Practice — Git Workflows with Claude Code (12 questions)

You ask Claude Code, running in acceptEdits mode, to "clean up the error handling in payments/charge.ts and commit the result." Claude makes several edits across the file and proposes this commit message:

Improve error handling in charge.ts

Wrap the Stripe API call in a try/catch and log failures.

1. Given that acceptEdits mode auto-approves file edits without prompting per-edit, what specific step should you take before this commit is created, and why does the agentic loop's own verification not cover it?
2. Name two concrete things that could be wrong with this diff that a fluent, well-formatted commit message would not reveal.
3. If you discover after the fact that the commit included a hardcoded Stripe test key that was already sitting in the file, what is the complete remediation — and why is "delete it in a follow-up commit" not sufficient on its own?

After a risky experiment goes wrong, you open the rewind menu (Esc Esc or /rewind).

Which of these can it restore you to?

Claude Code ran rm old-script.sh via Bash earlier in the session, and now you want to use /rewind to undo it.

Why might that not work the way you expect?

You spin up an isolated worktree for a subagent to work in packages/api/, but by default --worktree checks out the entire repository to disk — wasteful in a huge monorepo.

What setting scopes it down to just the directories you need?

You're using worktree.sparsePaths to scope several parallel worktrees to specific packages, but each one is still duplicating a heavy shared node_modules directory.

What setting pairs with sparsePaths to avoid that?