The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes the codex and claude CLI tools to delegate work between different agents. This activity is the core intended purpose of the skill and is governed by explicit user-defined arguments and script-level safety checks.
[PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface because it facilitates the execution of user-supplied instructions on a peer AI agent.
Ingestion points: The PROMPT_FILE argument passed to agent-delegate.sh acts as the primary ingestion point for untrusted instructions.
Boundary markers: While review mode utilizes the adversarial-review-prompt.md template to provide structure and instructions, the delegate mode passes the prompt directly to the peer agent without additional boundary markers or delimiters.
Capability inventory: Across its scripts, the skill utilizes codex exec, claude -p, git, and local file writing operations.
Sanitization: No specific sanitization or filtering is performed on the contents of the prompt file before it is passed to the peer agent via stdin.
This surface is integral to the skill's primary function of delegation and is managed within the host environment's security context.
[DATA_EXFILTRATION]: The script accesses local environment variables (e.g., CLAUDECODE, CODEX_HOME) and the ~/.codex/config.toml file to resolve agent configurations and verify workspace trust levels. These operations are confined to local configuration resolution and do not involve unauthorized external network communication.

agent-delegate