/**
 * darwin run <agent> "task"
 *
 * Runs an agent, optionally evaluates with critic, triggers Darwin evolution.
 */
import type { ProviderConfig } from '../providers/types.js';
import type { AgentDefinition, PromptVersion } from '../types.js';
import type { ABTest, EvolutionConfigOverride } from '../types.js';
export interface RunFlags {
    agentName: string;
    task: string;
    taskType: string;
    noEvolve: boolean;
    noCritic: boolean;
    model?: string;
    path?: string;
    verbose: boolean;
    /** Provider override: anthropic-api, openai, ollama */
    provider?: ProviderConfig['type'];
    /** Base URL for OpenAI-compatible / Ollama endpoints */
    baseUrl?: string;
    /** Advanced evolution-config toggles for this run (--gepa/--coverage/…) */
    evolutionOverride: EvolutionConfigOverride;
}
export declare function parseRunArgs(args: string[]): RunFlags;
/**
 * Which prompt version should THIS run use? (v0.14.0 — exported for tests.)
 *
 * - Active A/B test AND evolution enabled → round-robin the arm with fewer
 *   runs (the pre-v0.14 behaviour).
 * - Otherwise → the agent's ACTIVE version. Two real bugs lived here
 *   (cross-model review, v0.14.0):
 *     1. Without a running test the CLI always ran the STATIC v1 prompt —
 *        a promoted winner (v3 active in state) never actually ran, and its
 *        runs were recorded against v1, corrupting the version stats.
 *     2. `darwin evolve --disable` did not stop the A/B routing, so a
 *        disabled agent kept sending ~50% of its traffic through the
 *        challenger — with the test counters frozen, forever.
 */
export declare function pickRunVersion(abTest: ABTest | null, evolutionEnabled: boolean, activeVersion: string): string;
/**
 * Are both arms of an A/B test backed by a resolvable prompt? `v1` always
 * resolves (the agent definition is its floor); any other label needs a
 * stored prompt. (v0.14.0 — exported for tests.)
 *
 * Why this exists (R4 review, P0): when an open test references an arm whose
 * stored prompt has vanished (corrupted/foreign state), routing would fall
 * back to v1 and RELABEL the run as v1 — which `handleABTest` then counts
 * for whichever arm happens to be labeled v1, or for neither, so the test
 * either decides on wrong data or starves forever. An unresolvable test is
 * dead on arrival; the caller clears it and runs the active version.
 */
export declare function abTestArmsResolvable(abTest: ABTest, storedVersions: ReadonlyArray<Pick<PromptVersion, 'version'>>): boolean;
/**
 * Resolve the prompt TEXT + the version label the experiment must be recorded
 * under. (v0.14.0 — exported for tests; R3 review closed two gaps here.)
 *
 * Invariant: the returned `version` always names the prompt that actually
 * RUNS. When state requests a label with no stored prompt (corrupted/foreign
 * state), we fall back to the static v1 prompt AND relabel the run as v1 —
 * recording a static-v1 output under the missing label would corrupt that
 * label's stats and could decide an A/B test with wrong data.
 *
 * v1 itself prefers the STORED v1 prompt when one exists (the seeded copy),
 * falling back to the agent definition. `darwin eval` resolves versions the
 * same way, so a built-in prompt edited after seeding no longer makes live
 * runs and offline evals score different texts under the same label.
 */
export declare function resolveRunPrompt(agent: AgentDefinition, storedVersions: ReadonlyArray<Pick<PromptVersion, 'version' | 'promptText'>>, requestedVersion: string): {
    version: string;
    promptText: string;
    missingStored: boolean;
};
export declare function runCommand(args: string[]): Promise<void>;
//# sourceMappingURL=run.d.ts.map