The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions to the agent to use sudo for administrative tasks, such as creating symbolic links in system-protected directories (/usr/local/bin). It also utilizes a wide array of shell commands for environment setup and validation.
[CREDENTIALS_UNSAFE]: The instructions orchestrate the forwarding of sensitive secrets, specifically OPENAI_API_KEY and CRABBOX_COORDINATOR_TOKEN, into remote environments. While the skill includes guidelines for secret redaction in logs, the mechanism inherently facilitates the movement of credentials across network boundaries.
[EXTERNAL_DOWNLOADS]: Fetches the crabbox CLI from the vendor's official GitHub-hosted Homebrew tap and manages the installation of project-level dependencies using pnpm.
[REMOTE_CODE_EXECUTION]: The core functionality of the skill is the remote execution of local scripts and shell commands on diverse platforms, including AWS, Blacksmith Testbox, and remote SSH targets. It supports uploading local script files to these remote environments for execution.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection due to its processing of untrusted external content from GitHub pull requests (--fresh-pr) and local repository files which are used to determine the commands executed on remote infrastructure.
Ingestion points: Pulls data from GitHub PRs and reads local workspace files.
Boundary markers: No explicit markers or warnings are used to differentiate untrusted data from the command logic.
Capability inventory: Comprehensive remote command execution (crabbox run), script uploading (--script), and file system manipulation.
Sanitization: There is no evidence of sanitization for the commands or script contents derived from untrusted inputs.

crabbox