Kaya gets syntax highlighting

Every Robot Book Club book is generated by a program written in Kaya — a small scripting language built for orchestrating LLM work. Up to now Kaya code was just gray monospace everywhere it showed up: in the editor, in the repo, and in these posts. This week it got real editor support, and the same highlighter is now wired into this blog. Here's a shard of an actual cookbook pipeline, highlighted:

// A shard of a Robot Book Club cookbook pipeline.
import generateRecipe from "./generate-recipe"

type Recipe = {
  title: string + { max: 80 },
  servings: number + { min: 1 },
  intro: string + { is: "one warm sentence about the dish" }
}

#cached
function buildChapter(cuisine: string, count: number): Recipe[] {
  let recipes: Recipe[] = []
  for i in range(count) {
    let recipe = generateRecipe(cuisine: cuisine, model: "gemini-2.5-flash")
    recipes.push(value: recipe)
    "Drafted {{recipe.title}} ({{_index}} of {{count}})"
  }
  return recipes
}

[narrator] "The kitchen never sleeps."
let chapter = buildChapter(cuisine: "Sicilian-American", count: 12)
save(content: chapter, filename: "data/{{cuisine | slug}}/chapter.json")

A few things in there are pure Kaya and worth pointing at, because they're why the language exists:

• Refinement types. string + { max: 80 } is a string with a length cap; number + { min: 1 } is a positive number. And string + { is: "one warm sentence about the dish" } is the fun one — a semantic constraint the model is asked to satisfy, not a regex. The type system and the prompt are the same thing.
• {{ }} interpolation with filters. {{cuisine | slug}} runs the value through a slug filter inline — the same templating that keeps recipe filenames from blowing past the filesystem's name limit.
• Speaker tags. [narrator] "..." emits a line into the conversation as a named speaker. It reads like a screenplay because, often, that's what these programs are.
• Named arguments everywhere. generateRecipe(cuisine: ..., model: ...) — Kaya calls are always keyword args, which is what makes a program full of LLM calls legible six months later.

What actually got built

There are four pieces, and they all stay honest because they're checked against one shared test suite — a corpus of annotated Kaya snippets that asserts, token by token, what color each thing should be. Both highlighters are run against the same corpus, so VS Code and Neovim can't drift apart:

• A rebuilt TextMate grammar for VS Code. The old one had quietly rotted — it was highlighting a dialect Kaya doesn't use anymore (an @function() call syntax that's been dead for months, a set keyword that's now let/const, an old string() type form). It's now rebuilt against the real grammar and verified by the corpus.
• A tree-sitter grammar for Neovim, which can't use TextMate grammars at all.
• A smarter language server (LSP) that powers autocomplete and inline error squiggles in both editors — now offering the real keywords, the real built-in functions (generate, save, pick, …), and the real type forms.
• The web highlighter that rendered the block above — the very same grammar, run on the server, so a fenced Kaya block in one of these posts comes out colored.

The whole thing is tests-first: the corpus was written before the grammars were fixed, so "make Kaya highlight correctly" became a concrete, green-or-red target instead of a matter of squinting at a screenshot. 111 highlighting assertions, all passing, across both editors.

It's a small thing — colors on code — but it's the kind of small thing that makes the rest of the work faster. The cookbooks are written in this language; making the language pleasant to read is making the factory floor easier to walk.