Cookbook

The Cookbook is a collection of short, task-oriented recipes that cut across multiple Zeroth subsystems. Each recipe is around 200 words, has a When to use / When NOT to use block so you can scan for fit in under ten seconds, and embeds a real runnable file from examples/ so you can copy, paste, and run it locally.

If you're looking for an overview of a single subsystem, start in Concepts. If you're looking for a guided walkthrough of a single subsystem's API, see the Usage Guides. The Cookbook is where you come when you already know the pieces and want to see them snapped together.

Recipes

• Add a human approval step to a node — pause a node until a human approves, then resume.
• Attach memory to an agent — wire a memory connector so an agent can carry state across turns.
• Cap a run's cost budget — estimate LLM cost with CostEstimator and block calls that would exceed a per-run cap.
• Sandbox a tool call — run a tool in an isolated, allowlisted environment with SandboxManager.
• Retry a failing webhook with backoff — jittered exponential backoff, dead-letter thresholds, and HMAC verification.
• Block a tool call via policy — deny a Capability before the tool fires with PolicyGuard.
• Query the audit trail for a run — walk every NodeAuditRecord for a run with AuditRepository.
• Hand off between two agents mid-graph — chain a researcher agent into a writer agent.
• Branch execution on a condition — evaluate edge conditions with BranchResolver + ConditionContext.
• Inject a secret into an execution unit — resolve secret refs into env vars and redact them from audit.
• Govern LangGraph tool calls — decide every tool call with govern_tools or ZerothMiddleware, plus the compatibility matrix and what tool enforcement does not claim.
• LangGraph enforcement evidence — how a run or deployment is classified admission / observed / enforced, the 90-second stale threshold, and mixed-version handling.