/**
 * RuleBasedTruncationEngine — deterministic ContextEngine with no LLM dependency.
 *
 * Implements the {@link ContextEngine} contract via token-budget-aware FIFO
 * dropping: the oldest user/assistant pairs are discarded first, and system
 * messages (role `'tool'` treated as structural context) are never dropped.
 *
 * ## Algorithm
 *
 * 1. Split messages into three buckets:
 *    - **Leading system messages** — all leading `tool`-role messages
 *      (preserved unconditionally).
 *    - **Middle window** — the conversation body eligible for dropping.
 *    - **Tail messages** — the last {@link KEEP_TAIL} messages (preserved
 *      unconditionally to maintain recent context).
 * 2. Estimate tokens for all three buckets.
 * 3. If `total tokens <= targetBudget`, return messages unchanged.
 * 4. Otherwise, drop oldest middle messages (FIFO, whole pairs where possible)
 *    until the estimated token count fits within the budget.
 *
 * The engine is **side-effect-free** and **deterministic** — identical inputs
 * always produce identical outputs. It never makes network calls.
 *
 * @module llm/context-engines/rule-based-truncation
 * @task T9312
 * @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';
/** Compression triggers when `promptTokens / contextBudget >= this ratio`. */
export declare const TRUNCATION_RATIO = 0.75;
/** Minimum prompt tokens before compression is considered. */
export declare const MIN_TRUNCATION_TOKENS = 500;
/**
 * Number of messages at the tail of the history to preserve unconditionally.
 *
 * Keeping recent messages verbatim prevents the model from losing immediate
 * task context even after aggressive truncation of older history.
 */
export declare const KEEP_TAIL = 4;
/**
 * Deterministic, non-LLM {@link ContextEngine} that truncates history by
 * dropping the oldest eligible messages until the token estimate fits the
 * budget.
 *
 * @example
 * ```ts
 * const engine = new RuleBasedTruncationEngine();
 * const factory = new DefaultLlmExecutorFactory({ contextEngine: engine });
 * const executor = await factory.createForRole('orchestrator');
 * ```
 */
export declare class RuleBasedTruncationEngine implements ContextEngine {
    /**
     * Returns `true` when the prompt is large enough to warrant truncation.
     *
     * Truncation triggers when BOTH conditions hold:
     *   - `promptTokens >= MIN_TRUNCATION_TOKENS`
     *   - `promptTokens / contextBudget >= TRUNCATION_RATIO`
     *
     * @param promptTokens - Estimated token count of the current session history.
     * @param contextBudget - Effective context-window budget for the bound model.
     * @returns Whether truncation should be applied before the next model turn.
     */
    shouldCompress(promptTokens: number, contextBudget: number): boolean;
    /**
     * Truncates the session history by dropping the oldest non-system messages.
     *
     * The target budget is `contextBudget * (1 - TRUNCATION_RATIO)` worth of
     * headroom — i.e. after truncation the history should consume at most
     * `TRUNCATION_RATIO - 0.1` of the context window. In practice the method
     * drops messages FIFO from the middle window until the estimated token
     * count is at or below `targetTokens`.
     *
     * This method is intentionally synchronous-compatible (returns a resolved
     * promise) to satisfy the async contract without incurring any I/O.
     *
     * @param messages - Current read-only conversation history snapshot.
     * @param _focusTopic - Unused — included for interface compatibility.
     * @returns Resolved promise with the truncated context and token metrics.
     */
    compress(messages: readonly TransportMessage[], _focusTopic?: string): Promise<CompressedContext>;
}
//# sourceMappingURL=rule-based-truncation.d.ts.map