Remove the SSE terminal-event workaround once Foundry ends streams promptly

[shibayan]

Summary

OpenAIChatClient wraps streamed Responses API events in #untilTerminalEvent (packages/openai/src/chat-client.ts), which stops pulling the SSE stream as soon as a terminal event (response.completed / response.incomplete / response.failed / response.cancelled) has been yielded and releases the underlying SDK stream, instead of draining to the connection close the way the OpenAI SDK's own iterator does.

This is a deliberate workaround for a service-side behavior of Azure AI Foundry's OpenAI-compatible endpoint, not a permanent design choice. This issue tracks its removal once the service is fixed.

The service behavior it works around

Foundry's {project-endpoint}/openai/v1/responses (measured 2026-08-09):

• holds the HTTP connection open for ~5.0 seconds after sending the terminal SSE event (tail measured at 5008±4ms, identical across Accept-Encoding: identity / gzip / default), and
• never sends the data: [DONE] sentinel.

The OpenAI platform sends [DONE] and closes immediately after the terminal event. Official SDKs (openai-node, openai-python, openai-dotnet, openai-go) iterate SSE streams until [DONE] or connection close, so against Foundry every streamed request pays the fixed ~5s idle tail after response.completed. For a hosted agent, whose hosting layer streams every model call, that is ~5s of avoidable latency per model round.

Repro:

curl -sS -N -X POST "$FOUNDRY_PROJECT_ENDPOINT/openai/v1/responses" \
  -H "authorization: Bearer $(az account get-access-token --scope https://ai.azure.com/.default --query accessToken -o tsv)" \
  -H "content-type: application/json" \
  -d '{"model":"<deployment>","input":"Reply with exactly the word OK.","store":false,"stream":true}'

response.completed arrives in ~1.1–1.6s; the connection stays open for another ~5s with no further data.

Why the workaround is safe to keep in the meantime

The Responses protocol defines no events after the terminal one. Against a well-behaved endpoint the wrapper is a no-op: the close follows the terminal event at once, so stopping there changes nothing.

Removal steps (when Foundry ends streams promptly after the terminal event)

1. Delete #untilTerminalEvent in packages/openai/src/chat-client.ts and pass the SDK stream straight to #parseStream at its two call sites (the create-stream path and the background-resume path).
2. Delete the paired test releases the SSE stream at the terminal event instead of draining to the connection close in packages/openai/src/chat-client.test.ts.
3. Re-run the repro above and confirm the stream now ends promptly after the terminal event.

Feedback about the service behavior has been submitted to the Azure AI Foundry team.