Then when bundling a hadoop target that hasn't messed with the classpath (which is extremely rare), we would put a classpath entry in the generated jar manifest telling it where to find its third party classpath on the filesystem