import type { AgencyConfig } from "../../config/config.js";
import { type BatchStatistics, type IncompleteBatch } from "../../eval/batchStatistics.js";
import type { EvalRunGrading } from "../../eval/runTypes.js";
import type { GraderSource } from "../../eval/grading/gradeRun.js";
export type EvalGradeOptions = {
    /** Grade with each test's CURRENT graders from this suite (a file, a
     *  directory, or a git source), matched to runs by test id, instead of the
     *  copy each run directory stored. Improve a grader, re-score old runs. */
    suite?: string;
    /** Judge every trace against this goal with the bundled goal judge (a test's
     *  own recorded goal still wins). Not combined with `suite`: the suite's
     *  graders bring their own criteria, and `--goal` names the criterion. */
    goal?: string;
    /** Also write the grading summary here, as JSON. */
    out?: string;
    /** Grade up to this many run directories at once (default 1). Judge calls
     *  are LLM calls, so a big group grades much faster in parallel. */
    parallel?: number;
    /** Called once per graded run directory, in completion order — the CLI's
     *  progress line. Omitted (tests, programmatic use) = silent. */
    progress?: (message: string) => void;
    config?: AgencyConfig;
};
/** One graded run directory and the mean over them all. */
export type EvalGradeResult = {
    runs: {
        dir: string;
        grading: EvalRunGrading;
    }[];
    /** Mean objective over the runs graded. */
    mean: number;
    /** Total LLM spend of this pass's judge calls, in USD. */
    judgeCostUsd: number;
    /** Every run passed its gates. */
    gatesPassed: boolean;
    /** Trial statistics for each selected batch that ran more than one trial,
     *  in the order the batches were met. Unrelated batches never merge. */
    batches: BatchStatistics[];
    /** Batches with no statistics — an uneven trial grid, say — and why. Their
     *  runs are still graded above; only the batch-level numbers are missing. */
    incompleteBatches: IncompleteBatch[];
};
/** The command's own preconditions, checked before anything loads: the two
 *  ways of saying "judge against this" are exclusive, and the targets must
 *  hold run directories (a bare statelog copied into a folder is the common
 *  miss). Returns canonical run directories to grade, duplicates removed by
 *  physical identity: `eval grade runs/suite runs/suite/a` grades `a` once. */
export declare function validateGradeTarget(targets: string[], opts: EvalGradeOptions): string[];
/** Where `eval grade` takes each run's graders from: the suite named by
 *  `--suite`, else the copy the run directory stored. */
export declare function graderSourceFor(opts: EvalGradeOptions, config: AgencyConfig): GraderSource;
/**
 * Score run directories, or every run directory in groups. Never re-executes
 * the agent. Each run gets its own grading pass, appended to ITS
 * `annotations.jsonl` — a re-grade sits beside the earlier ones, never over
 * them.
 */
export declare function evalGrade(targets: string[], opts: EvalGradeOptions): Promise<EvalGradeResult>;
/**
 * The `agency eval grade` action: progress lines on stderr, the report on
 * stdout. Returns whether every run passed its gates — the caller owns the
 * exit code.
 */
export declare function runGradeCommand(paths: string[], opts: EvalGradeOptions): Promise<boolean>;