/**
 * oclif typed flags → `RunConfig` 装配。
 *
 * 主体 `parseRunConfig` 是「CLI > eval.yaml > 硬编码 default」精度链的统一落点 ——
 * 把 oclif strict 模式接受过的 flag values + 可选 eval.yaml 合并成 eval 子命令运行
 * 所需的完整 `RunConfig`,后续 `executeEvaluationPipeline` 与 doctor / batch 都从这
 * 里读字段。
 *
 * 3 个 sub-routine 拆到同级 parse-run-config/ 目录,主文件保留它们的 re-export
 * 让外部 import surface(`parseJudgeModelsArg` / `parseJudgeModelsArgOrExit`)
 * 不动:
 *   - judge-models.ts: --judge-models 字符串解析
 *   - samples-discovery.ts: 未传 --samples 时的路径发现
 *   - variant-resolution.ts: --control / --treatment / eval.yaml.variants 三态合并
 *
 * 主体 parseRunConfig 自身保留约 100 行的 field-default fanout —— 每个字段一行
 * `cli ?? evalConfig ?? default`,刻意不再细拆,否则只是把一长串 ?? 散到多文件,
 * 反而失去「精度链一目了然」的可读性。
 */
import type { EvalConfig, VariantSpec, JudgeConfig, EvalBudget, ProgressCallback } from '../../types/index.js';
import { type RuntimeResolutionOptions } from './runtime-defaults.js';
export { parseJudgeModelsArg, parseJudgeModelsArgOrExit } from './parse-run-config/judge-models.js';
export interface RunConfig {
    samplesPath: string;
    skillDir: string;
    variantSpecs: VariantSpec[];
    model: string;
    outputDir: string;
    noJudge: boolean | undefined;
    noCache: boolean | undefined;
    dryRun: boolean | undefined;
    concurrency: number;
    timeoutMs: number;
    executorName: string;
    /** 跳过 LLM 模型连通性检测。仅当 --resume 报告通过完整契约校验时自动 true。 */
    skipConnectivity: boolean | undefined;
    /** 跳过 doctor 健康检查门禁(--skip-doctor)。escape hatch — 默认 false。
     *  开启后 doctor 整段不跑(节省静态检查时间);doctor 失败也不再阻断 eval。
     *  典型场景:依赖在评测环境中通过 mock / stub 提供,doctor 的物理路径检查
     *  会误报。开启意味着用户接受 garbage-in 风险,自己负责依赖正确性。 */
    skipDoctor: boolean | undefined;
    /** 用户语言, 透传给 doctor 报告渲染。 */
    lang: 'zh' | 'en' | undefined;
    mcpConfig: string | undefined;
    verbose: boolean | undefined;
    retry?: number;
    resume?: string;
    layeredStats?: boolean;
    /** --holdout-ratio R (0 < R < 1). Hold out a deterministic sample slice; report-finalize
     *  computes train vs holdout composite (report.analysis.holdout) for the overfitting gate. */
    holdoutRatio?: number;
    /** --judge-repeat N. Calls LLM judge N times per (sample × dimension). Default 1. */
    judgeRepeat?: number;
    /** Unified judge config. Always non-empty; 1 entry = single judge, ≥ 2 = ensemble.
     *  Defaults to the selected runtime's judge model. */
    judgeModels: JudgeConfig[];
    /** --bootstrap. Adds bootstrap CI to summary (per-variant mean + pairwise diff). */
    bootstrap?: boolean;
    /** --bootstrap-samples N. Bootstrap resamples count, default 1000. */
    bootstrapSamples?: number;
    /** length-debias toggle. Default true; --no-debias-length sets false. */
    lengthDebias?: boolean;
    /** hard budget caps from CLI or config. */
    budget?: EvalBudget;
    /** Skill isolation default for baseline-kind variants. Default true.
     *  CLI flag --no-strict-baseline disables strict isolation. */
    strictBaseline?: boolean;
    /** Per-variant allowedSkills override extracted from eval.yaml. Always wins
     *  over strictBaseline default. Keyed by variant name. */
    variantAllowedSkills?: Record<string, string[]>;
    /** Reasoning effort for the executor LLM(被评测的模型,不是 judge)。
     *  Default 'low' — sonnet 默认会做大量扩展思考(13K thinking tokens / 单次),
     *  对结构化任务是浪费。低 effort 大幅省时间/成本但可能损失复杂推理质量,
     *  跨 effort 的报告不能严格比较。`undefined` 走 claude CLI / SDK 自身默认。 */
    effort?: 'low' | 'medium' | 'high' | 'xhigh' | 'max';
    /** 关闭 diagnostic LLM call。Default false(总是给 failed sample 跑诊断)。
     *  跟 noJudge 完全独立 — judge 答打分,diagnostic 答怎么改。 */
    noDiagnostic?: boolean;
    onProgress?: ProgressCallback | null;
}
export interface ParseRunConfigResult {
    values: Record<string, string | boolean | undefined>;
    config: RunConfig;
    /** Loaded eval.yaml when --config was provided. `commands/eval-runner.ts` uses it to apply
     *  CLI > eval.yaml > default fallback for fields not propagated by parseRunConfig
     *  (e.g. repeat / judgeRepeat / bootstrap — handled in `commands/eval-runner.ts` for input validation). */
    evalConfig: EvalConfig | null;
}
/**
 * 接 typed flags(来自 oclif Command.parse() 输出)。oclif strict 模式已经在
 * 上游对未知 flag 拦截 exit 2,这里不再 parseArgs。eval-runner 等业务 caller
 * 把 oclif flags 当 values 喂进来。
 */
export declare function parseRunConfig(values: Record<string, unknown>, runtimeOptions?: RuntimeResolutionOptions): ParseRunConfigResult;