ADE Cuts CLI Skill Size, Sonnet Holds

This is a strong case for routing agent context instead of dumping it wholesale, though the result remains an internal 14-task comparison.

• –On-demand references preserve specialized guidance without bloating every agent prompt.
• –Subagent-by-default behavior and one default per scenario reduce contradictory CLI instructions.
• –Cross-chat agent controls create a more coherent mesh for managing parallel work.
• –Typechecks passed, but broader quality tests and the GPT-6 Luna rerun were still pending.
• –No performance impact was measured, and remote-machine subagent reporting remains incomplete.