offload

SUSPICIOUS: the skill is purpose-aligned as an offload/orchestration tool, but it intentionally delegates work to external AI harnesses with dangerous permission bypass and optional arbitrary custom commands. No clear credential theft or covert exfiltration appears, yet the combination of broad execution, relaxed approvals, and third-party prompt forwarding makes the overall security risk high rather than benign.

This excerpt is a process/command orchestration script that injects $BLOCK and other config-derived values directly into commands executed via herdr/tmux/Terminal or (for custom) bash -c. It also intentionally disables/weakens tool safety controls using “dangerously-bypass-approvals-and-sandbox” (codex) and “dangerously-skip-permissions” (claude). No clear exfiltration/keylogging/backdoor is visible, but the direct input-to-execution wiring and permission bypass create a high security risk if any of the inputs (CANDIDATES/custom/BLOCK/env/config) are attacker-influenced.