The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on a bash script (scripts/smart_commit.sh) to execute several shell commands: git add ., git commit, git rev-parse, and git push. While these are standard git operations, the use of git add . is aggressive and could lead to the unintentional staging and pushing of sensitive files (like .env or credentials) that are not properly excluded in .gitignore.
[INDIRECT_PROMPT_INJECTION]: The skill accepts user-provided input for commit messages which are passed as arguments to a shell script. This creates a vulnerability surface where a malicious user could attempt to inject shell metacharacters.
Ingestion points: The SKILL.md instructions allow the agent to pass user-specified text as the first argument to smart_commit.sh.
Boundary markers: None. There are no delimiters or instructions to treat the message as inert data.
Capability inventory: The skill has the capability to write to the local filesystem (git commit) and perform network operations (git push).
Sanitization: The script uses double quotes ("$MESSAGE") which prevents word splitting, but it does not perform any validation or filtering of the input string to ensure it contains only expected commit message content.
[METADATA_POISONING]: The SKILL.md file claims the script handles a "Claude footer," yet the provided scripts/smart_commit.sh contains no logic to append such a footer or modify the commit message beyond the basic input. This inconsistency between the description and the actual script implementation can be misleading.

git-pushing