import type { GatewayProviderContext, GatewayResolvedProviderConfig } from "@cline/shared";
import type { ProviderFactoryResult } from "./types";
/**
 * Normalize a configured base URL to the native Ollama API root expected by
 * the provider (it appends endpoint paths such as `/chat`).
 *
 * Users configure hosts like `http://localhost:11434` or
 * `https://ollama.com`; configs saved by the 4.0.0 OpenAI-compatible
 * routing may carry a `/v1` suffix, and native-API configs an `/api` one.
 */
export declare function normalizeOllamaBaseUrl(baseUrl: string | undefined): string | undefined;
/**
 * Time to wait for the response to start when no timeout is configured.
 *
 * Deliberately generous: Ollama holds `/api/chat` open while it cold-loads
 * the model and only sends response headers once loading finishes, so with a
 * large model (or a large `num_ctx`, which this vendor requests) the first
 * request of a session routinely takes minutes before the stream starts.
 * A tight budget here turns every cold load into a user-facing timeout error
 * (see cline/cline#12829 — the legacy handler's 30s default was only
 * tolerable because its retry decorator silently re-issued the request until
 * the model was loaded). Unreachable servers are not this timeout's job:
 * connection-level failures (refused, DNS) reject on their own immediately,
 * and users can always cancel a request from the UI. This only bounds the
 * accepted-but-silent case, and 5 minutes matches the header-timeout default
 * other AI SDK-based agents use.
 */
export declare const OLLAMA_DEFAULT_TIMEOUT_MS = 300000;
/**
 * Read the configured request timeout (the legacy `requestTimeoutMs`
 * setting); zero/invalid values fall back to the default.
 */
export declare function readOllamaTimeoutMs(config: GatewayResolvedProviderConfig): number;
/**
 * Wrap a fetch so the *response* must start within `timeoutMs`. Once headers
 * arrive the timer is cleared — streaming the body is never interrupted.
 * Mirrors the legacy handler, which raced the chat call (stream start)
 * against a timeout rather than bounding the whole generation.
 */
export declare function withOllamaResponseTimeout(baseFetch: typeof fetch, timeoutMs: number): typeof fetch;
export declare function createOllamaProviderModule(config: GatewayResolvedProviderConfig, _context: GatewayProviderContext): Promise<ProviderFactoryResult>;
//# sourceMappingURL=ollama.d.ts.map