The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes the glab CLI to perform GitLab operations. Technical analysis of scripts/gitlab_ops.py confirms that commands are constructed as argument lists (e.g., ['mr', 'create', '--title', args.title, ...]) and executed without a shell, which prevents command injection vulnerabilities.
[PROMPT_INJECTION]: The skill processes untrusted GitLab comments, which is a surface for indirect prompt injection. However, the implementation includes significant mitigations:
Ingestion points: MR and issue comments are fetched in scripts/gitlab_pr.py via the fetch_findings command.
Boundary markers: Documentation and code implementation (in scripts/gitlab_pr.py) ensure that untrusted comment bodies are quarantined under a specific raw_input.body field to isolate them from active context.
Capability inventory: The skill can manage merge requests and issues (create, merge, close, reply) and retrieve CI logs using glab.
Sanitization: The skill uses a keyword-based pre-filter for noise reduction. While not a security sanitizer, the architectural decision to quarantine untrusted bodies provides a primary security boundary.
[SAFE]: No security issues were detected. The skill does not contain hardcoded credentials, obfuscation, or unauthorized data exfiltration. All external interactions are limited to the intended GitLab service and well-known repositories.

workflow-integration-gitlab