import type { Category, Mode } from '../../config/categories.js';
import type { OptimizationContext, OptimizationStrategy } from '../types.js';
import type { Intent } from '../../context/types.js';
import { type AcceptedExample, type PromptShape } from '../groundingContext.js';
import { type CurationResult } from '../curator.js';
export declare abstract class BaseStrategy implements OptimizationStrategy {
    abstract readonly name: string;
    abstract readonly category: Category;
    protected llmClient: import("../../llm/client.js").LLMClient;
    abstract buildSystemPrompt(context: OptimizationContext, platformInstructions?: string): string;
    protected getModeInstructions(mode: Mode): string;
    /**
     * Pass A: intent-specific overlay folded into the system prompt so strategies
     * actually branch on WHAT the user is trying to do, not just WHICH platform.
     */
    protected getIntentOverlay(intent: Intent | undefined): string;
    /**
     * Pass B: shape the system prompt to the downstream LLM's capabilities.
     * Compact budgets keep only the base prompt plus a one-line mode rule —
     * small models choke on multi-KB system prompts. (The intent overlay is
     * appended after shaping by the callers, so it always survives.)
     */
    protected applyShape(systemPrompt: string, shape: PromptShape, mode?: Mode): string;
    protected getBaseSystemPrompt(): string;
    renderLastSystemPrompt(context: OptimizationContext, platformInstructions?: string): string;
    optimize(prompt: string, context: OptimizationContext, platformHints?: string[], platformInstructions?: string): Promise<string>;
    /** Exposed for the engine to pull into the trace + response. */
    lastCuration?: CurationResult;
    protected runCurator(args: {
        context: OptimizationContext;
        systemPrompt: string;
        originalPrompt: string;
        acceptedExamples: AcceptedExample[];
        platformHints?: string[];
        platformInstructions?: string;
        shape: PromptShape;
    }): CurationResult;
}