agent-diff-view
v2.0.2
Published
Browser-based interactive git diff reviewer for AI coding agents
Maintainers
Readme
agent-diff-view
A browser-based interactive reviewer of AI agent changes, designed for humans. Understand and own the specific changes your agent makes — no git diffs, just clean before-and-after views of the actual code changes.

Type /changes in your AI coding harness and a local browser tab opens. Browse changed files,
navigate hunks, leave targeted comments on the ones that matter, hit ⌘↵ to submit —
the agent receives a structured prompt that maps every comment to its exact file and hunk.
Supported harnesses
| Harness | Command | Installs to |
|---------|---------|-------------|
| Claude Code | /changes | ~/.claude/commands/changes.md |
| OpenCode | /changes | ~/.config/opencode/commands/changes.md |
| Gemini CLI | /changes | ~/.gemini/commands/changes.toml |
| Codex CLI (OpenAI) | $changes | ~/.agents/skills/changes/SKILL.md |
| pi.dev | /changes | ~/.pi/agent/prompts/changes.md |
Note: Codex CLI uses
$changes(not/changes) because it uses a skills system rather than slash commands. The skill is configured to only trigger when explicitly invoked — it will not auto-activate based on task descriptions.
Install
npx agent-diff-view installThis opens an interactive picker — select your harness and the command is installed.
Usage
/changes # review uncommitted changes against HEAD
/changes --staged # review staged changes only
/changes main # review changes against a branch or commit
/changes v1.0..v2.0 # review a commit rangeFor Codex CLI, replace /changes with $changes.
Keyboard shortcuts
| Key | Action |
|-----|--------|
| ⌘← / ⌘→ | Previous / next hunk |
| ⌘↑ / ⌘↓ | Previous / next file |
| ⌘↵ | Submit |
How it works
You type /changes in your AI coding harness
The agent runs: npx agent-diff-view HEAD ← blocks
→ parses changes into before/after view
→ starts local HTTP server
→ opens browser tab
→ waits for you to submit
You: browse files, annotate hunks, hit ⌘↵
→ builds structured prompt
→ prints to stdout, exits
The agent receives the output and addresses each commentNo background processes are left running. No npm dependencies — only Node built-ins.
