OpenEnv gets a capture proxy for training open models across coding harnesses

A post says OpenEnv’s capture proxy lets users train open models with reinforcement learning across coding harnesses without modifying them. Using the same LFM2.5-2.6B weights, the author reports 62% of held-out tasks solved in mini-swe-agent versus 33% in Claude Code. After training across four harnesses, the reported average rose from 42% to 54%, and the Claude Code result from 33% to 49%.