The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documents the use of the atmos devcontainer exec and shell commands, which allow for the execution of arbitrary commands within managed containers.\n
Evidence: Examples show execution of commands like terraform plan and aws sts get-caller-identity inside the container.\n- [CREDENTIALS_UNSAFE]: The documentation includes configuration templates that demonstrate mounting sensitive host directories into containers.\n
Evidence: Mount specifications in SKILL.md include bind mounts for ${HOME}/.aws and ${HOME}/.ssh, which exposes host credentials to the container environment.\n- [EXTERNAL_DOWNLOADS]: The tool is designed to download container images from external registries.\n
Evidence: Configuration examples reference the cloudposse/geodesic image and support custom images via the image or build.dockerfile fields.\n- [PROMPT_INJECTION]: The tool's ingestion of configuration files (atmos.yaml, devcontainer.json) represents an indirect prompt injection surface.\n
Ingestion points: The skill reads configuration from atmos.yaml and .devcontainer/devcontainer.json (via the !include tag).\n
Boundary markers: No explicit boundary markers or 'ignore' instructions for configuration data are mentioned.\n
Capability inventory: The tool can execute arbitrary commands, mount sensitive host files, and run privileged containers.\n
Sanitization: No evidence of input validation or sanitization for the configuration fields is provided in the documentation.

atmos-devcontainer