The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions direct the agent to execute the external 'cmd' CLI with the '--yolo' (alias '--dangerously-skip-permissions') flag. This mode is explicitly designed to bypass all safety filters and permission gates, allowing the process to perform unrestricted shell commands and filesystem operations.\n- [PRIVILEGE_ESCALATION]: By promoting the use of '--yolo' mode, the skill enables the sub-agent to operate outside of its intended security boundaries. The documentation acknowledges that this mode has 'no filesystem sandbox and no path restriction,' effectively granting full system privileges to the automated process.\n- [EXTERNAL_DOWNLOADS]: The skill requires the installation and authentication of the third-party 'Command Code' CLI from commandcode.ai. Furthermore, it references an execution script ('scripts/relay.mjs') that is not provided in the skill package, which may prompt the agent to search for or fetch unverified code from external community repositories.\n- [INDIRECT_PROMPT_INJECTION]: The skill implements an orchestration loop where data and instructions are passed to an external agent. This creates a high-risk ingestion surface for indirect prompt injection, as the orchestrator is instructed to process the output and working tree changes produced by the delegate. The provided analysis checklist is a manual mitigation for an inherent architectural vulnerability.\n- [NO_CODE]: The skill package contains no executable code or scripts, despite the instructions providing command-line examples for a specific internal script. This 'docs-only' nature with references to missing functionality is a deceptive pattern that increases the risk of the agent interacting with untrusted external resources.

commandcode-delegate