Prompt Copilot CLI

A lightweight terminal coding agent built on LangChain and LangGraph. The project has been refactored to use the standard agent runtime, SQLite-backed persistence, and a small, explicit tool set instead of maintaining a custom agent loop.

What this project does

Prompt Copilot CLI lets you:

• open a workspace directory and chat with an agent in the terminal
• use a small set of coding tools for reading, searching, editing, and running Python commands
• persist conversation state and durable memories using LangGraph SQLite storage
• connect to an OpenAI-compatible model endpoint such as OpenAI, Azure OpenAI, OpenRouter, or local OpenAI-compatible gateways

Architecture

CLI
└── AgentRuntime
    ├── ChatOpenAI
    ├── create_agent()
    ├── SummarizationMiddleware
    ├── SqliteSaver
    ├── SqliteStore
    └── LangMem tools

The runtime is intentionally minimal:

• create_agent() provides the model/tool loop
• SummarizationMiddleware compresses long conversations
• SqliteSaver persists the current conversation thread
• SqliteStore + LangMem handle durable project memory
• tools remain focused on the practical coding workflow

Install

From the project root:

python -m pip install -e .

Or install from PyPI:

python -m pip install -U prompt-copilot-cli

Initial configuration

The first run creates a config file in:

• Windows: %USERPROFILE%\.prompt-copilot\config.json
• Linux/macOS: ~/.prompt-copilot/config.json

Example:

{
  "model": "gpt-4o-mini",
  "base_url": "https://api.openai.com/v1",
  "api_key": "YOUR_API_KEY",
  "temperature": 0.2,
  "timeout": 120,
  "memory": {
    "enabled": true,
    "max_recent_memories": 5
  },
  "context": {
    "summary_trigger_tokens": 12000,
    "keep_messages": 20
  }
}

You can point base_url to any OpenAI-compatible endpoint, and the project will use api_key from that config file.

Usage

Start interactive mode:

prompt-copilot -d D:\project

Run a single task and exit:

prompt-copilot -d D:\project -t "Inspect the repo, fix the failing tests, and run the relevant checks."

From the CLI, use these commands:

• /exit — quit the program
• /clear — reset the current workspace thread
• /memory — display persistent memories for this workspace

The prompt supports multiline input. Press Enter for a newline and Ctrl+Enter to submit.

Available tools

The agent exposes a compact tool set:

• read_file
• list_files
• search_code
• write_file
• edit_file
• delete_file
• execute_python_script
• execute_command

These are intentionally scoped to file work, code search, editing, and local execution.

Memory behavior

The runtime uses two persistence layers:

Short-term memory

SqliteSaver stores the live thread state for the current workspace. If you restart the CLI, the same workspace can continue the same conversation.

Long-term memory

SqliteStore stores durable project memory and LangMem exposes memory management tools so the agent can remember useful project conventions or user preferences.

Memory data is stored under:

~/.prompt-copilot/memory.db

Project structure

.
├── main.py
├── copilot/
│   ├── __init__.py
│   ├── agent.py
│   ├── cli.py
│   ├── config.py
│   ├── memory.py
│   └── tools.py
├── tests/
│   ├── test_config.py
│   └── test_tools.py
├── README.md
├── README.zh-CN.md
├── pyproject.toml
└── requirements.txt

Testing

The project is currently verified with:

python -m pytest -q

Fresh verification in this environment shows:

3 passed in 0.29s

The CLI entry point also starts successfully:

python main.py --help

This confirms the refactored program is runnable in the current environment.

License

Apache License 2.0