js",
     "scripts/priority-full-closure-check.js",
+    "scripts/prompt-placeholder-guard-check.js",
     "scripts/provider-badge-context-check.js",
     "scripts/provider-context-config-toml-check.js",
+    "scripts/python-codex-sdk-all-pipelines-check.js",
+    "scripts/python-codex-sdk-capability-check.js",
+    "scripts/python-codex-sdk-sandbox-policy-check.js",
+    "scripts/python-codex-sdk-stream-bridge-check.js",
     "scripts/python-tools-smoke-check.js",
     "scripts/qa-actual-route-backfill-check.js",
     "scripts/qa-backfill-route-blackbox.js",

package/dist/commands/doctor.js

@@ -17,6 +17,7 @@ import { inventoryCodexPermissionProfiles } from '../core/codex/codex-permission
 import { appendMigrationEvents, hashConfigText } from '../core/migration/migration-transaction-journal.js';
 import { repairCodexAppFastUi } from '../core/codex-app/codex-app-fast-ui-repair.js';
 import { resolveProviderContext } from '../core/provider/provider-context.js';
+import { readLocalModelConfig } from '../core/agents/ollama-worker-config.js';
 export async function run(_command, args = []) {
     let setupRepair = null;
     let migrationPreFix = null;
@@ -136,6 +137,7 @@ export async function run(_command, args = []) {
         }))
         : codexAppUiPlan;
     const zellij = await checkZellijCapability({ root, require: process.env.SKS_REQUIRE_ZELLIJ === '1' });
+    const localModel = await readLocalModelConfig().catch(() => null);
     const permissionProfiles = await inventoryCodexPermissionProfiles(root, { writeReport: true });
     const globalSksInstallCleanup = flag(args, '--fix') && !flag(args, '--local-only')
         ? await cleanDuplicateGlobalSksInstalls({ root, fix: true }).catch((err) => ({ schema: 'sks.global-sks-install-cleanup.v1', ok: false, fix: true, error: err?.message || String(err), blockers: ['global_sks_install_cleanup_exception'] }))
@@ -151,6 +153,7 @@ export async function run(_command, args = []) {
         codex_doctor: codexDoctor,
         require_codex_doctor: flag(args, '--fix') || flag(args, '--require-actual-codex'),
         zellij,
+        local_model: localModel,
         repair: configRepair,
         codex_app_ui: codexAppUi,
         require_codex_cli_config_load: flag(args, '--fix') || flag(args, '--require-actual-codex'),
@@ -175,6 +178,7 @@ export async function run(_command, args = []) {
         codex_doctor: codexDoctor,
         codex_doctor_diff: codexDoctorDiff,
         zellij,
+        local_model: localModel,
         zellij_readiness: zellijReadiness,
         codex_permission_profiles: permissionProfiles,
         imagegen: {
@@ -234,6 +238,16 @@ export async function run(_command, args = []) {
         }
     }
     console.log(`codex-lb:  ${codexLb.ok ? 'ok' : `warning ${codexLb.circuit?.state || 'unknown'}`}`);
+    if (localModel) {
+        console.log('Local LLM:');
+        console.log(`  enabled: ${localModel.enabled ? 'yes' : 'no'}`);
+        console.log(`  status: ${localModel.status}`);
+        console.log(`  provider: ${localModel.provider}`);
+        console.log(`  model: ${localModel.model}`);
+        console.log(`  endpoint: ${localModel.base_url}`);
+        console.log(`  last smoke: ${localModel.last_smoke?.ok ? `ok ${localModel.last_smoke.latency_ms || 0}ms ${localModel.last_smoke.tokens_per_second || 0} tok/s` : 'missing'}`);
+        console.log('  final arbiter: GPT required');
+    }
     console.log(`Permissions: config profile and permission profile are tracked separately (${permissionProfiles.codex_config_profile_field}, ${permissionProfiles.codex_permission_profile_field})`);
     console.log('Ready:');
     console.log(`  cli_ready: ${ready.cli_ready ? 'yes' : 'no'}`);

package/dist/core/agents/agent-proof-evidence.js

@@ -31,6 +31,13 @@ export async function writeAgentProofEvidence(root, input) {
     const patchSwarm = input.patchSwarm || await readJson(path.join(root, 'agent-patch-swarm-runtime.json'), null);
     const localCollaborationPolicy = input.localCollaborationPolicy || await readJson(path.join(root, 'local-collaboration-policy.json'), null) || resolveLocalCollaborationPolicy();
     const gptFinalArbiter = input.gptFinalArbiter || await readJson(path.join(root, 'gpt-final-arbiter', 'gpt-final-arbiter.json'), null);
+    const narutoWorkGraph = await readJson(path.join(root, 'naruto-work-graph.json'), null);
+    const narutoRoleDistribution = await readJson(path.join(root, 'naruto-role-distribution.json'), null);
+    const narutoConcurrencyGovernor = await readJson(path.join(root, 'naruto-concurrency-governor.json'), null);
+    const narutoActivePool = await readJson(path.join(root, 'naruto-active-pool.json'), null);
+    const narutoVerificationDag = await readJson(path.join(root, 'naruto-verification-dag.json'), null);
+    const narutoGptFinalPack = await readJson(path.join(root, 'naruto-gpt-final-pack.json'), null);
+    const narutoZellijDashboard = await readJson(path.join(root, 'naruto-zellij-dashboard.json'), null);
     const localParticipated = localCollaborationParticipated(input.results || []) || Number(gptFinalArbiter?.local_outputs_count || 0) > 0;
     const finalGptPatchStage = input.finalGptPatchStage || null;
     const localFinalGate = gptFinalArbiter?.final_gate || evaluateLocalCollaborationFinalGate({
@@ -74,6 +81,7 @@ export async function writeAgentProofEvidence(root, input) {
     const workQueueGoalRefsOk = Boolean(workQueue?.items?.length) && workQueue.items.every((item) => item.goal_mode_ref);
     const workQueueStrategyRefsOk = Boolean(workQueue?.items?.length) && workQueue.items.every((item) => item.slice?.strategy_refs);
     const route = String(input.route || taskGraph?.route_type || '$Agent');
+    const isNarutoRoute = route === '$Naruto';
     const routeCommand = String(input.routeCommand || 'sks agent run');
     const genericAgentRouteStandIn = !/\$?agent$/i.test(route) && /\bagent\s+run\b/i.test(routeCommand) && /--route/i.test(routeCommand);
     const realRouteCommandUsed = !genericAgentRouteStandIn;
@@ -177,6 +185,16 @@ export async function writeAgentProofEvidence(root, input) {
         ...(localParticipated && localFinalGate.ok !== true ? localFinalGate.blockers || ['gpt_final_arbiter_gate_not_ok'] : []),
         ...(localParticipated && gptFinalArbiter?.ok !== true ? gptFinalArbiter?.blockers || ['gpt_final_arbiter_not_ok'] : []),
         ...(localParticipated && finalGptPatchStage?.ok === false ? finalGptPatchStage.blockers || ['final_gpt_patch_stage_not_ok'] : []),
+        ...(isNarutoRoute && !narutoWorkGraph ? ['naruto_work_graph_missing'] : []),
+        ...(isNarutoRoute && narutoWorkGraph?.ok === false ? narutoWorkGraph.blockers || ['naruto_work_graph_not_ok'] : []),
+        ...(isNarutoRoute && !narutoRoleDistribution ? ['naruto_role_distribution_missing'] : []),
+        ...(isNarutoRoute && narutoRoleDistribution?.ok === false ? narutoRoleDistribution.blockers || ['naruto_role_distribution_not_ok'] : []),
+        ...(isNarutoRoute && !narutoConcurrencyGovernor ? ['naruto_concurrency_governor_missing'] : []),
+        ...(isNarutoRoute && !narutoActivePool ? ['naruto_active_pool_missing'] : []),
+        ...(isNarutoRoute && narutoActivePool?.ok === false ? narutoActivePool.blockers || ['naruto_active_pool_not_ok'] : []),
+        ...(isNarutoRoute && !narutoVerificationDag ? ['naruto_verification_dag_missing'] : []),
+        ...(isNarutoRoute && !narutoGptFinalPack ? ['naruto_gpt_final_pack_missing'] : []),
+        ...(isNarutoRoute && !narutoZellijDashboard ? ['naruto_zellij_dashboard_missing'] : []),
         ...agentChangedFileLeaseViolations(input.results || [], input.partition?.leases || [])
     ];
     const evidence = {
@@ -222,6 +240,23 @@ export async function writeAgentProofEvidence(root, input) {
         gpt_final_patch_source: finalGptPatchStage?.final_patch_source || (localParticipated ? 'blocked' : 'not_applicable'),
         gpt_final_gate_ok: localFinalGate.ok === true,
         gpt_final_gate: localFinalGate,
+        naruto_work_graph: narutoWorkGraph ? 'naruto-work-graph.json' : null,
+        naruto_total_work_items: Number(narutoWorkGraph?.total_work_items || 0),
+        naruto_mixed_work_kinds: narutoWorkGraph?.mixed_work_kinds || [],
+        naruto_write_allowed_count: Number(narutoWorkGraph?.write_allowed_count || 0),
+        naruto_role_distribution: narutoRoleDistribution ? 'naruto-role-distribution.json' : null,
+        naruto_role_distribution_entries: narutoRoleDistribution?.entries || [],
+        naruto_verifier_only: narutoRoleDistribution?.verifier_only === true,
+        naruto_implementation_like_ratio: Number(narutoRoleDistribution?.implementation_like_ratio || 0),
+        naruto_concurrency_governor: narutoConcurrencyGovernor ? 'naruto-concurrency-governor.json' : null,
+        naruto_safe_active_workers: Number(narutoConcurrencyGovernor?.safe_active_workers || 0),
+        naruto_safe_zellij_visible_panes: Number(narutoConcurrencyGovernor?.safe_zellij_visible_panes || 0),
+        naruto_headless_workers: Number(narutoConcurrencyGovernor?.headless_workers || 0),
+        naruto_active_pool: narutoActivePool ? 'naruto-active-pool.json' : null,
+        naruto_active_pool_refill_events: Number(narutoActivePool?.refill_events || 0),
+        naruto_verification_dag: narutoVerificationDag ? 'naruto-verification-dag.json' : null,
+        naruto_gpt_final_pack: narutoGptFinalPack ? 'naruto-gpt-final-pack.json' : null,
+        naruto_zellij_dashboard: narutoZellijDashboard ? 'naruto-zellij-dashboard.json' : null,
         patch_swarm_runtime: patchSwarm ? 'agent-patch-swarm-runtime.json' : null,
         patch_queue: patchSwarm ? 'agent-patch-queue.json' : null,
         patch_queue_events: patchSwarm ? 'agent-patch-queue-events.jsonl' : null,

package/dist/core/agents/agent-roster.js

@@ -2,6 +2,7 @@ import os from 'node:os';
 import { DEFAULT_AGENT_CONCURRENCY, DEFAULT_AGENT_COUNT, DEFAULT_NARUTO_CLONES, MAX_AGENT_COUNT, MAX_NARUTO_AGENT_COUNT, agentSessionId } from './agent-schema.js';
 import { defaultAgentPersonas, validatePersonaUniqueness } from './agent-persona.js';
 import { buildAgentEffortPolicy, decideAgentEffort, decideNarutoCloneEffort } from './agent-effort-policy.js';
+import { mapNarutoRoleToAgentRole, narutoRoleAllowsWrite } from '../naruto/naruto-role-policy.js';
 // $Naruto must never blindly spawn the full clone count at once, but the live
 // CONCURRENCY ceiling is NOT a function of CPU cores. Each clone is a separate CLI
 // worker process that spends ~all of its wall-clock awaiting the Codex API
@@ -156,28 +157,40 @@ export function buildNarutoCloneRoster(opts = {}) {
     const basePool = pool.length ? pool : defaultAgentPersonas(DEFAULT_AGENT_COUNT);
     const personas = [];
     const roster = [];
+    const roleCycle = narutoRoleCycle(readonly);
     for (let index = 0; index < cloneCount; index += 1) {
         const base = basePool[index % basePool.length];
         const cloneTag = 'clone-' + String(index + 1).padStart(3, '0');
         const id = 'naruto_' + cloneTag.replace(/-/g, '_');
-        const cloneReadonly = readonly || base.read_only;
+        const narutoRole = roleCycle[index % roleCycle.length] || 'verifier';
+        const writeAllowed = !readonly && narutoRoleAllowsWrite(narutoRole);
+        const cloneReadonly = readonly || !writeAllowed;
+        const role = mapNarutoRoleToAgentRole(narutoRole);
+        const allowedTools = writeAllowed ? ['read', 'search', 'edit', 'test'] : narutoRole === 'verifier' ? ['read', 'search', 'test'] : ['read', 'search'];
         // Dynamic per-clone effort like team mode, capped at low/medium and always fast.
-        const effort = decideNarutoCloneEffort({ persona: base, prompt: opts.prompt || '', agentId: id, readonly: cloneReadonly });
+        const effort = decideNarutoCloneEffort({ persona: { ...base, role, allowed_tools: allowedTools, read_only: cloneReadonly, write_policy: writeAllowed ? 'exclusive Naruto patch-envelope lease required' : 'read-only Naruto role' }, prompt: opts.prompt || '', agentId: id, readonly: cloneReadonly });
         const persona = {
             ...base,
             id,
             stable_id: base.stable_id + '-' + cloneTag,
-            read_only: readonly || base.read_only,
-            prompt: 'SHADOW CLONE: ' + cloneTag + ' (Kage Bunshin of ' + base.stable_id + ')\n' + base.prompt
+            role,
+            naruto_role: narutoRole,
+            write_allowed: writeAllowed,
+            read_only: cloneReadonly,
+            allowed_tools: allowedTools,
+            write_policy: writeAllowed ? 'exclusive Naruto patch-envelope lease required' : 'read-only Naruto role',
+            prompt: 'SHADOW CLONE: ' + cloneTag + ' (Kage Bunshin of ' + base.stable_id + ')\nNARUTO ROLE: ' + narutoRole + '\n' + base.prompt
         };
         personas.push(persona);
         roster.push({
             id,
             session_id: agentSessionId(id, index + 1),
             persona_id: id,
-            role: base.role,
+            role,
+            naruto_role: narutoRole,
+            write_allowed: writeAllowed,
             index: index + 1,
-            write_policy: cloneReadonly ? 'read-only' : base.write_policy,
+            write_policy: cloneReadonly ? 'read-only' : 'exclusive Naruto patch-envelope lease required',
             status: 'pending',
             reasoning_effort: effort.reasoning_effort,
             model_reasoning_effort: effort.model_reasoning_effort,
@@ -208,4 +221,20 @@ export function buildNarutoCloneRoster(opts = {}) {
     };
     return { ...result, effort_policy: buildAgentEffortPolicy(result) };
 }
+function narutoRoleCycle(readonly) {
+    if (readonly)
+        return ['verifier', 'researcher', 'verifier', 'gpt_final_arbiter'];
+    return [
+        'implementer',
+        'modifier',
+        'test_writer',
+        'verifier',
+        'researcher',
+        'conflict_resolver',
+        'rollback_planner',
+        'integrator',
+        'modifier',
+        'test_writer'
+    ];
+}
 //# sourceMappingURL=agent-roster.js.map

package/dist/core/agents/agent-schema.js

@@ -14,7 +14,7 @@ export const DEFAULT_AGENT_CONCURRENCY = 5;
 // cap; every other roster/scheduler caller keeps MAX_AGENT_COUNT as the default.
 export const MAX_NARUTO_AGENT_COUNT = 100;
 export const DEFAULT_NARUTO_CLONES = 12;
-export const AGENT_BACKENDS = ['fake', 'process', 'codex-sdk', 'zellij', 'ollama'];
+export const AGENT_BACKENDS = ['fake', 'process', 'codex-sdk', 'zellij', 'ollama', 'local-llm'];
 export function normalizeAgentBackend(input) {
     const value = String(input || 'codex-sdk');
     return AGENT_BACKENDS.includes(value) ? value : 'codex-sdk';

package/dist/core/agents/native-worker-backend-router.js

@@ -86,7 +86,8 @@ export async function runNativeWorkerBackendRouter(input) {
             verification: { status: ollamaRun.status === 'done' ? 'passed' : 'failed', checks: [...(ollamaRun.verification?.checks || []), 'native-worker-backend-router', 'ollama-api-generate'] }
         });
     }
-    else if (backend === 'codex-sdk' || backend === 'zellij') {
+    else if (backend === 'codex-sdk' || backend === 'zellij' || backend === 'local-llm') {
+        const localPreferred = backend === 'local-llm';
         const sdkTask = await runCodexTask({
             route: String(input.intake.route || '$Agent'),
             tier: 'worker',
@@ -111,6 +112,12 @@ export async function runNativeWorkerBackendRouter(input) {
                 user_confirmed_full_access: false,
                 mad_sks_authorized: input.intake.mad_sks_authorized === true || process.env.SKS_MAD_SKS_ACTIVE === '1'
             },
+            backendPreference: localPreferred ? ['local-llm', 'codex-sdk'] : ['codex-sdk'],
+            allowLocalLlm: localPreferred,
+            ...(localPreferred ? { localLlmPolicy: {
+                    mode: 'local_preferred',
+                    requiresGptFinal: true
+                } } : {}),
             mutationLedgerRoot: path.join(root, input.workerDirRel),
             zellijPaneId: await readZellijPaneId(root, input.workerDirRel),
             reliabilityPolicy: {
@@ -121,12 +128,14 @@ export async function runNativeWorkerBackendRouter(input) {
         outputLastMessagePath = sdkTask.workerResultPath;
         const sdkWorkerResult = await readJson(sdkTask.workerResultPath, null);
         patchEnvelopes = normalizeSdkPatchEnvelopes(sdkWorkerResult?.patch_envelopes || [], input, sdkTask.sdkThreadId);
-        proofLevel = sdkTask.ok ? (patchEnvelopes.length ? 'model_authored' : 'codex_sdk_thread_proven') : 'blocked';
+        proofLevel = sdkTask.ok ? (patchEnvelopes.length ? 'model_authored' : sdkTask.backend === 'local-llm' ? 'local_llm_worker_proven' : 'codex_sdk_thread_proven') : 'blocked';
         const sdkReport = {
             schema: 'sks.codex-sdk-worker-adapter.v1',
-            backend: 'codex-sdk',
+            backend: sdkTask.backend,
+            backend_family: sdkTask.backend_family,
             sdk_thread_id: sdkTask.sdkThreadId,
             sdk_run_id: sdkTask.sdkRunId,
+            local_llm_proof_path: sdkTask.localLlmProofPath || null,
             stream_event_count: sdkTask.streamEventCount,
             structured_output_valid: sdkTask.structuredOutputValid,
             worker_result_path: sdkTask.workerResultPath,
@@ -136,17 +145,30 @@ export async function runNativeWorkerBackendRouter(input) {
         childReports = [sdkReport];
         result = validateAgentWorkerResult({
             ...sdkWorkerResult,
-            backend: 'codex-sdk',
+            backend: sdkTask.backend === 'local-llm' ? 'local-llm' : 'codex-sdk',
             patch_envelopes: patchEnvelopes,
             codex_child_report: sdkReport,
             codex_sdk_thread: sdkReport,
             model_authored_patch_envelopes: patchEnvelopes.length > 0,
             fixture_patch_envelopes: false,
-            artifacts: [...new Set([...(sdkWorkerResult?.artifacts || []), path.relative(root, sdkTask.workerResultPath), path.join(input.workerDirRel, 'codex-control-proof.json'), path.join(input.workerDirRel, 'codex-thread-registry.json'), path.join(input.workerDirRel, 'codex-sdk-events.jsonl')])],
+            artifacts: [...new Set([
+                    ...(sdkWorkerResult?.artifacts || []),
+                    path.relative(root, sdkTask.workerResultPath),
+                    path.join(input.workerDirRel, 'codex-control-proof.json'),
+                    path.join(input.workerDirRel, 'codex-thread-registry.json'),
+                    sdkTask.backend === 'local-llm' ? path.join(input.workerDirRel, 'local-llm-events.jsonl') : path.join(input.workerDirRel, 'codex-sdk-events.jsonl'),
+                    ...(sdkTask.localLlmProofPath ? [path.relative(root, sdkTask.localLlmProofPath)] : [])
+                ])],
             blockers: [...(sdkWorkerResult?.blockers || []), ...sdkTask.blockers],
             verification: {
                 status: sdkTask.ok ? 'passed' : 'failed',
-                checks: [...(sdkWorkerResult?.verification?.checks || []), 'codex-sdk-control-plane', 'codex-sdk-event-stream', 'codex-sdk-structured-output']
+                checks: [
+                    ...(sdkWorkerResult?.verification?.checks || []),
+                    sdkTask.backend === 'local-llm' ? 'local-llm-control-plane' : 'codex-sdk-control-plane',
+                    sdkTask.backend === 'local-llm' ? 'local-llm-event-stream' : 'codex-sdk-event-stream',
+                    sdkTask.backend === 'local-llm' ? 'local-llm-structured-output' : 'codex-sdk-structured-output',
+                    ...(sdkTask.backend === 'local-llm' ? ['gpt-final-required-before-acceptance'] : [])
+                ]
             }
         });
     }
@@ -221,13 +243,13 @@ async function maybeAutoSelectOllamaBackend(backend, input) {
         model: input.intake?.ollama_model || null,
         baseUrl: input.intake?.ollama_base_url || null
     }).catch(() => null);
-    if (!config?.ok || config.enabled !== true)
+    if (!config?.ok || config.enabled !== true || config.status !== 'verified')
         return backend;
     const policy = classifyOllamaWorkerSlice(input.slice, { route: input.intake?.route, agent: input.agent });
-    return policy.ok ? 'ollama' : backend;
+    return policy.ok ? 'local-llm' : backend;
 }
 function normalizeBackend(value) {
-    return value === 'fake' || value === 'process' || value === 'codex-sdk' || value === 'zellij' || value === 'ollama' ? value : null;
+    return value === 'fake' || value === 'process' || value === 'codex-sdk' || value === 'zellij' || value === 'ollama' || value === 'local-llm' ? value : null;
 }
 function envelopeOpts(input, source, childPid) {
     return {