The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes the grok and git command-line tools via a Node.js relay script.
Argument Validation: Critical flags such as --model, --effort, and --session are validated against a strict alphanumeric allowlist (/^[A-Za-z0-9][A-Za-z0-9._:\/-]*$/) to prevent shell injection.
Shell Usage: On Windows systems, the script uses shell: true to execute grok to ensure compatibility with NPM shims. While path arguments are quoted, the script serializes the command string, which could present a risk if the orchestrating agent passes unverified path strings containing shell metacharacters.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a bridge, passing natural language instructions ("briefs") to an external implementer (grok) that possesses significant workspace permissions.
Ingestion points: scripts/relay.mjs reads the task brief from files or standard input.
Boundary markers: Documentation encourages the use of XML tags (e.g., <task>, <verification_loop>) to define task boundaries and safety constraints for the implementer.
Capability inventory: The grok tool can perform file writes, execute shell commands, and access the network.
Sanitization: No sanitization is performed on the brief's natural language content; the skill relies on the implementer's internal safety protocols and a mandatory human review of the generated changes.
[EXTERNAL_DOWNLOADS]: The skill documentation recommends installing the grok CLI from the official @xai-official NPM registry. As a well-known and reputable service, this dependency is considered safe.

grok-delegate