Streaming usage is terminal metadata, not a chunk. A stream decoder accumulates provider usage while chunks are read and publishes a stream.StreamResult only after Next observes clean io.EOF.

Terminal

for {
	_, err := reader.Next()
	if errors.Is(err, io.EOF) {
		break
	}
	if err != nil {
		return err // usage is not authoritative after a failed stream
	}
}
result, ok := reader.Result()
if ok && result.Usage != nil {
	fmt.Println(result.Usage.OutputTokens)
}

Result is false before EOF, after any non-EOF failure, or when the provider did not report terminal metadata. Anthropic requires message_stop; OpenAI uses its terminal usage event and [DONE]; Responses uses completed events; Gemini and Bedrock collect their final usage events. Missing terminal markers do not become fabricated usage.

Ownership

StreamReader.Result returns a fresh copy of Usage on every call. Close is idempotent and must still be called after EOF to release the response body. The retry decorator stamps establishment attempts into StreamResult.Attempts without changing mid-stream failure semantics.

Source and proof

Run go test ./stream ./codec/....