Choosing a recipe

A recipe is picked along two axes: what it evolves, and how it learns.

	Reactive: learns from the traffic it already serves	Proactive: generates its own attempts
Model weights	sao: feedback on each attempt over a stream of tasks openclawrl: multi-turn traffic, reward read from the next state	tttd: repeated attempts at one problem, at test time
Harness: prompts, rules, skills, config	skillclaw: grows a skill pool from the failures in its own served traffic gepa: rewrites the tree by reflecting on the transcripts it already served	not available

Reactive: learns from the traffic it already serves

Proactive: generates its own attempts

Model weights

sao: feedback on each attempt over a stream of tasks

openclawrl: multi-turn traffic, reward read from the next state

tttd: repeated attempts at one problem, at test time

Harness: prompts, rules, skills, config

skillclaw: grows a skill pool from the failures in its own served traffic

gepa: rewrites the tree by reflecting on the transcripts it already served

not available

Pick by the signal your workload can produce.

The signal you have	Recipe	Evolves	Needs GPUs
Feedback on each attempt, over a stream of tasks	sao	model weights	yes
A fixed grid of sibling attempts at one problem	tttd	model weights	yes
Agent conversations without reports	openclawrl	model weights	yes
Feedback on individual requests, and failures worth learning from	skillclaw (built into Reef)	harness tree	no
A score per request, and a stronger model to reflect with	gepa (built into Reef)	harness tree	no

How a recipe is selected #

A deployment serves exactly one recipe, named by reef.recipe in its config. Every scenario it creates uses that recipe. Requests never name a recipe, and scenario snapshots do not store one. The scenario header is the only routing a caller provides. The artifact repository is therefore deployment-owned: do not point deployments configured with different recipes at the same repository.

reef:
  recipe: recipes.sao.recipe:SAORecipe
  batch_size: 1

reef.recipe accepts the core value recipe, a dotted class, or a preset. Reef does not register or import learning methods. The recipes/ tree in this repository is a cookbook; installed method packages work the same way. Configuration describes each spelling.

Every recipe has a checkpoint strategy, defaulting to EveryNVersions(1). checkpoint_every_n_versions is the shorter spelling in deployment YAML.