/**
 * Metric-threshold graders — check trajectory metrics against configurable limits.
 *
 * These graders wrap the metrics already computed by {@link computeMetrics}
 * (token usage, tool calls, turns, errors, wall time) — plus `step-count`,
 * which counts agent steps directly from the event stream — and expose them as
 * first-class graders with a `max` threshold in eval.yaml:
 *
 * ```yaml
 * graders:
 *   - type: token-budget
 *     config: { max: 50000 }
 *   - type: tool-call-count
 *     config: { max: 20 }
 *   - type: step-count
 *     config: { max: 40 }
 *   - type: wall-time
 *     config: { max: "2m" }
 * ```
 *
 * `max` is required — these graders always enforce a budget.
 * `wall-time` accepts duration strings (`"30s"`, `"2m"`); the others accept numbers.
 *
 * All are deterministic and language-agnostic. Five are free — they read values
 * the pipeline has already computed. `step-count` is free but scans the event
 * stream (and applies regex matching when `tools` is set) rather than reading
 * a precomputed metric.
 */
import type { Grader, GraderInput, GraderMetadata, GraderResult } from "../types.js";
/**
 * Configuration for metric-threshold graders.
 *
 * @example
 * ```yaml
 * graders:
 *   - type: token-budget
 *     config: { max: 50000 }
 *   - type: wall-time
 *     config: { max: "2m" }
 * ```
 */
export interface MetricThresholdConfig {
    /**
     * Upper bound for the metric. The trajectory fails if the metric exceeds this value.
     *
     * For most graders this is a number (e.g. token count, tool calls).
     * For `wall-time` this is a duration string (e.g. `"30s"`, `"2m"`, `"1h"`).
     *
     * Required — omitting `max` is a validation error.
     */
    max?: number | string;
}
/** Shared config for metric graders that accept a `tools` regex filter. */
export interface ToolFilteredMetricThresholdConfig extends MetricThresholdConfig {
    /**
     * Regex patterns selecting entries to count. `step-count` also matches
     * `response` and system event names.
     */
    tools?: string[];
}
/** @deprecated Use {@link ToolFilteredMetricThresholdConfig}. */
export type ToolCallCountConfig = ToolFilteredMetricThresholdConfig;
interface MetricDef {
    /** Grader type string used in eval.yaml. */
    name: string;
    /** Human-readable description. */
    description: string;
    /** Extract the metric value from a trajectory. */
    extract: (input: GraderInput) => number;
    /** Unit label for evidence strings. */
    unit: string;
    /** Whether `max` requires a duration string (e.g. "30s", "2m") instead of a number. */
    acceptsDuration?: boolean;
    /** Whether this metric can be restricted to selected tool names. */
    acceptsTools?: boolean;
    /**
     * Noun describing what a `tools` pattern matches, for config-error messages
     * (e.g. "an empty regex matches every ${noun}"). Defaults to `"tool"`.
     */
    watchNoun?: string;
    /**
     * Count the metric directly from events when a tool filter is configured.
     * Metrics without it fall back to `metrics.toolCallBreakdown`.
     */
    extractFiltered?: (input: GraderInput, watch: (name: string) => boolean) => number;
}
export declare class MetricThresholdGrader implements Grader {
    metadata: GraderMetadata;
    private def;
    constructor(def: MetricDef);
    /** `token-budget max 5000` — the threshold is the only varying config. */
    defaultName(config: Record<string, unknown>): string;
    grade(input: GraderInput): Promise<GraderResult>;
    /** Resolve the metric value, optionally counting only selected tool calls. */
    private resolveMetricValue;
    /** Parse and validate the `max` config value. */
    private resolveMax;
}
/** Create all metric-threshold grader instances. */
export declare function createMetricThresholdGraders(): Grader[];
export {};
//# sourceMappingURL=metric-threshold-grader.d.ts.map