/**
 * ask Command - one round trip to another agent session (Issue #2376)
 *
 *   commandmate ask <worktree-id> "<message>" [--instance <id|alias>]
 *                   [--agent <tool>] [--timeout <seconds>] [--json]
 *
 * ## Why this exists next to send / wait / capture
 *
 * Delegation between two agent sessions was a HARNESS, not a command. Session A
 * had to `send`, then `wait`, then branch on four exit codes, then `capture` and
 * work out which part of the pane was the reply — and every one of those steps
 * is a place to get it wrong in a way that fails quietly. Dropping the `wait`
 * spins; reading the pane without one reads the previous turn; answering the
 * other session's prompt with `respond` picks whatever option happened to be
 * highlighted (#1681).
 *
 * `ask` is those three steps with the branch already taken. It is deliberately
 * NOT a new mechanism: the send is the same POST, the wait is `wait`'s own
 * {@link pollWorktree}, and the exit codes are `wait`'s exit codes. What it adds
 * is the last step nothing had — turning "the turn ended" into "here is what
 * they said" — and the two guarantees the brief in the GUI promises: `--auto-yes`
 * is not offered, and a prompt is reported rather than answered.
 *
 * ## Where the reply comes from
 *
 * The chat ledger first (`chat_messages`, what the transcript readers and the
 * structured-history gate write), because that is the agent's reply as text —
 * no box drawing, no composer, no spinner frames. A tool that has no transcript
 * reader falls back to the squeezed pane, and `--json` says which of the two
 * answered in `source` so a caller can tell a real reply from a screen scrape.
 */
import { Command } from 'commander';
export declare function createAskCommand(): Command;
//# sourceMappingURL=ask.d.ts.map