codeassay

CodeAssay

Git forensics tool for analyzing AI-authored code quality. Detects AI-generated commits, identifies rework, classifies root causes, measures code turnover rate, and generates quality reports.

Install

pip install codeassay

Quick Start

# Scan a repository
codeassay scan /path/to/repo

# View a report
codeassay report

# Scan multiple repos with combined output
codeassay scan ../repo1 ../repo2 ../repo3

Commands

Command	Purpose
codeassay scan <paths>	Scan repos (supports --with-scorer, --dry-run, --fail-on=turnover-red)
codeassay report	Generate quality report (CLI or markdown)
codeassay commits	List AI-authored commits (supports --source <glob>, --tool)
codeassay rework	List rework events with classification
codeassay reclassify <commit> <category>	Override a classification
codeassay export --format json	Export raw data
codeassay dashboard	Open interactive HTML dashboard in browser
codeassay config init	Scaffold .codeassay.toml in the current repo
codeassay config show	Print merged effective config
codeassay detect-test <hash>	Dry-run detection pipeline against one commit
codeassay tag [--tool <name>]	Add AI-Assisted: trailer to a commit (hook or amend)
codeassay install-hook [--tool <name>] [--mode always|prompt]	Install prepare-commit-msg hook
codeassay uninstall-hook	Remove the hook

How It Works

AI Commit Detection is layered — configurable in .codeassay.toml:

1. User-defined rules (highest priority) — match commits by author regex, branch pattern, commit-message regex, or author-plus-date-window.
2. Built-in profiles — shipped detectors for Claude Code, GitHub Copilot, Cursor, Aider, Windsurf, ChatGPT, and Gemini. Enabled by default, disable individually.
3. Probabilistic scorer (opt-in) — weights weak signals (diff shape, message structure, commit velocity, emoji, boilerplate headers, file diversity, punctuation) for commits no deterministic rule caught.

First match wins. Run codeassay config init to scaffold a starter config, codeassay config show to inspect the effective config, and codeassay detect-test <hash> to dry-run detection against a single commit.

Per-author fingerprint detection (opt-in) catches AI commits from authors with ≥20 prior commits by comparing each commit against the author's historical baseline on 5 metrics (diff size, comment ratio, identifier entropy, punctuation density, message length). Flags when ≥3 metrics diverge ≥2σ. Enable in .codeassay.toml:

[fingerprint]
enabled = true
min_prior_commits = 20
sigma_threshold = 2.0
min_divergent_metrics = 3

Use codeassay detect-test <hash> to see per-metric Z-scores when tuning thresholds.

To reliably tag AI commits going forward, install a git hook:

codeassay install-hook --tool cursor --mode always

This appends AI-Assisted: cursor to each commit message (idempotent — skips if a Co-Authored-By AI trailer is already present).

Rework Detection traces subsequent commits that modify AI-authored lines using git blame ancestry within a configurable time window (default: 14 days).

Classification categorizes rework into 7 types using commit message keywords and diff shape analysis:

• Bug fix, Misunderstanding, Test failure, Style/convention violation
• Security issue, Incomplete implementation, Over-engineering

Turnover Rate

Turnover Rate measures the fraction of lines added in a lookback window that are subsequently discarded or rewritten. Computed separately for AI and human cohorts; the AI/human ratio is the headline number. CodeAssay ships benchmarks.json with industry baselines (pre-AI 3.3%, 2026 avg 5.7%, healthy target <4%) so your report can show percentile vs industry.

Configure thresholds in .codeassay.toml:

[turnover]
lookback_days = 90
rewrite_window_days = 30
yellow_threshold = 0.04
red_threshold = 0.06

Fail CI builds on excessive turnover:

codeassay scan . --fail-on=turnover-red

Dashboard

Generate an interactive HTML dashboard with charts and visualizations:

codeassay dashboard

Opens a self-contained HTML file in your browser with:

• Summary metric cards (AI commit rate, first-pass success, rework rate, MTTR)
• Rework category doughnut chart with percentages
• Monthly trend line chart (AI commits vs rework events)
• Top rework file hotspots
• Rework by AI tool comparison

The dashboard works offline, requires no server, and is shareable — copy the HTML file to screenshot or embed in publications. Use --no-open to generate without opening the browser, or --output path.html to save to a custom location.

Ignoring Files