/** A message appended to the conversation. `ts` is the turn's boundary timestamp
 * (unix ms), captured when the turn happened — the server mints a K-sortable id
 * from it, so order is fixed independently of when the transcript text arrives. */
export interface VoiceTranscriptMessage {
    /** OpenAI conversation item id — the idempotency seed for the minted message id. */
    itemId: string;
    /** Turn boundary time (unix ms). Sorts the message; trusted within margin. */
    ts: number;
    role: "user" | "assistant";
    parts: Array<{
        type: "text";
        text: string;
    }>;
}
/**
 * Turns out-of-order Realtime transcript events into an **append-only** stream of
 * independently-ordered messages.
 *
 * User input transcription is asynchronous and can complete *after* the assistant
 * has already replied, so a user turn reserves its boundary timestamp at the VAD
 * speech boundary (`userTurn(ts)`) — that time sorts it before the assistant
 * reply even though its text lands later. Each settled transcript becomes one
 * message; there is no buffer holding the whole conversation and no
 * contiguous-prefix gate, so a missing or empty transcript drops only that one
 * message, never the rest.
 *
 * Messages stay in `unacked` (keyed by `itemId`) until the server confirms them;
 * that set doubles as the retry queue (failed POST), the defer queue (a greeting
 * that arrives before any user turn), and the `pagehide` beacon payload. It holds
 * only in-flight messages — normally zero or one — never history.
 */
export declare class VoiceTranscriptSequencer {
    /** Reserved boundary timestamps for user turns whose transcript hasn't arrived (FIFO). */
    private pendingUserTs;
    /** Item ids already recorded, to drop duplicate transcript-completion events. */
    private seenItemIds;
    /** Messages awaiting server confirmation, keyed by itemId. */
    private unacked;
    /** Reserve the boundary timestamp for a user turn (its transcript arrives later). */
    userTurn(ts: number): void;
    /** Release the oldest reserved user boundary when its input transcription failed. */
    userTranscriptFailed(): void;
    /**
     * Record a settled transcript as a message and queue it for sending. Returns the
     * message, or null if it was a duplicate or had no text. A user message takes its
     * reserved boundary timestamp (so it sorts before the assistant reply); anything
     * else takes `ts` (the moment it settled).
     */
    record(itemId: string, role: "user" | "assistant", text: string, ts: number): VoiceTranscriptMessage | null;
    /** Messages awaiting confirmation, oldest first — the batch to send or beacon. */
    pending(): VoiceTranscriptMessage[];
    /** Drop the given item ids once the server confirms them. */
    ack(itemIds: string[]): void;
}
//# sourceMappingURL=voice-transcript.d.ts.map