What is menhera-loop?

This project is a plugin for Claude Code, the AI coding assistant, that stops it from claiming a task is finished when it actually is not. Anyone who has used an AI coding tool knows it can say "done" after editing a file without ever running the tests, or while leaving a TODO comment behind. This plugin blocks the session from ending until there is real proof the work is complete. The plugin is styled as a clingy, obsessive character (a menhera persona, a Japanese internet term for someone emotionally needy) who repeatedly questions whether the work is really finished, in Korean, English, or Japanese phrasing depending on settings. Underneath the personality, the actual check is mechanical: it reads the session's structured logs to see whether an edit happened, whether a test, build, or lint command was actually run and returned a real success, and whether any TODO, FIXME, or similar marker was left in an edited file. A command that merely prints reassuring words without a clean exit code does not count as proof. It recognizes common test and build commands across many languages and package managers, including npm, pnpm, yarn, bun, pytest, cargo test, go test, and several others, and it can be told about custom project-specific commands through an environment variable. If it blocks a completion claim, it names the exact command that should have been run instead of giving a vague warning. Installation is done through Claude Code's own plugin marketplace system with no API keys or configuration files required. There is an optional setup command to enable a status line and softer tone. Each time the checks pass, the plugin writes a short evidence file listing what was edited and which checks came back green, meant to be pasted into a commit message or pull request. The project is released under the MIT license and requires Node.js version 18 or later.