v0.1.0 — first prototype release

First tagged release of terra-pilot: natural-language Terraform / Terragrunt generation grounded in your own repository. Pre-release: a working prototype. See STATUS.md for what is verified and what is not.

Highlights

• compose turns a plain-English request into inputs.hcl / module code that follows your repo's conventions (terragrunt, plain Terraform modules, flat Terraform), using any OpenAI-compatible LLM.
• Refuses to guess: when your repo has no module for the request it prints a module scaffold instead of letting an LLM invent inputs (--freeform opts out).
• Zero-dependency install (pip install .); optional extras for tree-sitter, embeddings, reranking, local models and PySpark data pipelines.
• Deterministic edit engine for mutations, a write-time safety gate, and a validation gate (fmt, HCL sanity, tflint, tfsec, validate/plan/checkov).

Verified

• 294 offline checks across 9 suites, plus opt-in real-tool and real-AWS-provider end-to-end checks.
• Reuse-vs-write decisions on 200 real GitHub Terraform repos: 0% false reuse on the compose path.
• Live runs against DeepSeek; real terragrunt, terraform, tflint, tfsec and checkov.

Not yet

Generation-quality evaluation across many requests, non-AWS providers, near-duplicate removal in the data pipeline, and a fine-tuned model.