/**
 * evaluation pipeline orchestrator —— 单次 evaluation run 的「执行+收尾」编排。
 *
 * 主入口 `executeEvaluationPipeline` 把以下几段串成线性时序:
 *   1. `initializeEvaluationRunState` 建 run / job 元数据并写 jobStore
 *   2. resolve judge models + executor map(必须先于 preflight,见函数内注释)
 *   3. `preflight` + `preflightAllJudges` LLM 连通性
 *   4. `emitPowerWarnings` + `emitIsolationWarnings` 结构性 / 隔离性预警
 *   5. `executeTasks` 真正跑用例
 *   6. `finalizeSuccessfulRun` + `finalizeEvaluationReport` 收尾 + 后置分析
 *   7. `persistReport` + `persistSuccessfulJob` 落盘
 *   失败路径走 `persistFailedJob`, finally 关 MCP server。
 *
 * 4 块被拆出去的 helper 都在 `./evaluation-pipeline/` 目录下,与本文件 1:1 协作:
 *   - run-state.ts: EvaluationRunState 生命周期(init / finalizeSuccess / persistSuccess / persistFailed)
 *   - test-set-hash.ts: 测试集水印 hash(spec §7.1)
 *   - preflight-warnings.ts: power + isolation 预警
 *   - report-finalize.ts: report 后置分析装配
 *
 * 兼容 re-export: `buildPowerWarnings` / `buildIsolationWarnings` /
 * `_computeTestSetHashForTest` 保留在本模块的 export surface,避免 test 与
 * `run-evaluation.ts:275` 动态 import 路径改动。
 */
import type { DependencyRequirements } from '../eval-core/dependency-checker.js';
import type { Artifact, ExecutorFn, JobStore, ProgressCallback, Report, Sample, Task } from '../types/index.js';
import type { Lang } from '../types/shared.js';
export { buildPowerWarnings, buildIsolationWarnings } from './evaluation-pipeline/preflight-warnings.js';
export { _computeTestSetHashForTest } from './evaluation-pipeline/test-set-hash.js';
export interface EvaluationPipelineOptions {
    samplesPath: string;
    /** Sample bundle 根目录。loadSamples 输出,缺省时 fallback dirname(samplesPath)。 */
    samplesBaseDir?: string;
    /** Sample bundle 内合并的源文件绝对路径列表。目录模式下用于 computeTestSetHash 遍历。 */
    samplesSourceFiles?: string[];
    skillDir: string;
    samples: Sample[];
    tasks: Task[];
    artifacts: Artifact[];
    model: string;
    judgeModel: string;
    noJudge: boolean;
    executorName: string;
    judgeExecutorName: string;
    executor: ExecutorFn;
    judgeExecutor: ExecutorFn;
    outputDir?: string | null;
    project?: string;
    owner?: string;
    tags?: string[];
    concurrency?: number;
    timeoutMs?: number;
    noCache?: boolean;
    jobStore?: JobStore | null;
    persistJob?: boolean;
    onProgress?: ProgressCallback | null;
    skipConnectivity?: boolean;
    verbose?: boolean;
    retry?: number;
    existingResults?: Record<string, Record<string, VariantResult>>;
    requires?: DependencyRequirements;
    layeredStats?: boolean;
    /** 透传到 meta.request.repeat */
    repeat?: number;
    /** 透传到 meta.request.holdoutRatio；> 0 时 report-finalize 算 train/holdout 子集综合分。 */
    holdoutRatio?: number;
    /** 透传到 meta.request.batch */
    batch?: boolean;
    /** 透传到 meta.request.judgeRepeat 与 grade()，每条 sample × dimension judge N 次 */
    judgeRepeat?: number;
    /** Multi-judge ensemble configs (≥ 2 entries triggers ensemble mode). */
    judgeModels?: import('../types/index.js').JudgeConfig[];
    /** --bootstrap. */
    bootstrap?: boolean;
    /** --bootstrap-samples N. Default 1000. */
    bootstrapSamples?: number;
    /** v0.21 length-debias toggle. Default true; --no-debias-length flips to false. */
    lengthDebias?: boolean;
    /** hard budget caps. */
    budget?: import('../types/index.js').EvalBudget;
    /** strict-baseline default state (only used to decide whether to emit
     *  isolation-disabled pre-flight warnings). True/undefined = default behavior
     *  (no warning); false = user explicitly disabled, warn if ~/.claude/skills/ has content. */
    strictBaseline?: boolean;
    /** Explicit persisted run id. Used by batch workflows that need stable child ids. */
    runId?: string;
    /** CLI output language for warnings emitted by the pipeline. */
    lang?: Lang;
    /** Reasoning effort 透传到 executor。默认 'low'(由 RunConfig 兜底)。 */
    effort?: 'low' | 'medium' | 'high' | 'xhigh' | 'max';
    /** 关闭 diagnostic。Default false。 */
    noDiagnostic?: boolean;
}
type VariantResult = import('../types/index.js').VariantResult;
export declare function executeEvaluationPipeline({ samplesPath, samplesBaseDir, samplesSourceFiles, skillDir, samples, tasks, artifacts, model, judgeModel, noJudge, executorName, judgeExecutorName, executor, judgeExecutor, outputDir, project, owner, tags, concurrency, timeoutMs, noCache, jobStore, persistJob, onProgress, skipConnectivity, verbose, retry, existingResults, requires: _requires, layeredStats, repeat, holdoutRatio, batch, judgeRepeat, judgeModels, bootstrap, bootstrapSamples, lengthDebias, budget, strictBaseline, runId, lang, effort, noDiagnostic, }: EvaluationPipelineOptions): Promise<{
    report: Report;
    filePath: string | null;
}>;