Cookbook

Mature, runnable examples for every major capability, in the cookbook/ tree. Each recipe is one focused file; each app is a realistic multi-feature application. Everything runs offline with a deterministic TestModel (so it and its CI test need no key) and runs live by setting YAAB_SAMPLE_MODEL.

export YAAB_SAMPLE_MODEL=gemini/gemini-2.5-flash   # any LiteLLM model id
python -m cookbook.recipes.agents
python -m cookbook.apps.support_desk

Flagship apps

Realistic applications that combine several capabilities end-to-end.

App	Scenario	Capabilities
support_desk	Grounded, governed, safe-to-act support	RAG · tools · HITL approval · per-tenant spend caps · OpenAI-compat serving
refund_pipeline	Durable workflow with a human gate	Flow · pause→decide→resume · durable checkpoint + approval store
research_assistant	Durable multi-step research	Flow · hybrid (BM25+dense) retrieval · streaming · browser + record/replay (live)
coding_agent	Plan, then run code in a sandbox	multi-agent (planner→coder) · sandboxed exec · tool approval

Recipes

One focused, self-checking example per capability (python -m cookbook.recipes.<name>).

Recipe	Capability	Docs
agents	The typed Agent unit of work	Agents
tools	Give an agent a typed function to call	Tools
typed_output	Get a validated Pydantic object back	Agents
sessions	Durable, multi-turn conversations	State
memory	Long-term recall, scoped per user	State
rag	Ground answers in your documents	RAG
hybrid_retrieval	Fuse sparse (BM25) + dense recall	RAG
streaming	Token-by-token output	Streaming
sequential	Fixed steps, each building on the last	Multi-agent
parallel	Fan out the same input to several agents	Multi-agent
map_agent	One agent across many inputs	Multi-agent
loop	Re-run until a quality bar (or a cap)	Multi-agent
swarm	Autonomous hand-off between peers	Multi-agent
conditions	Guard steps with when=/stop=/else_=	Conditions
router	Deterministic exclusive choice	Conditions
flow	Explicit, durable control flow	Flow
hitl	Gate a sensitive tool on approval	Human-in-the-loop
durable_runs	Runs as durable records; multi-pod	Durable runs
spend_governance	Per-identity / per-tenant budget caps	Governance
governance	Guardrails that block unsafe input	Governance
evaluation	Score outputs against expectations	Evaluation
optimization	Tune a prompt at build time, freeze it	Optimization
serving	Expose an agent over HTTP	Serving
openai_compat	Serve /v1/chat/completions	OpenAI-compatible API
record_replay	Deterministic offline tests of real runs	Record & replay
browser	Drive a real browser, safely	Tools
interop	Expose YAAB tools over MCP	Interop
harness	Sandboxed, approval-gated coding agent that edits a file	Coding harness

How they're tested

tests/test_cookbook.py auto-discovers every recipe and app and runs it offline, and each one self-checks its result — so an example that imports clean but does the wrong thing fails CI.