AI Coding Agent

A Python-based AI coding agent that uses Google's Gemini 2.0 Flash model to analyze code, execute tasks, and provide direct technical responses. The agent can interact with files and directories to understand and work with codebases.

Features

• File System Operations: List directories, read file contents, and write files
• Code Execution: Run Python scripts with optional arguments
• Intelligent Analysis: Uses Gemini 2.0 Flash to understand code structure and functionality
• Task-Focused: Provides direct, technical responses without conversational fluff
• Function Call Handling: Automatically executes tool calls and manages conversation state

Prerequisites

• Python 3.7+
• Google Gemini API key
• Required Python packages (see Installation)

Installation

1. Clone the repository:

git clone https://github.com/SSSM0602/python-ai-agent.git
cd aiagent

2. Install dependencies:

pip install google-genai python-dotenv

3. Set up your environment variables:

# Create a .env file in the root directory
GEMINI_API_KEY=your_gemini_api_key_here

Project Structure

aiagent/
├── main.py                 # Main application entry point
├── .env                    # Environment variables (create this)
├── functions/              # Function modules
│   ├── get_file_content.py # Read file contents
│   ├── get_files_info.py   # List files and directories
│   ├── run_python.py       # Execute Python files
│   └── write_file.py       # Write/create files
└── calculator/             # Example working directory
    ├── main.py
    └── pkg/
        ├── calculator.py
        └── render.py

Usage

Basic Usage

Run the agent with a question or task:

python main.py "your question or task here"

Verbose Mode

Get detailed information about function calls and token usage:

python main.py "your question" --verbose

Examples

Analyze how code works:

python main.py "how does the calculator render results to the console?"

Examine file structure:

python main.py "what files are in this project?"

Run and analyze code:

python main.py "run the calculator with expression '2 + 3 * 4'"

Available Functions

The agent has access to four main functions:

1. get_files_info: Lists files and directories in the working directory
2. get_file_content: Reads and returns the contents of specified files
3. run_python_file: Executes Python files with optional command-line arguments
4. write_file: Creates or overwrites files with specified content

Configuration

Working Directory

The agent operates within a specified working directory (default: ./calculator). Modify the working_directory variable in main.py to change this:

working_directory = "./your_project_directory"

System Prompt

The agent's behavior is controlled by the system prompt. The current configuration makes it:

• Provide direct, factual responses
• Avoid conversational language
• Focus on technical analysis
• Complete tasks efficiently

Function Schemas

Each function requires a schema definition that describes its parameters and return types. These are imported from the respective function modules:

• schema_get_files_info
• schema_get_file_content
• schema_run_python
• schema_write_file

How It Works

1. User Input: You provide a question or task via command line
2. AI Planning: Gemini analyzes the request and determines what functions to call
3. Function Execution: The agent executes necessary file operations or code runs
4. Response Generation: Based on the gathered information, provides a direct answer
5. Conversation Management: Maintains context across multiple function calls until task completion

Error Handling

The application includes comprehensive error handling:

• Function Call Errors: Caught and reported, with execution continuing
• API Errors: Network and API issues are handled gracefully
• Iteration Limits: Prevents infinite loops with a 20-iteration maximum
• Validation Errors: Function responses are validated and formatted correctly

Limitations

• Working Directory Scope: All file operations are relative to the configured working directory
• Python Execution Only: Can only execute Python files directly
• API Rate Limits: Subject to Google Gemini API rate limiting
• No Persistent State: No memory between separate program runs

Troubleshooting

Common Issues

"Need a cmd line argument": Provide a question or task as the first argument

"validation error for FunctionResponse": Function returned non-dictionary data - this is handled automatically in the current version

"Unknown function" errors: Check that all function modules are properly imported and defined in func_defs

API authentication errors: Verify your GEMINI_API_KEY is correctly set in the .env file

Debug Mode

Use the --verbose flag to see:

• Detailed function call information
• Token usage statistics
• Step-by-step execution flow

Contributing

To add new functions:

1. Create a new function module in the functions/ directory
2. Define both the function and its schema
3. Import and add to func_defs in main.py
4. Update the available_functions tool declaration