Code Step Author

What is Duvo?

Duvo is an AI-powered automation platform that handles repetitive business work across the systems a team already uses. An automation is built from two kinds of node, told apart by the guarantee each makes rather than the work it does:

• An Agent is judgement — an AOP (the markdown procedure that becomes its prompt), evals, memory, skills, and all human interaction.
• A Code Step is deterministic compute — a Python program the platform invokes. It reads and writes one case's data, runs exactly once per dispatch, and fails the case if it fails. No model runs when it executes.

A Code Step's versions are Builds and its executions are Runs, exactly as an Agent's are. Where an Agent's Build holds an AOP, a Code Step's Build holds source and an entrypoint.

What you're doing

The user is writing, reading, or debugging the program inside a Code Step, and needs to know what the duvo client can actually do. That surface has no autocomplete and no generated type declarations — it is only knowable from prose, and there is nothing in the public product documentation that covers it. That is what this skill is for.

Read references/sdk.md before answering anything specific. It is the same text the step's own AI generation is given, generated from one source, so an answer grounded in it and a program Duvo writes cannot disagree.

The one rule that matters: never invent SDK surface. A plausible-looking method that does not exist produces a step that fails on its first dispatch, and the author has no way to tell that from a real bug. If the user needs something references/sdk.md does not name, say it is not in the SDK rather than guessing at a neighbouring method.