Request encoders translate the provider-neutral inference.Request after feature validation. RequestModeInvoke and RequestModeStream are typed values, not an unstructured boolean.

Mode

type RequestMode uint8

const (
	RequestModeInvoke RequestMode = iota
	RequestModeStream
)

func (Codec) EncodeRequest(
	req inference.Request,
	mode codec.RequestMode,
) (codec.EncodedRequest, error)

OpenAI Chat, OpenAI Responses, and Anthropic put streaming in the JSON body. Gemini and Bedrock use the same JSON body in both modes; their route and response framing select streaming. Every bundled encoder returns Content-Type: application/json and a body reader.

Validation

Encoders call inference.ValidateRequestFeatures before marshaling. That check rejects a transient-message count outside the message slice, a named tool choice whose name matches no declared tool, required tool choice without tools, unsupported image input, invalid structured-output schemas, duplicate tool names, and capabilities the model does not advertise. Dialect encoders then reject blocks they cannot represent, returning a typed UnsupportedBlockError or equivalent rather than dropping data.

body, err := openaiapi.EncodeRequest(req, false)
if err != nil {
	var unsupported *openaiapi.UnsupportedBlockError
	if errors.As(err, &unsupported) {
		// Decide whether to remove the block or choose another model.
	}
	return err
}
_ = body // JSON bytes are consumed once by the transport.

Source and proof

Run go test ./codec/openaiapi ./codec/openairesponses ./codec/anthropicapi ./codec/geminiapi ./codec/bedrockconverse.