git-sift

An interactive CLI wizard for reviewing AI-generated code changes, ordered from highest to lowest risk.

git-sift walks you through every changed file in a diff — categorising each one by risk, showing you exactly why it was flagged, and letting you approve, raise a concern, mark a blocker, or ask Claude Code to fix it. Once everything is approved it can merge the PR for you.

Why

AI-generated code changes can bury high-risk edits — a removed auth check, a bumped dependency, an altered migration — inside large diffs full of low-risk formatting changes. git-sift forces you to confront the riskiest files first, every time.

Installation

pip install git-sift

Or from source:

git clone https://github.com/eamonnfaherty/git-compare-tool
cd git-compare-tool
pip install -e .

Requires Python 3.11+.

Quick start

# In any git repository — compares current branch against main
git-sift

That's it. git-sift detects your branch, diffs it against main (or master), and opens the wizard.

Usage

# Default: current branch vs main
git-sift

# Staged changes only
git-sift --staged

# Unstaged working-tree changes
git-sift --unstaged

# Specific refs
git-sift --ref-a main --ref-b feature/my-branch
git-sift --ref-a v1.0.0 --ref-b v1.1.0

# Non-interactive CI mode — exit 1 if security or dependency changes are detected
git-sift --fail-on security,dependency

# Install the Claude Code /git-sift skill
git-sift --install-skill

Risk categories

Files are assigned the highest-risk (lowest number) category that matches. Both filename patterns and diff content are checked — content rules can escalate a file beyond its path-based category.

Every matched rule adds a "Why this was flagged" reason shown above each file's diff so you always know what to look for before reading the code.

Interactive wizard

The wizard walks through files from risk 1 to 10, prompting for a decision on each:

After Concern or Blocker you are prompted for an optional note.

A compact status bar is pinned just above the prompt at the bottom of the screen, showing the file path, category, and risk level — so it is always visible regardless of diff length.

Interactive browser (second pass)

After the wizard completes its summary, an interactive file browser opens for a second pass:

Claude-assisted fixes

Pressing R on any file:

1. Prompts you for a description of what needs to change
2. Builds a structured prompt including the file path, risk category, the diff, and your recommendation
3. Launches an interactive Claude Code session (claude) in your terminal
4. When you exit Claude, the full wizard restarts with a fresh diff so the fix is reviewed from scratch

Every line Claude changes receives an inline # AI-REVIEW-FIX: <explanation> comment. Security fixes also get # AI-REVIEW-FIX(SECURITY): <vulnerability addressed>. This ensures AI-introduced changes are always visible to future reviewers.

Requires the claude CLI to be installed and on your PATH.

PR merge

Once all files are approved in the browser, git-sift offers to merge your open PR:

Merge this branch via PR? [Y/n]

Choose a merge strategy:

Requires the gh CLI and an open PR on the current branch. The branch is deleted after a successful merge.

Non-interactive / CI mode

git-sift --fail-on CATEGORIES

Runs the diff, categorises all files, prints a pass/fail summary, and exits with code 1 if any file matches a listed category. No prompts.

Category names (comma-separated, case-insensitive):

You can also use the numeric risk level directly (1–10).

# Fail on security or dependency changes (by name)
git-sift --fail-on security,dependency

# Same using risk levels
git-sift --fail-on 1,2

# Combine with a specific diff source
git-sift --staged --fail-on security
git-sift --ref-a main --ref-b HEAD --fail-on security,infrastructure,migrations

# In a CI pipeline
git-sift --fail-on security,dependency || exit 1

Exit codes:

Claude Code skill

Install the /git-sift slash command into Claude Code:

git-sift --install-skill

This writes ~/.claude/commands/git-sift.md so you can type /git-sift inside any Claude Code session to run the wizard, interpret results, and get help addressing blockers.

Detected patterns for incomplete AI code

git-sift specifically watches for signs that an AI left work unfinished or disabled existing functionality:

• Added # TODO, # FIXME, # HACK, or # XXX comments
• Added raise NotImplementedError
• Added a standalone pass or ... (ellipsis placeholder)
• Added comments containing words like disabled, not implemented, placeholder, temporary, workaround, for now
• Commented-out code: lines like # return something(), # validate_input(data)

These appear as reasons in the "Why this was flagged" panel even when a higher-risk rule (like SECURITY) already governs the file's category.

Requirements

• Python 3.11+
• Git
• claude CLI — for Claude-assisted fixes (brew install claude or see claude.ai/code)
• gh CLI — for PR merge (brew install gh)

License

MIT