@juspay/neurolink 12.12.2 → 12.12.3

package/dist/cli/loop/optionsSchema.js

@@ -118,6 +118,10 @@ export const textGenerationOptionsSchema = {
         type: "boolean",
         description: "Disable tool result caching for this request (overrides global mcp.cache.enabled).",
     },
+    disableInternalFallback: {
+        type: "boolean",
+        description: "Own fallback order yourself: skip NeuroLink's provider-priority walk and the catalog model fallback, so an invalid model or unavailable provider surfaces as its own error.",
+    },
     disableToolCallRepair: {
         type: "boolean",
         description: "Disable the schema-driven tool call repair mechanism (near-miss tool names, mis-typed arguments). Repair is enabled by default.",

package/dist/core/baseProvider.js

@@ -1299,11 +1299,10 @@ export class BaseProvider {
             const otelSpanState = { ended: false };
             return await context.with(activeCtx, async () => this.runGenerateInActiveContext(options, startTime, otelSpan, otelSpanState));
         };
-        // TextGenerationOptions does not declare the flag (it lives on
-        // StreamOptions), but callers that own fallback order pass it on both
-        // paths, so honour it here as well.
-        const callerOwnsFallback = "disableInternalFallback" in options &&
-            options.disableInternalFallback === true;
+        // Callers that own fallback order (providerFallback / modelChain callers,
+        // or a router that retries on its own) pass the flag on both paths;
+        // TextGenerationOptions declares it, so a plain read is enough here.
+        const callerOwnsFallback = options.disableInternalFallback === true;
         return await this.runGenerateWithModelFallback(attempt, callerOwnsFallback);
     }
     /**

package/dist/neurolink.js

@@ -4121,6 +4121,11 @@ Current user's request: ${currentInput}`;
             disableTools: options.disableTools,
             toolFilter: options.toolFilter,
             excludeTools: options.excludeTools,
+            // This explicit field list is the only road into the provider, so a
+            // flag left out here never reaches BaseProvider — which is exactly how
+            // disableInternalFallback was dropped on generate() while stream()
+            // (which spreads its options) honoured it.
+            disableInternalFallback: options.disableInternalFallback,
             maxSteps: options.maxSteps,
             toolChoice: options.toolChoice,
             prepareStep: options.prepareStep,
@@ -5930,11 +5935,23 @@ Current user's request: ${currentInput}`;
             : requestedProvider
                 ? [requestedProvider]
                 : providerPriority;
+        // The caller owns fallback order (a providerFallback / modelChain caller,
+        // or a router that retries on its own): bound the walk to its first
+        // candidate so an unavailable provider surfaces as its own error instead
+        // of a silent switch. An explicit provider is already a one-element
+        // list; this only changes the "auto" and orchestrated-preference walks.
+        const providersToTry = options.disableInternalFallback === true
+            ? tryProviders.slice(0, 1)
+            : tryProviders;
+        // Caller-owned fallback never enters tryProviders: providerFallback and
+        // modelChain are walked by runWithFallbackOrchestration around the public
+        // generate() call, and a configured ModelPool is consumed by the block
+        // above. Slicing here therefore never clips a caller's own list.
         logger.debug(`[${functionTag}] Starting direct generation`, {
             requestedProvider: requestedProvider || "auto",
             preferredOrchestrated: preferredOrchestrated || "none",
-            tryProviders,
-            allowFallback: !requestedProvider || !!preferredOrchestrated,
+            tryProviders: providersToTry,
+            allowFallback: providersToTry.length > 1,
         });
         // ─── ModelPool path ──────────────────────────────────────────────────────
         // When a ModelPool is configured, source the candidate sequence from the
@@ -6080,7 +6097,7 @@ Current user's request: ${currentInput}`;
         // ─── End ModelPool path ───────────────────────────────────────────────────
         let lastError = null;
         // Try each provider in order
-        for (const providerName of tryProviders) {
+        for (const providerName of providersToTry) {
             if (options.abortSignal?.aborted) {
                 throw new DOMException("The operation was aborted", "AbortError");
             }
@@ -6437,7 +6454,7 @@ Current user's request: ${currentInput}`;
         // All providers failed
         const responseTime = Date.now() - startTime;
         logger.error(`[${functionTag}] All providers failed`, {
-            triedProviders: tryProviders,
+            triedProviders: providersToTry,
             lastError: lastError?.message,
             responseTime,
         });

package/dist/types/generate.d.ts

@@ -459,6 +459,19 @@ export type GenerateOptions = {
     skipToolPromptInjection?: boolean;
     /** Disable tool result caching for this request (overrides global mcp.cache.enabled) */
     disableToolCache?: boolean;
+    /**
+     * Disable NeuroLink's internal fallback for this request: the static
+     * provider-priority walk that runs when no provider was requested, and the
+     * catalog model-fallback walk a provider performs when its model is
+     * rejected as invalid. Callers that own fallback order (a caller-supplied
+     * `providerFallback` / `modelChain`, or a router that retries on its own,
+     * as the Claude proxy does for its streams) set this so an invalid model
+     * or an unavailable provider surfaces as exactly that.
+     * A configured `ModelPool`, `providerFallback` and `modelChain` are the
+     * caller's own fallback and are unaffected. Mirrors the same flag on
+     * `StreamOptions`.
+     */
+    disableInternalFallback?: boolean;
     /** Maximum number of tool execution steps (default: 200) */
     maxSteps?: number;
     /**
@@ -1204,6 +1217,14 @@ export type TextGenerationOptions = {
     excludeTools?: string[];
     /** Disable tool result caching for this request (overrides global mcp.cache.enabled) */
     disableToolCache?: boolean;
+    /**
+     * Caller owns fallback order. Read in two places: `directProviderGeneration`
+     * bounds its static provider-priority walk to one candidate, and
+     * `BaseProvider.generate()` skips the catalog model-fallback walk so an
+     * invalid-model error surfaces as itself. Mapped from
+     * `GenerateOptions.disableInternalFallback`.
+     */
+    disableInternalFallback?: boolean;
     /**
      * Tool choice configuration for the generation.
      * Controls whether and which tools the model must call.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@juspay/neurolink",
-  "version": "12.12.2",
+  "version": "12.12.3",
   "packageManager": "pnpm@10.15.1",
   "description": "TypeScript AI SDK with 24+ LLM providers behind one consistent API. MCP-native (connect any MCP server), voice TTS/STT/realtime, RAG, agents, memory, context compaction. OpenAI · Anthropic · Gemini · Bedrock · Azure · Ollama · DeepSeek · NVIDIA NIM and more.",
   "author": {