The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/export_xlsx.js dynamically generates a Python script at runtime to facilitate Excel exports with embedded images. This script is written to a temporary location and executed via child_process.execFile using the host's python3 or python interpreter. This pattern of runtime code generation and execution constitutes a dynamic execution risk.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection in scripts/parse_input.js. User-provided captions for image uploads are interpolated directly into the prompt sent to the vision model: ${prompt}\n\n用户附带说明：「${caption}」. This lacks robust boundary markers or safety instructions for the model to ignore instructions within the user text.
Ingestion points: caption parameter in parseImageInput (scripts/parse_input.js).
Boundary markers: Chinese brackets (「 and 」) are used, but no explicit instructions to ignore embedded commands are present.
Capability inventory: The skill can write to the local file system (wrong_questions.json) and execute system commands (python via export_xlsx.js).
Sanitization: No sanitization or escaping of the caption string was identified.

shiyi