Autonomous crash-recovery agent for your terminal. Wrap any command; if it crashes, a headless Claude Code session diagnoses the bug, writes a failing test, patches it on a separate branch, verifies the fix, and leaves a post-mortem. Your branch is never touched.
Hey PH š I built Phantom because the loop of "app crashes ā copy stack trace
ā paste into a chat ā paste the fix back" was eating my afternoons.
The part I care about most: Phantom doesn't take Claude's word for it. After
the session finishes, Phantom re-runs your test command itself and diffs the
branch against the starting commit. "Fixed" means Phantom measured it.
Everything happens on a throwaway branch ā your working branch is never
touched, and it can't push anywhere. Secrets like .env and *.pem are blocked
three separate ways.
It's MIT, zero runtime dependencies, and runs on your own Claude subscription.
Would love feedback, especially from anyone who wants to poke holes in the
safety model.