The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the glab CLI to perform a wide range of GitLab operations. This includes read-only actions like glab mr view and glab mr diff, as well as actions that modify the repository or CI state, such as glab mr note, glab mr approve, and glab ci retry.
[DATA_EXFILTRATION]: To configure its sub-agents, the skill reads instruction files from the local filesystem at ~/.claude/agents/code-reviewer.agent.md and ~/.claude/agents/security.agent.md. It also reads the project's CLAUDE.md file. The contents of these files are included in prompts sent to external LLMs (Claude Opus and Sonnet).
[PROMPT_INJECTION]: The skill is subject to indirect prompt injection because it incorporates untrusted external data into agent prompts.
Ingestion points: Untrusted data enters the context through glab mr diff (the code being reviewed) and the project's CLAUDE.md file.
Boundary markers: The skill uses basic Markdown headers (e.g., ## Diff, ## Project Conventions) to delimit data, which does not provide strong protection against instructions embedded within that data.
Capability inventory: The agent has the ability to post comments, approve Merge Requests, and manage CI/CD pipelines via the glab CLI.
Sanitization: There is no evidence of sanitization or escaping of the diff content or project conventions before they are interpolated into the reviewer prompts.

gitlab