/**
 * The evals runner: discover `*.eval.ts` / `evals/*.ts` files, run each eval
 * through its scorer pipeline against the *real* agent loop, score, and report.
 *
 * It is the engine behind `agent-native eval` — when used as a CI deploy gate
 * the CLI exits non-zero if any eval scores below its threshold.
 *
 * Two layers:
 *   - `scoreEval` / `runEvals` — pure orchestration over an `AgentRunner` and
 *     a list of evals. Fully unit-testable with an injected runner (no model).
 *   - `discoverEvalFiles` / `loadEvals` — filesystem discovery + dynamic import
 *     of author-written eval modules.
 *
 * Results are also (best-effort) written to the observability eval store so a
 * dashboard can surface CI eval history next to production run evals.
 */
import type { ActionEntry } from "../agent/production-agent.js";
import type { AgentRunner } from "./agent-runner.js";
import type { Eval, EvalResultRow, EvalRunReport } from "./types.js";
/** Run a single eval: invoke the agent, then score with each scorer. */
export declare function scoreEval(evalCase: Eval, runner: AgentRunner, opts?: {
    thresholdOverride?: number;
}): Promise<EvalResultRow>;
/** Run a batch of evals against one runner and aggregate a report. */
export declare function runEvals(evals: Eval[], runner: AgentRunner, opts?: {
    thresholdOverride?: number;
    persist?: boolean;
}): Promise<EvalRunReport>;
/**
 * Walk `root` for eval files. Matches two conventions:
 *   - any `**\/*.eval.ts` (co-located with code), and
 *   - any `*.ts` directly inside an `evals/` directory.
 * `pattern` further filters by substring of the relative path.
 */
export declare function discoverEvalFiles(root: string, pattern?: string): Promise<string[]>;
/** Discover and import all eval files under `root`, returning their evals. */
export declare function loadEvals(root: string, pattern?: string): Promise<{
    files: string[];
    evals: Eval[];
}>;
export interface RunEvalSuiteOptions {
    /** App root to discover eval files + actions under. Defaults to cwd. */
    cwd?: string;
    /** Substring filter on the eval file path. */
    pattern?: string;
    /** Global threshold override (wins over per-eval thresholds). */
    thresholdOverride?: number;
    /** App actions to expose to the agent. Auto-discovered when omitted. */
    actions?: Record<string, ActionEntry>;
    /** System prompt for runs. */
    systemPrompt?: string;
    /** Write results to the observability eval store (default true). */
    persist?: boolean;
    /** Pre-built runner (tests inject this to avoid touching engine/loop). */
    runner?: AgentRunner;
    /** Pre-loaded evals (tests inject this to skip filesystem discovery). */
    evals?: Eval[];
}
/**
 * End-to-end: load evals, build a runner, score, report. The CLI wraps this
 * and maps `report.failed > 0` to a non-zero exit code (the CI gate).
 */
export declare function runEvalSuite(opts?: RunEvalSuiteOptions): Promise<{
    report: EvalRunReport;
    files: string[];
}>;
//# sourceMappingURL=runner.d.ts.map