AI-powered git diff review in your terminal.
The lazygaze TUI pipes git diffs to Claude CLI or Github Copilot CLI with
streaming output, prompt library, and persona system. Diff on the left. Analysis
on the right. No browser, reduced context switching. Fast workflow.
This software is licensed under the MIT LICENSE
+ IMPORTANT
- User is responsible for all Claude or Copilot costs incurred and meeting their terms.
- Lazygaze simply delegates to these CLI tools which you configure.
| Feature | Description |
|---|---|
| Split-pane TUI | Syntax-highlighted diff left, streaming markdown review right |
| Prompt library | Extensive library of curated prompts across architecture, security, performance, testing, and workflow |
| Personas | 50+ categorized reviewer archetypes inspired by legendary figures |
| Coding agent selection | Use either claude or copilot |
| Model cycling | Switch between Sonnet, Opus, and Haiku on the fly |
| Multiple diff sources | Git ranges, staged changes, initial commit |
| Clipboard | Copy diff or review to clipboard with a single key |
| CLI mode | --cli for scripting and CI pipelines |
| Persistent config | Remembers your preferred model and persona |
- Go 1.26+
At runtime, lazygaze delegates to Git and Claude
CLI running locally. It expects to
find git and claude shell commands in your PATH. Ensure Claude is
installed properly, authenticated (claude --version), and has
credits/subscription.
GitHub Copilot CLI is also supported
as an option instead of Claude Code. Ensure it is installed properly,
authenticated (copilot --version), and has credits/subscription.
brew install benstroud/tap/lazygazegit clone https://github.com/benstroud/lazygaze
cd lazygaze
make build
sudo mv lazygaze /usr/local/bin/ # or add to your PATHThe primary mode of operation is the TUI, where you can enter a git range and other options interactively:
lazygazeTypically, you'll set options interactively once the TUI is launched, but you can also initialize options when launching by using flags:
# Diff expression set at launch instead of interactively (Review the last 3 commits)
lazygaze HEAD~3..HEAD
# Prompt set a launch instead of interactively
lazygaze HEAD~1..HEAD "Focus on security vulnerabilities"
# Specify a model at launch instead of interactively
lazygaze --model opus HEAD~5..HEADIt is also possible to run non-interactively by using the --cli flag:
# Headless non-interactive output (pipes to stdout)
lazygaze --cli HEAD~1..HEAD| Key | Action |
|---|---|
tab |
Switch focus between diff and review panes |
? |
Open full keybinding help screen |
j / k |
Scroll focused pane |
q / ctrl+c |
Quit |
| Key | Action |
|---|---|
S |
Review staged changes |
D |
Review dirty changes (uncommitted) |
^ |
Review last commit (HEAD^..HEAD) |
U |
Review upstream changes (diff against upstream branch) |
~ |
Shorthand: enter n to review HEAD~n..HEAD |
: |
Enter a git range (e.g. HEAD~3..HEAD, main..feature, origin/master..HEAD) |
r |
Refresh / re-run current diff |
| Key | Action |
|---|---|
/ |
Set a custom prompt |
L |
Open prompt library |
P |
Select a reviewer persona |
m |
Cycle model: sonnet → opus → haiku |
c |
Copy focused pane to clipboard |
z |
Zoom in/out of the active pane. Useful before selecting text with mouse |
Press L to browse built-in prompts organized by category:
- Architecture — layering violations, structural improvements
- Bug Detection — nil dereferences, race conditions, off-by-ones
- Code Review — thorough review with actionable feedback
- Code Quality — DRY, naming, error handling
- Documentation — README generation, missing comments
- Performance — allocations, N+1 queries, hot paths
- Security — injection, hardcoded secrets, auth logic
- Testing — edge cases, unit test suggestions, Gherkin AC
- Workflow — commit messages, PR descriptions, Jira tickets
Press P to browse reviewer archetypes organized into categories:
- CS Foundations — structured programming, algorithmic rigor, formal correctness
- Legendary Creators — language designers, OS authors, paradigm inventors
- Clean Code & Design — readability, SOLID principles, refactoring
- Systems & Performance — low-level optimization, real-time constraints
- Educators & Evangelists — pedagogy, documentation, knowledge transfer
- Influencers — popular online voices and engineering culture
Special modes: (Critical Issues Only) suppresses style feedback and reports only bugs, security issues, and data loss risks. (Terse) returns bullet points only.
make build # compile
make test # run tests
make clean # remove binaryI took inspiration for this project idea and its name "lazygaze" from the open source TUI projects lazygit and lazydocker, both created by Jesse Duffied.
Also see my Zsh LLM code review helper snippets that were a precursor to this project. Example:
# Review the last commit of your feature branch
claudiff "HEAD^..HEAD" "code review"MIT