The Agent Skills Directory

[SAFE]: Local Processing: The skill's primary functions are performed by Python scripts that operate strictly on local JSON files. There are no network calls, remote downloads, or interactions with external APIs.
[SAFE]: Standard Libraries: The provided scripts (scenario_runner.py, eval_diff.py) rely exclusively on built-in Python modules (json, re, math, sys, argparse), avoiding the risks associated with third-party or unverifiable dependencies.
[SAFE]: Data Sanitization and Privacy: The skill includes extensive documentation and checklists (e.g., references/scenario-and-fixture-design.md) that instruct users to scrub Personally Identifiable Information (PII) and ensure that no credentials, tokens, or internal URLs are stored in test fixtures.
[SAFE]: Proactive Security Evaluation: The harness is designed to detect security regressions, such as prompt injection, by using specific assertion types like tool_not_called to enforce safety boundaries on irreversible agent actions.
[SAFE]: Transparent Metadata: All file paths and metadata are consistent with the stated author ('borghei') and the skill's purpose as an engineering evaluation tool.

agent-harness