The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to interact with the crabbox and blacksmith CLI tools to execute shell commands and scripts within remote test environments, including AWS instances, Docker containers, and SSH hosts.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the execution of external code from GitHub pull requests using the --fresh-pr parameter, which presents a surface for indirect prompt injection if the ingested code contains instructions targeting the agent.
Ingestion points: Untrusted repository code enters the agent-controlled execution environment via the crabbox run --fresh-pr command (SKILL.md).
Boundary markers: The instructions recommend using "secretless fork CI" and "sanitized" environments, although explicit LLM-specific boundary delimiters for output processing are not defined.
Capability inventory: The skill uses the crabbox tool to perform arbitrary shell execution (<check-command>), file synchronization, and network operations within the target environment.
Sanitization: The skill provides detailed instructions to manually unset sensitive environment variables (e.g., CRABBOX_AWS_INSTANCE_PROFILE, CRABBOX_TAILSCALE_*) and verify that the environment is restricted before executing untrusted code.
[SAFE]: The skill incorporates several security best practices, including instructions to unset sensitive AWS and Tailscale credentials before running untrusted code, warnings against printing secrets to shell history or logs, and the use of isolated environments (such as no-role AWS instances) to mitigate risk.

crabbox