import { type EvalTarget } from "../../agentTarget.js";
import type { AgencyConfig } from "../../config/config.js";
import type { SuiteRunResult, Test } from "../../eval/runTypes.js";
import type { SuiteIdentity } from "../../runDirectory/annotations.js";
import { type RunAgentOptions } from "./runAgent.js";
import type { EvalInputRunner } from "./subprocess.js";
/** Per-run knobs forwarded verbatim to every runAgent call in the suite.
 *  Pick, not re-declaration: the compiler keeps this in sync with
 *  RunAgentOptions. NOTE the omission semantics shift at suite level: with
 *  `seed` omitted, runSuite computes ONE closure walk for the whole suite and
 *  passes it to every run — never one walk per input (RunAgentOptions.seed's
 *  own doc describes the single-run case). */
export type PerRunOptions = Pick<RunAgentOptions, "seed" | "overlayFiles" | "pipeOutput">;
/** Options for running a LOADED suite: parsed Test[], resolved values.
 *  The raw-flags side lives in the evalRun command (EvalRunCliOptions). */
export type RunSuiteOptions = {
    /** The agent. A string is file-target convenience (path, path:node, or a
     *  directory meaning main.agency) and is resolved here; a resolved
     *  EvalTarget passes through with no re-validation — it already passed
     *  resolveEvalTarget, including the {input}-placeholder check. */
    agent: string | EvalTarget;
    inputs: Test[];
    /** The group directory; each test's run directory is written at
     *  `<out>/<testId>/`. Default: `<eval.runsDir or runs>/<timestamp>-<random suffix>`. */
    out?: string;
    /** Default true. */
    config?: AgencyConfig;
    /** Worker-pool size for input scheduling; 1 (default) = sequential with
     *  piped agent output, >1 = parallel with a status board instead. */
    parallel?: number;
    /** Default true: suite-level progress on stderr (the run-dir line, the
     *  per-input heartbeat and status lines, the parallel status board). The
     *  optimizer turns it off — its reporter owns the narrative, and
     *  `optimize --silent` must print nothing at all. */
    progress?: boolean;
    /** Where the suite came from, recorded on every test's `run` annotation. */
    suite?: SuiteIdentity;
    /** Run every test this many times, each as its own run directory at
     *  `<out>/<testId>/<trial>/`; 1 (the default) keeps `<out>/<testId>/`.
     *  A finite positive integer. */
    trials?: number;
    perRun?: PerRunOptions;
};
/** Test seam, same pattern as RunAgentDeps. */
export type RunSuiteDeps = {
    runner?: EvalInputRunner;
};
/**
 * Run an agent against a loaded suite and write one run directory per test
 * at `<out>/<testId>/`: the test's trace in `statelog.jsonl`, its workdir
 * under `workdir/`, the agent's code under `code/`, and one `run` annotation
 * saying which test it was and how it ended. A test whose directory already
 * exists is an error result; the others still run. Executes only — grading is
 * a separate pass over the finished directories (docs/dev/evals/eval-grading.md).
 */
export declare function runSuite(opts: RunSuiteOptions, deps?: RunSuiteDeps): Promise<SuiteRunResult>;