The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on the execution of dstack CLI commands to manage GPU workloads. This includes commands for applying configurations (dstack apply), monitoring runs (dstack ps, dstack logs), and managing infrastructure (dstack fleet, dstack volume). These are core to the tool's function.
[DATA_EXPOSURE]: The instructions note that dstack accesses and modifies configuration files in ~/.dstack/ and SSH configurations in ~/.ssh/config to manage tunnels and aliases. This is standard behavior for an orchestration tool that facilitates remote connections to compute resources.
[REMOTE_CODE_EXECUTION]: The YAML examples demonstrate the use of uv pip install within commands sections to set up environments on remote GPU nodes. This is a legitimate feature of the orchestration platform for preparing development and training environments.
[EXTERNAL_DOWNLOADS]: The skill includes an image selection policy referencing official rocm/* Docker images from Docker Hub. These are well-known, trusted sources for AMD GPU runtimes.
[INDIRECT_PROMPT_INJECTION]: The skill has a surface for indirect prompt injection as it processes user-provided *.dstack.yml configuration files and repository content. However, the instructions include guidance on verification, output accuracy, and avoiding automated execution without user confirmation, which serves as a mitigation surface.

dstack