IChatClient.GetStreamingResponseAsync yields nothing: streamGenerateContent is requested without ?alt=sse but parsed as SSE

[HavenDV]

Symptom

IChatClient.GetStreamingResponseAsync on GeminiClient (Google_Gemini 0.11.2) completes without yielding a single TextContent against the live API, while GetResponseAsync answers the same prompt normally. Observed 2026-09-15 from a production backend: every streamed call ended with an empty aggregate (gemini returned an empty response), model gemini-3.5-flash-lite, plain-text prompt, ThinkingLevel = Minimal via RawRepresentationFactory.

Root cause (from the generated code)

Google.Gemini.GeminiClient.ModelsStreamGenerateContentAsStream.g.cs builds the request as

var __pathBuilder = new PathBuilder(path: $"/models/{modelsId}:streamGenerateContent", baseUri: HttpClient.BaseAddress);
// … only the ApiKey query authorization is appended …

and then reads the body with

await foreach (var __sseEvent in System.Net.ServerSentEvents.SseParser.Create(__stream).EnumerateAsync(...))

The Gemini REST API only answers streamGenerateContent as server-sent events when ?alt=sse is on the query string. Without it the response is a chunked JSON array of GenerateContentResponse objects ([{…},\r\n{…}]), which contains no data: lines, so SseParser enumerates zero events and the method returns having yielded nothing — no exception, no text.

Nothing in the repo adds the parameter: GeminiClient.ChatClient.cs calls ModelsStreamGenerateContentAsStreamAsync(modelsId, request, cancellationToken) with no requestOptions, and src/libs/Google.Gemini/openapi.json / convert_discovery.py do not declare alt on the stream operations. (Dynamic… and TunedModels…StreamGenerateContentAsStreamAsync have the same shape.)

Reference: https://ai.google.dev/api/generate-content#method:-models.streamgeneratecontent — "alt=sse … streams the response as server-sent events".

Fix options

1. Cheapest, in the SDK: have the three …StreamGenerateContentAsStreamAsync requests always carry alt=sse — e.g. PrepareModelsStreamGenerateContentAsStreamArguments partial, or the ChatClient extension passing requestOptions: new AutoSDKRequestOptions { QueryParameters = { ["alt"] = "sse" } }.
2. In the spec: declare alt as a query parameter with default sse on the :streamGenerateContent operations in openapi.json (via convert_discovery.py), so the generator emits it.
3. Alternatively, fall back to parsing the JSON-array form when the response Content-Type is application/json.

Either way, src/tests/IntegrationTests/Examples/ChatClient.FiveRandomWords.Streaming.cs would have caught this if it ran against the live API — worth asserting updates.Any(u => u.Text.Length > 0) there.

Consumer-side workaround

A separate GeminiClient with Options.QueryParameters["alt"] = "sse" used only for streaming calls. Tracking on our side: HavenDV/Advantage (linked below).

[HavenDV]

Consumer tracking: https://github.com/HavenDV/Advantage/issues/521