The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/endpoint_evaluation.py uses subprocess.run to execute command-line tools. Specifically, it calls gcloud auth print-access-token to retrieve authentication tokens and gsutil cat to read dataset files from Google Cloud Storage. These operations are used for legitimate integration with Google Cloud services but represent an execution surface for external binaries.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external and potentially untrusted data for evaluation purposes.
Ingestion points: Data is ingested through EvaluationDataset objects, which can be populated from JSONL files, CSVs, or session traces parsed by scripts/parse_adk_traces.py.
Boundary markers: The parsing scripts lack explicit boundary markers or instructions to ignore embedded commands within the ingested data.
Capability inventory: The skill possesses capabilities to write files (Stage 3 results), perform network requests (scripts/endpoint_evaluation.py), and execute shell commands (scripts/endpoint_evaluation.py).
Sanitization: While structural validation is performed by scripts/validate_dataset.py, there is no content-level sanitization to prevent malicious instructions within the datasets from influencing the LLM-as-a-judge metrics.
[DYNAMIC_EXECUTION]: The skill documents and supports the use of CodeExecutionMetric (referenced in references/metric_registry.md and references/sdk_patterns.md). This feature allows users to provide custom Python code as a string to be executed either locally or in a remote sandbox during the evaluation process.
[EXTERNAL_DOWNLOADS]: The documentation in SKILL.md and references/deployment.md provides instructions to download and install several Python packages, including google-cloud-aiplatform and google-genai, from official registries using pip.

agent-platform-eval-flywheel