The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill invokes npx -y @zereight/mcp-gitlab to download and run code from a community-managed repository at runtime. This pattern allows for the execution of external code that is not version-pinned or verified for integrity on the host system.
[EXTERNAL_DOWNLOADS]: Dynamically fetches the @zereight/mcp-gitlab tool from the public NPM registry during the initialization of the MCP server.
[COMMAND_EXECUTION]: Spawns a local process via the shell to execute the npx command for server communication over stdio.
[PROMPT_INJECTION]: The skill contains an indirect prompt injection surface where untrusted data could influence agent behavior.
Ingestion points: Reads content from potentially untrusted sources including issue comments (list_issue_comments), merge request details (get_merge_request), wiki pages (get_wiki_page), and repository files (get_file_content).
Boundary markers: There are no defined delimiters or instructions to treat data from the repository as untrusted, increasing the risk that the agent may follow instructions embedded in comments or files.
Capability inventory: The skill has access to sensitive write operations such as merging code (merge_merge_request), triggering CI/CD pipelines (create_pipeline), and creating issues or releases.
Sanitization: The analysis found no evidence of sanitization or validation of the content retrieved from GitLab before it is processed by the agent.

gitlab-devops