🤖 Gemini Coding Agent

This project implements a Python-based interactive coding agent powered by Google's Gemini models (using the newer google-genai SDK).

Adapted from Thorsten Ball's "How to Build an Agent" https://ampcode.com/how-to-build-an-agent

The agent can:

• Understand and respond to natural language prompts.
• Interact with your local file system (read, list, edit files).
• Execute shell commands.
• Run commands within a secure Docker sandbox (no network access, resource limits).
• Find arxiv papers.
• Maintain conversation history.
• Display the token count of the current conversation context in the input prompt.
• Upload and discuss PDF documents using the Gemini File API.

✨ Features

• Interactive Chat: Engage in a conversational manner with the AI.
• File Operations: Ask the agent to read, list, or modify files within the project directory.
• Arxiv Search: Ask the agent to find arxiv papers from cs or stat categories matching keywords.
• Command Execution: Request the agent to run shell commands in the project's context.
• Sandboxed Execution: Safely run potentially risky commands in an isolated Docker container using the run_in_sandbox tool (requires Docker).
• Tool Integration: Leverages Gemini's function calling capabilities to use defined Python tools.
• Context Token Count: Displays the approximate token count for the next API call right in the prompt (e.g., You (123):).
• PDF Context Support: Upload and discuss PDF documents using the Gemini File API.
• Web Search: Query Google and return top results as a JSON list of titles and URLs using the google_search tool. Warning: This tool can be very slow to return results and consumes a lot of tokens. Also web-search-related token use is not shown in the coding-agent's token count.
• URL Browsing: Visit any webpage and extract its visible text using the open_url tool.

🛠️ Setup

1. Clone the repository:

   git clone https://github.com/voxmenthe/coding-agent.git
   cd coding-agent

2. Set up a virtual environment and install dependencies:

   python -m venv <your-env-name>
   source <your-env-name>/bin/activate
   sh project_setup.sh

3. Set up API Key:

   • The agent reads your Gemini API key exclusively from the GEMINI_API_KEY environment variable.
   • Before running, export your key:

     export GEMINI_API_KEY="your_key_here"

4. (Optional) Docker for Sandbox:

   • If you want to use the run_in_sandbox tool, ensure you have Docker installed and running on your system.
   • You might need to pull the base image specified in src/tools.py (e.g., python:3.12-slim) if it's not already available locally:

     docker pull python:3.12-slim

▶️ Running the Agent

Install the package locally and run the CLI from anywhere:

From the root directory of the project

pip install -e .

Run the agent

coding-agent

## 📝 Notes
- The agent operates relative to the directory it was started from.

## 💬 Usage

- Simply type your requests or questions at the `You (<token_count>):` prompt.
- To exit the agent, type `exit`, `quit`, `/exit` or `/q`.
- Ask it to perform tasks like:
    - "Read the file src/tools.py"
    - "List files in the root directory"
    - "Edit README.md and add a section about future plans"
    - "run the command 'ls -l'"
    - "run 'pip list' in the sandbox"
    - "Search Google for 'latest AI news'"
    - "Open https://example.com and extract visible text"
- The number in parentheses indicates the approximate token count of the conversation history that will be sent with your *next* message.

### 📄 Working with PDFs

The agent can include PDF documents in the conversation context, allowing you to discuss and ask questions about their content:

1. **Upload a PDF:** Use the `/upload` command followed by the path to the PDF (relative to the project directory):

/upload path/to/your/document.pdf

2. **Ask Questions:** Once uploaded, you can directly ask questions about the document:

What are the key points in this paper? Summarize the methodology section. Extract all tables from this document.

3. **PDF Ingestion:** Use the `/upload` command to seed an uploaded PDF’s text into the chat context (only once per file):

/upload path/to/your/document.pdf

After ingestion, the PDF is not reprocessed on each query.

4. **Reset Chat:** To clear the entire chat context (including any seeded PDF text) and start a fresh session:

/reset

5. **View Status:** The number of active PDFs is shown in the prompt (e.g., `You (123) [2 files]:`)

### 🔍 Example: Web Search & Browse

```bash
# Launch the agent
coding-agent

# Perform a Google search
🔵 You (0): Search Google for "Python 3.14 new features"
🟢 Agent: [
{"title": "What’s New in Python 3.14", "url": "https://docs.python.org/3/whatsnew/3.14.html"},
{"title": "Python 3.14 Release Notes", "url": "https://www.python.org/downloads/release/python-3140/"}
]

# Browse a webpage
🔵 You (123): Open https://docs.python.org/3/tutorial/
🟢 Agent: The Python Tutorial — Python 3.x ...