import { spawnHeadlessClaude, type ParsedTranscript, type SpawnResult } from '@vibe-agent-toolkit/utils';
export interface RunExecutorInput {
    evalId: string;
    task: string;
    subjectStagedDir: string;
    /** `<workspacesRoot>/<id>` when the eval declares input `files`. */
    workspaceDir?: string;
    pluginDirs: string[];
    env: NodeJS.ProcessEnv;
    model?: string;
    maxTurns: number;
    maxBudgetUsd: number;
    timeoutMs: number;
    stallMs?: number;
    /** 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;
}
export interface ExecutorOutcome {
    /** Full captured stdout (stream-json), held IN MEMORY only — never written to disk. */
    transcript: string;
    parsed: ParsedTranscript;
    spawnResult: SpawnResult;
    /** true = the executor exited non-zero WITHIN budget — a valid eval failure, not a harness break. */
    cleanFailure: boolean;
}
/**
 * Run ONE eval's executor: a blind `claude -p` spawn that performs `task`
 * against the staged subject (issue #145). The executor's entire stream-json
 * stdout is accumulated in memory (never written to a skill-writable disk
 * path — the "stop discarding stdout" fix) and echoed to `onProgress` chunk
 * by chunk for visibility.
 *
 * Encodes the R1 exit bifurcation at the executor boundary:
 * - A spawn error (the promise rejects), or our own watchdog firing
 *   (`timedOut`/`stalled`), is HARNESS breakage → thrown as
 *   {@link InternalHarnessError} (exit 1).
 * - A mid-stream rate-limit that cuts the run off before it completes (status
 *   !== 0) is NOT a harness failure but also not gradeable → thrown as
 *   {@link RateLimitSignal} so the caller's pool can back off and retry this
 *   eval. A rate-limit event that the run still completed past (status 0) is
 *   kept and returned normally.
 * - Anything else — including a clean non-zero exit within budget — is a
 *   valid eval OUTCOME, not a harness break: returned with `cleanFailure`
 *   reflecting whether the executor exited non-zero. Never laundered into a
 *   thrown error; the grader decides whether the transcript passes.
 */
export declare function runExecutorForEval(input: RunExecutorInput): Promise<ExecutorOutcome>;
//# sourceMappingURL=eval-executor.d.ts.map