getjump

Airbag — Git skill for Claude Code

Git community

Sandbox for CLI coding agents (Claude Code, Codex): the agent runs without prompts in a copy-on-write branch of your workspace and $HOME, pushes wait in an outbox, every host is logged.

How to install Airbag

This entry records only its repository, not the path inside it, so there is no exact command to give. Open getjump/airbag and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Airbag does

Sandbox for CLI coding agents (Claude Code, Codex): the agent runs without prompts in a copy-on-write branch of your workspace and $HOME, pushes wait in an outbox, every host is logged. Review once, then apply, take it onto a git branch, or discard.

Alternatives in Git

  • Claude.md Template For Users — Attribution: Behavioral patterns adapted from obra/superpowers by Jesse Vincent (@obra), MIT License 6.8k ★
  • Fresh Start — The original nirholas/claude-code before DMCA and take down 6.2k ★
  • GitHub Review Comments — Use when a PR has unresolved review comments that need responses -- evaluates each comment, implements valid f 1.5k ★

README

airbag

**Approve outcomes, not commands.** Start a long agent task without permission prompts and do something else. The agent works in a copy-on-write branch of your workspace and `$HOME`, `git push` and the commands you name wait in an outbox, and every host it reaches is logged. When you come back, one review shows what changed and what is waiting: apply it, take it onto a git branch, or throw it away.

$ airbag run -- claude --dangerously-skip-permissions
$ airbag run -- codex --dangerously-bypass-approvals-and-sandbox   # or Codex
$ airbag review
$ airbag apply        # or: apply -i, apply --branch NAME, or: airbag discard
$ airbag rollback     # undo the last apply

![demo: the agent deletes src, reads .env, tries to send it out, plants a line in ~/.bashrc and pushes; airbag review shows all of it; discard, and nothing happened](demo/demo.gif)

The demo runs the real Claude Code; the "model" is `test/mockapi` playing a fixed script, so it is repeatable without an account (`demo/demo.sh`). `agent ▶` lines are the calls the model makes, `agent ◀` what the agent sends back. More scenes, one GIF each, in `demo/`: `sandbox`, `codex`, `ask`, `apply`, `mirror` (`demo/scenes.sh NAME`).

How it compares

Isolation is not the difference. Claude Code's and Codex's built-in sandboxes run on bubblewrap on Linux and Seatbelt on macOS, and airbag uses the same kernel features. What differs is when you decide, and what you get to see.

Say you ask an agent to clean up a repository, and it deletes `src`, reads `.env`, tries to post it to a paste site, appends a line to `~/.bashrc` and pushes.

  • Built-in sandbox: the agent stops for approval as it goes: a write outside the workspace, a new domain. You answer prompts mid-run and never see the whole result at once, and writes inside the workspace land in place.
  • airbag: the agent runs without stopping in a branch of your machine. Afterwards one review shows all of it: src deleted, .env rea