/**
 * Where to start when every account looks out of room.
 *
 * ccx exists to keep you working across limits. The one thing it must never do
 * is become the reason you cannot start Claude at all, and that is exactly what
 * it did: with every account recorded as week-spent it refused to launch, so
 * `claude` did nothing and the only output was "every account has hit its
 * limit". The operator had usage the whole time.
 *
 * ccx is not the authority on whether a request will be served, and it is wrong
 * in more ways than it can enumerate:
 *
 *   - usage credits carry an account past its plan limit, and a cached snapshot
 *     knows nothing about them
 *   - a cap can have been recorded against the wrong account
 *   - a window can reset earlier than the reset time we stored
 *   - a plan can change under us
 *
 * Every one of those ends with ccx refusing work the server would have done. So
 * when there is any account left with a login, we start on it, say plainly why,
 * and let the server give the answer. Refusing is reserved for the case where
 * there is genuinely nothing to launch: no account with a usable login.
 */
export interface StartCandidate {
    name: string;
    dir: string;
}
export interface LastResortInput {
    /** Accounts that could be started, in the order rotation would try them. */
    usable: StartCandidate[];
    /** Preferred when it can still run, so the session stays where it was. */
    active?: string | null;
    /** Set only when every active limit is about ONE model. */
    modelOnly?: {
        model: string;
        resetsAt: number | null;
    } | null;
    /** Injected so the message is stable in tests. */
    formatTime?: (epochMs: number) => string;
}
export interface LastResortStart {
    account: StartCandidate;
    message: string;
}
/**
 * The account to start on anyway, and what to tell the operator, or null when
 * there is nothing at all to start.
 */
export declare function lastResortStart(input: LastResortInput): LastResortStart | null;