@juspay/neurolink 12.55.3 → 12.56.1

package/dist/types/providerCatalog.d.ts

@@ -262,12 +262,23 @@ export type ProviderCatalogJson = {
     evidence: CatalogEvidence;
 };
 export type DecideFormatterName = "fastapi-validation";
-export type DecideBaseURLTransform = "strip-trailing-slash";
+export type DecideBaseURLTransform = "strip-trailing-slash" | "strip-v1-suffix";
+/** A check a base URL must pass once normalised; `origin` allows only an http(s) URL with no userinfo, query string or fragment. */
+export type DecideBaseURLShape = "origin";
 export type DecideErrorRule = {
     /** One status or a list of them. A rule sets this or `statusMin`, never both. */
     status?: number | readonly number[];
     /** Matches every status at or above this value. */
     statusMin?: number;
+    /**
+     * With a status, the rule also needs the error message to match this regular
+     * expression (case-insensitive, searched in the first 2,000 characters of the
+     * message after the key is removed). Bounded: no unescaped `*`, `+` or
+     * `{n,}`, no quantifier on a group, no count above 255, and the counts side
+     * by side multiply to at most 4,096 ways (each `?` is 2, each `{n,m}` is
+     * m-n+1).
+     */
+    messagePattern?: string;
     kind: DecisionErrorKind;
     retryable?: boolean;
 };
@@ -289,16 +300,31 @@ export type DecideModel = {
     /** With `body-unless-alias`: the model name that means "send no model". */
     alias?: string;
 };
+/**
+ * The request body keys prepared media is written under. Set together with
+ * `decisionLimits.media`, which says how much of it a request may carry.
+ */
+export type DecideMedia = {
+    /** Written only when the request carries at least one image. */
+    imagesKey: string;
+    /** Written only when the request carries a video; required exactly when `decisionLimits.media.video` is true. */
+    videoKey?: string;
+};
 export type DecideBlock = {
     dialect: DecideDialectName;
     /** Decision order, lower runs first; unique across decide providers. */
     order: number;
     /** True when the provider has no built-in endpoint and the caller must supply one. */
     baseURLRequired?: boolean;
+    /** Applied to the base URL in list order. */
     baseURLNormalize?: readonly DecideBaseURLTransform[];
+    /** Refuses a base URL of another shape before any request. Omitted means no check. */
+    baseURLShape?: DecideBaseURLShape;
     /** Request URL; `{baseURL}` is the only placeholder. */
     endpoint: string;
     model: DecideModel;
+    /** Body keys for images and video, after `model`, `state` and `questions`. Omitted means the body never carries media. */
+    media?: DecideMedia;
     requestIdHeaders?: readonly string[];
     response?: {
         modelPaths?: readonly string[];

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

@@ -1,2 +1,2 @@
-export type CatalogProviderName = "a2agent" | "abliteration" | "above-dev" | "ai21" | "aiand" | "aihubmix" | "aionlabs" | "ambient" | "api-route" | "arcee" | "atlas-cloud" | "avian-io" | "baidu-qianfan" | "baseten" | "bee-heossi" | "byteplus-modelark" | "bytez" | "cerebras" | "charm-hyper" | "chutes" | "cloudflare" | "crusoe" | "dashscope" | "deepinfra" | "deepseek" | "empiriolabs" | "featherless-ai" | "fireworks" | "friendli" | "gmicloud" | "gradientai" | "groq" | "hetzner-inference" | "huggingface" | "inception-labs" | "inceptron" | "inco" | "inference-net" | "io-intelligence" | "koscompute" | "laya" | "lemonfox-ai" | "lilac" | "llmtech" | "longcat" | "mancer" | "meta-model-api" | "minimax" | "mistral" | "moark" | "modelscope" | "moonshot-ai" | "morph" | "naver-clova-studio" | "nebius" | "neuralwatt" | "novita" | "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 type CatalogCredentialKey = "a2agent" | "abliteration" | "aboveDev" | "ai21" | "aiand" | "aihubmix" | "aionlabs" | "ambient" | "apiRoute" | "arcee" | "atlasCloud" | "avianIo" | "baiduQianfan" | "baseten" | "beeHeossi" | "byteplusModelark" | "bytez" | "cerebras" | "charmHyper" | "chutes" | "cloudflare" | "crusoe" | "dashscope" | "deepinfra" | "deepseek" | "empiriolabs" | "featherlessAi" | "fireworks" | "friendli" | "gmicloud" | "gradientai" | "groq" | "hetznerInference" | "huggingFace" | "inceptionLabs" | "inceptron" | "inco" | "inferenceNet" | "ioIntelligence" | "koscompute" | "laya" | "lemonfoxAi" | "lilac" | "llmtech" | "longcat" | "mancer" | "metaModelApi" | "minimax" | "mistral" | "moark" | "modelscope" | "moonshotAi" | "morph" | "naverClovaStudio" | "nebius" | "neuralwatt" | "novita" | "ovhcloud" | "parasail" | "paretoInference" | "perplexity" | "poolside" | "primeIntellect" | "reka" | "runinfra" | "sakanaAi" | "sambanova" | "sarvam" | "scaleway" | "siliconflow" | "stackit" | "stepfun" | "subconscious" | "synthetic" | "telnyx" | "thinkingMachines" | "tinfoil" | "together" | "umansAi" | "upstage" | "veniceAi" | "vispark" | "vultrInference" | "wafer" | "wandbInference" | "xai" | "xiaomiMimo" | "zAi";
+export type CatalogProviderName = "a2agent" | "abliteration" | "above-dev" | "ai21" | "aiand" | "aihubmix" | "aionlabs" | "ambient" | "api-route" | "arcee" | "atlas-cloud" | "avian-io" | "baidu-qianfan" | "baseten" | "bee-heossi" | "byteplus-modelark" | "bytez" | "cerebras" | "charm-hyper" | "chutes" | "cloudflare" | "crusoe" | "dashscope" | "deepinfra" | "deepseek" | "empiriolabs" | "featherless-ai" | "fireworks" | "friendli" | "gmicloud" | "gradientai" | "groq" | "hetzner-inference" | "huggingface" | "inception-labs" | "inceptron" | "inco" | "inference-net" | "io-intelligence" | "koscompute" | "laya" | "lemonfox-ai" | "lilac" | "llmtech" | "longcat" | "mancer" | "meta-model-api" | "minimax" | "mistral" | "moark" | "modelscope" | "moonshot-ai" | "morph" | "naver-clova-studio" | "nebius" | "neuralwatt" | "novita" | "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" | "xor" | "z-ai";
+export type CatalogCredentialKey = "a2agent" | "abliteration" | "aboveDev" | "ai21" | "aiand" | "aihubmix" | "aionlabs" | "ambient" | "apiRoute" | "arcee" | "atlasCloud" | "avianIo" | "baiduQianfan" | "baseten" | "beeHeossi" | "byteplusModelark" | "bytez" | "cerebras" | "charmHyper" | "chutes" | "cloudflare" | "crusoe" | "dashscope" | "deepinfra" | "deepseek" | "empiriolabs" | "featherlessAi" | "fireworks" | "friendli" | "gmicloud" | "gradientai" | "groq" | "hetznerInference" | "huggingFace" | "inceptionLabs" | "inceptron" | "inco" | "inferenceNet" | "ioIntelligence" | "koscompute" | "laya" | "lemonfoxAi" | "lilac" | "llmtech" | "longcat" | "mancer" | "metaModelApi" | "minimax" | "mistral" | "moark" | "modelscope" | "moonshotAi" | "morph" | "naverClovaStudio" | "nebius" | "neuralwatt" | "novita" | "ovhcloud" | "parasail" | "paretoInference" | "perplexity" | "poolside" | "primeIntellect" | "reka" | "runinfra" | "sakanaAi" | "sambanova" | "sarvam" | "scaleway" | "siliconflow" | "stackit" | "stepfun" | "subconscious" | "synthetic" | "telnyx" | "thinkingMachines" | "tinfoil" | "together" | "umansAi" | "upstage" | "veniceAi" | "vispark" | "vultrInference" | "wafer" | "wandbInference" | "xai" | "xiaomiMimo" | "xor" | "zAi";

package/dist/types/providers.d.ts

@@ -541,6 +541,10 @@ export type NeurolinkCredentials = {
         apiKey?: string;
         baseURL?: string;
     };
+    xor?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
     zAi?: {
         apiKey?: string;
         baseURL?: string;
@@ -601,16 +605,6 @@ export type NeurolinkCredentials = {
          */
         gatewayURL?: string;
     };
-    /**
-     * XOR (Juspay, open weights) — the `decide` inference type. There is no
-     * built-in endpoint: `baseURL` (or XOR_BASE_URL) is required, pointing at the
-     * origin of an XOR deployment or a proxy route to one; requests go to
-     * `<baseURL>/v1/systemone`.
-     */
-    xor?: {
-        apiKey?: string;
-        baseURL?: string;
-    };
     /**
      * Perplexity Decisions API — the `decide` inference type. The key is the same
      * `PERPLEXITY_API_KEY` the `perplexity` text provider reads, but this slice

package/dist/utils/modelChoices.js

@@ -2,7 +2,7 @@
  * Centralized model choices for CLI commands
  * Derives choices from model enums to ensure consistency
  */
-import { AIProviderName, OpenAIModels, AnthropicModels, GoogleAIModels, BedrockModels, VertexModels, OllamaModels, AzureOpenAIModels, LiteLLMModels, SageMakerModels, OpenRouterModels, NvidiaNimModels, CohereModels, VoyageModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, TypeSafeModels, XorModels, PerplexityDeciderModels, CloudflareClefModels, ReplicateModels, } from "../constants/enums.js";
+import { AIProviderName, OpenAIModels, AnthropicModels, GoogleAIModels, BedrockModels, VertexModels, OllamaModels, AzureOpenAIModels, LiteLLMModels, SageMakerModels, OpenRouterModels, NvidiaNimModels, CohereModels, VoyageModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, TypeSafeModels, PerplexityDeciderModels, CloudflareClefModels, ReplicateModels, } from "../constants/enums.js";
 import { getCatalogJsonEntries, getDecideCatalogEntries, } from "../providers/catalog/loader.js";
 /**
  * Looks up the JSON catalog entry for a provider, if it is one of the 11
@@ -377,12 +377,6 @@ const TOP_MODELS_CONFIG = {
             description: "Preview build of Jev",
         },
     ],
-    [AIProviderName.XOR]: [
-        {
-            model: XorModels.XOR_1_1,
-            description: "Recommended - XOR 1.1, decision model that also reads images and video",
-        },
-    ],
     [AIProviderName.PERPLEXITY_DECIDER]: [
         {
             model: PerplexityDeciderModels.PPLX_DECIDER_V1_27B,
@@ -445,7 +439,6 @@ export const DEFAULT_MODELS = {
     [AIProviderName.IDEOGRAM]: IdeogramModels.IDEOGRAM_V3,
     [AIProviderName.RECRAFT]: RecraftModels.RECRAFT_V3,
     [AIProviderName.TYPESAFE]: TypeSafeModels.JEV_LATEST,
-    [AIProviderName.XOR]: XorModels.XOR_1_1,
     [AIProviderName.PERPLEXITY_DECIDER]: PerplexityDeciderModels.PPLX_DECIDER_V1_27B,
     [AIProviderName.CLOUDFLARE_CLEF]: CloudflareClefModels.CLEF,
 };
@@ -480,7 +473,6 @@ const MODEL_ENUMS = {
     [AIProviderName.IDEOGRAM]: IdeogramModels,
     [AIProviderName.RECRAFT]: RecraftModels,
     [AIProviderName.TYPESAFE]: TypeSafeModels,
-    [AIProviderName.XOR]: XorModels,
     [AIProviderName.PERPLEXITY_DECIDER]: PerplexityDeciderModels,
     [AIProviderName.CLOUDFLARE_CLEF]: CloudflareClefModels,
     [AIProviderName.AUTO]: null,

package/dist/utils/ollamaDiagnostics.d.ts

@@ -0,0 +1,69 @@
+import type { OllamaDiagnosticContext, OllamaDiagnosticGate, OllamaDiagnosticOutcome, OllamaDiagnosticPermit, OllamaDiagnosticSnapshot } from "../types/index.js";
+export declare const OLLAMA_DIAGNOSTIC_RESET_MS = 300000;
+/**
+ * Validate the shared failure threshold once and reuse that first-use value.
+ * Missing or invalid PROVIDER_FAILURE_THRESHOLD values use the existing default
+ * of three; ProviderHealthChecker retains its original initialization point.
+ */
+export declare function getValidatedProviderFailureThreshold(): number;
+/** Return diagnostic error text with URL secrets redacted for logs and status. */
+export declare function getOllamaDiagnosticErrorMessage(error: unknown): string;
+/**
+ * Normalize an endpoint identity without its fragment, retaining path and query.
+ * The key can contain credentials and must remain private like the source URL.
+ */
+export declare function getOllamaDiagnosticEndpointKey(url: string): string;
+/**
+ * Resolve native /api/tags at the configured Ollama origin, independently of a
+ * compatible /v1 base path. OLLAMA_BASE_URL precedes the legacy OLLAMA_API_BASE
+ * alias, then localhost; malformed URLs and non-HTTP(S) schemes are rejected.
+ */
+export declare function resolveOllamaNativeTagsUrl(): string;
+/**
+ * Capture endpoint identity and model before asynchronous diagnostic work.
+ * Later environment changes cannot retarget this context or its health-cache
+ * identity; failure accounting remains shared by endpoint rather than model.
+ */
+export declare function captureOllamaDiagnosticContext(model?: string): OllamaDiagnosticContext;
+/**
+ * Copy the current policy state after applying initialization or timed expiry.
+ * This observation issues no request and records neither failure nor recovery.
+ */
+export declare function getOllamaDiagnosticSnapshot(context: OllamaDiagnosticContext): OllamaDiagnosticSnapshot;
+/**
+ * Admit a native probe or report backoff at the current failure threshold.
+ * Timed expiry is applied before this decision. Admission captures the clear
+ * generation without reserving a request or recording an outcome; callers must
+ * record the actual outcome on completion.
+ */
+export declare function beginOllamaDiagnostic(context: OllamaDiagnosticContext): OllamaDiagnosticGate;
+/**
+ * Apply timed expiry, then an admitted outcome once to the latest endpoint state.
+ * Transport/protocol failures increment its current count; success clears a
+ * nonzero count; not-run records no outcome. Timed expiry preserves permits,
+ * including those admitted before the boundary and completed after it.
+ * Duplicate permits and permits invalidated by explicit clearing are ignored.
+ *
+ * @returns A copy of the resulting state, or undefined for an ignored permit.
+ */
+export declare function completeOllamaDiagnostic(attempt: OllamaDiagnosticPermit, outcome: OllamaDiagnosticOutcome): OllamaDiagnosticSnapshot | undefined;
+/**
+ * Clear every endpoint and advance generation, invalidating in-flight permits.
+ * Unlike timed expiry, this prevents an older completion from repopulating state.
+ */
+export declare function clearOllamaDiagnostics(): void;
+/**
+ * Separate cached health by captured endpoint and model without splitting the
+ * endpoint-wide diagnostic counter. This private key can contain URL secrets.
+ */
+export declare function getOllamaHealthCacheKey(context: OllamaDiagnosticContext): string;
+export declare class OllamaProtocolError extends Error {
+    /** Identify an answered tags response whose models envelope is malformed. */
+    constructor();
+}
+/**
+ * Require an object containing a models array, including an empty array.
+ * Individual model entries are left to existing normalization; a malformed
+ * envelope throws OllamaProtocolError rather than appearing as no models.
+ */
+export declare function readOllamaTagsModels(payload: unknown): unknown[];

package/dist/utils/ollamaDiagnostics.js

@@ -0,0 +1,192 @@
+/** Internal native-diagnostic policy; inference and provider selection are separate. */
+import { logger } from "./logger.js";
+import { DEFAULT_OLLAMA_MODEL } from "../providers/ollama/constants.js";
+import { redactUrlsInText } from "./logSanitize.js";
+export const OLLAMA_DIAGNOSTIC_RESET_MS = 300000;
+let validatedFailureThreshold;
+/**
+ * Validate the shared failure threshold once and reuse that first-use value.
+ * Missing or invalid PROVIDER_FAILURE_THRESHOLD values use the existing default
+ * of three; ProviderHealthChecker retains its original initialization point.
+ */
+export function getValidatedProviderFailureThreshold() {
+    if (validatedFailureThreshold !== undefined) {
+        return validatedFailureThreshold;
+    }
+    const envValue = process.env.PROVIDER_FAILURE_THRESHOLD;
+    if (!envValue) {
+        validatedFailureThreshold = 3;
+        return validatedFailureThreshold;
+    }
+    const parsed = Number(envValue);
+    if (isNaN(parsed) || parsed <= 0 || parsed > 10) {
+        logger.warn(`Invalid PROVIDER_FAILURE_THRESHOLD: ${envValue} (must be between 1 and 10), using default: 3`);
+        validatedFailureThreshold = 3;
+        return validatedFailureThreshold;
+    }
+    validatedFailureThreshold = parsed;
+    return validatedFailureThreshold;
+}
+/** Return diagnostic error text with URL secrets redacted for logs and status. */
+export function getOllamaDiagnosticErrorMessage(error) {
+    return redactUrlsInText(error instanceof Error ? error.message : String(error));
+}
+const states = new Map();
+const completed = new WeakSet();
+let clearGeneration = 0;
+/**
+ * Normalize an endpoint identity without its fragment, retaining path and query.
+ * The key can contain credentials and must remain private like the source URL.
+ */
+export function getOllamaDiagnosticEndpointKey(url) {
+    const parsed = new URL(url);
+    parsed.hash = "";
+    return parsed.toString();
+}
+/**
+ * Resolve native /api/tags at the configured Ollama origin, independently of a
+ * compatible /v1 base path. OLLAMA_BASE_URL precedes the legacy OLLAMA_API_BASE
+ * alias, then localhost; malformed URLs and non-HTTP(S) schemes are rejected.
+ */
+export function resolveOllamaNativeTagsUrl() {
+    const base = process.env.OLLAMA_BASE_URL ||
+        process.env.OLLAMA_API_BASE ||
+        "http://localhost:11434";
+    const resolved = new URL("/api/tags", base);
+    if (resolved.protocol !== "http:" && resolved.protocol !== "https:") {
+        throw new Error("Invalid OLLAMA_BASE_URL format (OLLAMA_API_BASE is still accepted as a legacy alias)");
+    }
+    resolved.hash = "";
+    return resolved.toString();
+}
+/**
+ * Capture endpoint identity and model before asynchronous diagnostic work.
+ * Later environment changes cannot retarget this context or its health-cache
+ * identity; failure accounting remains shared by endpoint rather than model.
+ */
+export function captureOllamaDiagnosticContext(model = process.env.OLLAMA_MODEL || DEFAULT_OLLAMA_MODEL) {
+    const url = resolveOllamaNativeTagsUrl();
+    return { url, key: getOllamaDiagnosticEndpointKey(url), model };
+}
+/**
+ * Initialize endpoint state or expire failures five minutes after the latest one.
+ * Expiry clears the count/timestamp and advances revision, retaining generation
+ * so existing in-flight permits remain valid. Only an explicit clear changes
+ * generation and invalidates those permits.
+ */
+function getState(context) {
+    let state = states.get(context.key);
+    if (!state) {
+        state = {
+            count: 0,
+            lastFailureAt: undefined,
+            revision: 0,
+            generation: clearGeneration,
+        };
+        states.set(context.key, state);
+    }
+    if (state.lastFailureAt !== undefined &&
+        Date.now() - state.lastFailureAt >= OLLAMA_DIAGNOSTIC_RESET_MS) {
+        // This is a fixed-window expiry, not an outcome from a cache hit/probe.
+        state.count = 0;
+        state.lastFailureAt = undefined;
+        state.revision++;
+    }
+    return state;
+}
+/**
+ * Copy the current policy state after applying initialization or timed expiry.
+ * This observation issues no request and records neither failure nor recovery.
+ */
+export function getOllamaDiagnosticSnapshot(context) {
+    return { ...getState(context) };
+}
+/**
+ * Admit a native probe or report backoff at the current failure threshold.
+ * Timed expiry is applied before this decision. Admission captures the clear
+ * generation without reserving a request or recording an outcome; callers must
+ * record the actual outcome on completion.
+ */
+export function beginOllamaDiagnostic(context) {
+    const state = getState(context);
+    if (state.count >= getValidatedProviderFailureThreshold()) {
+        return {
+            kind: "suppressed",
+            count: state.count,
+            reason: `Ollama native diagnostic backed off after ${state.count} consecutive transport failures`,
+            generation: state.generation,
+        };
+    }
+    return { kind: "allowed", context, generation: state.generation };
+}
+/**
+ * Apply timed expiry, then an admitted outcome once to the latest endpoint state.
+ * Transport/protocol failures increment its current count; success clears a
+ * nonzero count; not-run records no outcome. Timed expiry preserves permits,
+ * including those admitted before the boundary and completed after it.
+ * Duplicate permits and permits invalidated by explicit clearing are ignored.
+ *
+ * @returns A copy of the resulting state, or undefined for an ignored permit.
+ */
+export function completeOllamaDiagnostic(attempt, outcome) {
+    if (completed.has(attempt)) {
+        return undefined;
+    }
+    completed.add(attempt);
+    const storedState = states.get(attempt.context.key);
+    if (!storedState || storedState.generation !== attempt.generation) {
+        return undefined; // Explicit clearing invalidated this in-flight token.
+    }
+    const state = getState(attempt.context);
+    if (outcome === "transport-failure" || outcome === "protocol-failure") {
+        // Read and increment the latest completed state, never the starting count.
+        state.count++;
+        state.lastFailureAt = Date.now();
+        state.revision++;
+    }
+    else if (outcome === "success" && state.count !== 0) {
+        state.count = 0;
+        state.lastFailureAt = undefined;
+        state.revision++;
+    }
+    // Only admitted probes may reach this function. not-run cannot claim failure
+    // or recovery; genuine success while already clear retains the cache revision.
+    return { ...state };
+}
+/**
+ * Clear every endpoint and advance generation, invalidating in-flight permits.
+ * Unlike timed expiry, this prevents an older completion from repopulating state.
+ */
+export function clearOllamaDiagnostics() {
+    clearGeneration++;
+    states.clear();
+}
+/**
+ * Separate cached health by captured endpoint and model without splitting the
+ * endpoint-wide diagnostic counter. This private key can contain URL secrets.
+ */
+export function getOllamaHealthCacheKey(context) {
+    return JSON.stringify(["ollama", context.key, context.model]);
+}
+export class OllamaProtocolError extends Error {
+    /** Identify an answered tags response whose models envelope is malformed. */
+    constructor() {
+        super("Invalid models format from Ollama API");
+        this.name = "OllamaProtocolError";
+    }
+}
+/**
+ * Require an object containing a models array, including an empty array.
+ * Individual model entries are left to existing normalization; a malformed
+ * envelope throws OllamaProtocolError rather than appearing as no models.
+ */
+export function readOllamaTagsModels(payload) {
+    if (payload === null ||
+        typeof payload !== "object" ||
+        Array.isArray(payload) ||
+        !("models" in payload) ||
+        !Array.isArray(payload.models)) {
+        throw new OllamaProtocolError();
+    }
+    return payload.models;
+}

package/dist/utils/providerConfig.d.ts

@@ -400,11 +400,6 @@ export declare function resolveOpenAICompatConfig(entry: OpenAICompatConfigInput
  * TypeSafe (Jev) — the `decide` inference type. Not a text provider.
  */
 export declare function createTypeSafeConfig(): ProviderConfigOptions;
-/**
- * XOR — the `decide` inference type. Not a text provider, and it has no
- * built-in endpoint: XOR_BASE_URL is required alongside the key.
- */
-export declare function createXorConfig(): ProviderConfigOptions;
 /**
  * Perplexity Decisions API — the `decide` inference type, not the Sonar text
  * models. It reads the same PERPLEXITY_API_KEY as the `perplexity` text

package/dist/utils/providerConfig.js

@@ -1324,23 +1324,6 @@ export function createTypeSafeConfig() {
         ],
     };
 }
-/**
- * XOR — the `decide` inference type. Not a text provider, and it has no
- * built-in endpoint: XOR_BASE_URL is required alongside the key.
- */
-export function createXorConfig() {
-    return {
-        providerName: "XOR",
-        envVarName: "XOR_API_KEY",
-        setupUrl: "https://huggingface.co/juspay/xor",
-        description: "API key for your XOR endpoint",
-        instructions: [
-            "1. Reach an XOR deployment, or use a LiteLLM proxy with a pass-through route to one",
-            "2. Set XOR_BASE_URL to its origin (NeuroLink calls <XOR_BASE_URL>/v1/systemone)",
-            "3. Set XOR_API_KEY to the key it accepts (on LiteLLM, a virtual key whose team allows xor-1.1)",
-        ],
-    };
-}
 /**
  * Perplexity Decisions API — the `decide` inference type, not the Sonar text
  * models. It reads the same PERPLEXITY_API_KEY as the `perplexity` text

package/dist/utils/providerHealth.d.ts

@@ -1,7 +1,3 @@
-/**
- * Provider Health Checking System
- * Prevents 500 errors by validating provider availability and configuration
- */
 import { AIProviderName } from "../constants/enums.js";
 import type { ProviderHealthCheckOptions, ProviderHealthStatusOptions } from "../types/index.js";
 export declare class ProviderHealthChecker {
@@ -9,20 +5,16 @@ export declare class ProviderHealthChecker {
     private static readonly DEFAULT_TIMEOUT;
     private static readonly DEFAULT_CACHE_AGE;
     /**
-     * Fixed backoff window for the circuit breaker below, in milliseconds.
-     * Deliberately NOT the per-call `maxCacheAge` option: `consecutiveFailures`
-     * is a single process-wide map keyed only by provider name, shared by every
-     * caller (e.g. `checkFallbackProviderAvailability` hard-codes a 15s
-     * `maxCacheAge` for its own health-status cache). Evicting the breaker
-     * entry using whichever caller happens to read it next would let a
-     * short-TTL caller silently erase a backoff another caller is relying on
-     * — see `checkProviderHealth`. This constant governs breaker expiry alone
-     * and is intentionally independent of `maxCacheAge`/`DEFAULT_CACHE_AGE`.
+     * Fixed backoff for the existing non-Ollama provider breaker. Ollama uses
+     * the same fixed window in ollamaDiagnostics, shared by every native caller.
+     * Neither window depends on a caller's maxCacheAge: expiring a health cache
+     * must not silently erase a transport backoff owned by another caller.
      */
     private static readonly CIRCUIT_BREAKER_RESET_MS;
     private static readonly CONSECUTIVE_FAILURE_THRESHOLD;
     /**
-     * Circuit breaker state, keyed by provider.
+     * Non-Ollama circuit breaker state, keyed by provider. Ollama's separate
+     * native policy captures the actual endpoint and records aggregate outcomes.
      *
      * `count` is a count of consecutive failed CONNECTIVITY PROBES (plus the
      * rare case of the check itself throwing) — never of "the provider is
@@ -56,11 +48,14 @@ export declare class ProviderHealthChecker {
      */
     private static readonly MAX_CONCURRENT_HEALTH_CHECKS;
     /**
-     * Validate and return a safe failure threshold value
+     * Preserve the validated threshold's original first-use initialization.
      */
     private static getValidatedFailureThreshold;
     /**
-     * Comprehensive health check for a provider
+     * Check configuration and optional connectivity/model availability.
+     * Ollama captures its endpoint/model once and records only probes that run;
+     * a cache hit or missing configured model does not invent a transport failure.
+     * Its cached result must match the current diagnostic revision/generation.
      */
     static checkProviderHealth(providerName: AIProviderName, options?: ProviderHealthCheckOptions): Promise<ProviderHealthStatusOptions>;
     /**
@@ -189,18 +184,33 @@ export declare class ProviderHealthChecker {
     private static getLiteLLMModelsUrl;
     private static getConfiguredLiteLLMModel;
     private static getOllamaBaseUrl;
+    /** Resolve the shared native tags URL rather than the compatible models URL. */
     private static getOllamaTagsUrl;
     private static getConfiguredOllamaModel;
     private static fetchJsonWithTimeout;
     private static normalizeModelList;
     private static hasRequestedModel;
+    /**
+     * Fetch native tags from the captured endpoint and normalize its models.
+     * A malformed envelope rejects; a valid empty list remains an empty result.
+     */
     private static getOllamaAvailableModels;
     private static getLiteLLMAvailableModels;
+    /**
+     * Check a requested model against native tags while keeping diagnostic outcome
+     * separate from model availability. An answered valid list is diagnostic
+     * success even if that model is absent; malformed JSON/envelopes are protocol
+     * failures. A caller recording an aggregate health probe disables recording
+     * here so its admitted permit is completed only once at the outer boundary.
+     */
     private static checkOllamaAvailability;
     private static checkLiteLLMAvailability;
     private static checkLiteLLMConfig;
     /**
-     * Check Ollama configuration
+     * Report Ollama configuration and model availability using the captured context.
+     * Backoff can suppress native work; a missing installed model is a configuration
+     * issue after a successful probe. Return actual probe activity/failure for the
+     * enclosing health check to record, without completing a second policy permit.
      */
     private static checkOllamaConfig;
     /**
@@ -208,7 +218,9 @@ export declare class ProviderHealthChecker {
      */
     private static getCommonModelsForProvider;
     /**
-     * Get cached health status if still valid
+     * Return health within the caller's TTL and, for Ollama, the captured identity.
+     * Ollama revision/generation changes invalidate cached status; reading the
+     * snapshot may apply timed expiry but never records a probe outcome.
      */
     private static getCachedHealth;
     /**
@@ -267,7 +279,9 @@ export declare class ProviderHealthChecker {
      */
     static initializeBackgroundHealthChecks(): void;
     /**
-     * Clear health cache for a provider or all providers
+     * Clear cached health and failure state for one provider or every provider.
+     * Clearing Ollama removes all endpoint/model cache entries and explicitly
+     * advances diagnostic generation so older in-flight permits cannot commit.
      */
     static clearHealthCache(providerName?: AIProviderName): void;
     static checkFallbackProviderAvailability(providerName: string, model: string): Promise<{