@rigour-labs/core 4.3.3 → 4.3.4

package/dist/inference/cloud-provider.js

@@ -2,13 +2,13 @@
 const DEFAULT_MODELS = {
     claude: 'claude-opus-4-6',
     anthropic: 'claude-sonnet-4-6',
-    openai: 'gpt-4o-mini',
-    gemini: 'gemini-3-flash',
-    groq: 'llama-3.1-70b-versatile',
+    openai: 'gpt-5-mini',
+    gemini: 'gemini-2.5-flash',
+    groq: 'llama-3.3-70b-versatile',
     mistral: 'mistral-large-latest',
-    together: 'meta-llama/Llama-3.1-70B-Instruct-Turbo',
-    fireworks: 'accounts/fireworks/models/llama-v3p1-70b-instruct',
-    deepseek: 'deepseek-coder',
+    together: 'meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8',
+    fireworks: 'accounts/fireworks/models/llama-v3p3-70b-instruct',
+    deepseek: 'deepseek-v4',
     perplexity: 'llama-3.1-sonar-large-128k-online',
     ollama: 'qwen2.5-coder:7b',
     lmstudio: 'qwen2.5-coder-7b-instruct',

package/dist/inference/model-manager.d.ts

@@ -13,9 +13,17 @@ export declare function getModelPath(tier: ModelTier): string;
  * Get model info for a tier.
  */
 export declare function getModelInfo(tier: ModelTier): ModelInfo;
+/**
+ * Check HuggingFace for a newer model version (like antivirus signature updates).
+ * Reads latest_version.json from the RLAIF dataset repo. Non-blocking — if the
+ * check fails (offline, HF down), we silently use the cached/bundled version.
+ *
+ * Results are cached locally for 24 hours to avoid hammering HF on every run.
+ */
+export declare function checkForUpdates(onProgress?: (message: string, percent?: number) => void): Promise<string>;
 /**
  * Download a model from HuggingFace CDN.
- * Tries fine-tuned model first, falls back to stock Qwen if unavailable.
+ * Checks for updates first, then tries fine-tuned model, falls back to stock Qwen.
  */
 export declare function downloadModel(tier: ModelTier, onProgress?: (message: string, percent?: number) => void): Promise<string>;
 /**

package/dist/inference/model-manager.js

@@ -6,8 +6,11 @@ import path from 'path';
 import fs from 'fs-extra';
 import { createHash } from 'crypto';
 import { RIGOUR_DIR } from '../storage/db.js';
-import { MODELS, FALLBACK_MODELS } from './types.js';
+import { MODELS, FALLBACK_MODELS, VERSION_CHECK_URL, BUNDLED_MODEL_VERSION, updateModelVersion } from './types.js';
 const MODELS_DIR = path.join(RIGOUR_DIR, 'models');
+const VERSION_CACHE_PATH = path.join(MODELS_DIR, '.latest_version.json');
+const VERSION_CHECK_INTERVAL_MS = 24 * 60 * 60 * 1000; // Check once per day
+const VERSION_CHECK_TIMEOUT_MS = 5000; // 5s timeout — don't block startup
 const SHA256_RE = /^[a-f0-9]{64}$/i;
 function getModelMetadataPath(filename) {
     return path.join(MODELS_DIR, filename + '.meta.json');
@@ -173,12 +176,66 @@ async function downloadFromUrl(tier, model, onProgress) {
         throw error;
     }
 }
+/**
+ * Check HuggingFace for a newer model version (like antivirus signature updates).
+ * Reads latest_version.json from the RLAIF dataset repo. Non-blocking — if the
+ * check fails (offline, HF down), we silently use the cached/bundled version.
+ *
+ * Results are cached locally for 24 hours to avoid hammering HF on every run.
+ */
+export async function checkForUpdates(onProgress) {
+    fs.ensureDirSync(MODELS_DIR);
+    // Check local version cache first — avoid network on every run
+    try {
+        if (await fs.pathExists(VERSION_CACHE_PATH)) {
+            const cached = await fs.readJson(VERSION_CACHE_PATH);
+            const age = Date.now() - new Date(cached.checkedAt).getTime();
+            if (age < VERSION_CHECK_INTERVAL_MS && cached.version) {
+                const v = String(cached.version);
+                updateModelVersion(v);
+                return v;
+            }
+        }
+    }
+    catch {
+        // Corrupted cache — proceed to network check
+    }
+    // Fetch latest version from HuggingFace (with timeout)
+    try {
+        const controller = new AbortController();
+        const timeout = setTimeout(() => controller.abort(), VERSION_CHECK_TIMEOUT_MS);
+        const response = await fetch(VERSION_CHECK_URL, { signal: controller.signal });
+        clearTimeout(timeout);
+        if (response.ok) {
+            const data = await response.json();
+            const latestVersion = String(data.version || BUNDLED_MODEL_VERSION);
+            // Cache the result locally
+            await fs.writeJson(VERSION_CACHE_PATH, {
+                version: latestVersion,
+                checkedAt: new Date().toISOString(),
+                source: 'huggingface',
+            }, { spaces: 2 }).catch(() => { });
+            // Update in-memory model definitions
+            updateModelVersion(latestVersion);
+            if (latestVersion !== BUNDLED_MODEL_VERSION) {
+                onProgress?.(`Model update available: v${latestVersion}`, 0);
+            }
+            return latestVersion;
+        }
+    }
+    catch {
+        // Offline / HF down / timeout — use bundled version silently
+    }
+    return BUNDLED_MODEL_VERSION;
+}
 /**
  * Download a model from HuggingFace CDN.
- * Tries fine-tuned model first, falls back to stock Qwen if unavailable.
+ * Checks for updates first, then tries fine-tuned model, falls back to stock Qwen.
  */
 export async function downloadModel(tier, onProgress) {
     fs.ensureDirSync(MODELS_DIR);
+    // Check for newer model version (non-blocking, cached 24h)
+    await checkForUpdates(onProgress);
     if (await isModelCached(tier)) {
         onProgress?.(`Model ${MODELS[tier].name} already cached`, 100);
         return getModelPath(tier);

package/dist/inference/types.d.ts

@@ -62,7 +62,7 @@ export interface DeepAnalysisResult {
  * Available model tiers.
  *
  * - deep: Qwen2.5-Coder-1.5B fine-tuned — full power, company-hosted
- * - lite: Qwen3.5-0.8B fine-tuned — lightweight, ships as default CLI sidecar
+ * - lite: Qwen2.5-Coder-0.5B fine-tuned — lightweight, ships as default CLI sidecar
  * - legacy: Qwen2.5-Coder-0.5B fine-tuned — previous default, reproducibility
  */
 export type ModelTier = 'deep' | 'lite' | 'legacy';
@@ -78,13 +78,24 @@ export interface ModelInfo {
     sizeHuman: string;
 }
 /**
- * Model version — bump when new fine-tuned GGUF is published.
- * The RLAIF pipeline uploads new models to HuggingFace, and
- * model-manager checks this version to auto-update.
+ * Minimum bundled model version — used as fallback when the auto-update
+ * check fails (offline, HF down, first run). The RLAIF training pipeline
+ * publishes new versions to HuggingFace and updates latest_version.json.
+ * At startup, model-manager checks HF for the latest version and downloads
+ * it automatically (like antivirus signature updates).
  */
-export declare const MODEL_VERSION = "1";
-/** All supported model definitions */
+export declare const BUNDLED_MODEL_VERSION = "1";
+/** HuggingFace dataset repo where latest_version.json lives */
+export declare const VERSION_CHECK_URL = "https://huggingface.co/datasets/rigour-labs/rigour-rlaif-data/resolve/main/latest_version.json";
+/** Build model info for a given tier and version */
+export declare function buildModelInfo(tier: ModelTier, version: string): ModelInfo;
+/** Current model definitions — initialized with bundled version, updated at runtime */
 export declare const MODELS: Record<ModelTier, ModelInfo>;
+/**
+ * Update MODELS in-place to point to a newer version.
+ * Called by model-manager after checking latest_version.json.
+ */
+export declare function updateModelVersion(version: string): void;
 /**
  * Fallback stock models — used when fine-tuned model is not yet
  * available on HuggingFace (initial setup / first-time users).

package/dist/inference/types.js

@@ -1,36 +1,46 @@
 /**
- * Model version — bump when new fine-tuned GGUF is published.
- * The RLAIF pipeline uploads new models to HuggingFace, and
- * model-manager checks this version to auto-update.
+ * Minimum bundled model version — used as fallback when the auto-update
+ * check fails (offline, HF down, first run). The RLAIF training pipeline
+ * publishes new versions to HuggingFace and updates latest_version.json.
+ * At startup, model-manager checks HF for the latest version and downloads
+ * it automatically (like antivirus signature updates).
  */
-export const MODEL_VERSION = '1';
-/** All supported model definitions */
+export const BUNDLED_MODEL_VERSION = '1';
+/** HuggingFace dataset repo where latest_version.json lives */
+export const VERSION_CHECK_URL = 'https://huggingface.co/datasets/rigour-labs/rigour-rlaif-data/resolve/main/latest_version.json';
+/** Build model info for a given tier and version */
+export function buildModelInfo(tier, version) {
+    const meta = {
+        deep: { base: 'Qwen2.5-Coder-1.5B', size: 900_000_000, sizeH: '900MB' },
+        lite: { base: 'Qwen2.5-Coder-0.5B', size: 500_000_000, sizeH: '500MB' },
+        legacy: { base: 'Qwen2.5-Coder-0.5B', size: 350_000_000, sizeH: '350MB' },
+    };
+    const m = meta[tier];
+    return {
+        tier,
+        name: `Rigour-${tier[0].toUpperCase() + tier.slice(1)}-v${version} (${m.base} fine-tuned)`,
+        filename: `rigour-${tier}-v${version}-q4_k_m.gguf`,
+        url: `https://huggingface.co/rigour-labs/rigour-${tier}-v${version}-gguf/resolve/main/rigour-${tier}-v${version}-q4_k_m.gguf`,
+        sizeBytes: m.size,
+        sizeHuman: m.sizeH,
+    };
+}
+/** Current model definitions — initialized with bundled version, updated at runtime */
 export const MODELS = {
-    deep: {
-        tier: 'deep',
-        name: 'Rigour-Deep-v1 (Qwen2.5-Coder-1.5B fine-tuned)',
-        filename: `rigour-deep-v${MODEL_VERSION}-q4_k_m.gguf`,
-        url: `https://huggingface.co/rigour-labs/rigour-deep-v1-gguf/resolve/main/rigour-deep-v${MODEL_VERSION}-q4_k_m.gguf`,
-        sizeBytes: 900_000_000,
-        sizeHuman: '900MB',
-    },
-    lite: {
-        tier: 'lite',
-        name: 'Rigour-Lite-v1 (Qwen3.5-0.8B fine-tuned)',
-        filename: `rigour-lite-v${MODEL_VERSION}-q4_k_m.gguf`,
-        url: `https://huggingface.co/rigour-labs/rigour-lite-v1-gguf/resolve/main/rigour-lite-v${MODEL_VERSION}-q4_k_m.gguf`,
-        sizeBytes: 500_000_000,
-        sizeHuman: '500MB',
-    },
-    legacy: {
-        tier: 'legacy',
-        name: 'Rigour-Legacy-v1 (Qwen2.5-Coder-0.5B fine-tuned)',
-        filename: `rigour-legacy-v${MODEL_VERSION}-q4_k_m.gguf`,
-        url: `https://huggingface.co/rigour-labs/rigour-legacy-v1-gguf/resolve/main/rigour-legacy-v${MODEL_VERSION}-q4_k_m.gguf`,
-        sizeBytes: 350_000_000,
-        sizeHuman: '350MB',
-    },
+    deep: buildModelInfo('deep', BUNDLED_MODEL_VERSION),
+    lite: buildModelInfo('lite', BUNDLED_MODEL_VERSION),
+    legacy: buildModelInfo('legacy', BUNDLED_MODEL_VERSION),
 };
+/**
+ * Update MODELS in-place to point to a newer version.
+ * Called by model-manager after checking latest_version.json.
+ */
+export function updateModelVersion(version) {
+    for (const tier of ['deep', 'lite', 'legacy']) {
+        const updated = buildModelInfo(tier, version);
+        MODELS[tier] = updated;
+    }
+}
 /**
  * Fallback stock models — used when fine-tuned model is not yet
  * available on HuggingFace (initial setup / first-time users).
@@ -46,9 +56,9 @@ export const FALLBACK_MODELS = {
     },
     lite: {
         tier: 'lite',
-        name: 'Qwen3.5-0.8B (stock)',
-        filename: 'qwen3.5-0.8b-q4_k_m.gguf',
-        url: 'https://huggingface.co/Qwen/Qwen3.5-0.8B-GGUF/resolve/main/qwen3.5-0.8b-q4_k_m.gguf',
+        name: 'Qwen2.5-Coder-0.5B-Instruct (stock)',
+        filename: 'qwen2.5-coder-0.5b-instruct-q4_k_m.gguf',
+        url: 'https://huggingface.co/Qwen/Qwen2.5-Coder-0.5B-Instruct-GGUF/resolve/main/qwen2.5-coder-0.5b-instruct-q4_k_m.gguf',
         sizeBytes: 500_000_000,
         sizeHuman: '500MB',
     },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@rigour-labs/core",
-  "version": "4.3.3",
+  "version": "4.3.4",
   "description": "Deterministic quality gate engine for AI-generated code. AST analysis, drift detection, and Fix Packet generation across TypeScript, JavaScript, Python, Go, Ruby, and C#.",
   "license": "MIT",
   "homepage": "https://rigour.run",
@@ -59,11 +59,11 @@
     "@xenova/transformers": "^2.17.2",
     "better-sqlite3": "^11.0.0",
     "openai": "^4.104.0",
-    "@rigour-labs/brain-darwin-arm64": "4.3.3",
-    "@rigour-labs/brain-linux-arm64": "4.3.3",
-    "@rigour-labs/brain-darwin-x64": "4.3.3",
-    "@rigour-labs/brain-win-x64": "4.3.3",
-    "@rigour-labs/brain-linux-x64": "4.3.3"
+    "@rigour-labs/brain-darwin-arm64": "4.3.4",
+    "@rigour-labs/brain-darwin-x64": "4.3.4",
+    "@rigour-labs/brain-linux-x64": "4.3.4",
+    "@rigour-labs/brain-linux-arm64": "4.3.4",
+    "@rigour-labs/brain-win-x64": "4.3.4"
   },
   "devDependencies": {
     "@types/better-sqlite3": "^7.6.12",