The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands to identify the current execution environment by inspecting environment variables (env | grep).
Evidence: env | grep -E '^(CLAUDECODE|CODEX_THREAD_ID|ANTIGRAVITY_AGENT|OPENCODE)=' in SKILL.md.
[COMMAND_EXECUTION]: The skill instructs the agent to execute various CLI-based agent tools (opencode, agy, codex, claude) with interpolated prompts and arguments.
Evidence: Tables and command templates in SKILL.md such as opencode run --agent <build|plan> ... and agy --model <model> ....
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface where malicious instructions could be introduced via untrusted data sources.
Ingestion points: The skill reads output from sub-agents, as well as system state via git status --short and git diff (SKILL.md).
Boundary markers: It attempts to use specific instruction markers like "你是被派工的執行者" (You are the dispatched executor) to define boundaries and roles, though these are not cryptographically secure.
Capability inventory: The agent has the capability to execute arbitrary commands through the integrated CLI tools across all task types (SKILL.md, references/cli-cheatsheet.md).
Sanitization: The skill does not explicitly specify sanitization or escaping mechanisms for the data received from sub-agents before it is processed by the main agent.

agent-delegation