Architecture Introduction

CatInCup adopts an extremely decoupled design, with core code located in the src/ directory:

src/
├── agent/          # Agent core brain: contains Harness Engineering scheduling logic and state machine
├── loader/         # Data access layer: database connections and local memory RAG loader
├── models/         # Data structure layer: object-oriented definitions, including LLM model base classes and subtask (Task) abstractions
├── plugins/        # Tool and execution layer: MCP integration, native plugin system, local file whitelist verification, and Docker sandbox scheduling
├── sensor/         # Asynchronous sensing layer: environmental status monitoring (such as Feishu callbacks, RSS feeds, etc.)
└── utils/          # Infrastructure layer: common utility functions, logging, and retry mechanisms

Architecture Topology

graph TD
    Agent[Agent Brain<br>Harness Engineering] --> Loader[Data Access Layer<br>RAG Loader]
    Agent --> Models[Data Structure Layer<br>LLM & Task]
    Agent --> Plugins[Execution Layer<br>MCP & Sandbox]
    Sensor[Asynchronous Sensing Layer<br>Feishu Callback/RSS] -.Asynchronous Event Injection.-> Agent

Roadmap

1. Core Engine Optimization: Polish Harness Engineering, improve KV Cache hit rate, and further reduce Token consumption.

2. Multi-Model and High Availability: Deeply adapt to mainstream open source/closed source models, add API failure automatic transfer and hybrid scheduling.

3. Multi-Modal Fusion: Access visual/audio and other multi-modal large model interfaces to expand operation boundaries.

4. Distributed Execution: Support cross-node Docker container scheduling to handle high concurrency and heavy task computing.

5. Ecosystem and Memory System: Build a standardized plugin/Skill market; develop long-term structured memory retrieval trees that go beyond traditional RAG.