fix(streaming): reconcile completed Responses output items

[by-openai]

• I understand that this repository is auto-generated and my pull request may not be merged

Changes being requested

Update ResponseStreamState to replace provisional output items from response.output_item.added with the authoritative item received in response.output_item.done.

A reasoning item's initial encrypted_content, summary, and status can differ from its completed values. Ignoring the done event leaves the streaming snapshot with incomplete encrypted reasoning that cannot safely be replayed. Completed assistant messages and function calls likewise retain stale content, arguments, and status.

The reconciliation preserves ParsedResponseOutputMessage and ParsedResponseFunctionToolCall wrappers and applies to both synchronous and asynchronous Responses streams through their shared state.

Regression tests cover:

• Reasoning items whose preliminary ciphertext differs from the completed ciphertext.
• Completed assistant message content, status, and phase.
• Completed function-call arguments and status.

Validation:

• pytest tests/lib — 259 passed.
• ruff check and ruff format --check on both changed files.
• pyright on both changed files.
• mypy on the changed streaming implementation.

Additional context & links

• Related to responses.stream() crashes with NoneType iteration when response.completed payload has output=null #3321 and fix(parsing): guard against None response.output in parse_response and streaming accumulator #3420. This change does not modify null terminal-response handling; it ensures any future snapshot fallback receives completed output items.
• The stale-snapshot concern is already identified in fix(parsing): guard against None response.output in parse_response and streaming accumulator #3420 (comment).
• The OpenAI Node SDK already replaces added-event snapshots on done events: https://github.com/openai/openai-node/blob/main/src/lib/responses/ResponseAccumulator.ts#L33-L43.
• The implementation is under src/openai/lib/, which CONTRIBUTING.md identifies as excluded from SDK code generation.

[openai-sdks]

OkTest Summary

✅ 237/237 SDK tests passed in 11.408s for Python SDK PR #3615.

Test results — 42 files

Test	Result	Time
tests/chat-completions-complex-body.test.ts	✅ Passed	213ms
tests/chat-completions-create.test.ts	✅ Passed	209ms
tests/chat-completions-stream.test.ts	✅ Passed	145ms
tests/files-content-binary.test.ts	✅ Passed	234ms
tests/files-create-multipart.test.ts	✅ Passed	209ms
tests/files-list-pagination.test.ts	✅ Passed	146ms
tests/initialize-config.test.ts	✅ Passed	202ms
tests/instance-isolation.test.ts	✅ Passed	239ms
tests/models-list.test.ts	✅ Passed	225ms
tests/responses-background-lifecycle.test.ts	✅ Passed	213ms
tests/responses-body-method-errors.test.ts	✅ Passed	329ms
tests/responses-cancel-timeout.test.ts	✅ Passed	204ms
tests/responses-cancel.test.ts	✅ Passed	226ms
tests/responses-compact-retries.test.ts	✅ Passed	366ms
tests/responses-compact.test.ts	✅ Passed	289ms
tests/responses-create-advanced-stream.test.ts	✅ Passed	328ms
tests/responses-create-advanced.test.ts	✅ Passed	274ms
tests/responses-create-disconnect.test.ts	✅ Passed	1.182s
tests/responses-create-errors.test.ts	✅ Passed	193ms
tests/responses-create-malformed-api-responses.test.ts	✅ Passed	254ms
tests/responses-create-retries.test.ts	✅ Passed	262ms
tests/responses-create-stream-failures.test.ts	✅ Passed	137ms
tests/responses-create-stream-timeout.test.ts	✅ Passed	209ms
tests/responses-create-stream-wire.test.ts	✅ Passed	3.723s
tests/responses-create-stream.test.ts	✅ Passed	104ms
tests/responses-create-terminal-states.test.ts	✅ Passed	251ms
tests/responses-create-timeout.test.ts	✅ Passed	271ms
tests/responses-create.test.ts	✅ Passed	312ms
tests/responses-delete.test.ts	✅ Passed	214ms
tests/responses-input-items-errors.test.ts	✅ Passed	440ms
tests/responses-input-items-list.test.ts	✅ Passed	315ms
tests/responses-input-items-options.test.ts	✅ Passed	303ms
tests/responses-input-tokens-count-timeout.test.ts	✅ Passed	251ms
tests/responses-input-tokens-count.test.ts	✅ Passed	181ms
tests/responses-malformed-inputs.test.ts	✅ Passed	2.854s
tests/responses-not-found-errors.test.ts	✅ Passed	261ms
tests/responses-parse.test.ts	✅ Passed	509ms
tests/responses-retrieve-retries.test.ts	✅ Passed	242ms
tests/responses-retrieve.test.ts	✅ Passed	239ms
tests/responses-stored-method-errors.test.ts	✅ Passed	617ms
tests/retry-behavior.test.ts	✅ Passed	3.17s
tests/sdk-error-shape.test.ts	✅ Passed	271ms

View OkTest run #31767474355

_{SDK merge (6404589e77db) · head (9dad2ad84f53) · base (d9029e3ada3c) · OkTest (a845fa206fa4)}

[ting-hong-shieh]

Validated at 9dad2ad8 against base d9029e3a.

I ran the same synthetic response.created → response.output_item.added → response.output_item.done sequence for reasoning, message, and function-call items. The canonical snapshot hashes changed as follows:

Item	Base snapshot	Done item / head snapshot
Reasoning	aa8c8dd24e79	b897c980880a
Message	6a64d2e6f63e	1c0e1e631e2a
Function call	68c4bd453415	b0fc42c0be19

On the base, all three snapshots still held their provisional values after the done event. On this head, each snapshot matched the authoritative done item exactly. The completed message remained a ParsedResponseOutputMessage with ParsedResponseOutputText content, and the function call remained a ParsedResponseFunctionToolCall.

Validation on the exact head:

• pytest -q -n 0 tests/lib — 259 passed
• focused Responses file — 10 passed
• Ruff lint and format checks — passed
• Pyright — 0 errors; Mypy — passed
• compilation and git diff --check — passed

The three public workflow runs are also green. The snapshot used constructed local events only; it made no API request and used no credentials.