Devlooper is a program synthesis agent designed to autonomously generate and debug code by iterating through testing and fixing cycles. It targets developers and researchers interested in automated code generation and self-correcting AI agents, offering a novel approach to building complex software components with minimal human intervention.

How It Works

Devlooper extends the "smol developer" concept by integrating a sandboxed execution environment. The agent repeatedly generates code, runs tests within this sandbox, and uses the test results (stdout/stderr) to prompt an LLM for diagnosis and a fix plan. This plan can include code edits, package installations, or command execution, allowing for incremental image construction and robust error correction.

Quick Start & Requirements

• Install: pip install modal
• Prerequisites: Modal account, OpenAI API key (configured as a Modal secret named openai-secret).
• Usage: modal run src.main --prompt="your prompt" --template="[rust|react|python]"
• Output: Written to output/ by default.
• Docs: https://modal.com/docs

Highlighted Details

• Autonomous debugging loop: Iteratively fixes code and environment based on test failures.
• Environment templates: Supports Rust, React+Jest, and Python, extensible to other containerizable frameworks.
• Modal Sandbox: Leverages Modal's Sandbox primitive for isolated, incremental environment builds.
• Separate diagnosis step: Enhances LLM accuracy by isolating error diagnosis from code generation.

Maintenance & Community

• Project is a proof of concept with active contributions welcomed for new templates and features.
• Future directions include user feedback integration, embedding-based code retrieval, and support for open-source LLMs.

Licensing & Compatibility

• License: Not explicitly stated in the README.
• Compatibility: Requires Modal and OpenAI accounts. Primarily designed for use with Modal's infrastructure.

Limitations & Caveats

The project is described as a proof of concept with potential for infinite loops and limited LLM accuracy. Support for LLMs beyond OpenAI's models is a future direction.