PPCH Framework

The problem

Code-quality issues are usually found after a pull request merges. Can complexity risk be predicted when a pull request is submitted, and fixed before it merges?

How it works

1. 01

   Sense

   Mines PR history through the GitHub REST API and runs SonarScanner on before/after snapshots to build a leakage-safe dataset of 334 PRs.

2. 02

   Analyze

   Compares four models with 1,000-iteration bootstrap validation; the best, Logistic Regression, reached 0.78 AUC.

3. 03

   Act

   An ILP-based Extract Method engine proposes refactorings behind four safety checks; every applied suggestion was re-verified safe by a real SonarQube scan.

Results

• Dataset

  1. 1,767PRs scraped
  2. 500sampled
  3. 497scanned
  4. 334usable

• Model comparison

  Mean AUC, 1,000-iteration bootstrap validation

  Model	Mean AUC	95% CI
  Logistic Regressionbest	0.780	0.709–0.847
  XGBoost	0.750	–
  Random Forest	0.746	–
  Naive Bayes	0.660	–

• Act engine

  ILP-based Extract Method with four safety checks:

  • Return-forwarding
  • Break/continue
  • Independent-variable mutation
  • Non-productive extraction

  • 10 suggestions applied, each re-verified safe by a real SonarQube scan
  • 4 correctly declined
  • 0 unsafe

• Deployed GitHub Action

  A reactive scanner took a mean 82.4 h (median 12.9 h) to flag complexity-increasing PRs.

  The Action caught 22 of 52 (42.3%) before merge.

Screenshots

Limitations

The deployed Action runs Random Forest + XGBoost; Logistic Regression was found best later, in the four-model comparison.