python-harness 0.0.12

An agentic codebase evaluation and evolution tool for Python projects.

Python Harness

An agentic codebase evaluation and evolution tool for Python projects.

python-harness is designed to be a universal standard tool—just like pytest or ruff—but instead of just checking syntax or running tests, it evaluates the architecture, readability, and governance of your codebase using both static analysis and LLMs (DeepSeek/OpenAI).

Features

1. Hard Evaluation (First Fence): Enforces strict rules using ruff, mypy, and ty. Evaluates Cyclomatic Complexity (CC) and Maintainability Index (MI) via radon.
2. Governance QC (Second Fence): Checks if the changes violate core project governance or attempt to bypass the evaluation rules themselves.
3. Soft Evaluation (Third Fence):
   • Calculates architecture metrics like Fan-out (coupling).
   • Generates a holistic package understanding using LLMs.
   • Performs "Blind QA": Randomly samples functions/classes and tests the LLM's ability to understand them without context.
4. Actionable Output: Synthesizes the evaluation into a final Pass/Fail verdict with exactly 3 concrete, actionable refactoring suggestions.

Installation

You can install python-harness using uv or pip:

uv pip install python-harness

Configuration

python-harness requires an LLM to perform its soft evaluation. Create a .env file in the root of your project:

LLM_API_KEY=your_api_key_here
LLM_BASE_URL=https://api.deepseek.com/v1
LLM_MODEL_NAME=deepseek-reasoner
LLM_MINI_MODEL_NAME=deepseek-chat

(Note: If you don't provide an API key, the harness will safely run in Mock mode).

Usage

1. Measure

To evaluate your codebase, run the measure command in your project directory:

harness measure .

This will run the full 3-fence evaluation and output a report with a final verdict and top 3 improvement suggestions.

2. Refine (Evolution Loop - WIP)

The refine command is an Agentic Edit-Test-Improve loop. It takes the suggestions generated by measure, automatically creates branches (variants), applies the changes, runs the tests (pytest), and picks the best variant.

harness refine . --steps 1 --max-retries 3

License

MIT License. See LICENSE for more details.

A harness toolkit for Python projects