The Agent Skills Directory

Indirect Prompt Injection Surface: The skill evaluates data from various sources (CSV, JSONL, GCS, and ADK traces) using LLM-based metrics. This represents a potential surface where instructions embedded in the data could influence the judge model's evaluation process.
Ingestion Points: Data enters the context via EvaluationDataset constructors and GCS file reading in scripts/endpoint_evaluation.py and scripts/parse_adk_traces.py.
Boundary Markers: The skill relies on the underlying agentplatform SDK to handle prompt boundaries.
Capability Inventory: The skill can execute shell commands via subprocess, make network requests to GCP APIs, and write result artifacts to the local file system.
Sanitization: Explicit sanitization of processed data before evaluation is not detailed in the skill's high-level logic.
Command Execution via Subprocess: The skill uses subprocess.run to interact with Google Cloud CLI tools. While these are legitimate vendor resources, they involve executing commands in the host environment.
Authentication: scripts/endpoint_evaluation.py executes gcloud auth print-access-token to retrieve a temporary credential for API authentication. This is a standard pattern for GCP-integrated tools.
Data Retrieval: The same script uses gsutil cat to download datasets from GCS buckets specified by the user.

agent-platform-eval-flywheel