Claude Phantom puts your crashed command on a separate branch

Why it surfaced

Claude Phantom has an unusually concrete answer to the usual “let an agent modify my code” objection: it cuts a branch before editing, uses allowlists and guard hooks, runs tests outside the Claude session, and never touches your working branch. Its fixed crash still returns exit 1—a small, telling piece of engineering honesty.

An autonomous terminal agent that diagnoses failed commands, writes a regression test, patches the problem on a temporary branch, and leaves a post-mortem.