Gemma 4 Developer Agent (gemma4-dev-agent)

An autonomous software engineering agent built for the Google Gemma 4 Developer Agent Competition. This agent features custom query routing, static AST parsing, multi-file dependency refactoring, and automated test-driven healing powered by pytest.

🔗 GitHub Repository: EzioDEVio/gemma4-dev-agent

---

🚀 Key Features

• Intelligent Query Routing: Context-aware classification mapping user requests to specific developer workflows.
• Static AST Parsing: Inspects imports, classes, and function definitions using Python's native ast module.
• Multi-File Dependency Tracking: Manages cross-file code references and updates dependencies automatically across modules.
• Automated Pytest Healing (Test-Driven Repair): Detects failing test assertions, patches the underlying logic, and re-runs test suites until all checks pass successfully.
• Comprehensive Toolset: Includes robust execution wrappers for bash commands, file reading/writing, AST analysis, and test coverage measurement (pytest-cov).

---

📊 Benchmark Results

The agent has been rigorously tested against an advanced evaluation suite covering core software engineering tasks:

Task ID	Description	Status
task_01_refactor	Code refactoring & function signature updates	PASSED
task_02_syntax_fix	Automatic syntax error detection & fixing	PASSED
task_03_multi_file_dependency	Cross-file function renaming & dependency tracking	PASSED
task_04_pytest_healing	Test-driven error patching & test suite validation	PASSED

• Overall Score: 4/4 (100.0%)

---

🛠️ Supported Tools

1. execute_bash - Executes arbitrary bash shell commands on the host system.
2. file_read - Safely reads file contents with error handling.
3. file_write - Creates or overwrites files and directories.
4. repo_ast_parser - Analyzes Python syntax trees for structural overview.
5. run_pytest - Executes test suites with optional code coverage flags.

---

🏃 Quick Start & Evaluation

To run the full evaluation suite and verify the agent's 100% benchmark score locally:

# Set up your virtual environment if not already active
source .venv/bin/activate

# Run the advanced benchmark evaluation
python3 scripts/evaluate.py