The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script q-multimodal/scripts/gemini/batch/utils.py employs importlib.util to dynamically load and execute a configuration file (pipeline_config.py). This pattern allows for the execution of arbitrary Python code defined within the configuration file at runtime.
[COMMAND_EXECUTION]: The media processing scripts q-multimodal/scripts/opensmile/audio_features.py and q-multimodal/scripts/pillow/video_features.py utilize subprocess.run to invoke ffmpeg for audio and frame extraction. While these calls use list-based arguments (which is generally safer), they operate on file paths derived directly from user-provided datasets (Excel/CSV), which could lead to unexpected behavior if path names are maliciously crafted.
[INDIRECT_PROMPT_INJECTION]: The q-multimodal and q-tf sub-skills ingest untrusted data from user-provided tabular files (e.g., image captions, document text) and interpolate this content directly into prompts sent to the Gemini API.
Ingestion points: Dataset files loaded via pandas.read_excel in q-multimodal/scripts/gemini/batch/0uploadMedia.py and q-tf/scripts/classify_outliers.py.
Boundary markers: The skill lacks explicit delimiters or instructions to ignore embedded commands within the ingested metadata and text fields.
Capability inventory: The skill has the ability to write files (checkpoints), perform network operations (Gemini API), and execute shell commands (ffmpeg).
Sanitization: There is no evidence of sanitization or escaping for the user-provided text before it is sent to the LLM.
[DATA_EXPOSURE]: The skill requires the use of API keys (GOOGLE_API_KEY, GEMINI_API_KEY) stored in .env files. While this is a standard and recommended practice for secret management, the skill's media upload scripts have broad read access to the local filesystem to process datasets and media files.

q-scholar