fix: preserve chat stream moderation results

[jbeckwith-oai]

Problem

Moderated Chat Completions can deliver the existing moderation field on streamed raw chunks, including a final moderation-only chunk with empty choices. The chat streaming helper accumulated content and usage but dropped moderation, so current snapshots and get_final_completion returned nil even though stream_raw exposed the result.

User impact

Applications using client.chat.completions.stream for moderated completions now receive the same moderation result already available through client.chat.completions.stream_raw. Unmoderated streams remain nil.

Fix

The helper now carries moderation into the initial parsed snapshot and, for later non-nil chunks, deep-dumps and coerces the field into the existing final ChatCompletion moderation type. Omitted or explicit nil later values do not erase a previously received moderation result. Raw chunks are not mutated.

Synthetic regression proof

The focused WebMock SSE test uses only fake local data through the public stream_raw and stream APIs. Before this fix, a final empty-choices moderation chunk retained nested moderation data in stream_raw while helper snapshots and final completion returned nil. After this fix, snapshots and final completion retain the same nested input/output success and error variants, boolean flags, and numeric scores. The test also covers initial moderation surviving later omitted and nil fields, and the unmoderated nil control.

Compatibility and scope

No public signatures, generated models, dependencies, transport behavior, parsing timing, event types, safety policy, tool-call accumulation, logprob accumulation, or payload limits change. Content, choices, usage, event ordering, and raw chunk values remain covered. The diff is limited to the handwritten stream helper and one focused regression test.

Verification

• Focused regression: 3 runs, 21 assertions, 0 failures, 0 errors
• Affected chat streaming checks: 36 runs, 234 assertions, 0 failures, 0 errors
• Full offline suite against the current-spec mock: 1626 runs, 14450 assertions, 0 failures, 0 errors, 1 skip
• rake lint: RuboCop inspected 2870 files with no offenses; Sorbet and 1278 RBS validations passed
• Trusted public custom-code budget on committed candidate: 3363 of 4000 custom lines, 637 headroom
• Dedicated bounded response deserialization and isolation security review: no new sensitive sink, logging path, policy decision, or raw-chunk mutation
• Adversarial review: two consecutive clean rounds, each with two new independent read-only reviewers

Limitations

This is an offline synthetic regression proof. It does not claim live server observation and does not introduce moderation enforcement.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-04T16:42:24.914861Z	ba771ce	PR opened
🔒 Security Review	✅ Completed 2026-09-04T16:44:10.320030Z	ba771ce	PR opened

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[github-actions]

Castiron custom code

✅ No new custom-code files detected.

47 mixed files remain; 0 existing customizations changed.

Compared 91d0e86ed995 → ba771cec1145. Generated baselines verified.

47 existing customizations unchanged

• lib/openai.rb
• lib/openai/client.rb
• lib/openai/models/chat/chat_completion_message.rb
• lib/openai/models/chat/chat_completion_message_function_tool_call.rb
• lib/openai/models/chat/completion_create_params.rb
• lib/openai/models/response_format_json_schema.rb
• lib/openai/models/responses/function_tool.rb
• lib/openai/models/responses/response.rb
• lib/openai/models/responses/response_create_params.rb
• lib/openai/models/responses/response_format_text_config.rb
• lib/openai/models/responses/response_format_text_json_schema_config.rb
• lib/openai/models/responses/response_function_tool_call.rb
• lib/openai/models/responses/response_function_web_search.rb
• lib/openai/models/responses/response_output_text.rb
• lib/openai/models/responses/tool.rb
• lib/openai/resources/beta/threads.rb
• lib/openai/resources/chat/completions.rb
• lib/openai/resources/files.rb
• lib/openai/resources/responses.rb
• lib/openai/resources/vector_stores/file_batches.rb
• lib/openai/resources/vector_stores/files.rb
• lib/openai/resources/webhooks.rb
• rbi/openai/client.rbi
• rbi/openai/models/chat/chat_completion_message.rbi
• rbi/openai/models/chat/chat_completion_message_function_tool_call.rbi
• rbi/openai/models/chat/completion_create_params.rbi
• rbi/openai/models/response_format_json_schema.rbi
• rbi/openai/models/responses/response.rbi
• rbi/openai/models/responses/response_create_params.rbi
• rbi/openai/models/responses/response_function_tool_call.rbi
• rbi/openai/models/responses/response_function_web_search.rbi
• rbi/openai/models/responses/response_output_text.rbi
• rbi/openai/resources/chat/completions.rbi
• rbi/openai/resources/files.rbi
• rbi/openai/resources/responses.rbi
• rbi/openai/resources/vector_stores/file_batches.rbi
• rbi/openai/resources/vector_stores/files.rbi
• scripts/castiron/README.md
• scripts/castiron/custom_code_report.py
• scripts/castiron/test_custom_code_report.py

7 more in the full report.

A changed generated baseline means this report cannot reliably identify which handwritten lines changed.

Inspect the custom-code diff

Download the exact patch produced by this run (requires repository access):

gh run download 33896536957 --repo openai/openai-ruby \
  --name castiron-custom-code-33896536957-1 --dir /tmp/castiron-custom-code-33896536957-1
git apply --stat /tmp/castiron-custom-code-33896536957-1/custom-code.patch
cat /tmp/castiron-custom-code-33896536957-1/custom-code.patch

Or reproduce it from an SDK checkout containing the vendored reporter:

git fetch --no-tags origin 91d0e86ed995f54361b329d7641af9f20163f27d ba771cec1145abc878a730b91f701e529e15d063
python3 scripts/castiron/custom_code_report.py report \
  --base 91d0e86ed995f54361b329d7641af9f20163f27d \
  --head ba771cec1145abc878a730b91f701e529e15d063 --fetch --require-head-hash --public \
  --out /tmp/castiron-custom-code-ba771cec1145
cat /tmp/castiron-custom-code-ba771cec1145/custom-code.patch

This is the current full custom patch for mixed files, not an attribution of only the handwritten lines changed by this PR.

Full report and patch

[openai-sdks]

OkTest Summary

✅ 237/237 SDK tests passed in 8.504s for Ruby SDK PR #616.

Test results — 42 files

Test	Result	Time
tests/chat-completions-complex-body.test.ts	✅ Passed	152ms
tests/chat-completions-create.test.ts	✅ Passed	159ms
tests/chat-completions-stream.test.ts	✅ Passed	106ms
tests/files-content-binary.test.ts	✅ Passed	141ms
tests/files-create-multipart.test.ts	✅ Passed	204ms
tests/files-list-pagination.test.ts	✅ Passed	127ms
tests/initialize-config.test.ts	✅ Passed	160ms
tests/instance-isolation.test.ts	✅ Passed	153ms
tests/models-list.test.ts	✅ Passed	159ms
tests/responses-background-lifecycle.test.ts	✅ Passed	182ms
tests/responses-body-method-errors.test.ts	✅ Passed	347ms
tests/responses-cancel-timeout.test.ts	✅ Passed	256ms
tests/responses-cancel.test.ts	✅ Passed	185ms
tests/responses-compact-retries.test.ts	✅ Passed	228ms
tests/responses-compact.test.ts	✅ Passed	204ms
tests/responses-create-advanced-stream.test.ts	✅ Passed	153ms
tests/responses-create-advanced.test.ts	✅ Passed	164ms
tests/responses-create-disconnect.test.ts	✅ Passed	1.076s
tests/responses-create-errors.test.ts	✅ Passed	241ms
tests/responses-create-malformed-api-responses.test.ts	✅ Passed	113ms
tests/responses-create-retries.test.ts	✅ Passed	242ms
tests/responses-create-stream-failures.test.ts	✅ Passed	129ms
tests/responses-create-stream-timeout.test.ts	✅ Passed	208ms
tests/responses-create-stream-wire.test.ts	✅ Passed	1.943s
tests/responses-create-stream.test.ts	✅ Passed	83ms
tests/responses-create-terminal-states.test.ts	✅ Passed	218ms
tests/responses-create-timeout.test.ts	✅ Passed	241ms
tests/responses-create.test.ts	✅ Passed	204ms
tests/responses-delete.test.ts	✅ Passed	225ms
tests/responses-input-items-errors.test.ts	✅ Passed	150ms
tests/responses-input-items-list.test.ts	✅ Passed	147ms
tests/responses-input-items-options.test.ts	✅ Passed	137ms
tests/responses-input-tokens-count-timeout.test.ts	✅ Passed	224ms
tests/responses-input-tokens-count.test.ts	✅ Passed	182ms
tests/responses-malformed-inputs.test.ts	✅ Passed	1.604s
tests/responses-not-found-errors.test.ts	✅ Passed	245ms
tests/responses-parse.test.ts	✅ Passed	187ms
tests/responses-retrieve-retries.test.ts	✅ Passed	260ms
tests/responses-retrieve.test.ts	✅ Passed	177ms
tests/responses-stored-method-errors.test.ts	✅ Passed	548ms
tests/retry-behavior.test.ts	✅ Passed	3.19s
tests/sdk-error-shape.test.ts	✅ Passed	344ms

View OkTest run #33896496684

_{SDK merge (726a76666e08) · head (ba771cec1145) · base (91d0e86ed995) · OkTest (2b1bdfd25e98)}

[sylvesterkaczmarek]

Treating moderation as sticky stream state is the correct accumulation rule: a moderation-only terminal chunk must update the snapshot, while later omitted or nil values must not erase an earlier result. The raw-stream parity assertions cover that boundary well.

[markstuart-oai]

The stream accumulator now carries moderation into the initial snapshot and updates it only from later non-nil chunks, preserving the existing sticky stream-state semantics. The focused public-API coverage exercises final moderation-only chunks, initial values, nil/omitted updates, and raw-stream isolation. Exact-head CI is green across the Ruby test matrix, lint, types, packaging, CodeQL, and Castiron. I could not rerun the focused test locally because the required Ruby 4.0.6 toolchain is not installed on this review box.