fix: handle Azure filter stream chunks

[HAYDEN-OAI]

Summary

• tolerate Azure async content-filter annotation chunks that omit delta
• keep empty-ID filter chunks attached to the current streamed completion
• add a regression test for ChatCompletionStreamingRunner.fromReadableStream()

Why

Azure's asynchronous content filter can interleave annotation-only chunks with normal chat completion chunks. Those annotations may omit delta and use an empty id, which caused the streaming runner to dereference choice.delta and could also make it treat the annotation as a new completion.

Impact

Applications proxying Azure chat completion streams through fromReadableStream() no longer crash or finalize early when async filter annotations arrive. Normal OpenAI and Azure streaming behavior is unchanged.

Validation

• pnpm exec jest tests/lib/ChatCompletionStream.test.ts --runInBand
• pnpm exec tsc --noEmit
• pnpm run lint
• git diff --check

[openai-sdks]

OkTest Summary

✅ 237/237 SDK tests passed in 10.242s for Node SDK PR #1982.

Test results — 42 files

Test	Result	Time
tests/chat-completions-complex-body.test.ts	✅ Passed	140ms
tests/chat-completions-create.test.ts	✅ Passed	241ms
tests/chat-completions-stream.test.ts	✅ Passed	204ms
tests/files-content-binary.test.ts	✅ Passed	162ms
tests/files-create-multipart.test.ts	✅ Passed	137ms
tests/files-list-pagination.test.ts	✅ Passed	160ms
tests/initialize-config.test.ts	✅ Passed	189ms
tests/instance-isolation.test.ts	✅ Passed	107ms
tests/models-list.test.ts	✅ Passed	140ms
tests/responses-background-lifecycle.test.ts	✅ Passed	197ms
tests/responses-body-method-errors.test.ts	✅ Passed	378ms
tests/responses-cancel-timeout.test.ts	✅ Passed	205ms
tests/responses-cancel.test.ts	✅ Passed	178ms
tests/responses-compact-retries.test.ts	✅ Passed	244ms
tests/responses-compact.test.ts	✅ Passed	204ms
tests/responses-create-advanced-stream.test.ts	✅ Passed	122ms
tests/responses-create-advanced.test.ts	✅ Passed	165ms
tests/responses-create-disconnect.test.ts	✅ Passed	1.193s
tests/responses-create-errors.test.ts	✅ Passed	185ms
tests/responses-create-malformed-api-responses.test.ts	✅ Passed	129ms
tests/responses-create-retries.test.ts	✅ Passed	267ms
tests/responses-create-stream-failures.test.ts	✅ Passed	257ms
tests/responses-create-stream-timeout.test.ts	✅ Passed	2.173s
tests/responses-create-stream-wire.test.ts	✅ Passed	2.586s
tests/responses-create-stream.test.ts	✅ Passed	70ms
tests/responses-create-terminal-states.test.ts	✅ Passed	197ms
tests/responses-create-timeout.test.ts	✅ Passed	226ms
tests/responses-create.test.ts	✅ Passed	138ms
tests/responses-delete.test.ts	✅ Passed	132ms
tests/responses-input-items-errors.test.ts	✅ Passed	200ms
tests/responses-input-items-list.test.ts	✅ Passed	152ms
tests/responses-input-items-options.test.ts	✅ Passed	185ms
tests/responses-input-tokens-count-timeout.test.ts	✅ Passed	328ms
tests/responses-input-tokens-count.test.ts	✅ Passed	218ms
tests/responses-malformed-inputs.test.ts	✅ Passed	2.161s
tests/responses-not-found-errors.test.ts	✅ Passed	213ms
tests/responses-parse.test.ts	✅ Passed	159ms
tests/responses-retrieve-retries.test.ts	✅ Passed	286ms
tests/responses-retrieve.test.ts	✅ Passed	209ms
tests/responses-stored-method-errors.test.ts	✅ Passed	492ms
tests/retry-behavior.test.ts	✅ Passed	3.021s
tests/sdk-error-shape.test.ts	✅ Passed	248ms

View OkTest run #28827072571

_{SDK merge (fd0e0f69b49b) · head (1762e8fabc5f) · base (95b54e589491) · OkTest (7fc645091b1a)}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1a5b96f38e

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[jbeckwith-oai]

Reviewed against the original Azure async-filter report in #1015 and the documented Azure annotation stream shape.

Validation performed:

• Reproduced the pre-fix failure at base 95b54e58: the proxied fromReadableStream() sequence rejects with Cannot read properties of undefined (reading 'content') when an empty-ID annotation omits delta.
• Ran the same sequence at 1762e8fa: it completes successfully and preserves id, created, model, content, finish reason, and choice-level filter annotations.
• Exercised Azure's documented leading prompt-annotation chunk, interleaved/trailing completion annotations, two proxied completions separated by an empty-ID annotation, and the direct createChatCompletion() path. All passed; annotation fields remained attached to the intended choice/completion and completion boundaries remained correct.
• jest tests/lib/ChatCompletionStream.test.ts --runInBand: 6/6 tests and 4/4 snapshots passed.
• git diff --check: clean.
• GitHub checks are green for build, lint, tests on Node 20/22/24/26, ecosystem tests, breaking-change detection, CodeQL, and Agents SDK regressions; OkTest reports 237/237 passed.

The changes are defensive for an Azure wire-format divergence without changing the public TypeScript API. Non-empty IDs retain existing completion-boundary and metadata behavior; empty-ID annotation chunks now preserve active completion identity while still merging choice-level moderation data. I did not find a blocking or nonblocking defect.

[jbeckwith-oai]

LGTM - the rename/destructure did make it a tad annoying to review tho