samuelfaj/distill v2.0.29 on GitHub

Fixes

Auto effort (/effort auto) now applies Jev's choice on every model call. Before, a split or unsure answer ran the call at the session's pinned effort; with default_reasoning_effort = "max" that was 81% of the calls in a measured session.
The pick is the cheapest level that holds at least half of Jev's probability mass, so a split answer lands in the middle (medium) instead of on max. The keep_session_effort option and the 0.40 confidence floor are gone.
Effort levels without a catalog description (Claude models) reach Jev with a standard description of cost and use; catalog descriptions still win.
Worker subagents with auto effort get the same fix (Sonnet workers used to fall back to high).

Claude Opus 5.5, Opus 5, Sonnet 5.5, Fable 5.1 and Mythos 5.1 now receive the effort as a per-message marker (Anthropic beta mid-conversation-output-config-2026-07-01) instead of the top-level output_config.effort.
A top-level effort change used to rewrite the cached conversation; in a measured session the rounds that changed effort were 5.5% of the calls and 23% of the cost. With the marker, an effort switch read 98.6% of the prompt from cache.
Other models are unchanged.