The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local Python scripts provided within the skill package (scripts/eval_state.py and the referenced scripts/skillsrc.py) using uv run. It also utilizes the built-in Python module http.server to host a local web server for viewing evaluation reports.
[PERSISTENCE]: The skill maintains evaluation state and history by creating a .draekien/ directory and subdirectories (activation/, impact/) within the project root. It versions test suites and appends run logs to JSONL files to track performance over time.
[INDIRECT_PROMPT_INJECTION]: As a testing framework, the skill ingests content from external skills (descriptions and instructions) to generate evaluation scenarios. This presents a potential attack surface where a malicious skill being tested could attempt to influence the evaluator. However, the skill implements a robust mitigation strategy by delegating all generation and evaluation tasks to isolated subagents with restricted contexts to prevent contamination or prompt-based attacks.
Ingestion points: Processes external skill files provided via the [skill-name-or-path] argument.
Boundary markers: Explicitly uses isolated subagents to separate the evaluation logic from the content being evaluated.
Capability inventory: Limited to local file system writes in specific directories and execution of provided scripts.
Sanitization: Uses context isolation as the primary security mechanism.

skill-evals