FixFlow : GitHub Issues, Resolved Autonomously.

FixFlow: Project Submission Details

Inspiration

Manual debugging is the most time-consuming part of software development. Developers often spend hours tracing issue reports, navigating unfamiliar codebases, and identifying the exact line where a bug occurs. We wanted to build a tool that could automate this "detective work," allowing developers to focus on building features rather than hunting for typos or compatibility breaks.

What it does

FixFlow is an autonomous agent that takes a GitHub issue URL and returns a full root cause analysis along with a pull-request-ready fix. It uses a structured 5-step pipeline to parse the bug report, map out the repository to find suspect files, perform deep chain-of-thought analysis, and generate a precise unified diff for the solution.

How we built it

LLM Orchestration: We utilized GLM 5.1 by Z.ai to handle high-reasoning tasks like codebase mapping and root cause tracing.
Backend: Developed with Python 3.11 using the PyGithub library to interact with the GitHub REST API and difflib for generating standard patch files.
Frontend: Built a dark-themed, glassmorphism-style dashboard using Streamlit that supports real-time streaming of the agent's thought process.
Multi-Step Agent: We designed a custom orchestrator in agent.py that manages state between issue parsing, file ranking, and fix generation to ensure the model stays within context limits.

Challenges we ran into

Context Window Management: Feeding an entire repository into an LLM is impossible for large projects. We overcame this by building a Codebase Mapping step that pre-filters and ranks the top 5-10 most relevant files before the deep analysis begins.
Hallucination Prevention: Early versions occasionally "invented" file paths. We implemented a validation layer that checks every LLM-suggested path against the actual repository tree before proceeding.
Rate Limiting: Managing burst API requests during the multi-step reasoning process required implementing strategic sleep intervals to avoid rate limits.

Accomplishments that we're proud of

Precision: Successfully pinpointing deep logic errors, such as Pydantic v2 compatibility breaks, and generating minimal, non-breaking code changes.
Interactive UI: Creating a seamless experience where users can see the agent "thinking" and provide feedback to refine the fix.
Autonomous Reasoning: Achieving a pipeline where the agent cites specific line numbers and function calls with high confidence.

What we learned

Reasoning Density: We learned that GLM 5.1 excels at "mapping" structures—it can understand how different modules interact even when only provided with a high-level file tree.
Agentic Workflows: We discovered that breaking a complex task like "fixing a bug" into five discrete, verifiable steps significantly improves the reliability of the output compared to a single-shot prompt.

What's next for FixFlow: GitHub Issues, Resolved Autonomously.

The next step for FixFlow is moving beyond just suggesting fixes to Autonomous Testing—automatically generating and running test cases to verify the fix before the user even sees it. We also plan to integrate FixFlow directly as a GitHub Action, so it can comment on new issues with a proposed solution the moment they are opened.