cookbook & recipes

A cookbook is the shelf every recipe in a project sits on. A recipe is the individual item on that shelf: a saved, repeatable procedure built once from Kitchen nodes, then run again on demand without being rebuilt from scratch. A cookbook holds recipes; a recipe holds nodes.

Building a recipe once and reusing it is the point of the split. Reasoning that would otherwise get redone by hand every session gets wired once on the Kitchen canvas — prompts, gates, asks, or the newer node kinds — and saved as a callable procedure. What ships to the catalog is exactly what a recipe run pulls back out, node for node.

Every recipe is built from Kitchennodes, and today those nodes come from two generations. The older vocabulary is three primitives — prompt, ask, gate — enough to chain a question, wait on an answer, and branch on a condition. It’s what most of the catalog is actually built from. The newer vocabulary is fifteen distinct v2 node kinds, built to cover more of what a recipe can do than prompt, ask, and gate ever tried to.

A recipe doesn’t have to touch the newer vocabulary to work. The three-primitive set closes plenty of loops on its own — a prompt that asks a question, a gate that waits for a yes — and nothing about the newer kinds retires it.

Fourteen recipes live in the catalog right now (#13048), confirmed by calling the running endpoint that serves them, not by counting a spec. Thirteen of the fourteen are built entirely from the older three-primitive vocabulary: prompt, ask, gate covers the whole recipe, start to finish, for all thirteen.

That leaves a real gap between what the catalog could look like and what it holds today. One recipe uses the newer vocabulary. The other thirteen do the same job the older vocabulary has always done, and they do it fine. The newer node kinds are proven on one recipe, not standard across the catalog yet.

An earlier read of the catalog counted higher: eight additional “primitive” recipes, plus a recipe that wrote other recipes, a self-styled Chef’s Table. Neither was ever seeded into production. The correction landed as #12503, and the real count held at fourteen, re-confirmed live (#13048).

One recipe in the catalog, scout-and-brief, is built differently from the other thirteen. It runs fifteen nodes, and those fifteen nodes span all fifteen distinct v2 node kinds: not a sample of the newer vocabulary, the entire set, in one graph (#13048).

It’s walker-eligible, and that’s confirmed by static analysis of its live graph: a check of the graph’s shape, not a completed run. Nobody has taken scout-and-brief start to finish in production yet. What’s verified is that it’s built correctly enough to; what’s not yet verified is what happens when it does.

The count above didn’t come from a spec or a slide. It came from calling GET /api/cookbook/recipeson the running server and counting what came back (#13048): fourteen recipes, thirteen on the older vocabulary, one, scout-and-brief, spanning the newer one. The catalog once carried a bigger number: eight extra “primitive” recipes and a recipe-writing “Chef’s Table” recipe, both confirmed never seeded (#12503). The corrected count is the one still standing.