The Agent Skills Directory

[COMMAND_EXECUTION]: Several automation scripts (e.g., setup.sh, configure.py, gate-check.py, run-and-report.py) utilize subprocess.run() to interface with the platform's CLI tool (cxas), git for environment checks, and other internal utility scripts. These operations are limited to orchestrating the skill's own functionality and interacting with standard developer tools.
[EXTERNAL_DOWNLOADS]: The skill interacts with official Google Cloud Platform endpoints (ces.googleapis.com) and utilizes Google's Gemini LLM models for analysis tasks. It provides instructions for installing dependencies like the cxas-scrapi package from local source directories. These interactions are directed toward established, well-known service domains and originate from a trusted vendor.
[DYNAMIC_EXECUTION]: The unit testing framework (scripts/tests/test_cluster_failures.py) employs importlib.util to dynamically load internal modules. This is a standard practice used to test scripts that have non-standard filenames (such as those containing hyphens), ensuring proper isolation during the testing phase.
[INDIRECT_PROMPT_INJECTION]: The skill analyzes conversational logs and instructions using LLM-based evaluation judges (agents/triage-failure.md, agents/coverage-analyst.md). To prevent potential obedience to malicious instructions embedded in the data, the skill utilizes architectural safeguards such as XML delimiters, explicit schema validation via Pydantic models, and strict role isolation to differentiate between instructions and the data being processed.

cxas-agent-foundry