@revealui/ai 0.10.0 → 1.0.0

package/dist/llm/client.js

@@ -10,6 +10,7 @@ import { decryptApiKey } from '@revealui/db/crypto';
 import { tenantProviderConfigs, userApiKeys } from '@revealui/db/schema';
 import { CircuitBreaker, CircuitBreakerOpenError, } from '@revealui/resilience';
 import { and, eq } from 'drizzle-orm';
+import { defaultBaseURLForProvider, defaultModelForProvider, isGroqCatalogModel, resolveInferenceRoute, } from './inference-route.js';
 import { applyLocalAiProfileToEnv } from './local-ai-profile.js';
 import { AnthropicProvider } from './providers/anthropic.js';
 import { GroqProvider } from './providers/groq.js';
@@ -17,11 +18,10 @@ import { InferenceSnapsProvider, } from './providers/inference-snaps.js';
 import { OllamaProvider } from './providers/ollama.js';
 import { OpenAIProvider } from './providers/openai.js';
 import { OpenAICompatProvider } from './providers/openai-compat.js';
-import { DEFAULT_US_ORIGIN_INFERENCE_SNAP } from './providers/us-origin-snaps.js';
 import { XaiProvider } from './providers/xai.js';
 import { ResponseCache } from './response-cache.js';
-import { SemanticCache, } from './semantic-cache.js';
 import { estimateRequest as _estimateRequestTokens } from './token-counter.js';
+export { defaultBaseURLForProvider, defaultModelForProvider, GROQ_DEFAULT_BASE_URL, GROQ_DEFAULT_MODEL, groqAcceptedModel, isGroqCatalogModel, resolveInferenceRoute, resolveModelForProvider, } from './inference-route.js';
 /**
  * Providers reachable from a hosted (serverless) deployment. Localhost-only
  * providers are false. Consumed by PR-2/PR-3 (resolver + settings UI) to pick
@@ -51,13 +51,20 @@ export class LLMClient {
     rateLimitState;
     responseCache;
     semanticCache;
+    semanticCacheLoading;
     healthMonitor;
     circuitBreaker;
     fallbackCircuitBreaker;
     /** Tracks the last resolved API key so we only recreate the provider when it changes */
     currentApiKey;
     constructor(config) {
-        this.config = config;
+        const route = resolveInferenceRoute({
+            provider: config.provider,
+            model: config.model,
+            baseURL: config.baseURL,
+            groqCredentialAvailable: config.provider === 'groq',
+        });
+        this.config = { ...config, ...route };
         this.currentApiKey = config.apiKey;
         this.rateLimitState = {
             requests: [],
@@ -68,10 +75,10 @@ export class LLMClient {
         if (config.enableResponseCache) {
             this.responseCache = new ResponseCache(config.responseCacheOptions);
         }
-        // Initialize semantic cache if enabled
-        if (config.enableSemanticCache) {
-            this.semanticCache = new SemanticCache(config.semanticCacheOptions);
-        }
+        // SemanticCache is loaded on first chat() when enableSemanticCache is
+        // set. A static import pulls VectorMemoryService → @revealui/db/client →
+        // @revealui/config, which throws REVEALUI_PUBLIC_SERVER_URL in production
+        // (Apify Store 0.1.8). BYOK construction must not evaluate that graph.
         // Wire health monitor if provided
         this.healthMonitor = config.healthMonitor;
         // Per-provider circuit breakers — isolate outages so one provider's failure
@@ -87,22 +94,22 @@ export class LLMClient {
         }
         // Wire dedicated embed provider if supplied
         this.embedProviderOverride = config.embedProvider;
-        // Create primary provider
-        this.provider = this.createProvider(config.provider, {
-            apiKey: config.apiKey,
-            baseURL: config.baseURL,
-            model: config.model,
-            temperature: config.temperature,
-            maxTokens: config.maxTokens,
+        // Create primary provider from the resolved route (never a Groq id on OpenAI)
+        this.provider = this.createProvider(this.config.provider, {
+            apiKey: this.config.apiKey,
+            baseURL: this.config.baseURL,
+            model: this.config.model,
+            temperature: this.config.temperature,
+            maxTokens: this.config.maxTokens,
         });
         // Create fallback provider if specified
-        if (config.fallbackProvider) {
-            this.fallbackProvider = this.createProvider(config.fallbackProvider, {
-                apiKey: config.apiKey, // Note: In practice, you'd want separate API keys
-                baseURL: config.baseURL,
-                model: config.model,
-                temperature: config.temperature,
-                maxTokens: config.maxTokens,
+        if (this.config.fallbackProvider) {
+            this.fallbackProvider = this.createProvider(this.config.fallbackProvider, {
+                apiKey: this.config.apiKey, // Note: In practice, you'd want separate API keys
+                baseURL: this.config.baseURL,
+                model: this.config.model,
+                temperature: this.config.temperature,
+                maxTokens: this.config.maxTokens,
             });
         }
     }
@@ -185,13 +192,29 @@ export class LLMClient {
         this.rateLimitState.requests.push(now);
         this.rateLimitState.dailyRequests++;
     }
+    async ensureSemanticCache() {
+        if (!this.config.enableSemanticCache) {
+            return undefined;
+        }
+        if (this.semanticCache) {
+            return this.semanticCache;
+        }
+        if (!this.semanticCacheLoading) {
+            this.semanticCacheLoading = import('./semantic-cache.js').then(({ SemanticCache }) => {
+                this.semanticCache = new SemanticCache(this.config.semanticCacheOptions);
+                return this.semanticCache;
+            });
+        }
+        return this.semanticCacheLoading;
+    }
     async chat(messages, options) {
         await this.refreshProviderIfNeeded();
         // Check semantic cache first (if enabled)
         // Semantic cache is more powerful - matches similar queries, not just exact matches
-        if (this.semanticCache) {
-            const query = this.semanticCache.extractQuery(messages);
-            const cached = await this.semanticCache.get(query);
+        const semanticCache = await this.ensureSemanticCache();
+        if (semanticCache) {
+            const query = semanticCache.extractQuery(messages);
+            const cached = await semanticCache.get(query);
             if (cached) {
                 // Semantic cache hit - return immediately without API call
                 return {
@@ -241,9 +264,9 @@ export class LLMClient {
             const response = await this.circuitBreaker.execute(() => this.provider.chat(messages, options));
             this.healthMonitor?.recordCall(this.config.provider, Date.now() - callStart);
             // Store in semantic cache (if enabled)
-            if (this.semanticCache) {
-                const query = this.semanticCache.extractQuery(messages);
-                await this.semanticCache.set(query, response.content, response.usage);
+            if (semanticCache) {
+                const query = semanticCache.extractQuery(messages);
+                await semanticCache.set(query, response.content, response.usage);
             }
             // Store in response cache (if enabled)
             if (this.responseCache) {
@@ -464,13 +487,13 @@ export class LLMClient {
  *
  * Canonical Inference Snaps is the reference local provider on Ubuntu — offline,
  * silicon-optimized, no API key required. Product usage is **US-origin snaps
- * only** (nemotron-3-nano default; gemma3/4 also allowlisted). See
+ * only** (gemma3 default; gemma4 and nemotron also allowlisted). See
  * `providers/us-origin-snaps.ts` and `providers/inference-snaps.ts`.
  *
  * Provider defaults:
- *   inference-snaps → nemotron-3-nano   (base URL defaults to http://localhost:9090/v1)
- *   groq            → qwen/qwen3-32b
- *   ollama          → gemma4:e2b        (base URL defaults to http://localhost:11434)
+ *   inference-snaps → gemma3   (base URL defaults to http://localhost:9090/v1)
+ *   groq            → openai/gpt-oss-120b (Groq-accepted default; retired llama ids remap)
+ *   ollama          → DEFAULT_DAILY_OLLAMA_MODEL (qwen2.5:3b; base URL http://localhost:11434)
  *   anthropic       → claude-sonnet-4-6 (base URL defaults to https://api.anthropic.com/v1)
  *   openai          → gpt-4o            (base URL defaults to https://api.openai.com/v1)
  *   xai             → grok-4.5          (base URL defaults to https://api.x.ai/v1)
@@ -521,18 +544,18 @@ export function createLLMClientFromEnv() {
     let defaultModel;
     if (provider === 'anthropic') {
         apiKey = process.env.ANTHROPIC_API_KEY;
-        baseURL = process.env.ANTHROPIC_BASE_URL ?? 'https://api.anthropic.com/v1';
-        defaultModel = 'claude-sonnet-4-6';
+        baseURL = process.env.ANTHROPIC_BASE_URL ?? defaultBaseURLForProvider('anthropic');
+        defaultModel = defaultModelForProvider('anthropic');
     }
     else if (provider === 'openai') {
         apiKey = process.env.OPENAI_API_KEY;
-        baseURL = process.env.OPENAI_BASE_URL ?? 'https://api.openai.com/v1';
-        defaultModel = 'gpt-4o';
+        baseURL = process.env.OPENAI_BASE_URL ?? defaultBaseURLForProvider('openai');
+        defaultModel = defaultModelForProvider('openai');
     }
     else if (provider === 'xai') {
         apiKey = process.env.XAI_API_KEY;
-        baseURL = process.env.XAI_BASE_URL ?? 'https://api.x.ai/v1';
-        defaultModel = 'grok-4.5';
+        baseURL = process.env.XAI_BASE_URL ?? defaultBaseURLForProvider('xai');
+        defaultModel = defaultModelForProvider('xai');
     }
     else if (provider === 'huggingface') {
         apiKey = process.env.HF_TOKEN;
@@ -540,34 +563,43 @@ export function createLLMClientFromEnv() {
     }
     else if (provider === 'groq') {
         apiKey = process.env.GROQ_API_KEY;
-        baseURL = process.env.GROQ_BASE_URL;
-        defaultModel = 'qwen/qwen3-32b';
+        baseURL = process.env.GROQ_BASE_URL ?? defaultBaseURLForProvider('groq');
+        defaultModel = defaultModelForProvider('groq');
     }
     else if (provider === 'ollama') {
         apiKey = 'ollama'; // Ollama ignores the API key
         // Ollama's OpenAI-compatible endpoint lives at /v1
         const ollamaBase = process.env.OLLAMA_BASE_URL ?? 'http://localhost:11434';
         baseURL = ollamaBase.endsWith('/v1') ? ollamaBase : `${ollamaBase}/v1`;
-        defaultModel = 'gemma4:e2b';
+        defaultModel = defaultModelForProvider('ollama');
     }
     else if (provider === 'inference-snaps') {
         apiKey = 'inference-snaps'; // inference-snaps ignores the API key
         // Defaults to Canonical's Inference Snap local service on port 9090; override
         // via INFERENCE_SNAPS_BASE_URL when the snap listens on a non-default port.
         // Model defaults to the US-origin allowlist default (asserted in provider ctor).
-        baseURL = process.env.INFERENCE_SNAPS_BASE_URL ?? 'http://localhost:9090/v1';
-        defaultModel = DEFAULT_US_ORIGIN_INFERENCE_SNAP;
+        baseURL = process.env.INFERENCE_SNAPS_BASE_URL ?? defaultBaseURLForProvider('inference-snaps');
+        defaultModel = defaultModelForProvider('inference-snaps');
+    }
+    const route = resolveInferenceRoute({
+        provider,
+        model: process.env.LLM_MODEL ?? defaultModel,
+        baseURL,
+        groqCredentialAvailable: Boolean(process.env.GROQ_API_KEY),
+    });
+    if (route.provider === 'groq' && provider !== 'groq') {
+        apiKey = process.env.GROQ_API_KEY;
     }
     if (!apiKey) {
-        throw new Error(`API key not found for provider "${provider}". Set the corresponding env var ` +
+        throw new Error(`API key not found for provider "${route.provider}". Set the corresponding env var ` +
             `(INFERENCE_SNAPS_BASE_URL, GROQ_API_KEY, OLLAMA_BASE_URL, HF_TOKEN, ` +
             `ANTHROPIC_API_KEY, OPENAI_API_KEY, or XAI_API_KEY).`);
     }
     return new LLMClient({
-        provider,
+        provider: route.provider,
         apiKey,
-        baseURL,
-        model: process.env.LLM_MODEL ?? defaultModel,
+        baseURL: route.baseURL,
+        model: route.model,
         temperature: process.env.LLM_TEMPERATURE ? parseFloat(process.env.LLM_TEMPERATURE) : undefined,
         maxTokens: process.env.LLM_MAX_TOKENS ? parseInt(process.env.LLM_MAX_TOKENS, 10) : undefined,
         enableCacheByDefault: process.env.LLM_ENABLE_CACHE === 'true' || process.env.ANTHROPIC_ENABLE_CACHE === 'true',
@@ -580,9 +612,12 @@ export function createLLMClientFromEnv() {
 /**
  * Create an LLM client using a user's stored BYOK API key.
  *
- * Looks up the user's preferred provider from `tenant_provider_configs`
- * (falling back to the first key in `user_api_keys`), decrypts the key
- * with AES-256-GCM, and returns a configured LLMClient.
+ * Looks up the user's preferred provider from `tenant_provider_configs`.
+ * If that provider has no saved key, uses another saved key (hosted-viable
+ * when `opts.hostedViableOnly` is set — e.g. a saved Groq key). Decrypts
+ * with AES-256-GCM and returns a configured LLMClient whose model and
+ * base URL match the **key's** provider, never a stale preferred-provider
+ * model (a Groq id must not be sent to OpenAI).
  *
  * Returns `null` if the user has no stored keys (callers should fall back
  * to `createLLMClientFromEnv()` or return a 402/feature-unavailable error).
@@ -606,15 +641,22 @@ auditStore, opts) {
         .from(tenantProviderConfigs)
         .where(and(eq(tenantProviderConfigs.userId, userId), eq(tenantProviderConfigs.isDefault, true)))
         .limit(1);
-    // Find the matching API key (preferred provider, or any available key)
-    const keyQuery = db
-        .select()
-        .from(userApiKeys)
-        .where(preferredConfig
-        ? and(eq(userApiKeys.userId, userId), eq(userApiKeys.provider, preferredConfig.provider))
-        : eq(userApiKeys.userId, userId))
-        .limit(1);
-    const [keyRow] = await keyQuery;
+    // All of this user's keys. Preferred provider wins when it has a key;
+    // otherwise use another saved key (hosted-viable on hosted).
+    const keyRows = await db.select().from(userApiKeys).where(eq(userApiKeys.userId, userId));
+    const preferredKey = preferredConfig
+        ? keyRows.find((row) => row.provider === preferredConfig.provider)
+        : undefined;
+    const groqKey = keyRows.find((row) => row.provider === 'groq');
+    const fallbackKey = keyRows.find((row) => {
+        if (opts?.hostedViableOnly && !isHostedViable(row.provider)) {
+            return false;
+        }
+        return true;
+    });
+    const preferredModel = preferredConfig?.model ?? undefined;
+    const groqModelSelected = Boolean(preferredModel && isGroqCatalogModel(preferredModel));
+    const keyRow = groqModelSelected && groqKey ? groqKey : (preferredKey ?? fallbackKey);
     if (!keyRow)
         return null;
     const provider = keyRow.provider;
@@ -623,7 +665,12 @@ auditStore, opts) {
     if (opts?.hostedViableOnly && !isHostedViable(provider))
         return null;
     const plaintext = decryptApiKey(keyRow.encryptedKey);
-    const model = preferredConfig?.model ?? undefined;
+    const requestedModel = preferredConfig?.provider === provider || groqModelSelected ? preferredModel : undefined;
+    const route = resolveInferenceRoute({
+        provider,
+        model: requestedModel,
+        groqCredentialAvailable: provider === 'groq' || Boolean(groqKey),
+    });
     // Fire-and-forget: record when this key was last used (best-effort, never blocks)
     db.update(userApiKeys)
         .set({ lastUsedAt: new Date() })
@@ -638,10 +685,15 @@ auditStore, opts) {
             eventType: 'byok:key:accessed',
             severity: 'info',
             agentId: 'system',
-            payload: { userId, provider, keyId: keyRow.id },
+            payload: { userId, provider: route.provider, keyId: keyRow.id },
             policyViolations: [],
         })
             .catch(() => undefined);
     }
-    return new LLMClient({ provider, apiKey: plaintext, model });
+    return new LLMClient({
+        provider: route.provider,
+        apiKey: plaintext,
+        model: route.model,
+        baseURL: route.baseURL,
+    });
 }

package/dist/llm/inference-route.d.ts

@@ -0,0 +1,46 @@
+/**
+ * Provider + model + host pairing for LLM clients.
+ *
+ * A2A Send Task and /api/agent-stream both construct an LLMClient through
+ * resolveLLMClientForRequest. This module is the single place that refuses
+ * the production pairing: a Groq catalog id on an OpenAI host
+ * (`llama-3.3-70b-versatile` posted to api.openai.com).
+ */
+export type LLMProviderType = 'anthropic' | 'openai' | 'groq' | 'huggingface' | 'ollama' | 'inference-snaps' | 'xai';
+/** Groq's recommended replacement for llama-3.3-70b-versatile (retired 2026-08-16). */
+export declare const GROQ_DEFAULT_MODEL = "openai/gpt-oss-120b";
+export declare const GROQ_DEFAULT_BASE_URL = "https://api.groq.com/openai/v1";
+export declare function isGroqCatalogModel(model: string): boolean;
+/** Default catalog model for a provider. */
+export declare function defaultModelForProvider(provider: LLMProviderType): string | undefined;
+/** Default OpenAI-compatible base URL for a provider. */
+export declare function defaultBaseURLForProvider(provider: LLMProviderType): string | undefined;
+/** Current Groq-accepted model. Retired catalog ids map to the Groq default. */
+export declare function groqAcceptedModel(requested: string | undefined): string;
+/**
+ * Pick a model that belongs to `provider`. A Groq catalog id is never
+ * forwarded to OpenAI (walk residual: `llama-3.3-70b-versatile` on api.openai.com).
+ */
+export declare function resolveModelForProvider(provider: LLMProviderType, requested: string | undefined): string | undefined;
+export interface InferenceRouteInput {
+    provider: LLMProviderType;
+    model?: string;
+    baseURL?: string;
+    /** True when a Groq credential is available (saved key or GROQ_API_KEY). */
+    groqCredentialAvailable?: boolean;
+}
+export interface InferenceRoute {
+    provider: LLMProviderType;
+    model?: string;
+    baseURL?: string;
+}
+/**
+ * Resolve provider, model, and host together.
+ *
+ * When the selected model is a Groq catalog id and a Groq credential exists,
+ * the request goes to Groq's OpenAI-compatible host with a Groq-accepted
+ * model. When the OpenAI client is actually used, a Groq id is replaced
+ * with the OpenAI default — never posted to api.openai.com.
+ */
+export declare function resolveInferenceRoute(input: InferenceRouteInput): InferenceRoute;
+//# sourceMappingURL=inference-route.d.ts.map

package/dist/llm/inference-route.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"inference-route.d.ts","sourceRoot":"","sources":["../../src/llm/inference-route.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAOH,MAAM,MAAM,eAAe,GACvB,WAAW,GACX,QAAQ,GACR,MAAM,GACN,aAAa,GACb,QAAQ,GACR,iBAAiB,GACjB,KAAK,CAAC;AAEV,uFAAuF;AACvF,eAAO,MAAM,kBAAkB,wBAAwB,CAAC;AACxD,eAAO,MAAM,qBAAqB,mCAAmC,CAAC;AAyCtE,wBAAgB,kBAAkB,CAAC,KAAK,EAAE,MAAM,GAAG,OAAO,CASzD;AAED,4CAA4C;AAC5C,wBAAgB,uBAAuB,CAAC,QAAQ,EAAE,eAAe,GAAG,MAAM,GAAG,SAAS,CAiBrF;AAED,yDAAyD;AACzD,wBAAgB,yBAAyB,CAAC,QAAQ,EAAE,eAAe,GAAG,MAAM,GAAG,SAAS,CAiBvF;AAED,gFAAgF;AAChF,wBAAgB,iBAAiB,CAAC,SAAS,EAAE,MAAM,GAAG,SAAS,GAAG,MAAM,CAKvE;AAED;;;GAGG;AACH,wBAAgB,uBAAuB,CACrC,QAAQ,EAAE,eAAe,EACzB,SAAS,EAAE,MAAM,GAAG,SAAS,GAC5B,MAAM,GAAG,SAAS,CAQpB;AAED,MAAM,WAAW,mBAAmB;IAClC,QAAQ,EAAE,eAAe,CAAC;IAC1B,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,4EAA4E;IAC5E,uBAAuB,CAAC,EAAE,OAAO,CAAC;CACnC;AAED,MAAM,WAAW,cAAc;IAC7B,QAAQ,EAAE,eAAe,CAAC;IAC1B,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,OAAO,CAAC,EAAE,MAAM,CAAC;CAClB;AAiBD;;;;;;;GAOG;AACH,wBAAgB,qBAAqB,CAAC,KAAK,EAAE,mBAAmB,GAAG,cAAc,CAkBhF"}

package/dist/llm/inference-route.js

@@ -0,0 +1,154 @@
+/**
+ * Provider + model + host pairing for LLM clients.
+ *
+ * A2A Send Task and /api/agent-stream both construct an LLMClient through
+ * resolveLLMClientForRequest. This module is the single place that refuses
+ * the production pairing: a Groq catalog id on an OpenAI host
+ * (`llama-3.3-70b-versatile` posted to api.openai.com).
+ */
+import { DEFAULT_DAILY_OLLAMA_MODEL, DEFAULT_US_ORIGIN_INFERENCE_SNAP, } from './providers/us-origin-snaps.js';
+/** Groq's recommended replacement for llama-3.3-70b-versatile (retired 2026-08-16). */
+export const GROQ_DEFAULT_MODEL = 'openai/gpt-oss-120b';
+export const GROQ_DEFAULT_BASE_URL = 'https://api.groq.com/openai/v1';
+/**
+ * Groq catalog ids — current and retired — that must never be sent to
+ * api.openai.com. Set lookup plus prefix checks; no regex (M2).
+ */
+const GROQ_CATALOG_MODELS = new Set([
+    'llama-3.3-70b-versatile',
+    'llama-3.3-70b-specdec',
+    'llama-3.1-70b-versatile',
+    'llama-3.1-8b-instant',
+    'llama-3.2-90b-vision-preview',
+    'llama-3.2-11b-vision-preview',
+    'llama-3.2-3b-preview',
+    'llama-3.2-1b-preview',
+    'mixtral-8x7b-32768',
+    'gemma2-9b-it',
+    'gemma-7b-it',
+    'qwen/qwen3-32b',
+    'qwen/qwen3.6-27b',
+    'openai/gpt-oss-120b',
+    'openai/gpt-oss-20b',
+    'openai/gpt-oss-safeguard-20b',
+]);
+/** Retired Groq ids that the public API rejects (same error text as OpenAI). */
+const GROQ_RETIRED_MODELS = new Set([
+    'llama-3.3-70b-versatile',
+    'llama-3.3-70b-specdec',
+    'llama-3.1-70b-versatile',
+    'llama-3.1-8b-instant',
+    'llama-3.2-90b-vision-preview',
+    'llama-3.2-11b-vision-preview',
+    'llama-3.2-3b-preview',
+    'llama-3.2-1b-preview',
+    'mixtral-8x7b-32768',
+    'gemma2-9b-it',
+    'gemma-7b-it',
+    'qwen/qwen3-32b',
+]);
+export function isGroqCatalogModel(model) {
+    if (GROQ_CATALOG_MODELS.has(model))
+        return true;
+    return (model.startsWith('llama-3.') ||
+        model.startsWith('mixtral-') ||
+        model.startsWith('gemma2-') ||
+        model.startsWith('openai/gpt-oss') ||
+        model.startsWith('qwen/qwen3'));
+}
+/** Default catalog model for a provider. */
+export function defaultModelForProvider(provider) {
+    switch (provider) {
+        case 'anthropic':
+            return 'claude-sonnet-4-6';
+        case 'openai':
+            return 'gpt-4o';
+        case 'xai':
+            return 'grok-4.5';
+        case 'groq':
+            return GROQ_DEFAULT_MODEL;
+        case 'ollama':
+            return DEFAULT_DAILY_OLLAMA_MODEL;
+        case 'inference-snaps':
+            return DEFAULT_US_ORIGIN_INFERENCE_SNAP;
+        default:
+            return undefined;
+    }
+}
+/** Default OpenAI-compatible base URL for a provider. */
+export function defaultBaseURLForProvider(provider) {
+    switch (provider) {
+        case 'anthropic':
+            return 'https://api.anthropic.com/v1';
+        case 'openai':
+            return 'https://api.openai.com/v1';
+        case 'xai':
+            return 'https://api.x.ai/v1';
+        case 'groq':
+            return GROQ_DEFAULT_BASE_URL;
+        case 'ollama':
+            return 'http://localhost:11434/v1';
+        case 'inference-snaps':
+            return 'http://localhost:9090/v1';
+        default:
+            return undefined;
+    }
+}
+/** Current Groq-accepted model. Retired catalog ids map to the Groq default. */
+export function groqAcceptedModel(requested) {
+    if (requested && isGroqCatalogModel(requested) && !GROQ_RETIRED_MODELS.has(requested)) {
+        return requested;
+    }
+    return GROQ_DEFAULT_MODEL;
+}
+/**
+ * Pick a model that belongs to `provider`. A Groq catalog id is never
+ * forwarded to OpenAI (walk residual: `llama-3.3-70b-versatile` on api.openai.com).
+ */
+export function resolveModelForProvider(provider, requested) {
+    if (provider === 'groq') {
+        return groqAcceptedModel(requested);
+    }
+    if (requested && !isGroqCatalogModel(requested)) {
+        return requested;
+    }
+    return defaultModelForProvider(provider);
+}
+const OPENAI_API_HOSTNAME = 'api.openai.com';
+/**
+ * True only when `url` is a parseable absolute URL whose hostname is exactly
+ * api.openai.com. Prefix/substring checks are not used — `api.openai.com.evil.com`
+ * must not count as OpenAI (CodeQL js/incomplete-url-substring-sanitization).
+ */
+function isOpenAiHost(url) {
+    try {
+        return new URL(url).hostname === OPENAI_API_HOSTNAME;
+    }
+    catch {
+        return false;
+    }
+}
+/**
+ * Resolve provider, model, and host together.
+ *
+ * When the selected model is a Groq catalog id and a Groq credential exists,
+ * the request goes to Groq's OpenAI-compatible host with a Groq-accepted
+ * model. When the OpenAI client is actually used, a Groq id is replaced
+ * with the OpenAI default — never posted to api.openai.com.
+ */
+export function resolveInferenceRoute(input) {
+    const groqModel = Boolean(input.model && isGroqCatalogModel(input.model));
+    if (groqModel && (input.provider === 'groq' || input.groqCredentialAvailable)) {
+        const keepCustomGroqUrl = Boolean(input.provider === 'groq' && input.baseURL && !isOpenAiHost(input.baseURL));
+        return {
+            provider: 'groq',
+            model: groqAcceptedModel(input.model),
+            baseURL: keepCustomGroqUrl ? input.baseURL : defaultBaseURLForProvider('groq'),
+        };
+    }
+    return {
+        provider: input.provider,
+        model: resolveModelForProvider(input.provider, input.model),
+        baseURL: input.baseURL ?? defaultBaseURLForProvider(input.provider),
+    };
+}

package/dist/llm/local-ai-profile.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"local-ai-profile.d.ts","sourceRoot":"","sources":["../../src/llm/local-ai-profile.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAWH,MAAM,MAAM,WAAW,GAAG,MAAM,GAAG,OAAO,GAAG,OAAO,GAAG,OAAO,CAAC;AAE/D,MAAM,MAAM,eAAe,GAAG,QAAQ,GAAG,iBAAiB,GAAG,IAAI,CAAC;AAElE,MAAM,WAAW,cAAc;IAC7B,IAAI,EAAE,WAAW,CAAC;IAClB,QAAQ,EAAE,eAAe,CAAC;IAC1B,KAAK,EAAE,MAAM,GAAG,IAAI,CAAC;IACrB,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;IACvB,0CAA0C;IAC1C,eAAe,EAAE,MAAM,GAAG,IAAI,CAAC;IAC/B,8DAA8D;IAC9D,SAAS,EAAE,MAAM,GAAG,IAAI,CAAC;IACzB,SAAS,EAAE,MAAM,CAAC;IAClB,iDAAiD;IACjD,IAAI,CAAC,EAAE,MAAM,CAAC;CACf;AAED,wBAAgB,qBAAqB,IAAI,MAAM,CAK9C;AAED,gEAAgE;AAChE,wBAAgB,uBAAuB,IAAI,MAAM,CAEhD;AAED,6DAA6D;AAC7D,wBAAgB,gBAAgB,IAAI,cAAc,CAiBjD;AAED,wBAAgB,kBAAkB,CAAC,IAAI,GAAE,MAAgC,GAAG,cAAc,GAAG,IAAI,CAShG;AAED,wBAAgB,kBAAkB,CAChC,OAAO,EAAE,cAAc,EACvB,IAAI,GAAE,MAAgC,GACrC,IAAI,CAqCN;AAED;;;;GAIG;AACH,wBAAgB,wBAAwB,CACtC,GAAG,GAAE,MAAM,CAAC,UAAwB,EACpC,OAAO,GAAE,cAAc,GAAG,IAA2B,GACpD,IAAI,CAyBN;AAED,wBAAgB,sBAAsB,CACpC,IAAI,EAAE,WAAW,GAChB,IAAI,CAAC,cAAc,EAAE,UAAU,GAAG,OAAO,GAAG,SAAS,GAAG,WAAW,GAAG,MAAM,CAAC,CAuC/E"}
+{"version":3,"file":"local-ai-profile.d.ts","sourceRoot":"","sources":["../../src/llm/local-ai-profile.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAUH,MAAM,MAAM,WAAW,GAAG,MAAM,GAAG,OAAO,GAAG,OAAO,GAAG,OAAO,CAAC;AAE/D,MAAM,MAAM,eAAe,GAAG,QAAQ,GAAG,iBAAiB,GAAG,IAAI,CAAC;AAElE,MAAM,WAAW,cAAc;IAC7B,IAAI,EAAE,WAAW,CAAC;IAClB,QAAQ,EAAE,eAAe,CAAC;IAC1B,KAAK,EAAE,MAAM,GAAG,IAAI,CAAC;IACrB,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;IACvB,0CAA0C;IAC1C,eAAe,EAAE,MAAM,GAAG,IAAI,CAAC;IAC/B,8DAA8D;IAC9D,SAAS,EAAE,MAAM,GAAG,IAAI,CAAC;IACzB,SAAS,EAAE,MAAM,CAAC;IAClB,iDAAiD;IACjD,IAAI,CAAC,EAAE,MAAM,CAAC;CACf;AAED,wBAAgB,qBAAqB,IAAI,MAAM,CAK9C;AAED,gEAAgE;AAChE,wBAAgB,uBAAuB,IAAI,MAAM,CAEhD;AAED,6DAA6D;AAC7D,wBAAgB,gBAAgB,IAAI,cAAc,CAiBjD;AAED,wBAAgB,kBAAkB,CAAC,IAAI,GAAE,MAAgC,GAAG,cAAc,GAAG,IAAI,CAShG;AAED,wBAAgB,kBAAkB,CAChC,OAAO,EAAE,cAAc,EACvB,IAAI,GAAE,MAAgC,GACrC,IAAI,CAqCN;AAED;;;;GAIG;AACH,wBAAgB,wBAAwB,CACtC,GAAG,GAAE,MAAM,CAAC,UAAwB,EACpC,OAAO,GAAE,cAAc,GAAG,IAA2B,GACpD,IAAI,CAyBN;AAED,wBAAgB,sBAAsB,CACpC,IAAI,EAAE,WAAW,GAChB,IAAI,CAAC,cAAc,EAAE,UAAU,GAAG,OAAO,GAAG,SAAS,GAAG,WAAW,GAAG,MAAM,CAAC,CAuC/E"}

package/dist/llm/local-ai-profile.js

@@ -10,7 +10,7 @@
 import { existsSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs';
 import { homedir } from 'node:os';
 import { dirname, join } from 'node:path';
-import { DEFAULT_DAILY_OLLAMA_MODEL, DEFAULT_LOW_RAM_INFERENCE_SNAP, DEFAULT_US_ORIGIN_INFERENCE_SNAP, } from './providers/us-origin-snaps.js';
+import { DEFAULT_DAILY_OLLAMA_MODEL, DEFAULT_LOW_RAM_INFERENCE_SNAP, } from './providers/us-origin-snaps.js';
 export function getLocalAiProfilePath() {
     return (process.env.REVEALUI_INFERENCE_PROFILE_PATH ??
         join(homedir(), '.local', 'share', 'revealui', 'inference-profile.json'));
@@ -139,7 +139,7 @@ export function profileDefaultsForTier(tier) {
                 model: DEFAULT_DAILY_OLLAMA_MODEL,
                 baseURL: 'http://127.0.0.1:11434',
                 keepAlive: '0',
-                note: 'Ollama small US model; weights unload after each request',
+                note: 'Ollama daily default (qwen2.5:3b); weights unload after each request',
             };
         case 'snaps':
             return {
@@ -152,7 +152,7 @@ export function profileDefaultsForTier(tier) {
         case 'heavy':
             return {
                 provider: 'inference-snaps',
-                model: DEFAULT_US_ORIGIN_INFERENCE_SNAP,
+                model: 'nemotron-3-nano',
                 baseURL: null,
                 keepAlive: null,
                 note: 'Heavy snap — needs substantial RAM; avoid concurrent IDE load on 4GB WSL',

package/dist/llm/providers/groq.d.ts

@@ -10,7 +10,7 @@ export interface GroqProviderConfig extends Omit<LLMProviderConfig, 'apiKey'> {
     apiKey: string;
     /** Defaults to https://api.groq.com/openai/v1 */
     baseURL?: string;
-    /** Defaults to qwen/qwen3-32b */
+    /** Defaults to openai/gpt-oss-120b (current Groq-accepted catalog) */
     model?: string;
 }
 export declare class GroqProvider implements LLMProvider {

package/dist/llm/providers/groq.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"groq.d.ts","sourceRoot":"","sources":["../../../src/llm/providers/groq.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,OAAO,KAAK,EACV,SAAS,EACT,cAAc,EACd,QAAQ,EACR,eAAe,EACf,WAAW,EACX,iBAAiB,EACjB,WAAW,EACX,gBAAgB,EAChB,OAAO,EACP,oBAAoB,EACrB,MAAM,WAAW,CAAC;AAGnB,MAAM,WAAW,kBAAmB,SAAQ,IAAI,CAAC,iBAAiB,EAAE,QAAQ,CAAC;IAC3E,MAAM,EAAE,MAAM,CAAC;IACf,iDAAiD;IACjD,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,iCAAiC;IACjC,KAAK,CAAC,EAAE,MAAM,CAAC;CAChB;AAED,qBAAa,YAAa,YAAW,WAAW;IAC9C,OAAO,CAAC,KAAK,CAAuB;gBAExB,MAAM,EAAE,kBAAkB;IAQtC,YAAY,IAAI,oBAAoB;IAepC,IAAI,CAAC,QAAQ,EAAE,OAAO,EAAE,EAAE,OAAO,CAAC,EAAE,cAAc,GAAG,OAAO,CAAC,WAAW,CAAC;IAIzE,MAAM,CAAC,QAAQ,EAAE,OAAO,EAAE,EAAE,OAAO,CAAC,EAAE,gBAAgB,GAAG,aAAa,CAAC,QAAQ,CAAC;IAIhF,KAAK,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM,EAAE,EAAE,QAAQ,CAAC,EAAE,eAAe,GAAG,OAAO,CAAC,SAAS,GAAG,SAAS,EAAE,CAAC;CAG9F"}
+{"version":3,"file":"groq.d.ts","sourceRoot":"","sources":["../../../src/llm/providers/groq.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAGH,OAAO,KAAK,EACV,SAAS,EACT,cAAc,EACd,QAAQ,EACR,eAAe,EACf,WAAW,EACX,iBAAiB,EACjB,WAAW,EACX,gBAAgB,EAChB,OAAO,EACP,oBAAoB,EACrB,MAAM,WAAW,CAAC;AAGnB,MAAM,WAAW,kBAAmB,SAAQ,IAAI,CAAC,iBAAiB,EAAE,QAAQ,CAAC;IAC3E,MAAM,EAAE,MAAM,CAAC;IACf,iDAAiD;IACjD,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,sEAAsE;IACtE,KAAK,CAAC,EAAE,MAAM,CAAC;CAChB;AAED,qBAAa,YAAa,YAAW,WAAW;IAC9C,OAAO,CAAC,KAAK,CAAuB;gBAExB,MAAM,EAAE,kBAAkB;IAUtC,YAAY,IAAI,oBAAoB;IAepC,IAAI,CAAC,QAAQ,EAAE,OAAO,EAAE,EAAE,OAAO,CAAC,EAAE,cAAc,GAAG,OAAO,CAAC,WAAW,CAAC;IAIzE,MAAM,CAAC,QAAQ,EAAE,OAAO,EAAE,EAAE,OAAO,CAAC,EAAE,gBAAgB,GAAG,aAAa,CAAC,QAAQ,CAAC;IAIhF,KAAK,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM,EAAE,EAAE,QAAQ,CAAC,EAAE,eAAe,GAAG,OAAO,CAAC,SAAS,GAAG,SAAS,EAAE,CAAC;CAG9F"}

package/dist/llm/providers/groq.js

@@ -5,18 +5,21 @@
  * Free tier: 6,000 TPM / 500k TPD.
  * Sign up: console.groq.com
  */
+import { GROQ_DEFAULT_BASE_URL, GROQ_DEFAULT_MODEL } from '../inference-route.js';
 import { OpenAICompatProvider } from './openai-compat.js';
 export class GroqProvider {
     inner;
     constructor(config) {
         this.inner = new OpenAICompatProvider({
             ...config,
-            baseURL: config.baseURL ?? 'https://api.groq.com/openai/v1',
-            model: config.model ?? 'qwen/qwen3-32b',
+            baseURL: config.baseURL ?? GROQ_DEFAULT_BASE_URL,
+            // Retired catalog ids (llama-3.3-70b-versatile, qwen/qwen3-32b) 404 on
+            // Groq's public API as of 2026-08-16. Prefer the current Groq default.
+            model: config.model ?? GROQ_DEFAULT_MODEL,
         });
     }
     capabilities() {
-        // Profile for the default model (qwen/qwen3-32b). Groq exposes no embeddings endpoint.
+        // Profile for the default model. Groq exposes no embeddings endpoint.
         return {
             providerTag: 'groq',
             tools: true,

package/dist/llm/providers/inference-snaps.d.ts

@@ -6,17 +6,18 @@
  *
  * Product hardline: US-origin snap models only (see us-origin-snaps.ts).
  * Allowlisted snaps (install these):
- *   nemotron-3-nano        -  NVIDIA general + tools (default)
+ *   gemma3                 -  Google general + vision (default; WSL-supported)
+ *   gemma4                 -  Google general + vision + tools
+ *   nemotron-3-nano        -  NVIDIA general + tools (heavy / capable hosts)
  *   nemotron-3-nano-omni   -  NVIDIA multimodal
- *   gemma3 / gemma4        -  Google general + vision + tools
  *
  * Canonical also publishes non-US snaps (deepseek-r1, qwen-*, glm-*). Those
  * are rejected at construction unless REVEALUI_ALLOW_NON_US_MODELS=1.
  *
  * Install a model:
- *   sudo snap install nemotron-3-nano
- *   nemotron-3-nano set http.port=9090   # optional: change port
- *   nemotron-3-nano status              # shows base URL and available models
+ *   sudo snap install gemma3
+ *   gemma3 set http.port=9090 --assume-yes
+ *   gemma3 status              # shows base URL and available models
  *
  * Set env vars:
  *   INFERENCE_SNAPS_BASE_URL=http://localhost:9090/v1

package/dist/llm/providers/inference-snaps.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"inference-snaps.d.ts","sourceRoot":"","sources":["../../../src/llm/providers/inference-snaps.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;GA0BG;AAEH,OAAO,KAAK,EACV,SAAS,EACT,cAAc,EACd,QAAQ,EACR,eAAe,EACf,WAAW,EACX,iBAAiB,EACjB,WAAW,EACX,gBAAgB,EAChB,OAAO,EACP,oBAAoB,EACrB,MAAM,WAAW,CAAC;AAEnB,OAAO,EAEL,gCAAgC,EACjC,MAAM,sBAAsB,CAAC;AAE9B,MAAM,WAAW,4BAA6B,SAAQ,IAAI,CAAC,iBAAiB,EAAE,QAAQ,CAAC;IACrF,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,6EAA6E;IAC7E,OAAO,EAAE,MAAM,CAAC;IAChB;;;OAGG;IACH,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,qEAAqE;IACrE,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB;;;OAGG;IACH,gBAAgB,CAAC,EAAE,OAAO,CAAC;CAC5B;AAED,qBAAa,sBAAuB,YAAW,WAAW;IACxD,OAAO,CAAC,KAAK,CAAuB;IACpC,OAAO,CAAC,UAAU,CAAS;IAC3B,OAAO,CAAC,OAAO,CAAS;gBAEZ,MAAM,EAAE,4BAA4B;IAchD,YAAY,IAAI,oBAAoB;IAgBpC,IAAI,CAAC,QAAQ,EAAE,OAAO,EAAE,EAAE,OAAO,CAAC,EAAE,cAAc,GAAG,OAAO,CAAC,WAAW,CAAC;IAIzE,MAAM,CAAC,QAAQ,EAAE,OAAO,EAAE,EAAE,OAAO,CAAC,EAAE,gBAAgB,GAAG,aAAa,CAAC,QAAQ,CAAC;IAI1E,KAAK,CACT,IAAI,EAAE,MAAM,GAAG,MAAM,EAAE,EACvB,QAAQ,CAAC,EAAE,eAAe,GACzB,OAAO,CAAC,SAAS,GAAG,SAAS,EAAE,CAAC;CAsBpC;AAED,OAAO,EAAE,gCAAgC,EAAE,CAAC"}
+{"version":3,"file":"inference-snaps.d.ts","sourceRoot":"","sources":["../../../src/llm/providers/inference-snaps.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;GA2BG;AAEH,OAAO,KAAK,EACV,SAAS,EACT,cAAc,EACd,QAAQ,EACR,eAAe,EACf,WAAW,EACX,iBAAiB,EACjB,WAAW,EACX,gBAAgB,EAChB,OAAO,EACP,oBAAoB,EACrB,MAAM,WAAW,CAAC;AAEnB,OAAO,EAEL,gCAAgC,EACjC,MAAM,sBAAsB,CAAC;AAE9B,MAAM,WAAW,4BAA6B,SAAQ,IAAI,CAAC,iBAAiB,EAAE,QAAQ,CAAC;IACrF,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,6EAA6E;IAC7E,OAAO,EAAE,MAAM,CAAC;IAChB;;;OAGG;IACH,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,qEAAqE;IACrE,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB;;;OAGG;IACH,gBAAgB,CAAC,EAAE,OAAO,CAAC;CAC5B;AAED,qBAAa,sBAAuB,YAAW,WAAW;IACxD,OAAO,CAAC,KAAK,CAAuB;IACpC,OAAO,CAAC,UAAU,CAAS;IAC3B,OAAO,CAAC,OAAO,CAAS;gBAEZ,MAAM,EAAE,4BAA4B;IAchD,YAAY,IAAI,oBAAoB;IAgBpC,IAAI,CAAC,QAAQ,EAAE,OAAO,EAAE,EAAE,OAAO,CAAC,EAAE,cAAc,GAAG,OAAO,CAAC,WAAW,CAAC;IAIzE,MAAM,CAAC,QAAQ,EAAE,OAAO,EAAE,EAAE,OAAO,CAAC,EAAE,gBAAgB,GAAG,aAAa,CAAC,QAAQ,CAAC;IAI1E,KAAK,CACT,IAAI,EAAE,MAAM,GAAG,MAAM,EAAE,EACvB,QAAQ,CAAC,EAAE,eAAe,GACzB,OAAO,CAAC,SAAS,GAAG,SAAS,EAAE,CAAC;CAsBpC;AAED,OAAO,EAAE,gCAAgC,EAAE,CAAC"}