Condense.chat's proxy compresses coding-agent context with two in-house models, cutting token bills by up to 72 percent on deep sessions.