The Agent Skills Directory

[COMMAND_EXECUTION]: The skill directs the agent to perform standard git operations, including git status, git diff, git log, git add, and git commit. These commands are used to analyze the working tree and persist changes.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from the local repository (diffs, logs, and configuration files) to inform the generation of commit messages. This is an inherent attack surface for indirect prompt injection if the repository content is untrusted.
Ingestion points: Files are read from the repository using git diff, git log, and by inspecting configuration files like package.json or .commitlintrc.
Boundary markers: The instructions mandate the use of quoted shell heredocs (<<'EOF') for the git commit command, which provides a layer of protection against shell command injection from the generated message content.
Capability inventory: The skill allows the agent to modify the repository state through git add and git commit commands.
Sanitization: The skill provides specific rules to filter out unwanted content (like AI attribution strings) and warns against staging files that look like secrets (e.g., .env, credentials.json).

commit