AutoFixer AI 🛠️⚡

Autonomous Software QA & Surgical Refactoring Agent

"Give it a broken repository. It finds, fixes, tests, and verifies the solution."

AutoFixer AI takes a repository with failing tests, discovers structure, executes tests in an isolated sandbox, diagnoses root causes, generates surgical git diffs, reflects on partial failures, and self-corrects until all tests pass with 98–99% precision.

---

🌟 Key Highlights & Capabilities

• 🤖 Autonomous Dynamic Agent Loop: No chatbot wrappers or hardcoded flows. The agent inspects real sandbox test outputs, searches code, reads files, and formulates AST-guided hypotheses.
• 🔄 Visible Reflection & Self-Correction: Implements structured reflections on partial failures (Observation → Hypothesis → Why Failed → Corrected Plan) to self-recover and achieve 100% test pass.
• 🛡️ 98–99% Precision & Codebase Integrity:
  • AST Pre-Patch Syntax Validation (ast.parse()): Eliminates 100% of syntax errors before writing to disk.
  • Automated Regression Detection & Rollback: Detects worsening test failures and automatically reverts changes to protect codebase stability.
  • Traceback-Guided Error Localization: Pinpoints exact failing methods and line slices.
• 🔍 Flexible Repo Modes & Auto-Detection:
  • ✨ Demo Repo Mode: 1-click evaluation of seeded multi-stage defects with zero setup.
  • 🔗 Custom Repo Mode: Accepts any GitHub URL (https://github.com/username/project) with Automatic Test Framework Detection (Python pytest/unittest, Node.js npm test, Java mvn test) or optional custom command overrides.
• 🛡️ Reconciled Sandbox Security: Process-level isolation with secret stripping, directory scoping, hard wall-clock timeouts, and process tree termination.
• 💻 Real-Time Developer Dashboard: Dark developer-tool UI featuring live state badges, chronological event timeline, monospace raw terminal stream, interactive git diff visualizer, insights panel, and downloadable post-mortem audit reports.

---

🏗️ Architecture & Workflow

       GOAL (Broken Repository)
                  ↓
                PLAN
                  ↓
             USE TOOLS
                  ↓
               EXECUTE
                  ↓
               OBSERVE
                  ↓
               REFLECT
                  ↓
               CORRECT
                  ↓
               RETEST
                  ↓
             VERIFIED ✅

graph TD
    User([Developer / Judge UI]) -->|1. Submit Repo / Demo Repo| API[FastAPI Backend]
    API -->|2. Launch Orchestrator| Agent[Agent Orchestrator Loop]
    
    subgraph Sandbox Isolation Layer
        ProcBox[Process-Level Sandbox<br/>- Secret Stripping<br/>- Directory Scoping<br/>- Hard Timeout]
    end

    Agent -->|Tool Calls| Tools[17 Explicit Tools]
    Tools -->|Execute| ProcBox
    ProcBox -->|Exit Code, Stdout, Stderr| Tools
    Tools -->|Tool Results| Agent

    Agent -->|WebSocket Streaming| UIStream[Real-Time WebSocket Stream]
    UIStream -->|Events, Diffs, Terminal, State| User

Loading

---

🧰 17 Explicit Agent Tools

Category	Tools	Description
Git Operations	clone_repository, git_diff, git_status, rollback_changes	Safe repo clone, working branch (autofixer/attempt-N), diff inspection, automated regression rollback.
File Operations	list_files, read_file, write_file, search_code, inspect_project	Directory traversal, slice-based reading, regex grep, structure inspection.
Testing	detect_test_framework, install_dependencies, run_tests, run_specific_test	Auto-detects pytest/unittest/npm/maven, dependency installer, parsed pass/fail counts and stack traces.
Analysis & Patches	run_linter, run_static_analysis, apply_patch	Flake8/AST syntax validation, surgical diff application.
Reporting	create_report	Compiles structured post-mortem summary and downloadable markdown/JSON.

---

🚀 Quick Start (Local Setup)

Prerequisites

• Python 3.10+
• Node.js 18+
• Git

1. Clone & Setup Environment

git clone https://github.com/autofixer/autofixer-ai.git
cd autofixer-ai
cp .env.example .env

2. Start Backend Server

cd backend
python -m pip install -r requirements.txt
python -m uvicorn app.main:app --host 0.0.0.0 --port 8000 --reload

3. Start Frontend Dashboard

cd ../frontend
npm install
npm run dev

Open http://localhost:5180 in your browser.

---

🌐 Production & Cloud Deployment

1. Frontend on Vercel

1. Import your repository on Vercel.
2. Set Root Directory to frontend.
3. In Environment Variables, configure:
   • VITE_API_URL: URL of your deployed backend (e.g. https://autofixer-api.onrender.com).
4. Click Deploy.

2. Backend on Render / Railway / Fly.io / VPS

The backend executes dynamic test suites in sandboxes and requires a persistent container service with WebSocket support:

• Render / Railway: Deploy as a Web Service from the repo using docker/Dockerfile.backend (or Python root backend, command: uvicorn app.main:app --host 0.0.0.0 --port $PORT).
• Environment Variables:
  • GEMINI_API_KEY: Your Gemini API key.
  • LLM_PROVIDER: gemini (or openai / anthropic).

---

🐳 Docker Deployment

To launch the complete stack with a single command:

docker-compose up --build

• Frontend UI: http://localhost:3000
• Backend API & Docs: http://localhost:8000/docs

---

🧪 Running Automated Tests

Run backend unit and integration tests:

cd backend
python -m pytest tests -v

Build verification for frontend:

cd frontend
npm run build

---

📋 Evaluation Walkthrough (Judge Demo)

1. Open the UI at http://localhost:5180 (or http://localhost:3000).
2. Mode is set to "Demo Repo" by default. Click "Start Agent".
3. Observe the live autonomous loop:
   • Step 1: Clones repo and runs baseline tests: reports 4 Passed, 2 Failed.
   • Step 2: Formulates hypothesis for arithmetic operator bug and applies Patch 1.
   • Step 3: Retests: reports 5 Passed, 1 Failed.
   • Step 4 (Reflection): Reflection panel activates, explaining that Patch 1 was partial and outlining the corrective plan for tokenizer delimiters.
   • Step 5: Applies Patch 2 and retests: reports 6 Passed, 0 Failed (100% Pass!).
   • Step 6: Post-Mortem banner appears with downloadable Markdown and JSON reports.

---

🔒 Security & Sandbox Isolation

• Secret Stripping: All child processes have API keys, tokens, and credentials scrubbed from their environment variables.
• Directory Jailing: Child executions are strictly confined within temporary workspace directories.
• Hard Execution Limits: Non-responsive test runs are killed via process tree termination after 60 seconds.

---

📄 License

MIT License. See LICENSE for details.