import type { ArvoSemanticVersion } from 'arvo-core';
import z from 'zod';
import type { AgentInternalTool } from '../AgentTool/types.js';
import type { PromiseAble } from '../types.js';
import type { AgentContextBuilder, AgentOutputBuilder, AgentServiceContract, AnyArvoOrchestratorContract } from './types.js';
/**
 * Creates a Zod schema validator for base64-encoded data URLs with MIME type restrictions.
 *
 * Validates that strings match the data URL format (data:mime/type;base64,encodedContent)
 * and optionally restricts to specific MIME types.
 *
 * @example
 * ```typescript
 * const imageSchema = dataUrlString(['image/jpeg', 'image/png']);
 * imageSchema.parse('data:image/jpeg;base64,/9j/4AAQ...'); // Valid
 * imageSchema.parse('data:application/pdf;base64,...');     // Throws - wrong MIME type
 * ```
 */
export declare const dataUrlString: (allowedMimeTypes: string[]) => z.ZodEffects<z.ZodString, string, string>;
/**
 * Default schemas and builders for common agent patterns.
 *
 * Provides ready-to-use implementations for typical agent configurations,
 * reducing boilerplate while remaining customizable. Use these defaults
 * to get started quickly, then customize as needed for specific use cases.
 */
export declare const AgentDefaults: {
    /**
     * Default initialization schema accepting a simple text message.
     *
     * Use for basic conversational agents that only need text input.
     */
    readonly INIT_SCHEMA: z.ZodObject<{
        message: z.ZodString;
    }, "strip", z.ZodTypeAny, {
        message: string;
    }, {
        message: string;
    }>;
    /**
     * Multimodal initialization schema supporting text, images, and PDFs.
     *
     * Accepts base64-encoded images (JPEG, PNG, GIF, WebP) and PDF documents
     * alongside the text message. Media content is visible to the LLM only once
     * and should be extracted into conversation history via internal tools
     * if needed for later reference.
     *
     * @remarks
     * The media masking optimization automatically replaces viewed media with
     * placeholder text in subsequent LLM calls to reduce token consumption.
     */
    readonly INIT_MULTIMODAL_SCHEMA: z.ZodObject<{
        message: z.ZodString;
        imageBase64: z.ZodOptional<z.ZodArray<z.ZodEffects<z.ZodString, string, string>, "many">>;
        pdfBase64: z.ZodOptional<z.ZodArray<z.ZodEffects<z.ZodString, string, string>, "many">>;
    }, "strip", z.ZodTypeAny, {
        message: string;
        imageBase64?: string[] | undefined;
        pdfBase64?: string[] | undefined;
    }, {
        message: string;
        imageBase64?: string[] | undefined;
        pdfBase64?: string[] | undefined;
    }>;
    /**
     * Default completion schema outputting a simple text response.
     *
     * Use when agents produce conversational text rather than structured data.
     */
    readonly COMPLETE_SCHEMA: z.ZodObject<{
        response: z.ZodString;
    }, "strip", z.ZodTypeAny, {
        response: string;
    }, {
        response: string;
    }>;
    /**
     * Default context builder transforming initialization events into LLM context.
     *
     * Converts the message field into a user message and processes any attached
     * media (images, PDFs) into the conversation history (Assumes the media
     * content to be base64 - ArvoAgent always assumes this internally). Accepts an optional
     * system prompt builder function that receives initialization parameters
     * and available tools for dynamic prompt construction.
     *
     * @param systemPromptBuilder - Optional function returning the system prompt.
     *                              Receives input event, tools catalog, and contract reference.
     *
     * @returns Context builder function compatible with agent handler configuration
     *
     * @example
     * ```typescript
     * context: AgentDefaults.CONTEXT_BUILDER(({ tools, input }) =>
     *   `You are a helpful agent. Use ${tools.tools.dateTool.name} for dates.`
     * )
     * ```
     */
    readonly CONTEXT_BUILDER: <T extends AnyArvoOrchestratorContract, V extends ArvoSemanticVersion, TServiceContract extends Record<string, AgentServiceContract>, TTools extends Record<string, AgentInternalTool>>(systemPromptBuilder?: (param: Parameters<AgentContextBuilder<T, V, TServiceContract, TTools>>[0]) => PromiseAble<string | {
        system: string;
        enabledTools?: Record<string, boolean>;
    }>) => AgentContextBuilder<T, V, TServiceContract, TTools>;
    /**
     * Default output builder validating agent responses against contract schemas.
     *
     * Handles both text and JSON output modes:
     * - **Text mode:** wraps the LLM's raw response string in the default complete schema structure.
     * - **JSON mode:** attempts to recover a valid JSON object from the LLM's output using a
     *   two-stage strategy:
     *   1. Use `parsedContent` if the LLM integration already parsed it successfully.
     *   2. If `parsedContent` is null (parse failed), run `jsonrepair` on the raw `content` to
     *      recover from common LLM output issues: markdown code fences, prose mixed with JSON,
     *      and structurally malformed JSON. Non-object results (arrays, primitives) are rejected.
     *   3. If recovery fails, return a natural language error that feeds into the self-correction
     *      loop so the LLM can retry with explicit formatting instructions.
     *
     * @remarks
     * Returns `{ data }` on success or `{ error }` on failure. Errors trigger the agent's
     * self-correction loop, allowing the LLM to fix its output before the cycle limit is hit.
     */
    readonly OUTPUT_BUILDER: AgentOutputBuilder;
};
//# sourceMappingURL=AgentDefaults.d.ts.map