I would suggest a backtracking LS coupled with a filter strategy. A filter forces one of two quantities (objective and constraint violation) to decrease and, together with the LS, enforces global convergence.
You can have a look at my latest slides about the NLP solver I’m developing: https://www.researchgate.net/publication/354921761
However, I’m not aware of any libraries that provide these components (yet. My C++ framework will allow that but hasn’t been published yet).