vibediff 0.1.0

Knows when a PR doesn't belong in your codebase.

VibeDiff

AI agents write code that compiles, passes tests, and feels completely foreign in your codebase. It works — but it doesn't fit. Your naming is snake_case, the PR is camelCase. Your project uses properties, the PR is full of getter/setter methods straight out of a Java textbook. The author committed 500 lines without changing a single AI-generated variable name.

VibeDiff learns how your project actually writes code, then tells you when a PR doesn't match.

What it catches

• Style drift — your codebase uses one set of conventions, the PR uses another
• Idiom contamination — AI bleeds patterns from other languages into your code (Java-style Python, Go-style error handling in JS, etc.)
• Raw AI dumps — 500 lines committed with zero human editing, generic variable names, TODO placeholders left in

How it works

VibeDiff runs 4 analyzers on every diff:

AI Detection (0-100) — heuristic signals that code was AI-generated:

• Comment patterns (restating code, section headers, excessive density)
• Naming patterns (verbose 4+ word names, uniform naming, no short vars)
• Burstiness (AI has uniform line complexity; humans vary wildly)
• Structural patterns (excessive is None guards, broad except Exception)

Style Drift (0-100) — how much the diff deviates from your codebase conventions:

• Naming convention mismatch (camelCase in a snake_case codebase)
• Comment density deviation from baseline
• Function length drift (2x longer or 0.4x shorter than average)
• Import style mismatch (bare imports vs from-imports)

Run vibediff learn first to build a fingerprint of your codebase conventions.

Collaboration Quality (0-100) — signs the human didn't actually review the AI output:

• Unresolved TODO comments left in
• Generic variable names (data, result, output, value)
• Generic test names (test_function_1, test_method_2)
• Placeholder stubs (pass, ..., NotImplementedError)
• Uniform comment density across files (no human editing trace)

Idiom Contamination (0-100) — AI bleeding patterns from other languages:

• Java in Python: getter/setter methods, Interface/Abstract class naming
• Go in Python: error-return patterns (return data, err)
• C/C++ in Python: == None instead of is None
• JavaScript in Python: callback parameters

Install

pip install vibediff

Usage

# Review last commit
vibediff review HEAD~1

# Review a range
vibediff review main..feature

# Review a GitHub PR (requires gh CLI)
vibediff review --pr 42

# Learn your codebase conventions first (enables drift detection)
vibediff learn

# Skip drift analysis
vibediff review HEAD~1 --no-fingerprint

# JSON output (for CI pipelines)
vibediff review HEAD~1 --format json

# Markdown output (for PR comments)
vibediff review HEAD~1 --format md

Every diff gets a letter grade (A-F) synthesized from all 4 analyzers:

╭──────────────────────────────── VibeDiff ─────────────────────────────────╮
│                                                                           │
│  ████      3 file(s)  +180  -12                                           │
│  ██ ██                                                                    │
│  ██  ██      AI Detection           █████████░░░░░░      58    medium     │
│  ██ ██       Style Drift            ██████░░░░░░░░░      42    medium     │
│  ████        Collaboration          ██████████░░░░░      65    mixed      │
│              Idiom Contamination    ██████░░░░░░░░░      40    medium     │
│                                                                           │
╰──────────────────────────────────────────────────────────────────────────╯

──────────────── AI Detection  58/100 (medium) ─────────────────
  restating_comments    4 comments that restate the code  ███░░
  verbose_names         3/5 functions have 4+ word names  ██░░░
  low_burstiness        CV 0.31 (expected >0.6)           ██░░░

────────────────── Style Drift  42/100 (medium) ────────────────
  naming_convention     expected snake_case, got camelCase ███░░
  comment_density       expected 8%, got 22%               ██░░░

─────────────── Collaboration  65/100 (mixed) ──────────────────
  unresolved_todos      3 TODO comments left in            ██░░░
  generic_names         4/12 variables have generic names  ██░░░

──────────── Idiom Contamination  40/100 (medium) ──────────────
  getter_setter         3 getter/setter methods [java]     ███░░
  error_return_pattern  4 Go-style error returns [go]      ██░░░

GitHub Action

Add VibeDiff to your PR workflow:

# .github/workflows/vibediff.yml
name: VibeDiff
on:
  pull_request:

jobs:
  review:
    runs-on: ubuntu-latest
    permissions:
      pull-requests: read
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0
      - uses: doncarbon/VibeDiff@main

The action posts a markdown report to the PR's job summary. To enable style drift detection, commit a fingerprint and pass it:

      - uses: doncarbon/VibeDiff@main
        with:
          fingerprint: .vibediff-cache/fingerprint.json

Author

Hamza

License

MIT