[BenjaminMichaelis]

Why

Chat on dev fails ("Unable to connect to the chat service"). The server logs show ClientResultException: HTTP 404 (Resource not found) from ResponsesClient.CreateResponseAsync.

#1282 replaced AzureOpenAIClient.GetResponsesClient() (binary-incompatible after the OpenAI 2.12 bump) with a hand-built ResponsesClient whose base is {endpoint}/openai/deployments/{name}. Responses has no deployment-scoped route: that path returns 404 for every deployment (gpt-5 included). The SDK's own client called {endpoint}/openai/responses?api-version=2025-04-01-preview (I read the Azure.AI.OpenAI 2.9.0-beta.1 source). Prod still runs the August code (77c809fb) with the SDK client and works, so the next prod deploy would break prod chat.

Azure.AI.OpenAI has had no release since 2026-03-13, and its changelog/migration guide recommend using the OpenAI SDK directly against the v1 routes. (Azure.AI.Extensions.OpenAI targets Foundry agents, not this use.)

What changed

• New AzureOpenAIV1Clients: Responses and embedding clients at {endpoint}/openai/v1, no api-version, deployment name as model, Entra bearer auth. Token scope is unchanged (cognitiveservices.azure.com) so only one variable changes.
• AIChatService and the embedding generator use it. Removed the hand-built deployment URL, the api-version policy, the AzureOpenAIClient registration and two unused Semantic Kernel chat registrations (AddAzureOpenAIChatClient/AddAzureOpenAIChatCompletion; nothing resolves them).
• Microsoft.SemanticKernel -> Microsoft.SemanticKernel.Core (Kernel and TextChunker still come from it). The metapackage was the only thing pulling in Azure.AI.OpenAI, which is gone from the dependency graph. Removed its version entry.
• AzureOpenAIV1ClientsTests: pin method, URL, query, auth header and model for both clients through a stub transport.

Not verified (needs dev)

The resource is private, so I could not call it from CI or my machine. Please send one chat message on dev after deploy; contextual search is on by default, so it exercises both chat and embeddings. Things that could still fail:

• the v1 route on ecs-cog-prod.cognitiveservices.azure.com with grok-4-1-fast-non-reasoning (Microsoft documents Grok via the Foundry project route .../api/projects/<project>/openai/v1; the project essentialcsharp exists but is not in Terraform);
• the token scope on the v1 route (a 401 would point at https://ai.azure.com/.default);
• embeddings now use /openai/v1/embeddings instead of the classic deployment route.

If the account host does not work, the next step is a Terraform change (the endpoint comes from AIOptions--Endpoint Key Vault secrets, source of truth is EssentialCSharp.AzureResourceManagement), dev first, since prod still runs the old code.

Hold the pending Production approval until chat works on dev.

Test plan

• dotnet test EssentialCSharp.Chat.Tests (35 passed, incl. 4 new)
• dotnet test EssentialCSharp.Web.Tests (190 passed)
• dotnet nuget why Azure.AI.OpenAI reports no dependency in the Web and Chat projects
• Chat message works on dev