GDBuddy (1st Place - Austin 8VC/Meta Llama Stack Hacakthon)

A FastAPI server that integrates with a VS Code extension to provide AI-assisted debugging for C code using GDB and an LLM. This solution streams live trace output and later serves a summary of the debugging session—all without having to wait for the full session to complete.

Overview

This project implements a debugging workflow that combines several technologies:

• GDB & Python Agent:
  A Python script (originally in gdb_debugger_agent.py) launches GDB with the Machine Interface (MI) to run your C executable and capture its output.

• LLM Integration:
  An integrated language model (via Llama Stack, for example) analyzes the GDB output, determines the next debugging step, and provides the appropriate GDB command.

• FastAPI Streaming Server:
  The FastAPI server (in this project) provides two endpoints:

  • /debug_target (POST):
    Launches a debugging session and streams the trace output (i.e. everything printed before a delimiter).
  • /get_summary (GET):
    Once the session ends, returns the summary—i.e. the content between two delimiters.

• VS Code Extension Integration:
  A VS Code extension calls these endpoints to receive real-time feedback and summaries, thereby assisting you in debugging C code quickly.

Features

• Real-Time Streaming:
  Uses HTTP chunked transfer encoding with FastAPI’s StreamingResponse to stream trace logs as they are produced by the agent.

• Summary Extraction:
  Separates detailed trace output (live debugging logs) from a post-execution summary using a custom delimiter.

• Seamless Integration:
  Designed to work in conjunction with a VS Code extension that sends debugging requests and displays streamed responses.

Architecture

1. Agent Execution:
   The agent’s main routine is executed in a background thread. Standard output (sys.stdout) is redirected to a custom writer that both collects the output in a global buffer and pushes every text chunk to an asyncio queue.

2. Streaming Endpoint:
   The /debug_target endpoint starts the debugging session. An async generator yields output chunks until the first delimiter is detected, sending only the trace.

3. Summary Endpoint:
   The /get_summary endpoint waits until the agent has completed and then parses the global output. It extracts and returns only the portion between the first and second occurrence of the delimiter.

Requirements

• Python 3.8+
• GDB

The following are in the requirements.txt:

llama-stack-client
llama-stack
pygdbmi
rich
fastapi
uvicorn
pydantic

Installation

1. Clone this repository:

   git clone <repository_url>
   cd <repository_directory>

2. Install Python dependencies from the agent_src directory:

   pip install -r requirements.txt

Usage

Running the Server

Run the FastAPI server from the agent_src directory:

python3 debugging_agent.py

This will start the debugging backend on port 8000.

Endpoints

1. POST /debug_target

Starts a debugging session and streams live trace output until the delimiter appears.

Example using curl:

curl --no-buffer -X POST "http://localhost:8000/debug_target" \
     -H "Content-Type: application/json" \
     -d '{"executable": "test_executables/test_1", "bug_description": "This is segfaulting, help"}'

The trace output (everything before the delimiter) will stream back to the client.

2. GET /get_summary

Returns the summary of the debugging session; that is, the content between the first and second occurrence of the delimiter.

Example using curl:

curl "http://localhost:8000/get_summary"

VS Code Extension Integration

Our VS Code extension can call these endpoints:

• Use /debug_target for real-time trace information.
• Use /get_summary to display a concise summary after the session.

Customization

• Delimiters:
  The agent output is split using the delimiter string ========================================. Modify this string in both the agent code and the endpoints if needed.
• Output Redirection:
  The custom QueueWriter class ensures that all printed output is captured and pushed to both a global buffer and an asyncio queue.

Contributing

Contributions are welcome! Feel free to open issues or submit pull requests to improve:

• The integration between the GDB agent and FastAPI.
• Additional parsing and formatting of output.
• VS Code extension compatibility and user experience.

License

MIT