type ModelCapabilities = {
    temperature?: boolean;
    [key: string]: unknown;
};
type ModelInfo = {
    modelID?: string;
    capabilities?: ModelCapabilities;
    [key: string]: unknown;
};
type ChatParamsInput = {
    sessionID: string;
    agent: string;
    model: ModelInfo;
    provider: unknown;
    message: unknown;
};
type ChatParamsOutput = {
    temperature: number;
    topP: number;
    top_p?: number;
    topK: number;
    maxOutputTokens: number | undefined;
    options: Record<string, unknown>;
};
/**
 * Sanitize LLM sampling parameters to avoid API rejections.
 *
 * Two classes of problems are addressed:
 *
 * 1. **Model does not support temperature** — Reasoning models (o1/o3/o4-mini,
 *    Claude thinking, etc.) report `capabilities.temperature === false`.
 *    For these models we strip both `temperature` and `topP`/`top_p` because
 *    the provider will reject the request if either is present.
 *
 * 2. **temperature and top_p both present** — Some providers (e.g. certain
 *    OpenAI endpoints) refuse requests that set both simultaneously.
 *    We keep `temperature` (the more commonly configured knob) and drop `topP`.
 *
 * This hook runs for every LLM call — main agent and subagents alike — so it
 * covers the child-session path that previously caused
 * "`temperature` and `top_p` cannot both be specified" errors.
 */
export declare function createChatParamsHook(): (input: ChatParamsInput, output: ChatParamsOutput) => Promise<void>;
export {};