.NET: terminate chat completions SSE stream with data: [DONE]

[Manjunath Janardhan (manjunathshiva)]

Motivation & Context

The agent chat completions endpoint ends its SSE stream by simply running out of frames. It never
writes data: [DONE], which is the terminator the OpenAI wire format specifies and which every
other party in this repo already assumes:

• A live Azure OpenAI chat completion with stream=true ends with exactly data: [DONE].
• This repo's own recorded OpenAI trace for the endpoint ends with it too:
  tests/Microsoft.Agents.AI.Hosting.OpenAI.UnitTests/ConformanceTraces/ChatCompletions/streaming/response.txt.
• [DONE] appears exactly once anywhere in dotnet/src, and it is a reader:
  Aspire.Hosting.AgentFramework.DevUI/SseResponseIdCapture.cs:91 matches "[DONE]"u8.

So the framework ships a consumer that waits for a sentinel the server never produces. Today callers
have to add their own middleware to append the frame, which is what the issue reports.

One correction to the issue, so reviewers are not looking for the wrong symptom: I could not
reproduce "clients will hang or timeout". The response uses Transfer-Encoding: chunked, so the
zero-length terminating chunk gives a definitive end-of-body, and the official Python openai SDK
completes cleanly against the unfixed endpoint. The real defect is protocol conformance — a
consumer that keys on the sentinel rather than end-of-body, such as the DevUI capture above, cannot
detect completion.

Description & Review Guide

• What are the major changes?

  GetStreamingChunksAsync yields a null-payload sentinel after the agent stream completes, and the
  item formatter writes [DONE] for it. The item type widens to SseItem<ChatCompletionChunk?>;
  that method is private and ChatCompletionChunk is a record, so this is a nullable annotation
  only.

  The terminator goes through SseFormatter rather than being appended to response.Body after
  WriteAsync returns, for two reasons. It keeps SseFormatter the single owner of the data:
  prefix and the blank-line frame terminator, instead of copying that framing to a second site. And
  because the sentinel sits after the await foreach, an agent exception or an aborted request
  propagates out of the enumerator and no terminator is written — a client must never be able to read
  a truncated stream as a complete one.

• What is the impact of these changes?

  One additional SSE frame at the end of a streaming chat completion, matching the upstream format.
  AIAgentChatCompletionsProcessor is internal, so there is no public API change. The
  non-streaming path is untouched.

  The Responses endpoint is deliberately unchanged. That API signals completion with typed
  response.* events — StreamingResponseCompleted at Responses/AgentResponseUpdateExtensions.cs:232
  — and its recorded trace contains no [DONE] frame, so adding one there would be wrong. AG-UI has
  its own protocol and is likewise untouched.

• What do you want reviewers to focus on?

Whether the null-payload sentinel is the mechanism you want, or you would rather see an explicit
marker instance. And whether the terminator should be unconditional as implemented — the issue
author's own workaround was opt-in middleware, but making it configurable here would keep the
non-conforming behaviour as the default.

Two notes that may save a review cycle. The existing StreamingRequestResponseAsync could not have
caught this: it asserts through ParseChatCompletionChunksFromSse, which contains
// Skip [DONE] marker and continue (OpenAIChatCompletionsConformanceTests.cs:605, duplicated
at OpenAIChatCompletionsSerializationTests.cs:554), so it loads a trace ending in the terminator
and then discards the only line that would have failed. I left that shared parser alone — eight
call sites depend on it — and the new test asserts on the raw body instead. Separately, the new
test pins the full frame including its blank line, data: [DONE]\n\n; that matches live wire
behaviour and this endpoint's output, but note the recorded trace file itself has no trailing
newline, so the assertion cannot be derived from the trace.

Related Issue

Fixes #8526

Contribution Checklist

• The code builds clean without any errors or warnings
• All unit tests pass, and I have added new tests where possible
• The PR follows the Contribution Guidelines
• This PR is linked to an issue and there is no other open PR for this issue (see Related Issue above).
• This is not a breaking change. If it is a breaking change, add the breaking change label (or add "[BREAKING]" to the title prefix, before or after any language prefix) — a workflow keeps the label and title prefix in sync automatically.

[Copilot]

🟡 Changes recommended

The failure-path invariant needs coverage ensuring truncated streams never emit [DONE].

Get a fresh assessment by requesting another Copilot review.

Pull request overview

Adds OpenAI-compatible [DONE] termination to .NET chat-completion SSE streams.

Changes:

• Emits [DONE] only after successful stream completion.
• Adds raw-response conformance coverage.

File summaries

File	Description
AIAgentChatCompletionsProcessor.cs	Formats and emits the terminal SSE frame.
OpenAIChatCompletionsConformanceTests.cs	Verifies sentinel framing and placement.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 2
• Review effort level: Balanced

---

💡 Add a code-review agent skill for context-aware, tailored reviews. Learn more in the docs.