@juspay/neurolink 13.1.0 → 13.2.0

package/dist/cli/commands/decide.d.ts

@@ -10,7 +10,8 @@
  * Credentials are env-only here, exactly as every other CLI command: the
  * decision provider reads its own settings (TYPESAFE_API_KEY, LAYA_API_KEY and
  * LAYA_BASE_URL, XOR_API_KEY and XOR_BASE_URL, SUBQ_API_KEY, PERPLEXITY_API_KEY,
- * CLOUDFLARE_API_KEY and CLOUDFLARE_ACCOUNT_ID, or OPENAI_API_KEY) itself. An
+ * CLOUDFLARE_API_KEY and CLOUDFLARE_ACCOUNT_ID, OPENAI_API_KEY, or
+ * SCALEDOWN_API_KEY) itself. An
  * OPENAI_API_KEY beside an OPENAI_BASE_URL, OPENAI_API_BASE or credentials.openai.baseURL that names a host
  * other than https://api.openai.com is a gateway's and is not used by default.
  */

package/dist/cli/commands/decide.js

@@ -10,7 +10,8 @@
  * Credentials are env-only here, exactly as every other CLI command: the
  * decision provider reads its own settings (TYPESAFE_API_KEY, LAYA_API_KEY and
  * LAYA_BASE_URL, XOR_API_KEY and XOR_BASE_URL, SUBQ_API_KEY, PERPLEXITY_API_KEY,
- * CLOUDFLARE_API_KEY and CLOUDFLARE_ACCOUNT_ID, or OPENAI_API_KEY) itself. An
+ * CLOUDFLARE_API_KEY and CLOUDFLARE_ACCOUNT_ID, OPENAI_API_KEY, or
+ * SCALEDOWN_API_KEY) itself. An
  * OPENAI_API_KEY beside an OPENAI_BASE_URL, OPENAI_API_BASE or credentials.openai.baseURL that names a host
  * other than https://api.openai.com is a gateway's and is not used by default.
  */

package/dist/cli/commands/setup.d.ts

@@ -13,7 +13,7 @@ import type { ProviderConfigOptions, SetupArgs } from "../../types/index.js";
  * perplexity, sambanova, together-ai, xai — mistral is the 9th catalog
  * provider but keeps its own handleMistralSetup() flow above, so it's
  * excluded here), spread directly from the catalog JSON via
- * buildCatalogConfigOptions(); decide-kind catalog entries (laya, xor, subq, openai-decisions) are
+ * buildCatalogConfigOptions(); decide-kind catalog entries (laya, xor, subq, openai-decisions, scaledown) are
  * spread the same way from getDecideCatalogEntries(). The hand rows listed
  * first reuse the existing createXConfig() factories in providerConfig.ts; the
  * remaining 5 (ollama, litellm, sagemaker, lm-studio, llamacpp) don't have a

package/dist/cli/commands/setup.js

@@ -150,7 +150,7 @@ const PROVIDERS = [
  * perplexity, sambanova, together-ai, xai — mistral is the 9th catalog
  * provider but keeps its own handleMistralSetup() flow above, so it's
  * excluded here), spread directly from the catalog JSON via
- * buildCatalogConfigOptions(); decide-kind catalog entries (laya, xor, subq, openai-decisions) are
+ * buildCatalogConfigOptions(); decide-kind catalog entries (laya, xor, subq, openai-decisions, scaledown) are
  * spread the same way from getDecideCatalogEntries(). The hand rows listed
  * first reuse the existing createXConfig() factories in providerConfig.ts; the
  * remaining 5 (ollama, litellm, sagemaker, lm-studio, llamacpp) don't have a

package/dist/cli/factories/commandFactory.js

@@ -649,7 +649,7 @@ export class CLICommandFactory {
         },
         classifierStrategy: {
             type: "string",
-            description: "Classifier strategy: 'auto' (default — 'jev' when a decision provider is configured, such as 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 unless OPENAI_BASE_URL, OPENAI_API_BASE or credentials.openai.baseURL names a host other than https://api.openai.com, else 'heuristic'), 'heuristic' (no LLM), 'llm' (a cheap model picks per prompt), or 'jev' (a decision model — whichever decision provider is configured: TypeSafe Jev, Laya, XOR, SubQ, Perplexity, Cloudflare Clef or OpenAI Decisions — with the confidence it reports).",
+            description: "Classifier strategy: 'auto' (default — 'jev' when a decision provider is configured, such as 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, OPENAI_API_KEY unless OPENAI_BASE_URL, OPENAI_API_BASE or credentials.openai.baseURL names a host other than https://api.openai.com, or SCALEDOWN_API_KEY, else 'heuristic'), 'heuristic' (no LLM), 'llm' (a cheap model picks per prompt), or 'jev' (a decision model — whichever decision provider is configured: TypeSafe Jev, Laya, XOR, SubQ, Perplexity, Cloudflare Clef, OpenAI Decisions or ScaleDown — with the confidence it reports).",
             choices: ["auto", "heuristic", "llm", "jev"],
             alias: "classifier-strategy",
         },
@@ -2303,10 +2303,14 @@ export class CLICommandFactory {
             }
             // Use SDK's provider diagnostic method instead of manual testing
             const results = await sdk.getProviderStatus({ quiet: !!argv.quiet });
+            // A configured provider whose live check the SDK skips comes back as
+            // "not-configured" with configured: true. It is neither working nor
+            // unconfigured, so it is shown as skipped and left out of the count.
+            const isSkipped = (r) => r.configured && r.status === "not-configured";
             if (spinner) {
                 const working = results.filter((r) => r.status === "working").length;
-                const configured = results.filter((r) => r.configured).length;
-                spinner.succeed(`Provider check complete: ${working}/${configured} providers working`);
+                const checked = results.filter((r) => r.configured && !isSkipped(r)).length;
+                spinner.succeed(`Provider check complete: ${working}/${checked} providers working`);
             }
             // Display results
             for (const result of results) {
@@ -2314,11 +2318,16 @@ export class CLICommandFactory {
                     ? chalk.green("✅ Working")
                     : result.status === "failed"
                         ? chalk.red("❌ Failed")
-                        : chalk.gray("⚪ Not configured");
+                        : isSkipped(result)
+                            ? chalk.gray("⏭ Skipped")
+                            : chalk.gray("⚪ Not configured");
                 const time = result.responseTime ? ` (${result.responseTime}ms)` : "";
                 const model = result.model ? ` [${result.model}]` : "";
                 logger.always(`${result.provider}: ${status}${time}${model}`);
-                if (argv.verbose && result.error) {
+                if (isSkipped(result) && result.error) {
+                    logger.always(`  ${chalk.gray(result.error)}`);
+                }
+                else if (argv.verbose && result.error) {
                     logger.always(`  Error: ${chalk.red(result.error)}`);
                 }
             }

package/dist/constants/contextWindows.d.ts

@@ -11,6 +11,13 @@
  * - Others: Provider documentation as of Feb 2026
  */
 export { stripBedrockGeoPrefix } from "./bedrockIdentifiers.js";
+/**
+ * The output-token ceiling a JSON-catalog provider declares for one model, or
+ * undefined when the provider is not a catalog provider, the model is not
+ * listed, or the listing records no maxOutputTokens for it. Never falls back
+ * to the provider-wide default (see {@link CATALOG_OUTPUT_CEILINGS}).
+ */
+export declare function getCatalogOutputCeiling(provider: string, model?: string): number | undefined;
 /** Default context window when provider/model is unknown */
 export declare const DEFAULT_CONTEXT_WINDOW = 128000;
 /** Maximum output reserve when maxTokens not specified */

package/dist/constants/contextWindows.js

@@ -37,6 +37,54 @@ const CATALOG_CONTEXT_WINDOWS = Object.fromEntries(getCatalogJsonEntries().map((
             : [])),
     },
 ]));
+/**
+ * Per-model output-token ceilings for the JSON-catalog providers, derived
+ * from models.catalog[*].maxOutputTokens and keyed by lowercased provider id
+ * and alias. Each provider's map holds every model id as written plus its
+ * lowercased form, so a lookup tries the exact spelling first.
+ *
+ * Deliberately NOT models.defaultMaxOutputTokens: for about 28 catalog
+ * providers (groq, xai, together-ai, fireworks, perplexity, cloudflare,
+ * mistral ...) that is a generic 4096 placeholder (see manifestRegistry.ts),
+ * far below the real ceilings of their models. A model with no recorded
+ * maxOutputTokens has no entry here, so it is never capped by a guess.
+ */
+const CATALOG_OUTPUT_CEILINGS = new Map(getCatalogJsonEntries().flatMap((entry) => {
+    const ceilings = Object.entries(entry.models.catalog).flatMap(([modelId, spec]) => spec.maxOutputTokens !== undefined &&
+        Number.isFinite(spec.maxOutputTokens) &&
+        spec.maxOutputTokens > 0
+        ? [[modelId, spec.maxOutputTokens]]
+        : []);
+    if (ceilings.length === 0) {
+        return [];
+    }
+    // Lowercased spellings go in first so an exact id always overrides a
+    // case-folded one; two ids differing only by case keep the smaller cap.
+    const byModel = new Map();
+    for (const [modelId, ceiling] of ceilings) {
+        const folded = modelId.toLowerCase();
+        byModel.set(folded, Math.min(byModel.get(folded) ?? ceiling, ceiling));
+    }
+    for (const [modelId, ceiling] of ceilings) {
+        byModel.set(modelId, ceiling);
+    }
+    return [entry.id, ...entry.aliases].map((name) => [name.toLowerCase(), byModel]);
+}));
+/**
+ * The output-token ceiling a JSON-catalog provider declares for one model, or
+ * undefined when the provider is not a catalog provider, the model is not
+ * listed, or the listing records no maxOutputTokens for it. Never falls back
+ * to the provider-wide default (see {@link CATALOG_OUTPUT_CEILINGS}).
+ */
+export function getCatalogOutputCeiling(provider, model) {
+    // getSafeMaxTokens has always tolerated a missing provider, so this lookup
+    // must too (callers outside the typed option normalisers can pass one).
+    if (!model || typeof provider !== "string") {
+        return undefined;
+    }
+    const ceilings = CATALOG_OUTPUT_CEILINGS.get(provider.toLowerCase());
+    return ceilings?.get(model) ?? ceilings?.get(model.toLowerCase());
+}
 /** Default context window when provider/model is unknown */
 export const DEFAULT_CONTEXT_WINDOW = 128_000;
 /** Maximum output reserve when maxTokens not specified */

package/dist/constants/enums.d.ts

@@ -85,6 +85,7 @@ export declare enum AIProviderName {
     SAKANA_AI = "sakana-ai",
     SAMBANOVA = "sambanova",
     SARVAM = "sarvam",
+    SCALEDOWN = "scaledown",
     SCALEWAY = "scaleway",
     SILICONFLOW = "siliconflow",
     STACKIT = "stackit",
@@ -1570,6 +1571,10 @@ export declare enum SarvamModels {
     SARVAM_105B = "sarvam-105b",
     SARVAM_105B_CONVERSATIONS = "sarvam-105b-conversations"
 }
+export declare enum ScaleDownModels {
+    CLASSIFY_1 = "classify-1",
+    CLASSIFY = "classify"
+}
 export declare enum ScalewayModels {
     MISTRAL_SMALL_3_2_24B_INSTRUCT_2506 = "mistral-small-3.2-24b-instruct-2506",
     LLAMA_3_3_70B_INSTRUCT = "llama-3.3-70b-instruct",

package/dist/constants/enums.js

@@ -90,6 +90,7 @@ export var AIProviderName;
     AIProviderName["SAKANA_AI"] = "sakana-ai";
     AIProviderName["SAMBANOVA"] = "sambanova";
     AIProviderName["SARVAM"] = "sarvam";
+    AIProviderName["SCALEDOWN"] = "scaledown";
     AIProviderName["SCALEWAY"] = "scaleway";
     AIProviderName["SILICONFLOW"] = "siliconflow";
     AIProviderName["STACKIT"] = "stackit";
@@ -1875,6 +1876,11 @@ export var SarvamModels;
     SarvamModels["SARVAM_105B"] = "sarvam-105b";
     SarvamModels["SARVAM_105B_CONVERSATIONS"] = "sarvam-105b-conversations";
 })(SarvamModels || (SarvamModels = {}));
+export var ScaleDownModels;
+(function (ScaleDownModels) {
+    ScaleDownModels["CLASSIFY_1"] = "classify-1";
+    ScaleDownModels["CLASSIFY"] = "classify";
+})(ScaleDownModels || (ScaleDownModels = {}));
 export var ScalewayModels;
 (function (ScalewayModels) {
     ScalewayModels["MISTRAL_SMALL_3_2_24B_INSTRUCT_2506"] = "mistral-small-3.2-24b-instruct-2506";

package/dist/factories/providerDescriptors.js

@@ -469,18 +469,22 @@ const HAND_DESCRIPTORS = [
     },
     // Decision precedence is decideOrder, not position: typesafe 1, laya 2 and
     // xor 3 (from catalog/laya.json and catalog/xor.json), subq 4 (from
-    // catalog/subq.json), perplexity-decider 5, cloudflare-clef 6 and
-    // openai-decisions 7 (catalog/openai-decisions.json).
+    // catalog/subq.json), perplexity-decider 5, cloudflare-clef 6,
+    // openai-decisions 7 (catalog/openai-decisions.json) and scaledown 8
+    // (catalog/scaledown.json).
     // resolveDefaultDecisionProvider() returns the first configured
     // DECISION_PROVIDERS entry in that order, so a host configured for several
     // keeps Jev for every built-in consumer and reaches the others only by
     // naming them. Renumbering silently changes which model routes, compacts
     // and plans for every such host.
-    // perplexity-decider, cloudflare-clef and openai-decisions come last because
-    // each shares its environment key with a text provider: a key set only for
-    // that text provider must not displace a decision provider a host set up on
+    // perplexity-decider, cloudflare-clef and openai-decisions come after subq
+    // because each shares its environment key with a text provider: a key set only
+    // for that text provider must not displace a decision provider a host set up on
     // purpose. subq is ahead of them only because SUBQ_API_KEY belongs to no other
-    // provider.
+    // provider. scaledown is last by the owner's decision of 2026-10-15, not by
+    // that rule: SCALEDOWN_API_KEY belongs to no other provider either, so an
+    // ambient OPENAI_API_KEY, PERPLEXITY_API_KEY or Cloudflare pair outranks a
+    // ScaleDown key a host set on purpose, and ScaleDown is reached by naming it.
     {
         name: AIProviderName.PERPLEXITY_DECIDER,
         aliases: [],
@@ -552,9 +556,9 @@ const HAND_DESCRIPTORS = [
         credentialsKey: "cloudflareClef",
         envVars: {
             // The same token and account id the `cloudflare` text provider reads, so
-            // ambient Workers AI settings configure this provider too. It sits last
-            // in this list, which is what keeps that from displacing any other
-            // decision provider a host has configured.
+            // ambient Workers AI settings configure this provider too. Its place in
+            // the decision order (see the comment above) keeps that from displacing
+            // TypeSafe, Laya, XOR or SubQ.
             apiKey: "CLOUDFLARE_API_KEY",
             // Cloudflare's own API is the endpoint, so a base URL is optional; but
             // the path carries the account id, so a token alone is useless. Both are

package/dist/neurolink.d.ts

@@ -1845,6 +1845,10 @@ export declare class NeuroLink {
     /**
      * Get comprehensive status of all AI providers
      * Primary method for provider health checking and diagnostics
+     *
+     * Bedrock and SageMaker are never called live: their entries come back as
+     * `status: "not-configured"` with `configured: true` and the reason in
+     * `error`. Use `testProvider()` to check either one explicitly.
      */
     getProviderStatus(options?: {
         quiet?: boolean;

package/dist/neurolink.js

@@ -11870,6 +11870,10 @@ Current user's request: ${currentInput}`;
     /**
      * Get comprehensive status of all AI providers
      * Primary method for provider health checking and diagnostics
+     *
+     * Bedrock and SageMaker are never called live: their entries come back as
+     * `status: "not-configured"` with `configured: true` and the reason in
+     * `error`. Use `testProvider()` to check either one explicitly.
      */
     async getProviderStatus(options) {
         // Track memory and timing for provider status checks
@@ -11912,6 +11916,19 @@ Current user's request: ${currentInput}`;
                         responseTime: Date.now() - startTime,
                     };
                 }
+                // These probes are billable generate() calls signed with ambient AWS
+                // credentials, against models the account may not be allowed to
+                // call, so they are never made here.
+                if (providerName === "bedrock" || providerName === "sagemaker") {
+                    return {
+                        provider: providerName,
+                        status: "not-configured",
+                        configured: true,
+                        authenticated: false,
+                        error: `Live status check is skipped for ${providerName}`,
+                        responseTime: Date.now() - startTime,
+                    };
+                }
                 // Special handling for Ollama
                 if (providerName === "ollama") {
                     let ollamaAttempt;
@@ -14281,7 +14298,13 @@ Current user's request: ${currentInput}`;
                     timeoutMs: options.timeoutMs,
                 });
                 const answerCount = Object.keys(result.answers).length;
-                const cost = calculateCost(providerName, result.model, {
+                // `providerName` is what the caller passed, so it can be an alias
+                // (`jev`, `scaledown-classify`). The pricing table is keyed by the
+                // canonical id and does not read the decide catalog, so an alias
+                // priced the decision at 0.
+                const pricedProvider = PROVIDER_ALIAS_INDEX.get(providerName.toLowerCase()) ??
+                    providerName;
+                const cost = calculateCost(pricedProvider, result.model, {
                     input: result.usage.inputTokens,
                     output: result.usage.outputTokens,
                     total: result.usage.inputTokens + result.usage.outputTokens,

package/dist/providers/catalog/charm-hyper.json

@@ -89,11 +89,10 @@
                 },
                 "vision": true,
                 "status": "production",
-                "description": "Moonshot Kimi K3 — vision-capable per the live roster; reasoning effort levels low/high/xhigh (default high). Context window, max output tokens, vision and pricing per GET https://hyper.charm.land/v1/models (unauthenticated, 2026-09-28)."
+                "description": "Moonshot Kimi K3 — vision-capable per the live roster; reasoning effort levels low/high/xhigh (default high). Context window, max output tokens, vision and pricing per GET https://hyper.charm.land/v1/models (unauthenticated, 2026-09-28). The 16000 max output tokens is Hyper's own cap for this model: Hyper's List Models docs (https://hyper.charm.land/docs/api/list-models.html) define `max_output_tokens` as the model's max output tokens, and the roster still reported 16000 on 2026-10-18. Moonshot's own API allows more (`max_completion_tokens` defaults to 131072 and can be set up to 1048576, per https://platform.kimi.ai/docs/guide/kimi-k3-quickstart), so the cap belongs to this gateway, not to the model."
             },
             "kimi-k2-thinking": {
                 "contextWindow": 262144,
-                "maxOutputTokens": 26214,
                 "pricingPerMTok": {
                     "input": 0.6,
                     "output": 2.5,
@@ -101,7 +100,7 @@
                 },
                 "vision": false,
                 "status": "production",
-                "description": "Moonshot Kimi K2 Thinking — reasoning-first Kimi variant; reasoning effort levels low/high/xhigh (default high). Context window, max output tokens, vision and pricing per GET https://hyper.charm.land/v1/models (unauthenticated, 2026-09-28)."
+                "description": "Moonshot Kimi K2 Thinking — reasoning-first Kimi variant; reasoning effort levels low/high/xhigh (default high). Context window, max output tokens, vision and pricing per GET https://hyper.charm.land/v1/models (unauthenticated, 2026-09-28). No maxOutputTokens is recorded: the roster reports 26214, which is a fixed fraction of the 262144-token context window and reads like a gateway default rather than a model limit."
             },
             "kimi-k2.7-code": {
                 "contextWindow": 256000,
@@ -113,7 +112,7 @@
                 },
                 "vision": true,
                 "status": "production",
-                "description": "Moonshot Kimi K2.7 Code — code-focused Kimi variant, vision-capable per the live roster; reasoning effort levels low/high/xhigh (default high). Context window, max output tokens, vision and pricing per GET https://hyper.charm.land/v1/models (unauthenticated, 2026-09-28)."
+                "description": "Moonshot Kimi K2.7 Code — code-focused Kimi variant, vision-capable per the live roster; reasoning effort levels low/high/xhigh (default high). Context window, max output tokens, vision and pricing per GET https://hyper.charm.land/v1/models (unauthenticated, 2026-09-28). The 16000 max output tokens is Hyper's own cap for this model: Hyper's List Models docs (https://hyper.charm.land/docs/api/list-models.html) define `max_output_tokens` as the model's max output tokens, and the roster still reported 16000 on 2026-10-18. Moonshot's own API defaults `max_tokens` to 32768 (https://platform.kimi.ai/docs/guide/kimi-k2-7-code-quickstart), so the cap belongs to this gateway, not to the model."
             },
             "glm-5.3": {
                 "contextWindow": 1000000,
@@ -165,7 +164,6 @@
             },
             "minimax-m2.7": {
                 "contextWindow": 262100,
-                "maxOutputTokens": 6553,
                 "pricingPerMTok": {
                     "input": 0.484,
                     "output": 1.852,
@@ -173,7 +171,7 @@
                 },
                 "vision": false,
                 "status": "production",
-                "description": "MiniMax M2.7 — prior MiniMax generation. Context window, max output tokens, vision and pricing per GET https://hyper.charm.land/v1/models (unauthenticated, 2026-09-28)."
+                "description": "MiniMax M2.7 — prior MiniMax generation. Context window, max output tokens, vision and pricing per GET https://hyper.charm.land/v1/models (unauthenticated, 2026-09-28). No maxOutputTokens is recorded: the roster reports 6553, which is a fixed fraction of the 262100-token context window and reads like a gateway default rather than a model limit."
             },
             "inkling": {
                 "contextWindow": 1048576,
@@ -273,7 +271,6 @@
             },
             "gpt-oss-120b": {
                 "contextWindow": 131072,
-                "maxOutputTokens": 13107,
                 "pricingPerMTok": {
                     "input": 0.178,
                     "output": 0.68,
@@ -281,11 +278,10 @@
                 },
                 "vision": false,
                 "status": "production",
-                "description": "OpenAI's open-weight gpt-oss-120b. Context window, max output tokens, vision and pricing per GET https://hyper.charm.land/v1/models (unauthenticated, 2026-09-28)."
+                "description": "OpenAI's open-weight gpt-oss-120b. Context window, max output tokens, vision and pricing per GET https://hyper.charm.land/v1/models (unauthenticated, 2026-09-28). No maxOutputTokens is recorded: the roster reports 13107, which is a fixed fraction of the 131072-token context window and reads like a gateway default rather than a model limit."
             },
             "gemma-4-26b-a4b-it": {
                 "contextWindow": 256000,
-                "maxOutputTokens": 25600,
                 "pricingPerMTok": {
                     "input": 0.098,
                     "output": 0.334,
@@ -293,7 +289,7 @@
                 },
                 "vision": false,
                 "status": "production",
-                "description": "Google Gemma 4 26B A4B, instruction-tuned. Context window, max output tokens, vision and pricing per GET https://hyper.charm.land/v1/models (unauthenticated, 2026-09-28)."
+                "description": "Google Gemma 4 26B A4B, instruction-tuned. Context window, max output tokens, vision and pricing per GET https://hyper.charm.land/v1/models (unauthenticated, 2026-09-28). No maxOutputTokens is recorded: the roster reports 25600, which is a fixed fraction of the 256000-token context window and reads like a gateway default rather than a model limit."
             }
         },
         "topModels": [

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

@@ -2,4 +2,4 @@ import type { DecideCatalogJson, ProviderCatalogJson } from "../../types/index.j
 export declare const CATALOG_JSON_ENTRIES: ProviderCatalogJson[];
 export declare const CATALOG_PROVIDER_IDS: readonly ["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", "lemonfox-ai", "lilac", "llmtech", "longcat", "mancer", "meta-model-api", "minimax", "mistral", "moark", "modelscope", "moonshot-ai", "morph", "naver-clova-studio", "nebius", "neuralwatt", "novita", "ovhcloud", "parasail", "pareto-inference", "perplexity", "poolside", "prime-intellect", "reka", "runinfra", "sakana-ai", "sambanova", "sarvam", "scaleway", "siliconflow", "stackit", "stepfun", "subconscious", "synthetic", "telnyx", "thinking-machines", "tinfoil", "together-ai", "umans-ai", "upstage", "venice-ai", "vispark", "vultr-inference", "wafer", "wandb-inference", "xai", "xiaomi-mimo", "z-ai"];
 export declare const DECIDE_CATALOG_JSON_ENTRIES: DecideCatalogJson[];
-export declare const DECIDE_CATALOG_PROVIDER_IDS: readonly ["laya", "openai-decisions", "subq", "xor"];
+export declare const DECIDE_CATALOG_PROVIDER_IDS: readonly ["laya", "openai-decisions", "scaledown", "subq", "xor"];

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

@@ -69,6 +69,7 @@ import runinfraJson from "./runinfra.json" with { type: "json" };
 import sakanaAiJson from "./sakana-ai.json" with { type: "json" };
 import sambanovaJson from "./sambanova.json" with { type: "json" };
 import sarvamJson from "./sarvam.json" with { type: "json" };
+import scaledownJson from "./scaledown.json" with { type: "json" };
 import scalewayJson from "./scaleway.json" with { type: "json" };
 import siliconflowJson from "./siliconflow.json" with { type: "json" };
 import stackitJson from "./stackit.json" with { type: "json" };
@@ -280,12 +281,14 @@ export const CATALOG_PROVIDER_IDS = [
 export const DECIDE_CATALOG_JSON_ENTRIES = [
     layaJson,
     openaiDecisionsJson,
+    scaledownJson,
     subqJson,
     xorJson,
 ].map(asDecideEntry);
 export const DECIDE_CATALOG_PROVIDER_IDS = [
     "laya",
     "openai-decisions",
+    "scaledown",
     "subq",
     "xor",
 ];

package/dist/providers/catalog/openai-decisions.json

@@ -60,7 +60,7 @@
         "instructions": [
             "1. Create an API key at https://platform.openai.com/api-keys",
             "2. Set {apiKeyEnvVar}; it is shared with the openai chat provider",
-            "3. With no other decision provider configured, {apiKeyEnvVar} alone makes openai-decisions the default decision provider (the last resort, after every other one, as PERPLEXITY_API_KEY does for perplexity-decider), unless OPENAI_BASE_URL, OPENAI_API_BASE or credentials.openai.baseURL names a host other than https://api.openai.com: the key is then taken to be a gateway's and is not used by default. To use it while another provider is configured, name it in a call (provider: \"openai-decisions\" or --provider openai-decisions), but only if the key is OpenAI's, because a named call sends the environment key to api.openai.com. Behind a gateway, route it on purpose with OPENAI_DECISIONS_BASE_URL or credentials.openaiDecisions, or do not use it",
+            "3. With no other decision provider configured, {apiKeyEnvVar} alone makes openai-decisions the default decision provider (a last resort, after every other one except ScaleDown, as PERPLEXITY_API_KEY does for perplexity-decider), unless OPENAI_BASE_URL, OPENAI_API_BASE or credentials.openai.baseURL names a host other than https://api.openai.com: the key is then taken to be a gateway's and is not used by default. To use it while another provider is configured, name it in a call (provider: \"openai-decisions\" or --provider openai-decisions), but only if the key is OpenAI's, because a named call sends the environment key to api.openai.com. Behind a gateway, route it on purpose with OPENAI_DECISIONS_BASE_URL or credentials.openaiDecisions, or do not use it",
             "4. Optional: OPENAI_DECISIONS_BASE_URL (an origin) and OPENAI_DECISIONS_MODEL (default gpt-6-luna)"
         ]
     },

package/dist/providers/catalog/provider-catalog.schema.json

@@ -142,6 +142,14 @@
             }
           }
         },
+        "authHeader": {
+          "enum": ["bearer", "x-api-key"],
+          "description": "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."
+        },
+        "stateEncoding": {
+          "enum": ["text"],
+          "description": "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."
+        },
         "media": {
           "type": "object",
           "additionalProperties": false,

package/dist/providers/catalog/scaledown.json

@@ -0,0 +1,68 @@
+{
+    "$schema": "./provider-catalog.schema.json",
+    "id": "scaledown",
+    "displayName": "ScaleDown",
+    "enumTypeName": "ScaleDownModels",
+    "aliases": ["scaledown-classify"],
+    "inferenceKinds": ["decide"],
+    "wire": { "baseURL": "https://api.scaledown.xyz" },
+    "models": {
+        "default": "classify-1",
+        "topModels": ["classify-1", "classify"],
+        "catalog": {
+            "classify-1": {
+                "contextWindow": 128000,
+                "maxOutputTokens": 1,
+                "vision": false,
+                "status": "preview",
+                "description": "Recommended - ScaleDown Classify 1, decision model for boolean, choice and score questions; context window and output limit not established"
+            },
+            "classify": {
+                "vision": false,
+                "status": "preview",
+                "description": "ScaleDown's name for its latest Classify model; whether it moves to a newer model is not established"
+            }
+        }
+    },
+    "decide": {
+        "dialect": "system-one",
+        "order": 8,
+        "baseURLNormalize": ["strip-trailing-slash"],
+        "endpoint": "{baseURL}/v1/scaledown",
+        "model": { "placement": "body" },
+        "authHeader": "x-api-key",
+        "stateEncoding": "text",
+        "errors": {
+            "rules": [
+                { "status": 401, "kind": "authentication" },
+                { "status": [402, 422], "kind": "invalid_request" },
+                { "status": 429, "kind": "rate_limit" },
+                { "status": 502, "kind": "server" },
+                { "status": 504, "kind": "timeout" },
+                { "statusMin": 500, "kind": "server" }
+            ],
+            "default": "invalid_request",
+            "bodyShapes": [{ "path": "detail.message" }, { "path": "detail" }]
+        }
+    },
+    "setup": {
+        "url": "https://docs.scaledown.ai",
+        "apiKeyFormat": null,
+        "billingPolicy": "promotional-credit",
+        "description": "API key from ScaleDown (see its documentation). Sign-up gives a free allowance; whether it expires is not established",
+        "instructions": [
+            "1. Create an account and an API key, following https://docs.scaledown.ai",
+            "2. Set {apiKeyEnvVar} in your .env file",
+            "3. ScaleDown's documentation says the key goes in an x-api-key header rather than Authorization: Bearer, and NeuroLink sends it that way",
+            "4. ScaleDown's documentation says a request made with no credit left returns HTTP 402"
+        ]
+    },
+    "evidence": {
+        "rosterVerified": {
+            "date": "2026-10-15",
+            "method": "Checked by the marketing session, not in this change: ScaleDown's docs at docs.scaledown.ai were read and probes were made against api.scaledown.xyz. No request was made to ScaleDown in this change"
+        },
+        "liveMatrix": null,
+        "addedInPR": "https://github.com/juspay/neurolink/pull/2005"
+    }
+}

package/dist/providers/catalog/schema.js

@@ -600,6 +600,8 @@ const decideBlockSchema = z.strictObject({
         placement: z.enum(["body", "body-unless-alias"]),
         alias: z.string().min(1).optional(),
     }),
+    authHeader: z.enum(["bearer", "x-api-key"]).optional(),
+    stateEncoding: z.enum(["text"]).optional(),
     media: decideMediaSchema.optional(),
     ambientKeyGuard: ambientKeyGuardSchema.optional(),
     requestIdHeaders: z.array(decideHeaderNameSchema).optional(),

package/dist/providers/configuredDecision.d.ts

@@ -15,6 +15,11 @@ export declare class ConfiguredDecisionProvider extends SystemOneDecisionProvide
     protected missingConfigMessage(): string | undefined;
     protected decisionEndpoint(): string;
     protected decisionHeaders(): Record<string, string>;
+    /**
+     * The state the dialect is given. Only an entry that sets `stateEncoding`
+     * changes it, so every other entry sends exactly what the caller passed.
+     */
+    private encodeState;
     protected buildDecisionBody(state: DecisionState, questions: Record<string, Record<string, unknown>>, model: string, media?: DecisionPreparedMedia): Record<string, unknown>;
     protected parseDecisionError(status: number, payload: unknown, requestId: string | undefined): DecisionError;
     protected readRequestId(headers: Headers): string | undefined;

package/dist/providers/configuredDecision.js

@@ -10,6 +10,9 @@ import { readPath } from "./decide/paths.js";
 import { fillEndpointTemplate } from "./decide/template.js";
 import { SystemOneDecisionProvider } from "./systemOneDecision.js";
 const readText = (value) => typeof value === "string" ? value.trim() : "";
+// JSON.stringify answers undefined for a value JSON cannot hold, such as a
+// state a JavaScript caller left out, although its own return type says string.
+const stateText = (state) => typeof state === "string" ? state : JSON.stringify(state);
 /** A `decide` provider described entirely by its catalog entry. */
 export class ConfiguredDecisionProvider extends SystemOneDecisionProvider {
     entry;
@@ -68,15 +71,34 @@ export class ConfiguredDecisionProvider extends SystemOneDecisionProvider {
         return this.endpoint;
     }
     decisionHeaders() {
-        return {
-            Authorization: `Bearer ${this.apiKey}`,
-            "Content-Type": "application/json",
-        };
+        const credential = this.entry.decide.authHeader === "x-api-key"
+            ? { "x-api-key": this.apiKey }
+            : { Authorization: `Bearer ${this.apiKey}` };
+        return { ...credential, "Content-Type": "application/json" };
+    }
+    /**
+     * The state the dialect is given. Only an entry that sets `stateEncoding`
+     * changes it, so every other entry sends exactly what the caller passed.
+     */
+    encodeState(state) {
+        if (this.entry.decide.stateEncoding !== "text") {
+            return state;
+        }
+        const text = stateText(state);
+        if (text === undefined || text.trim() === "") {
+            throw this.decisionError({
+                kind: "invalid_request",
+                // Never the state itself, which can hold anything the caller was sent.
+                message: `${this.vendorLabel()} needs non-blank state text; the state was empty or only whitespace.`,
+                retryable: false,
+            });
+        }
+        return text;
     }
     buildDecisionBody(state, questions, model, media) {
         const { placement, alias } = this.entry.decide.model;
         return this.dialect().buildBody({
-            state,
+            state: this.encodeState(state),
             questions,
             model: placement === "body-unless-alias" && model === alias
                 ? undefined