poliac

A Terraform linter, delivered over LSP, whose recommendations are grounded in production metrics instead of hardcoded rules.

Give poliac a Terraform file, a CSV of production/runtime metrics, and a short statement of what you're trying to achieve — it asks an LLM to recommend changes to the infrastructure code, with a rationale, and with every claim checked against the real data before it's ever shown to you.

main.tf        (your infrastructure code)   ┐
metrics.csv    (production/runtime data)    ├──▶  LLM  ──▶  validator  ──▶  diagnostics,
specification.md    (your stated intent)         ┘                                hover, quick-fix

What it supports today

• Infrastructure: libvirt_domain (vcpu/memory) and aws_instance (instance type, checked against a real catalog so the model can't invent one). Adding a resource type is a config file, not a code change.
• LLM providers: Gemini, Claude (Anthropic), ChatGPT (OpenAI), and DeepSeek — pick per run, no code changes.
• Interfaces: a CLI (poliac inspect, poliac analyse) and an LSP server (poliac --stdio) with diagnostics, hover, quick-fix, and code lens. No packaged editor extension yet — any LSP-capable client configured to launch poliac --stdio works.

Install

python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev,gemini]"          # swap/add: anthropic, openai, deepseek

Set the API key for whichever provider you're using (never in a config file):

export GEMINI_API_KEY=...        # or ANTHROPIC_API_KEY / OPENAI_API_KEY / DEEPSEEK_API_KEY

Quickstart

poliac inspect examples/libvirt/main.tf --metrics-csv examples/libvirt/analysis.csv

Parses your Terraform and metrics, prints the resource inventory. Free — no LLM call.

cd examples/libvirt
poliac analyse main.tf --dry-run     # see the exact prompt that would be sent, still free
poliac analyse main.tf               # the real thing

Prints validated recommendations as JSON, each with a rationale and the exact metrics cells it cites, plus an overall assessment. Anything the model claimed that didn't check out against your actual files is silently dropped from the output and logged to .poliac/rejected.jsonl instead.

Every setting above can be overridden per-run instead of coming from .poliac.yaml — useful for trying a different contract, metrics snapshot, or LLM without touching the config file:

poliac analyse main.tf \
  --metrics-csv examples/libvirt/analysis.csv \
  --contract examples/libvirt/specification-cost-cutting.md \
  --llm-provider anthropic --llm-model claude-opus-5

Configuration

Optional .poliac.yaml at your project root — everything here can also be passed as a CLI flag for a one-off run:

metrics:
  path: ./analysis.csv
contract:
  path: ./specification.md      # what are you trying to achieve?
llm:
  provider: gemini          # gemini | anthropic | openai | deepseek
  model: gemini-3.6-flash

Run poliac analyse --help / poliac inspect --help for the full flag list.

As an LSP server

poliac --stdio

Point an editor at this for *.tf files (namespaced poliac.*, source "poliac", so it runs alongside terraform-ls). Trigger analysis via the poliac.analyse command; results show up as diagnostics, hover text, and quick-fixes.