RACT

The problem

A 2024 analysis of 623 million commits by GitClear and GitKraken found that AI-assisted code is already rotting codebases: more copy/paste, less refactoring, and a measurable decline in code movement. The tools that write code fastest are not the tools that keep code healthy.

What RACT does differently

RACT is an agentic coding tool that keeps a human in the loop while a small management LM routes work to the right provider — local, cheap frontier, or cloud. Every plan and result is Rooted[T]: it carries the assumption, confidence, and provenance that justify it.

Every generated file carries the Root Knot (_ROOT_KNOT = object()). The knot is both a coder signature and a loop invariant: if the recursion loop produces an artifact without it, the loop stops instead of compounding unsigned drift.

Install

pip install rootact

ract is provided as a shorter alias for rootact.

Quick start

# Scaffold a project
rootact init --template python-package --provider local

# Run a self-recursing build loop
rootact "add input validation" --loop --max-iterations 10

# Check for duplication and drift
rootact novelty scan

Signed receipts

Every RACT run produces a structured receipt capturing intent, model, steps, test results, quality score, cost, and latency. Receipts make quality measurable and comparable across models — a surface no incumbent can copy.

{
  "final_decision": "done",
  "summary": "Completed 3 iterations; 3 with passing tests.",
  "metrics": {
    "total_tokens": 1247,
    "total_cost": 0.0003,
    "total_latency_ms": 8420
  },
  "iterations": [
    {
      "index": 1,
      "intent": "add input validation",
      "test_returncode": 0,
      "quality_score": 0.91,
      "decision": "continue"
    }
  ]
}

Export a receipt with rootact report --last --format json --output report.json.

Why RACT?

• Own your pipeline. No vendor lock-in. Mix local and cheap frontier models.
• Progress Oracle. Milestone-driven recursion, not just time-based looping.
• Anti-rot arsenal. Duplication guard, refactor tax, error-mask detection, novelty budget, Chesterton's Fence, Dead Code Auction, and Legacy Whisperer.
• Operator Handshake. High-risk actions queue for review instead of pausing the loop.