samuelfaj/distill v2.0.20 on GitHub

Highlights

Main and worker models: the main model (/model) plans, specifies, delegates and reviews; the new optional worker model (/worker-model) runs delegated work from precise specs. The reasoning tier and /reasoning-model are gone. On first start, [models].reasoning becomes the main model and the old main becomes the worker.
Effort for both tiers: /model <model> [effort|auto], /worker-model <model> [effort|auto] and the /tiers screen. auto lets Jev pick the effort for every call.
Worker-first delegation: fresh general-purpose and explore subagents run on the worker, start with a discipline reminder (do exactly the spec, stop and report on mismatches), and the main model reviews by diff and check output.
Caveman is always on: there is no off level. off in GROK_CAVEMAN, [agent] caveman or /caveman is ignored; /caveman lite|full|ultra still changes the intensity. It now reaches codex, custom and full-mode prompts and hosts that render only the subagent prompt, which had no terse style before.
Ponytail built in: every system prompt carries the lazy-senior-developer rules (reuse, standard library, one line, minimum code) adapted from ponytail (MIT). Subagents get a shorter variant that shapes how they work, not what the assignment asks.

Full Changelog: v2.0.19...v2.0.20