May 31 Stream, Telemetry, and Provider Migration

This guide maps the May 31 TypeScript AI SDK changes to the Go-AI SDK API names.

Stream Helpers

Use standalone stream helpers with result.Stream() for new code:

result, err := ai.StreamText(ctx, ai.StreamTextOptions{
    Model:  model,
    Prompt: "Write a concise release note.",
})
if err != nil {
    log.Fatal(err)
}
defer result.Close()

textStream, errStream := ai.ToTextStream(ctx, result.Stream())
for text := range textStream {
    fmt.Print(text)
}
if err := <-errStream; err != nil {
    log.Fatal(err)
}

For HTTP responses, prefer helpers that accept a raw provider.TextStream:

response, err := ai.CreateTextStreamResponseFromStream(ctx, result.Stream(), &ai.TextStreamResponseInit{
    Headers: map[string]string{"Cache-Control": "no-cache"},
})

StreamTextResult.ToTextStreamResponse, StreamTextResult.ToUIMessageStream, StreamTextResult.ToUIMessageStreamResponse, StreamTextResult.PipeTextStreamToResponse, and StreamTextResult.PipeUIMessageStreamToResponse are retained as deprecated compatibility wrappers.

Stream and FullStream

Stream() is the canonical Go accessor for the underlying provider stream. FullStream() remains available as a deprecated alias for TypeScript fullStream migration.

stream := result.Stream()

// Deprecated compatibility:
legacy := result.FullStream()
_ = legacy

UI Message Conversion

Use ToUIMessageChunk for one chunk and ToUIMessageStream for a full stream:

sendStart := true
sendFinish := true
uiChunks, uiErrs := ai.ToUIMessageStream(ctx, result.Stream(), ai.UIMessageStreamResultOptions{
    SendStart:  &sendStart,
    SendFinish: &sendFinish,
})

for chunk := range uiChunks {
    fmt.Printf("%s\n", chunk.Type)
}
if err := <-uiErrs; err != nil {
    log.Fatal(err)
}

The standalone stream helper is the Go equivalent of TypeScript's helper surface. Result-bound methods delegate to these helpers only for source compatibility.

Telemetry

Telemetry is the canonical field. ExperimentalTelemetry still works as a deprecated alias.

Telemetry integrations now receive abort events through OnAbort, and OpenTelemetry spans are closed when generation is canceled. Model calls run inside the telemetry context, so integrations that attach spans or values in OnStart can make them visible downstream.

ctx, cancel := context.WithCancel(context.Background())
defer cancel()
enabled := true

result, err := ai.StreamText(ctx, ai.StreamTextOptions{
    Model:  model,
    Prompt: "Stream until canceled.",
    Telemetry: &ai.TelemetrySettings{
        IsEnabled: &enabled,
    },
})
if err != nil {
    log.Fatal(err)
}
defer result.Close()

cancel()
_ = result.ConsumeStream()

Output timing fields use the TypeScript-compatible names:

• TimeToFirstOutputMs
• TimeBetweenOutputChunksMs

File outputs and tool calls count as first output events, matching the TypeScript SDK. Earlier token-specific naming is not canonical in Go.

Nested runtime/tool context objects are split into separate telemetry attributes when opted in through IncludeRuntimeContext or IncludeToolsContext.