/**
 * Agent measurement core (Spec 25 Phase D / Q2) — shipped so `openlore prove`
 * can run a WITH/WITHOUT pass on the user's own repo and print a personal
 * token-value scorecard. The live agent call is behind an injectable runner so
 * the deterministic parts (JSON parsing, aggregation) are unit-tested without
 * spending API budget; the default runner shells out to `claude -p`, mirroring
 * the Spec 14 benchmark harness exactly.
 */
export type Condition = 'with' | 'without';
export interface Metrics {
    /** input_tokens + cache_creation — what the model processed fresh. */
    freshInputTokens: number;
    /** cache_read_input_tokens — ~10× cheaper amortized reads. */
    cacheReadTokens: number;
    outputTokens: number;
    costUsd: number;
    /** num_turns — round-trips, the most consistent WITH/WITHOUT signal. */
    numTurns: number;
    durationMs: number;
    answer: string;
    correct: boolean;
    error?: string;
}
/** Parse a `claude -p --output-format json` result blob into raw metrics (no scoring). */
export declare function parseAgentJson(raw: string): Omit<Metrics, 'correct'>;
export interface AgentRunInput {
    prompt: string;
    mcpConfigPath: string;
    cwd: string;
    model: string;
    maxBudgetUsd: number;
    /** Appended only on the WITH arm — the shipped orient nudge. */
    systemPrompt?: string;
}
/** Pluggable agent invoker — returns the raw stdout JSON. Default = `claude -p`. */
export type AgentRunner = (input: AgentRunInput) => string;
/** Default runner: shell out to the `claude` CLI (mirrors the Spec 14 harness). */
export declare const claudeRunner: AgentRunner;
/** Write the WITH (openlore navigation) + WITHOUT (empty) strict-mcp configs. */
export declare function writeProveMcpConfigs(workDir: string, cliEntry: string): {
    withPath: string;
    withoutPath: string;
    systemPrompt: string;
};
export declare const median: (xs: number[]) => number;
export interface Cell {
    costUsd: number;
    freshInputTokens: number;
    cacheReadTokens: number;
    numTurns: number;
    durationMs: number;
    correctRate: number;
    runs: number;
}
/** Median-aggregate N runs of one arm into a comparable cell. */
export declare function summarize(runs: Metrics[]): Cell;
//# sourceMappingURL=measure.d.ts.map