@juspay/neurolink 12.37.4 → 12.38.0

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

@@ -8,8 +8,8 @@
  * for the full vocabulary.
  *
  * Credentials are env-only here, exactly as every other CLI command: the
- * decision provider reads its own settings (TYPESAFE_API_KEY, or LAYA_API_KEY
- * and LAYA_BASE_URL) itself.
+ * decision provider reads its own settings (TYPESAFE_API_KEY, LAYA_API_KEY and
+ * LAYA_BASE_URL, or XOR_API_KEY and XOR_BASE_URL) itself.
  */
 import type { CommandModule } from "yargs";
 import type { CliDecideArgs } from "../../types/index.js";

package/dist/cli/commands/decide.js

@@ -8,8 +8,8 @@
  * for the full vocabulary.
  *
  * Credentials are env-only here, exactly as every other CLI command: the
- * decision provider reads its own settings (TYPESAFE_API_KEY, or LAYA_API_KEY
- * and LAYA_BASE_URL) itself.
+ * decision provider reads its own settings (TYPESAFE_API_KEY, LAYA_API_KEY and
+ * LAYA_BASE_URL, or XOR_API_KEY and XOR_BASE_URL) itself.
  */
 import chalk from "chalk";
 import ora from "ora";
@@ -186,6 +186,17 @@ export const decideCommand = {
         .option("model", {
         type: "string",
         describe: "Overrides the provider's configured model for this call",
+    })
+        .option("image", {
+        type: "string",
+        array: true,
+        nargs: 1,
+        describe: "Image for the model to read: a file path or a data: URL. Repeat for several (XOR takes up to 8)",
+    })
+        .option("video", {
+        type: "string",
+        nargs: 1,
+        describe: "One video for the model to read: a file path or a data: URL",
     })
         .option("timeout", {
         type: "number",
@@ -205,7 +216,8 @@ export const decideCommand = {
         describe: "Enable debug logging (written to stderr when --format json is used)",
     })
         .example('$0 decide "Refund request for a damaged item" --questions \'{"urgent":{"type":"boolean","instructions":"Is this urgent?"}}\'', "Ask a single yes/no question")
-        .example("$0 decide --state-file ticket.json --questions-file questions.json --format json", "Read state and questions from files, emit raw JSON"),
+        .example("$0 decide --state-file ticket.json --questions-file questions.json --format json", "Read state and questions from files, emit raw JSON")
+        .example('$0 decide "What color is this?" --provider xor --image ./photo.png --questions \'{"color":{"type":"choice","instructions":"What color is the image?","criteria":{"red":"red","blue":"blue"}}}\'', "Ask about an image (XOR only)"),
     handler: async (argv) => {
         const outputFormat = argv.format ?? "text";
         // --- Validate before any provider work ---------------------------------
@@ -235,6 +247,11 @@ export const decideCommand = {
             logger.error(chalk.red("Error: --timeout must be a positive number of milliseconds."));
             process.exit(1);
         }
+        const videos = [].concat(argv.video ?? []);
+        if (videos.length > 1) {
+            logger.error(chalk.red("Error: pass at most one --video."));
+            process.exit(1);
+        }
         if (!isDecisionQuestionMap(parsedQuestions)) {
             logger.error(chalk.red('Error: questions must be a non-empty object of { type: "boolean" | "choice" | "score", ... }.'));
             process.exit(1);
@@ -261,6 +278,8 @@ export const decideCommand = {
                 provider: argv.provider,
                 model: argv.model,
                 timeoutMs: argv.timeout,
+                ...(argv.image && argv.image.length > 0 ? { images: argv.image } : {}),
+                ...(videos.length === 1 ? { video: videos[0] } : {}),
             });
             spinner?.succeed("Decision complete");
             if (outputFormat === "json") {

package/dist/cli/commands/setup.js

@@ -20,7 +20,7 @@ import { handleGCPSetup } from "./setup-gcp.js";
 import { handleHuggingFaceSetup } from "./setup-huggingface.js";
 import { handleMistralSetup } from "./setup-mistral.js";
 import { PROVIDER_DESCRIPTORS_BY_NAME } from "../../factories/providerDescriptors.js";
-import { createCohereConfig, createDeepSeekConfig, createIdeogramConfig, createJinaConfig, createNvidiaNimConfig, createOpenAICompatibleConfig, createRecraftConfig, createReplicateConfig, createStabilityConfig, createVoyageConfig, createTypeSafeConfig, createLayaConfig, satisfiesFallbacks, } from "../../utils/providerConfig.js";
+import { createCohereConfig, createDeepSeekConfig, createIdeogramConfig, createJinaConfig, createNvidiaNimConfig, createOpenAICompatibleConfig, createRecraftConfig, createReplicateConfig, createStabilityConfig, createVoyageConfig, createTypeSafeConfig, createLayaConfig, createXorConfig, satisfiesFallbacks, } from "../../utils/providerConfig.js";
 import { getCatalogJsonEntries, buildCatalogConfigOptions, } from "../../providers/catalog/loader.js";
 // Provider information database
 const PROVIDERS = [
@@ -170,6 +170,7 @@ export const EXTRA_PROVIDER_CONFIGS = {
     recraft: createRecraftConfig(),
     typesafe: createTypeSafeConfig(),
     laya: createLayaConfig(),
+    xor: createXorConfig(),
     ...Object.fromEntries(getCatalogJsonEntries()
         .filter((e) => e.id !== "mistral")
         .map((e) => [e.id, buildCatalogConfigOptions(e)])),

package/dist/cli/factories/commandFactory.js

@@ -616,7 +616,7 @@ export class CLICommandFactory {
         },
         classifierStrategy: {
             type: "string",
-            description: "Classifier strategy: 'auto' (default — 'jev' when a decision provider is configured, such as TYPESAFE_API_KEY, or LAYA_API_KEY with LAYA_BASE_URL, else 'heuristic'), 'heuristic' (no LLM), 'llm' (a cheap model picks per prompt), or 'jev' (a System One decision model — TypeSafe Jev, or Laya — with calibrated confidence).",
+            description: "Classifier strategy: 'auto' (default — 'jev' when a decision provider is configured, such as TYPESAFE_API_KEY, LAYA_API_KEY with LAYA_BASE_URL, or XOR_API_KEY with XOR_BASE_URL, else 'heuristic'), 'heuristic' (no LLM), 'llm' (a cheap model picks per prompt), or 'jev' (a System One decision model — TypeSafe Jev, Laya or XOR — with calibrated confidence).",
             choices: ["auto", "heuristic", "llm", "jev"],
             alias: "classifier-strategy",
         },

package/dist/constants/enums.d.ts

@@ -48,6 +48,8 @@ export declare enum AIProviderName {
     TYPESAFE = "typesafe",
     /** Laya (Convai, open weights) — serves the `decide` inference type only. */
     LAYA = "laya",
+    /** XOR (Juspay, open weights) — serves the `decide` inference type only. */
+    XOR = "xor",
     AUTO = "auto"
 }
 /**
@@ -1097,3 +1099,10 @@ export declare enum LayaModels {
     MULTILINGUAL = "multilingual",
     AUTO = "auto"
 }
+/**
+ * XOR decision models. Hand-written: XOR is a Tier-3 provider, so it is not in
+ * the provider catalog and codegen never touches this.
+ */
+export declare enum XorModels {
+    XOR_1_1 = "xor-1.1"
+}

package/dist/constants/enums.js

@@ -54,6 +54,8 @@ export var AIProviderName;
     AIProviderName["TYPESAFE"] = "typesafe";
     /** Laya (Convai, open weights) — serves the `decide` inference type only. */
     AIProviderName["LAYA"] = "laya";
+    /** XOR (Juspay, open weights) — serves the `decide` inference type only. */
+    AIProviderName["XOR"] = "xor";
     AIProviderName["AUTO"] = "auto";
 })(AIProviderName || (AIProviderName = {}));
 /**
@@ -1357,3 +1359,11 @@ export var LayaModels;
     LayaModels["MULTILINGUAL"] = "multilingual";
     LayaModels["AUTO"] = "auto";
 })(LayaModels || (LayaModels = {}));
+/**
+ * XOR decision models. Hand-written: XOR is a Tier-3 provider, so it is not in
+ * the provider catalog and codegen never touches this.
+ */
+export var XorModels;
+(function (XorModels) {
+    XorModels["XOR_1_1"] = "xor-1.1";
+})(XorModels || (XorModels = {}));

package/dist/core/baseProvider.d.ts

@@ -262,12 +262,21 @@ export declare abstract class BaseProvider implements AIProvider {
      * and a retired default model stops degrading to the next live one in the
      * catalog — which is exactly what the provider contract gate checks.
      */
-    protected runGenerateWithModelFallback(attempt: () => Promise<EnhancedGenerateResult | null>, callerOwnsFallback: boolean): Promise<EnhancedGenerateResult | null>;
+    protected runGenerateWithModelFallback(attempt: () => Promise<EnhancedGenerateResult | null>, callerOwnsFallback: boolean, options?: TextGenerationOptions): Promise<EnhancedGenerateResult | null>;
     /**
      * Model ids to try when this provider rejects its current model as invalid.
      * Empty by default; catalog-driven providers return their `fallbacks`.
      */
     protected getModelFallbacks(): string[];
+    /**
+     * The fallbacks an invalid-model retry may try for this request. A catalog
+     * orders `fallbacks` for text, so a request carrying an image keeps only the
+     * vision-capable ones: a text-only candidate fails the vision guard, and
+     * that error names the fallback instead of the retired model the caller
+     * asked for. With no vision-capable candidate left, the caller gets the
+     * original invalid-model error.
+     */
+    protected modelFallbacksFor(options: TextGenerationOptions | StreamOptions, requestedModel: string): string[];
     /**
      * Alias for generate method - implements AIProvider interface
      */

package/dist/core/baseProvider.js

@@ -46,6 +46,7 @@ import { StreamHandler } from "./modules/StreamHandler.js";
 import { TelemetryHandler } from "./modules/TelemetryHandler.js";
 import { ToolsManager } from "./modules/ToolsManager.js";
 import { Utilities } from "./modules/Utilities.js";
+import { ProviderImageAdapter } from "../adapters/providerImageAdapter.js";
 import { generateOnceNative } from "../utils/nativeSingleShot.js";
 import { validateExecutionControl } from "../utils/parameterValidation.js";
 import { extractTokenUsage } from "../utils/tokenUtils.js";
@@ -107,6 +108,18 @@ function isProviderErrorClassified(error) {
     }
     return error[PROVIDER_ERROR_CLASSIFIED] === true;
 }
+function requestCarriesImages(options) {
+    const input = options.input;
+    if (!input) {
+        return false;
+    }
+    if ((input.images?.length ?? 0) > 0) {
+        return true;
+    }
+    return ("content" in input &&
+        Array.isArray(input.content) &&
+        input.content.some((part) => part.type === "image"));
+}
 export class BaseProvider {
     // Not `readonly` because providers that auto-discover the model from a
     // /v1/models endpoint (lm-studio, llamacpp) need to update modelName after
@@ -523,9 +536,7 @@ export class BaseProvider {
                     throw error;
                 }
                 const requestedModel = provider.modelName;
-                const candidates = provider
-                    .getModelFallbacks()
-                    .filter((model) => model !== requestedModel);
+                const candidates = provider.modelFallbacksFor(options, requestedModel);
                 for (const candidate of candidates) {
                     logger.warn(`[${provider.providerName}] model "${requestedModel}" was rejected as invalid — retrying stream with fallback "${candidate}". This provider's catalog entry is stale; run "pnpm run check:models".`);
                     provider.refreshHandlersForModel(candidate);
@@ -671,7 +682,7 @@ export class BaseProvider {
             return undefined;
         }
         const requestedModel = this.modelName;
-        const candidates = this.getModelFallbacks().filter((model) => model !== requestedModel);
+        const candidates = this.modelFallbacksFor(options, requestedModel);
         for (const candidate of candidates) {
             logger.warn(`[${this.providerName}] model "${requestedModel}" was rejected as invalid — retrying stream with fallback "${candidate}". This provider's catalog entry is stale; run "pnpm run check:models".`);
             this.refreshHandlersForModel(candidate);
@@ -1493,7 +1504,7 @@ export class BaseProvider {
         // or a router that retries on its own) pass the flag on both paths;
         // TextGenerationOptions declares it, so a plain read is enough here.
         const callerOwnsFallback = options.disableInternalFallback === true;
-        return await this.runGenerateWithModelFallback(attempt, callerOwnsFallback);
+        return await this.runGenerateWithModelFallback(attempt, callerOwnsFallback, options);
     }
     /**
      * Models the catalog names are the models a vendor served when the entry was
@@ -1520,7 +1531,7 @@ export class BaseProvider {
      * and a retired default model stops degrading to the next live one in the
      * catalog — which is exactly what the provider contract gate checks.
      */
-    async runGenerateWithModelFallback(attempt, callerOwnsFallback) {
+    async runGenerateWithModelFallback(attempt, callerOwnsFallback, options) {
         const requestedModel = this.modelName;
         try {
             return await attempt();
@@ -1529,7 +1540,9 @@ export class BaseProvider {
             if (callerOwnsFallback || !isInvalidModelError(error)) {
                 throw error;
             }
-            const candidates = this.getModelFallbacks().filter((model) => model !== requestedModel);
+            const candidates = options
+                ? this.modelFallbacksFor(options, requestedModel)
+                : this.getModelFallbacks().filter((model) => model !== requestedModel);
             if (candidates.length === 0) {
                 throw error;
             }
@@ -1563,6 +1576,21 @@ export class BaseProvider {
     getModelFallbacks() {
         return [];
     }
+    /**
+     * The fallbacks an invalid-model retry may try for this request. A catalog
+     * orders `fallbacks` for text, so a request carrying an image keeps only the
+     * vision-capable ones: a text-only candidate fails the vision guard, and
+     * that error names the fallback instead of the retired model the caller
+     * asked for. With no vision-capable candidate left, the caller gets the
+     * original invalid-model error.
+     */
+    modelFallbacksFor(options, requestedModel) {
+        const candidates = this.getModelFallbacks().filter((model) => model !== requestedModel);
+        if (!requestCarriesImages(options)) {
+            return candidates;
+        }
+        return candidates.filter((model) => ProviderImageAdapter.supportsVision(this.providerName, model));
+    }
     /**
      * Alias for generate method - implements AIProvider interface
      */

package/dist/factories/providerDescriptors.d.ts

@@ -30,6 +30,8 @@ export declare const DECISION_PROVIDERS: readonly ProviderDescriptor[];
  * internal consumer treats as "carry on exactly as before".
  */
 export declare function resolveDefaultDecisionProvider(credentials?: NeurolinkCredentials): string | undefined;
+/** Names of the decision providers that accept images or video. */
+export declare function listMediaDecisionProviders(): string[];
 /**
  * The variables that would configure each decision provider, one clause per
  * provider in precedence order, e.g. "TYPESAFE_API_KEY or AI_GATEWAY_API_KEY

package/dist/factories/providerDescriptors.js

@@ -1,5 +1,5 @@
 import { AIProviderName } from "../constants/enums.js";
-import { GoogleAIModels, OpenAIModels, AnthropicModels, VertexModels, OllamaModels, LiteLLMModels, NvidiaNimModels, OpenRouterModels, CohereModels, VoyageModels, TypeSafeModels, LayaModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, ReplicateModels, } from "../constants/enums.js";
+import { GoogleAIModels, OpenAIModels, AnthropicModels, VertexModels, OllamaModels, LiteLLMModels, NvidiaNimModels, OpenRouterModels, CohereModels, VoyageModels, TypeSafeModels, LayaModels, XorModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, ReplicateModels, } from "../constants/enums.js";
 import { API_KEY_FORMATS } from "../utils/providerConfig.js";
 import { getCatalogJsonEntries, catalogCredentialsKey, catalogEnvVar, } from "../providers/catalog/loader.js";
 import { DEFAULT_INFERENCE_KINDS } from "../types/index.js";
@@ -451,11 +451,12 @@ const HAND_DESCRIPTORS = [
         timeouts: { decideMs: 5000 },
         setupUrl: "https://console.typesafe.ai/keys",
     },
-    // Laya MUST stay after TypeSafe. resolveDefaultDecisionProvider() returns
-    // the first configured DECISION_PROVIDERS entry, in this order, so a host
-    // configured for both keeps Jev for every built-in consumer and reaches
-    // Laya only by naming it. Reordering these two silently changes which model
-    // routes, compacts and plans for every such host.
+    // Laya MUST stay after TypeSafe, and XOR after Laya.
+    // resolveDefaultDecisionProvider() returns the first configured
+    // DECISION_PROVIDERS entry, in this order, so a host configured for several
+    // keeps Jev for every built-in consumer and reaches the others only by
+    // naming them. Reordering these silently changes which model routes,
+    // compacts and plans for every such host.
     {
         name: AIProviderName.LAYA,
         aliases: [],
@@ -501,6 +502,47 @@ const HAND_DESCRIPTORS = [
         },
         setupUrl: "https://github.com/NandhaKishorM/laya",
     },
+    {
+        name: AIProviderName.XOR,
+        aliases: [],
+        credentialsKey: "xor",
+        envVars: {
+            apiKey: "XOR_API_KEY",
+            baseURL: "XOR_BASE_URL",
+            // XOR has no built-in endpoint (it is reached at a deployment or behind a
+            // proxy), so a key is useless without a base URL. Both are required for
+            // it to count as configured, from the environment or credentials.xor.
+            extraRequired: ["XOR_BASE_URL"],
+            model: "XOR_MODEL",
+        },
+        defaultModel: XorModels.XOR_1_1,
+        // Serves only `decide`, like TypeSafe and Laya — the one declaration that
+        // keeps it out of every generation code path.
+        inferenceKinds: ["decide"],
+        toolSupport: "none",
+        localRuntime: false,
+        healthCheck: "env-only",
+        // Deliberately NO autoSelectPriority / autoSelectPreference /
+        // defaultHealthSweepPriority, for the same reason as TypeSafe above.
+        timeouts: { decideMs: 5000 },
+        // The deployment's prefill is 250,000 tokens, shared by the state, the
+        // questions and any media. 200,000 leaves 20% for those, and a state
+        // between this and the server's 4 MB would otherwise pass validation and
+        // fail inside the engine as a 500, which the base class would retry. There
+        // is no question cap (the server only limits a choice to 2..255 options).
+        // Non-ASCII text is charged per character because the four-characters-
+        // per-token estimate is several times too generous for CJK on this
+        // tokenizer. One token per character is a deliberate upper bound, to be
+        // replaced by a live measurement.
+        decisionLimits: {
+            maxStateTokens: 200_000,
+            nonAsciiTokensPerChar: 1,
+            // The server takes 1..8 images and one video, and refuses a body over
+            // 8 MB.
+            media: { maxImages: 8, video: true, maxRequestBytes: 8 * 1024 * 1024 },
+        },
+        setupUrl: "https://huggingface.co/juspay/xor",
+    },
 ];
 /**
  * Builds a ProviderDescriptor for every JSON-catalog provider. Every field
@@ -636,6 +678,10 @@ function isDecisionProviderConfigured(descriptor, credentials) {
 export function resolveDefaultDecisionProvider(credentials) {
     return DECISION_PROVIDERS.find((descriptor) => isDecisionProviderConfigured(descriptor, credentials))?.name;
 }
+/** Names of the decision providers that accept images or video. */
+export function listMediaDecisionProviders() {
+    return DECISION_PROVIDERS.filter((descriptor) => descriptor.decisionLimits?.media !== undefined).map((descriptor) => descriptor.name);
+}
 /**
  * The variables that would configure each decision provider, one clause per
  * provider in precedence order, e.g. "TYPESAFE_API_KEY or AI_GATEWAY_API_KEY

package/dist/factories/providerRegistry.js

@@ -1,6 +1,6 @@
 import { ProviderFactory } from "./providerFactory.js";
 import { logger } from "../utils/logger.js";
-import { AIProviderName, GoogleAIModels, OpenAIModels, AnthropicModels, VertexModels, OllamaModels, LiteLLMModels, NvidiaNimModels, OpenRouterModels, CohereModels, VoyageModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, TypeSafeModels, LayaModels, ReplicateModels, } from "../constants/enums.js";
+import { AIProviderName, GoogleAIModels, OpenAIModels, AnthropicModels, VertexModels, OllamaModels, LiteLLMModels, NvidiaNimModels, OpenRouterModels, CohereModels, VoyageModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, TypeSafeModels, LayaModels, XorModels, ReplicateModels, } from "../constants/enums.js";
 import { PROVIDER_DESCRIPTORS_BY_NAME } from "./providerDescriptors.js";
 import { OPENAI_COMPAT_CATALOG } from "../providers/openaiCompatCatalog.js";
 import { providerChoicesFor } from "./mediaHandlerCatalog.js";
@@ -242,6 +242,14 @@ export class ProviderRegistry {
                 const { LayaProvider } = await import("../providers/laya.js");
                 return new LayaProvider(modelName, sdk, undefined, layaCreds);
             }, process.env.LAYA_MODEL || LayaModels.TYPED_DECISIONS, [], PROVIDER_DESCRIPTORS_BY_NAME.get(AIProviderName.LAYA));
+            // Register XOR — the third `decide` provider, reached at whatever base
+            // URL its config names. Its descriptor declares inferenceKinds:
+            // ["decide"], so nothing in the generation fallback chain can reach it.
+            ProviderFactory.registerProvider(AIProviderName.XOR, async (modelName, _providerName, sdk, _region, credentials) => {
+                const xorCreds = credentials;
+                const { XorProvider } = await import("../providers/xor.js");
+                return new XorProvider(modelName, sdk, undefined, xorCreds);
+            }, process.env.XOR_MODEL || XorModels.XOR_1_1, [], PROVIDER_DESCRIPTORS_BY_NAME.get(AIProviderName.XOR));
             logger.debug("All AI providers registered successfully");
             // ===== MEDIA HANDLER REGISTRATION =====
             // Single registration path (Task 11): each ecosystem barrel (voice,

package/dist/models/manifestRegistry.js

@@ -21,6 +21,7 @@ import { xaiManifest } from "./manifests/xai.js";
 import { groqManifest } from "./manifests/groq.js";
 import { typesafeManifest } from "./manifests/typesafe.js";
 import { layaManifest } from "./manifests/laya.js";
+import { xorManifest } from "./manifests/xor.js";
 import { cohereManifest } from "./manifests/cohere.js";
 import { togetherAiManifest } from "./manifests/together-ai.js";
 import { fireworksManifest } from "./manifests/fireworks.js";
@@ -120,6 +121,7 @@ export const MANIFEST_REGISTRY = {
     groq: groqManifest,
     typesafe: typesafeManifest,
     laya: layaManifest,
+    xor: xorManifest,
     cerebras: catalogManifest("cerebras"),
     sambanova: catalogManifest("sambanova"),
     cohere: cohereManifest,

package/dist/models/manifests/anthropic.js

@@ -38,7 +38,7 @@ export const anthropicManifest = {
             samplingParams: false, // matches SAMPLING_PARAM_REJECTING_FAMILIES /opus[-_.]?5(?![0-9])/i
         },
         "claude-sonnet-5-5": {
-            aliases: ["sonnet-5.5"],
+            aliases: ["sonnet-5.5", "claude-sonnet"],
             displayName: "Claude Sonnet 5.5",
             contextWindow: 1_000_000,
             maxOutputTokens: 128_000,
@@ -66,7 +66,7 @@ export const anthropicManifest = {
             samplingParams: false, // matches SAMPLING_PARAM_REJECTING_FAMILIES /fable/i
         },
         "claude-sonnet-5": {
-            aliases: ["sonnet-5", "claude-sonnet"],
+            aliases: ["sonnet-5"],
             displayName: "Claude Sonnet 5",
             contextWindow: 1_000_000,
             // Was 64_000; corrected to 128_000 to match the official spec

package/dist/models/manifests/xor.d.ts

@@ -0,0 +1,12 @@
+import type { ProviderModelManifest } from "../../types/index.js";
+/**
+ * XOR — the `decide` inference type, not text generation.
+ *
+ * `contextWindow` is the deployment's maximum prefill, shared by the state,
+ * the questions and any images or video. `vision` describes `generate()`
+ * input, which XOR does not serve; its image input is `DecisionRequest.images`.
+ *
+ * `maxOutputTokens` is required by the manifest type but has no honest value
+ * for a model that emits no text, so a small non-zero figure is used.
+ */
+export declare const xorManifest: ProviderModelManifest;

package/dist/models/manifests/xor.js

@@ -0,0 +1,30 @@
+/**
+ * XOR — the `decide` inference type, not text generation.
+ *
+ * `contextWindow` is the deployment's maximum prefill, shared by the state,
+ * the questions and any images or video. `vision` describes `generate()`
+ * input, which XOR does not serve; its image input is `DecisionRequest.images`.
+ *
+ * `maxOutputTokens` is required by the manifest type but has no honest value
+ * for a model that emits no text, so a small non-zero figure is used.
+ */
+export const xorManifest = {
+    defaultContextWindow: 250_000,
+    models: {
+        _default: {
+            aliases: [],
+            contextWindow: 250_000,
+            maxOutputTokens: 256,
+            vision: false,
+            functionCalling: false,
+        },
+        "xor-1.1": {
+            aliases: [],
+            displayName: "XOR 1.1",
+            contextWindow: 250_000,
+            maxOutputTokens: 256,
+            vision: false,
+            functionCalling: false,
+        },
+    },
+};

package/dist/models/modelRegistry.d.ts

@@ -4,7 +4,7 @@
  * Part of Phase 4.1 - Models Command System
  */
 import { AIProviderName } from "../constants/enums.js";
-import type { JsonValue, ModelCapabilities, ModelInfo } from "../types/index.js";
+import type { ClaudeDisabledThinkingReplacement, JsonValue, ModelCapabilities, ModelInfo } from "../types/index.js";
 /**
  * Public MODEL_REGISTRY: the manifest-derived entries merged over the
  * hand-authored LEGACY_MODEL_REGISTRY base. Spread order matters — manifest
@@ -90,6 +90,17 @@ export declare function resolveSamplingParams(provider: string, model: string |
     temperature?: number;
     topP?: number;
 };
+/**
+ * Whether a Claude model accepts a forced `tool_choice` (`any` or a named
+ * `tool`). Unknown models default to accepting it.
+ */
+export declare function modelSupportsForcedToolChoice(model: string | undefined): boolean;
+/**
+ * What to send instead of `thinking: {type: "disabled"}` to a Claude model.
+ * `between_tools` is refused at `xhigh` and `max` effort, so those fall back
+ * to adaptive thinking (`omit`).
+ */
+export declare function claudeDisabledThinkingReplacement(model: string | undefined, effort: string | undefined): ClaudeDisabledThinkingReplacement;
 /**
  * Get models by provider
  */

package/dist/models/modelRegistry.js

@@ -1168,14 +1168,14 @@ const LEGACY_MODEL_REGISTRY = {
             translation: 9,
             summarization: 9,
         },
-        aliases: ["sonnet-5", "claude-sonnet-latest"],
+        aliases: ["sonnet-5"],
         deprecated: false,
         isLocal: false,
         releaseDate: "2026-06-30",
         category: "general",
     },
-    // "claude-sonnet-latest" stays on CLAUDE_SONNET_5 until someone decides to
-    // move it; this entry sits after that one so resolver ties still go to it.
+    // Owns "claude-sonnet-latest", but sits after CLAUDE_SONNET_5 so resolver
+    // score ties still go to Sonnet 5.
     [AnthropicModels.CLAUDE_SONNET_5_5]: {
         id: AnthropicModels.CLAUDE_SONNET_5_5,
         name: "Claude Sonnet 5.5",
@@ -1210,7 +1210,7 @@ const LEGACY_MODEL_REGISTRY = {
             translation: 9,
             summarization: 9,
         },
-        aliases: ["sonnet-5.5"],
+        aliases: ["sonnet-5.5", "claude-sonnet-latest"],
         deprecated: false,
         isLocal: false,
         releaseDate: "2026-09-28",
@@ -1374,7 +1374,7 @@ const LEGACY_MODEL_REGISTRY = {
             translation: 8,
             summarization: 8,
         },
-        // "claude-sonnet-latest" now belongs to CLAUDE_SONNET_5 — see the note
+        // "claude-sonnet-latest" now belongs to CLAUDE_SONNET_5_5 — see the note
         // on CLAUDE_OPUS_4_5's aliases.
         aliases: ["claude-4.5-sonnet", "sonnet-4.5"],
         deprecated: false,
@@ -2846,6 +2846,58 @@ export function resolveSamplingParams(provider, model, params, site) {
     }
     return {};
 }
+/**
+ * Claude models that answer a forced `tool_choice` (`any` or `tool`) with a
+ * 400: "tool_choice: type "tool" and "any" are not supported for this model."
+ * Sonnet 5.5 and Opus 5.5 were probed live on Vertex; the Fable 5.1 docs state
+ * the same, and Mythos 5.1 shares Fable 5.1's capabilities. The same families
+ * also refuse `thinking: {type: "disabled"}`.
+ */
+const FORCED_TOOL_CHOICE_REJECTING_FAMILIES = [
+    /sonnet[-_.]?5[-_.]5(?![0-9])/i,
+    /opus[-_.]?5[-_.]5(?![0-9])/i,
+    /(?:fable|mythos)[-_.]?5[-_.]1(?![0-9])/i,
+];
+/**
+ * Of those, the ones that accept `thinking: {type: "between_tools"}` as their
+ * lowest setting. The rest think adaptively, always on, and take no
+ * `thinking` field in its place.
+ */
+const BETWEEN_TOOLS_THINKING_FAMILIES = [
+    /sonnet[-_.]?5[-_.]5(?![0-9])/i,
+];
+function matchesClaudeFamily(families, model) {
+    const normalizedModel = (model ?? "").toLowerCase();
+    if (!normalizedModel) {
+        return false;
+    }
+    const modelId = MODEL_ALIASES[normalizedModel] ?? normalizedModel;
+    return families.some((family) => family.test(modelId) || family.test(normalizedModel));
+}
+/**
+ * Whether a Claude model accepts a forced `tool_choice` (`any` or a named
+ * `tool`). Unknown models default to accepting it.
+ */
+export function modelSupportsForcedToolChoice(model) {
+    return !matchesClaudeFamily(FORCED_TOOL_CHOICE_REJECTING_FAMILIES, model);
+}
+/**
+ * What to send instead of `thinking: {type: "disabled"}` to a Claude model.
+ * `between_tools` is refused at `xhigh` and `max` effort, so those fall back
+ * to adaptive thinking (`omit`).
+ */
+export function claudeDisabledThinkingReplacement(model, effort) {
+    if (!matchesClaudeFamily(FORCED_TOOL_CHOICE_REJECTING_FAMILIES, model)) {
+        return "keep";
+    }
+    const level = effort?.toLowerCase();
+    if (matchesClaudeFamily(BETWEEN_TOOLS_THINKING_FAMILIES, model) &&
+        level !== "xhigh" &&
+        level !== "max") {
+        return "between_tools";
+    }
+    return "omit";
+}
 /**
  * Get models by provider
  */

package/dist/neurolink.js

@@ -12956,6 +12956,7 @@ Current user's request: ${currentInput}`;
                 "gen_ai.operation.name": "decide",
                 "ai.provider": providerName,
                 "decision.question_count": questionCount,
+                "decision.images.count": options.images?.length ?? 0,
             },
         }, async (otelSpan) => {
             const { traceId, parentSpanId } = getActiveTraceContext();
@@ -12964,6 +12965,7 @@ Current user's request: ${currentInput}`;
                 "ai.provider": providerName,
                 "ai.model": options.model ?? "",
                 "decision.question_count": questionCount,
+                "decision.images.count": options.images?.length ?? 0,
             }, parentSpanId, traceId);
             try {
                 // enableMCP: false — a decision model has no tools and never will.
@@ -12974,6 +12976,8 @@ Current user's request: ${currentInput}`;
                 const result = await provider.decide({
                     state: options.state,
                     questions: options.questions,
+                    images: options.images,
+                    video: options.video,
                     model: options.model,
                     signal: options.signal,
                     timeoutMs: options.timeoutMs,
@@ -12988,6 +12992,7 @@ Current user's request: ${currentInput}`;
                 otelSpan.setAttribute("gen_ai.usage.input_tokens", result.usage.inputTokens);
                 otelSpan.setAttribute("gen_ai.usage.output_tokens", result.usage.outputTokens);
                 otelSpan.setAttribute("decision.answer_count", answerCount);
+                otelSpan.setAttribute("decision.media.bytes", result.mediaBytes ?? 0);
                 otelSpan.setAttribute("ai.cost.total", cost);
                 if (result.upstreamMs !== undefined) {
                     otelSpan.setAttribute("decision.upstream_ms", result.upstreamMs);
@@ -13008,6 +13013,7 @@ Current user's request: ${currentInput}`;
                     "ai.cost.output": 0,
                     "ai.cost.total": cost,
                     "decision.answer_count": answerCount,
+                    "decision.media.bytes": result.mediaBytes ?? 0,
                     "decision.latency_ms": result.latencyMs,
                     ...(result.upstreamMs !== undefined
                         ? { "decision.upstream_ms": result.upstreamMs }

package/dist/providers/amazonSagemaker.js

@@ -126,7 +126,7 @@ export class AmazonSageMakerProvider extends BaseProvider {
             options.disableInternalFallback === true;
         // The native loop bypasses BaseProvider.executeGeneration, so the turn
         // budget has to be composed here or it stops existing for this provider.
-        return this.runGenerateWithModelFallback(() => this.withTurnTimeout({ ...options, tools: mergedTools }, this.getDescriptorGenerateMs(), (timedOptions) => this.executeNativeGenerate(timedOptions)), callerOwnsFallback);
+        return this.runGenerateWithModelFallback(() => this.withTurnTimeout({ ...options, tools: mergedTools }, this.getDescriptorGenerateMs(), (timedOptions) => this.executeNativeGenerate(timedOptions)), callerOwnsFallback, options);
     }
     async executeNativeGenerate(options) {
         const startTime = Date.now();