Project Overview

A terminal-based AI coding assistant in TypeScript. Intended as a learning project to discover how AI agents work internally and to explore OpenAI’s Responses API.

Features

• AI-Powered (LLM) Coding Assistant in the Unix Command-line interface(CLI).
• Tool Calling: The model invokes tools (Read, Write, Bash) to write files and execute commands.
• Conversation History: The agent maintains context by storing user messages.
• Interactive Modes:
  • Prompt mode: Run with -p flag then “your prompt” for single tasks.
  • REPL mode: Interactive chat session for ongoing conversation.

Project Structure

src/
├── index.ts          # Entry point
├── core/
│   ├── agent.ts      # AI client & agent loop
│   └── config.ts     # Configuration management
└── tools/
    ├── definitions.ts   # Tool schemas (definitions)
    └── handlers.ts       # Tool execution logic

What I Learned

• Setting up an AI agent - Initializing an AI client with API keys and base URLs.
• Standard API for AI agents - Using the OpenAI Responses API (not Chat API) for agentic workflows.
• Tool advertising - Defining tools using OpenAI’s FunctionTool format with name, description, and JSON schema parameters.
• Tool implementation - Creating handlers (Read, Write, Bash) that execute actual filesystem/shell operations.
• Agent loop - Building a conversation loop that sends history to the model, handles tool calls by executing handlers and appending results, and continues until the model returns a final message.
• History management - Maintaining conversation context by appending user messages, assistant responses, and tool outputs.

Future Improvements

• Switch to WebSocket mode for the Responses API instead of HTTP/REST
• Introduce Sessions & resume capability
• Improve UI