The Agent Skills Directory

[COMMAND_EXECUTION]: The skill facilitates the execution of shell commands through two main mechanisms: lifecycle hooks (hooks.host, hooks.sandbox) for environment preparation and prompt shell expressions (using the !"command" syntax) that run inside the sandbox to provide real-time context to the agent.
[DYNAMIC_EXECUTION]: Prompt files support dynamic evaluation where shell expressions are expanded at runtime before the final prompt is sent to the agent. This allows for the injection of dynamic environment state, such as test results or git diffs, directly into the agent's instructions.
[PRIVILEGE_ESCALATION]: The SDK supports executing hook commands with elevated privileges using the sudo: true option in sandbox hooks. It also allows configuration of specific container UIDs and GIDs for Docker/Podman providers.
[INDIRECT_PROMPT_INJECTION]: The skill architecture defines an ingestion surface for external data via promptArgs and the output of shell expressions. These values are interpolated into agent prompts without explicit sanitization mentioned in the provided documentation, representing a potential indirect injection vector if processing untrusted content.

sandcastle