The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The ensure_container_writable_directory function in scripts/paidf_common.py modifies directory permissions using a bitwise OR with stat.S_IRWXU | stat.S_IRWXG | stat.S_IRWXO. This effectively sets the output, metadata, and caption directories to be world-readable, world-writable, and world-searchable (equivalent to chmod 0777). This is performed to ensure the Docker container can write results regardless of the host's user configuration.
[COMMAND_EXECUTION]: The skill requires the agent to execute docker run with significant privileges, including host network access (--network host) and GPU resources (--gpus). It also involves mounting host media and output directories into the container.
[INDIRECT_PROMPT_INJECTION]: The skill processes external media lists and prompt templates that are interpolated into a YAML configuration file used by a VLM/video generation model.
Ingestion points: The skill reads input.jsonl for media metadata and a user-provided prompt file in scripts/prepare_paidf_config.py.
Boundary markers: The processed data is structured into a config.yaml file, providing some structural separation.
Capability inventory: The skill possesses the capability to execute Docker containers, write to the host filesystem, and perform network preflight checks.
Sanitization: Input media paths are normalized and validated to ensure they reside within a specified media directory. YAML scalar values are serialized using json.dumps to prevent structural injection into the configuration file.

paidf-cosmos-predict