Builds the adversarial chat messages.
The model is asked to find the assumptions a piece of code makes that could break, tied to specific lines, and to answer with a strict JSON array so the result is machine-parseable. The skepticism tier only swaps one phrase: how aggressively to flag.
Summary
Types
Functions
@spec messages(String.t(), String.t(), skepticism()) :: [ %{role: String.t(), content: String.t()} ]