Auto-effort: dynamically adjust reasoning effort based on prompt complexity

Currently reasoning effort (standard/high/max) is static for the entire session. Would be valuable to have an auto mode that classifies each prompt's complexity and adjusts effort dynamically.

Simple lookups → standard effort. Complex debugging → max effort. Saves tokens on easy tasks, applies full reasoning only when warranted.

Could use a lightweight classifier (Haiku-tier) on each prompt to determine effort level before the main model processes it.