import { type UserContent, type TranscriptionModel } from 'ai';
/**
 * User-turn content the runtime accepts: plain text, or AI SDK multimodal parts
 * (text + file/image/audio). This is exactly the `content` of a user `ModelMessage`,
 * so it threads into the model with no translation.
 *
 * Durability invariant: any `FilePart.data` flowing through the runtime must be
 * JSON-serializable (a base64 string, data URL, or https URL) — never a raw
 * `Buffer`/`Uint8Array`. `RunState.messages`, `session.messages`, and the pending
 * input buffer are all persisted through the `SessionStore` (JSON/Redis/Postgres).
 */
export type UserInputContent = UserContent;
/** Merge multiple user inputs into one turn (ingress coalescing / mid-turn drain). */
export declare function mergeUserInputContents(items: UserInputContent[]): UserInputContent | undefined;
/** Text projection of user input — for confirm-gate parsing, choice matching, and
 *  extraction hints. Non-text parts are dropped. A plain string returns as-is. */
export declare function userInputToText(input: UserInputContent): string;
/** Whether the input carries any non-text (file/image/audio) parts. */
export declare function hasMediaParts(input: UserInputContent): boolean;
/**
 * Replace audio file parts with their transcript (a text part) using an AI SDK
 * transcription model. With no model configured, audio parts pass through
 * unchanged — audio-capable models (e.g. Gemini) accept them directly. Non-audio
 * parts (images, documents) are always left untouched.
 */
export declare function transcribeAudioParts(input: UserInputContent, transcriptionModel: TranscriptionModel | undefined): Promise<UserInputContent>;