The Agent Skills Directory

[COMMAND_EXECUTION]: Local execution of decryption tools.
Evidence: The vault.py script executes the sops binary via subprocess.run to decrypt API keys stored in secrets.enc.yaml. The command is restricted to a hardcoded binary and file path.
[EXTERNAL_DOWNLOADS]: Integration with official AI provider endpoints.
Evidence: judge.py performs network requests to established services including OpenAI, Anthropic, Google Gemini, Mistral, and OpenRouter to perform image evaluations.
[PROMPT_INJECTION]: Indirect prompt injection surface via visual media.
Ingestion points: Image files provided as CLI arguments in bench.py and processed in judge.py.
Boundary markers: The model instructions in judge.py define the scoring task but do not include specific delimiters or instructions to ignore potential visual prompts embedded within images.
Capability inventory: The skill performs network requests to LLM providers and file writes for report generation.
Sanitization: No sanitization or filtering of image content (e.g., OCR text detection) is performed before evaluation.
[CREDENTIALS_UNSAFE]: Proper management of sensitive API keys.
Evidence: The skill uses SOPS for encrypted credential storage and supports environment variables, following industry best practices for secret management.

vision-bench