Ghost: Autonomous Test Generation Agent

Ghost is a local-first autonomous agent designed to automate the unit testing lifecycle for Python applications. Functioning as a background daemon, it monitors filesystem events, generates context-aware pytest(or any other framework) suites, and autonomously resolves execution errors through a continuous feedback loop.

It is engineered for privacy and latency, supporting Ollama and LM Studio for fully local execution, alongside Groq, OpenAI, and Anthropic for cloud-based inference.

Core Capabilities

Context-Aware Code Analysis

Unlike generic AI coding assistants, Ghost parses the project's Abstract Syntax Tree (AST) to construct a dependency map. This ensures generated tests utilize correct imports, class hierarchies, and function signatures, significantly reducing runtime errors.

Autonomous Error Resolution (Self-Healing)

Ghost implements a closed-loop control system for test execution. When a generated test fails:

1. Capture: stdout, stderr, and stack traces are intercepted.
2. Analysis: The error context is analyzed against the source code.
3. Patching: The agent generates and applies a fix (e.g., correcting imports, adjusting mocks).
4. Verification: The test is re-executed to validate the patch.

Logic Verification Protocol ("The Judge")

To prevent regression validation—where tests are modified to match incorrect implementation logic—Ghost employs a secondary verification step. If an AssertionError occurs, a specialized "Judge" agent determines if the discrepancy stems from the test expectation or the source implementation. If the source code is deemed buggy, the agent halts and alerts the developer rather than altering the test.

Provider-Agnostic Architecture

Ghost decouples the agentic logic from the underlying LLM, allowing seamless switching between providers:

• Groq: Recommended for high-frequency watch modes (low latency).
• Ollama: Recommended for air-gapped or privacy-constrained environments.
• OpenAI / Anthropic: Available for complex reasoning tasks.

Installation

Ghost is published to PyPI as ghosttest.

Global Installation (Recommended)

We recommend installing Ghost as a standalone tool using uv to ensure environment isolation.

uv tool install ghosttest

Standard Installation

pip install ghosttest

Quick Start

1. Initialization

Navigate to your project root and initialize the Ghost configuration. This generates a ghost.toml file and performs an initial AST scan of the codebase.

ghost init

2. Environment Configuration

If using cloud providers, export the necessary API keys. Local providers (Ollama/LM Studio) require no configuration.

# Example for Groq
export GROQ_API_KEY=gsk_...

# Example for OpenAI
export OPENAI_API_KEY=sk-...

3. Execution

Start the daemon. Ghost will monitor for file modifications and trigger the generation/healing loop automatically.

ghost watch

Configuration

Ghost is configured via ghost.toml.

[project]
name = "my-application"
language = "python"

[ai]
# Options: groq, ollama, openai, anthropic, lmstudio
provider = "groq"
model = "llama-3.3-70b-versatile"
rate_limit_rpm = 30

[scanner]
# Directories to exclude from context analysis
ignore_dirs = [".venv", "node_modules", "dist", "__pycache__"]
ignore_files = ["setup.py", "conftest.py"]

[tests]
framework = "pytest"
output_dir = "tests"
auto_heal = true
max_heal_attempts = 3
use_judge = true

Command Reference

License

This project is licensed under the MIT License.