AI Code Assistant (Gemini Function-Calling)

A small command-line AI coding agent that uses Google's Gemini models with function calling to safely inspect and manipulate a sandboxed project directory. The agent can:

• List files and directories within a configured working directory
• Read file contents (with size limits)
• Run Python files with optional arguments
• Write or overwrite files (creating directories as needed)

All tool operations are constrained to the configured working directory to prevent access outside the sandbox.

Demo

• 
• Fallback: Watch the full MP4

Requirements

• Python 3.13+
• A Google Gemini API key

Python dependencies are managed via pyproject.toml and installed using your preferred tool (e.g., uv, pip).

Installation

Using uv (recommended):

# From the project root
uv sync

Configuration

Environment variables are loaded via .env (using python-dotenv). Create a .env file in the project root containing:

GEMINI_API_KEY=your_api_key_here

Other runtime configuration lives in config.py:

• WORKING_DIR: sandboxed directory for all tool calls (default: ./calculator)
• MAX_CHARS: max characters returned when reading files (default: 10000)
• MAX_ITERS: safety limit for agent turns (default: 20)

Usage

From the project root:

python main.py "Your request here" [--verbose]

Examples:

python main.py "List files in the project"
python main.py "Open calculator/main.py and explain the bug" --verbose
python main.py "Run calculator/tests.py"

• --verbose prints token usage, tool invocations, and intermediate outputs.

How it works

Entry point: main.py

• Loads .env
• Builds a conversation with your user prompt
• Calls Gemini (gemini-2.0-flash-001) with a system instruction (prompts.py) and tool declarations
• Handles model tool calls via dispatcher/call_function.py
• Continues until the model returns a final text response or MAX_ITERS is reached

System instruction

prompts.py defines the agent’s capabilities and constraints for tool usage.

Tools (Functions)

Declared in dispatcher/call_function.py and implemented under functions/:

• get_files_info(directory='.')

  • Lists files in a directory within WORKING_DIR, returning name, size, and is_dir.
  • Enforced sandbox: prevents escaping the working directory.

• get_file_content(file_path)

  • Reads a file within WORKING_DIR up to 10,000 characters; truncates if larger.

• run_python_file(file_path, args=[])

  • Executes a Python script inside WORKING_DIR with optional args.
  • Captures stdout/stderr; reports non-zero exit codes.

• write_file(file_path, content)

  • Writes text to a file inside WORKING_DIR; creates parent dirs as needed.

Development

• Adjust WORKING_DIR in config.py to target a different sandbox.
• Extend tools by adding new functions in functions/ and wiring their schemas into dispatcher/call_function.py.
• Keep I/O constrained to WORKING_DIR to maintain safety.