/**
 * kosha-discovery — Shared LiteLLM catalog loader.
 *
 * Single source of truth for the community-maintained LiteLLM model catalog
 * used by both the discovery seed (origin providers without an API key) and
 * the post-discovery pricing enricher.
 *
 * Hardening:
 *  - HTTPS-only, fixed source URL — no caller-supplied URLs.
 *  - Promise-deduplicated singleton load — concurrent callers share one fetch.
 *  - AbortController-bounded fetch with explicit timeout.
 *  - Response size cap before parsing to prevent memory bombs.
 *  - {@link quarantineEntries} runs before any field is read, dropping an
 *    unclean model entry rather than the whole catalog.
 *  - Entry-count cap after parse to bound downstream work.
 *
 * The catalog source is the same one already trusted by the existing
 * enricher; this module just centralises the load.
 * @module
 */
import { type QuarantinedEntry } from "../security.js";
/** Pinned upstream catalog URL — HTTPS only, no user override. */
export declare const LITELLM_CATALOG_URL = "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json";
/**
 * Shape of a single entry in the LiteLLM pricing JSON.
 * Fields are all optional because the upstream schema evolves over time;
 * consumers must defensively handle missing fields.
 */
export interface LiteLLMModelEntry {
    max_tokens?: number;
    max_input_tokens?: number;
    max_output_tokens?: number;
    input_cost_per_token?: number;
    output_cost_per_token?: number;
    input_cost_per_reasoning_token?: number;
    output_cost_per_reasoning_token?: number;
    reasoning_input_cost_per_token?: number;
    reasoning_output_cost_per_token?: number;
    cache_read_input_token_cost?: number;
    cache_creation_input_token_cost?: number;
    input_cost_per_token_batches?: number;
    output_cost_per_token_batches?: number;
    input_cost_per_image?: number;
    output_cost_per_image?: number;
    input_cost_per_audio_token?: number;
    output_cost_per_audio_token?: number;
    input_cost_per_audio_per_second?: number;
    output_cost_per_audio_per_second?: number;
    input_cost_per_video_per_second?: number;
    output_cost_per_video_per_second?: number;
    input_cost_per_video_token?: number;
    input_cost_per_character?: number;
    output_cost_per_character?: number;
    input_cost_per_token_above_128k_tokens?: number;
    output_cost_per_token_above_128k_tokens?: number;
    input_cost_per_token_above_200k_tokens?: number;
    output_cost_per_token_above_200k_tokens?: number;
    deprecation_date?: string;
    litellm_provider?: string;
    mode?: string;
    supports_function_calling?: boolean;
    supports_parallel_function_calling?: boolean;
    supports_tool_choice?: boolean;
    supports_response_schema?: boolean;
    supports_system_messages?: boolean;
    supports_vision?: boolean;
    supports_audio_input?: boolean;
    supports_audio_output?: boolean;
    supports_video_input?: boolean;
    supports_prompt_caching?: boolean;
    supports_reasoning?: boolean;
    output_vector_size?: number;
}
/**
 * Model entries the threat scan dropped from the last successful load.
 *
 * A quarantined entry is a supply-chain signal — a base64 credential or a
 * script payload someone committed to the upstream catalog. Dropping the row
 * keeps the other few thousand models usable, but the drop has to be
 * reportable rather than silent, which is what this exposes.
 */
export declare function liteLLMQuarantined(): QuarantinedEntry[];
/**
 * Fetch the LiteLLM catalog with full hardening, returning a defensively
 * filtered map. Concurrent callers share the same in-flight promise.
 */
export declare function loadLiteLLMCatalog(): Promise<Record<string, LiteLLMModelEntry>>;
/**
 * Reset the cached catalog. Test-only helper — callers in production code
 * should rely on the process-lifetime cache.
 */
export declare function resetLiteLLMCatalogCache(): void;
//# sourceMappingURL=litellm-catalog.d.ts.map