The Codex app (macOS desktop) works differently from hook-based tools — it's a multi-agent orchestration layer with built-in Git worktree management. openai
Each agent runs on an isolated Git worktree — a separate copy of your repo. The agent makes changes in its worktree, and you review the diff in the app's review pane before deciding to commit or merge. The model is not the one running git commit — it's a "do the work → human reviews the diff → human commits" flow. intuitionlabs
This means there isn't a git commit call to intercept the way there is with Claude Code.
1. Create an Omni skill
The integration point in the Codex app is Skills — bundled instructions + scripts that Codex discovers and uses automatically or on demand. You can create skills through the app's built-in Skills UI, or commit them as markdown files in your repo (e.g. .agents/skills/omni-commit.md) so they're shared across threads, team members, and across the App, CLI, and IDE Extension. developers.openai
---
name: omni-commit
description: Commit staged changes using the Omni CLI instead of git
---
# Omni Commit
When you need to commit changes, use the `omni` CLI instead of `git commit`.
## Commit
```bash
omni commit --yes --json
omni pr --yes --json
- Never use
git commitdirectly. - Never use
gh pr createdirectly. - Always pass
--jsonfor structured output. - Always pass
--yesfor non-interactive execution.
**2. Pair it with `AGENTS.md` for reinforcement**
```markdown
# AGENTS.md
## Git Policy
- Use the `omni-commit` skill for all commits and PRs.
- Never run `git commit`, `git push`, or `gh pr create` directly.
- Use `omni commit --yes --json` for commits.
- Use `omni pr --yes --json` for pull requests.
3. Use Automations for scheduled flows
The Codex app supports Automations — scheduled tasks that run agents on a cadence. You could set one up like: intuitionlabs
"Every day at 5pm, review uncommitted work across all threads and run
omni commit --yes --jsonfor any completed tasks."
The automation deposits results in the review queue for you to approve. intuitionlabs
How it differs from Claude Code:
| Aspect | Claude Code | Codex App |
|---|---|---|
Who runs git commit |
The agent, automatically | The human, after reviewing the diff intuitionlabs |
| Interception mechanism | PreToolUse hooks block & redirect code.claude |
Not needed — agent doesn't commit on its own developers.openai |
How to teach omni |
CLAUDE.md rules + hooks |
Skills + AGENTS.md developers.openai |
| Multi-agent git safety | N/A (single agent) | Built-in worktrees isolate each agent intuitionlabs |
| CI/scheduled work | Hooks + --yes --json |
Automations with review queue intuitionlabs |
The Codex app is a "do the body of work, then the human reviews and commits" tool. You don't need to intercept git calls — you teach Codex about omni through a Skill, and when you (or an Automation) decide it's time to commit, the agent runs omni commit instead of git commit. developers.openai