@juspay/neurolink 12.15.1 → 12.17.0

package/dist/core/nativeGenerateLoop.js

@@ -118,6 +118,20 @@ export async function runNativeGenerateLoop(args, toolExecutionSummaries) {
     const hasTools = Boolean(args.tools && args.tools.length > 0);
     for (let step = 0; step < args.maxSteps; step++) {
         steps = step + 1;
+        // Per-step context reclaim. The pre-dispatch budget check runs ONCE and
+        // never sees the assistant turns and tool results this loop appends, which
+        // is how a long agentic run overflows the model window mid-loop and loses
+        // every completed step to a provider 400. The streaming paths have guarded
+        // this for a while; generate() did not, so the identical turn reclaimed on
+        // stream() and overflowed on generate().
+        //
+        // The provider supplies the guard because only it knows its wire shape.
+        // Returning undefined leaves the conversation byte-identical, so a turn
+        // that fits never pays a prompt-cache invalidation.
+        const guarded = args.guardConversation?.(args.conversation);
+        if (guarded) {
+            args.conversation.splice(0, args.conversation.length, ...guarded);
+        }
         const runThisStep = () => args.runStep(() => args.doGenerate({
             prompt: args.conversation,
             ...(args.tools && args.tools.length > 0 ? { tools: args.tools } : {}),
@@ -169,6 +183,7 @@ export async function runNativeGenerateLoop(args, toolExecutionSummaries) {
             throw stepError;
         }
         reaskPending = false;
+        args.observeUsage?.(res.usage);
         const parts = asParts(res.content);
         // Each step REPLACES the text rather than appending: the final step's
         // answer is the turn's answer, matching what generateText reported.
@@ -177,11 +192,16 @@ export async function runNativeGenerateLoop(args, toolExecutionSummaries) {
             .map((p) => p.text)
             .join("");
         // Reasoner models (DeepSeek `reasoning_content`, gateway `reasoning`,
-        // OpenAI o-series) emit reasoning as its own V3 content part. Join it on
-        // the same replace-per-step rule as the text so the caller can surface
-        // `result.reasoning`; without this the provider builds the part and the
-        // loop drops it.
-        reasoning = parts
+        // OpenAI o-series) emit reasoning as its own V3 content part.
+        //
+        // Unlike `text`, this ACCUMULATES across steps. Each step's answer
+        // supersedes the last, but each step's reasoning is its own thought:
+        // replacing meant a tool loop reported only whatever the final step
+        // happened to think, which is usually the shortest part and frequently
+        // empty once the model just answers from the tool result. Concatenating
+        // with nothing between them is what a stream() consumer accumulates from
+        // the per-step reasoning chunks, so both surfaces agree on the turn.
+        reasoning += parts
             .filter((p) => p.type === "reasoning" && typeof p.text === "string")
             .map((p) => p.text)
             .join("");

package/dist/providers/catalog/api-route.json

@@ -0,0 +1,130 @@
+{
+    "$schema": "./provider-catalog.schema.json",
+    "id": "api-route",
+    "displayName": "API Route",
+    "aliases": ["apiroute"],
+    "tier": 2,
+    "wire": {
+        "baseURL": "https://global.api-route.com/v1"
+    },
+    "models": {
+        "default": "claude-sonnet-4-6",
+        "fallbacks": ["claude-haiku-4-5", "deepseek-v4-flash", "gemini-3.8-flash"],
+        "defaultContextWindow": 1000000,
+        "defaultMaxOutputTokens": 64000,
+        "catalog": {
+            "claude-sonnet-4-6": {
+                "contextWindow": 1000000,
+                "maxOutputTokens": 64000,
+                "vision": true,
+                "status": "production",
+                "description": "Claude Sonnet 4.6 — flagship model with 1M context, tool calling, and multimodal vision"
+            },
+            "claude-haiku-4-5": {
+                "contextWindow": 200000,
+                "maxOutputTokens": 32768,
+                "vision": true,
+                "status": "production",
+                "description": "Claude Haiku 4.5 — fast, high-efficiency model with 200K context, tool calling, and vision"
+            },
+            "deepseek-v4-flash": {
+                "contextWindow": 1000000,
+                "maxOutputTokens": 64000,
+                "vision": false,
+                "status": "production",
+                "description": "DeepSeek V4 Flash — fast, cost-effective reasoning model with 1M context"
+            },
+            "deepseek-v4-pro": {
+                "contextWindow": 1000000,
+                "maxOutputTokens": 64000,
+                "vision": false,
+                "status": "production",
+                "description": "DeepSeek V4 Pro — powerful reasoning model with 1M context"
+            },
+            "gemini-3.8-flash": {
+                "contextWindow": 1000000,
+                "maxOutputTokens": 64000,
+                "vision": true,
+                "status": "production",
+                "description": "Gemini 3.8 Flash — high-speed multimodal model supporting image input and streaming"
+            },
+            "qwen3.8-flash": {
+                "contextWindow": 128000,
+                "maxOutputTokens": 16384,
+                "vision": false,
+                "status": "production",
+                "description": "Qwen 3.8 Flash — versatile, fast model for general instruction and coding tasks"
+            },
+            "kimi-k2.7-code": {
+                "contextWindow": 256000,
+                "maxOutputTokens": 32768,
+                "vision": false,
+                "status": "production",
+                "description": "Kimi K2.7 Code — specialized for coding and multi-step reasoning"
+            },
+            "glm-5.3-flash": {
+                "contextWindow": 128000,
+                "maxOutputTokens": 16384,
+                "vision": false,
+                "status": "production",
+                "description": "GLM 5.3 Flash — lightweight, fast conversational and instruction-following model"
+            }
+        },
+        "visionModel": "claude-sonnet-4-6",
+        "topModels": [
+            "claude-sonnet-4-6",
+            "claude-haiku-4-5",
+            "deepseek-v4-flash",
+            "gemini-3.8-flash",
+            "qwen3.8-flash"
+        ]
+    },
+    "capabilities": {
+        "text": true,
+        "streaming": true,
+        "tools": true,
+        "toolsWithStreaming": true,
+        "structuredOutput": true,
+        "structuredOutputWithTools": true,
+        "embeddings": false,
+        "thinking": false
+    },
+    "errorRules": [
+        {
+            "pattern": "该分组下模型无可用渠道|model_not_found",
+            "class": "invalid-model",
+            "message": "Model '{model}' not found or has no available channels on API Route."
+        },
+        {
+            "pattern": "无效的令牌",
+            "class": "authentication",
+            "message": "Invalid API Route API key. Check {apiKeyEnvVar}."
+        }
+    ],
+    "setup": {
+        "url": "https://api-route.com",
+        "apiKeyFormat": null,
+        "billingPolicy": "free-tier",
+        "instructions": [
+            "1. Visit: https://api-route.com and create an account",
+            "2. Generate an API key in your console / dashboard",
+            "3. Set {apiKeyEnvVar} in your .env file"
+        ]
+    },
+    "evidence": {
+        "rosterVerified": {
+            "date": "2026-09-17",
+            "method": "authenticated GET /v1/models",
+            "status": 200
+        },
+        "authProbe": {
+            "date": "2026-09-17",
+            "status": 401
+        },
+        "liveMatrix": {
+            "date": "2026-09-17",
+            "result": "Verified live against https://global.api-route.com/v1: GET /v1/models returned 70 models. Chat completions, streaming, function calling, streamed function calling, and structured JSON output (json_object and json_schema) verified. Multimodal vision verified on claude-sonnet-4-6, claude-haiku-4-5, and gemini-3.8-flash. Invalid API key returns 401 (无效的令牌) and invalid model returns 503 (model_not_found)."
+        },
+        "addedInPR": "pending"
+    }
+}

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

@@ -1,3 +1,3 @@
 import type { ProviderCatalogJson } from "../../types/index.js";
 export declare const CATALOG_JSON_ENTRIES: ProviderCatalogJson[];
-export declare const CATALOG_PROVIDER_IDS: readonly ["baseten", "cerebras", "cloudflare", "fireworks", "gmicloud", "groq", "inception-labs", "io-intelligence", "mancer", "mistral", "perplexity", "sambanova", "together-ai", "upstage", "xai"];
+export declare const CATALOG_PROVIDER_IDS: readonly ["api-route", "baseten", "cerebras", "cloudflare", "fireworks", "gmicloud", "groq", "inception-labs", "io-intelligence", "mancer", "mistral", "perplexity", "sambanova", "together-ai", "upstage", "xai"];

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

@@ -1,5 +1,6 @@
 // GENERATED FILE — do not edit. Regenerate with `pnpm run codegen:catalog`.
 // Source of truth: the per-provider JSON files in this directory.
+import apiRouteJson from "./api-route.json" with { type: "json" };
 import basetenJson from "./baseten.json" with { type: "json" };
 import cerebrasJson from "./cerebras.json" with { type: "json" };
 import cloudflareJson from "./cloudflare.json" with { type: "json" };
@@ -16,6 +17,7 @@ import togetherAiJson from "./together-ai.json" with { type: "json" };
 import upstageJson from "./upstage.json" with { type: "json" };
 import xaiJson from "./xai.json" with { type: "json" };
 export const CATALOG_JSON_ENTRIES = [
+    apiRouteJson,
     basetenJson,
     cerebrasJson,
     cloudflareJson,
@@ -33,6 +35,7 @@ export const CATALOG_JSON_ENTRIES = [
     xaiJson,
 ];
 export const CATALOG_PROVIDER_IDS = [
+    "api-route",
     "baseten",
     "cerebras",
     "cloudflare",

package/dist/providers/catalog/schema.d.ts

@@ -73,8 +73,8 @@ export declare const providerCatalogJsonSchema: z.ZodObject<{
         class: z.ZodEnum<{
             network: "network";
             provider: "provider";
-            authentication: "authentication";
             "invalid-model": "invalid-model";
+            authentication: "authentication";
             "rate-limit": "rate-limit";
         }>;
         message: z.ZodString;

package/dist/providers/openaiChatCompletionsBase.js

@@ -20,6 +20,8 @@
 import { trace } from "@opentelemetry/api";
 import { getAvailableInputTokens, getRuntimeContextWindow, getRuntimeOutputCeiling, registerRuntimeContextWindow, } from "../constants/contextWindows.js";
 import { guardOpenAICompatConversation } from "../context/openaiCompatLoopGuard.js";
+import { estimateTokens, serializeForEstimate, } from "../utils/tokenEstimation.js";
+import { createNativeGenerateGuard } from "../context/nativeGenerateGuard.js";
 import { isContextOverflowError, parseProviderOverflowDetails, } from "../context/errorDetection.js";
 import { ContextBudgetExceededError } from "../context/errors.js";
 import { BaseProvider } from "../core/baseProvider.js";
@@ -899,6 +901,19 @@ export class OpenAIChatCompletionsProvider extends BaseProvider {
         // legal again on a vendor that refuses the two together.
         withTools = true) => runNativeGenerateLoop({
             doGenerate,
+            ...createNativeGenerateGuard({
+                provider: this.providerName,
+                availableInputTokens: getAvailableInputTokens(this.providerName, modelId, options.maxTokens ?? undefined),
+                // Runs on every step of the generate loop, so a cyclic or
+                // oversized tool schema must not throw the turn away. Shares the
+                // estimator's own serializer: its fallback over-reports, which
+                // makes the guard fire rather than silently size the overhead at
+                // nothing.
+                getFixedOverheadTokens: () => estimateTokens(serializeForEstimate({
+                    tools: v3Tools,
+                    responseFormat: format,
+                }), this.providerName),
+            }),
             conversation: conv,
             ...(withTools && v3Tools ? { tools: v3Tools } : {}),
             toolsRecord,

package/dist/types/config.d.ts

@@ -415,6 +415,16 @@ export type ToolConfig = {
     enableMCPTools?: boolean;
     /** Whether the bash command execution tool should be enabled (opt-in, defaults to false) */
     enableBashTool?: boolean;
+    /**
+     * Byte ceiling for the safety-net truncation `ToolsManager` applies to
+     * every direct/custom/external-MCP tool result before it reaches the AI
+     * SDK accumulator (BZ-666). Independent of `mcp.outputLimits`, which only
+     * governs external MCP results earlier in the pipeline via its own
+     * externalize-to-artifact-store strategy. Default: 51200 (50 KB) —
+     * unchanged from the previous hard-coded ceiling, so existing consumers
+     * see no behavior change unless they set this explicitly.
+     */
+    outputTruncationMaxBytes?: number;
 };
 /**
  * Backup metadata information

package/dist/types/context.d.ts

@@ -373,6 +373,15 @@ export type ToolOutputPreviewOptions = {
     headRatio?: number;
     /** Fraction of preview budget allocated to the tail (default: 0.75) */
     tailRatio?: number;
+    /**
+     * Override the omission notice spliced between head and tail. A string is
+     * used verbatim; a function receives the omitted byte count and returns
+     * the notice text. When omitted, the built-in default is used, which
+     * names the `retrieve_context` tool — pass this when that tool is not
+     * registered on the calling instance so the model isn't pointed at a tool
+     * that doesn't exist.
+     */
+    notice?: string | ((omittedBytes: number) => string);
 };
 /** Result of tool output preview generation. */
 export type ToolOutputPreviewResult = {
@@ -562,4 +571,10 @@ export type FileSummarizationServiceOptions = {
     provider?: string;
     model?: string;
 };
+/** Per-turn configuration for reclaiming a native V3 conversation. */
+export type NativeGenerateGuardConfig = {
+    provider: string;
+    availableInputTokens: number;
+    getFixedOverheadTokens: () => number;
+};
 export {};

package/dist/types/generate.d.ts

@@ -1604,6 +1604,15 @@ export type GenerateOptionsNormalized = GenerateOptions & {
  * `doGenerate`; the provider supplies the wire details.
  */
 export type NativeGenerateLoopArgs = {
+    /** Observed usage for calibrating the next step against the last request. */
+    observeUsage?: (usage: unknown) => void;
+    /**
+     * Per-step context reclaim, called before every model call with the
+     * conversation as it now stands. Return a replacement to have the loop adopt
+     * it, or undefined to leave it untouched. The provider owns this because the
+     * reclaim has to understand its wire shape.
+     */
+    guardConversation?: (conversation: Array<Record<string, unknown>>) => Array<Record<string, unknown>> | undefined;
     doGenerate: (options: Record<string, unknown>) => Promise<Record<string, unknown>>;
     /** Conversation in the message-builder shape each doGenerate converts itself. */
     conversation: Array<Record<string, unknown>>;

package/dist/types/providerCatalog.generated.d.ts

@@ -1,2 +1,2 @@
-export type CatalogProviderName = "baseten" | "cerebras" | "cloudflare" | "fireworks" | "gmicloud" | "groq" | "inception-labs" | "io-intelligence" | "mancer" | "mistral" | "perplexity" | "sambanova" | "together-ai" | "upstage" | "xai";
-export type CatalogCredentialKey = "baseten" | "cerebras" | "cloudflare" | "fireworks" | "gmicloud" | "groq" | "inceptionLabs" | "ioIntelligence" | "mancer" | "mistral" | "perplexity" | "sambanova" | "together" | "upstage" | "xai";
+export type CatalogProviderName = "api-route" | "baseten" | "cerebras" | "cloudflare" | "fireworks" | "gmicloud" | "groq" | "inception-labs" | "io-intelligence" | "mancer" | "mistral" | "perplexity" | "sambanova" | "together-ai" | "upstage" | "xai";
+export type CatalogCredentialKey = "apiRoute" | "baseten" | "cerebras" | "cloudflare" | "fireworks" | "gmicloud" | "groq" | "inceptionLabs" | "ioIntelligence" | "mancer" | "mistral" | "perplexity" | "sambanova" | "together" | "upstage" | "xai";

package/dist/types/providers.d.ts

@@ -178,6 +178,10 @@ export type NeurolinkCredentials = {
         apiKey?: string;
         baseURL?: string;
     };
+    apiRoute?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
     baseten?: {
         apiKey?: string;
         baseURL?: string;

package/dist/utils/tokenEstimation.d.ts

@@ -47,6 +47,19 @@ export declare function getProviderMultiplier(provider?: string): number;
  * @returns Estimated token count
  */
 export declare function estimateTokens(text: string, provider?: string, isCode?: boolean): number;
+/**
+ * Serialize an arbitrary value for estimation. Never throws: `JSON.stringify`
+ * raises RangeError once a value exceeds V8's max string length, and a tool
+ * argument blob is exactly the shape that gets there. A payload that large is
+ * charged at the fallback size rather than aborting the estimate (and with it
+ * the whole turn).
+ */
+/**
+ * Serialize a value for token estimation. The fallback is deliberately huge:
+ * an estimator that cannot measure a value must over-report, so a budget guard
+ * fires rather than waving through content it failed to size.
+ */
+export declare function serializeForEstimate(value: unknown): string;
 /**
  * Estimate token count for a single ChatMessage.
  * Includes message framing overhead.

package/dist/utils/tokenEstimation.js

@@ -93,7 +93,12 @@ export function estimateTokens(text, provider, isCode) {
  * charged at the fallback size rather than aborting the estimate (and with it
  * the whole turn).
  */
-function serializeForEstimate(value) {
+/**
+ * Serialize a value for token estimation. The fallback is deliberately huge:
+ * an estimator that cannot measure a value must over-report, so a budget guard
+ * fires rather than waving through content it failed to size.
+ */
+export function serializeForEstimate(value) {
     if (typeof value === "string") {
         return value;
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@juspay/neurolink",
-  "version": "12.15.1",
+  "version": "12.17.0",
   "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": {
@@ -166,7 +166,9 @@
     "test:agents-live": "tsx test/continuous-test-agents-live.ts",
     "test:tool-routing-cli": "pnpm exec tsx test/continuous-test-suite-tool-routing-cli.ts",
     "test:tool-dedup": "pnpm exec tsx test/continuous-test-suite-tool-dedup.ts",
+    "test:tools-manager-truncation": "pnpm exec tsx test/continuous-test-suite-tools-manager-truncation.ts",
     "test:model-not-found-retryable": "pnpm exec tsx test/continuous-test-suite-model-not-found-retryable.ts",
+    "test:reasoning-parity-live": "pnpm exec tsx test/continuous-test-suite-reasoning-parity-live.ts",
     "test:ci": "pnpm run test && pnpm run test:client && pnpm run test:hitl",
     "// CI tier — fast, no live AI calls, safe for every commit (test:unit; also see the separate provider-safety-net CI job, which runs build + test:providers-mocked + test:provider-structure + test:error-classifier-contract on every PR)": "",
     "test:tool-routing": "pnpm exec tsx test/continuous-test-suite-tool-routing.ts",