Cloud-Gym

Scalable Training Data Generation for Infrastructure-as-Code Repair via Environment Inversion.

Cloud-Gym generates (broken_config, error_message, fix) training pairs for IaC repair by applying environment inversion — taking working Terraform, CloudFormation, and OpenTofu configs and systematically breaking them using a defined fault taxonomy. It includes a benchmark (188 entries across 8 error categories) and fine-tuned models that run entirely on CPU.

stackfix: AI-Powered IaC Repair

The stackfix CLI tool validates and repairs broken IaC files using fine-tuned local models. No API keys, no cloud costs, no data leaves your machine.

Install

pip install stackfix

Download a Model

# Recommended: 3B Q4 (1.8 GB, 87% pass@1)
python -c "
from huggingface_hub import hf_hub_download
hf_hub_download('Tetsuto/iac-repair-3b-gguf', 'iac-repair-3b-q4.gguf', local_dir='.')
"

Usage

# Check files for errors
stackfix check main.tf template.yaml

# Repair a broken file (show diff)
stackfix repair main.tf --backend gguf --model iac-repair-3b-q4.gguf

# Repair and apply fix in place
stackfix repair main.tf --apply --backend gguf --model iac-repair-3b-q4.gguf

# Explain errors in plain language
stackfix discuss main.tf --backend gguf --model iac-repair-3b-q4.gguf

# Pipe mode (stdin/stdout)
cat broken.tf | stackfix repair - --backend gguf --model iac-repair-3b-q4.gguf > fixed.tf

# Check all changed IaC files in git
stackfix git-diff --backend gguf --model iac-repair-3b-q4.gguf

Models

All models are fine-tuned Qwen2.5-Coder with LoRA, exported to GGUF. They run on any CPU (Linux, macOS, Windows).

Backends

CI/CD Integration

Add to your GitHub Actions workflow to catch IaC errors on every PR:

- name: Check IaC
  run: |
    pip install stackfix
    python -c "
    from huggingface_hub import hf_hub_download
    hf_hub_download('Tetsuto/iac-repair-3b-gguf', 'iac-repair-3b-q4.gguf', local_dir='.')
    "
    stackfix check **/*.tf **/*.yaml

See examples/USE_CASES.md for more deployment scenarios (pre-commit hooks, Lambda, pipeline integration).

Pre-Commit Hook

# .pre-commit-config.yaml
repos:
  - repo: local
    hooks:
      - id: stackfix
        name: stackfix
        entry: stackfix pre-commit --backend gguf --model iac-repair-3b-q4.gguf
        language: python
        types_or: [terraform, yaml]
        additional_dependencies: ['stackfix[gguf]']

Benchmark

188 entries across 8 error categories, 3 difficulty levels, and 2 formats (Terraform + CloudFormation).

Results Summary

Fine-tuning a 0.5B model outperforms a 26B base model by 80x.

Training Data Generation

Cloud-Gym generates training data via environment inversion:

1. Collect working IaC configs from GitHub, Terraform Registry, AWS samples
2. Break them systematically using a fault taxonomy (28+ fault types across 8 categories)
3. Validate broken configs to capture real error messages
4. Pair (broken + errors) with the original working config as the gold fix

# Generate training data
cloud-gym taxonomy          # View fault types
python scripts/scrape.py    # Collect gold configs
cloud-gym invert            # Generate broken variants
cloud-gym export            # Export training pairs

Project Structure

cloudgym/
  taxonomy/     Fault type definitions (28+ types, 8 categories)
  scraper/      Gold config collection
  validator/    IaC validation wrappers (terraform, cfn-lint)
  inverter/     Fault injection engines
  generator/    Training data pipeline
  benchmark/    Evaluation harness
  fixer/        stackfix CLI tool + model backends
scripts/        Training, evaluation, and export scripts
examples/       Broken IaC examples + use case docs

Supported Formats

• Terraform (.tf) — validated with terraform validate
• CloudFormation (.yaml, .yml, .json) — validated with cfn-lint
• OpenTofu (.tf) — same as Terraform

License

MIT