/**
 * The `/eval` TUI command: run a benchmark task inside the interactive session, so every
 * tool call and observation is visible live, then score it with the same harness the
 * batch runner uses.
 *
 *   /eval                     list every installed harness's benchmarks, suites, tasks
 *   /eval 24                  run a task of this session's default benchmark
 *   /eval webarena 24         run task 24 of a named benchmark
 *   /eval androidworld ContactsAddContact
 *
 * All benchmark knowledge (environments, scoring, reports) stays in the eval harnesses;
 * this file only shells out to them. Each benchmark maps to a harness CLI and to the
 * DRIVER it needs — the session's driver is frozen at bootstrap, so a mismatch is an
 * honest "relaunch with PI_GUI_DRIVER=…", never a mid-session driver swap.
 *
 * Scoring goes through the prepare process's hold protocol (pge_core/protocol.py):
 * benchmarks like AndroidWorld judge from LIVE environment state, so the verdict must
 * come from the process that holds the environment.
 */
import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
import type { Profile } from "../../agent/session.ts";
export declare function registerEvalCommand(pi: ExtensionAPI, profile: Profile): void;