Inference streaming has two layers: a wire framer yields StreamFrame values, then a semantic decoder maps frames to sealed content.Chunk values. A StreamReader exposes both as a pull iterator and keeps terminal metadata separate from the last chunk.

Stream lifecycle

%%{init: {"theme":"base","themeVariables":{"background":"#111827","primaryColor":"#1f2937","primaryTextColor":"#f8fafc","primaryBorderColor":"#64748b","lineColor":"#94a3b8","secondaryColor":"#0f172a","tertiaryColor":"#172033","fontFamily":"Inter, ui-sans-serif, system-ui, sans-serif"}}}%%
sequenceDiagram
    participant C as Caller
    participant R as StreamReader
    participant P as Provider body
    C->>R: Next()
    R->>P: read frame
    P-->>R: Chunk
    R-->>C: content.Chunk
    C->>R: Next() until io.EOF
    R-->>C: clean EOF and Result()
    C->>R: Close()

Always call Close, even after clean EOF. Any non-EOF failure permanently fails the reader and makes Result unavailable. Clean EOF is the only point at which a terminal result producer is consulted.

reader, err := client.Stream(ctx, request)
if err != nil {
	return err
}
defer reader.Close()
for {
	chunk, err := reader.Next()
	if errors.Is(err, io.EOF) {
		break
	}
	if err != nil {
		return err
	}
	consume(chunk) // render or accumulate the provider-neutral chunk
}

The Harness streaming response step is the canonical consumer when a Harness loop renders these chunks.

Overview

Stream lifecycle

Proof