The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/relay.mjs script executes the Command Code CLI (cmd) and git binaries to perform tasks, check repository status, and verify changes. All command-line arguments derived from user input are strictly validated against restrictive regular expressions to prevent command injection.
[DYNAMIC_EXECUTION]: When the --lane option is utilized, scripts/relay.mjs executes a sibling script (delegate-setup/scripts/lane.mjs) using the Node.js runtime to resolve configuration dials. This is a controlled execution of local code required for the skill's fleet management functionality.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted instructions (the task brief) which are then executed by a delegated AI agent.
Ingestion points: scripts/relay.mjs reads the task brief from a file path or stdin.
Boundary markers: references/writing-the-brief.md provides a structured XML-based template (<task>, <verification_loop>, etc.) to isolate different types of instructions and minimize obedience to embedded malicious data.
Capability inventory: The delegated agent (cmd) is typically executed with --yolo permissions, allowing full file system access and shell execution within the working directory.
Sanitization: The relay script validates all CLI argument tokens and pipes the brief content directly to the child process's stdin, effectively preventing it from being interpreted as shell command components.
[DATA_EXPOSURE]: To provide security monitoring during read-only runs, the script calculates SHA-256 fingerprints of files in the working directory and monitors git status output. This data is used locally to report potential 'read-only violations' to the user and is not exfiltrated.

commandcode-delegate