The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Python's subprocess module to manage repository state and execute sub-agent tasks. It primarily interacts with git and a codex binary. To mitigate risk, the skill provides sophisticated management of these executions, including serialized locks for write operations and process group termination logic to handle timeouts.
[PROMPT_INJECTION]: The main agent orchestrates tasks by interpolating user instructions into role-specific templates (explorer, worker, verifier). While this creates an indirect prompt injection surface, the skill implements defensive boundaries, such as appending "Do not delegate further" and "Stay within the exact task and authority" to all delegated prompts, reducing the risk of sub-agent misbehavior.
[SAFE]: The skill demonstrates a security-first design by implementing a "guarded fallback" mechanism. Instead of allowing sub-agents to modify the primary repository directly, it clones the repository into an isolated candidate worktree, performs edits there, and requires the main agent to manually review and integrate the resulting patch bundle. Furthermore, it explicitly validates that symlinks do not point outside the repository boundaries before allowing write operations.

gpt-engineer-spark