import type { AgencyConfig } from "../../config/config.js";
import type { EvalInputRunner } from "../../eval/run/subprocess.js";
import type { SuiteRunResult, Test } from "../../eval/runTypes.js";
import type { SuiteIdentity } from "../../runDirectory/annotations.js";
export type EvalRunCliOptions = {
    /** File agent target. Exactly one of agent / agentCmd. */
    agent?: string;
    /** Command agent target: the command string with a {input} placeholder. */
    agentCmd?: string;
    /** The test suite: a JSON file, a directory, or a git source. */
    suite?: string;
    /** One inline test with this input text; no suite file needed. */
    input?: string;
    /** Directory to write the run directories into, one per test at `<out>/<testId>/`
     *  (default `<eval.runsDir or runs>/<timestamp>-<random suffix>`). */
    out?: string;
    config?: AgencyConfig;
    /** Worker-pool size (-n/--parallel); default 1 = sequential. */
    parallel?: number;
    /** `--test` id patterns; only matching tests run (suite runs only). */
    test?: string[];
    /** `--tags` values; only tests carrying every one run (suite runs only). */
    tags?: string[];
    /** `--trials`: run every selected test this many times (default 1). */
    trials?: number;
};
/** `--suite` runs a suite; otherwise the run is one inline test, with
 *  `--input`'s text or (for an agent that takes no argument) no input. */
export declare function validateInputSelection(opts: {
    suite?: string;
    input?: string;
}): "suite" | "input";
/**
 * The `agency eval run` command: load the suite, run it, write the run
 * directory. It never grades — `agency eval grade <dir>` does that, whenever
 * you like, and no `goal` is needed to run.
 */
export declare function evalRun(opts: EvalRunCliOptions, 
/** Test seam — the CLI has no flag for it. */
deps?: {
    runner?: EvalInputRunner;
}): Promise<SuiteRunResult>;
export type LoadedSuite = {
    tests: Test[];
    identity: SuiteIdentity;
};
/** Load the suite named by --suite/--input, resolving a git source when given
 *  one and recording the resolved sha as the suite's identity. Running never
 *  needs a `goal`, so none is required here; `eval grade` takes one (`--goal`)
 *  or reads each test's own when its judge needs it. Exported for `eval ls`,
 *  which must see the suite exactly as a run would. */
export declare function loadSuite(args: {
    selection: "suite" | "input";
    source?: string;
    input?: string;
    cacheRoot?: string;
}): LoadedSuite;