@juspay/neurolink 12.23.2 → 12.24.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@juspay/neurolink",
-  "version": "12.23.2",
+  "version": "12.24.0",
   "packageManager": "pnpm@10.15.1",
   "description": "The pipe layer of an AI nervous system: one interface connecting provider neurons to your application, across three inference types — generate, stream and decide. `decide` returns typed, calibrated judgements from a non-generative model (~400ms, ~$0.00002/call) for routing, tool selection and context budgeting. MCP-native (4 transports), voice TTS/STT/realtime, RAG, agents, memory, compaction, 9 observability exporters. OpenAI · Anthropic · Gemini · Bedrock · Azure · Ollama · TypeSafe Jev and more.",
   "author": {
@@ -99,6 +99,7 @@
     "test:openai-compat-streaming-retry": "pnpm exec tsx test/continuous-test-suite-openai-compat-streaming-retry.ts",
     "test:anthropic-streaming-retry": "pnpm exec tsx test/continuous-test-suite-anthropic-streaming-retry.ts",
     "test:anthropic-silent-truncation": "pnpm exec tsx test/continuous-test-suite-anthropic-silent-truncation.ts",
+    "test:cache-cost-accounting": "pnpm exec tsx test/continuous-test-suite-cache-cost-accounting.ts",
     "test:adjust-body-after-400": "pnpm exec tsx test/continuous-test-suite-adjust-body-after-400.ts",
     "test:zod3-schema-native-path": "pnpm exec tsx test/continuous-test-suite-zod3-schema-native-path.ts",
     "test:error-classification-e2e": "pnpm exec tsx test/continuous-test-suite-error-classification-e2e.ts",

package/dist/providers/deepseek.d.ts

@@ -1,47 +0,0 @@
-import type { AIProviderName } from "../constants/enums.js";
-import type { NeurolinkCredentials, OpenAICompatResponseFormat } from "../types/index.js";
-import { OpenAIChatCompletionsProvider } from "./openaiChatCompletionsBase.js";
-/**
- * DeepSeek Provider — direct HTTP, no AI SDK.
- *
- * OpenAI-compatible chat completions at api.deepseek.com (deepseek-chat /
- * deepseek-reasoner). All request/stream/tool-loop orchestration lives in
- * `OpenAIChatCompletionsProvider`; this class declares configuration and
- * provider-specific quirks:
- *
- *   1. Structured-output downgrade — DeepSeek rejects `response_format:
- *      { type: "json_schema" }` ("This response_format type is unavailable
- *      now"), so `adjustResponseFormat` downgrades it to `json_object` —
- *      matching the `supportsStructuredOutputs: false` behaviour of the
- *      `@ai-sdk/openai-compatible` path this migration replaced. The base
- *      client injects the literal "json" word the API requires for that mode.
- *
- *   2. Reasoning support — `reasoning_content` (deepseek-reasoner / R1) is
- *      surfaced automatically by the native base client: streamed deltas
- *      arrive as `{ content: "", reasoning }` chunks and the non-streaming
- *      result carries `result.reasoning`. The opt-in `thinking` request
- *      param (non-reasoner chat models) still needs thinking-signal plumbing
- *      and is tracked as a follow-up. All other behavior is preserved.
- *
- * @see https://api-docs.deepseek.com
- */
-export declare class DeepSeekProvider extends OpenAIChatCompletionsProvider {
-    constructor(modelName?: string, sdk?: unknown, _region?: string, credentials?: NeurolinkCredentials["deepseek"]);
-    protected getProviderName(): AIProviderName;
-    protected getDefaultModel(): string;
-    protected formatProviderError(error: unknown): Error;
-    protected getFallbackModelName(): string;
-    protected getFallbackModels(): string[];
-    /**
-     * DeepSeek's /chat/completions rejects `response_format: { type:
-     * "json_schema" }` outright ("This response_format type is unavailable
-     * now"). The `@ai-sdk/openai-compatible` provider this migration replaced
-     * ran with `supportsStructuredOutputs: false`, which downgraded structured-
-     * output requests to `{ type: "json_object" }`. Replicate that downgrade so
-     * `generate({ schema })` keeps working. (DeepSeek's json_object mode also
-     * requires the word "json" somewhere in the messages; the base client's
-     * `ensureJsonWordInBody` injects a minimal instruction when the prompt
-     * lacks it.)
-     */
-    protected adjustResponseFormat(rf: OpenAICompatResponseFormat | undefined, _modelId: string): OpenAICompatResponseFormat | undefined;
-}

package/dist/providers/deepseek.js

@@ -1,120 +0,0 @@
-import { DeepSeekModels } from "../constants/enums.js";
-import { AuthenticationError, InvalidModelError, ProviderError, } from "../types/index.js";
-import { logger } from "../utils/logger.js";
-import { redactUrlCredentials } from "../utils/logSanitize.js";
-import { classifyProviderError, DEFAULT_ERROR_RULES, } from "../utils/errorClassifier.js";
-import { createDeepSeekConfig, getProviderModel, validateApiKey, } from "../utils/providerConfig.js";
-import { OpenAIChatCompletionsProvider } from "./openaiChatCompletionsBase.js";
-const DEEPSEEK_BASE_URL = "https://api.deepseek.com";
-const getDeepSeekApiKey = () => {
-    return validateApiKey(createDeepSeekConfig());
-};
-const getDefaultDeepSeekModel = () => {
-    return getProviderModel("DEEPSEEK_MODEL", DeepSeekModels.DEEPSEEK_CHAT);
-};
-/**
- * DeepSeek Provider — direct HTTP, no AI SDK.
- *
- * OpenAI-compatible chat completions at api.deepseek.com (deepseek-chat /
- * deepseek-reasoner). All request/stream/tool-loop orchestration lives in
- * `OpenAIChatCompletionsProvider`; this class declares configuration and
- * provider-specific quirks:
- *
- *   1. Structured-output downgrade — DeepSeek rejects `response_format:
- *      { type: "json_schema" }` ("This response_format type is unavailable
- *      now"), so `adjustResponseFormat` downgrades it to `json_object` —
- *      matching the `supportsStructuredOutputs: false` behaviour of the
- *      `@ai-sdk/openai-compatible` path this migration replaced. The base
- *      client injects the literal "json" word the API requires for that mode.
- *
- *   2. Reasoning support — `reasoning_content` (deepseek-reasoner / R1) is
- *      surfaced automatically by the native base client: streamed deltas
- *      arrive as `{ content: "", reasoning }` chunks and the non-streaming
- *      result carries `result.reasoning`. The opt-in `thinking` request
- *      param (non-reasoner chat models) still needs thinking-signal plumbing
- *      and is tracked as a follow-up. All other behavior is preserved.
- *
- * @see https://api-docs.deepseek.com
- */
-export class DeepSeekProvider extends OpenAIChatCompletionsProvider {
-    constructor(modelName, sdk, _region, credentials) {
-        // Trim the override before applying precedence. A blank/whitespace
-        // `credentials.apiKey` must NOT bypass `getDeepSeekApiKey()` — that
-        // would build a client with an unusable bearer token and fail at
-        // request time with a confusing 401 instead of at construction time.
-        const overrideApiKey = credentials?.apiKey?.trim();
-        const apiKey = overrideApiKey && overrideApiKey.length > 0
-            ? overrideApiKey
-            : getDeepSeekApiKey();
-        // Treat blank/whitespace overrides as unset so an empty
-        // `credentials.baseURL` or `DEEPSEEK_BASE_URL=` cannot silently override
-        // the default with "" (mirrors the apiKey precedence above).
-        const baseURL = credentials?.baseURL?.trim() ||
-            process.env.DEEPSEEK_BASE_URL?.trim() ||
-            DEEPSEEK_BASE_URL;
-        super("deepseek", modelName, sdk, { baseURL, apiKey });
-        logger.debug("DeepSeek Provider initialized", {
-            modelName: this.modelName,
-            providerName: this.providerName,
-            baseURL: redactUrlCredentials(this.config.baseURL),
-        });
-    }
-    // ===========================================================================
-    // Abstract hooks (required)
-    // ===========================================================================
-    getProviderName() {
-        return "deepseek";
-    }
-    getDefaultModel() {
-        return getDefaultDeepSeekModel();
-    }
-    formatProviderError(error) {
-        const rules = [
-            {
-                match: (ctx) => ctx.statusCode === 401 ||
-                    /Invalid API key|Authentication/i.test(ctx.message),
-                errorClass: AuthenticationError,
-                message: "Invalid DeepSeek API key. Please check your DEEPSEEK_API_KEY environment variable.",
-            },
-            {
-                match: (ctx) => ctx.statusCode === 402 ||
-                    /Insufficient Balance|insufficient_balance/i.test(ctx.message),
-                errorClass: ProviderError,
-                message: "DeepSeek account has insufficient balance. Top up at https://platform.deepseek.com/usage",
-            },
-            {
-                match: (ctx) => /model_not_found/i.test(ctx.message),
-                errorClass: InvalidModelError,
-                message: (ctx) => `DeepSeek model '${ctx.modelName}' not found. Use 'deepseek-chat' or 'deepseek-reasoner'.`,
-            },
-            ...DEFAULT_ERROR_RULES,
-        ];
-        return classifyProviderError(error, rules, this.providerName, this.modelName);
-    }
-    // ===========================================================================
-    // Optional hooks — provider-specific quirks
-    // ===========================================================================
-    getFallbackModelName() {
-        return DeepSeekModels.DEEPSEEK_CHAT;
-    }
-    getFallbackModels() {
-        return [DeepSeekModels.DEEPSEEK_CHAT, DeepSeekModels.DEEPSEEK_REASONER];
-    }
-    /**
-     * DeepSeek's /chat/completions rejects `response_format: { type:
-     * "json_schema" }` outright ("This response_format type is unavailable
-     * now"). The `@ai-sdk/openai-compatible` provider this migration replaced
-     * ran with `supportsStructuredOutputs: false`, which downgraded structured-
-     * output requests to `{ type: "json_object" }`. Replicate that downgrade so
-     * `generate({ schema })` keeps working. (DeepSeek's json_object mode also
-     * requires the word "json" somewhere in the messages; the base client's
-     * `ensureJsonWordInBody` injects a minimal instruction when the prompt
-     * lacks it.)
-     */
-    adjustResponseFormat(rf, _modelId) {
-        if (rf?.type === "json_schema") {
-            return { type: "json_object" };
-        }
-        return rf;
-    }
-}

package/dist/providers/huggingFace/client.d.ts

@@ -1,54 +0,0 @@
-import type { AIProviderName } from "../../constants/enums.js";
-import type { NeurolinkCredentials } from "../../types/index.js";
-import { OpenAIChatCompletionsProvider } from "../openaiChatCompletionsBase.js";
-/**
- * HuggingFace Provider — direct HTTP, no AI SDK.
- *
- * OpenAI-compatible chat completions at router.huggingface.co/v1 (unified
- * router endpoint, 2025). Supports the full HuggingFace model hub including
- * Llama 3.x, Qwen 2.5, Mistral, DeepSeek, and tool-calling capable variants.
- * All request/stream/tool-loop orchestration lives in
- * `OpenAIChatCompletionsProvider`; this class only declares configuration
- * and provider-specific error mapping.
- *
- * @see https://huggingface.co/docs/api-inference/index
- */
-export declare class HuggingFaceProvider extends OpenAIChatCompletionsProvider {
-    constructor(modelName?: string, sdk?: unknown, _region?: string, credentials?: NeurolinkCredentials["huggingFace"]);
-    protected getProviderName(): AIProviderName;
-    protected getDefaultModel(): string;
-    protected getFallbackModelName(): string;
-    /**
-     * Tool support is resolved by the base through MODEL_REGISTRY's
-     * `modelSupports()` facade, which defaults unknown models to supported.
-     *
-     * This class used to override that with a 13-entry allowlist of model-name
-     * substrings, written for the old per-model Inference API where many
-     * endpoints rejected the OpenAI `tools` field. The router replaced that, and
-     * the list did not keep up. Measured against the 142 models
-     * `GET router.huggingface.co/v1/models` served on 2026-09-13 (the
-     * catalogue drifts, so treat the totals as a snapshot; the ratio is the
-     * point):
-     *
-     *   allowlist entries matching ANY served model   1 of 13
-     *   served models the allowlist ADMITS            1 of 142
-     *   served models the allowlist BLOCKS          141
-     *
-     * So it suppressed tool calling for 99% of the catalogue, including
-     * Qwen/Qwen2.5-72B-Instruct, which the router demonstrably tool-calls. A
-     * spread of the blocked models — GLM-5.3-Flash, gemma-4-31B-it,
-     * DeepSeek-V4-Flash — all accept `tools` with HTTP 200 and return
-     * `tool_calls`; none rejected the field.
-     *
-     * The allowlist's own comment named `microsoft/DialoGPT-medium` — which
-     * `getDefaultModel()`, not `getFallbackModelName()`, used to return — as a
-     * model that rejects `tools`. The router does not serve it under any
-     * request shape, so that path was never reachable; the default now points
-     * at a served model (see `getDefaultHuggingFaceModel`).
-     *
-     * A model that genuinely cannot use tools simply does not emit `tool_calls`,
-     * which the loop already handles, so the optimistic default costs nothing
-     * that the allowlist was protecting.
-     */
-    protected formatProviderError(error: unknown): Error;
-}

package/dist/providers/huggingFace/client.js

@@ -1,120 +0,0 @@
-import { HuggingFaceModels } from "../../constants/enums.js";
-import { AuthenticationError, InvalidModelError, ProviderError, RateLimitError, } from "../../types/index.js";
-import { classifyProviderError, DEFAULT_ERROR_RULES, } from "../../utils/errorClassifier.js";
-import { logger } from "../../utils/logger.js";
-import { redactUrlCredentials } from "../../utils/logSanitize.js";
-import { createHuggingFaceConfig, getProviderModel, validateApiKey, } from "../../utils/providerConfig.js";
-import { OpenAIChatCompletionsProvider } from "../openaiChatCompletionsBase.js";
-const HUGGINGFACE_DEFAULT_BASE_URL = "https://router.huggingface.co/v1";
-const getHuggingFaceApiKey = () => validateApiKey(createHuggingFaceConfig());
-/**
- * Must stay the same id `providerRegistry` passes as the registration default
- * (`process.env.HUGGINGFACE_MODEL || HuggingFaceModels.QWEN_2_5_72B_INSTRUCT`),
- * hence the shared constant rather than a second literal.
- *
- * The registry supplies `modelName` on every factory-created provider, so this
- * only decides a directly-constructed `new HuggingFaceProvider()`. It still
- * matters: it previously read `microsoft/DialoGPT-medium`, which the router
- * does not serve under any request shape — it answers 400 "not supported by
- * any provider you have enabled" with or without `tools` — and that stale id
- * was what the docs and setup guidance advertised as the default.
- */
-const getDefaultHuggingFaceModel = () => getProviderModel("HUGGINGFACE_MODEL", HuggingFaceModels.QWEN_2_5_72B_INSTRUCT);
-/**
- * HuggingFace Provider — direct HTTP, no AI SDK.
- *
- * OpenAI-compatible chat completions at router.huggingface.co/v1 (unified
- * router endpoint, 2025). Supports the full HuggingFace model hub including
- * Llama 3.x, Qwen 2.5, Mistral, DeepSeek, and tool-calling capable variants.
- * All request/stream/tool-loop orchestration lives in
- * `OpenAIChatCompletionsProvider`; this class only declares configuration
- * and provider-specific error mapping.
- *
- * @see https://huggingface.co/docs/api-inference/index
- */
-export class HuggingFaceProvider extends OpenAIChatCompletionsProvider {
-    constructor(modelName, sdk, _region, credentials) {
-        const apiKey = credentials?.apiKey?.trim()
-            ? credentials.apiKey.trim()
-            : getHuggingFaceApiKey();
-        // Treat blank/whitespace overrides as unset so an empty
-        // `credentials.baseURL` or `HUGGINGFACE_BASE_URL=` cannot override the
-        // default with "" (mirrors the apiKey precedence above).
-        const baseURL = credentials?.baseURL?.trim() ||
-            process.env.HUGGINGFACE_BASE_URL?.trim() ||
-            HUGGINGFACE_DEFAULT_BASE_URL;
-        super("huggingface", modelName, sdk, { baseURL, apiKey });
-        logger.debug("HuggingFaceProvider initialized", {
-            modelName: this.modelName,
-            providerName: this.providerName,
-            baseURL: redactUrlCredentials(this.config.baseURL),
-        });
-    }
-    getProviderName() {
-        return "huggingface";
-    }
-    getDefaultModel() {
-        return getDefaultHuggingFaceModel();
-    }
-    getFallbackModelName() {
-        return "meta-llama/Llama-3.1-8B-Instruct";
-    }
-    /**
-     * Tool support is resolved by the base through MODEL_REGISTRY's
-     * `modelSupports()` facade, which defaults unknown models to supported.
-     *
-     * This class used to override that with a 13-entry allowlist of model-name
-     * substrings, written for the old per-model Inference API where many
-     * endpoints rejected the OpenAI `tools` field. The router replaced that, and
-     * the list did not keep up. Measured against the 142 models
-     * `GET router.huggingface.co/v1/models` served on 2026-09-13 (the
-     * catalogue drifts, so treat the totals as a snapshot; the ratio is the
-     * point):
-     *
-     *   allowlist entries matching ANY served model   1 of 13
-     *   served models the allowlist ADMITS            1 of 142
-     *   served models the allowlist BLOCKS          141
-     *
-     * So it suppressed tool calling for 99% of the catalogue, including
-     * Qwen/Qwen2.5-72B-Instruct, which the router demonstrably tool-calls. A
-     * spread of the blocked models — GLM-5.3-Flash, gemma-4-31B-it,
-     * DeepSeek-V4-Flash — all accept `tools` with HTTP 200 and return
-     * `tool_calls`; none rejected the field.
-     *
-     * The allowlist's own comment named `microsoft/DialoGPT-medium` — which
-     * `getDefaultModel()`, not `getFallbackModelName()`, used to return — as a
-     * model that rejects `tools`. The router does not serve it under any
-     * request shape, so that path was never reachable; the default now points
-     * at a served model (see `getDefaultHuggingFaceModel`).
-     *
-     * A model that genuinely cannot use tools simply does not emit `tool_calls`,
-     * which the loop already handles, so the optimistic default costs nothing
-     * that the allowlist was protecting.
-     */
-    formatProviderError(error) {
-        const rules = [
-            {
-                match: (ctx) => /API_TOKEN_INVALID|Invalid token/.test(ctx.message),
-                errorClass: AuthenticationError,
-                message: "Invalid HuggingFace API token. Please check your HUGGINGFACE_API_KEY environment variable.",
-            },
-            {
-                match: (ctx) => /rate limit/.test(ctx.message),
-                errorClass: RateLimitError,
-                message: "HuggingFace rate limit exceeded. Consider using a paid plan or try again later.",
-            },
-            {
-                match: (ctx) => /model/.test(ctx.message) && /not found/.test(ctx.message),
-                errorClass: InvalidModelError,
-                message: () => `HuggingFace model '${this.modelName}' not found.\n\nSuggestions:\n1. Check model name spelling\n2. Ensure model exists on HuggingFace Hub\n3. For tool calling, use: Llama-3.1-8B-Instruct, Hermes-3-Llama-3.2-3B, or CodeLlama-34b-Instruct-hf`,
-            },
-            {
-                match: (ctx) => /function|tool/.test(ctx.message),
-                errorClass: ProviderError,
-                message: (ctx) => `HuggingFace tool calling error: ${ctx.message}\n\nNotes:\n1. Ensure you're using a tool-capable model (Llama-3.1+, Hermes-3+, CodeLlama)\n2. Check that your model supports function calling\n3. Verify tool schema format is correct`,
-            },
-            ...DEFAULT_ERROR_RULES,
-        ];
-        return classifyProviderError(error, rules, this.providerName, this.modelName);
-    }
-}

package/dist/providers/huggingFace/index.d.ts

@@ -1 +0,0 @@
-export * from "./client.js";

package/dist/providers/huggingFace/index.js

@@ -1 +0,0 @@
-export * from "./client.js";