import type { Api, AssistantMessage, Model, SimpleStreamOptions } from "../types.js";
import type { TextBackend } from "./text.js";
/**
 * Output budget for keyword replies. Sized against two independent constraints:
 *   - Backends that ignore `disableReasoning` still emit a thinking preamble
 *     (e.g. Qwen3 via llama.cpp catalogued `reasoning: false` but still thinking;
 *     Anthropic via LiteLLM/Vertex, whose `openai-completions` route downgrades a
 *     disabled request to the lowest reasoning effort instead of turning thinking
 *     off). The keyword must have room to land after that preamble (issue #4355).
 *   - Anthropic-dialect proxies reject `max_tokens <= thinking.budget_tokens`. The
 *     pinned lowest effort maps to at least Anthropic's 1024-token minimum budget,
 *     so the cap MUST comfortably exceed 1024 or every call 400s with
 *     `max_tokens must be greater than thinking.budget_tokens` (issue #8610).
 * `maxTokens` is a hard cap — non-thinking completions still return in a handful
 * of tokens.
 */
export declare const JUDGMENT_CHAT_MAX_TOKENS = 4096;
export type ChatTextBackendOptions = Pick<SimpleStreamOptions, "apiKey" | "sessionId" | "metadata"> & {
    /** Receives every completed attempt (including transient failures) for usage accounting. */
    onAttempt?: (message: AssistantMessage) => void;
};
/** Chat completions with reasoning disabled, temperature 0, and transient-failure retry. */
export declare function chatTextBackend(model: Model<Api>, options: ChatTextBackendOptions): TextBackend;