You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
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.