• Changed: with a worker model configured, the main model hands multi-step work to the worker and reviews it, whichever model it is (#67). Before, DeepSeek and Opus mains did the work themselves. The turn-start delegation hint now covers any request that takes more than one call, not only one with parallel parts; the prompt delegates creative and from-scratch work and the investigation of an unexplained failure; and a main model that keeps working hands-on is reminded every 8 rounds. A one-call request is still done directly. Off: [jev.ladder] b6_delegation_hint = false.
• Fix: a pinned mode without subagents (UltraFast) no longer carries "delegate by default" in its instructions with no tool to delegate with.
• Measured on one prompt (DeepSeek V4.1 Flash main, GLM 5.3 Flash worker, one run per mode): main-model calls fell from 175 to 92 in Normal and from 200 to 108 in UltraCode, and cost from $0.142 to $0.116 and from $0.207 to $0.096. Wall time rose in Normal (44 to 58 min) and fell in UltraCode (64 to 39 min). Not rerun on Opus or GPT mains.