@juspay/neurolink 12.49.20 → 12.51.0

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

@@ -7,16 +7,18 @@
  */
 import type { ProviderConfigOptions, SetupArgs } from "../../types/index.js";
 /**
- * ProviderConfigOptions for the 21 canonical AIProviderName values the
+ * ProviderConfigOptions for the canonical AIProviderName values the
  * interactive wizard above doesn't have a bespoke handleXSetup() for. 8 are
  * the JSON-catalog providers (cerebras, cloudflare, fireworks, groq,
  * 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(). 10 more reuse the existing createXConfig()
- * factories in providerConfig.ts; the remaining 5 (ollama, litellm,
- * sagemaker, lm-studio, llamacpp) don't have a factory and are defined
- * inline using their real env var names. lm-studio and llamacpp previously
+ * buildCatalogConfigOptions(); decide-kind catalog entries (laya) 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
+ * factory and are defined inline using their real env var names. lm-studio
+ * and llamacpp previously
  * had createLmStudioConfig()/createLlamaCppConfig() factories, but those
  * were removed as dead code in a later cleanup — this mirrors that removal
  * rather than re-adding them.

package/dist/cli/commands/setup.js

@@ -21,8 +21,8 @@ import { handleHuggingFaceSetup } from "./setup-huggingface.js";
 import { handleMistralSetup } from "./setup-mistral.js";
 import { OpenRouterModels, } from "../../constants/enums.js";
 import { PROVIDER_DESCRIPTORS_BY_NAME } from "../../factories/providerDescriptors.js";
-import { createCohereConfig, createDeepSeekConfig, createIdeogramConfig, createJinaConfig, createNvidiaNimConfig, createOpenAICompatibleConfig, createRecraftConfig, createReplicateConfig, createStabilityConfig, createVoyageConfig, createTypeSafeConfig, createLayaConfig, createXorConfig, createPerplexityDeciderConfig, createCloudflareClefConfig, satisfiesFallbacks, } from "../../utils/providerConfig.js";
-import { getCatalogJsonEntries, buildCatalogConfigOptions, } from "../../providers/catalog/loader.js";
+import { createCohereConfig, createDeepSeekConfig, createIdeogramConfig, createJinaConfig, createNvidiaNimConfig, createOpenAICompatibleConfig, createRecraftConfig, createReplicateConfig, createStabilityConfig, createVoyageConfig, createTypeSafeConfig, createXorConfig, createPerplexityDeciderConfig, createCloudflareClefConfig, satisfiesFallbacks, } from "../../utils/providerConfig.js";
+import { getCatalogJsonEntries, getDecideCatalogEntries, buildCatalogConfigOptions, } from "../../providers/catalog/loader.js";
 // Provider information database
 const PROVIDERS = [
     {
@@ -144,16 +144,18 @@ const PROVIDERS = [
     },
 ];
 /**
- * ProviderConfigOptions for the 21 canonical AIProviderName values the
+ * ProviderConfigOptions for the canonical AIProviderName values the
  * interactive wizard above doesn't have a bespoke handleXSetup() for. 8 are
  * the JSON-catalog providers (cerebras, cloudflare, fireworks, groq,
  * 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(). 10 more reuse the existing createXConfig()
- * factories in providerConfig.ts; the remaining 5 (ollama, litellm,
- * sagemaker, lm-studio, llamacpp) don't have a factory and are defined
- * inline using their real env var names. lm-studio and llamacpp previously
+ * buildCatalogConfigOptions(); decide-kind catalog entries (laya) 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
+ * factory and are defined inline using their real env var names. lm-studio
+ * and llamacpp previously
  * had createLmStudioConfig()/createLlamaCppConfig() factories, but those
  * were removed as dead code in a later cleanup — this mirrors that removal
  * rather than re-adding them.
@@ -170,10 +172,10 @@ export const EXTRA_PROVIDER_CONFIGS = {
     ideogram: createIdeogramConfig(),
     recraft: createRecraftConfig(),
     typesafe: createTypeSafeConfig(),
-    laya: createLayaConfig(),
     xor: createXorConfig(),
     "perplexity-decider": createPerplexityDeciderConfig(),
     "cloudflare-clef": createCloudflareClefConfig(),
+    ...Object.fromEntries(getDecideCatalogEntries().map((e) => [e.id, buildCatalogConfigOptions(e)])),
     ...Object.fromEntries(getCatalogJsonEntries()
         .filter((e) => e.id !== "mistral")
         .map((e) => [e.id, buildCatalogConfigOptions(e)])),

package/dist/constants/enums.d.ts

@@ -56,6 +56,7 @@ export declare enum AIProviderName {
     INFERENCE_NET = "inference-net",
     IO_INTELLIGENCE = "io-intelligence",
     KOSCOMPUTE = "koscompute",
+    LAYA = "laya",
     LEMONFOX_AI = "lemonfox-ai",
     LILAC = "lilac",
     LLMTECH = "llmtech",
@@ -109,8 +110,6 @@ export declare enum AIProviderName {
     RECRAFT = "recraft",
     /** TypeSafe (Jev) — serves the `decide` inference type only. */
     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",
     /**
@@ -1327,6 +1326,12 @@ export declare enum IoIntelligenceModels {
 export declare enum KoscomputeModels {
     QWEN_QWEN3_8_27B = "qwen/qwen3.8-27b"
 }
+export declare enum LayaModels {
+    TYPED_DECISIONS = "typed-decisions",
+    ENGLISH = "english",
+    MULTILINGUAL = "multilingual",
+    AUTO = "auto"
+}
 export declare enum LemonfoxAiModels {
     DEEPSEEK_V4_FLASH = "deepseek-v4-flash",
     MIMO_V2_5_PRO = "mimo-v2.5-pro"
@@ -1771,19 +1776,6 @@ export declare enum TypeSafeModels {
     JEV_LATEST = "jev-latest",
     JEV_PREVIEW = "jev-preview"
 }
-/**
- * Laya checkpoints. Hand-written: Laya is a Tier-3 provider, so it is not in
- * the provider catalog and codegen never touches this.
- *
- * `auto` sends no model, which lets Laya's own router pick a checkpoint by
- * the detected language of the state.
- */
-export declare enum LayaModels {
-    TYPED_DECISIONS = "typed-decisions",
-    ENGLISH = "english",
-    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.

package/dist/constants/enums.js

@@ -61,6 +61,7 @@ export var AIProviderName;
     AIProviderName["INFERENCE_NET"] = "inference-net";
     AIProviderName["IO_INTELLIGENCE"] = "io-intelligence";
     AIProviderName["KOSCOMPUTE"] = "koscompute";
+    AIProviderName["LAYA"] = "laya";
     AIProviderName["LEMONFOX_AI"] = "lemonfox-ai";
     AIProviderName["LILAC"] = "lilac";
     AIProviderName["LLMTECH"] = "llmtech";
@@ -115,8 +116,6 @@ export var AIProviderName;
     AIProviderName["RECRAFT"] = "recraft";
     /** TypeSafe (Jev) — serves the `decide` inference type only. */
     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";
     /**
@@ -1603,6 +1602,13 @@ export var KoscomputeModels;
 (function (KoscomputeModels) {
     KoscomputeModels["QWEN_QWEN3_8_27B"] = "qwen/qwen3.8-27b";
 })(KoscomputeModels || (KoscomputeModels = {}));
+export var LayaModels;
+(function (LayaModels) {
+    LayaModels["TYPED_DECISIONS"] = "typed-decisions";
+    LayaModels["ENGLISH"] = "english";
+    LayaModels["MULTILINGUAL"] = "multilingual";
+    LayaModels["AUTO"] = "auto";
+})(LayaModels || (LayaModels = {}));
 export var LemonfoxAiModels;
 (function (LemonfoxAiModels) {
     LemonfoxAiModels["DEEPSEEK_V4_FLASH"] = "deepseek-v4-flash";
@@ -2093,20 +2099,6 @@ export var TypeSafeModels;
     TypeSafeModels["JEV_LATEST"] = "jev-latest";
     TypeSafeModels["JEV_PREVIEW"] = "jev-preview";
 })(TypeSafeModels || (TypeSafeModels = {}));
-/**
- * Laya checkpoints. Hand-written: Laya is a Tier-3 provider, so it is not in
- * the provider catalog and codegen never touches this.
- *
- * `auto` sends no model, which lets Laya's own router pick a checkpoint by
- * the detected language of the state.
- */
-export var LayaModels;
-(function (LayaModels) {
-    LayaModels["TYPED_DECISIONS"] = "typed-decisions";
-    LayaModels["ENGLISH"] = "english";
-    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.

package/dist/core/conversationMemoryManager.js

@@ -11,6 +11,7 @@ import { tracers, withSpan } from "../telemetry/index.js";
 import { ConversationMemoryError } from "../types/index.js";
 import { buildContextFromPointer, getEffectiveTokenThreshold, } from "../utils/conversationMemory.js";
 import { logger } from "../utils/logger.js";
+import { assistantTurnIdentity, isTurnAlreadyStored, userTurnIdentity, } from "../utils/conversationTurnIdentity.js";
 import { extractIsErrorText } from "../utils/toolResultStatus.js";
 export class ConversationMemoryManager {
     sessions = new Map();
@@ -82,11 +83,14 @@ export class ConversationMemoryManager {
                     session = this.createNewSession(options.sessionId, options.userId);
                     this.sessions.set(options.sessionId, session);
                 }
+                if (isTurnAlreadyStored(session.messages, options.turnId)) {
+                    return;
+                }
                 const tokenThreshold = options.providerDetails
                     ? getEffectiveTokenThreshold(options.providerDetails.provider, options.providerDetails.model, this.config.tokenThreshold, session.tokenThreshold)
                     : this.config.tokenThreshold || 50000;
-                const userMsg = await this.validateAndPrepareMessage(options.userMessage, "user", tokenThreshold);
-                const assistantMsg = await this.validateAndPrepareMessage(options.aiResponse, "assistant", tokenThreshold);
+                const userMsg = await this.validateAndPrepareMessage(options.userMessage, "user", tokenThreshold, userTurnIdentity(options));
+                const assistantMsg = await this.validateAndPrepareMessage(options.aiResponse, "assistant", tokenThreshold, assistantTurnIdentity(options));
                 if (options.events && options.events.length > 0) {
                     assistantMsg.events = options.events;
                 }
@@ -147,8 +151,8 @@ export class ConversationMemoryManager {
      * Validate and prepare a message before adding to session
      * Truncates if message exceeds token limit
      */
-    async validateAndPrepareMessage(content, role, threshold) {
-        const id = randomUUID();
+    async validateAndPrepareMessage(content, role, threshold, identity) {
+        const id = identity?.id ?? randomUUID();
         const tokenCount = TokenUtils.estimateTokenCount(content);
         const maxMessageSize = Math.floor(threshold * MEMORY_THRESHOLD_PERCENTAGE);
         if (tokenCount > maxMessageSize) {
@@ -166,6 +170,7 @@ export class ConversationMemoryManager {
                 content: truncated,
                 timestamp: new Date().toISOString(),
                 metadata: {
+                    ...identity?.metadata,
                     truncated: true,
                 },
             };
@@ -175,6 +180,7 @@ export class ConversationMemoryManager {
             role,
             content,
             timestamp: new Date().toISOString(),
+            ...(identity?.metadata ? { metadata: identity.metadata } : {}),
         };
     }
     /**

package/dist/core/redisConversationMemoryManager.js

@@ -15,6 +15,7 @@ import { withTimeout } from "../utils/errorHandling.js";
 import { buildContextFromPointer, getEffectiveTokenThreshold, } from "../utils/conversationMemory.js";
 import { runWithCurrentLangfuseContext } from "../services/server/ai/observability/instrumentation.js";
 import { logger } from "../utils/logger.js";
+import { assistantTurnIdentity, isTurnAlreadyStored, turnMessageFields, userTurnIdentity, } from "../utils/conversationTurnIdentity.js";
 import { extractIsErrorText } from "../utils/toolResultStatus.js";
 import { deserializeConversation, encodeStoredMessages, getSessionMessagesKey, MESSAGES_KEY_SUFFIX, isSessionMessagesKey, parseStoredMessages, serializeConversationMetadata, usesSplitMessageStorage, getNormalizedConfig, getPooledRedisClient, getSessionKey, getUserSessionsKey, releasePooledRedisClient, scanKeys, serializeConversation, } from "../utils/redis.js";
 const redisTracer = tracers.redis;
@@ -393,6 +394,9 @@ export class RedisConversationMemoryManager {
                 const redisKey = getSessionKey(this.redisConfig, options.sessionId, options.userId);
                 const conversationData = await this.redisClient.get(redisKey);
                 let conversation = await this.hydrateMessages(deserializeConversation(conversationData), redisKey);
+                if (isTurnAlreadyStored(conversation?.messages ?? [], options.turnId)) {
+                    return;
+                }
                 // Split-storage bookkeeping: whether this session already keeps its
                 // messages in the companion LIST, and how many it held before this
                 // turn — so only the new ones are appended rather than rewriting the
@@ -457,7 +461,7 @@ export class RedisConversationMemoryManager {
                     ? getEffectiveTokenThreshold(options.providerDetails.provider, options.providerDetails.model, this.config.tokenThreshold, conversation.tokenThreshold)
                     : this.config.tokenThreshold || 50000;
                 const userMsg = {
-                    id: randomUUID(),
+                    ...turnMessageFields(userTurnIdentity(options)),
                     timestamp: options.startTimeStamp?.toISOString() || this.generateTimestamp(),
                     role: "user",
                     content: options.userMessage,
@@ -471,14 +475,11 @@ export class RedisConversationMemoryManager {
                     conversation.messages.push(...options.skillMessages);
                 }
                 const assistantMsg = {
-                    id: randomUUID(),
+                    ...turnMessageFields(assistantTurnIdentity(options, options.thoughtSignature)),
                     timestamp: this.generateTimestamp(),
                     role: "assistant",
                     content: options.aiResponse,
                     events: options.events || undefined,
-                    ...(options.thoughtSignature && {
-                        metadata: { thoughtSignature: options.thoughtSignature },
-                    }),
                 };
                 conversation.messages.push(assistantMsg);
                 // Store API-reported token counts if available

package/dist/factories/providerDescriptors.d.ts

@@ -1,5 +1,11 @@
 import { AIProviderName } from "../constants/enums.js";
-import type { NeurolinkCredentials, ProviderDescriptor } from "../types/index.js";
+import type { DecideCatalogJson, NeurolinkCredentials, ProviderDescriptor } from "../types/index.js";
+/**
+ * Builds the ProviderDescriptor of a JSON `decide` provider. It serves only
+ * `decide`, so it declares no auto-select, preference or health-sweep rank:
+ * those feed generation fallback chains a text-less provider must never join.
+ */
+export declare function buildDecideDescriptor(entry: DecideCatalogJson): ProviderDescriptor;
 /**
  * Single source of truth for provider identity, credentials, defaults, and
  * runtime behavior classification — the hand-maintained providers plus the
@@ -14,10 +20,18 @@ export declare const PROVIDER_DESCRIPTORS_BY_NAME: ReadonlyMap<AIProviderName, P
  * ProviderFactory.normalizeProviderName().
  */
 export declare const PROVIDER_ALIAS_INDEX: ReadonlyMap<string, AIProviderName>;
+/**
+ * Throws unless every descriptor that serves `decide` has its own finite
+ * `decideOrder`. The order is the precedence of resolveDefaultDecisionProvider(),
+ * so a missing or repeated number would make that choice depend on array position.
+ */
+export declare function assertUniqueDecideOrder(descriptors: readonly ProviderDescriptor[]): void;
+/** The descriptors that serve `decide`, lowest `decideOrder` first. */
+export declare function orderDecisionProviders(descriptors: readonly ProviderDescriptor[]): readonly ProviderDescriptor[];
 /**
  * Providers that serve the `decide` inference type, derived from each
- * descriptor's `inferenceKinds`. A provider that declares nothing is a text
- * provider, so it is not here.
+ * descriptor's `inferenceKinds` and sorted by `decideOrder`, lowest first. A
+ * provider that declares nothing is a text provider, so it is not here.
  */
 export declare const DECISION_PROVIDERS: readonly ProviderDescriptor[];
 /**

package/dist/factories/providerDescriptors.js

@@ -1,7 +1,7 @@
 import { AIProviderName } from "../constants/enums.js";
-import { GoogleAIModels, OpenAIModels, AnthropicModels, VertexModels, OllamaModels, LiteLLMModels, NvidiaNimModels, OpenRouterModels, CohereModels, VoyageModels, TypeSafeModels, LayaModels, XorModels, PerplexityDeciderModels, CloudflareClefModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, ReplicateModels, } from "../constants/enums.js";
+import { GoogleAIModels, OpenAIModels, AnthropicModels, VertexModels, OllamaModels, LiteLLMModels, NvidiaNimModels, OpenRouterModels, CohereModels, VoyageModels, TypeSafeModels, XorModels, PerplexityDeciderModels, CloudflareClefModels, 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 { getCatalogJsonEntries, getDecideCatalogEntries, catalogCredentialsKey, catalogEnvVar, } from "../providers/catalog/loader.js";
 import { DEFAULT_INFERENCE_KINDS } from "../types/index.js";
 /**
  * Hand-maintained provider identity, credentials, defaults, and runtime
@@ -440,6 +440,7 @@ const HAND_DESCRIPTORS = [
         // what keeps it out of every generation code path, rather than each of
         // those having to special-case the provider name.
         inferenceKinds: ["decide"],
+        decideOrder: 1,
         // It emits no text, so it cannot call tools and cannot answer a
         // "live-generate" health probe.
         toolSupport: "none",
@@ -466,58 +467,13 @@ const HAND_DESCRIPTORS = [
         },
         setupUrl: "https://console.typesafe.ai/keys",
     },
-    // Laya MUST stay after TypeSafe, XOR after Laya, Perplexity after XOR, and
-    // Cloudflare Clef after Perplexity.
+    // Decision precedence is decideOrder, not position: typesafe 1, laya 2 (from
+    // catalog/laya.json), xor 3, perplexity-decider 4, cloudflare-clef 5.
     // resolveDefaultDecisionProvider() returns the first configured
-    // DECISION_PROVIDERS entry, in this order, so a host configured for several
+    // 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. Reordering these silently changes which model routes,
-    // compacts and plans for every such host.
-    {
-        name: AIProviderName.LAYA,
-        aliases: [],
-        credentialsKey: "laya",
-        envVars: {
-            apiKey: "LAYA_API_KEY",
-            baseURL: "LAYA_BASE_URL",
-            // Laya has no built-in endpoint (it is self-hosted 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.laya.
-            extraRequired: ["LAYA_BASE_URL"],
-            model: "LAYA_MODEL",
-        },
-        defaultModel: LayaModels.TYPED_DECISIONS,
-        // Serves only `decide`, like TypeSafe — 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 },
-        // Laya's encoders cut the state off silently past their window (measured
-        // live: a 60,000-character state answered 200 on 1,024 tokens). 768 is
-        // the 1,024-token window minus the 256-token budget Laya reserves for a
-        // question and its options; 320 is the English checkpoint's 512 minus
-        // 192. `auto` may route to English, and an unlisted name (Laya's server
-        // accepts aliases such as `en`) could be any checkpoint, so both get 320.
-        // 64 is the question cap in Laya's own server. Non-ASCII rates were
-        // measured live: Chinese is ~1.4 tokens per character on the English
-        // tokenizer (typed-decisions, english) and ~0.56 on multilingual.
-        decisionLimits: {
-            maxStateTokens: 320,
-            maxQuestions: 64,
-            nonAsciiTokensPerChar: 1.5,
-            models: {
-                "typed-decisions": { maxStateTokens: 768 },
-                multilingual: { maxStateTokens: 768, nonAsciiTokensPerChar: 0.6 },
-                english: { maxStateTokens: 320 },
-                auto: { maxStateTokens: 320 },
-            },
-        },
-        setupUrl: "https://github.com/NandhaKishorM/laya",
-    },
+    // naming them. Renumbering silently changes which model routes, compacts
+    // and plans for every such host.
     {
         name: AIProviderName.XOR,
         aliases: [],
@@ -535,6 +491,7 @@ const HAND_DESCRIPTORS = [
         // Serves only `decide`, like TypeSafe and Laya — the one declaration that
         // keeps it out of every generation code path.
         inferenceKinds: ["decide"],
+        decideOrder: 3,
         toolSupport: "none",
         localRuntime: false,
         healthCheck: "env-only",
@@ -579,6 +536,7 @@ const HAND_DESCRIPTORS = [
         // declaration that keeps it out of every generation code path, and out of
         // the way of the `perplexity` text provider.
         inferenceKinds: ["decide"],
+        decideOrder: 4,
         toolSupport: "none",
         localRuntime: false,
         healthCheck: "env-only",
@@ -646,6 +604,7 @@ const HAND_DESCRIPTORS = [
         // declaration that keeps it out of every generation code path, and out of
         // the way of the `cloudflare` text provider.
         inferenceKinds: ["decide"],
+        decideOrder: 5,
         toolSupport: "none",
         localRuntime: false,
         healthCheck: "env-only",
@@ -778,8 +737,48 @@ function buildCatalogDescriptor(entry) {
             : {}),
     };
 }
+/**
+ * Builds the ProviderDescriptor of a JSON `decide` provider. It serves only
+ * `decide`, so it declares no auto-select, preference or health-sweep rank:
+ * those feed generation fallback chains a text-less provider must never join.
+ */
+export function buildDecideDescriptor(entry) {
+    const fallbacks = entry.wire.apiKeyFallbackEnvVars?.length
+        ? { fallbacks: [...entry.wire.apiKeyFallbackEnvVars] }
+        : {};
+    return {
+        name: entry.id,
+        aliases: entry.aliases,
+        credentialsKey: catalogCredentialsKey(entry),
+        envVars: {
+            apiKey: catalogEnvVar(entry, "apiKey"),
+            ...fallbacks,
+            baseURL: catalogEnvVar(entry, "baseURL"),
+            // A key is useless without an endpoint when the provider has no built-in one.
+            ...(entry.decide.baseURLRequired
+                ? { extraRequired: [catalogEnvVar(entry, "baseURL")] }
+                : {}),
+            model: catalogEnvVar(entry, "model"),
+        },
+        defaultModel: entry.models.default,
+        inferenceKinds: ["decide"],
+        decideOrder: entry.decide.order,
+        toolSupport: "none",
+        localRuntime: false,
+        healthCheck: "env-only",
+        setupUrl: entry.setup.url,
+        ...(entry.setup.apiKeyFormat
+            ? { apiKeyFormatPattern: new RegExp(entry.setup.apiKeyFormat) }
+            : {}),
+        ...(entry.timeouts ? { timeouts: { ...entry.timeouts } } : {}),
+        ...(entry.decisionLimits ? { decisionLimits: entry.decisionLimits } : {}),
+    };
+}
 function buildCatalogDescriptors() {
-    return getCatalogJsonEntries().map(buildCatalogDescriptor);
+    return [
+        ...getCatalogJsonEntries().map(buildCatalogDescriptor),
+        ...getDecideCatalogEntries().map(buildDecideDescriptor),
+    ];
 }
 /**
  * Single source of truth for provider identity, credentials, defaults, and
@@ -801,12 +800,38 @@ export const PROVIDER_ALIAS_INDEX = new Map(PROVIDER_DESCRIPTORS.flatMap((d) =>
     [d.name.toLowerCase(), d.name],
     ...d.aliases.map((alias) => [alias.toLowerCase(), d.name]),
 ]));
+const servesDecide = (descriptor) => (descriptor.inferenceKinds ?? DEFAULT_INFERENCE_KINDS).includes("decide");
+/**
+ * Throws unless every descriptor that serves `decide` has its own finite
+ * `decideOrder`. The order is the precedence of resolveDefaultDecisionProvider(),
+ * so a missing or repeated number would make that choice depend on array position.
+ */
+export function assertUniqueDecideOrder(descriptors) {
+    const owners = new Map();
+    for (const descriptor of descriptors.filter(servesDecide)) {
+        const order = descriptor.decideOrder;
+        if (order === undefined || !Number.isFinite(order)) {
+            throw new Error(`decision provider "${descriptor.name}" has no usable decideOrder`);
+        }
+        const clash = owners.get(order);
+        if (clash !== undefined) {
+            throw new Error(`decideOrder ${order} is used by both "${clash}" and "${descriptor.name}"`);
+        }
+        owners.set(order, descriptor.name);
+    }
+}
+/** The descriptors that serve `decide`, lowest `decideOrder` first. */
+export function orderDecisionProviders(descriptors) {
+    const rank = (d) => d.decideOrder ?? Number.MAX_SAFE_INTEGER;
+    return descriptors.filter(servesDecide).sort((a, b) => rank(a) - rank(b));
+}
+assertUniqueDecideOrder(PROVIDER_DESCRIPTORS);
 /**
  * Providers that serve the `decide` inference type, derived from each
- * descriptor's `inferenceKinds`. A provider that declares nothing is a text
- * provider, so it is not here.
+ * descriptor's `inferenceKinds` and sorted by `decideOrder`, lowest first. A
+ * provider that declares nothing is a text provider, so it is not here.
  */
-export const DECISION_PROVIDERS = PROVIDER_DESCRIPTORS.filter((d) => (d.inferenceKinds ?? DEFAULT_INFERENCE_KINDS).includes("decide"));
+export const DECISION_PROVIDERS = orderDecisionProviders(PROVIDER_DESCRIPTORS);
 /**
  * Whether a decision provider is fully configured: its key and every
  * extraRequired variable are set, each either in the environment or in

package/dist/factories/providerRegistry.d.ts

@@ -1,4 +1,12 @@
-import type { ProviderRegistryOptions } from "../types/index.js";
+import type { DecideCatalogJson, ProviderRegistryOptions } from "../types/index.js";
+/**
+ * Registers the config-driven `decide` providers described by catalog JSON
+ * (inferenceKinds ["decide"]). Each one is built by ConfiguredDecisionProvider,
+ * loaded lazily like every other provider; its descriptor declares
+ * inferenceKinds ["decide"], so nothing in the generation fallback chain can
+ * reach it.
+ */
+export declare function registerDecideCatalogProviders(entries: readonly DecideCatalogJson[]): void;
 /**
  * Provider Registry - registers all providers with the factory
  * This is where we migrate providers one by one to the new pattern

package/dist/factories/providerRegistry.js

@@ -1,9 +1,25 @@
 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, XorModels, PerplexityDeciderModels, CloudflareClefModels, ReplicateModels, } from "../constants/enums.js";
+import { AIProviderName, GoogleAIModels, OpenAIModels, AnthropicModels, VertexModels, OllamaModels, LiteLLMModels, NvidiaNimModels, OpenRouterModels, CohereModels, VoyageModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, TypeSafeModels, XorModels, PerplexityDeciderModels, CloudflareClefModels, 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";
+import { catalogEnvVar, getDecideCatalogEntries, } from "../providers/catalog/loader.js";
+/**
+ * Registers the config-driven `decide` providers described by catalog JSON
+ * (inferenceKinds ["decide"]). Each one is built by ConfiguredDecisionProvider,
+ * loaded lazily like every other provider; its descriptor declares
+ * inferenceKinds ["decide"], so nothing in the generation fallback chain can
+ * reach it.
+ */
+export function registerDecideCatalogProviders(entries) {
+    for (const entry of entries) {
+        ProviderFactory.registerProvider(entry.id, async (modelName, _providerName, sdk, _region, credentials) => {
+            const { ConfiguredDecisionProvider } = await import("../providers/configuredDecision.js");
+            return new ConfiguredDecisionProvider(entry, modelName, sdk, credentials);
+        }, process.env[catalogEnvVar(entry, "model")] || entry.models.default, [...entry.aliases], PROVIDER_DESCRIPTORS_BY_NAME.get(entry.id));
+    }
+}
 /**
  * Provider Registry - registers all providers with the factory
  * This is where we migrate providers one by one to the new pattern
@@ -234,14 +250,6 @@ export class ProviderRegistry {
                 const { TypeSafeProvider } = await import("../providers/typesafe.js");
                 return new TypeSafeProvider(modelName, sdk, undefined, typesafeCreds);
             }, process.env.TYPESAFE_MODEL || TypeSafeModels.JEV_LATEST, ["jev", "typesafe-ai"], PROVIDER_DESCRIPTORS_BY_NAME.get(AIProviderName.TYPESAFE));
-            // Register Laya — the second `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.LAYA, async (modelName, _providerName, sdk, _region, credentials) => {
-                const layaCreds = credentials;
-                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.
@@ -269,6 +277,8 @@ export class ProviderRegistry {
                 const { CloudflareClefProvider } = await import("../providers/cloudflareClef.js");
                 return new CloudflareClefProvider(modelName, sdk, undefined, cloudflareClefCreds);
             }, process.env.CLOUDFLARE_CLEF_MODEL || CloudflareClefModels.CLEF, [], PROVIDER_DESCRIPTORS_BY_NAME.get(AIProviderName.CLOUDFLARE_CLEF));
+            // The decide providers described by catalog JSON (laya.json and later).
+            registerDecideCatalogProviders(getDecideCatalogEntries());
             logger.debug("All AI providers registered successfully");
             // ===== MEDIA HANDLER REGISTRATION =====
             // Single registration path (Task 11): each ecosystem barrel (voice,

package/dist/models/manifestRegistry.js

@@ -1,5 +1,5 @@
 import { PROVIDER_MAX_TOKENS } from "../core/constants.js";
-import { getCatalogJsonEntries } from "../providers/catalog/loader.js";
+import { getCatalogJsonEntries, getDecideCatalogEntries, } from "../providers/catalog/loader.js";
 import { anthropicManifest } from "./manifests/anthropic.js";
 import { openaiManifest } from "./manifests/openai.js";
 import { azureManifest } from "./manifests/azure.js";
@@ -20,7 +20,6 @@ import { llamacppManifest } from "./manifests/llamacpp.js";
 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 { perplexityDeciderManifest } from "./manifests/perplexityDecider.js";
 import { cloudflareClefManifest } from "./manifests/cloudflareClef.js";
@@ -94,6 +93,44 @@ const catalogManifest = (id) => {
     }
     return buildCatalogManifest(entry);
 };
+/**
+ * Builds a manifest for a decide catalog provider from its catalog entry.
+ *
+ * A decide model emits no text, but the manifest type requires a
+ * `maxOutputTokens`, so the figure a decide entry declares is carried over
+ * as is. A model that omits `contextWindow` or `maxOutputTokens` takes the
+ * default model's, and `_default` is the default model's own row.
+ * `functionCalling` is always false: a decide model has no tool calling.
+ * No row gets a `displayName`: the decide catalog has no field for one and
+ * no reader shows a decide model's manifest name.
+ */
+function buildDecideManifest(entry) {
+    const base = entry.models.catalog[entry.models.default];
+    if (base === undefined ||
+        base.contextWindow === undefined ||
+        base.maxOutputTokens === undefined) {
+        throw new Error(`Decide catalog entry "${entry.id}": models.default needs contextWindow and maxOutputTokens for its manifest`);
+    }
+    const baseWindow = base.contextWindow;
+    const baseMaxOutput = base.maxOutputTokens;
+    const toEntry = (spec) => ({
+        aliases: [],
+        contextWindow: spec.contextWindow ?? baseWindow,
+        maxOutputTokens: spec.maxOutputTokens ?? baseMaxOutput,
+        vision: spec.vision,
+        functionCalling: false,
+    });
+    return {
+        defaultContextWindow: baseWindow,
+        models: {
+            _default: toEntry(base),
+            ...Object.fromEntries(Object.entries(entry.models.catalog).map(([modelId, spec]) => [
+                modelId,
+                toEntry(spec),
+            ])),
+        },
+    };
+}
 /**
  * Every provider's model manifest, keyed by the exact AIProviderName enum
  * value (kebab-case) — e.g. "google-ai", "nvidia-nim". Manifests are pure
@@ -122,7 +159,7 @@ export const MANIFEST_REGISTRY = {
     xai: xaiManifest,
     groq: groqManifest,
     typesafe: typesafeManifest,
-    laya: layaManifest,
+    ...Object.fromEntries(getDecideCatalogEntries().map((e) => [e.id, buildDecideManifest(e)])),
     xor: xorManifest,
     "perplexity-decider": perplexityDeciderManifest,
     "cloudflare-clef": cloudflareClefManifest,