Concrete examples are needed for live programming while editing abstract code. However, coming up with examples takes effort and usually happens after writing code. But there are ways to execute code without specifying examples, such as fuzzing. Fuzzing is a technique that synthesizes program inputs to find bugs in security-critical software.

While fuzzing focuses on finding crashes, it also produces valid inputs that can serve as examples for developers. Our approach is to make use of this and show examples including edge cases directly in the editor. For this, we have to make fuzzing faster and make it produce more natural examples. To implement fuzzing and make it fast enough, we implement it on the granularity of functions. We integrate it into the compiler pipeline and language tooling of Martinaise, a custom programming language with a limited feature set. Initially, our examples are random, and then mutate based on coverage feedback to reach interesting code locations and be more pleasing to humans.

We evaluate our tool in small case studies, showing generated examples for numbers, strings, and composite objects. Our fuzzed examples still feel synthetic, but since they are grounded in the dynamic behavior of code, they can cover the entire execution and show edge cases.