import type { AgentTool } from "@kenkaiiii/gg-agent";
import { type ContextLimits } from "../context-limits.js";
/** Clamp a tool's description to the MCP description budget (eager or deferred). */
export declare function clampMcpToolDescription(tool: AgentTool, limits?: ContextLimits): AgentTool;
/**
 * Holds MCP tools OUT of the per-turn request payload until the model asks
 * for them via `tool_search`.
 *
 * Measured: injecting every MCP tool schema
 * eagerly cost ~33KB (~8.3k tokens) per cache-miss turn with just two MCP
 * servers connected — 56% of all billed input tokens in a 6-turn session.
 * Deferring keeps the tool prefix small and byte-stable; promotion is a
 * one-time cache break paid only when a capability is actually needed.
 *
 * Descriptions are clamped at `add` time (server-controlled content): a
 * hostile or bloated MCP server must not be able to inflate every
 * tool_search result — and, once promoted, every request's tool schema.
 */
export declare class DeferredToolCatalog {
    private readonly limits;
    private byName;
    constructor(limits?: ContextLimits);
    add(tools: AgentTool[]): void;
    /** Remove tools (e.g. when their MCP server is reloaded/removed). */
    removeWhere(predicate: (name: string) => boolean): void;
    get size(): number;
    /** Names still waiting in the catalog (for "no match" tool output). */
    names(): string[];
    /**
     * Rank catalog tools against a free-text capability query.
     * Word-overlap scoring: name hits weigh 3×, description hits 1×.
     */
    search(query: string, limit?: number): AgentTool[];
    /** Remove the named tools from the catalog and return them for activation. */
    promote(names: string[]): AgentTool[];
}
//# sourceMappingURL=deferred-catalog.d.ts.map