import type { AgencyConfig } from "../config/config.js";
import type { GraderSource } from "../eval/grading/gradeRun.js";
import type { Test } from "../eval/grading/types.js";
import type { OptimizeResult } from "./types.js";
/** What to optimize: an agent (file[:node]) and the inputs to run it on,
 *  plus an optional held-out validation set used to pick the champion. */
export type OptimizeTarget = {
    agent: string;
    inputs: Test[];
    validationInputs?: Test[];
};
/** Cross-cutting config every optimizer needs; each optimizer may extend it. */
export type BaseOptimizerConfig = {
    /** The objective. "snapshot" (the CLI default with --suite): each input is
     *  graded by its own graders — the test's graders.ts and harness pairs,
     *  stored in its run directory the way eval run stores them, the goal
     *  judge for tests with neither. Candidate runs read the suite's grader
     *  files as they happen, so the same rule applies as to eval run: do not
     *  edit the suite while a search is running. "override" (--graders): one
     *  set for every input, the experiment knob. The grading layer's third
     *  source ("suite", re-grading old runs against live graders) has no
     *  meaning here, so the type excludes it. */
    graders: Extract<GraderSource, {
        kind: "snapshot";
    } | {
        kind: "override";
    }>;
    iterations: number;
    seed?: number;
    config: AgencyConfig;
    runsDir: string;
    runId: string;
    writeback?: boolean;
    mutatorModel?: string;
    /** Progress output verbosity. Defaults to silent (programmatic use); the CLI sets "default". */
    verbosity?: "silent" | "default";
};
/** A pluggable optimization strategy. */
export type Optimizer = {
    readonly name: string;
    optimize(target: OptimizeTarget): Promise<OptimizeResult>;
};
export type OptimizerFactory = (config: BaseOptimizerConfig) => Optimizer;