I see "direct manipulation of output" and some boredom/frustration from reading internals of the source code, and I think the solution you seek is a kind of rep

I see "direct manipulation of output" and some boredom/frustration from reading internals of the source code, and I think the solution you seek is a kind of repl LLM, that would understand and rewrite on user input every internal code in order to produce the effect you may want.