/**
 * LlmSummarizationEngine — default ContextEngine backed by an LLM.
 *
 * Implements the Hermes §7.6 SUMMARY_PREFIX compression pattern:
 *   1. Keep the first {@link KEEP_FIRST} messages verbatim (system context).
 *   2. Keep the last {@link KEEP_LAST} messages verbatim (recent context).
 *   3. Summarize the middle window via a single {@link LlmExecutor.auxiliary} call
 *      using the `'compression'` role (cheap haiku-class model).
 *
 * Threshold: compress when `promptTokens / contextBudget >= 0.75`.
 *
 * @module memory/context-engines/llm-summarizer
 * @task T9304
 * @epic T9261 T-LLM-CRED-CENTRALIZATION
 * @see packages/contracts/src/memory/context-engine.ts
 */
import type { TransportMessage } from '@cleocode/contracts/llm/normalized-response.js';
import type { CompressedContext, ContextEngine } from '@cleocode/contracts/memory/context-engine.js';
/** Minimum estimated tokens before compression is even attempted. */
export declare const MIN_SUMMARY_TOKENS = 2000;
/** Compression threshold ratio — compress when prompt fills >= 75% of budget. */
export declare const SUMMARY_RATIO = 0.2;
/** Upper bound on summary output tokens to keep the compressed history small. */
export declare const SUMMARY_TOKENS_CEILING = 12000;
/**
 * Default {@link ContextEngine} implementation using LLM-backed summarization.
 *
 * Uses the `'compression'` role (resolved via `getLlmExecutor('compression')`)
 * so the provider + model + credential come from `config.llm.roles.compression`
 * → `config.llm.default` → `config.llm.daemon` → implicit haiku fallback.
 *
 * @example
 * ```ts
 * const engine = new LlmSummarizationEngine();
 * const factory = new DefaultLlmExecutorFactory({ contextEngine: engine });
 * const executor = await factory.createForRole('orchestrator');
 * ```
 */
export declare class LlmSummarizationEngine implements ContextEngine {
    /**
     * Returns `true` when the prompt is large enough to warrant compression.
     *
     * Compression triggers when BOTH conditions hold:
     *   - `promptTokens >= MIN_SUMMARY_TOKENS`
     *   - `promptTokens / contextBudget >= 0.75`
     *
     * @param promptTokens - Estimated token count of the current session history.
     * @param contextBudget - Effective context-window budget for the bound model.
     */
    shouldCompress(promptTokens: number, contextBudget: number): boolean;
    /**
     * Compresses the session history using LLM summarization.
     *
     * Preserves the first {@link KEEP_FIRST} and last {@link KEEP_LAST} messages
     * verbatim. The middle window is summarized via an auxiliary LLM call with the
     * {@link SUMMARY_PREFIX} header, producing a single synthetic `assistant`
     * message that replaces the middle turns.
     *
     * @param messages - Current read-only conversation history snapshot.
     * @param focusTopic - Optional topic hint appended to the summarization prompt.
     * @returns The compressed context with before/after token estimates.
     */
    compress(messages: readonly TransportMessage[], focusTopic?: string): Promise<CompressedContext>;
}
//# sourceMappingURL=llm-summarizer.d.ts.map