@justinechang39/maki/dist/core/prompt-cache.js

AI-powered CLI agent for file operations, CSV manipulation, todo management, and web content fetching using OpenRouter

import { SELECTED_MODEL } from './config.js';
import { SYSTEM_PROMPT } from './system-prompt.js';
/**
 * Creates cache-optimized system message based on model type
 *
 * For different models:
 * - Gemini: Implicit caching (automatic) - no special handling needed
 * - OpenAI: Automatic caching - no special handling needed
 * - Anthropic: Explicit cache_control required for large prompts
 */
export function createCacheOptimizedSystemMessage() {
    // For Anthropic models, use explicit cache control on the system prompt
    // since it's large and static across conversations
    if (SELECTED_MODEL.includes('anthropic')) {
        return {
            role: 'system',
            content: [
                {
                    type: 'text',
                    text: SYSTEM_PROMPT,
                    cache_control: { type: 'ephemeral' }
                }
            ]
        };
    }
    // For Gemini and OpenAI, use standard format - they handle caching automatically
    return {
        role: 'system',
        content: SYSTEM_PROMPT
    };
}
/**
 * Creates cache-optimized user message for large content
 * Useful for RAG data, large CSV files, or extensive context
 */
export function createCacheOptimizedUserMessage(content, largeContext) {
    if (!largeContext) {
        return {
            role: 'user',
            content: content
        };
    }
    // For Anthropic models with large context, cache the context
    if (SELECTED_MODEL.includes('anthropic')) {
        return {
            role: 'user',
            content: [
                {
                    type: 'text',
                    text: content
                },
                {
                    type: 'text',
                    text: largeContext,
                    cache_control: { type: 'ephemeral' }
                }
            ]
        };
    }
    // For other models, combine normally - they handle implicit caching
    return {
        role: 'user',
        content: `${content}\n\n${largeContext}`
    };
}
/**
 * Gets cache optimization info for current model
 */
export function getCacheInfo() {
    if (SELECTED_MODEL.includes('gemini')) {
        return {
            type: 'implicit',
            description: 'Automatic caching (0.25x cost for cached tokens)',
            minTokens: SELECTED_MODEL.includes('2.5-pro') ? 2048 : 1028
        };
    }
    if (SELECTED_MODEL.includes('openai') || SELECTED_MODEL.includes('gpt')) {
        return {
            type: 'implicit',
            description: 'Automatic caching (0.5x-0.75x cost for cached tokens)',
            minTokens: 1024
        };
    }
    if (SELECTED_MODEL.includes('anthropic')) {
        return {
            type: 'explicit',
            description: 'Manual cache control (0.1x reads, 1.25x writes)',
            minTokens: 1024,
            maxBreakpoints: 4,
            ttl: '5 minutes'
        };
    }
    return {
        type: 'none',
        description: 'No caching support detected'
    };
}