import { BaseOptimizer, type BaseOptimizerDeps } from "../baseOptimizer.js";
import type { Input } from "../grading/types.js";
import { type ProposeMutationArgs } from "../mutator.js";
import type { BaseOptimizerConfig } from "../optimizer.js";
import { type OptimizeMutationOperation, type OptimizeMutationPreview } from "../sourceMutator.js";
import { type OptimizeTargetSet } from "../targets.js";
import type { MutationProposal, OptimizeResult } from "../types.js";
/**
 * Optional injection points, so the optimizer can be unit-tested without an LLM
 * or real file edits. The registry constructs it with none — `new ExampleOptimizer(config)`.
 */
export type ExampleDeps = BaseOptimizerDeps & {
    propose?: (args: ProposeMutationArgs) => Promise<MutationProposal>;
    preview?: (targetSet: OptimizeTargetSet, operations: OptimizeMutationOperation[]) => OptimizeMutationPreview;
};
/**
 * The smallest useful optimizer — a copy-paste template for writing your own.
 *
 * Every optimizer extends {@link BaseOptimizer} and implements
 * {@link optimizeTargets}. By the time it runs, the base class has already
 * resolved the agent file and discovered its `optimize` targets; your job is to
 * search for better target values and return the best candidate.
 *
 * This one runs a single round: score the agent as-is, ask the built-in mutator
 * for one new set of values, score those, and keep the candidate if it beats the
 * baseline. The real optimizers (greedy, gepa) loop and search more cleverly, but
 * they all follow the same shape — fork → apply → evaluate → compare → report → return.
 *
 * Protected helpers available from BaseOptimizer:
 *   - `this.scoreFiles(source, files, inputs)` — fork + apply + grade, returns a Scorecard
 *   - `this.finishPointwise(source, candidates, trainChampion, attempts, startedAt)` —
 *     pick the writeback champion (by validation when configured), write it back,
 *     and build the OptimizeResult with train/validation objectives + breakdown
 *   - `this.reporter` — progress output (silent unless the CLI sets verbosity)
 *   - `this.config` — graders, runId, writeback, mutatorModel, …
 *
 * Register it (see registry.ts):
 *   registerOptimizer("example", (config) => new ExampleOptimizer(config));
 */
export declare class ExampleOptimizer extends BaseOptimizer {
    private readonly exampleDeps;
    readonly name = "example";
    constructor(config: BaseOptimizerConfig, exampleDeps?: ExampleDeps);
    protected optimizeTargets(source: OptimizeTargetSet, inputs: Input[]): Promise<OptimizeResult>;
    /** Apply a candidate file set into a fresh workspace, run + grade it. The
     *  `targetSet` reflects this candidate's targets (the baseline's for the
     *  baseline candidate, `outcome.preview.targetSet` for an accepted mutation),
     *  so `finalTargets` in the reporter accurately describes the champion. */
    private makeCandidate;
}