expo-repro-cleanup

expo-repro-cleanup

Cleanup tool for Expo reproduction examples - removes potentially unsafe files and configurations before running untrusted code.

Usage

bunx expo-repro-cleanup

By default it runs interactively, prompting before each change. Pass --non-interactive (or set CI, so it works out of the box in CI) to clean without prompts: it removes attack-surface and noise files automatically, and for a bare project runs expo prebuild --clean to regenerate native code. Since the repo is under git, you can review everything with git diff.

Options:
  -h, --help          Show this help message
  --version           Show version number
  --non-interactive   Run without prompts, cleaning automatically
                      (also enabled when the CI env var is set)
  --no-prebuild       Do not run `expo prebuild --clean` for bare projects

Example Workflow

# 1. Clone a reproduction repo
git clone https://github.com/someone/expo-issue-repro.git
cd expo-issue-repro

# 2. Step through each change interactively (default)
bunx expo-repro-cleanup

# ...or clean automatically, then review what changed
bunx expo-repro-cleanup --non-interactive    # or: CI=1 bunx expo-repro-cleanup
git diff

What it checks

In interactive mode (the default) you're prompted to keep or remove each item below. With --non-interactive or CI set, the tool decides automatically:

Removed automatically (attack surface / noise — not needed to run the app):

• Lock files and IDE settings (.vscode/)
• Build configs — metro.config.js, babel.config.js, .eslintrc.js, eslint.config.js, tsconfig.json, etc. A config whose contents exactly match the pristine Expo default is left untouched; only customized configs are removed.
• Git hooks — scripts that run automatically during git operations
• AI agent files — CLAUDE.md, CLAUDE.local.md, AGENTS.md, .mcp.json, and .claude/. A malicious repro can use these to inject prompts into (or run commands through) any AI coding agent you point at it. Their contents are never printed, since echoing them could inject the agent running this tool.

Kept and printed for you to review (integral to the repro):

• app.config.*, package.json scripts, and root source files — including source files flagged with suspicious patterns.