Bridging dissimilar message busses with data-driven Lua scripts. Creating a Jenkins SCM plugin that exposes the sources packages of a Debian repository (complete with binary dependency tracking) as an organization folder to turn it into a package builder. Improvising a Git proxy/cache with a hundred lines of Bash to lessen the load on the lab uplink (still load-bearing to this day). Writing a toolchain in Python that takes OpenAPI documents as an abstract syntax tree and run passes on it to parse, validate, aggregate, transform and format it for various needs (such as generating C++ code for data models, dataframe bindings and so on). Delinking programs back into object files and creating Frankenstein monsters from salvaged pieces, and somehow landing a poster presentation about that at ACM CCS 2025 as a hobbyist (this one outside of office hours, but it still triggers brain meltdowns when I talk about it). And so many, many more sins.
I honestly don't know how they are putting up with the incarnation of chaos that is me.
I'm in the process of writing a toolchain of sorts, with the OpenAPI document as an abstract syntax tree that goes through various passes (parsing, validation, aggregation, analysis, transformation, serialization...). My immediate use-case is generating C++ type/class headers from component schemas, with the intent to eventually auto-generate as much code as I can from a single source of truth specification (like binding these generated C++ data classes with serializers/deserializers, generating a command-line interface...).
JSON schema is so flexible that I have several passes to normalize/canonicalize the component schemas of an OpenAPI document into something that I can then project into the C++ language. It works, but this was significantly trickier to accomplish than I anticipated.