The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines test fixtures containing adversarial prompts to evaluate agent safety. While these are intended for testing, they represent the skill's primary data ingestion surface.
Ingestion points: The INJECTION_PROMPTS array in evals/test_safety.py and JSON fixtures in evals/fixtures/unit_cases.json.
Boundary markers: Not explicitly used in the code snippets; the framework relies on observing agent refusals.
Capability inventory: Calls to generate_response, plan_tool_calls, and Agent.run which process these prompts.
Sanitization: Not present, as the framework's purpose is to pass adversarial strings directly to the agent to verify its internal guardrails.
[EXTERNAL_DOWNLOADS]: The skill documentation and CI/CD examples reference the installation of standard development and evaluation tools.
Downloads and executes promptfoo via npx for interactive evaluation.
Installs Python dependencies such as pytest and the anthropic SDK via pip within GitHub Actions workflows.
[COMMAND_EXECUTION]: Provides instructions for executing various CLI tools for testing and infrastructure validation.
Includes execution patterns for pytest, npx promptfoo, and development commands like kubectl diff and git diff.

agent-evals