import type { InputBreakdown } from "../eval/grading/gradeBreakdown.js";
import type { OptimizeMutationOperation } from "./sourceMutator.js";
export type OptimizeDecision = "baseline" | "accepted" | "rejected" | "validation-failed";
export type MutationProposal = {
    operations: OptimizeMutationOperation[];
    rationale: string;
    /** What the proposing model call cost, when the proposer can tell. */
    costUsd?: number;
};
export type ValidationResult = {
    ok: true;
} | {
    ok: false;
    reason: string;
};
export type IterationResult = {
    iter: number;
    decision: OptimizeDecision;
    agentDir?: string;
    mutationPath?: string;
    evalRunDir?: string;
    verdictPath?: string;
    /** Human-readable reason for the decision: validation diagnostics for
     *  "validation-failed", otherwise the proposal rationale. */
    detail?: string;
    /** The training objective the decision rested on: the full training set
     *  for the baseline and accepted candidates, the minibatch for a rejection. */
    objective?: number;
    /** The held-out objective, scored for the baseline and every accepted
     *  candidate when a validation set is configured. */
    validationObjective?: number;
};
export type OptimizeResult = {
    runId: string;
    runDir: string;
    championIter: number | "baseline";
    /** The champion's full Agency file set, keyed by relative path. */
    championFiles: Record<string, string>;
    acceptedCount: number;
    rejectedCount: number;
    validationFailedCount: number;
    iterations: IterationResult[];
    /** Champion's gate-aware train objective (`scorecard.gatedObjective()`,
     *  i.e. raw objective or 0 if any `mustPass` gate failed) — set by
     *  pointwise optimizers. Matches the score optimizers use to compare
     *  candidates. */
    trainObjective?: number;
    /** Baseline candidate's gate-aware train objective, so consumers can see
     *  the improvement without re-running. Computed identically to
     *  `trainObjective` so the two are directly comparable. */
    baselineObjective?: number;
    /** Champion's validation objective, when a validation set was used (Phase 3). */
    validationObjective?: number;
    /** Per-input grade breakdown for the champion — the reward-hacking lens. */
    championBreakdown?: InputBreakdown[];
    /** What the run spent, in USD, split by who spent it. */
    cost?: OptimizeCost;
};
export type OptimizeCost = {
    /** The agent under test, across every candidate and input. */
    agentUsd: number;
    /** The judges, across every grading pass (including validation scoring). */
    gradingUsd: number;
    /** The model that proposed mutations. */
    mutatorUsd: number;
    totalUsd: number;
};