The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting data from external, potentially untrusted sources such as git branch names and git log output.
Ingestion points: git branch --show-current, git log <target>...HEAD --oneline (in SKILL.md).
Boundary markers: The skill provides structural templates for titles and descriptions, but does not explicitly instruct the agent to ignore instructions embedded within the git logs.
Capability inventory: The skill utilizes git push and glab mr create via the Bash tool.
Sanitization: The instructions use a HEREDOC (cat <<'BODY') for the merge request body, which mitigates simple shell injection during command construction, but does not prevent the LLM from being influenced by malicious text within commit messages.
[COMMAND_EXECUTION]: The skill dynamically constructs and executes shell commands using the Bash tool to interact with the local filesystem and remote GitLab instances via git and glab. While the skill includes safety constraints (e.g., forbidding force-pushes and requiring commits before push), the execution path relies on the agent's correct interpretation of repository metadata.

create-mr