A Go-based CLI agent that fetches open Erdős problems from the community database, sends them to an LLM for proof or disproof attempts, and securely saves the generated solutions.

📺 Overview

✨ Features

• Automated Scraping: Fetches the latest open problems from the community database (problems.yaml) and scrapes detailed problem descriptions directly from erdosproblems.com.
• LLM Integration: Uses the any-llm library to interface seamlessly with various LLM providers (OpenAI, Anthropic, Gemini, Groq, Ollama) and orchestrate proof generation.
• Secure Credentials: Prompts securely for API keys and stores them via OS keychain encryption for seamless future runs.
• Continuous Solve Loop: Allows for selecting and solving multiple problems in a continuous execution loop without needing to restart the CLI.
• Robust Error Handling: Retries automatically on transient errors (timeouts, rate limits, network issues) with exponential backoff, and supports long-running LLM requests (up to 120 minutes).
• Solution Archiving: Automatically saves generated proofs as markdown files in a designated solns/ directory.

🚀 Getting Started

Prerequisites

• Go 1.21 or higher installed on your system.

Build and Run

1. Clone the repository:

   git clone https://github.com/demirbey05/erdos-agent.git
   cd erdos-agent

2. Build the executable binary:

   go build -o erdos-agent .

3. Run the agent:

   ./erdos-agent

🛠 Usage & Workflow

1. Configuration: On the first run, the agent will prompt you to select an LLM provider (e.g., openai, anthropic, gemini, groq, ollama), a specific model, and your API key (hidden input). This configuration is securely encrypted and saved to ~/.erdos-agent/config.enc.
2. Problem Selection: The agent retrieves and displays a list of currently open Erdős problems. You can then:
   • Enter a comma-separated list of problem numbers (e.g., 1, 42, 108).
   • Type all to attempt all open problems.
   • Type prize to attempt only problems that have an associated cash prize.
3. Execution: The agent will automatically fetch the full problem details, query the configured LLM with a specialized prompt to generate an unconditional proof or disproof, and wait for the response.
4. Results: Generated solutions are saved in the solns/ directory. Each markdown file includes the attempt details and is easily readable.

🧠 Architecture Overview

The codebase is organized as follows:

• main.go: The entry point that orchestrates the CLI flow, fetches metadata, and handles the continuous execution loop.
• internal/scraper/: Retrieves and parses YAML metadata from GitHub and scrapes HTML descriptions from the web.
• internal/solver/: Interfaces with the LLM via gollm, constructs the mathematical prompts, and handles saving the solutions.
• internal/keystore/: Securely manages API keys using AES-256-GCM encryption.
• internal/models/: Defines core domain data types like Problem and Status.

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.