@revealui/ai 0.6.4 → 0.7.0

package/README.md

@@ -16,7 +16,7 @@ AI system for RevealUI - memory, LLM, orchestration, and tools.
 ## Features
 
 - **Memory System**: CRDT-based persistent memory (Working, Episodic, Semantic)
-- **LLM Integration**: Provider abstractions for Anthropic, GROQ, Ollama, Canonical Inference Snaps, and more
+- **LLM Integration**: Provider abstractions for Ollama and Canonical Inference Snaps (open-model defaults), plus Groq, Anthropic, OpenAI, and HuggingFace as bring-your-own-key cloud adapters. Every adapter calls the vendor's OpenAI-compatible HTTP endpoint with your own key. No proprietary Anthropic or OpenAI SDK.
 - **Agent Orchestration**: Runtime and execution engine for AI agents
 - **Tool Calling**: Tool registry + standard-MCP-client integration (Stage 5.1a)
 - **Vector Search**: Semantic search with pgvector
@@ -64,8 +64,9 @@ const provider = new InferenceSnapsProvider({
 const client = new LLMClient({ provider })
 ```
 
-Cloud providers (Anthropic, OpenAI, GROQ) remain supported; the local
-inference path is the documented default for self-hosted deployments.
+Groq, Anthropic, OpenAI, and HuggingFace remain supported as pluggable,
+bring-your-own-key cloud providers. The local inference path is the
+documented default for self-hosted deployments.
 
 ## MCP tool integration
 
@@ -313,7 +314,7 @@ const state = await persistence.loadCRDTState(crdtId, 'lww_register')
 
 ### LLM Integration
 
-Provider abstractions and unified client for Anthropic, GROQ, and Ollama.
+Provider abstractions and unified client for Ollama, Canonical Inference Snaps, Groq, Anthropic, OpenAI, and HuggingFace.
 
 ```typescript
 import { LLMClient, createLLMClientFromEnv } from '@revealui/ai/llm/client'

package/dist/embeddings/index.d.ts

@@ -7,6 +7,7 @@
  * Note: Groq and Anthropic do not support embeddings  -  use Ollama for embedding tasks.
  */
 import z from 'zod/v4';
+import { type LLMClient } from '../llm/client.js';
 declare const EmbeddingSchema: z.ZodObject<{
     model: z.ZodString;
     vector: z.ZodArray<z.ZodNumber>;
@@ -23,6 +24,13 @@ export interface GenerateEmbeddingOptions {
      */
     model?: string;
     cache?: boolean;
+    /**
+     * Pre-resolved LLM client (GAP-360). When supplied — by an authenticated
+     * caller that ran `resolveLLMClientForRequest` — embeddings use it (per-user
+     * BYOK / hosted resolution). When omitted, embeddings fall back to the
+     * env-configured client, preserving self-hosted and internal-caller behavior.
+     */
+    client?: Pick<LLMClient, 'embed'>;
 }
 /**
  * Generate an embedding for the given text using the configured LLM provider.

package/dist/embeddings/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/embeddings/index.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,OAAO,CAAC,MAAM,QAAQ,CAAC;AAGvB,QAAA,MAAM,eAAe;;;;;iBAgBjB,CAAC;AACL,KAAK,SAAS,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,eAAe,CAAC,CAAC;AAEjD,MAAM,WAAW,wBAAwB;IACvC;;;;;OAKG;IACH,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,KAAK,CAAC,EAAE,OAAO,CAAC;CACjB;AAED;;;;;;;;;;;;;GAaG;AACH,wBAAsB,iBAAiB,CACrC,IAAI,EAAE,MAAM,EACZ,OAAO,GAAE,wBAA6B,GACrC,OAAO,CAAC,SAAS,CAAC,CA4BpB;AAED;;;;;;GAMG;AACH,wBAAsB,kBAAkB,CACtC,KAAK,EAAE,MAAM,EAAE,EACf,OAAO,GAAE,wBAA6B,GACrC,OAAO,CAAC,SAAS,EAAE,CAAC,CAMtB"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/embeddings/index.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,OAAO,CAAC,MAAM,QAAQ,CAAC;AACvB,OAAO,EAA0B,KAAK,SAAS,EAAE,MAAM,kBAAkB,CAAC;AAE1E,QAAA,MAAM,eAAe;;;;;iBAgBjB,CAAC;AACL,KAAK,SAAS,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,eAAe,CAAC,CAAC;AAEjD,MAAM,WAAW,wBAAwB;IACvC;;;;;OAKG;IACH,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,KAAK,CAAC,EAAE,OAAO,CAAC;IAChB;;;;;OAKG;IACH,MAAM,CAAC,EAAE,IAAI,CAAC,SAAS,EAAE,OAAO,CAAC,CAAC;CACnC;AAED;;;;;;;;;;;;;GAaG;AACH,wBAAsB,iBAAiB,CACrC,IAAI,EAAE,MAAM,EACZ,OAAO,GAAE,wBAA6B,GACrC,OAAO,CAAC,SAAS,CAAC,CA6BpB;AAED;;;;;;GAMG;AACH,wBAAsB,kBAAkB,CACtC,KAAK,EAAE,MAAM,EAAE,EACf,OAAO,GAAE,wBAA6B,GACrC,OAAO,CAAC,SAAS,EAAE,CAAC,CAMtB"}

package/dist/embeddings/index.js

@@ -41,8 +41,9 @@ export async function generateEmbedding(text, options = {}) {
     if (!text || typeof text !== 'string' || text.trim().length === 0) {
         throw new Error('Text must be a non-empty string');
     }
-    // Use unified LLM client  -  auto-detects provider from env vars
-    const client = createLLMClientFromEnv();
+    // Prefer the caller-resolved client (BYOK / hosted). Fall back to the
+    // env-configured client for self-hosted and internal (no-user) callers.
+    const client = options.client ?? createLLMClientFromEnv();
     // Ask client to embed  -  each provider uses its own default model when model is undefined
     const result = await client.embed(text, model ? { model } : undefined);
     // If provider returned batch, pick first

package/dist/index.d.ts

@@ -54,6 +54,7 @@ export * from './llm/provider-health.js';
 export * from './llm/providers/base.js';
 export * from './llm/providers/inference-snaps.js';
 export * from './llm/providers/openai-compat.js';
+export * from './llm/resolve.js';
 export * from './llm/token-counter.js';
 export * from './llm/workspace-provider-config.js';
 export * from './memory/index.js';

package/dist/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8CG;AAGH,cAAc,gBAAgB,CAAC;AAE/B,cAAc,kBAAkB,CAAC;AAIjC,cAAc,uBAAuB,CAAC;AAEtC,cAAc,sBAAsB,CAAC;AAErC,cAAc,iBAAiB,CAAC;AAChC,cAAc,0BAA0B,CAAC;AACzC,cAAc,yBAAyB,CAAC;AACxC,cAAc,oCAAoC,CAAC;AACnD,cAAc,kCAAkC,CAAC;AACjD,cAAc,wBAAwB,CAAC;AACvC,cAAc,oCAAoC,CAAC;AAEnD,cAAc,mBAAmB,CAAC;AAElC,cAAc,0BAA0B,CAAC;AACzC,cAAc,6BAA6B,CAAC;AAC5C,cAAc,uCAAuC,CAAC;AACtD,cAAc,iCAAiC,CAAC;AAChD,cAAc,4BAA4B,CAAC;AAC3C,cAAc,sCAAsC,CAAC;AACrD,cAAc,iCAAiC,CAAC;AAEhD,cAAc,mBAAmB,CAAC;AAElC,cAAc,sBAAsB,CAAC;AAErC,cAAc,iBAAiB,CAAC;AAChC,cAAc,yBAAyB,CAAC;AACxC,cAAc,gCAAgC,CAAC;AAC/C,cAAc,wBAAwB,CAAC;AACvC,cAAc,4BAA4B,CAAC;AAC3C,cAAc,uBAAuB,CAAC;AACtC,cAAc,yBAAyB,CAAC;AACxC,cAAc,yBAAyB,CAAC;AACxC,cAAc,qBAAqB,CAAC;AACpC,cAAc,yBAAyB,CAAC;AACxC,cAAc,sBAAsB,CAAC"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8CG;AAGH,cAAc,gBAAgB,CAAC;AAE/B,cAAc,kBAAkB,CAAC;AAIjC,cAAc,uBAAuB,CAAC;AAEtC,cAAc,sBAAsB,CAAC;AAErC,cAAc,iBAAiB,CAAC;AAChC,cAAc,0BAA0B,CAAC;AACzC,cAAc,yBAAyB,CAAC;AACxC,cAAc,oCAAoC,CAAC;AACnD,cAAc,kCAAkC,CAAC;AACjD,cAAc,kBAAkB,CAAC;AACjC,cAAc,wBAAwB,CAAC;AACvC,cAAc,oCAAoC,CAAC;AAEnD,cAAc,mBAAmB,CAAC;AAElC,cAAc,0BAA0B,CAAC;AACzC,cAAc,6BAA6B,CAAC;AAC5C,cAAc,uCAAuC,CAAC;AACtD,cAAc,iCAAiC,CAAC;AAChD,cAAc,4BAA4B,CAAC;AAC3C,cAAc,sCAAsC,CAAC;AACrD,cAAc,iCAAiC,CAAC;AAEhD,cAAc,mBAAmB,CAAC;AAElC,cAAc,sBAAsB,CAAC;AAErC,cAAc,iBAAiB,CAAC;AAChC,cAAc,yBAAyB,CAAC;AACxC,cAAc,gCAAgC,CAAC;AAC/C,cAAc,wBAAwB,CAAC;AACvC,cAAc,4BAA4B,CAAC;AAC3C,cAAc,uBAAuB,CAAC;AACtC,cAAc,yBAAyB,CAAC;AACxC,cAAc,yBAAyB,CAAC;AACxC,cAAc,qBAAqB,CAAC;AACpC,cAAc,yBAAyB,CAAC;AACxC,cAAc,sBAAsB,CAAC"}

package/dist/index.js

@@ -61,6 +61,7 @@ export * from './llm/provider-health.js';
 export * from './llm/providers/base.js';
 export * from './llm/providers/inference-snaps.js';
 export * from './llm/providers/openai-compat.js';
+export * from './llm/resolve.js';
 export * from './llm/token-counter.js';
 export * from './llm/workspace-provider-config.js';
 // Re-export memory system

package/dist/llm/client.d.ts

@@ -10,7 +10,15 @@ import type { ProviderHealthMonitor } from './provider-health.js';
 import type { Embedding, LLMChatOptions, LLMChunk, LLMEmbedOptions, LLMProvider, LLMResponse, LLMStreamOptions, Message } from './providers/base.js';
 import { type CacheStats, type ResponseCacheOptions } from './response-cache.js';
 import { type SemanticCacheOptions, type SemanticCacheStats } from './semantic-cache.js';
-export type LLMProviderType = 'groq' | 'ollama' | 'huggingface' | 'inference-snaps';
+export type LLMProviderType = 'anthropic' | 'openai' | 'groq' | 'huggingface' | 'ollama' | 'inference-snaps';
+/**
+ * Providers reachable from a hosted (serverless) deployment. Localhost-only
+ * providers are false. Consumed by PR-2/PR-3 (resolver + settings UI) to pick
+ * which providers a hosted account may configure; no runtime consumer yet.
+ */
+export declare const hostedViable: Record<LLMProviderType, boolean>;
+/** True when the provider can serve a hosted (serverless) deployment. */
+export declare function isHostedViable(provider: LLMProviderType): boolean;
 export interface LLMClientConfig {
     provider: LLMProviderType;
     apiKey: string;
@@ -129,20 +137,24 @@ export declare class LLMClient {
  * Create an LLM client from environment variables.
  *
  * When LLM_PROVIDER is not set, auto-detects the provider by checking env vars
- * in priority order: INFERENCE_SNAPS → GROQ → OLLAMA. If none are set, defaults
- * to Inference Snaps at http://localhost:9090/v1 (Ubuntu local) and emits a
- * one-line stderr warning so the implicit default is discoverable in logs.
+ * in priority order: INFERENCE_SNAPS → GROQ → OLLAMA → ANTHROPIC_API_KEY →
+ * OPENAI_API_KEY. If none are set, defaults to Inference Snaps at
+ * http://localhost:9090/v1 (Ubuntu local) and emits a one-line stderr warning so
+ * the implicit localhost default is discoverable in logs.
  *
- * All providers use OpenAI-compatible APIs. No proprietary provider SDKs.
+ * All providers use OpenAI-compatible APIs. No proprietary provider SDKs
+ * (Anthropic + OpenAI ride their OpenAI-compatible endpoints).
  *
  * Canonical Inference Snaps is the reference local provider on Ubuntu — offline,
  * silicon-optimized, no API key required. See `providers/inference-snaps.ts` for
  * install docs (`sudo snap install gemma3`, etc.).
  *
  * Provider defaults:
- *   inference-snaps → gemma3       (base URL defaults to http://localhost:9090/v1)
+ *   inference-snaps → gemma3            (base URL defaults to http://localhost:9090/v1)
  *   groq            → qwen/qwen3-32b
- *   ollama          → gemma4:e2b   (base URL defaults to http://localhost:11434)
+ *   ollama          → gemma4:e2b        (base URL defaults to 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)
  */
 export declare function createLLMClientFromEnv(): LLMClient;
 /**
@@ -155,8 +167,16 @@ export declare function createLLMClientFromEnv(): LLMClient;
  * Returns `null` if the user has no stored keys (callers should fall back
  * to `createLLMClientFromEnv()` or return a 402/feature-unavailable error).
  *
+ * When `opts.hostedViableOnly` is set, a stored key whose provider is not
+ * hosted-viable (ollama / inference-snaps are localhost-only) resolves to
+ * `null` instead of a client. The resolver passes this on hosted deployments so
+ * a localhost-only BYOK key can never yield a localhost client — the exact
+ * silent-localhost defect GAP-360 closes (spec §6.5, fail-closed).
+ *
  * @param userId - The user's ID from the `users` table
  * @param db - A Drizzle NeonDB client instance
  */
-export declare function createLLMClientForUser(userId: string, db: Database, auditStore?: AuditStore): Promise<LLMClient | null>;
+export declare function createLLMClientForUser(userId: string, db: Database, auditStore?: Pick<AuditStore, 'append'>, opts?: {
+    hostedViableOnly?: boolean;
+}): Promise<LLMClient | null>;
 //# sourceMappingURL=client.d.ts.map

package/dist/llm/client.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"client.d.ts","sourceRoot":"","sources":["../../src/llm/client.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAKH,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,qBAAqB,CAAC;AAGpD,OAAO,EACL,cAAc,EAGf,MAAM,sBAAsB,CAAC;AAE9B,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,mBAAmB,CAAC;AACpD,OAAO,KAAK,EAAE,qBAAqB,EAAE,MAAM,sBAAsB,CAAC;AAClE,OAAO,KAAK,EACV,SAAS,EACT,cAAc,EACd,QAAQ,EACR,eAAe,EACf,WAAW,EACX,WAAW,EACX,gBAAgB,EAChB,OAAO,EACR,MAAM,qBAAqB,CAAC;AAQ7B,OAAO,EAAE,KAAK,UAAU,EAAiB,KAAK,oBAAoB,EAAE,MAAM,qBAAqB,CAAC;AAChG,OAAO,EAEL,KAAK,oBAAoB,EACzB,KAAK,kBAAkB,EACxB,MAAM,qBAAqB,CAAC;AAG7B,MAAM,MAAM,eAAe,GAAG,MAAM,GAAG,QAAQ,GAAG,aAAa,GAAG,iBAAiB,CAAC;AAEpF,MAAM,WAAW,eAAe;IAC9B,QAAQ,EAAE,eAAe,CAAC;IAC1B,MAAM,EAAE,MAAM,CAAC;IACf;;;;OAIG;IACH,QAAQ,CAAC,EAAE,MAAM,OAAO,CAAC,MAAM,CAAC,CAAC;IACjC,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf;;;;OAIG;IACH,aAAa,CAAC,EAAE,WAAW,CAAC;IAC5B,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,gBAAgB,CAAC,EAAE,eAAe,CAAC;IACnC,SAAS,CAAC,EAAE;QACV,iBAAiB,CAAC,EAAE,MAAM,CAAC;QAC3B,cAAc,CAAC,EAAE,MAAM,CAAC;KACzB,CAAC;IACF,oFAAoF;IACpF,oBAAoB,CAAC,EAAE,OAAO,CAAC;IAC/B,wEAAwE;IACxE,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B,6BAA6B;IAC7B,oBAAoB,CAAC,EAAE,oBAAoB,CAAC;IAC5C,iEAAiE;IACjE,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B,6BAA6B;IAC7B,oBAAoB,CAAC,EAAE,oBAAoB,CAAC;IAC5C,4EAA4E;IAC5E,aAAa,CAAC,EAAE,qBAAqB,CAAC;IACtC,qEAAqE;IACrE,8BAA8B,CAAC,EAAE,MAAM,CAAC;IACxC,kFAAkF;IAClF,0BAA0B,CAAC,EAAE,MAAM,CAAC;CACrC;AAQD,qBAAa,SAAS;IACpB,OAAO,CAAC,QAAQ,CAAc;IAC9B,OAAO,CAAC,gBAAgB,CAAC,CAAc;IACvC,OAAO,CAAC,qBAAqB,CAAC,CAAc;IAC5C,OAAO,CAAC,MAAM,CAAkB;IAChC,OAAO,CAAC,cAAc,CAAiB;IACvC,OAAO,CAAC,aAAa,CAAC,CAAgB;IACtC,OAAO,CAAC,aAAa,CAAC,CAAgB;IACtC,OAAO,CAAC,aAAa,CAAC,CAAwB;IAC9C,OAAO,CAAC,cAAc,CAAiB;IACvC,OAAO,CAAC,sBAAsB,CAAC,CAAiB;IAChD,wFAAwF;IACxF,OAAO,CAAC,aAAa,CAAS;gBAElB,MAAM,EAAE,eAAe;IA0DnC,OAAO,CAAC,cAAc;IAoBtB;;;OAGG;YACW,uBAAuB;IAmBrC,OAAO,CAAC,cAAc;IAoCtB,OAAO,CAAC,aAAa;IAMf,IAAI,CAAC,QAAQ,EAAE,OAAO,EAAE,EAAE,OAAO,CAAC,EAAE,cAAc,GAAG,OAAO,CAAC,WAAW,CAAC;IAwHzE,KAAK,CACT,IAAI,EAAE,MAAM,GAAG,MAAM,EAAE,EACvB,OAAO,CAAC,EAAE,eAAe,GACxB,OAAO,CAAC,SAAS,GAAG,SAAS,EAAE,CAAC;IAsD5B,MAAM,CAAC,QAAQ,EAAE,OAAO,EAAE,EAAE,OAAO,CAAC,EAAE,gBAAgB,GAAG,aAAa,CAAC,QAAQ,CAAC;IA4EvF;;;OAGG;IACH,eAAe,CAAC,QAAQ,EAAE,OAAO,EAAE,GAAG;QAAE,MAAM,EAAE,MAAM,CAAC;QAAC,gBAAgB,EAAE,MAAM,CAAA;KAAE;IAIlF;;OAEG;IACH,sBAAsB,IAAI;QACxB,OAAO,EAAE;YAAE,IAAI,EAAE,MAAM,CAAC;YAAC,KAAK,EAAE,MAAM,CAAC;YAAC,KAAK,EAAE,UAAU,CAAC,cAAc,CAAC,UAAU,CAAC,CAAC,CAAA;SAAE,CAAC;QACxF,QAAQ,CAAC,EAAE;YAAE,IAAI,EAAE,MAAM,CAAC;YAAC,KAAK,EAAE,MAAM,CAAC;YAAC,KAAK,EAAE,UAAU,CAAC,cAAc,CAAC,UAAU,CAAC,CAAC,CAAA;SAAE,CAAC;KAC3F;IAkBD;;OAEG;IACH,gBAAgB,IAAI,qBAAqB,GAAG,SAAS;IAIrD;;;;OAIG;IACH,qBAAqB,IAAI,UAAU,GAAG,SAAS;IAI/C;;OAEG;IACH,kBAAkB,IAAI,IAAI;IAI1B;;;;OAIG;IACH,qBAAqB,IAAI,kBAAkB,GAAG,SAAS;IAIvD;;OAEG;IACH,kBAAkB,IAAI,IAAI;CAG3B;AAED;;;;;;;;;;;;;;;;;;GAkBG;AACH,wBAAgB,sBAAsB,IAAI,SAAS,CAiElD;AAED;;;;;;;;;;;;GAYG;AACH,wBAAsB,sBAAsB,CAC1C,MAAM,EAAE,MAAM,EACd,EAAE,EAAE,QAAQ,EACZ,UAAU,CAAC,EAAE,UAAU,GACtB,OAAO,CAAC,SAAS,GAAG,IAAI,CAAC,CAiD3B"}
+{"version":3,"file":"client.d.ts","sourceRoot":"","sources":["../../src/llm/client.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAMH,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,qBAAqB,CAAC;AAGpD,OAAO,EACL,cAAc,EAGf,MAAM,sBAAsB,CAAC;AAE9B,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,mBAAmB,CAAC;AACpD,OAAO,KAAK,EAAE,qBAAqB,EAAE,MAAM,sBAAsB,CAAC;AAElE,OAAO,KAAK,EACV,SAAS,EACT,cAAc,EACd,QAAQ,EACR,eAAe,EACf,WAAW,EACX,WAAW,EACX,gBAAgB,EAChB,OAAO,EACR,MAAM,qBAAqB,CAAC;AAS7B,OAAO,EAAE,KAAK,UAAU,EAAiB,KAAK,oBAAoB,EAAE,MAAM,qBAAqB,CAAC;AAChG,OAAO,EAEL,KAAK,oBAAoB,EACzB,KAAK,kBAAkB,EACxB,MAAM,qBAAqB,CAAC;AAG7B,MAAM,MAAM,eAAe,GACvB,WAAW,GACX,QAAQ,GACR,MAAM,GACN,aAAa,GACb,QAAQ,GACR,iBAAiB,CAAC;AAEtB;;;;GAIG;AACH,eAAO,MAAM,YAAY,EAAE,MAAM,CAAC,eAAe,EAAE,OAAO,CAOzD,CAAC;AAEF,yEAAyE;AACzE,wBAAgB,cAAc,CAAC,QAAQ,EAAE,eAAe,GAAG,OAAO,CAEjE;AAMD,MAAM,WAAW,eAAe;IAC9B,QAAQ,EAAE,eAAe,CAAC;IAC1B,MAAM,EAAE,MAAM,CAAC;IACf;;;;OAIG;IACH,QAAQ,CAAC,EAAE,MAAM,OAAO,CAAC,MAAM,CAAC,CAAC;IACjC,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf;;;;OAIG;IACH,aAAa,CAAC,EAAE,WAAW,CAAC;IAC5B,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,gBAAgB,CAAC,EAAE,eAAe,CAAC;IACnC,SAAS,CAAC,EAAE;QACV,iBAAiB,CAAC,EAAE,MAAM,CAAC;QAC3B,cAAc,CAAC,EAAE,MAAM,CAAC;KACzB,CAAC;IACF,oFAAoF;IACpF,oBAAoB,CAAC,EAAE,OAAO,CAAC;IAC/B,wEAAwE;IACxE,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B,6BAA6B;IAC7B,oBAAoB,CAAC,EAAE,oBAAoB,CAAC;IAC5C,iEAAiE;IACjE,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B,6BAA6B;IAC7B,oBAAoB,CAAC,EAAE,oBAAoB,CAAC;IAC5C,4EAA4E;IAC5E,aAAa,CAAC,EAAE,qBAAqB,CAAC;IACtC,qEAAqE;IACrE,8BAA8B,CAAC,EAAE,MAAM,CAAC;IACxC,kFAAkF;IAClF,0BAA0B,CAAC,EAAE,MAAM,CAAC;CACrC;AAQD,qBAAa,SAAS;IACpB,OAAO,CAAC,QAAQ,CAAc;IAC9B,OAAO,CAAC,gBAAgB,CAAC,CAAc;IACvC,OAAO,CAAC,qBAAqB,CAAC,CAAc;IAC5C,OAAO,CAAC,MAAM,CAAkB;IAChC,OAAO,CAAC,cAAc,CAAiB;IACvC,OAAO,CAAC,aAAa,CAAC,CAAgB;IACtC,OAAO,CAAC,aAAa,CAAC,CAAgB;IACtC,OAAO,CAAC,aAAa,CAAC,CAAwB;IAC9C,OAAO,CAAC,cAAc,CAAiB;IACvC,OAAO,CAAC,sBAAsB,CAAC,CAAiB;IAChD,wFAAwF;IACxF,OAAO,CAAC,aAAa,CAAS;gBAElB,MAAM,EAAE,eAAe;IA0DnC,OAAO,CAAC,cAAc;IAgCtB;;;OAGG;YACW,uBAAuB;IAmBrC,OAAO,CAAC,cAAc;IAoCtB,OAAO,CAAC,aAAa;IAMf,IAAI,CAAC,QAAQ,EAAE,OAAO,EAAE,EAAE,OAAO,CAAC,EAAE,cAAc,GAAG,OAAO,CAAC,WAAW,CAAC;IAwHzE,KAAK,CACT,IAAI,EAAE,MAAM,GAAG,MAAM,EAAE,EACvB,OAAO,CAAC,EAAE,eAAe,GACxB,OAAO,CAAC,SAAS,GAAG,SAAS,EAAE,CAAC;IAsD5B,MAAM,CAAC,QAAQ,EAAE,OAAO,EAAE,EAAE,OAAO,CAAC,EAAE,gBAAgB,GAAG,aAAa,CAAC,QAAQ,CAAC;IA4EvF;;;OAGG;IACH,eAAe,CAAC,QAAQ,EAAE,OAAO,EAAE,GAAG;QAAE,MAAM,EAAE,MAAM,CAAC;QAAC,gBAAgB,EAAE,MAAM,CAAA;KAAE;IAIlF;;OAEG;IACH,sBAAsB,IAAI;QACxB,OAAO,EAAE;YAAE,IAAI,EAAE,MAAM,CAAC;YAAC,KAAK,EAAE,MAAM,CAAC;YAAC,KAAK,EAAE,UAAU,CAAC,cAAc,CAAC,UAAU,CAAC,CAAC,CAAA;SAAE,CAAC;QACxF,QAAQ,CAAC,EAAE;YAAE,IAAI,EAAE,MAAM,CAAC;YAAC,KAAK,EAAE,MAAM,CAAC;YAAC,KAAK,EAAE,UAAU,CAAC,cAAc,CAAC,UAAU,CAAC,CAAC,CAAA;SAAE,CAAC;KAC3F;IAkBD;;OAEG;IACH,gBAAgB,IAAI,qBAAqB,GAAG,SAAS;IAIrD;;;;OAIG;IACH,qBAAqB,IAAI,UAAU,GAAG,SAAS;IAI/C;;OAEG;IACH,kBAAkB,IAAI,IAAI;IAI1B;;;;OAIG;IACH,qBAAqB,IAAI,kBAAkB,GAAG,SAAS;IAIvD;;OAEG;IACH,kBAAkB,IAAI,IAAI;CAG3B;AAED;;;;;;;;;;;;;;;;;;;;;;GAsBG;AACH,wBAAgB,sBAAsB,IAAI,SAAS,CAyFlD;AAED;;;;;;;;;;;;;;;;;;GAkBG;AACH,wBAAsB,sBAAsB,CAC1C,MAAM,EAAE,MAAM,EACd,EAAE,EAAE,QAAQ,EAGZ,UAAU,CAAC,EAAE,IAAI,CAAC,UAAU,EAAE,QAAQ,CAAC,EACvC,IAAI,CAAC,EAAE;IAAE,gBAAgB,CAAC,EAAE,OAAO,CAAA;CAAE,GACpC,OAAO,CAAC,SAAS,GAAG,IAAI,CAAC,CAsD3B"}

package/dist/llm/client.js

@@ -3,16 +3,42 @@
  *
  * Single interface for all LLM providers with fallback and rate limiting
  */
+// Log redaction lives in @revealui/security — import `redactLogContext`
+// (recursive walker) or `redactLogField` (single key/value).
+import { createLogger } from '@revealui/core/observability/logger';
 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 { AnthropicProvider } from './providers/anthropic.js';
 import { GroqProvider } from './providers/groq.js';
 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 { ResponseCache } from './response-cache.js';
 import { SemanticCache, } from './semantic-cache.js';
 import { estimateRequest as _estimateRequestTokens } from './token-counter.js';
+/**
+ * Providers reachable from a hosted (serverless) deployment. Localhost-only
+ * providers are false. Consumed by PR-2/PR-3 (resolver + settings UI) to pick
+ * which providers a hosted account may configure; no runtime consumer yet.
+ */
+export const hostedViable = {
+    anthropic: true,
+    openai: true,
+    groq: true,
+    huggingface: true,
+    ollama: false,
+    'inference-snaps': false,
+};
+/** True when the provider can serve a hosted (serverless) deployment. */
+export function isHostedViable(provider) {
+    return hostedViable[provider];
+}
+/** Emitted once per process when the zero-config localhost default is selected. */
+let warnedLocalhostDefault = false;
+const envFactoryLogger = createLogger({ component: 'createLLMClientFromEnv' });
 export class LLMClient {
     provider;
     fallbackProvider;
@@ -78,12 +104,22 @@ export class LLMClient {
     }
     createProvider(type, config) {
         switch (type) {
+            case 'anthropic':
+                return new AnthropicProvider(config);
+            case 'openai':
+                return new OpenAIProvider(config);
             case 'groq':
                 return new GroqProvider(config);
             case 'ollama':
                 return new OllamaProvider(config);
             case 'inference-snaps':
                 return new InferenceSnapsProvider(config);
+            case 'huggingface':
+                // HuggingFace exposes an OpenAI-compatible inference endpoint; baseURL is
+                // per-model (HF_MODEL_URL), so it has no dedicated wrapper — the compat
+                // base serves it directly. Fixes the latent defect where the env factory
+                // accepted 'huggingface' but createProvider threw 'Unknown provider type'.
+                return new OpenAICompatProvider(config);
             default:
                 throw new Error(`Unknown provider type: ${String(type)}`);
         }
@@ -412,23 +448,29 @@ export class LLMClient {
  * Create an LLM client from environment variables.
  *
  * When LLM_PROVIDER is not set, auto-detects the provider by checking env vars
- * in priority order: INFERENCE_SNAPS → GROQ → OLLAMA. If none are set, defaults
- * to Inference Snaps at http://localhost:9090/v1 (Ubuntu local) and emits a
- * one-line stderr warning so the implicit default is discoverable in logs.
+ * in priority order: INFERENCE_SNAPS → GROQ → OLLAMA → ANTHROPIC_API_KEY →
+ * OPENAI_API_KEY. If none are set, defaults to Inference Snaps at
+ * http://localhost:9090/v1 (Ubuntu local) and emits a one-line stderr warning so
+ * the implicit localhost default is discoverable in logs.
  *
- * All providers use OpenAI-compatible APIs. No proprietary provider SDKs.
+ * All providers use OpenAI-compatible APIs. No proprietary provider SDKs
+ * (Anthropic + OpenAI ride their OpenAI-compatible endpoints).
  *
  * Canonical Inference Snaps is the reference local provider on Ubuntu — offline,
  * silicon-optimized, no API key required. See `providers/inference-snaps.ts` for
  * install docs (`sudo snap install gemma3`, etc.).
  *
  * Provider defaults:
- *   inference-snaps → gemma3       (base URL defaults to http://localhost:9090/v1)
+ *   inference-snaps → gemma3            (base URL defaults to http://localhost:9090/v1)
  *   groq            → qwen/qwen3-32b
- *   ollama          → gemma4:e2b   (base URL defaults to http://localhost:11434)
+ *   ollama          → gemma4:e2b        (base URL defaults to 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)
  */
 export function createLLMClientFromEnv() {
-    // Auto-detect provider when LLM_PROVIDER is not explicitly set
+    // Auto-detect provider when LLM_PROVIDER is not explicitly set. The existing
+    // priority order (INFERENCE_SNAPS → GROQ → OLLAMA) is preserved so existing
+    // deployments resolve identically; the frontier providers are appended after.
     let provider;
     if (process.env.LLM_PROVIDER) {
         provider = process.env.LLM_PROVIDER;
@@ -442,15 +484,38 @@ export function createLLMClientFromEnv() {
     else if (process.env.OLLAMA_BASE_URL) {
         provider = 'ollama';
     }
+    else if (process.env.ANTHROPIC_API_KEY) {
+        provider = 'anthropic';
+    }
+    else if (process.env.OPENAI_API_KEY) {
+        provider = 'openai';
+    }
     else {
-        // Zero-config Ubuntu default: assume Inference Snaps on the standard
-        // local port. Operator sees the warning once at boot.
+        // Zero-config Ubuntu default: assume Inference Snaps on the standard local
+        // port. This localhost default is unreachable inside a hosted serverless
+        // function, so warn once per process to make the implicit choice visible.
         provider = 'inference-snaps';
+        if (!warnedLocalhostDefault) {
+            warnedLocalhostDefault = true;
+            envFactoryLogger.warn('No LLM provider env var set — defaulting to inference-snaps at ' +
+                'http://localhost:9090/v1. This localhost endpoint is unreachable on a ' +
+                'hosted deployment; set LLM_PROVIDER or a provider key.');
+        }
     }
     let apiKey;
     let baseURL;
     let defaultModel;
-    if (provider === 'huggingface') {
+    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';
+    }
+    else if (provider === 'openai') {
+        apiKey = process.env.OPENAI_API_KEY;
+        baseURL = process.env.OPENAI_BASE_URL ?? 'https://api.openai.com/v1';
+        defaultModel = 'gpt-4o';
+    }
+    else if (provider === 'huggingface') {
         apiKey = process.env.HF_TOKEN;
         baseURL = process.env.HF_MODEL_URL;
     }
@@ -475,7 +540,8 @@ export function createLLMClientFromEnv() {
     }
     if (!apiKey) {
         throw new Error(`API key not found for provider "${provider}". Set the corresponding env var ` +
-            `(INFERENCE_SNAPS_BASE_URL, GROQ_API_KEY, OLLAMA_BASE_URL, or HF_TOKEN).`);
+            `(INFERENCE_SNAPS_BASE_URL, GROQ_API_KEY, OLLAMA_BASE_URL, HF_TOKEN, ` +
+            `ANTHROPIC_API_KEY, or OPENAI_API_KEY).`);
     }
     return new LLMClient({
         provider,
@@ -501,10 +567,19 @@ export function createLLMClientFromEnv() {
  * Returns `null` if the user has no stored keys (callers should fall back
  * to `createLLMClientFromEnv()` or return a 402/feature-unavailable error).
  *
+ * When `opts.hostedViableOnly` is set, a stored key whose provider is not
+ * hosted-viable (ollama / inference-snaps are localhost-only) resolves to
+ * `null` instead of a client. The resolver passes this on hosted deployments so
+ * a localhost-only BYOK key can never yield a localhost client — the exact
+ * silent-localhost defect GAP-360 closes (spec §6.5, fail-closed).
+ *
  * @param userId - The user's ID from the `users` table
  * @param db - A Drizzle NeonDB client instance
  */
-export async function createLLMClientForUser(userId, db, auditStore) {
+export async function createLLMClientForUser(userId, db, 
+// Only `append` is used; the narrow type lets a persistent store whose entry
+// type widens eventType/severity to `string` (e.g. DrizzleAuditStore) fit.
+auditStore, opts) {
     // Find the user's preferred provider config
     const [preferredConfig] = await db
         .select()
@@ -522,8 +597,12 @@ export async function createLLMClientForUser(userId, db, auditStore) {
     const [keyRow] = await keyQuery;
     if (!keyRow)
         return null;
-    const plaintext = decryptApiKey(keyRow.encryptedKey);
     const provider = keyRow.provider;
+    // Fail-closed hosted filter (§6.5): reject a localhost-only provider before
+    // decrypting, so no plaintext is touched for a key we will not use.
+    if (opts?.hostedViableOnly && !isHostedViable(provider))
+        return null;
+    const plaintext = decryptApiKey(keyRow.encryptedKey);
     const model = preferredConfig?.model ?? undefined;
     // Fire-and-forget: record when this key was last used (best-effort, never blocks)
     db.update(userApiKeys)

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

@@ -0,0 +1,30 @@
+/**
+ * Anthropic Provider
+ *
+ * Thin wrapper over OpenAICompatProvider using Anthropic's OpenAI-compatible
+ * surface at https://api.anthropic.com/v1. No proprietary Anthropic SDK — the
+ * fleet posture is "No proprietary provider SDKs" (see client.ts). Anthropic
+ * exposes an OpenAI-compatible chat/completions endpoint, so the same base
+ * implementation that serves Groq/Ollama/inference-snaps serves Anthropic too.
+ *
+ * Docs: https://docs.anthropic.com/en/api/openai-sdk
+ */
+import type { Embedding, LLMChatOptions, LLMChunk, LLMEmbedOptions, LLMProvider, LLMProviderConfig, LLMResponse, LLMStreamOptions, Message, ReasonerCapabilities } from './base.js';
+export interface AnthropicProviderConfig extends Omit<LLMProviderConfig, 'apiKey'> {
+    apiKey: string;
+    /** Defaults to https://api.anthropic.com/v1 */
+    baseURL?: string;
+    /** Defaults to claude-sonnet-4-6 (conservative documented default; overridable) */
+    model?: string;
+    timeout?: number;
+    maxRetries?: number;
+}
+export declare class AnthropicProvider implements LLMProvider {
+    private inner;
+    constructor(config: AnthropicProviderConfig);
+    capabilities(): ReasonerCapabilities;
+    chat(messages: Message[], options?: LLMChatOptions): Promise<LLMResponse>;
+    stream(messages: Message[], options?: LLMStreamOptions): AsyncIterable<LLMChunk>;
+    embed(_text: string | string[], _options?: LLMEmbedOptions): Promise<Embedding | Embedding[]>;
+}
+//# sourceMappingURL=anthropic.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"anthropic.d.ts","sourceRoot":"","sources":["../../../src/llm/providers/anthropic.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;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,uBAAwB,SAAQ,IAAI,CAAC,iBAAiB,EAAE,QAAQ,CAAC;IAChF,MAAM,EAAE,MAAM,CAAC;IACf,+CAA+C;IAC/C,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,mFAAmF;IACnF,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,UAAU,CAAC,EAAE,MAAM,CAAC;CACrB;AAED,qBAAa,iBAAkB,YAAW,WAAW;IACnD,OAAO,CAAC,KAAK,CAAuB;gBAExB,MAAM,EAAE,uBAAuB;IAQ3C,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/anthropic.js

@@ -0,0 +1,45 @@
+/**
+ * Anthropic Provider
+ *
+ * Thin wrapper over OpenAICompatProvider using Anthropic's OpenAI-compatible
+ * surface at https://api.anthropic.com/v1. No proprietary Anthropic SDK — the
+ * fleet posture is "No proprietary provider SDKs" (see client.ts). Anthropic
+ * exposes an OpenAI-compatible chat/completions endpoint, so the same base
+ * implementation that serves Groq/Ollama/inference-snaps serves Anthropic too.
+ *
+ * Docs: https://docs.anthropic.com/en/api/openai-sdk
+ */
+import { OpenAICompatProvider } from './openai-compat.js';
+export class AnthropicProvider {
+    inner;
+    constructor(config) {
+        this.inner = new OpenAICompatProvider({
+            ...config,
+            baseURL: config.baseURL ?? 'https://api.anthropic.com/v1',
+            model: config.model ?? 'claude-sonnet-4-6',
+        });
+    }
+    capabilities() {
+        return {
+            providerTag: 'anthropic',
+            tools: true,
+            parallelToolCalls: false,
+            vision: false,
+            streaming: true,
+            // Anthropic exposes no embeddings endpoint on its OpenAI-compat surface.
+            embeddings: false,
+            reasoningEffort: false,
+            promptCache: false,
+            structuredOutput: false,
+        };
+    }
+    chat(messages, options) {
+        return this.inner.chat(messages, options);
+    }
+    stream(messages, options) {
+        return this.inner.stream(messages, options);
+    }
+    embed(_text, _options) {
+        throw new Error('Anthropic does not expose an embeddings endpoint. Use OpenAI or Ollama.');
+    }
+}

package/dist/llm/providers/openai-compat.d.ts

@@ -7,11 +7,27 @@
  */
 import type { Embedding, LLMChatOptions, LLMChunk, LLMEmbedOptions, LLMProvider, LLMProviderConfig, LLMResponse, LLMStreamOptions, Message, ReasonerCapabilities } from './base.js';
 export interface OpenAICompatConfig extends LLMProviderConfig {
+    /** Per-request timeout in ms applied to connection setup. Default 60_000. */
+    timeout?: number;
+    /** Retries on connection/timeout errors (not on HTTP error responses). Default 1. */
+    maxRetries?: number;
 }
 export declare class OpenAICompatProvider implements LLMProvider {
     private config;
     private baseURL;
+    private timeoutMs;
+    private maxRetries;
     constructor(config: OpenAICompatConfig);
+    /**
+     * fetch() with a connection-setup timeout and bounded retries. The timeout
+     * aborts a hung connection (the localhost-default failure mode that otherwise
+     * burns the whole serverless duration) and is cleared once headers arrive, so
+     * a legitimately long streaming body is not truncated. Retries fire only on a
+     * thrown fetch error (connection refused, DNS, timeout) — never on an HTTP
+     * error response, which the caller inspects — so a POST is re-sent only when
+     * no response was received.
+     */
+    private fetchWithResilience;
     capabilities(): ReasonerCapabilities;
     chat(messages: Message[], options?: LLMChatOptions): Promise<LLMResponse>;
     embed(text: string | string[], options?: LLMEmbedOptions): Promise<Embedding | Embedding[]>;

package/dist/llm/providers/openai-compat.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"openai-compat.d.ts","sourceRoot":"","sources":["../../../src/llm/providers/openai-compat.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,OAAO,KAAK,EACV,SAAS,EAET,cAAc,EACd,QAAQ,EACR,eAAe,EACf,WAAW,EACX,iBAAiB,EACjB,WAAW,EACX,gBAAgB,EAChB,OAAO,EACP,oBAAoB,EAErB,MAAM,WAAW,CAAC;AAEnB,MAAM,WAAW,kBAAmB,SAAQ,iBAAiB;CAAG;AAqDhE,qBAAa,oBAAqB,YAAW,WAAW;IACtD,OAAO,CAAC,MAAM,CAAqB;IACnC,OAAO,CAAC,OAAO,CAAS;gBAEZ,MAAM,EAAE,kBAAkB;IAUtC,YAAY,IAAI,oBAAoB;IAgB9B,IAAI,CAAC,QAAQ,EAAE,OAAO,EAAE,EAAE,OAAO,CAAC,EAAE,cAAc,GAAG,OAAO,CAAC,WAAW,CAAC;IA6EzE,KAAK,CACT,IAAI,EAAE,MAAM,GAAG,MAAM,EAAE,EACvB,OAAO,CAAC,EAAE,eAAe,GACxB,OAAO,CAAC,SAAS,GAAG,SAAS,EAAE,CAAC;IAyC5B,MAAM,CAAC,QAAQ,EAAE,OAAO,EAAE,EAAE,OAAO,CAAC,EAAE,gBAAgB,GAAG,aAAa,CAAC,QAAQ,CAAC;IAuFvF,OAAO,CAAC,cAAc;CA4BvB"}
+{"version":3,"file":"openai-compat.d.ts","sourceRoot":"","sources":["../../../src/llm/providers/openai-compat.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,OAAO,KAAK,EACV,SAAS,EAET,cAAc,EACd,QAAQ,EACR,eAAe,EACf,WAAW,EACX,iBAAiB,EACjB,WAAW,EACX,gBAAgB,EAChB,OAAO,EACP,oBAAoB,EAErB,MAAM,WAAW,CAAC;AAEnB,MAAM,WAAW,kBAAmB,SAAQ,iBAAiB;IAC3D,6EAA6E;IAC7E,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,qFAAqF;IACrF,UAAU,CAAC,EAAE,MAAM,CAAC;CACrB;AAwDD,qBAAa,oBAAqB,YAAW,WAAW;IACtD,OAAO,CAAC,MAAM,CAAqB;IACnC,OAAO,CAAC,OAAO,CAAS;IACxB,OAAO,CAAC,SAAS,CAAS;IAC1B,OAAO,CAAC,UAAU,CAAS;gBAEf,MAAM,EAAE,kBAAkB;IAYtC;;;;;;;;OAQG;YACW,mBAAmB;IAgBjC,YAAY,IAAI,oBAAoB;IAgB9B,IAAI,CAAC,QAAQ,EAAE,OAAO,EAAE,EAAE,OAAO,CAAC,EAAE,cAAc,GAAG,OAAO,CAAC,WAAW,CAAC;IA6EzE,KAAK,CACT,IAAI,EAAE,MAAM,GAAG,MAAM,EAAE,EACvB,OAAO,CAAC,EAAE,eAAe,GACxB,OAAO,CAAC,SAAS,GAAG,SAAS,EAAE,CAAC;IAyC5B,MAAM,CAAC,QAAQ,EAAE,OAAO,EAAE,EAAE,OAAO,CAAC,EAAE,gBAAgB,GAAG,aAAa,CAAC,QAAQ,CAAC;IAuFvF,OAAO,CAAC,cAAc;CA4BvB"}

package/dist/llm/providers/openai-compat.js

@@ -5,6 +5,8 @@
  * format. Used by: Ollama, Groq, Inference Snaps.
  * NOT for direct OpenAI usage  -  RevealUI uses open-source models only.
  */
+const DEFAULT_TIMEOUT_MS = 60_000;
+const DEFAULT_MAX_RETRIES = 1;
 const authorizationHeader = 'Authorization';
 const maxTokensKey = 'max_tokens';
 const toolChoiceKey = 'tool_choice';
@@ -31,12 +33,42 @@ const isFunctionToolCall = (call) => {
 export class OpenAICompatProvider {
     config;
     baseURL;
+    timeoutMs;
+    maxRetries;
     constructor(config) {
         this.config = config;
         if (!config.baseURL) {
             throw new Error('OpenAICompatProvider requires a baseURL  -  use a specific provider (InferenceSnapsProvider, OllamaProvider, etc.)');
         }
         this.baseURL = config.baseURL;
+        this.timeoutMs = config.timeout ?? DEFAULT_TIMEOUT_MS;
+        this.maxRetries = config.maxRetries ?? DEFAULT_MAX_RETRIES;
+    }
+    /**
+     * fetch() with a connection-setup timeout and bounded retries. The timeout
+     * aborts a hung connection (the localhost-default failure mode that otherwise
+     * burns the whole serverless duration) and is cleared once headers arrive, so
+     * a legitimately long streaming body is not truncated. Retries fire only on a
+     * thrown fetch error (connection refused, DNS, timeout) — never on an HTTP
+     * error response, which the caller inspects — so a POST is re-sent only when
+     * no response was received.
+     */
+    async fetchWithResilience(url, init) {
+        let lastError;
+        for (let attempt = 0; attempt <= this.maxRetries; attempt++) {
+            const controller = new AbortController();
+            const timeoutId = setTimeout(() => controller.abort(), this.timeoutMs);
+            try {
+                return await fetch(url, { ...init, signal: controller.signal });
+            }
+            catch (error) {
+                lastError = error;
+            }
+            finally {
+                clearTimeout(timeoutId);
+            }
+        }
+        throw lastError instanceof Error ? lastError : new Error(String(lastError));
     }
     capabilities() {
         // Base OpenAI-compatible profile. Concrete providers (Groq, Ollama, inference-snaps)
@@ -54,7 +86,7 @@ export class OpenAICompatProvider {
         };
     }
     async chat(messages, options) {
-        const response = await fetch(`${this.baseURL}/chat/completions`, {
+        const response = await this.fetchWithResilience(`${this.baseURL}/chat/completions`, {
             method: 'POST',
             headers: {
                 'Content-Type': 'application/json',
@@ -122,7 +154,7 @@ export class OpenAICompatProvider {
     async embed(text, options) {
         const texts = Array.isArray(text) ? text : [text];
         const model = options?.model || 'text-embedding-3-small';
-        const response = await fetch(`${this.baseURL}/embeddings`, {
+        const response = await this.fetchWithResilience(`${this.baseURL}/embeddings`, {
             method: 'POST',
             headers: {
                 'Content-Type': 'application/json',
@@ -155,7 +187,7 @@ export class OpenAICompatProvider {
         return Array.isArray(text) ? embeddings : embeddings[0];
     }
     async *stream(messages, options) {
-        const response = await fetch(`${this.baseURL}/chat/completions`, {
+        const response = await this.fetchWithResilience(`${this.baseURL}/chat/completions`, {
             method: 'POST',
             headers: {
                 'Content-Type': 'application/json',

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

@@ -0,0 +1,27 @@
+/**
+ * OpenAI Provider
+ *
+ * Thin wrapper over OpenAICompatProvider targeting OpenAI's native
+ * chat/completions + embeddings API at https://api.openai.com/v1. OpenAI is the
+ * reference OpenAI-compatible surface, so the shared base implementation serves
+ * it directly — no proprietary SDK (fleet posture in client.ts).
+ */
+import type { Embedding, LLMChatOptions, LLMChunk, LLMEmbedOptions, LLMProvider, LLMProviderConfig, LLMResponse, LLMStreamOptions, Message, ReasonerCapabilities } from './base.js';
+export interface OpenAIProviderConfig extends Omit<LLMProviderConfig, 'apiKey'> {
+    apiKey: string;
+    /** Defaults to https://api.openai.com/v1 */
+    baseURL?: string;
+    /** Defaults to gpt-4o (conservative documented default; overridable) */
+    model?: string;
+    timeout?: number;
+    maxRetries?: number;
+}
+export declare class OpenAIProvider implements LLMProvider {
+    private inner;
+    constructor(config: OpenAIProviderConfig);
+    capabilities(): ReasonerCapabilities;
+    chat(messages: Message[], options?: LLMChatOptions): Promise<LLMResponse>;
+    stream(messages: Message[], options?: LLMStreamOptions): AsyncIterable<LLMChunk>;
+    embed(text: string | string[], options?: LLMEmbedOptions): Promise<Embedding | Embedding[]>;
+}
+//# sourceMappingURL=openai.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"openai.d.ts","sourceRoot":"","sources":["../../../src/llm/providers/openai.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;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,oBAAqB,SAAQ,IAAI,CAAC,iBAAiB,EAAE,QAAQ,CAAC;IAC7E,MAAM,EAAE,MAAM,CAAC;IACf,4CAA4C;IAC5C,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,wEAAwE;IACxE,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,UAAU,CAAC,EAAE,MAAM,CAAC;CACrB;AAED,qBAAa,cAAe,YAAW,WAAW;IAChD,OAAO,CAAC,KAAK,CAAuB;gBAExB,MAAM,EAAE,oBAAoB;IAQxC,YAAY,IAAI,oBAAoB;IAcpC,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,IAAI,EAAE,MAAM,GAAG,MAAM,EAAE,EAAE,OAAO,CAAC,EAAE,eAAe,GAAG,OAAO,CAAC,SAAS,GAAG,SAAS,EAAE,CAAC;CAG5F"}

package/dist/llm/providers/openai.js

@@ -0,0 +1,41 @@
+/**
+ * OpenAI Provider
+ *
+ * Thin wrapper over OpenAICompatProvider targeting OpenAI's native
+ * chat/completions + embeddings API at https://api.openai.com/v1. OpenAI is the
+ * reference OpenAI-compatible surface, so the shared base implementation serves
+ * it directly — no proprietary SDK (fleet posture in client.ts).
+ */
+import { OpenAICompatProvider } from './openai-compat.js';
+export class OpenAIProvider {
+    inner;
+    constructor(config) {
+        this.inner = new OpenAICompatProvider({
+            ...config,
+            baseURL: config.baseURL ?? 'https://api.openai.com/v1',
+            model: config.model ?? 'gpt-4o',
+        });
+    }
+    capabilities() {
+        return {
+            providerTag: 'openai',
+            tools: true,
+            parallelToolCalls: false,
+            vision: false,
+            streaming: true,
+            embeddings: true,
+            reasoningEffort: false,
+            promptCache: false,
+            structuredOutput: false,
+        };
+    }
+    chat(messages, options) {
+        return this.inner.chat(messages, options);
+    }
+    stream(messages, options) {
+        return this.inner.stream(messages, options);
+    }
+    embed(text, options) {
+        return this.inner.embed(text, options);
+    }
+}

package/dist/llm/resolve.d.ts

@@ -0,0 +1,97 @@
+/**
+ * Per-request LLM client resolver (GAP-360 PR-2).
+ *
+ * The single home for key resolution at every dispatch site (spec §5.2 / §5.4).
+ * One resolver, consumed by all sites — no site-local copies.
+ *
+ * Resolution order (spec §5.2):
+ *   1. Per-user BYOK        — createLLMClientForUser (preferred)
+ *   2. Site inference config — workspace_inference_configs (hostedViable on hosted)
+ *   3. Deployment env        — createLLMClientFromEnv (SELF-HOSTED ONLY)
+ *   4. Hosted + nothing above → throw LLMNotConfiguredError (typed → HTTP 409)
+ *
+ * Security invariants (spec §6, guardrail-2):
+ *   - §6.1 userId is authenticated-identity-scoped; the caller derives it from
+ *     session/entitlement context — or, for the durable worker, the
+ *     authenticated dispatcher captured server-side at enqueue time — never
+ *     from request params/body, and never from a client-writable DB column
+ *     (e.g. a ticket's `reporterId`, which has no ownership check on the
+ *     general tickets API). The resolver takes userId as an argument and
+ *     never reads it from a request. The same rule binds the step-2 site
+ *     inference key: the workspaceId is client-writable, so its stored key is
+ *     decrypted only after userCanAccessSite confirms the caller owns or
+ *     collaborates on that site. Otherwise a request could name another
+ *     tenant's site id and run on that site's key.
+ *   - §6.2 plaintext lifetime = request scope. The client is constructed per
+ *     request; no key cache. Nothing here logs, serializes, or returns a key.
+ *   - §6.3 decryption is server-side only via the existing decryptApiKey.
+ *   - §6.4 the byok:key:accessed audit event fires inside createLLMClientForUser
+ *     when an audit store is wired (ctx.auditStore).
+ *   - §6.5 fail-closed: unknown provider / failed decrypt / a non-hostedViable
+ *     provider on hosted resolves to LLMNotConfiguredError, never an env
+ *     fallthrough on hosted (the exact silent-localhost defect class).
+ *   - §6.6 no new secret surface.
+ *
+ * Feature flag (spec §7): HOSTED_BYOK_DISPATCH. Default ON for hosted, absent
+ * (off) for self-hosted so self-hosted env-first behavior is byte-unchanged.
+ * The flag is the one-release rollback lever.
+ */
+import type { Database } from '@revealui/db/client';
+import type { AuditStore } from '../audit/store.js';
+import { LLMClient } from './client.js';
+/**
+ * Thrown when a hosted deployment has no usable LLM configuration for the
+ * request. Maps to HTTP 409 (configuration is the remedy, not payment).
+ * Callers name {@link settingsPath} in the machine-readable response body.
+ */
+export declare class LLMNotConfiguredError extends Error {
+    /** Stable machine-readable code for API response bodies. */
+    readonly code: "LLM_NOT_CONFIGURED";
+    /** Where the account owner configures a key. */
+    readonly settingsPath: "/settings/api-keys";
+    constructor(message?: string);
+}
+/** The 409 response body every dispatch site returns for an unconfigured account. */
+export interface LLMNotConfiguredBody {
+    success: false;
+    error: string;
+    code: 'LLM_NOT_CONFIGURED';
+    settingsPath: '/settings/api-keys';
+}
+/** Build the machine-readable 409 body for {@link LLMNotConfiguredError}. */
+export declare function llmNotConfiguredBody(err: LLMNotConfiguredError): LLMNotConfiguredBody;
+export interface ResolveLLMContext {
+    /**
+     * True on the hosted revealui.com SaaS deployment. Derived by the caller from
+     * the existing deployment-mode signal (server: detectDeploymentMode; admin:
+     * REVEALUI_LICENSE_PRIVATE_KEY presence) — never sniffed here.
+     */
+    isHosted: boolean;
+    /** Site id for step-2 site-level config lookup. Omit when the site has none. */
+    workspaceId?: string;
+    /**
+     * Durable audit sink for the byok:key:accessed event (§6.4). Only `append` is
+     * used, so a persistent store (DrizzleAuditStore) fits. Omit where none is
+     * wired (e.g. the admin process, in-memory until GAP-338 closes).
+     */
+    auditStore?: Pick<AuditStore, 'append'>;
+}
+/**
+ * Whether the BYOK dispatch order (spec §5.2) is active. Default follows the
+ * deployment: ON when hosted, OFF when self-hosted. HOSTED_BYOK_DISPATCH
+ * overrides explicitly; an unrecognized value falls back to the default.
+ */
+export declare function hostedByokDispatchEnabled(isHosted: boolean): boolean;
+/**
+ * Resolve the LLM client for a single request.
+ *
+ * @param userId - Authenticated user id, from session/entitlement context for
+ *   request-scoped callers, or the authenticated dispatcher captured
+ *   server-side at enqueue time for the durable worker. Never a request
+ *   param/body value, and never a client-writable DB column. Null when there
+ *   is no authenticated user.
+ * @param db - Drizzle client.
+ * @param ctx - Deployment mode, optional site id, optional audit sink.
+ */
+export declare function resolveLLMClientForRequest(userId: string | null, db: Database, ctx: ResolveLLMContext): Promise<LLMClient>;
+//# sourceMappingURL=resolve.d.ts.map

package/dist/llm/resolve.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"resolve.d.ts","sourceRoot":"","sources":["../../src/llm/resolve.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAqCG;AAGH,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,qBAAqB,CAAC;AAIpD,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,mBAAmB,CAAC;AACpD,OAAO,EAIL,SAAS,EAEV,MAAM,aAAa,CAAC;AAMrB;;;;GAIG;AACH,qBAAa,qBAAsB,SAAQ,KAAK;IAC9C,4DAA4D;IAC5D,QAAQ,CAAC,IAAI,EAAG,oBAAoB,CAAU;IAC9C,gDAAgD;IAChD,QAAQ,CAAC,YAAY,EAAG,oBAAoB,CAAU;gBAE1C,OAAO,SAAoD;CAIxE;AAED,qFAAqF;AACrF,MAAM,WAAW,oBAAoB;IACnC,OAAO,EAAE,KAAK,CAAC;IACf,KAAK,EAAE,MAAM,CAAC;IACd,IAAI,EAAE,oBAAoB,CAAC;IAC3B,YAAY,EAAE,oBAAoB,CAAC;CACpC;AAED,6EAA6E;AAC7E,wBAAgB,oBAAoB,CAAC,GAAG,EAAE,qBAAqB,GAAG,oBAAoB,CAOrF;AAED,MAAM,WAAW,iBAAiB;IAChC;;;;OAIG;IACH,QAAQ,EAAE,OAAO,CAAC;IAClB,gFAAgF;IAChF,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB;;;;OAIG;IACH,UAAU,CAAC,EAAE,IAAI,CAAC,UAAU,EAAE,QAAQ,CAAC,CAAC;CACzC;AAKD;;;;GAIG;AACH,wBAAgB,yBAAyB,CAAC,QAAQ,EAAE,OAAO,GAAG,OAAO,CAOpE;AAED;;;;;;;;;;GAUG;AACH,wBAAsB,0BAA0B,CAC9C,MAAM,EAAE,MAAM,GAAG,IAAI,EACrB,EAAE,EAAE,QAAQ,EACZ,GAAG,EAAE,iBAAiB,GACrB,OAAO,CAAC,SAAS,CAAC,CAsDpB"}

package/dist/llm/resolve.js

@@ -0,0 +1,218 @@
+/**
+ * Per-request LLM client resolver (GAP-360 PR-2).
+ *
+ * The single home for key resolution at every dispatch site (spec §5.2 / §5.4).
+ * One resolver, consumed by all sites — no site-local copies.
+ *
+ * Resolution order (spec §5.2):
+ *   1. Per-user BYOK        — createLLMClientForUser (preferred)
+ *   2. Site inference config — workspace_inference_configs (hostedViable on hosted)
+ *   3. Deployment env        — createLLMClientFromEnv (SELF-HOSTED ONLY)
+ *   4. Hosted + nothing above → throw LLMNotConfiguredError (typed → HTTP 409)
+ *
+ * Security invariants (spec §6, guardrail-2):
+ *   - §6.1 userId is authenticated-identity-scoped; the caller derives it from
+ *     session/entitlement context — or, for the durable worker, the
+ *     authenticated dispatcher captured server-side at enqueue time — never
+ *     from request params/body, and never from a client-writable DB column
+ *     (e.g. a ticket's `reporterId`, which has no ownership check on the
+ *     general tickets API). The resolver takes userId as an argument and
+ *     never reads it from a request. The same rule binds the step-2 site
+ *     inference key: the workspaceId is client-writable, so its stored key is
+ *     decrypted only after userCanAccessSite confirms the caller owns or
+ *     collaborates on that site. Otherwise a request could name another
+ *     tenant's site id and run on that site's key.
+ *   - §6.2 plaintext lifetime = request scope. The client is constructed per
+ *     request; no key cache. Nothing here logs, serializes, or returns a key.
+ *   - §6.3 decryption is server-side only via the existing decryptApiKey.
+ *   - §6.4 the byok:key:accessed audit event fires inside createLLMClientForUser
+ *     when an audit store is wired (ctx.auditStore).
+ *   - §6.5 fail-closed: unknown provider / failed decrypt / a non-hostedViable
+ *     provider on hosted resolves to LLMNotConfiguredError, never an env
+ *     fallthrough on hosted (the exact silent-localhost defect class).
+ *   - §6.6 no new secret surface.
+ *
+ * Feature flag (spec §7): HOSTED_BYOK_DISPATCH. Default ON for hosted, absent
+ * (off) for self-hosted so self-hosted env-first behavior is byte-unchanged.
+ * The flag is the one-release rollback lever.
+ */
+import { createLogger } from '@revealui/core/observability/logger';
+import { decryptApiKey } from '@revealui/db/crypto';
+import { siteCollaborators, sites, workspaceInferenceConfigs } from '@revealui/db/schema';
+import { and, eq } from 'drizzle-orm';
+import { createLLMClientForUser, createLLMClientFromEnv, isHostedViable, LLMClient, } from './client.js';
+const resolverLogger = createLogger({ component: 'resolveLLMClientForRequest' });
+/** Emitted once per process when the hosted BYOK rollback lever is pulled. */
+let warnedBreakGlass = false;
+/**
+ * Thrown when a hosted deployment has no usable LLM configuration for the
+ * request. Maps to HTTP 409 (configuration is the remedy, not payment).
+ * Callers name {@link settingsPath} in the machine-readable response body.
+ */
+export class LLMNotConfiguredError extends Error {
+    /** Stable machine-readable code for API response bodies. */
+    code = 'LLM_NOT_CONFIGURED';
+    /** Where the account owner configures a key. */
+    settingsPath = '/settings/api-keys';
+    constructor(message = 'No LLM provider is configured for this account.') {
+        super(message);
+        this.name = 'LLMNotConfiguredError';
+    }
+}
+/** Build the machine-readable 409 body for {@link LLMNotConfiguredError}. */
+export function llmNotConfiguredBody(err) {
+    return {
+        success: false,
+        error: err.message,
+        code: err.code,
+        settingsPath: err.settingsPath,
+    };
+}
+const FLAG_ON = new Set(['true', '1', 'on', 'yes']);
+const FLAG_OFF = new Set(['false', '0', 'off', 'no']);
+/**
+ * Whether the BYOK dispatch order (spec §5.2) is active. Default follows the
+ * deployment: ON when hosted, OFF when self-hosted. HOSTED_BYOK_DISPATCH
+ * overrides explicitly; an unrecognized value falls back to the default.
+ */
+export function hostedByokDispatchEnabled(isHosted) {
+    const raw = process.env.HOSTED_BYOK_DISPATCH;
+    if (raw === undefined || raw.trim() === '')
+        return isHosted;
+    const value = raw.trim().toLowerCase();
+    if (FLAG_OFF.has(value))
+        return false;
+    if (FLAG_ON.has(value))
+        return true;
+    return isHosted;
+}
+/**
+ * Resolve the LLM client for a single request.
+ *
+ * @param userId - Authenticated user id, from session/entitlement context for
+ *   request-scoped callers, or the authenticated dispatcher captured
+ *   server-side at enqueue time for the durable worker. Never a request
+ *   param/body value, and never a client-writable DB column. Null when there
+ *   is no authenticated user.
+ * @param db - Drizzle client.
+ * @param ctx - Deployment mode, optional site id, optional audit sink.
+ */
+export async function resolveLLMClientForRequest(userId, db, ctx) {
+    const hosted = ctx.isHosted;
+    // Feature-flag gate. When disabled (self-hosted default), behavior is
+    // byte-unchanged: env-first, exactly as before this PR.
+    if (!hostedByokDispatchEnabled(hosted)) {
+        if (hosted && !warnedBreakGlass) {
+            warnedBreakGlass = true;
+            // Break-glass: HOSTED_BYOK_DISPATCH is explicitly off on a hosted
+            // deployment. Every account now shares the deployment env client while
+            // this lever is pulled — a deliberate one-release rollback, but an
+            // operator must know it is active.
+            resolverLogger.warn('HOSTED_BYOK_DISPATCH is disabled on a hosted deployment — all accounts are ' +
+                'sharing the deployment env LLM client instead of per-account BYOK keys.');
+        }
+        return createLLMClientFromEnv();
+    }
+    // 1. Per-user BYOK (preferred). On hosted, filter to hostedViable providers
+    //    so a localhost-only BYOK key (e.g. ollama) can never yield a localhost
+    //    client — that is the silent-localhost defect (§6.5, fail-closed).
+    if (userId) {
+        try {
+            const byok = await createLLMClientForUser(userId, db, ctx.auditStore, {
+                hostedViableOnly: hosted,
+            });
+            if (byok)
+                return byok;
+        }
+        catch {
+            // Failed decrypt / unknown-provider row / CHECK-violating row. Fail-closed
+            // on hosted — never fall through to env. Self-hosted may continue.
+            if (hosted) {
+                throw new LLMNotConfiguredError('Your stored API key could not be used. Re-add it under /settings/api-keys.');
+            }
+        }
+    }
+    // 2. Site-level inference config. userId is passed so the site's stored key
+    //    is decrypted only for a caller authorized on that site (§6.1) — the
+    //    workspaceId reaching here is a client-writable value (e.g. a request
+    //    body field), so it carries no ownership guarantee on its own.
+    const siteClient = await resolveSiteInferenceClient(db, userId, ctx.workspaceId, hosted);
+    if (siteClient)
+        return siteClient;
+    // 3. Deployment env — SELF-HOSTED ONLY. Forbidden on hosted (§5.2 step 3).
+    if (!hosted) {
+        return createLLMClientFromEnv();
+    }
+    // 4. Hosted + nothing above → fail loud, actionable.
+    throw new LLMNotConfiguredError();
+}
+/**
+ * Whether `userId` is authorized to act on `siteId` — the site's owner, or a
+ * row in `site_collaborators`. Gate for the site inference key (§6.1): the
+ * site's stored provider key must never be decrypted for a caller who does not
+ * belong to the site, even though the workspaceId is client-supplied.
+ */
+async function userCanAccessSite(db, userId, siteId) {
+    const [owned] = await db
+        .select({ id: sites.id })
+        .from(sites)
+        .where(and(eq(sites.id, siteId), eq(sites.ownerId, userId)))
+        .limit(1);
+    if (owned)
+        return true;
+    const [collaborator] = await db
+        .select({ id: siteCollaborators.id })
+        .from(siteCollaborators)
+        .where(and(eq(siteCollaborators.siteId, siteId), eq(siteCollaborators.userId, userId)))
+        .limit(1);
+    return Boolean(collaborator);
+}
+/**
+ * Build a client from the site's workspace_inference_configs row (spec §5.2
+ * step 2). Returns null when the site has no config, when the caller is not
+ * authorized on the site, or (on hosted) when its provider is not hostedViable.
+ * Fail-closed: a malformed row throws on hosted and is skipped on self-hosted.
+ */
+async function resolveSiteInferenceClient(db, userId, workspaceId, hosted) {
+    if (!workspaceId)
+        return null;
+    // §6.1 authorization: only decrypt a site's key for a caller who belongs to
+    // that site. An unauthorized (or anonymous) request skips the site config
+    // entirely; on hosted it then falls through to the step-4 409, never a
+    // decrypt of another tenant's key.
+    if (!(userId && (await userCanAccessSite(db, userId, workspaceId))))
+        return null;
+    const [config] = await db
+        .select()
+        .from(workspaceInferenceConfigs)
+        .where(eq(workspaceInferenceConfigs.workspaceId, workspaceId))
+        .limit(1);
+    if (!config)
+        return null;
+    const provider = config.provider;
+    // On hosted, only hostedViable providers are reachable. A non-viable site
+    // config is skipped so resolution fails closed rather than hitting localhost.
+    if (hosted && !isHostedViable(provider))
+        return null;
+    try {
+        // Keyless providers (ollama / inference-snaps) carry a NULL encrypted key
+        // and use the provider name as the placeholder key — matches the env
+        // factory. Keyed providers decrypt server-side at dispatch time (§6.3).
+        const apiKey = config.encryptedApiKey ? decryptApiKey(config.encryptedApiKey) : provider;
+        return new LLMClient({
+            provider,
+            apiKey,
+            model: config.model ?? undefined,
+            baseURL: config.baseURL ?? undefined,
+            temperature: config.temperature ?? undefined,
+            maxTokens: config.maxTokens ?? undefined,
+        });
+    }
+    catch {
+        // Unknown provider / failed decrypt. Fail-closed on hosted; skip on self-hosted.
+        if (hosted) {
+            throw new LLMNotConfiguredError('This site’s inference configuration could not be used. Update it under /settings/api-keys.');
+        }
+        return null;
+    }
+}

package/dist/llm/server.d.ts

@@ -10,5 +10,6 @@ export * from './providers/groq.js';
 export * from './providers/inference-snaps.js';
 export * from './providers/ollama.js';
 export * from './providers/openai-compat.js';
+export * from './resolve.js';
 export * from './workspace-provider-config.js';
 //# sourceMappingURL=server.d.ts.map

package/dist/llm/server.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"server.d.ts","sourceRoot":"","sources":["../../src/llm/server.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAGH,cAAc,aAAa,CAAC;AAG5B,cAAc,qBAAqB,CAAC;AACpC,cAAc,qBAAqB,CAAC;AACpC,cAAc,gCAAgC,CAAC;AAC/C,cAAc,uBAAuB,CAAC;AACtC,cAAc,8BAA8B,CAAC;AAI7C,cAAc,gCAAgC,CAAC"}
+{"version":3,"file":"server.d.ts","sourceRoot":"","sources":["../../src/llm/server.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAGH,cAAc,aAAa,CAAC;AAE5B,cAAc,qBAAqB,CAAC;AACpC,cAAc,qBAAqB,CAAC;AACpC,cAAc,gCAAgC,CAAC;AAC/C,cAAc,uBAAuB,CAAC;AACtC,cAAc,8BAA8B,CAAC;AAE7C,cAAc,cAAc,CAAC;AAI7B,cAAc,gCAAgC,CAAC"}

package/dist/llm/server.js

@@ -12,6 +12,8 @@ export * from './providers/groq.js';
 export * from './providers/inference-snaps.js';
 export * from './providers/ollama.js';
 export * from './providers/openai-compat.js';
+// Export the per-request client resolver (GAP-360 PR-2)
+export * from './resolve.js';
 // Export per-workspace provider registry (used by admin inference-config route
 // to hydrate per-site config at boot + apply changes immediately on PUT)
 export * from './workspace-provider-config.js';

package/dist/skills/types.d.ts

@@ -27,13 +27,13 @@ export declare const SkillMetadataSchema: z.ZodObject<{
     repository: z.ZodOptional<z.ZodString>;
     tags: z.ZodOptional<z.ZodArray<z.ZodString>>;
     compatibility: z.ZodOptional<z.ZodArray<z.ZodEnum<{
+        anthropic: "anthropic";
+        openai: "openai";
         "claude-code": "claude-code";
         cursor: "cursor";
         windsurf: "windsurf";
         cline: "cline";
         copilot: "copilot";
-        openai: "openai";
-        anthropic: "anthropic";
         universal: "universal";
     }>>>;
     allowedTools: z.ZodOptional<z.ZodArray<z.ZodString>>;
@@ -78,13 +78,13 @@ export declare const SkillSchema: z.ZodObject<{
         repository: z.ZodOptional<z.ZodString>;
         tags: z.ZodOptional<z.ZodArray<z.ZodString>>;
         compatibility: z.ZodOptional<z.ZodArray<z.ZodEnum<{
+            anthropic: "anthropic";
+            openai: "openai";
             "claude-code": "claude-code";
             cursor: "cursor";
             windsurf: "windsurf";
             cline: "cline";
             copilot: "copilot";
-            openai: "openai";
-            anthropic: "anthropic";
             universal: "universal";
         }>>>;
         allowedTools: z.ZodOptional<z.ZodArray<z.ZodString>>;
@@ -143,13 +143,13 @@ export declare const SkillActivationResultSchema: z.ZodObject<{
             repository: z.ZodOptional<z.ZodString>;
             tags: z.ZodOptional<z.ZodArray<z.ZodString>>;
             compatibility: z.ZodOptional<z.ZodArray<z.ZodEnum<{
+                anthropic: "anthropic";
+                openai: "openai";
                 "claude-code": "claude-code";
                 cursor: "cursor";
                 windsurf: "windsurf";
                 cline: "cline";
                 copilot: "copilot";
-                openai: "openai";
-                anthropic: "anthropic";
                 universal: "universal";
             }>>>;
             allowedTools: z.ZodOptional<z.ZodArray<z.ZodString>>;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@revealui/ai",
-  "version": "0.6.4",
+  "version": "0.7.0",
   "description": "AI runtime for agent-driven products — agents, memory, LLM providers (Inference Snaps, Ollama, OpenAI-compatible), tools, and orchestration. Anthropic-SDK-free.",
   "keywords": [
     "agent",
@@ -25,9 +25,9 @@
   "dependencies": {
     "lru-cache": "^11.5.1",
     "zod": "^4.4.3",
-    "@revealui/contracts": "0.6.2",
-    "@revealui/core": "0.11.0",
-    "@revealui/db": "0.7.3",
+    "@revealui/contracts": "0.7.0",
+    "@revealui/core": "0.11.1",
+    "@revealui/db": "0.8.0",
     "@revealui/resilience": "0.2.4"
   },
   "devDependencies": {