SudoDev

SudoDev is an autonomous AI Software Engineer capable of solving complex coding tasks, refactoring code, and debugging issues. It combines a recursive Feedback Loop for self-correction with a Smart Context Search engine to operate effectively on large codebases.

---

System Architecture

The system uses a decoupled architecture with a React-based frontend and a Python agentic backend.

graph TD
    User((User)) -->|HTTPS| UI["Client Layer<br/>(Next.js 16 / React 19)"]
    
    subgraph "Server Side (FastAPI)"
        UI -->|POST /api/run| API[API Gateway]
        UI -->|GET /api/status| API
        API -->|Async| BG[Background Worker]
        BG -->|Start| Agent[("Improved Agent")]
    end
    
    subgraph "Agent Logic"
        Agent -->|1. Analyze| Context["Context Search (AST)"]
        Context -->|AST Parse| FileSys
        
        Agent -->|2. Reproduce| Repro["Reproduction Script"]
        Repro -->|Run| Sandbox
        
        Agent -->|3. Loop| Feedback{"Feedback Loop"}
        Feedback -->|Fix| CodeGen["LLM Code Gen"]
        CodeGen -->|Apply| Sandbox
        
        Feedback -->|Verify| Test[Test Runner]
        Test -->|Result| Feedback
    end
    
    subgraph "Runtime Sandbox (Docker)"
        Sandbox[Containerized Env]
        FileSys[File System]
    end
    
    subgraph "External Intelligence"
        Context -->|Query| Groq["Groq API (Llama 3.3)"]
        CodeGen -->|Prompt| Groq
        Repro -->|Prompt| Groq
    end

Loading

Key Design Decisions

• Recursive Feedback Loop: The agent doesn't just guess; it writes code, runs it, captures errors, and self-corrects iteratively (up to 3 attempts).
• Smart Context Search: Instead of relying solely on vector databases, SudoDev uses AST (Abstract Syntax Tree) parsing combined with LLM-based relevance scoring to identify critical code sections without overhead.
• Sandboxed Execution: Every agent run happens inside an isolated Docker container, ensuring safety and reproducibility.

Key Features

• Dual Mode Operation:
  • SWE-bench Mode: Solves standard benchmark issues for evaluation.
  • GitHub Mode: Connects to any public repository to fix reported issues.
• Deep Debugging: Auto-generates reproduction scripts to confirm bugs before fixing them.
• Context-Aware: Intelligently extracts only relevant classes and functions from large files to fit within LLM context windows.
• Live Observation: Real-time streaming of logs, terminal outputs, and agent thoughts to the UI.

Tech Stack

• Frontend: Next.js 16 (App Router), React 19, Tailwind CSS
• Backend API: Python, FastAPI
• AI Model: Groq (Llama 3.3)
• Code Analysis: Python AST
• Runtime: Docker
• Styling: Tailwind CSS

TO-DO

• Bench Mode testing
• Github pipeline
• Docker Sandbox
• AST context
• Recursive feedback
• IDE-integration
• Auto PRs

Contributing

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

License

Apache-2.0 License. See LICENSE for more information.1