@juspay/neurolink 12.55.3 → 12.56.1

package/dist/cli/commands/setup.d.ts

@@ -13,9 +13,9 @@ import type { ProviderConfigOptions, SetupArgs } from "../../types/index.js";
  * perplexity, sambanova, together-ai, xai — mistral is the 9th catalog
  * provider but keeps its own handleMistralSetup() flow above, so it's
  * excluded here), spread directly from the catalog JSON via
- * buildCatalogConfigOptions(); decide-kind catalog entries (laya) are spread
- * the same way from getDecideCatalogEntries(). The hand rows listed first
- * reuse the existing createXConfig() factories in providerConfig.ts; the
+ * buildCatalogConfigOptions(); decide-kind catalog entries (laya, xor) are
+ * spread the same way from getDecideCatalogEntries(). The hand rows listed
+ * first reuse the existing createXConfig() factories in providerConfig.ts; the
  * remaining 5 (ollama, litellm, sagemaker, lm-studio, llamacpp) don't have a
  * factory and are defined inline using their real env var names. lm-studio
  * and llamacpp previously

package/dist/cli/commands/setup.js

@@ -21,7 +21,7 @@ import { handleHuggingFaceSetup } from "./setup-huggingface.js";
 import { handleMistralSetup } from "./setup-mistral.js";
 import { OpenRouterModels, } from "../../constants/enums.js";
 import { PROVIDER_DESCRIPTORS_BY_NAME } from "../../factories/providerDescriptors.js";
-import { createCohereConfig, createDeepSeekConfig, createIdeogramConfig, createJinaConfig, createNvidiaNimConfig, createOpenAICompatibleConfig, createRecraftConfig, createReplicateConfig, createStabilityConfig, createVoyageConfig, createTypeSafeConfig, createXorConfig, createPerplexityDeciderConfig, createCloudflareClefConfig, satisfiesFallbacks, } from "../../utils/providerConfig.js";
+import { createCohereConfig, createDeepSeekConfig, createIdeogramConfig, createJinaConfig, createNvidiaNimConfig, createOpenAICompatibleConfig, createRecraftConfig, createReplicateConfig, createStabilityConfig, createVoyageConfig, createTypeSafeConfig, createPerplexityDeciderConfig, createCloudflareClefConfig, satisfiesFallbacks, } from "../../utils/providerConfig.js";
 import { getCatalogJsonEntries, getDecideCatalogEntries, buildCatalogConfigOptions, } from "../../providers/catalog/loader.js";
 // Provider information database
 const PROVIDERS = [
@@ -150,9 +150,9 @@ const PROVIDERS = [
  * perplexity, sambanova, together-ai, xai — mistral is the 9th catalog
  * provider but keeps its own handleMistralSetup() flow above, so it's
  * excluded here), spread directly from the catalog JSON via
- * buildCatalogConfigOptions(); decide-kind catalog entries (laya) are spread
- * the same way from getDecideCatalogEntries(). The hand rows listed first
- * reuse the existing createXConfig() factories in providerConfig.ts; the
+ * buildCatalogConfigOptions(); decide-kind catalog entries (laya, xor) are
+ * spread the same way from getDecideCatalogEntries(). The hand rows listed
+ * first reuse the existing createXConfig() factories in providerConfig.ts; the
  * remaining 5 (ollama, litellm, sagemaker, lm-studio, llamacpp) don't have a
  * factory and are defined inline using their real env var names. lm-studio
  * and llamacpp previously
@@ -172,7 +172,6 @@ export const EXTRA_PROVIDER_CONFIGS = {
     ideogram: createIdeogramConfig(),
     recraft: createRecraftConfig(),
     typesafe: createTypeSafeConfig(),
-    xor: createXorConfig(),
     "perplexity-decider": createPerplexityDeciderConfig(),
     "cloudflare-clef": createCloudflareClefConfig(),
     ...Object.fromEntries(getDecideCatalogEntries().map((e) => [e.id, buildCatalogConfigOptions(e)])),

package/dist/constants/enums.d.ts

@@ -103,6 +103,7 @@ export declare enum AIProviderName {
     WANDB_INFERENCE = "wandb-inference",
     XAI = "xai",
     XIAOMI_MIMO = "xiaomi-mimo",
+    XOR = "xor",
     Z_AI = "z-ai",
     COHERE = "cohere",
     REPLICATE = "replicate",
@@ -113,8 +114,6 @@ export declare enum AIProviderName {
     RECRAFT = "recraft",
     /** TypeSafe (Jev) — serves the `decide` inference type only. */
     TYPESAFE = "typesafe",
-    /** XOR (Juspay, open weights) — serves the `decide` inference type only. */
-    XOR = "xor",
     /**
      * Perplexity Decisions API (`pplx-decider-v1-27b`) — serves the `decide`
      * inference type only. Distinct from `PERPLEXITY`, the Sonar text provider.
@@ -1763,6 +1762,9 @@ export declare enum XaiModels {
 export declare enum XiaomiMimoModels {
     MIMO_V2_6_FLASH = "mimo-v2.6-flash"
 }
+export declare enum XorModels {
+    XOR_1_1 = "xor-1.1"
+}
 export declare enum ZAiModels {
     GLM_5_3 = "glm-5.3",
     GLM_5_3_FLASH = "glm-5.3-flash",
@@ -1788,13 +1790,6 @@ export declare enum TypeSafeModels {
     JEV_LATEST = "jev-latest",
     JEV_PREVIEW = "jev-preview"
 }
-/**
- * XOR decision models. Hand-written: XOR is a Tier-3 provider, so it is not in
- * the provider catalog and codegen never touches this.
- */
-export declare enum XorModels {
-    XOR_1_1 = "xor-1.1"
-}
 /**
  * Perplexity decision models. Hand-written: the Decisions API is a Tier-3
  * provider, so it is not in the provider catalog and codegen never touches

package/dist/constants/enums.js

@@ -108,6 +108,7 @@ export var AIProviderName;
     AIProviderName["WANDB_INFERENCE"] = "wandb-inference";
     AIProviderName["XAI"] = "xai";
     AIProviderName["XIAOMI_MIMO"] = "xiaomi-mimo";
+    AIProviderName["XOR"] = "xor";
     AIProviderName["Z_AI"] = "z-ai";
     // ── END GENERATED(provider-members) ──
     AIProviderName["COHERE"] = "cohere";
@@ -119,8 +120,6 @@ export var AIProviderName;
     AIProviderName["RECRAFT"] = "recraft";
     /** TypeSafe (Jev) — serves the `decide` inference type only. */
     AIProviderName["TYPESAFE"] = "typesafe";
-    /** XOR (Juspay, open weights) — serves the `decide` inference type only. */
-    AIProviderName["XOR"] = "xor";
     /**
      * Perplexity Decisions API (`pplx-decider-v1-27b`) — serves the `decide`
      * inference type only. Distinct from `PERPLEXITY`, the Sonar text provider.
@@ -2086,6 +2085,10 @@ export var XiaomiMimoModels;
 (function (XiaomiMimoModels) {
     XiaomiMimoModels["MIMO_V2_6_FLASH"] = "mimo-v2.6-flash";
 })(XiaomiMimoModels || (XiaomiMimoModels = {}));
+export var XorModels;
+(function (XorModels) {
+    XorModels["XOR_1_1"] = "xor-1.1";
+})(XorModels || (XorModels = {}));
 export var ZAiModels;
 (function (ZAiModels) {
     ZAiModels["GLM_5_3"] = "glm-5.3";
@@ -2114,14 +2117,6 @@ export var TypeSafeModels;
     TypeSafeModels["JEV_LATEST"] = "jev-latest";
     TypeSafeModels["JEV_PREVIEW"] = "jev-preview";
 })(TypeSafeModels || (TypeSafeModels = {}));
-/**
- * XOR decision models. Hand-written: XOR is a Tier-3 provider, so it is not in
- * the provider catalog and codegen never touches this.
- */
-export var XorModels;
-(function (XorModels) {
-    XorModels["XOR_1_1"] = "xor-1.1";
-})(XorModels || (XorModels = {}));
 /**
  * Perplexity decision models. Hand-written: the Decisions API is a Tier-3
  * provider, so it is not in the provider catalog and codegen never touches

package/dist/factories/providerDescriptors.js

@@ -1,5 +1,5 @@
 import { AIProviderName } from "../constants/enums.js";
-import { GoogleAIModels, OpenAIModels, AnthropicModels, VertexModels, OllamaModels, LiteLLMModels, NvidiaNimModels, OpenRouterModels, CohereModels, VoyageModels, TypeSafeModels, XorModels, PerplexityDeciderModels, CloudflareClefModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, ReplicateModels, } from "../constants/enums.js";
+import { GoogleAIModels, OpenAIModels, AnthropicModels, VertexModels, OllamaModels, LiteLLMModels, NvidiaNimModels, OpenRouterModels, CohereModels, VoyageModels, TypeSafeModels, PerplexityDeciderModels, CloudflareClefModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, ReplicateModels, } from "../constants/enums.js";
 import { API_KEY_FORMATS } from "../utils/providerConfig.js";
 import { getCatalogJsonEntries, getDecideCatalogEntries, catalogCredentialsKey, catalogEnvVar, } from "../providers/catalog/loader.js";
 import { DEFAULT_INFERENCE_KINDS } from "../types/index.js";
@@ -467,64 +467,23 @@ const HAND_DESCRIPTORS = [
         },
         setupUrl: "https://console.typesafe.ai/keys",
     },
-    // Decision precedence is decideOrder, not position: typesafe 1, laya 2 (from
-    // catalog/laya.json), xor 3, perplexity-decider 4, cloudflare-clef 5.
+    // Decision precedence is decideOrder, not position: typesafe 1, laya 2 and
+    // xor 3 (from catalog/laya.json and catalog/xor.json), perplexity-decider 4,
+    // cloudflare-clef 5.
     // resolveDefaultDecisionProvider() returns the first configured
     // DECISION_PROVIDERS entry in that order, so a host configured for several
     // keeps Jev for every built-in consumer and reaches the others only by
     // naming them. Renumbering silently changes which model routes, compacts
     // and plans for every such host.
-    {
-        name: AIProviderName.XOR,
-        aliases: [],
-        credentialsKey: "xor",
-        envVars: {
-            apiKey: "XOR_API_KEY",
-            baseURL: "XOR_BASE_URL",
-            // XOR has no built-in endpoint (it is reached at a deployment or behind a
-            // proxy), so a key is useless without a base URL. Both are required for
-            // it to count as configured, from the environment or credentials.xor.
-            extraRequired: ["XOR_BASE_URL"],
-            model: "XOR_MODEL",
-        },
-        defaultModel: XorModels.XOR_1_1,
-        // Serves only `decide`, like TypeSafe and Laya — the one declaration that
-        // keeps it out of every generation code path.
-        inferenceKinds: ["decide"],
-        decideOrder: 3,
-        toolSupport: "none",
-        localRuntime: false,
-        healthCheck: "env-only",
-        // Deliberately NO autoSelectPriority / autoSelectPreference /
-        // defaultHealthSweepPriority, for the same reason as TypeSafe above.
-        timeouts: { decideMs: 5000 },
-        // The deployment's prefill is 250,000 tokens, shared by the state, the
-        // questions and any media. 200,000 leaves 20% for those, and a state
-        // between this and the server's 4 MB would otherwise pass validation and
-        // fail inside the engine as a 500, which the base class would retry. There
-        // is no question cap (the server only limits a choice to 2..255 options).
-        // Non-ASCII text is charged per character because the four-characters-
-        // per-token estimate is several times too generous for CJK on this
-        // tokenizer. One token per character is a deliberate upper bound, to be
-        // replaced by a live measurement.
-        decisionLimits: {
-            maxStateTokens: 200_000,
-            nonAsciiTokensPerChar: 1,
-            // The server takes 1..8 images and one video, and refuses a body over
-            // 8 MB.
-            media: { maxImages: 8, video: true, maxRequestBytes: 8 * 1024 * 1024 },
-        },
-        setupUrl: "https://huggingface.co/juspay/xor",
-    },
     {
         name: AIProviderName.PERPLEXITY_DECIDER,
         aliases: [],
         credentialsKey: "perplexityDecider",
         envVars: {
             // The same key the `perplexity` text provider reads, so an ambient
-            // PERPLEXITY_API_KEY configures this provider too. It sits after XOR in
-            // this list, which is what keeps that from displacing any other decision
-            // provider a host has configured.
+            // PERPLEXITY_API_KEY configures this provider too. Its decideOrder (4)
+            // sits after XOR's (3), which is what keeps that from displacing any other
+            // decision provider a host has configured.
             apiKey: "PERPLEXITY_API_KEY",
             // Optional: the API is hosted at a public endpoint, so a key alone is
             // enough and there is no extraRequired.

package/dist/factories/providerRegistry.js

@@ -1,6 +1,6 @@
 import { ProviderFactory } from "./providerFactory.js";
 import { logger } from "../utils/logger.js";
-import { AIProviderName, GoogleAIModels, OpenAIModels, AnthropicModels, VertexModels, OllamaModels, LiteLLMModels, NvidiaNimModels, OpenRouterModels, CohereModels, VoyageModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, TypeSafeModels, XorModels, PerplexityDeciderModels, CloudflareClefModels, ReplicateModels, } from "../constants/enums.js";
+import { AIProviderName, GoogleAIModels, OpenAIModels, AnthropicModels, VertexModels, OllamaModels, LiteLLMModels, NvidiaNimModels, OpenRouterModels, CohereModels, VoyageModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, TypeSafeModels, PerplexityDeciderModels, CloudflareClefModels, ReplicateModels, } from "../constants/enums.js";
 import { PROVIDER_DESCRIPTORS_BY_NAME } from "./providerDescriptors.js";
 import { OPENAI_COMPAT_CATALOG } from "../providers/openaiCompatCatalog.js";
 import { providerChoicesFor } from "./mediaHandlerCatalog.js";
@@ -250,14 +250,6 @@ export class ProviderRegistry {
                 const { TypeSafeProvider } = await import("../providers/typesafe.js");
                 return new TypeSafeProvider(modelName, sdk, undefined, typesafeCreds);
             }, process.env.TYPESAFE_MODEL || TypeSafeModels.JEV_LATEST, ["jev", "typesafe-ai"], PROVIDER_DESCRIPTORS_BY_NAME.get(AIProviderName.TYPESAFE));
-            // Register XOR — the third `decide` provider, reached at whatever base
-            // URL its config names. Its descriptor declares inferenceKinds:
-            // ["decide"], so nothing in the generation fallback chain can reach it.
-            ProviderFactory.registerProvider(AIProviderName.XOR, async (modelName, _providerName, sdk, _region, credentials) => {
-                const xorCreds = credentials;
-                const { XorProvider } = await import("../providers/xor.js");
-                return new XorProvider(modelName, sdk, undefined, xorCreds);
-            }, process.env.XOR_MODEL || XorModels.XOR_1_1, [], PROVIDER_DESCRIPTORS_BY_NAME.get(AIProviderName.XOR));
             // Register the Perplexity Decisions API — a `decide` provider reached at
             // Perplexity's public endpoint. Its descriptor declares inferenceKinds:
             // ["decide"], so nothing in the generation fallback chain can reach it,

package/dist/models/manifestRegistry.js

@@ -20,7 +20,6 @@ import { llamacppManifest } from "./manifests/llamacpp.js";
 import { xaiManifest } from "./manifests/xai.js";
 import { groqManifest } from "./manifests/groq.js";
 import { typesafeManifest } from "./manifests/typesafe.js";
-import { xorManifest } from "./manifests/xor.js";
 import { perplexityDeciderManifest } from "./manifests/perplexityDecider.js";
 import { cloudflareClefManifest } from "./manifests/cloudflareClef.js";
 import { cohereManifest } from "./manifests/cohere.js";
@@ -160,7 +159,6 @@ export const MANIFEST_REGISTRY = {
     groq: groqManifest,
     typesafe: typesafeManifest,
     ...Object.fromEntries(getDecideCatalogEntries().map((e) => [e.id, buildDecideManifest(e)])),
-    xor: xorManifest,
     "perplexity-decider": perplexityDeciderManifest,
     "cloudflare-clef": cloudflareClefManifest,
     cerebras: catalogManifest("cerebras"),

package/dist/neurolink.d.ts

@@ -1,10 +1,3 @@
-/**
- * NeuroLink - Unified AI Interface with Real MCP Tool Integration
- *
- * REDESIGNED FALLBACK CHAIN - NO CIRCULAR DEPENDENCIES
- * Enhanced AI provider system with natural MCP tool access.
- * Uses real MCP infrastructure for tool discovery and execution.
- */
 import type { DecisionCallerOptions, DecisionLimitsQuery, DecisionLimitsReading, DecisionOptions, DecisionResult, AgentDefinition, AgentNetworkConfig, AgentRunOptions, AgentRunOutcome, AgentToolRegistrationOptions, IsolatedAgentDefinition, NetworkExecutionInput, NetworkExecutionOptions, NetworkExecutionResult, NetworkStreamChunk, WorkerInstanceOptions } from "./types/index.js";
 import type { CompactionConfig, CompactionResult, ArtifactStore, SpanData, ObservabilityConfig, MetricsSummary, MCPToolAnnotations, TraceView, AuthenticatedContext, AuthProvider, JsonObject, NeuroLinkEvents, TypedEventEmitter, MCPEnhancementsConfig, NeuroLinkAuthConfig, NeurolinkConstructorConfig, ChatMessage, ConversationSeedMessage, ConversationSeedOptions, RecordConversationTurnOptions, RecordConversationTurnResult, ExternalMCPOperationResult, ExternalMCPServerInstance, ExternalMCPServerHealth, ExternalMCPToolInfo, GenerateOptions, GenerateResult, ProviderStatus, TextGenerationOptions, TextGenerationResult, MCPExecutableTool, MCPServerInfo, MCPStatus, StreamOptions, StreamResult, ToolExecutionContext, ToolExecutionSummary, ToolInfo, ToolRegistrationOptions, BatchOperationResult, StreamGenerationEndContext, ToolRoutingServerDescriptor, ToolDedupConfig, ToolConfig, KnowledgeEngineStatus, ChecklistState, ArtifactPageRequest, BankArtifactOptions, BankedArtifactRef, DelegateCollectRequest, DelegateCollectResult, DelegateHandle, DelegateRegistrationOptions, DelegateSpawnOptions, BackgroundCommandHandle, BackgroundCommandOptions, BackgroundCommandOutputPage, BackgroundCommandPageRequest, BackgroundCommandPolicy, BackgroundCommandStatus, GitToolResult, GitToolsetOptions } from "./types/index.js";
 import { ConversationMemoryManager } from "./core/conversationMemoryManager.js";
@@ -612,6 +605,12 @@ export declare class NeuroLink {
      * @returns Modified options with orchestrated provider marked in context, or empty object if validation fails
      */
     private applyStreamOrchestration;
+    /**
+     * Read a nonempty orchestration preference only when provider is absent/auto.
+     * An explicit provider always takes precedence over incoming or generated
+     * context; empty, non-string and auto preferences are ignored.
+     */
+    private getOrchestratedPreferredProvider;
     /**
      * MAIN ENTRY POINT: Enhanced generate method with new function signature
      * Replaces both generateText and legacy methods
@@ -937,6 +936,7 @@ export declare class NeuroLink {
      *
      * The router is skipped when:
      *   - no `requestRouter` is configured on this instance, or
+     *   - a valid orchestration preference already owns the provider/model pair, or
      *   - the caller explicitly set both `options.provider` AND `options.model`
      *     (we only skip the router if BOTH are set; a caller setting only one
      *     still lets the router fill in the other field).
@@ -1118,6 +1118,8 @@ export declare class NeuroLink {
     private generateWithMCPProvider;
     /**
      * Direct provider generation (no MCP, no recursion)
+     * An explicit provider precedes orchestration context; otherwise a valid
+     * orchestration preference is tried before the existing provider fallback list.
      */
     private directProviderGeneration;
     /**
@@ -1324,7 +1326,9 @@ export declare class NeuroLink {
      */
     private emitStreamStartEvents;
     /**
-     * Create MCP stream
+     * Create the provider-backed MCP stream with the existing tool lifecycle.
+     * An explicit provider precedes a valid orchestration context preference;
+     * absent either, getBestProvider retains its normal automatic selection.
      */
     private createMCPStream;
     /**

package/dist/neurolink.js

@@ -1,3 +1,4 @@
+import { beginOllamaDiagnostic, captureOllamaDiagnosticContext, completeOllamaDiagnostic, getOllamaDiagnosticErrorMessage, } from "./utils/ollamaDiagnostics.js";
 /**
  * NeuroLink - Unified AI Interface with Real MCP Tool Integration
  *
@@ -203,13 +204,6 @@ const isNonRetryableForPool = sharedIsNonRetryableForPool;
  */
 const MEMORY_WRITE_TIMEOUT_MS = 30_000;
 /** Native tags share the health check's endpoint precedence and origin. */
-function getOllamaTagsUrl() {
-    const baseUrl = process.env.OLLAMA_BASE_URL ||
-        process.env.OLLAMA_API_BASE ||
-        "http://localhost:11434";
-    // Native /api/tags is at the origin even if the compatible base ends /v1.
-    return new URL("/api/tags", baseUrl).toString();
-}
 /**
  * NeuroLink - Universal AI Development Platform
  *
@@ -2737,8 +2731,17 @@ Current user's request: ${currentInput}`;
             }
             // For Ollama, check if service is running and model is available
             if (route.provider === "ollama") {
+                let ollamaAttempt;
+                let ollamaOutcome = "not-run";
                 try {
-                    const response = await fetch(getOllamaTagsUrl(), {
+                    const ollamaContext = captureOllamaDiagnosticContext();
+                    const gate = beginOllamaDiagnostic(ollamaContext);
+                    if (gate.kind === "suppressed") {
+                        return {};
+                    }
+                    ollamaAttempt = gate;
+                    ollamaOutcome = "transport-failure";
+                    const response = await fetch(ollamaContext.url, {
                         method: "GET",
                         signal: AbortSignal.timeout(2000),
                     });
@@ -2752,6 +2755,7 @@ Current user's request: ${currentInput}`;
                         });
                         return {}; // Return empty object to preserve existing fallback behavior
                     }
+                    ollamaOutcome = "protocol-failure";
                     const responseData = await response.json();
                     const models = responseData?.models;
                     // Runtime-safe guard: ensure models is an array with valid objects
@@ -2762,6 +2766,7 @@ Current user's request: ${currentInput}`;
                         });
                         return {}; // Return empty object for fallback behavior
                     }
+                    ollamaOutcome = "success";
                     // Filter and validate models before comparison
                     const validModels = models.filter((m) => m && typeof m === "object" && typeof m.name === "string");
                     const targetModel = route.model;
@@ -2807,12 +2812,17 @@ Current user's request: ${currentInput}`;
                         routedProvider: route.provider,
                         routedModel: route.model,
                         reason: error instanceof Error
-                            ? error.message
+                            ? getOllamaDiagnosticErrorMessage(error)
                             : "Ollama service check failed",
                         orchestrationTime: `${Date.now() - startTime}ms`,
                     });
                     return {}; // Return empty object to preserve existing fallback behavior
                 }
+                finally {
+                    if (ollamaAttempt) {
+                        completeOllamaDiagnostic(ollamaAttempt, ollamaOutcome);
+                    }
+                }
             }
             logger.debug("Orchestration route determined", {
                 taskType: classification.type,
@@ -2875,8 +2885,17 @@ Current user's request: ${currentInput}`;
             }
             // For Ollama, check if service is running and model is available
             if (route.provider === "ollama") {
+                let ollamaAttempt;
+                let ollamaOutcome = "not-run";
                 try {
-                    const response = await fetch(getOllamaTagsUrl(), {
+                    const ollamaContext = captureOllamaDiagnosticContext();
+                    const gate = beginOllamaDiagnostic(ollamaContext);
+                    if (gate.kind === "suppressed") {
+                        return {};
+                    }
+                    ollamaAttempt = gate;
+                    ollamaOutcome = "transport-failure";
+                    const response = await fetch(ollamaContext.url, {
                         method: "GET",
                         signal: AbortSignal.timeout(2000),
                     });
@@ -2890,6 +2909,7 @@ Current user's request: ${currentInput}`;
                         });
                         return {}; // Return empty object to preserve existing fallback behavior
                     }
+                    ollamaOutcome = "protocol-failure";
                     const responseData = await response.json();
                     const models = responseData?.models;
                     // Runtime-safe guard: ensure models is an array with valid objects
@@ -2900,6 +2920,7 @@ Current user's request: ${currentInput}`;
                         });
                         return {}; // Return empty object for fallback behavior
                     }
+                    ollamaOutcome = "success";
                     // Filter and validate models before comparison
                     const validModels = models.filter((m) => m && typeof m === "object" && typeof m.name === "string");
                     const targetModel = route.model;
@@ -2945,12 +2966,17 @@ Current user's request: ${currentInput}`;
                         routedProvider: route.provider,
                         routedModel: route.model,
                         reason: error instanceof Error
-                            ? error.message
+                            ? getOllamaDiagnosticErrorMessage(error)
                             : "Ollama service check failed",
                         orchestrationTime: `${Date.now() - startTime}ms`,
                     });
                     return {}; // Return empty object to preserve existing fallback behavior
                 }
+                finally {
+                    if (ollamaAttempt) {
+                        completeOllamaDiagnostic(ollamaAttempt, ollamaOutcome);
+                    }
+                }
             }
             logger.debug("Stream orchestration route determined", {
                 taskType: classification.type,
@@ -2978,6 +3004,22 @@ Current user's request: ${currentInput}`;
             throw error;
         }
     }
+    /**
+     * Read a nonempty orchestration preference only when provider is absent/auto.
+     * An explicit provider always takes precedence over incoming or generated
+     * context; empty, non-string and auto preferences are ignored.
+     */
+    getOrchestratedPreferredProvider(options) {
+        if (options.provider && options.provider !== "auto") {
+            return undefined;
+        }
+        const preferred = options.context?.__orchestratedPreferredProvider;
+        return typeof preferred === "string" &&
+            preferred.length > 0 &&
+            preferred !== "auto"
+            ? preferred
+            : undefined;
+    }
     /**
      * MAIN ENTRY POINT: Enhanced generate method with new function signature
      * Replaces both generateText and legacy methods
@@ -4339,6 +4381,7 @@ Current user's request: ${currentInput}`;
      *
      * The router is skipped when:
      *   - no `requestRouter` is configured on this instance, or
+     *   - a valid orchestration preference already owns the provider/model pair, or
      *   - the caller explicitly set both `options.provider` AND `options.model`
      *     (we only skip the router if BOTH are set; a caller setting only one
      *     still lets the router fill in the other field).
@@ -4358,8 +4401,11 @@ Current user's request: ${currentInput}`;
         }
         // Stand down if the classifier router already selected a provider/model on
         // this turn — one explicit precedence order: classifierRouter wins.
-        if (options.context
-            ?.__classifierRouted) {
+        if (options.context?.__classifierRouted) {
+            return;
+        }
+        // Keep an orchestration-selected model with the provider that owns it.
+        if (this.getOrchestratedPreferredProvider(options)) {
             return;
         }
         // Skip if the caller explicitly set both provider AND model — they know
@@ -6577,6 +6623,8 @@ Current user's request: ${currentInput}`;
     }
     /**
      * Direct provider generation (no MCP, no recursion)
+     * An explicit provider precedes orchestration context; otherwise a valid
+     * orchestration preference is tried before the existing provider fallback list.
      */
     async directProviderGeneration(options) {
         const startTime = Date.now();
@@ -6593,13 +6641,8 @@ Current user's request: ${currentInput}`;
             "ollama",
         ];
         const requestedProvider = options.provider === "auto" ? undefined : options.provider;
-        // Check for orchestrated preferred provider in context
-        const preferredOrchestrated = options.context &&
-            typeof options.context === "object" &&
-            "__orchestratedPreferredProvider" in options.context
-            ? options.context
-                .__orchestratedPreferredProvider
-            : undefined;
+        // Read the shared preference while keeping an explicit provider first.
+        const preferredOrchestrated = this.getOrchestratedPreferredProvider(options);
         // Build provider list with orchestrated preference first, then fallback to full list
         const tryProviders = preferredOrchestrated
             ? [
@@ -9425,11 +9468,16 @@ Current user's request: ${currentInput}`;
         this.emitter.emit("message", `Starting ${options.provider || "auto"} stream...`);
     }
     /**
-     * Create MCP stream
+     * Create the provider-backed MCP stream with the existing tool lifecycle.
+     * An explicit provider precedes a valid orchestration context preference;
+     * absent either, getBestProvider retains its normal automatic selection.
      */
     async createMCPStream(options) {
-        // Simplified placeholder - in the actual implementation this would contain the complex MCP stream logic
-        const providerName = await getBestProvider(options.provider);
+        // Consume the validated orchestration choice after an explicit provider.
+        const requestedProvider = options.provider && options.provider !== "auto"
+            ? options.provider
+            : this.getOrchestratedPreferredProvider(options);
+        const providerName = await getBestProvider(requestedProvider);
         const provider = await AIProviderFactory.createProvider(providerName, options.model, !options.disableTools, // Pass disableTools as inverse of enableMCP
         this, // Pass SDK instance
         options.region, // Pass region parameter
@@ -11900,14 +11948,24 @@ Current user's request: ${currentInput}`;
                 }
                 // Special handling for Ollama
                 if (providerName === "ollama") {
+                    let ollamaAttempt;
+                    let ollamaOutcome = "not-run";
                     try {
-                        const response = await fetch(getOllamaTagsUrl(), {
+                        const ollamaContext = captureOllamaDiagnosticContext();
+                        const gate = beginOllamaDiagnostic(ollamaContext);
+                        if (gate.kind === "suppressed") {
+                            throw new Error(gate.reason);
+                        }
+                        ollamaAttempt = gate;
+                        ollamaOutcome = "transport-failure";
+                        const response = await fetch(ollamaContext.url, {
                             method: "GET",
                             signal: AbortSignal.timeout(PROVIDER_TIMEOUTS.AUTH_MS),
                         });
                         if (!response.ok) {
                             throw new Error("Ollama service not responding");
                         }
+                        ollamaOutcome = "protocol-failure";
                         const responseData = await response.json();
                         const models = responseData?.models;
                         // Runtime-safe guard: ensure models is an array with valid objects
@@ -11918,6 +11976,7 @@ Current user's request: ${currentInput}`;
                             });
                             throw new Error("Invalid models format from Ollama API");
                         }
+                        ollamaOutcome = "success";
                         // Filter and validate models before comparison
                         const validModels = models.filter((m) => m && typeof m === "object" && typeof m.name === "string");
                         if (validModels.length > 0) {
@@ -11948,11 +12007,16 @@ Current user's request: ${currentInput}`;
                             configured: false,
                             authenticated: false,
                             error: error instanceof Error
-                                ? error.message
+                                ? getOllamaDiagnosticErrorMessage(error)
                                 : "Ollama service not running",
                             responseTime: Date.now() - startTime,
                         };
                     }
+                    finally {
+                        if (ollamaAttempt) {
+                            completeOllamaDiagnostic(ollamaAttempt, ollamaOutcome);
+                        }
+                    }
                 }
                 // Test other providers with actual generation call. The losing
                 // side of the race is aborted in `finally` below — previously the

package/dist/providers/catalog/index.generated.d.ts

@@ -2,4 +2,4 @@ import type { DecideCatalogJson, ProviderCatalogJson } from "../../types/index.j
 export declare const CATALOG_JSON_ENTRIES: ProviderCatalogJson[];
 export declare const CATALOG_PROVIDER_IDS: readonly ["a2agent", "abliteration", "above-dev", "ai21", "aiand", "aihubmix", "aionlabs", "ambient", "api-route", "arcee", "atlas-cloud", "avian-io", "baidu-qianfan", "baseten", "bee-heossi", "byteplus-modelark", "bytez", "cerebras", "charm-hyper", "chutes", "cloudflare", "crusoe", "dashscope", "deepinfra", "deepseek", "empiriolabs", "featherless-ai", "fireworks", "friendli", "gmicloud", "gradientai", "groq", "hetzner-inference", "huggingface", "inception-labs", "inceptron", "inco", "inference-net", "io-intelligence", "koscompute", "lemonfox-ai", "lilac", "llmtech", "longcat", "mancer", "meta-model-api", "minimax", "mistral", "moark", "modelscope", "moonshot-ai", "morph", "naver-clova-studio", "nebius", "neuralwatt", "novita", "ovhcloud", "parasail", "pareto-inference", "perplexity", "poolside", "prime-intellect", "reka", "runinfra", "sakana-ai", "sambanova", "sarvam", "scaleway", "siliconflow", "stackit", "stepfun", "subconscious", "synthetic", "telnyx", "thinking-machines", "tinfoil", "together-ai", "umans-ai", "upstage", "venice-ai", "vispark", "vultr-inference", "wafer", "wandb-inference", "xai", "xiaomi-mimo", "z-ai"];
 export declare const DECIDE_CATALOG_JSON_ENTRIES: DecideCatalogJson[];
-export declare const DECIDE_CATALOG_PROVIDER_IDS: readonly ["laya"];
+export declare const DECIDE_CATALOG_PROVIDER_IDS: readonly ["laya", "xor"];

package/dist/providers/catalog/index.generated.js

@@ -87,6 +87,7 @@ import waferJson from "./wafer.json" with { type: "json" };
 import wandbInferenceJson from "./wandb-inference.json" with { type: "json" };
 import xaiJson from "./xai.json" with { type: "json" };
 import xiaomiMimoJson from "./xiaomi-mimo.json" with { type: "json" };
+import xorJson from "./xor.json" with { type: "json" };
 import zAiJson from "./z-ai.json" with { type: "json" };
 import { isDecideCatalogEntry, parseCatalogEntry } from "./schema.js";
 const asDecideEntry = (raw) => {
@@ -274,5 +275,8 @@ export const CATALOG_PROVIDER_IDS = [
     "xiaomi-mimo",
     "z-ai",
 ];
-export const DECIDE_CATALOG_JSON_ENTRIES = [layaJson].map(asDecideEntry);
-export const DECIDE_CATALOG_PROVIDER_IDS = ["laya"];
+export const DECIDE_CATALOG_JSON_ENTRIES = [
+    layaJson,
+    xorJson,
+].map(asDecideEntry);
+export const DECIDE_CATALOG_PROVIDER_IDS = ["laya", "xor"];