Reward design: agent prompts, full grading with judge prompts, and a Reward design page (source 1f6b1ba)

Compare rollouts of one task side by side

Warn about missing sandbox credit before a run; only finished, scored rollouts in Community

Live progress while the model writes, thinking low by default, stuck chart tooltip

Public and private rollouts, community, provenance, repository browser, Xiaomi-matched configuration

Bring your own endpoint, run settings, thinking in traces

Redesign, token sign-in, new name and logo

Explorer with rollouts: Docker Space, HF sign-in, private bucket for traces