The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the gh (GitHub) CLI via the subprocess module to interact with GitHub Actions workflows. It fetches run lists, job details, and downloads diagnostic bundles. The script constructs these commands using a list-based argument passing method rather than a shell string, which mitigates command injection risks. Furthermore, a slug() function is implemented to sanitize data (like reservation names or GPU types) before using them as directory names for downloads, preventing path traversal attacks.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted data from UAT logs and JSON reports, which represents a structural indirect prompt injection surface. Content from these logs is echoed to the agent's context for summarization.
Ingestion points: Data enters the agent's context through GitHub API responses and the contents of diagnostic files such as MANIFEST.yaml, report.json, and readiness-gate.log within uat_report.py.
Boundary markers: The skill uses Markdown tables and headers to delimit tool outputs, and the instructions require the agent to cite specific file_path:line markers when reporting findings.
Capability inventory: The skill's capabilities include running gh CLI commands (read-only for runs) and writing downloaded artifacts to a localized directory (e.g., /tmp/uat-debug).
Sanitization: Path-level sanitization is present via the regex-based slug() function. While the content of the logs themselves is not sanitized, the risk is typical for diagnostic tools and is managed by the agent's core instructions to report, rather than execute, content found in logs.

aicr-uat-report