import { spawnHeadlessClaude } from '@vibe-agent-toolkit/utils';
import { type EvalFragment } from './eval-fragment.js';
import type { ToolExpectations } from './eval-inputs.js';
export interface RunGraderInput {
    evalId: string;
    /** The FULL captured transcript (stream-json) for this eval's executor run. */
    transcript: string;
    expectations: string[];
    expectedOutput?: string;
    /** Path to the vendored skill-creator grader rubric (references/grader.md). */
    rubricPath: string;
    /**
     * The eval's declared tool expectations (issue #145 Phase T). Optional — when
     * absent, `buildGraderPrompt` emits no tool-verdict instruction and the
     * returned fragment omits `tool`. See `grader-prompt.ts`'s
     * `BuildGraderPromptOptions.toolExpectations`.
     */
    toolExpectations?: ToolExpectations;
    /** Name → invocation hint for declared executables, a recognition aid alongside `toolExpectations`. */
    declaredExecutables?: Array<{
        name: string;
        howInvoked: string;
        kind: string;
    }>;
    /**
     * VAT-ONLY directory the grader fragment is written to — deliberately
     * OUTSIDE the skill-writable sandbox (never the executor's staged/workspace
     * dir) so untrusted skill code from Task 7's executor run cannot forge or
     * delete the fragment this run relies on.
     */
    graderOutDir: string;
    graderModel: string;
    /** Per-run integrity nonce the grader must echo back in the fragment. */
    nonce: string;
    maxTurns: number;
    maxBudgetUsd: number;
    timeoutMs: number;
    stallMs?: number;
    /** Grader auth env — no skill secrets are needed beyond auth. */
    env: NodeJS.ProcessEnv;
    /** Injectable seam for tests; defaults to the real {@link spawnHeadlessClaude}. */
    spawn?: typeof spawnHeadlessClaude;
    /** Called with each stdout chunk as it streams (for progress echo). */
    onProgress?: (chunk: string) => void;
    /**
     * Reports this grader session's `total_cost_usd` (parsed from its stream-json
     * transcript's terminal result; `undefined` when the transcript carried none,
     * e.g. a mock spawn). Called once after a successful grader spawn so the run can
     * aggregate total spend across all executor+grader sessions (adopter follow-up).
     */
    costSink?: (totalCostUsd: number | undefined) => void;
}
export declare function runGraderForEval(input: RunGraderInput): Promise<EvalFragment>;
//# sourceMappingURL=eval-grader.d.ts.map