The Agent Skills Directory

[COMMAND_EXECUTION]: The skill orchestrates complex workflows by executing local Python scripts and Docker commands. It uses standard practices for argument passing in subprocess calls within its test suite and main instructions.
[PRIVILEGE_ESCALATION]: The helper script paidf_common.py includes a function that generates a sudo command string to help users resolve file permission issues. The skill's instructions specifically mandate that the agent must request explicit user approval before executing any such corrective commands.
[CREDENTIALS_UNSAFE]: The skill requires access to sensitive tokens (HF_TOKEN for HuggingFace and VLM_API_KEY for captioning services). It implements security best practices by instructing the agent to never print these tokens, never ask the user to paste them in chat, and only forward them to the Docker container via environment variables.
[EXTERNAL_DOWNLOADS]: The workflow involves interactions with external services, including pulling Docker images, downloading model weights from HuggingFace, and communicating with user-provided OpenAI-compatible VLM endpoints. These operations are essential to the skill's primary function and are handled through standard interfaces.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data from JSONL files and user-defined prompt files. To mitigate injection risks when generating the PAIDF YAML configuration, the scripts employ JSON-based string serialization to ensure that user inputs are correctly escaped and constrained within their intended data fields.

paidf-cosmos-predict