Recipes

These overlays are opt-in. Neither one replaces the default pipeline. The factory does not switch a default on its own.

RoutingCopy Anchor Link

Routing is for skipping spec review on a small change. Everything else in the pipeline stays. Small means the classified size is at most small_max (default S). A classifier failure runs the full pipeline.

Reviewer modelCopy Anchor Link

Reviewer-model adds a spec review that passes config.model through to the user's Claude Code. Replace the placeholder with a setting that Claude Code accepts. The factory does not name a model. An empty model keeps that user's default.

Before either recipe replaces a defaultCopy Anchor Link

Before either recipe replaces a default, run kusaw bench run against baseline on at least 3 benchmark tasks with --repeat at least 3, then kusaw bench report <batch> --markdown, and read the table. The factory does not switch the default itself.

Installing a recipeCopy Anchor Link

kusaw bench init-recipes copies the two overlays into <data_dir>/benchmarks/variants/ for kusaw bench run <task> --variants routing (or reviewer-model). Using one in production means merging that file into kusaw.yaml.

Provider as a benchmark variantCopy Anchor Link

Provider is a benchmark variant the same way routing and reviewer-model are: an overlay file in <data_dir>/benchmarks/variants/, then kusaw bench run <task> --variants cursor (or another name). Enable Cursor on the machine first; a kusaw.yaml overlay cannot turn it on. Existing factories and the two shipped recipes stay on Claude Code until you add a variant.

A whole-factory overlay:

provider: cursor

One step:

steps:
  review:
    provider: cursor

Compare that variant against baseline with kusaw bench run on at least 3 tasks and --repeat at least 3, then kusaw bench report <batch> --markdown. The factory does not switch a default on its own.