The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's primary function is to execute arbitrary commands within isolated containers through the workshop exec, workshop run, and workshop shell tools. These are used for standard development tasks like building and testing code.
[PRIVILEGE_ESCALATION]: The instructions guide the agent to suggest system-level administrative commands to the user when necessary for the tool's operation. Examples include using sudo to manage LXD storage pools (sudo lxc storage set ...), restarting the Workshop daemon (snap restart workshop), and adding users to the lxd system group (sudo usermod -a -G lxd $USER). These commands are legitimate for managing the dev environment infrastructure but require elevated permissions.
[INDIRECT_PROMPT_INJECTION]: The skill processes project-local files which creates a potential vector for indirect injection.
Ingestion points: The agent establishes its context by reading workshop.yaml and sdk.yaml files from the project repository (Step 1 of bootstrap-project.md).
Boundary markers: The instructions do not provide specific delimiters or warnings to isolate logic from data within these project-local definition files.
Capability inventory: The agent has extensive capabilities including writing to the local filesystem, executing shell commands in containers, and querying the official SDK Store.
Sanitization: There is no mention of sanitizing or validating the content of these configuration files before the agent acts on the instructions they contain.
[DYNAMIC_EXECUTION]: The skill involves the dynamic creation and execution of shell scripts for container hooks and actions. The agent is instructed to write these scripts to the .workshop/ directory at runtime and mark them as executable using chmod +x (Step 4 of author-in-project-sdk.md).

use-workshop