codechurn

Track AI vs human code quality — enhanced AI churn detection, hotspot scoring, and configuration system.

You're using Copilot, Claude, Cursor. But is the AI-generated code actually surviving? Or does it get rewritten in the next PR? codechurn answers that by analyzing your git history and attributing code to AI or human authors, then tracking what happens to it over time.

Why?

• 66% of devs are frustrated with "almost right" AI code (Stack Overflow 2025)
• AI code has 1.7x more issues and 2.74x more security vulnerabilities (Exceeds AI research)
• No open-source tool tracks AI code quality over time — until now
• Enhanced version now with configurable AI detection, improved hotspot scoring, and CI-friendly outputs

What it does

1. Enhanced Attribution — reads git history with configurable AI detection:
   • Co-authored-by headers (Co-authored-by: Copilot)
   • Commit message keywords (claude, copilot, cursor, aider)
   • Author name patterns
   • Customizable detection patterns and thresholds
2. Improved Hotspot Scoring — ranks files using weighted algorithm:
   • Churn rate (40%) — how much AI code gets rewritten
   • Survival rate (30%) — how long AI code persists
   • Volume weight (30%) — amount of AI code in file
3. Churn tracking — for each file, tracks lines added/removed by AI vs human
4. Survival analysis — what percentage of AI-written lines are still there?
5. Model breakdown — compare Claude vs Copilot vs Cursor code quality
6. Configuration system — customizable thresholds, patterns, and output formats

Install

npm install -g codechurn

Usage

Enhanced analysis with configuration

codechurn analyze
codechurn analyze --since "3 months ago"
codechurn analyze --since "2024-01-01" --json
codechurn analyze --path ../my-project --top 10 --md

# Custom thresholds and scoring
codechurn analyze --high-churn-threshold 0.6 --low-survival-threshold 0.8 --min-lines 100

# Config file support
codechurn analyze --config ./codechurn.config.json

Quick scan (summary only)

codechurn scan
codechurn scan --since "1 month ago"

Enhanced Hotspots with improved scoring

codechurn hotspots
codechurn hotspots --top 5 --json

# Filter by minimum hotspot score
codechurn hotspots --min-score 0.5

# See detailed scoring breakdown
codechurn hotspots --json | jq '.[] | {file, hotspotScore, aiChurnRate, aiSurvivalRate}'

Model comparison

codechurn models
codechurn models --json

Output

codechurn report — my-project (main)
period: 3 months ago → now

overview
  commits analyzed: 247
  AI commits:       89 (36.0%)
  Human commits:    158 (64.0%)

survival rates
  AI lines:    1243 — 52.3% survived (0.477 churn)
  Human lines: 3891 — 78.1% survived (0.219 churn)

revert rates
  AI:    8.9% of AI commits were reverts
  Human: 3.2% of human commits were reverts

avg code lifetime
  AI:    5 days
  Human: 30 days

⚠ 3 hotspots (improved scoring)
  src/auth.ts      score: 0.742 (high churn + low survival)
  src/api/handlers.ts score: 0.621 (high churn + volume)
  src/utils/helpers.ts score: 0.589 (moderate churn + low survival)

CI/CD Integration

# Fail if AI churn rate exceeds threshold
codechurn analyze --since "1 month ago" --json | \
  jq -e '.summary.aiSurvivalRate > 0.5'

Enhanced Configuration

Threshold customization

codechurn analyze --high-churn-threshold 0.6 --low-survival-threshold 0.8 --min-lines 100

How attribution works

codechurn looks for these signals in commits:

| Signal | Example | |--------|---------| | Co-authored-by | Co-authored-by: Copilot <[email protected]> | | Keywords in message | "used claude to generate", "copilot suggested" | | Author name | Author contains "copilot", "cursor", "aider" |

Hotspot scoring algorithm

Hotspots are ranked using a weighted formula:

score = churn × 0.4 + (1 - survival) × 0.3 + volume × 0.3

Where:

• churn = lines_removed / lines_added (0-1)
• survival = lines_still_present / lines_added (0-1)
• volume = min(ai_lines / min_lines, 1) (0-1)

As a library

import { ChurnAnalyzer, Reporter } from 'codechurn';

const analyzer = new ChurnAnalyzer('/path/to/repo');
const report = analyzer.analyze({ since: '3 months ago' });

console.log(report.summary.aiSurvivalRate); // 0.52
console.log(report.summary.humanSurvivalRate); // 0.78

License

MIT