nemo-relay-instrument-typed-wrappers

Use Typed Wrappers And Codecs

Use this skill when an application wants stronger domain types than raw JSON for tool or LLM integration. Keep typed boundaries explicit so middleware still sees predictable JSON.

Default Guidance

• Prefer plain JSON first for initial adoption.
• Reach for typed wrappers when the application already has stable domain models.
• Keep in mind that middleware still operates on JSON, not typed objects.

Embedded Codec Model

• A typed value codec is a pure boundary translator. It converts application-facing values to JSON before NeMo Relay emits events or runs middleware, then converts JSON back into the framework callback or caller type.
• Python exposes JsonPassthrough, DataclassCodec, PydanticCodec, and BestEffortAnyCodec. Node.js exposes JsonPassthrough plus custom Codec<T> implementations.
• Use BestEffortAnyCodec only at boundaries where strict schemas are not available. Prefer dataclass, Pydantic, or explicit Node.js codecs when the framework owns a stable schema.
• Provider codecs are different from typed value codecs: they normalize provider-specific LLM requests and responses so middleware and subscribers can inspect messages, tools, model names, generation parameters, and response annotations.
• Built-in provider codecs include OpenAIChatCodec, OpenAIResponsesCodec, and AnthropicMessagesCodec in Python, Node.js, and Rust. Choose the codec that matches the actual provider payload shape.
• Response codecs annotate LLM end events with fields such as id, model, message, tool_calls, finish_reason, usage, provider-specific data, and extra unmodeled fields. They do not rewrite the caller-visible response.
• Request codecs run before LLM request intercepts. Intercepts receive both the raw LLMRequest and optional annotated request; encode merges annotated edits back before execution intercepts and the provider callback run.
• Built-in request codecs guarantee JSON-value identity for an unchanged annotation. They compare edits with a decoded baseline and patch only changed fields, preserving native representation details and unknown fields.
• Use instructions, portable messages and components, and the tagged api_specific request surface for normalized edits. Provider-only union members use explicit { provider, kind, value } native components. Reserve top-level extra for unknown future fields.
• The nemo-relay gateway always supplies matching request codecs on Anthropic Messages, OpenAI Chat Completions, and OpenAI Responses generation routes. On those routes, treat raw request.content as read-only and return body edits through the annotated request. Header edits still use the raw request.