The Agent Skills Directory

[SAFE]: The skill exhibits an exceptional security posture, specifically designed to defend against common attack vectors in AI agent environments. It utilizes a 'Git Environment Policy' that unsets a comprehensive list of over 20 GIT_* variables (derived from git's own internal list) to prevent environment-based hijacking of repository roots or configurations.
[COMMAND_EXECUTION]: All shell commands are routed through a dedicated dispatcher (smart-commit-dispatch.sh) that enforces a strict allowlist: bash, git, mktemp, and rm. This prevents the execution of unauthorized tools or external network utilities like curl or wget.
[SAFE]: The bundled scripts (smart-commit-execute.sh and smart-commit-inspect.sh) establish privileged mode at launch using exec /usr/bin/env -u ... /bin/bash -p. This ensures that the scripts ignore BASH_ENV and potentially malicious shell functions exported from the environment.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data such as git logs, diffs, and configuration values. To mitigate risks, it includes an esc() function that sanitizes control characters and ANSI escape sequences. It also features 'AI trailer sanitization' to prevent the accidental or malicious inclusion of AI attribution headers in commit messages.
[SAFE]: The skill implements a robust sensitive-file exclusion policy, automatically warning the user and blocking the commit of files like .env, *.pem, *.key, id_rsa, and cloud credentials.
[SAFE]: Execution of commits in --execute mode requires explicit user approval via AskUserQuestion, with a structured 'Goal mode' that only relaxes this for pre-approved user goals under strict conditions.

smart-commit