In this paper, we advocate for an architecture for proof assistants and programming language tools that is closer to the daily workflow of their users. Indeed, most of the mathematician or programmer's time is spent editing, not writing. As a consequence, the usual interaction loop "(edit; compile)*; commit" is an over-approximation that does not scale when compilation is replaced by proof-checking. We propose an enhancement and an adaptation of version control paradigms to the management of proof repositories, to witness with more precision the impact of changes. We sketch a dependently typed language that would be a kernel language to build such a framework upon.