The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's evaluation harness processes external datasets containing adversarial inputs and prompt injection payloads designed to test agent robustness.
Ingestion points: The resources/eval-benchmarks/eval-cases.json file provides input payloads for the evaluation script.
Boundary markers: The harness treats inputs as structured JSON data, using specific fields for prompt content.
Capability inventory: The run-eval-benchmark.py script reads the local filesystem for role definitions, schemas, and indices. The skill is permitted to use write_file, run_tests, and other high-privilege tools.
Sanitization: Grading logic relies on schema validation (Draft 2020-12), tokenization, and regex matching to verify expected behaviors without executing the ingested strings as code.
[DYNAMIC_EXECUTION]: The test suite in tests/test_golden_benchmarks.py uses importlib.util to dynamically load the run-eval-benchmark.py runner script from the project directory at runtime for unit testing.
[EXTERNAL_DOWNLOADS]: The evaluation cases and manifest files reference the official JSON Schema specification at json-schema.org for data validation purposes, which is a trusted industry standard.

agent-evaluation