Your prompt runs before every command, so whetū renders in about 2 ms. One compiled binary, no config file, nothing to theme, nothing to tune — the same curated prompt for everyone. Git branch, state and status. Toolchain versions for 39 languages. A syntax-highlighted history picker on the up arrow. Written in Zig, every disk probe running at once, so a render costs what its slowest one costs. No network, no telemetry, no account. macOS and Linux, on fish, bash and zsh.
Formal verification for AI-generated code. Automatically extracts correctness properties from pure functions, translates them into Lean 4 theorems, and machine-checks them with Mathlib — so you get mathematical proof, not just tests, for the logic your AI coding agent produces.
Works with any LLM — Claude, GPT-4, Gemini, Llama, Mistral, or any OpenAI-compatible endpoint.
