@juspay/neurolink 12.9.6 → 12.11.0

package/dist/constants/enums.d.ts

@@ -18,14 +18,19 @@ export declare enum AIProviderName {
     NVIDIA_NIM = "nvidia-nim",
     LM_STUDIO = "lm-studio",
     LLAMACPP = "llamacpp",
+    BASETEN = "baseten",
     CEREBRAS = "cerebras",
     CLOUDFLARE = "cloudflare",
     FIREWORKS = "fireworks",
+    GMICLOUD = "gmicloud",
     GROQ = "groq",
+    INCEPTION_LABS = "inception-labs",
+    IO_INTELLIGENCE = "io-intelligence",
     MISTRAL = "mistral",
     PERPLEXITY = "perplexity",
     SAMBANOVA = "sambanova",
     TOGETHER_AI = "together-ai",
+    UPSTAGE = "upstage",
     XAI = "xai",
     COHERE = "cohere",
     REPLICATE = "replicate",
@@ -809,6 +814,24 @@ export declare enum ReplicateModels {
  * authentication failures during ongoing operations
  */
 export declare const TOKEN_EXPIRY_BUFFER_MS: number;
+export declare enum BasetenModels {
+    OPENAI_GPT_OSS_120B = "openai/gpt-oss-120b",
+    ZAI_ORG_GLM_4_7 = "zai-org/GLM-4.7",
+    MOONSHOTAI_KIMI_K2_6 = "moonshotai/Kimi-K2.6",
+    DEEPSEEK_AI_DEEPSEEK_V4_PRO = "deepseek-ai/DeepSeek-V4-Pro",
+    NVIDIA_NVIDIA_NEMOTRON_3_ULTRA_550B_A55B = "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B",
+    ZAI_ORG_GLM_5_2 = "zai-org/GLM-5.2",
+    MOONSHOTAI_KIMI_K2_7_CODE = "moonshotai/Kimi-K2.7-Code",
+    DEEPSEEK_AI_DEEPSEEK_V4_FLASH_0731 = "deepseek-ai/DeepSeek-V4-Flash-0731",
+    THINKINGMACHINES_INKLING = "thinkingmachines/inkling",
+    ZAI_ORG_GLM_5_2_FAST = "zai-org/GLM-5.2-Fast",
+    MOONSHOTAI_KIMI_K3 = "moonshotai/Kimi-K3",
+    THINKINGMACHINES_INKLING_SMALL = "thinkingmachines/inkling-small",
+    DEEPSEEK_AI_DEEPSEEK_V4_PRO_0813 = "deepseek-ai/DeepSeek-V4-Pro-0813",
+    ZAI_ORG_GLM_5_3_FLASH = "zai-org/GLM-5.3-Flash",
+    ZAI_ORG_GLM_5_3 = "zai-org/GLM-5.3",
+    ZAI_ORG_GLM_5_3_FAST = "zai-org/GLM-5.3-Fast"
+}
 export declare enum CerebrasModels {
     GPT_OSS_120B = "gpt-oss-120b",
     GEMMA_4_31B = "gemma-4-31b"
@@ -838,6 +861,9 @@ export declare enum FireworksModels {
     LLAMA_V3P2_11B_VISION_INSTRUCT = "accounts/fireworks/models/llama-v3p2-11b-vision-instruct",
     PHI_3_VISION_128K_INSTRUCT = "accounts/fireworks/models/phi-3-vision-128k-instruct"
 }
+export declare enum GmicloudModels {
+    MINIMAXAI_MINIMAX_M3 = "MiniMaxAI/MiniMax-M3"
+}
 export declare enum GroqModels {
     GPT_OSS_120B = "openai/gpt-oss-120b",
     GPT_OSS_20B = "openai/gpt-oss-20b",
@@ -854,6 +880,45 @@ export declare enum GroqModels {
     LLAMA_3_2_90B_VISION_PREVIEW = "llama-3.2-90b-vision-preview",
     LLAMA_3_2_11B_VISION_PREVIEW = "llama-3.2-11b-vision-preview"
 }
+export declare enum InceptionLabsModels {
+    MERCURY_2 = "mercury-2"
+}
+export declare enum IoIntelligenceModels {
+    ZAI_ORG_GLM_5_3_FLASH = "zai-org/GLM-5.3-Flash",
+    ZAI_ORG_GLM_5_3 = "zai-org/GLM-5.3",
+    QWEN_QWEN3_8_27B = "Qwen/Qwen3.8-27B",
+    DEEPSEEK_AI_DEEPSEEK_V4_FLASH_0731 = "deepseek-ai/DeepSeek-V4-Flash-0731",
+    MOONSHOTAI_KIMI_K3 = "moonshotai/Kimi-K3",
+    XIAOMIMIMO_MIMO_V2_5 = "XiaomiMiMo/MiMo-V2.5",
+    ZAI_ORG_GLM_5_2 = "zai-org/GLM-5.2",
+    MOONSHOTAI_KIMI_K2_7_CODE = "moonshotai/Kimi-K2.7-Code",
+    QWEN_QWEN3_6_35B_A3B = "Qwen/Qwen3.6-35B-A3B",
+    QWEN_QWEN3_6_27B = "Qwen/Qwen3.6-27B",
+    MINIMAXAI_MINIMAX_M2_7 = "MiniMaxAI/MiniMax-M2.7",
+    DEEPSEEK_AI_DEEPSEEK_V4_FLASH = "deepseek-ai/DeepSeek-V4-Flash",
+    DEEPSEEK_AI_DEEPSEEK_V4_PRO = "deepseek-ai/DeepSeek-V4-Pro",
+    MOONSHOTAI_KIMI_K2_6 = "moonshotai/Kimi-K2.6",
+    ZAI_ORG_GLM_5_1 = "zai-org/GLM-5.1",
+    MINIMAXAI_MINIMAX_M2_5 = "MiniMaxAI/MiniMax-M2.5",
+    MOONSHOTAI_KIMI_K2_5 = "moonshotai/Kimi-K2.5",
+    ZAI_ORG_GLM_5 = "zai-org/GLM-5",
+    DEEPSEEK_AI_DEEPSEEK_V3_2 = "deepseek-ai/DeepSeek-V3.2",
+    MOONSHOTAI_KIMI_K2_THINKING = "moonshotai/Kimi-K2-Thinking",
+    ZAI_ORG_GLM_4_5_AIR = "zai-org/GLM-4.5-Air",
+    GOOGLE_GEMMA_4_26B_A4B_IT = "google/gemma-4-26b-a4b-it",
+    ZAI_ORG_GLM_4_7_FLASH = "zai-org/GLM-4.7-Flash",
+    ZAI_ORG_GLM_4_7 = "zai-org/GLM-4.7",
+    MOONSHOTAI_KIMI_K2_INSTRUCT_0905 = "moonshotai/Kimi-K2-Instruct-0905",
+    OPENAI_GPT_OSS_120B = "openai/gpt-oss-120b",
+    DEEPSEEK_AI_DEEPSEEK_R1_0528 = "deepseek-ai/DeepSeek-R1-0528",
+    ZAI_ORG_GLM_4_6 = "zai-org/GLM-4.6",
+    QWEN_QWEN3_NEXT_80B_A3B_INSTRUCT = "Qwen/Qwen3-Next-80B-A3B-Instruct",
+    INTEL_QWEN3_CODER_480B_A35B_INSTRUCT_INT4_MIXED_AR = "Intel/Qwen3-Coder-480B-A35B-Instruct-int4-mixed-ar",
+    META_LLAMA_LLAMA_4_MAVERICK_17B_128E_INSTRUCT_FP8 = "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8",
+    MISTRALAI_MISTRAL_NEMO_INSTRUCT_2407 = "mistralai/Mistral-Nemo-Instruct-2407",
+    OPENAI_GPT_OSS_20B = "openai/gpt-oss-20b",
+    META_LLAMA_LLAMA_3_3_70B_INSTRUCT = "meta-llama/Llama-3.3-70B-Instruct"
+}
 export declare enum MistralModels {
     MISTRAL_LARGE_LATEST = "mistral-large-latest",
     MISTRAL_LARGE_2512 = "mistral-large-2512",
@@ -918,6 +983,18 @@ export declare enum TogetherAIModels {
     GEMMA_2_27B_IT = "google/gemma-2-27b-it",
     WIZARDLM_2_8X22B = "microsoft/WizardLM-2-8x22B"
 }
+export declare enum UpstageModels {
+    SOLAR_PRO4 = "solar-pro4",
+    SOLAR_PRO4_260806 = "solar-pro4-260806",
+    SOLAR_PRO3 = "solar-pro3",
+    SOLAR_PRO3_260323 = "solar-pro3-260323",
+    SOLAR_PRO2 = "solar-pro2",
+    SOLAR_PRO2_251215 = "solar-pro2-251215",
+    SOLAR_MINI = "solar-mini",
+    SOLAR_MINI_250422 = "solar-mini-250422",
+    SYN_PRO = "syn-pro",
+    SYN_PRO_251021 = "syn-pro-251021"
+}
 export declare enum XaiModels {
     GROK_4_6 = "grok-4.6",
     GROK_4_5 = "grok-4.5",

package/dist/constants/enums.js

@@ -23,14 +23,19 @@ export var AIProviderName;
     AIProviderName["LM_STUDIO"] = "lm-studio";
     AIProviderName["LLAMACPP"] = "llamacpp";
     // ── BEGIN GENERATED(provider-members): provider catalog (pnpm run codegen:catalog) ──
+    AIProviderName["BASETEN"] = "baseten";
     AIProviderName["CEREBRAS"] = "cerebras";
     AIProviderName["CLOUDFLARE"] = "cloudflare";
     AIProviderName["FIREWORKS"] = "fireworks";
+    AIProviderName["GMICLOUD"] = "gmicloud";
     AIProviderName["GROQ"] = "groq";
+    AIProviderName["INCEPTION_LABS"] = "inception-labs";
+    AIProviderName["IO_INTELLIGENCE"] = "io-intelligence";
     AIProviderName["MISTRAL"] = "mistral";
     AIProviderName["PERPLEXITY"] = "perplexity";
     AIProviderName["SAMBANOVA"] = "sambanova";
     AIProviderName["TOGETHER_AI"] = "together-ai";
+    AIProviderName["UPSTAGE"] = "upstage";
     AIProviderName["XAI"] = "xai";
     // ── END GENERATED(provider-members) ──
     AIProviderName["COHERE"] = "cohere";
@@ -1057,6 +1062,25 @@ export var ReplicateModels;
  */
 export const TOKEN_EXPIRY_BUFFER_MS = 5 * 60 * 1000; // 5 minutes
 // ── BEGIN GENERATED(models-enums): provider catalog (pnpm run codegen:catalog) ──
+export var BasetenModels;
+(function (BasetenModels) {
+    BasetenModels["OPENAI_GPT_OSS_120B"] = "openai/gpt-oss-120b";
+    BasetenModels["ZAI_ORG_GLM_4_7"] = "zai-org/GLM-4.7";
+    BasetenModels["MOONSHOTAI_KIMI_K2_6"] = "moonshotai/Kimi-K2.6";
+    BasetenModels["DEEPSEEK_AI_DEEPSEEK_V4_PRO"] = "deepseek-ai/DeepSeek-V4-Pro";
+    BasetenModels["NVIDIA_NVIDIA_NEMOTRON_3_ULTRA_550B_A55B"] = "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B";
+    BasetenModels["ZAI_ORG_GLM_5_2"] = "zai-org/GLM-5.2";
+    BasetenModels["MOONSHOTAI_KIMI_K2_7_CODE"] = "moonshotai/Kimi-K2.7-Code";
+    BasetenModels["DEEPSEEK_AI_DEEPSEEK_V4_FLASH_0731"] = "deepseek-ai/DeepSeek-V4-Flash-0731";
+    BasetenModels["THINKINGMACHINES_INKLING"] = "thinkingmachines/inkling";
+    BasetenModels["ZAI_ORG_GLM_5_2_FAST"] = "zai-org/GLM-5.2-Fast";
+    BasetenModels["MOONSHOTAI_KIMI_K3"] = "moonshotai/Kimi-K3";
+    BasetenModels["THINKINGMACHINES_INKLING_SMALL"] = "thinkingmachines/inkling-small";
+    BasetenModels["DEEPSEEK_AI_DEEPSEEK_V4_PRO_0813"] = "deepseek-ai/DeepSeek-V4-Pro-0813";
+    BasetenModels["ZAI_ORG_GLM_5_3_FLASH"] = "zai-org/GLM-5.3-Flash";
+    BasetenModels["ZAI_ORG_GLM_5_3"] = "zai-org/GLM-5.3";
+    BasetenModels["ZAI_ORG_GLM_5_3_FAST"] = "zai-org/GLM-5.3-Fast";
+})(BasetenModels || (BasetenModels = {}));
 export var CerebrasModels;
 (function (CerebrasModels) {
     CerebrasModels["GPT_OSS_120B"] = "gpt-oss-120b";
@@ -1089,6 +1113,10 @@ export var FireworksModels;
     FireworksModels["LLAMA_V3P2_11B_VISION_INSTRUCT"] = "accounts/fireworks/models/llama-v3p2-11b-vision-instruct";
     FireworksModels["PHI_3_VISION_128K_INSTRUCT"] = "accounts/fireworks/models/phi-3-vision-128k-instruct";
 })(FireworksModels || (FireworksModels = {}));
+export var GmicloudModels;
+(function (GmicloudModels) {
+    GmicloudModels["MINIMAXAI_MINIMAX_M3"] = "MiniMaxAI/MiniMax-M3";
+})(GmicloudModels || (GmicloudModels = {}));
 export var GroqModels;
 (function (GroqModels) {
     GroqModels["GPT_OSS_120B"] = "openai/gpt-oss-120b";
@@ -1106,6 +1134,47 @@ export var GroqModels;
     GroqModels["LLAMA_3_2_90B_VISION_PREVIEW"] = "llama-3.2-90b-vision-preview";
     GroqModels["LLAMA_3_2_11B_VISION_PREVIEW"] = "llama-3.2-11b-vision-preview";
 })(GroqModels || (GroqModels = {}));
+export var InceptionLabsModels;
+(function (InceptionLabsModels) {
+    InceptionLabsModels["MERCURY_2"] = "mercury-2";
+})(InceptionLabsModels || (InceptionLabsModels = {}));
+export var IoIntelligenceModels;
+(function (IoIntelligenceModels) {
+    IoIntelligenceModels["ZAI_ORG_GLM_5_3_FLASH"] = "zai-org/GLM-5.3-Flash";
+    IoIntelligenceModels["ZAI_ORG_GLM_5_3"] = "zai-org/GLM-5.3";
+    IoIntelligenceModels["QWEN_QWEN3_8_27B"] = "Qwen/Qwen3.8-27B";
+    IoIntelligenceModels["DEEPSEEK_AI_DEEPSEEK_V4_FLASH_0731"] = "deepseek-ai/DeepSeek-V4-Flash-0731";
+    IoIntelligenceModels["MOONSHOTAI_KIMI_K3"] = "moonshotai/Kimi-K3";
+    IoIntelligenceModels["XIAOMIMIMO_MIMO_V2_5"] = "XiaomiMiMo/MiMo-V2.5";
+    IoIntelligenceModels["ZAI_ORG_GLM_5_2"] = "zai-org/GLM-5.2";
+    IoIntelligenceModels["MOONSHOTAI_KIMI_K2_7_CODE"] = "moonshotai/Kimi-K2.7-Code";
+    IoIntelligenceModels["QWEN_QWEN3_6_35B_A3B"] = "Qwen/Qwen3.6-35B-A3B";
+    IoIntelligenceModels["QWEN_QWEN3_6_27B"] = "Qwen/Qwen3.6-27B";
+    IoIntelligenceModels["MINIMAXAI_MINIMAX_M2_7"] = "MiniMaxAI/MiniMax-M2.7";
+    IoIntelligenceModels["DEEPSEEK_AI_DEEPSEEK_V4_FLASH"] = "deepseek-ai/DeepSeek-V4-Flash";
+    IoIntelligenceModels["DEEPSEEK_AI_DEEPSEEK_V4_PRO"] = "deepseek-ai/DeepSeek-V4-Pro";
+    IoIntelligenceModels["MOONSHOTAI_KIMI_K2_6"] = "moonshotai/Kimi-K2.6";
+    IoIntelligenceModels["ZAI_ORG_GLM_5_1"] = "zai-org/GLM-5.1";
+    IoIntelligenceModels["MINIMAXAI_MINIMAX_M2_5"] = "MiniMaxAI/MiniMax-M2.5";
+    IoIntelligenceModels["MOONSHOTAI_KIMI_K2_5"] = "moonshotai/Kimi-K2.5";
+    IoIntelligenceModels["ZAI_ORG_GLM_5"] = "zai-org/GLM-5";
+    IoIntelligenceModels["DEEPSEEK_AI_DEEPSEEK_V3_2"] = "deepseek-ai/DeepSeek-V3.2";
+    IoIntelligenceModels["MOONSHOTAI_KIMI_K2_THINKING"] = "moonshotai/Kimi-K2-Thinking";
+    IoIntelligenceModels["ZAI_ORG_GLM_4_5_AIR"] = "zai-org/GLM-4.5-Air";
+    IoIntelligenceModels["GOOGLE_GEMMA_4_26B_A4B_IT"] = "google/gemma-4-26b-a4b-it";
+    IoIntelligenceModels["ZAI_ORG_GLM_4_7_FLASH"] = "zai-org/GLM-4.7-Flash";
+    IoIntelligenceModels["ZAI_ORG_GLM_4_7"] = "zai-org/GLM-4.7";
+    IoIntelligenceModels["MOONSHOTAI_KIMI_K2_INSTRUCT_0905"] = "moonshotai/Kimi-K2-Instruct-0905";
+    IoIntelligenceModels["OPENAI_GPT_OSS_120B"] = "openai/gpt-oss-120b";
+    IoIntelligenceModels["DEEPSEEK_AI_DEEPSEEK_R1_0528"] = "deepseek-ai/DeepSeek-R1-0528";
+    IoIntelligenceModels["ZAI_ORG_GLM_4_6"] = "zai-org/GLM-4.6";
+    IoIntelligenceModels["QWEN_QWEN3_NEXT_80B_A3B_INSTRUCT"] = "Qwen/Qwen3-Next-80B-A3B-Instruct";
+    IoIntelligenceModels["INTEL_QWEN3_CODER_480B_A35B_INSTRUCT_INT4_MIXED_AR"] = "Intel/Qwen3-Coder-480B-A35B-Instruct-int4-mixed-ar";
+    IoIntelligenceModels["META_LLAMA_LLAMA_4_MAVERICK_17B_128E_INSTRUCT_FP8"] = "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8";
+    IoIntelligenceModels["MISTRALAI_MISTRAL_NEMO_INSTRUCT_2407"] = "mistralai/Mistral-Nemo-Instruct-2407";
+    IoIntelligenceModels["OPENAI_GPT_OSS_20B"] = "openai/gpt-oss-20b";
+    IoIntelligenceModels["META_LLAMA_LLAMA_3_3_70B_INSTRUCT"] = "meta-llama/Llama-3.3-70B-Instruct";
+})(IoIntelligenceModels || (IoIntelligenceModels = {}));
 export var MistralModels;
 (function (MistralModels) {
     MistralModels["MISTRAL_LARGE_LATEST"] = "mistral-large-latest";
@@ -1174,6 +1243,19 @@ export var TogetherAIModels;
     TogetherAIModels["GEMMA_2_27B_IT"] = "google/gemma-2-27b-it";
     TogetherAIModels["WIZARDLM_2_8X22B"] = "microsoft/WizardLM-2-8x22B";
 })(TogetherAIModels || (TogetherAIModels = {}));
+export var UpstageModels;
+(function (UpstageModels) {
+    UpstageModels["SOLAR_PRO4"] = "solar-pro4";
+    UpstageModels["SOLAR_PRO4_260806"] = "solar-pro4-260806";
+    UpstageModels["SOLAR_PRO3"] = "solar-pro3";
+    UpstageModels["SOLAR_PRO3_260323"] = "solar-pro3-260323";
+    UpstageModels["SOLAR_PRO2"] = "solar-pro2";
+    UpstageModels["SOLAR_PRO2_251215"] = "solar-pro2-251215";
+    UpstageModels["SOLAR_MINI"] = "solar-mini";
+    UpstageModels["SOLAR_MINI_250422"] = "solar-mini-250422";
+    UpstageModels["SYN_PRO"] = "syn-pro";
+    UpstageModels["SYN_PRO_251021"] = "syn-pro-251021";
+})(UpstageModels || (UpstageModels = {}));
 export var XaiModels;
 (function (XaiModels) {
     XaiModels["GROK_4_6"] = "grok-4.6";

package/dist/core/modules/GenerationHandler.d.ts

@@ -39,9 +39,6 @@ export declare function resolveTurnBudget(options: TextGenerationOptions, turnSt
     wrapupLeadMs: number;
     turnDeadline: number | undefined;
 };
-/**
- * GenerationHandler class - Handles text generation operations for AI providers
- */
 export declare class GenerationHandler {
     private readonly providerName;
     private readonly modelName;
@@ -127,6 +124,21 @@ export declare class GenerationHandler {
     /**
      * Format the enhanced result
      */
+    /**
+     * One re-ask with `toolChoice: "none"` when a tool loop ends on a
+     * `tool-calls` finish that carries neither a tool call nor any text.
+     *
+     * io.net's Llama endpoint does exactly this on the step after a tool result
+     * when the caller asked for JSON: the model's JSON-shaped answer trips the
+     * vendor's tool-call parser, which drops it and reports
+     * `finish_reason: tool_calls` with `content: null` and no `tool_calls`.
+     * The AI-SDK loop has nothing to execute and stops, so the caller gets an
+     * empty turn although the tool ran. Replaying that request with
+     * `tool_choice: "none"` (or no tool list) returns the answer — verified on
+     * the wire, 4/4 — so the recovery is one bounded extra step that keeps the
+     * executed tool steps and usage in the returned result.
+     */
+    private recoverEmptyToolCallsFinish;
     formatEnhancedResult(generateResult: Awaited<ReturnType<typeof generateText>>, tools: Record<string, Tool>, toolsUsed: string[], toolExecutions: ToolExecutionRecord[], options: TextGenerationOptions): EnhancedGenerateResult;
     /**
      * Analyze AI response structure and log detailed debugging information

package/dist/core/modules/GenerationHandler.js

@@ -157,6 +157,53 @@ function buildWrapupStepResult(prepared, stepMessages) {
 /**
  * GenerationHandler class - Handles text generation operations for AI providers
  */
+/**
+ * Append a schema-derived JSON instruction to the hoisted system prompt for
+ * the structured-output fallback retry. Phrased for tool loops too: the model
+ * may still call tools first; only the final answer must be the object.
+ */
+function appendJsonInstruction(system, schema) {
+    const jsonSchema = JSON.stringify(convertZodToJsonSchema(schema));
+    const instruction = "When you give your final answer, respond with only a single JSON object that conforms to the following JSON Schema. " +
+        "No prose before or after it, and no markdown code fence. " +
+        `JSON Schema: ${jsonSchema}`;
+    const existing = system ?? [];
+    const last = existing[existing.length - 1];
+    // Merge into the caller's own system message when there is one, the way
+    // messageBuilder folds STRUCTURED_OUTPUT_INSTRUCTIONS into a single system
+    // turn — several self-hosted OpenAI-compatible stacks only honour one
+    // leading system message.
+    if (last !== undefined && typeof last.content === "string") {
+        return [
+            ...existing.slice(0, -1),
+            { ...last, content: `${last.content}\n\n${instruction}` },
+        ];
+    }
+    return [...existing, { role: "system", content: instruction }];
+}
+/** Sum two AI-SDK usage records leaf by leaf (the v6 shape nests token
+ *  details); undefined counts are treated as absent, not zero. */
+function addUsage(a, b) {
+    const sum = (x, y) => {
+        if (typeof x === "number" || typeof y === "number") {
+            return (typeof x === "number" ? x : 0) + (typeof y === "number" ? y : 0);
+        }
+        if (x && y && typeof x === "object" && typeof y === "object") {
+            const left = x;
+            const right = y;
+            const out = {};
+            for (const key of new Set([
+                ...Object.keys(left),
+                ...Object.keys(right),
+            ])) {
+                out[key] = sum(left[key], right[key]);
+            }
+            return out;
+        }
+        return x ?? y;
+    };
+    return sum(a, b);
+}
 export class GenerationHandler {
     providerName;
     modelName;
@@ -192,7 +239,7 @@ export class GenerationHandler {
      * @private
      */
     async callGenerateText(model, messages, tools, options, callConfig) {
-        const { shouldUseTools, includeStructuredOutput, turnStartMs } = callConfig;
+        const { shouldUseTools, includeStructuredOutput, turnStartMs, promptJsonInstruction, } = callConfig;
         // Check if this is a Google provider (for provider-specific options)
         const isGoogleProvider = this.providerName === "google-ai" || this.providerName === "vertex";
         // Check if this is an Anthropic provider (includes Vertex+Claude)
@@ -259,7 +306,17 @@ export class GenerationHandler {
         // Hoist system-role messages into generateText's top-level `system` option
         // rather than passing them inside `messages` (deprecated by the AI SDK,
         // rejected in v7). See extractSystemMessages for the rationale. (#1024)
-        const { system, messages: nonSystemMessages } = extractSystemMessages(messages);
+        const { system: hoistedSystem, messages: nonSystemMessages } = extractSystemMessages(messages);
+        // Structured-output fallback: spell the schema out in the prompt. The
+        // native `response_format` attempt already failed to yield an object, and
+        // a vendor that ignores `response_format` outright (GMI Cloud's MiniMax
+        // endpoint answers a strict json_schema request in markdown) would answer
+        // a silent retry the same way. With the schema in the system prompt the
+        // same text-coercion path in formatEnhancedResult recovers a schema-valid
+        // object — verified live on that endpoint, 3/3.
+        const system = promptJsonInstruction && options.schema
+            ? appendJsonInstruction(hoistedSystem, options.schema)
+            : hoistedSystem;
         // Per-step context budget guard: the tool loop appends assistant turns and
         // tool results on every step — growth the pre-call budget check never
         // sees. Estimate each step's projected request and deterministically
@@ -509,12 +566,19 @@ export class GenerationHandler {
                     });
                 }
                 // Set token usage and completion attributes on span
-                this.setUsageSpanAttributes(span, result);
-                if (result.finishReason) {
-                    span.setAttribute("gen_ai.response.finish_reason", result.finishReason);
+                // Span attributes come from the RECOVERED result: a successful
+                // toolChoice:"none" re-ask changes the finish reason and adds usage.
+                const recovered = await this.recoverEmptyToolCallsFinish(model, messages, tools, options, {
+                    shouldUseTools,
+                    includeStructuredOutput: true,
+                    turnStartMs: genStartTime,
+                }, { result, span });
+                this.setUsageSpanAttributes(span, recovered);
+                if (recovered.finishReason) {
+                    span.setAttribute("gen_ai.response.finish_reason", recovered.finishReason);
                 }
                 span.setStatus({ code: SpanStatusCode.OK });
-                return result;
+                return recovered;
             }
             catch (error) {
                 // Fall back to text-mode (no experimental_output) when structured
@@ -567,11 +631,13 @@ export class GenerationHandler {
                         // includeStructuredOutput intentionally omitted
                         includeStructuredOutput: false,
                         turnStartMs: genStartTime,
+                        promptJsonInstruction: true,
                     }), span, "generateText(fallback)");
                     // NLK-GAP-007: Record recovery event after successful fallback
                     span.addEvent("retry.recovered", {
                         "retry.attempts": 2,
                         "retry.strategy": "structured_output_disabled",
+                        "retry.prompt_json_instruction": true,
                     });
                     span.setAttribute("retry.count", 1);
                     logger.info("[GenerationHandler] generateText returned (fallback)", {
@@ -582,12 +648,20 @@ export class GenerationHandler {
                         toolCallsTotal: result.toolCalls?.length || 0,
                         responseChars: result.text?.length || 0,
                     });
-                    this.setUsageSpanAttributes(span, result);
-                    if (result.finishReason) {
-                        span.setAttribute("gen_ai.response.finish_reason", result.finishReason);
+                    // Span attributes come from the RECOVERED result: a successful
+                    // toolChoice:"none" re-ask changes the finish reason and adds usage.
+                    const recovered = await this.recoverEmptyToolCallsFinish(model, messages, tools, options, {
+                        shouldUseTools,
+                        includeStructuredOutput: false,
+                        turnStartMs: genStartTime,
+                        promptJsonInstruction: true,
+                    }, { result, span });
+                    this.setUsageSpanAttributes(span, recovered);
+                    if (recovered.finishReason) {
+                        span.setAttribute("gen_ai.response.finish_reason", recovered.finishReason);
                     }
                     span.setStatus({ code: SpanStatusCode.OK });
-                    return result;
+                    return recovered;
                 }
                 // Retry once without `temperature` when the model deprecated it. The
                 // newest Anthropic models (e.g. claude-opus-4-8 with tools + advanced
@@ -619,12 +693,19 @@ export class GenerationHandler {
                         "retry.strategy": "temperature_omitted",
                     });
                     span.setAttribute("retry.count", 1);
-                    this.setUsageSpanAttributes(span, result);
-                    if (result.finishReason) {
-                        span.setAttribute("gen_ai.response.finish_reason", result.finishReason);
+                    // Span attributes come from the RECOVERED result: a successful
+                    // toolChoice:"none" re-ask changes the finish reason and adds usage.
+                    const recovered = await this.recoverEmptyToolCallsFinish(model, messages, tools, { ...options, temperature: undefined }, {
+                        shouldUseTools,
+                        includeStructuredOutput: true,
+                        turnStartMs: genStartTime,
+                    }, { result, span });
+                    this.setUsageSpanAttributes(span, recovered);
+                    if (recovered.finishReason) {
+                        span.setAttribute("gen_ai.response.finish_reason", recovered.finishReason);
                     }
                     span.setStatus({ code: SpanStatusCode.OK });
-                    return result;
+                    return recovered;
                 }
                 span.setStatus({
                     code: SpanStatusCode.ERROR,
@@ -774,6 +855,83 @@ export class GenerationHandler {
     /**
      * Format the enhanced result
      */
+    /**
+     * One re-ask with `toolChoice: "none"` when a tool loop ends on a
+     * `tool-calls` finish that carries neither a tool call nor any text.
+     *
+     * io.net's Llama endpoint does exactly this on the step after a tool result
+     * when the caller asked for JSON: the model's JSON-shaped answer trips the
+     * vendor's tool-call parser, which drops it and reports
+     * `finish_reason: tool_calls` with `content: null` and no `tool_calls`.
+     * The AI-SDK loop has nothing to execute and stops, so the caller gets an
+     * empty turn although the tool ran. Replaying that request with
+     * `tool_choice: "none"` (or no tool list) returns the answer — verified on
+     * the wire, 4/4 — so the recovery is one bounded extra step that keeps the
+     * executed tool steps and usage in the returned result.
+     */
+    async recoverEmptyToolCallsFinish(model, messages, tools, options, callConfig, attempt) {
+        const { result, span } = attempt;
+        // The re-ask is one more loop step, so it must fit the caller's step
+        // budget: when the loop already spent every step, the honest answer is
+        // the step-cap stop the caller configured, not an extra request.
+        const maxSteps = options.maxSteps ?? DEFAULT_MAX_STEPS;
+        const emptyToolCallsFinish = callConfig.shouldUseTools &&
+            !callConfig.isToolReask &&
+            Object.keys(tools).length > 0 &&
+            result.finishReason === "tool-calls" &&
+            result.toolCalls.length === 0 &&
+            (result.text ?? "").trim().length === 0 &&
+            result.steps.length < maxSteps;
+        if (!emptyToolCallsFinish) {
+            return result;
+        }
+        logger.warn("[GenerationHandler] tool loop ended on a tool-calls finish with no tool call and no text — re-asking once with toolChoice: none", {
+            provider: this.providerName,
+            model: this.modelName,
+            stepsSoFar: result.steps.length,
+        });
+        span.setAttribute("neurolink.has_fallback", true);
+        span.addEvent("retry.initial_failure", {
+            "retry.attempt": 1,
+            "retry.reason": "empty_tool_calls_finish",
+        });
+        let reask;
+        try {
+            reask = await withProviderRetry(() => this.callGenerateText(model, [...messages, ...result.response.messages], tools, { ...options, toolChoice: "none", maxSteps: 1 }, { ...callConfig, isToolReask: true }), span, "generateText(tool-choice-none)");
+        }
+        catch (reaskError) {
+            // The re-ask is a bonus request on top of a call that already
+            // succeeded. If it fails for any reason, hand back the original result
+            // — the executed tool steps are still in it — rather than turning a
+            // degraded turn into a thrown one.
+            logger.warn("[GenerationHandler] toolChoice: none re-ask failed; returning the original result", {
+                provider: this.providerName,
+                model: this.modelName,
+                error: reaskError instanceof Error
+                    ? reaskError.message
+                    : String(reaskError),
+            });
+            span.addEvent("retry.failed", {
+                "retry.attempts": 2,
+                "retry.strategy": "tool_choice_none_reask",
+            });
+            return result;
+        }
+        span.addEvent("retry.recovered", {
+            "retry.attempts": 2,
+            "retry.strategy": "tool_choice_none_reask",
+        });
+        span.setAttribute("retry.count", 1);
+        // Keep the executed tool steps in front of the re-ask so
+        // extractToolInformation still reports them and usage stays a true
+        // cross-step total. `steps` and `totalUsage` are constructor-assigned
+        // fields on the AI-SDK result; every other accessor derives from the
+        // final step, which is now the re-ask's answer.
+        const merged = reask;
+        merged.steps = [...result.steps, ...reask.steps];
+        merged.totalUsage = addUsage(result.totalUsage, reask.totalUsage);
+        return reask;
+    }
     formatEnhancedResult(generateResult, tools, toolsUsed, toolExecutions, options) {
         // Structured output check — schema alone is sufficient to activate
         const useStructuredOutput = !!options.schema ||

package/dist/index.d.ts

@@ -38,6 +38,9 @@ export { PROVIDER_DESCRIPTORS, PROVIDER_DESCRIPTORS_BY_NAME, PROVIDER_ALIAS_INDE
 export { NeuroLinkConfigManager as ConfigManager } from "./config/configManager.js";
 export { BaseFactory, BaseRegistry, NeuroLinkFeatureError, createErrorFactory, withRetry, TypedEventEmitter, } from "./core/infrastructure/index.js";
 export { NeuroLinkClient, createClient, NeuroLinkApiError, } from "./client/httpClient.js";
+export { LocalTempArtifactStore } from "./artifacts/artifactStore.js";
+export { RedisArtifactStore } from "./artifacts/redisArtifactStore.js";
+export { createArtifactStore, resolveArtifactStorageType, } from "./artifacts/artifactStoreFactory.js";
 export { NeuroLinkLanguageModel, NeuroLinkProvider as NeuroLinkAIProvider, createNeuroLinkProvider, createNeuroLinkModel, createStreamingResponse, neurolink as neuroLinkAIInstance, } from "./client/aiSdkAdapter.js";
 export { createApiKeyAuthInterceptor, createBearerAuthInterceptor, createDynamicAuthInterceptor, createLoggingInterceptor, createRetryInterceptor, createRateLimitInterceptor, createRequestTransformInterceptor, createResponseTransformInterceptor, createCacheInterceptor, createTimeoutInterceptor, createErrorHandlerInterceptor, composeMiddleware, conditionalMiddleware, } from "./client/interceptors.js";
 export { SSEClient, WebSocketStreamingClient, createStreamingClient, createAsyncStream, collectStream, } from "./client/streamingClient.js";

package/dist/index.js

@@ -51,6 +51,13 @@ export { BaseFactory, BaseRegistry, NeuroLinkFeatureError, createErrorFactory, w
 export { 
 // HTTP Client
 NeuroLinkClient, createClient, NeuroLinkApiError, } from "./client/httpClient.js";
+// ============================================================================
+// ARTIFACT STORAGE — backends for externalized tool outputs and banked payloads.
+// Pick one with `artifacts.storage` / STORAGE_TYPE, or implement ArtifactStore.
+// ============================================================================
+export { LocalTempArtifactStore } from "./artifacts/artifactStore.js";
+export { RedisArtifactStore } from "./artifacts/redisArtifactStore.js";
+export { createArtifactStore, resolveArtifactStorageType, } from "./artifacts/artifactStoreFactory.js";
 export { 
 // AI SDK Adapter
 NeuroLinkLanguageModel, NeuroLinkProvider as NeuroLinkAIProvider, createNeuroLinkProvider, createNeuroLinkModel, createStreamingResponse, neurolink as neuroLinkAIInstance, } from "./client/aiSdkAdapter.js";