Repository navigation
What's new
- Auto effort now picks the most efficient level for every main model call: the lowest level that gets the step right the first time, then the fastest, then the cheapest. Jev re-decides on every call and can move up or down.
- Each execution mode biases the pick. UltraSave leans lower, UltraFast favors faster levels and avoids xhigh/max, UltraCode leans higher for planning, design and review, and Normal stays balanced.
- The effort question includes a short note on how each provider's levels behave (Anthropic, OpenAI, xAI, DeepSeek, Gemini, GLM).
- Fixed: when Jev could not answer, a session configured with
default_reasoning_effort = "max"ran every call at max. Distill now lowers that call to the mode baseline (low for UltraSave/UltraFast, medium for Normal, high for UltraCode). It never raises an effort that is already lower. - In UltraSave and UltraFast, Claude models with per-message effort now re-pick effort on later rounds of a turn, because switching does not break their prompt cache.
- Effort decisions and fallbacks are written to
unified.jsonl. - Benchmark aggregates report the main-call effort mix and the number of effort switches.
Benchmarks for this release ran while Jev was unavailable, so they cover the fallback path only. Details are in PR #48.