ianymu/claude-verify-before-stop

verify-before-stop is a small shell script that plugs into Claude Code, Anthropic's command line coding assistant, and stops a coding session from ending until the AI has actually proved it tested its work. The problem the author describes is familiar to anyone who has used these tools: the model says all tests are passing, you merge the change, and production breaks. The script does not try to make the model smarter, it adds a workflow guard at the moment the session would end. Claude Code has a system called Stop hooks: a script the tool runs when a session is about to end. This project hooks into that. When the model tries to stop, the script first checks git for changed or untracked files. If files were changed, it looks for a recent VERIFIED entry in a log file at .claude/state/stop-verify.log written within the last five minutes. If that entry is missing, the script blocks the stop and prints exact instructions telling the model what it has to do. If no files were touched, the stop goes through with no friction. Inside a session, the agent is expected to log proof of what it ran. The README shows an example: run npm test, then append a VERIFY_ACTION line and a VERIFIED line to the log. The same pattern works with curl for HTTP services, psql for database checks, or Playwright for UI tests. Installation is short. You make a .claude/hooks directory in your project, download the script with curl, mark it executable, and add a Stop hook entry to .claude/settings.json that points at the script. The dependencies are bash and Python 3 from the standard library. The script is MIT licensed and pitched as the free sample from a paid pack of six Claude Code productivity hooks. The README also discusses compatibility with related projects and an academic reference on agent failure modes.