context-chef-core

What problem does it solve?

ContextChef helps you stop AI agent context from drifting and overflowing by compiling your system prompt, compressed history, dynamic state, and (optionally) memory and tool handling into a provider-ready payload for each LLM call.

Core Features & Use Cases

• History compression (Janitor): keeps conversations within the model’s context window via token-aware preservation and LLM-based summarization when needed.
• Tool management (Pruner): reduces hallucinated or irrelevant tool calls by filtering tool sets for the current task.
• Dynamic state with Zod (XML injection): prevents state drift by validating user/task state and injecting it efficiently for recency bias.
• Provider-adapter compilation: produces correctly formatted payloads for OpenAI, Anthropic, or Gemini without rewriting prompts per provider.
• Optional memory + VFS offloading: persist key facts across turns and truncate/offload oversized tool results to keep responses usable.

Quick Start

Ask the AI to integrate @context-chef/core into your existing TypeScript/JavaScript agent loop by detecting your provider, installing the package, wiring compile() per turn, and enabling history compression plus dynamic Zod state injection.