Rozenite Agent SDK

Use this skill when the user wants code-first access to Rozenite for Agents.

Read references/code-patterns.md for copy-pastable examples.

Rules

• Prefer throwaway Node ESM scripts that import @rozenite/agent-sdk.
• Default flow: createAgentClient() -> client.withSession(...) -> inspect or call tools -> exit.
• Use session.domains.list() as the source of truth for live built-in and runtime domains.
• Use session.tools.list({ domain }) and session.tools.getSchema({ domain, tool }) only when you need live inspection or argument confirmation.
• Prefer typed plugin SDK descriptors from any available official plugin ./sdk export and call tools as session.tools.call(descriptor, args).
• Fall back to session.tools.call({ domain, tool, args }) only when no matching ./sdk descriptor is available or the tool is discovered dynamically at runtime.
• When you discover tools through session.tools.list({ domain }), treat the returned shortName as the canonical tool name to pass back into a later call-by-name invocation. Do not invent camelCase aliases or normalize tool names yourself.
• Prefer stable SDK domain identifiers such as built-in domain IDs (network, react, memory) and plugin IDs (@rozenite/storage-plugin, @rozenite/tanstack-query-plugin) over CLI-only live domain tokens like at-rozenite__storage-plugin.
• If paged results should be merged automatically, use autoPaginate.
• If a plugin only mounts after navigation, navigate first, then refresh the live view with session.domains.list() or session.tools.list(...) before calling the plugin tool.
• For advanced session control with client.openSession() or client.attachSession(sessionId), see the reference patterns.
• If a script encounters an unexpected runtime error, let the script fail clearly. Do not hide the failure by printing placeholder JSON.

Handoff

• Use rozenite-agent instead when the task is shell-driven, needs a reusable CLI session, operates directly through rozenite agent ..., or requires target enumeration before choosing a deviceId.