🐾 Compatibillabuddy

An autonomous AI agent that diagnoses AND repairs broken Python ML environments — with self-correction.

Built for the Google DeepMind × Devpost Gemini API Hackathon — Marathon Agent Track

🧠 An autonomous system that plans, executes, verifies, and self-corrects — not a chatbot, not a prompt wrapper.

The Problem

$ pip install torch numpy pandas scikit-learn
# ✅ Successfully installed ...

$ python -c "import torch; print(torch.cuda.is_available())"
# ❌ False — wrong CUDA version for your GPU driver

$ python -c "import sklearn"
# ❌ ImportError: numpy ABI incompatibility

ML environments don't just have version conflicts — they have hardware mismatches, ABI breaks, and runtime incompatibilities that pip, uv, and poetry can't catch. These failures waste hours of developer time across the ML ecosystem.

Compatibillabuddy treats hardware as a first-class dependency and ships an autonomous Gemini-powered agent that can diagnose AND repair these issues without human intervention.

How It Works

flowchart TD
    A["🧑 User: 'Fix my environment'"] --> B["📸 Snapshot Environment"]
    B --> C["🩺 Run Doctor (diagnose)"]
    C --> D{Issues Found?}
    D -- No --> E["✅ Report: Environment Healthy"]
    D -- Yes --> F["🧠 Gemini Plans Fix Order"]
    F --> G["🔧 Execute Fix (pip install/uninstall)"]
    G --> H["🔍 Verify Fix (re-diagnose)"]
    H --> I{Improved?}
    I -- Yes --> J{More Issues?}
    J -- Yes --> G
    J -- No --> K["✅ Report: All Fixed + Changelog"]
    I -- No --> L["⏪ Rollback to Snapshot"]
    L --> M["🔄 Try Alternative Fix"]
    M --> H

The agent doesn't guess — it runs real diagnostics via structured tools, executes real pip commands with safety guardrails, and verifies every fix before moving on. If a fix makes things worse, it rolls back automatically and tries an alternative.

Installation

# Core framework (no AI, works offline)
pip install compatibillabuddy

# With Gemini-powered AI agent
pip install "compatibillabuddy[agent]"

Quick Start

1. Diagnose Your Environment (No AI Required)

# Human-readable console report
compatibuddy doctor

# Machine-readable JSON
compatibuddy doctor --format json

# Save to file
compatibuddy doctor --output report.json --format json

2. Autonomous Repair (Gemini Agent)

# Set your API key
export GEMINI_API_KEY="your-key-here"  # Linux/Mac
$env:GEMINI_API_KEY = "your-key-here"  # PowerShell

# Dry run — see what the agent WOULD do (safe, default)
compatibuddy repair

# Live mode — actually execute fixes
compatibuddy repair --live

# JSON output for programmatic use
compatibuddy repair --format json

# Choose model and retry limits
compatibuddy repair --model gemini-2.5-pro --max-retries 5

3. Interactive Chat

# Start an interactive session with the agent
compatibuddy agent

# You: What GPU do I have?
# Agent: [calls tool_probe_hardware] You have an NVIDIA RTX 4090 with CUDA 12.3...
# You: Run a diagnosis
# Agent: [calls tool_run_doctor] Found 2 issues...
# You: Fix the CUDA mismatch
# Agent: [calls tool_snapshot_environment, tool_run_pip, tool_verify_fix] ...

Architecture

compatibillabuddy/
├── hardware/          # GPU detection, nvidia-smi parsing, platform info
│   ├── probe.py       # probe_hardware() → HardwareProfile
│   └── inspector.py   # inspect_environment() → EnvironmentInventory
├── kb/                # Knowledge base of known-bad combinations
│   ├── engine.py      # TOML rulepack loader + evaluator
│   └── rulepacks/     # Community-extensible rule definitions
│       └── ml_core.toml
├── engine/            # Core diagnosis orchestrator
│   ├── models.py      # Pydantic v2 models (GPU, packages, issues)
│   ├── doctor.py      # diagnose() → DiagnosisResult
│   └── report.py      # Rich console + JSON report formatters
├── agent/             # Gemini-powered autonomous agent
│   ├── config.py      # API key resolution, model selection
│   ├── tools.py       # 9 tools: diagnose, repair, snapshot, rollback
│   └── core.py        # AgentSession: chat loop + auto_repair()
└── cli/               # Typer CLI
    ├── doctor.py      # compatibuddy doctor
    ├── agent.py       # compatibuddy agent (interactive)
    └── repair.py      # compatibuddy repair (autonomous)

Agent Tools

The Gemini agent has access to 9 structured tools — not arbitrary shell access:

Safety Guardrails

The repair agent operates under strict safety constraints:

• Virtual environment isolation — refuses to modify system Python
• Snapshot before every change — full rollback capability
• Dry-run by default — shows what it would do without executing
• Protected package blocklist — never uninstalls pip, setuptools, wheel, or itself
• Operation limit — stops after 10 pip commands per session
• Only pip install/uninstall — no arbitrary shell commands
• Automatic rollback — if a fix introduces new problems, reverts immediately

What Makes This Different

Known Compatibility Rules

The bundled ml_core.toml rulepack covers:

• CUDA mismatches — PyTorch, TensorFlow, JAX vs. detected CUDA version
• NumPy ABI breaks — pandas, scikit-learn, scipy built against incompatible NumPy
• Driver conflicts — GPU driver version vs. framework requirements

Rules are written in TOML and can be extended by the community:

[[rule]]
id = "cuda-torch-mismatch"
description = "PyTorch installed without CUDA support on a CUDA-capable system"
severity = "error"
category = "cuda-mismatch"
fix_suggestion = "pip install torch --index-url https://download.pytorch.org/whl/cu121"

[rule.when]
package_installed = "torch"
gpu_vendor = "nvidia"

Development

git clone https://github.com/jemsbhai/compatibillabuddy.git
cd compatibillabuddy
pip install -e ".[all]"

# Run tests
pytest

# Lint
ruff check src/ tests/
ruff format --check src/ tests/

# Integration tests (requires GEMINI_API_KEY)
pytest tests/integration/ -m integration -v

Tech Stack

• Python 3.10+ (tested on 3.10–3.13)
• Pydantic v2 — structured data models with JSON schema export
• Typer + Rich — CLI with beautiful terminal output
• google-genai SDK — Gemini function calling for the agent
• packaging — PEP 440 version matching
• pytest — 280+ tests with CI on 3 OS × 4 Python versions
• Hatchling — modern Python build backend

License

MIT — see LICENSE.

Author

Muntaser Syed — muntaser@ieee.org