import type { Test, TestInput } from "../../eval/runTypes.js";
import type { EvalRecord } from "../../eval/types.js";
import type { AgencyRunner } from "./agencyRunner.js";
/** The unified run-spec type lives in the eval layer; re-export it so the many
 *  `import { Test } from "./types.js"` sites in the optimizer keep working. */
export type { Test };
/** A JSON-compatible value. */
export type JSON = string | number | boolean | null | JSON[] | {
    [key: string]: JSON;
};
/** A path of object keys / array indices into a JSON value. */
export type JSONPath = (string | number)[];
/** One trace's evidence as grading sees it — as opposed to `AgentRun`
 *  (lib/eval/run/runAgent.ts), which is the outcome of EXECUTING. */
export type LoadedRun = {
    output: JSON;
    traceId: string;
    workdir: string;
    record: EvalRecord;
};
/** A grader's score: pass/fail or a continuous value. */
export type Score = {
    kind: "binary";
    pass: boolean;
} | {
    kind: "scalar";
    value: number;
};
/** One answer from a checklist judge: the point as the test wrote it,
 *  whether the work met it, and the words from the work that decided it. */
export type Point = {
    point: string;
    pass: boolean;
    evidence: string;
};
/** A grader's output: a score plus optional natural-language feedback, and
 *  for a checklist grader the answer per point. */
export type Grade = {
    score: Score;
    feedback?: string;
    points?: Point[];
};
/** Options common to every grader; subclasses extend this with their own fields. */
export type GraderOptions = {
    mustPass?: boolean;
    threshold?: number;
    weight?: number;
    samples?: number;
    aggregate?: "any" | "all";
    name?: string;
};
/** What a grader's `_run` receives. */
export type GraderInput<T = TestInput> = {
    test: Test<T>;
    run: LoadedRun;
    runAgency: AgencyRunner;
    /** The test's grader-only files, as a directory; absent when it has none. */
    graderFiles?: string;
};