@juspay/neurolink 13.1.0 → 13.2.0

package/dist/types/classifierRouter.d.ts

@@ -24,16 +24,16 @@ export type ClassifierDifficulty = "trivial" | "simple" | "moderate" | "hard" |
  * - `heuristic` — keyword/length scoring. Deterministic, zero latency.
  * - `llm` — a cheap classifier model via the injected `generate`.
  * - `jev` — a decision model, whichever decision provider is configured
- *   (TypeSafe's Jev, Laya, XOR, SubQ, Perplexity, Cloudflare Clef or OpenAI
- *   Decisions); one round trip that returns the confidence the provider
+ *   (TypeSafe's Jev, Laya, XOR, SubQ, Perplexity, Cloudflare Clef, OpenAI
+ *   Decisions or ScaleDown); one round trip that returns the confidence the provider
  *   reports (*calibrated* on TypeSafe's Jev, ~400ms there) rather than a
  *   self-reported one.
  * - `auto` — `jev` when a decision provider is configured, in the environment
  *   or in SDK credentials (`TYPESAFE_API_KEY`, `LAYA_API_KEY` with
  *   `LAYA_BASE_URL`, `XOR_API_KEY` with `XOR_BASE_URL`, `SUBQ_API_KEY`, `PERPLEXITY_API_KEY`,
- *   `CLOUDFLARE_API_KEY` with `CLOUDFLARE_ACCOUNT_ID`, or `OPENAI_API_KEY`,
+ *   `CLOUDFLARE_API_KEY` with `CLOUDFLARE_ACCOUNT_ID`, `OPENAI_API_KEY`,
  *   which does not count while `OPENAI_BASE_URL`, `OPENAI_API_BASE` or `credentials.openai.baseURL` names a
- *   host other than `https://api.openai.com`), otherwise `heuristic`.
+ *   host other than `https://api.openai.com`, or `SCALEDOWN_API_KEY`), otherwise `heuristic`.
  */
 export type ClassifierStrategyKind = "heuristic" | "llm" | "jev" | "auto";
 /**
@@ -195,15 +195,16 @@ export type ClassifierRouterConfig = {
      * decision provider is configured, in the environment or in SDK credentials
      * (`TYPESAFE_API_KEY`, `LAYA_API_KEY` with `LAYA_BASE_URL`,
      * `XOR_API_KEY` with `XOR_BASE_URL`, `SUBQ_API_KEY`, `PERPLEXITY_API_KEY`,
-     * `CLOUDFLARE_API_KEY` with `CLOUDFLARE_ACCOUNT_ID`, or `OPENAI_API_KEY`) and
-     * "heuristic" otherwise, so configuring one upgrades routing without any code
+     * `CLOUDFLARE_API_KEY` with `CLOUDFLARE_ACCOUNT_ID`, `OPENAI_API_KEY` or
+     * `SCALEDOWN_API_KEY`) and "heuristic" otherwise, so configuring one upgrades routing without any code
      * change. Behaviour for a caller with none of these is unchanged. Three of them
      * are shared with a text provider and count here too: `PERPLEXITY_API_KEY`
      * (Perplexity's), the Cloudflare pair (Workers AI's) and `OPENAI_API_KEY` (the
      * OpenAI chat provider's). A caller whose only key is `OPENAI_API_KEY` now
-     * resolves to "jev", answered by OpenAI Decisions as the last-resort decision
-     * provider; each classification is then a paid call that still falls back to
-     * the heuristic on any failure. The exception is a key that sits beside an
+     * resolves to "jev", answered by OpenAI Decisions (the default decision
+     * provider whenever none ahead of it in the order is configured); each
+     * classification is then a paid call that still falls back to the heuristic
+     * on any failure. The exception is a key that sits beside an
      * `OPENAI_BASE_URL`, `OPENAI_API_BASE` or `credentials.openai.baseURL` naming a host other than
      * `https://api.openai.com`: that key is a gateway's and does not count.
      */

package/dist/types/cli.d.ts

@@ -1770,9 +1770,9 @@ export type CliClassifierRouterFlags = {
      * (--classifier-strategy). "auto" resolves to "jev" when a decision
      * provider is configured (TYPESAFE_API_KEY, LAYA_API_KEY with
      * LAYA_BASE_URL, XOR_API_KEY with XOR_BASE_URL, SUBQ_API_KEY, PERPLEXITY_API_KEY,
-     * CLOUDFLARE_API_KEY with CLOUDFLARE_ACCOUNT_ID, or OPENAI_API_KEY, which does
+     * CLOUDFLARE_API_KEY with CLOUDFLARE_ACCOUNT_ID, OPENAI_API_KEY, which does
      * not count beside an OPENAI_BASE_URL, OPENAI_API_BASE or credentials.openai.baseURL naming a host other
-     * than https://api.openai.com) and "heuristic" otherwise.
+     * than https://api.openai.com, or SCALEDOWN_API_KEY) and "heuristic" otherwise.
      */
     classifierStrategy?: string;
     /**

package/dist/types/providerCatalog.d.ts

@@ -25,6 +25,14 @@ export type CatalogTimeouts = {
 };
 export type CatalogModelSpec = {
     contextWindow?: number;
+    /**
+     * The model's own output-token limit. Enforced at runtime: for a provider
+     * with no PROVIDER_MAX_TOKENS row, getSafeMaxTokens clamps a larger
+     * requested maxTokens to this value. Record only a number the vendor states
+     * as a hard limit; omit it when the source is a roster field that looks
+     * derived from the context window, because an omitted value leaves the
+     * model uncapped while a wrong one silently truncates callers.
+     */
     maxOutputTokens?: number;
     pricingPerMTok?: CatalogPricingPerMTok;
     vision: boolean;
@@ -314,6 +322,20 @@ export type DecideBlock = {
     /** Request URL; `{baseURL}` is the only placeholder. */
     endpoint: string;
     model: DecideModel;
+    /**
+     * How the API key is sent. `bearer`, the default when omitted, sends
+     * `Authorization: Bearer <key>`. `x-api-key` sends `x-api-key: <key>` and no
+     * `Authorization` header at all. This is the decide request's own setting;
+     * `quirks.authHeaderStyle` belongs to the chat path and is not read here.
+     */
+    authHeader?: "bearer" | "x-api-key";
+    /**
+     * `text` makes the state that goes out always a string: a string is sent
+     * unchanged (not trimmed), any other value is sent as its JSON text, and a
+     * state that is blank after trimming is refused before any request, as an
+     * `invalid_request`. Omitted means the state goes out as given.
+     */
+    stateEncoding?: "text";
     /** Body keys for images and video, after `model`, `state` and `questions`. Omitted means the body never carries media. */
     media?: DecideMedia;
     /**

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

@@ -1,2 +1,2 @@
-export type CatalogProviderName = "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" | "laya" | "lemonfox-ai" | "lilac" | "llmtech" | "longcat" | "mancer" | "meta-model-api" | "minimax" | "mistral" | "moark" | "modelscope" | "moonshot-ai" | "morph" | "naver-clova-studio" | "nebius" | "neuralwatt" | "novita" | "openai-decisions" | "ovhcloud" | "parasail" | "pareto-inference" | "perplexity" | "poolside" | "prime-intellect" | "reka" | "runinfra" | "sakana-ai" | "sambanova" | "sarvam" | "scaleway" | "siliconflow" | "stackit" | "stepfun" | "subconscious" | "subq" | "synthetic" | "telnyx" | "thinking-machines" | "tinfoil" | "together-ai" | "umans-ai" | "upstage" | "venice-ai" | "vispark" | "vultr-inference" | "wafer" | "wandb-inference" | "xai" | "xiaomi-mimo" | "xor" | "z-ai";
-export type CatalogCredentialKey = "a2agent" | "abliteration" | "aboveDev" | "ai21" | "aiand" | "aihubmix" | "aionlabs" | "ambient" | "apiRoute" | "arcee" | "atlasCloud" | "avianIo" | "baiduQianfan" | "baseten" | "beeHeossi" | "byteplusModelark" | "bytez" | "cerebras" | "charmHyper" | "chutes" | "cloudflare" | "crusoe" | "dashscope" | "deepinfra" | "deepseek" | "empiriolabs" | "featherlessAi" | "fireworks" | "friendli" | "gmicloud" | "gradientai" | "groq" | "hetznerInference" | "huggingFace" | "inceptionLabs" | "inceptron" | "inco" | "inferenceNet" | "ioIntelligence" | "koscompute" | "laya" | "lemonfoxAi" | "lilac" | "llmtech" | "longcat" | "mancer" | "metaModelApi" | "minimax" | "mistral" | "moark" | "modelscope" | "moonshotAi" | "morph" | "naverClovaStudio" | "nebius" | "neuralwatt" | "novita" | "openaiDecisions" | "ovhcloud" | "parasail" | "paretoInference" | "perplexity" | "poolside" | "primeIntellect" | "reka" | "runinfra" | "sakanaAi" | "sambanova" | "sarvam" | "scaleway" | "siliconflow" | "stackit" | "stepfun" | "subconscious" | "subq" | "synthetic" | "telnyx" | "thinkingMachines" | "tinfoil" | "together" | "umansAi" | "upstage" | "veniceAi" | "vispark" | "vultrInference" | "wafer" | "wandbInference" | "xai" | "xiaomiMimo" | "xor" | "zAi";
+export type CatalogProviderName = "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" | "laya" | "lemonfox-ai" | "lilac" | "llmtech" | "longcat" | "mancer" | "meta-model-api" | "minimax" | "mistral" | "moark" | "modelscope" | "moonshot-ai" | "morph" | "naver-clova-studio" | "nebius" | "neuralwatt" | "novita" | "openai-decisions" | "ovhcloud" | "parasail" | "pareto-inference" | "perplexity" | "poolside" | "prime-intellect" | "reka" | "runinfra" | "sakana-ai" | "sambanova" | "sarvam" | "scaledown" | "scaleway" | "siliconflow" | "stackit" | "stepfun" | "subconscious" | "subq" | "synthetic" | "telnyx" | "thinking-machines" | "tinfoil" | "together-ai" | "umans-ai" | "upstage" | "venice-ai" | "vispark" | "vultr-inference" | "wafer" | "wandb-inference" | "xai" | "xiaomi-mimo" | "xor" | "z-ai";
+export type CatalogCredentialKey = "a2agent" | "abliteration" | "aboveDev" | "ai21" | "aiand" | "aihubmix" | "aionlabs" | "ambient" | "apiRoute" | "arcee" | "atlasCloud" | "avianIo" | "baiduQianfan" | "baseten" | "beeHeossi" | "byteplusModelark" | "bytez" | "cerebras" | "charmHyper" | "chutes" | "cloudflare" | "crusoe" | "dashscope" | "deepinfra" | "deepseek" | "empiriolabs" | "featherlessAi" | "fireworks" | "friendli" | "gmicloud" | "gradientai" | "groq" | "hetznerInference" | "huggingFace" | "inceptionLabs" | "inceptron" | "inco" | "inferenceNet" | "ioIntelligence" | "koscompute" | "laya" | "lemonfoxAi" | "lilac" | "llmtech" | "longcat" | "mancer" | "metaModelApi" | "minimax" | "mistral" | "moark" | "modelscope" | "moonshotAi" | "morph" | "naverClovaStudio" | "nebius" | "neuralwatt" | "novita" | "openaiDecisions" | "ovhcloud" | "parasail" | "paretoInference" | "perplexity" | "poolside" | "primeIntellect" | "reka" | "runinfra" | "sakanaAi" | "sambanova" | "sarvam" | "scaledown" | "scaleway" | "siliconflow" | "stackit" | "stepfun" | "subconscious" | "subq" | "synthetic" | "telnyx" | "thinkingMachines" | "tinfoil" | "together" | "umansAi" | "upstage" | "veniceAi" | "vispark" | "vultrInference" | "wafer" | "wandbInference" | "xai" | "xiaomiMimo" | "xor" | "zAi";

package/dist/types/providers.d.ts

@@ -469,6 +469,10 @@ export type NeurolinkCredentials = {
         apiKey?: string;
         baseURL?: string;
     };
+    scaledown?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
     scaleway?: {
         apiKey?: string;
         baseURL?: string;

package/dist/utils/pricing.js

@@ -812,6 +812,18 @@ const PRICING = {
         // name, so a `"gpt-6-luna"` key would price that chat model at this rate.
         _default: { input: 0.1 / 1_000_000, output: 0 },
     },
+    scaledown: {
+        // $0.05 per million input tokens, output free: `output: 0` is a price, not
+        // a token count, so do not "correct" it from `usage.output_tokens`. A list
+        // price from ScaleDown's launch post and from `usage.cost` in a response,
+        // both as reported to this project on 2026-10-15; this project has not
+        // priced a call itself. Per ScaleDown's docs each choice option is a
+        // separate internal call, so input tokens grow with options times state
+        // length (not measured here). Only `_default`: LiteLLM, OpenRouter and
+        // openai-compatible search every table by model name, so a `"classify"` key
+        // would price an unrelated model at this rate.
+        _default: { input: 0.05 / 1_000_000, output: 0 },
+    },
     "cloudflare-clef": {
         // Workers AI bills Clef per million INPUT tokens ($0.24 for clef, $0.09 for
         // clef-flash); no output price is listed. `output: 0` is a price, not a token

package/dist/utils/tokenLimits.js

@@ -2,7 +2,7 @@
  * Provider-specific token limit utilities
  * Provides safe maxTokens values based on provider and model capabilities
  */
-import { getRuntimeOutputCeiling } from "../constants/contextWindows.js";
+import { getCatalogOutputCeiling, getRuntimeOutputCeiling, } from "../constants/contextWindows.js";
 import { bedrockModelIdForLookup, stripBedrockGeoPrefix, } from "../constants/bedrockIdentifiers.js";
 import { PROVIDER_MAX_TOKENS } from "../core/constants.js";
 import { logger } from "./logger.js";
@@ -68,6 +68,25 @@ export function getSafeMaxTokens(provider, model, requestedMaxTokens) {
     // Get provider-specific limits
     const providerLimits = PROVIDER_MAX_TOKENS[canonicalProvider];
     if (!providerLimits) {
+        // A JSON-catalog provider has no PROVIDER_MAX_TOKENS row, but its catalog
+        // entry can declare the model's own output ceiling. Only that per-model
+        // value counts: the provider-wide default is a 4096 placeholder for many
+        // providers. An over-large request is clamped. Unlike the runtime ceiling
+        // above, this ceiling is not injected as a default: an unset request stays
+        // unset, because a ceiling-sized default would shrink usable input on
+        // total-context backends.
+        const catalogCeiling = getCatalogOutputCeiling(provider, model);
+        if (catalogCeiling !== undefined) {
+            if (requestedMaxTokens === undefined || requestedMaxTokens === null) {
+                return undefined;
+            }
+            if (requestedMaxTokens > catalogCeiling) {
+                logger.warn(`Requested maxTokens ${requestedMaxTokens} exceeds the catalog output ceiling ` +
+                    `of ${catalogCeiling} for ${provider}/${model}. Using ${catalogCeiling} instead.`);
+                return catalogCeiling;
+            }
+            return requestedMaxTokens;
+        }
         logger.warn(`Unknown provider ${provider}, no token limits enforced`);
         // Explicit undefined/null check so a caller-supplied 0 is preserved
         // (truthy guard would silently drop 0 here as it does in the restricted branch).