feat(openai): instrument Audio API (speech, transcriptions, translations)

[Abhijeet Prasad (AbhiPrasad)]

Add tracing wrappers for OpenAI's three Audio API methods so that
wrap_openai() and OpenAIIntegration.setup() produce Braintrust spans
for audio.speech.create, audio.transcriptions.create, and
audio.translations.create (sync + async).

Closes#174

[starfolkai]

Drop into this review session: sfk devbox session 68

PR #232 — Audio API Instrumentation: Code Review

Verdict: Approve with minor suggestions

All 438 tests pass (nox -s "test_openai(latest)"), including the 6 new audio tests. The implementation is clean and consistent with the repo's established patterns.

---

What's good

• Follows existing conventions exactly: VCR cassettes, sync/async patcher pairs, registration in both integration.py (for setup()) and _WRAP_TARGETS (for wrap_openai()). The test_wrap_openai_and_setup_use_same_wrappers enforcement is satisfied.
• _AudioFileWrapper._parse_params correctly prevents logging binary audio data by popping file from the prettified copy of params, so the original kwargs passed to the API call are unaffected.
• _make_base_wrapper_callback is a reasonable factory that reduces the repetitive sync/async branching seen in the inline wrappers for chat/embeddings/moderation.
• SpeechWrapper logging {"type": "audio"} for output is the right call — streaming binary audio into a span would be impractical.
• TranscriptionWrapper correctly handles the {"type":"duration","seconds":1} usage shape from Whisper — _parse_metrics_from_usage returns an empty dict gracefully when token fields are absent.

---

Issues

1. _extract_text fallback can silently produce wrong output (minor bug)

tracing.py, _AudioFileWrapper._extract_text:

@staticmethoddef_extract_text(response: Any) ->str|None:
ifisinstance(response, dict):
returnresponse.get("text")
ifisinstance(response, str):
returnresponsereturnstr(response) # ← problem here

When _try_to_dict can't convert the response to a dict (e.g., an object type it doesn't recognize), and the response isn't a plain string, the fallback str(response) will return something like Transcription(text='hello world', logprobs=None, usage=None) — Pydantic's repr — instead of the actual transcribed text. This won't crash, but the span output would be garbage.

A safer fallback:

returngetattr(response, "text", None)

This handles OpenAI Pydantic model types (Transcription, TranscriptionVerbose, Translation, etc.) that have a .text attribute directly, without relying on dict conversion.

---

2. Weak test assertions for transcription/translation

assertspan["output"] isnotNone

For transcription and translation tests, the output is never checked beyond being non-None. Given that the cassette response is {"text":"you","usage":{"type":"duration","seconds":1}}, you can assert the actual value:

assertspan["output"] =="you"

This would have caught the _extract_text fallback issue at test-recording time if anything went wrong.

---

Nits / observations (non-blocking)

_make_base_wrapper_callback is placed ~600 lines before BaseWrapper — the string annotation wrapper_cls: type["BaseWrapper"] is correct, but it's a bit of a readability surprise. The existing per-method wrapper functions (_embedding_create_wrapper, etc.) aren't converted to use the factory, which leaves a small inconsistency. Neither is a blocker, but if the factory pattern proves useful it might be worth a follow-up to unify the three inline wrappers.

No coverage for response_format="text" — when the caller passes response_format="text", OpenAI returns a plain str rather than a JSON object. The isinstance(response, str) branch in _extract_text handles this, but there's no test for it. A cassette for this path would be a good addition.

---

Summary

The core logic is correct and all existing + new tests pass. I'd address the _extract_text fallback (issue 1) and strengthen the transcription/translation output assertions (issue 2) before merge. Everything else is minor.

[Abhijeet Prasad (AbhiPrasad)]

Add tracing wrappers for OpenAI's three Audio API methods so that
wrap_openai() and OpenAIIntegration.setup() produce Braintrust spans
for audio.speech.create, audio.transcriptions.create, and
audio.translations.create (sync + async).

Closes#174

[starfolkai]

Drop into this review session: sfk devbox session 68

PR #232 — Audio API Instrumentation: Code Review

Verdict: Approve with minor suggestions

All 438 tests pass (nox -s "test_openai(latest)"), including the 6 new audio tests. The implementation is clean and consistent with the repo's established patterns.

---

What's good

• Follows existing conventions exactly: VCR cassettes, sync/async patcher pairs, registration in both integration.py (for setup()) and _WRAP_TARGETS (for wrap_openai()). The test_wrap_openai_and_setup_use_same_wrappers enforcement is satisfied.
• _AudioFileWrapper._parse_params correctly prevents logging binary audio data by popping file from the prettified copy of params, so the original kwargs passed to the API call are unaffected.
• _make_base_wrapper_callback is a reasonable factory that reduces the repetitive sync/async branching seen in the inline wrappers for chat/embeddings/moderation.
• SpeechWrapper logging {"type": "audio"} for output is the right call — streaming binary audio into a span would be impractical.
• TranscriptionWrapper correctly handles the {"type":"duration","seconds":1} usage shape from Whisper — _parse_metrics_from_usage returns an empty dict gracefully when token fields are absent.

---

Issues

1. _extract_text fallback can silently produce wrong output (minor bug)

tracing.py, _AudioFileWrapper._extract_text:

@staticmethoddef_extract_text(response: Any) ->str|None:
ifisinstance(response, dict):
returnresponse.get("text")
ifisinstance(response, str):
returnresponsereturnstr(response) # ← problem here

When _try_to_dict can't convert the response to a dict (e.g., an object type it doesn't recognize), and the response isn't a plain string, the fallback str(response) will return something like Transcription(text='hello world', logprobs=None, usage=None) — Pydantic's repr — instead of the actual transcribed text. This won't crash, but the span output would be garbage.

A safer fallback:

returngetattr(response, "text", None)

This handles OpenAI Pydantic model types (Transcription, TranscriptionVerbose, Translation, etc.) that have a .text attribute directly, without relying on dict conversion.

---

2. Weak test assertions for transcription/translation

assertspan["output"] isnotNone

For transcription and translation tests, the output is never checked beyond being non-None. Given that the cassette response is {"text":"you","usage":{"type":"duration","seconds":1}}, you can assert the actual value:

assertspan["output"] =="you"

This would have caught the _extract_text fallback issue at test-recording time if anything went wrong.

---

Nits / observations (non-blocking)

_make_base_wrapper_callback is placed ~600 lines before BaseWrapper — the string annotation wrapper_cls: type["BaseWrapper"] is correct, but it's a bit of a readability surprise. The existing per-method wrapper functions (_embedding_create_wrapper, etc.) aren't converted to use the factory, which leaves a small inconsistency. Neither is a blocker, but if the factory pattern proves useful it might be worth a follow-up to unify the three inline wrappers.

No coverage for response_format="text" — when the caller passes response_format="text", OpenAI returns a plain str rather than a JSON object. The isinstance(response, str) branch in _extract_text handles this, but there's no test for it. A cassette for this path would be a good addition.

---

Summary

The core logic is correct and all existing + new tests pass. I'd address the _extract_text fallback (issue 1) and strengthen the transcription/translation output assertions (issue 2) before merge. Everything else is minor.

[Abhijeet Prasad (AbhiPrasad)]

Add tracing wrappers for OpenAI's three Audio API methods so that
wrap_openai() and OpenAIIntegration.setup() produce Braintrust spans
for audio.speech.create, audio.transcriptions.create, and
audio.translations.create (sync + async).

Closes#174

[starfolkai]

Drop into this review session: sfk devbox session 68

PR #232 — Audio API Instrumentation: Code Review

Verdict: Approve with minor suggestions

All 438 tests pass (nox -s "test_openai(latest)"), including the 6 new audio tests. The implementation is clean and consistent with the repo's established patterns.

---

What's good

• Follows existing conventions exactly: VCR cassettes, sync/async patcher pairs, registration in both integration.py (for setup()) and _WRAP_TARGETS (for wrap_openai()). The test_wrap_openai_and_setup_use_same_wrappers enforcement is satisfied.
• _AudioFileWrapper._parse_params correctly prevents logging binary audio data by popping file from the prettified copy of params, so the original kwargs passed to the API call are unaffected.
• _make_base_wrapper_callback is a reasonable factory that reduces the repetitive sync/async branching seen in the inline wrappers for chat/embeddings/moderation.
• SpeechWrapper logging {"type": "audio"} for output is the right call — streaming binary audio into a span would be impractical.
• TranscriptionWrapper correctly handles the {"type":"duration","seconds":1} usage shape from Whisper — _parse_metrics_from_usage returns an empty dict gracefully when token fields are absent.

---

Issues

1. _extract_text fallback can silently produce wrong output (minor bug)

tracing.py, _AudioFileWrapper._extract_text:

@staticmethoddef_extract_text(response: Any) ->str|None:
ifisinstance(response, dict):
returnresponse.get("text")
ifisinstance(response, str):
returnresponsereturnstr(response) # ← problem here

When _try_to_dict can't convert the response to a dict (e.g., an object type it doesn't recognize), and the response isn't a plain string, the fallback str(response) will return something like Transcription(text='hello world', logprobs=None, usage=None) — Pydantic's repr — instead of the actual transcribed text. This won't crash, but the span output would be garbage.

A safer fallback:

returngetattr(response, "text", None)

This handles OpenAI Pydantic model types (Transcription, TranscriptionVerbose, Translation, etc.) that have a .text attribute directly, without relying on dict conversion.

---

2. Weak test assertions for transcription/translation

assertspan["output"] isnotNone

For transcription and translation tests, the output is never checked beyond being non-None. Given that the cassette response is {"text":"you","usage":{"type":"duration","seconds":1}}, you can assert the actual value:

assertspan["output"] =="you"

This would have caught the _extract_text fallback issue at test-recording time if anything went wrong.

---

Nits / observations (non-blocking)

_make_base_wrapper_callback is placed ~600 lines before BaseWrapper — the string annotation wrapper_cls: type["BaseWrapper"] is correct, but it's a bit of a readability surprise. The existing per-method wrapper functions (_embedding_create_wrapper, etc.) aren't converted to use the factory, which leaves a small inconsistency. Neither is a blocker, but if the factory pattern proves useful it might be worth a follow-up to unify the three inline wrappers.

No coverage for response_format="text" — when the caller passes response_format="text", OpenAI returns a plain str rather than a JSON object. The isinstance(response, str) branch in _extract_text handles this, but there's no test for it. A cassette for this path would be a good addition.

---

Summary

The core logic is correct and all existing + new tests pass. I'd address the _extract_text fallback (issue 1) and strengthen the transcription/translation output assertions (issue 2) before merge. Everything else is minor.

[Abhijeet Prasad (AbhiPrasad)]

Add tracing wrappers for OpenAI's three Audio API methods so that
wrap_openai() and OpenAIIntegration.setup() produce Braintrust spans
for audio.speech.create, audio.transcriptions.create, and
audio.translations.create (sync + async).

Closes#174

[starfolkai]

Drop into this review session: sfk devbox session 68

PR #232 — Audio API Instrumentation: Code Review

Verdict: Approve with minor suggestions

All 438 tests pass (nox -s "test_openai(latest)"), including the 6 new audio tests. The implementation is clean and consistent with the repo's established patterns.

---

What's good

• Follows existing conventions exactly: VCR cassettes, sync/async patcher pairs, registration in both integration.py (for setup()) and _WRAP_TARGETS (for wrap_openai()). The test_wrap_openai_and_setup_use_same_wrappers enforcement is satisfied.
• _AudioFileWrapper._parse_params correctly prevents logging binary audio data by popping file from the prettified copy of params, so the original kwargs passed to the API call are unaffected.
• _make_base_wrapper_callback is a reasonable factory that reduces the repetitive sync/async branching seen in the inline wrappers for chat/embeddings/moderation.
• SpeechWrapper logging {"type": "audio"} for output is the right call — streaming binary audio into a span would be impractical.
• TranscriptionWrapper correctly handles the {"type":"duration","seconds":1} usage shape from Whisper — _parse_metrics_from_usage returns an empty dict gracefully when token fields are absent.

---

Issues

1. _extract_text fallback can silently produce wrong output (minor bug)

tracing.py, _AudioFileWrapper._extract_text:

@staticmethoddef_extract_text(response: Any) ->str|None:
ifisinstance(response, dict):
returnresponse.get("text")
ifisinstance(response, str):
returnresponsereturnstr(response) # ← problem here

When _try_to_dict can't convert the response to a dict (e.g., an object type it doesn't recognize), and the response isn't a plain string, the fallback str(response) will return something like Transcription(text='hello world', logprobs=None, usage=None) — Pydantic's repr — instead of the actual transcribed text. This won't crash, but the span output would be garbage.

A safer fallback:

returngetattr(response, "text", None)

This handles OpenAI Pydantic model types (Transcription, TranscriptionVerbose, Translation, etc.) that have a .text attribute directly, without relying on dict conversion.

---

2. Weak test assertions for transcription/translation

assertspan["output"] isnotNone

For transcription and translation tests, the output is never checked beyond being non-None. Given that the cassette response is {"text":"you","usage":{"type":"duration","seconds":1}}, you can assert the actual value:

assertspan["output"] =="you"

This would have caught the _extract_text fallback issue at test-recording time if anything went wrong.

---

Nits / observations (non-blocking)

_make_base_wrapper_callback is placed ~600 lines before BaseWrapper — the string annotation wrapper_cls: type["BaseWrapper"] is correct, but it's a bit of a readability surprise. The existing per-method wrapper functions (_embedding_create_wrapper, etc.) aren't converted to use the factory, which leaves a small inconsistency. Neither is a blocker, but if the factory pattern proves useful it might be worth a follow-up to unify the three inline wrappers.

No coverage for response_format="text" — when the caller passes response_format="text", OpenAI returns a plain str rather than a JSON object. The isinstance(response, str) branch in _extract_text handles this, but there's no test for it. A cassette for this path would be a good addition.

---

Summary

The core logic is correct and all existing + new tests pass. I'd address the _extract_text fallback (issue 1) and strengthen the transcription/translation output assertions (issue 2) before merge. Everything else is minor.

[Abhijeet Prasad (AbhiPrasad)]

Add tracing wrappers for OpenAI's three Audio API methods so that
wrap_openai() and OpenAIIntegration.setup() produce Braintrust spans
for audio.speech.create, audio.transcriptions.create, and
audio.translations.create (sync + async).

Closes#174

[starfolkai]

Drop into this review session: sfk devbox session 68

PR #232 — Audio API Instrumentation: Code Review

Verdict: Approve with minor suggestions

All 438 tests pass (nox -s "test_openai(latest)"), including the 6 new audio tests. The implementation is clean and consistent with the repo's established patterns.

---

What's good

• Follows existing conventions exactly: VCR cassettes, sync/async patcher pairs, registration in both integration.py (for setup()) and _WRAP_TARGETS (for wrap_openai()). The test_wrap_openai_and_setup_use_same_wrappers enforcement is satisfied.
• _AudioFileWrapper._parse_params correctly prevents logging binary audio data by popping file from the prettified copy of params, so the original kwargs passed to the API call are unaffected.
• _make_base_wrapper_callback is a reasonable factory that reduces the repetitive sync/async branching seen in the inline wrappers for chat/embeddings/moderation.
• SpeechWrapper logging {"type": "audio"} for output is the right call — streaming binary audio into a span would be impractical.
• TranscriptionWrapper correctly handles the {"type":"duration","seconds":1} usage shape from Whisper — _parse_metrics_from_usage returns an empty dict gracefully when token fields are absent.

---

Issues

1. _extract_text fallback can silently produce wrong output (minor bug)

tracing.py, _AudioFileWrapper._extract_text:

@staticmethoddef_extract_text(response: Any) ->str|None:
ifisinstance(response, dict):
returnresponse.get("text")
ifisinstance(response, str):
returnresponsereturnstr(response) # ← problem here

When _try_to_dict can't convert the response to a dict (e.g., an object type it doesn't recognize), and the response isn't a plain string, the fallback str(response) will return something like Transcription(text='hello world', logprobs=None, usage=None) — Pydantic's repr — instead of the actual transcribed text. This won't crash, but the span output would be garbage.

A safer fallback:

returngetattr(response, "text", None)

This handles OpenAI Pydantic model types (Transcription, TranscriptionVerbose, Translation, etc.) that have a .text attribute directly, without relying on dict conversion.

---

2. Weak test assertions for transcription/translation

assertspan["output"] isnotNone

For transcription and translation tests, the output is never checked beyond being non-None. Given that the cassette response is {"text":"you","usage":{"type":"duration","seconds":1}}, you can assert the actual value:

assertspan["output"] =="you"

This would have caught the _extract_text fallback issue at test-recording time if anything went wrong.

---

Nits / observations (non-blocking)

_make_base_wrapper_callback is placed ~600 lines before BaseWrapper — the string annotation wrapper_cls: type["BaseWrapper"] is correct, but it's a bit of a readability surprise. The existing per-method wrapper functions (_embedding_create_wrapper, etc.) aren't converted to use the factory, which leaves a small inconsistency. Neither is a blocker, but if the factory pattern proves useful it might be worth a follow-up to unify the three inline wrappers.

No coverage for response_format="text" — when the caller passes response_format="text", OpenAI returns a plain str rather than a JSON object. The isinstance(response, str) branch in _extract_text handles this, but there's no test for it. A cassette for this path would be a good addition.

---

Summary

The core logic is correct and all existing + new tests pass. I'd address the _extract_text fallback (issue 1) and strengthen the transcription/translation output assertions (issue 2) before merge. Everything else is minor.

[Abhijeet Prasad (AbhiPrasad)]

Add tracing wrappers for OpenAI's three Audio API methods so that
wrap_openai() and OpenAIIntegration.setup() produce Braintrust spans
for audio.speech.create, audio.transcriptions.create, and
audio.translations.create (sync + async).

Closes#174

[starfolkai]

Drop into this review session: sfk devbox session 68

PR #232 — Audio API Instrumentation: Code Review

Verdict: Approve with minor suggestions

All 438 tests pass (nox -s "test_openai(latest)"), including the 6 new audio tests. The implementation is clean and consistent with the repo's established patterns.

---

What's good

• Follows existing conventions exactly: VCR cassettes, sync/async patcher pairs, registration in both integration.py (for setup()) and _WRAP_TARGETS (for wrap_openai()). The test_wrap_openai_and_setup_use_same_wrappers enforcement is satisfied.
• _AudioFileWrapper._parse_params correctly prevents logging binary audio data by popping file from the prettified copy of params, so the original kwargs passed to the API call are unaffected.
• _make_base_wrapper_callback is a reasonable factory that reduces the repetitive sync/async branching seen in the inline wrappers for chat/embeddings/moderation.
• SpeechWrapper logging {"type": "audio"} for output is the right call — streaming binary audio into a span would be impractical.
• TranscriptionWrapper correctly handles the {"type":"duration","seconds":1} usage shape from Whisper — _parse_metrics_from_usage returns an empty dict gracefully when token fields are absent.

---

Issues

1. _extract_text fallback can silently produce wrong output (minor bug)

tracing.py, _AudioFileWrapper._extract_text:

@staticmethoddef_extract_text(response: Any) ->str|None:
ifisinstance(response, dict):
returnresponse.get("text")
ifisinstance(response, str):
returnresponsereturnstr(response) # ← problem here

When _try_to_dict can't convert the response to a dict (e.g., an object type it doesn't recognize), and the response isn't a plain string, the fallback str(response) will return something like Transcription(text='hello world', logprobs=None, usage=None) — Pydantic's repr — instead of the actual transcribed text. This won't crash, but the span output would be garbage.

A safer fallback:

returngetattr(response, "text", None)

This handles OpenAI Pydantic model types (Transcription, TranscriptionVerbose, Translation, etc.) that have a .text attribute directly, without relying on dict conversion.

---

2. Weak test assertions for transcription/translation

assertspan["output"] isnotNone

For transcription and translation tests, the output is never checked beyond being non-None. Given that the cassette response is {"text":"you","usage":{"type":"duration","seconds":1}}, you can assert the actual value:

assertspan["output"] =="you"

This would have caught the _extract_text fallback issue at test-recording time if anything went wrong.

---

Nits / observations (non-blocking)

_make_base_wrapper_callback is placed ~600 lines before BaseWrapper — the string annotation wrapper_cls: type["BaseWrapper"] is correct, but it's a bit of a readability surprise. The existing per-method wrapper functions (_embedding_create_wrapper, etc.) aren't converted to use the factory, which leaves a small inconsistency. Neither is a blocker, but if the factory pattern proves useful it might be worth a follow-up to unify the three inline wrappers.

No coverage for response_format="text" — when the caller passes response_format="text", OpenAI returns a plain str rather than a JSON object. The isinstance(response, str) branch in _extract_text handles this, but there's no test for it. A cassette for this path would be a good addition.

---

Summary

The core logic is correct and all existing + new tests pass. I'd address the _extract_text fallback (issue 1) and strengthen the transcription/translation output assertions (issue 2) before merge. Everything else is minor.

[Abhijeet Prasad (AbhiPrasad)]

Add tracing wrappers for OpenAI's three Audio API methods so that
wrap_openai() and OpenAIIntegration.setup() produce Braintrust spans
for audio.speech.create, audio.transcriptions.create, and
audio.translations.create (sync + async).

Closes#174

[starfolkai]

Drop into this review session: sfk devbox session 68

PR #232 — Audio API Instrumentation: Code Review

Verdict: Approve with minor suggestions

All 438 tests pass (nox -s "test_openai(latest)"), including the 6 new audio tests. The implementation is clean and consistent with the repo's established patterns.

---

What's good

• Follows existing conventions exactly: VCR cassettes, sync/async patcher pairs, registration in both integration.py (for setup()) and _WRAP_TARGETS (for wrap_openai()). The test_wrap_openai_and_setup_use_same_wrappers enforcement is satisfied.
• _AudioFileWrapper._parse_params correctly prevents logging binary audio data by popping file from the prettified copy of params, so the original kwargs passed to the API call are unaffected.
• _make_base_wrapper_callback is a reasonable factory that reduces the repetitive sync/async branching seen in the inline wrappers for chat/embeddings/moderation.
• SpeechWrapper logging {"type": "audio"} for output is the right call — streaming binary audio into a span would be impractical.
• TranscriptionWrapper correctly handles the {"type":"duration","seconds":1} usage shape from Whisper — _parse_metrics_from_usage returns an empty dict gracefully when token fields are absent.

---

Issues

1. _extract_text fallback can silently produce wrong output (minor bug)

tracing.py, _AudioFileWrapper._extract_text:

@staticmethoddef_extract_text(response: Any) ->str|None:
ifisinstance(response, dict):
returnresponse.get("text")
ifisinstance(response, str):
returnresponsereturnstr(response) # ← problem here

When _try_to_dict can't convert the response to a dict (e.g., an object type it doesn't recognize), and the response isn't a plain string, the fallback str(response) will return something like Transcription(text='hello world', logprobs=None, usage=None) — Pydantic's repr — instead of the actual transcribed text. This won't crash, but the span output would be garbage.

A safer fallback:

returngetattr(response, "text", None)

This handles OpenAI Pydantic model types (Transcription, TranscriptionVerbose, Translation, etc.) that have a .text attribute directly, without relying on dict conversion.

---

2. Weak test assertions for transcription/translation

assertspan["output"] isnotNone

For transcription and translation tests, the output is never checked beyond being non-None. Given that the cassette response is {"text":"you","usage":{"type":"duration","seconds":1}}, you can assert the actual value:

assertspan["output"] =="you"

This would have caught the _extract_text fallback issue at test-recording time if anything went wrong.

---

Nits / observations (non-blocking)

_make_base_wrapper_callback is placed ~600 lines before BaseWrapper — the string annotation wrapper_cls: type["BaseWrapper"] is correct, but it's a bit of a readability surprise. The existing per-method wrapper functions (_embedding_create_wrapper, etc.) aren't converted to use the factory, which leaves a small inconsistency. Neither is a blocker, but if the factory pattern proves useful it might be worth a follow-up to unify the three inline wrappers.

No coverage for response_format="text" — when the caller passes response_format="text", OpenAI returns a plain str rather than a JSON object. The isinstance(response, str) branch in _extract_text handles this, but there's no test for it. A cassette for this path would be a good addition.

---

Summary

The core logic is correct and all existing + new tests pass. I'd address the _extract_text fallback (issue 1) and strengthen the transcription/translation output assertions (issue 2) before merge. Everything else is minor.

[Abhijeet Prasad (AbhiPrasad)]

Add tracing wrappers for OpenAI's three Audio API methods so that
wrap_openai() and OpenAIIntegration.setup() produce Braintrust spans
for audio.speech.create, audio.transcriptions.create, and
audio.translations.create (sync + async).

Closes#174

[starfolkai]

Drop into this review session: sfk devbox session 68

PR #232 — Audio API Instrumentation: Code Review

Verdict: Approve with minor suggestions

All 438 tests pass (nox -s "test_openai(latest)"), including the 6 new audio tests. The implementation is clean and consistent with the repo's established patterns.

---

What's good

• Follows existing conventions exactly: VCR cassettes, sync/async patcher pairs, registration in both integration.py (for setup()) and _WRAP_TARGETS (for wrap_openai()). The test_wrap_openai_and_setup_use_same_wrappers enforcement is satisfied.
• _AudioFileWrapper._parse_params correctly prevents logging binary audio data by popping file from the prettified copy of params, so the original kwargs passed to the API call are unaffected.
• _make_base_wrapper_callback is a reasonable factory that reduces the repetitive sync/async branching seen in the inline wrappers for chat/embeddings/moderation.
• SpeechWrapper logging {"type": "audio"} for output is the right call — streaming binary audio into a span would be impractical.
• TranscriptionWrapper correctly handles the {"type":"duration","seconds":1} usage shape from Whisper — _parse_metrics_from_usage returns an empty dict gracefully when token fields are absent.

---

Issues

1. _extract_text fallback can silently produce wrong output (minor bug)

tracing.py, _AudioFileWrapper._extract_text:

@staticmethoddef_extract_text(response: Any) ->str|None:
ifisinstance(response, dict):
returnresponse.get("text")
ifisinstance(response, str):
returnresponsereturnstr(response) # ← problem here

When _try_to_dict can't convert the response to a dict (e.g., an object type it doesn't recognize), and the response isn't a plain string, the fallback str(response) will return something like Transcription(text='hello world', logprobs=None, usage=None) — Pydantic's repr — instead of the actual transcribed text. This won't crash, but the span output would be garbage.

A safer fallback:

returngetattr(response, "text", None)

This handles OpenAI Pydantic model types (Transcription, TranscriptionVerbose, Translation, etc.) that have a .text attribute directly, without relying on dict conversion.

---

2. Weak test assertions for transcription/translation

assertspan["output"] isnotNone

For transcription and translation tests, the output is never checked beyond being non-None. Given that the cassette response is {"text":"you","usage":{"type":"duration","seconds":1}}, you can assert the actual value:

assertspan["output"] =="you"

This would have caught the _extract_text fallback issue at test-recording time if anything went wrong.

---

Nits / observations (non-blocking)

_make_base_wrapper_callback is placed ~600 lines before BaseWrapper — the string annotation wrapper_cls: type["BaseWrapper"] is correct, but it's a bit of a readability surprise. The existing per-method wrapper functions (_embedding_create_wrapper, etc.) aren't converted to use the factory, which leaves a small inconsistency. Neither is a blocker, but if the factory pattern proves useful it might be worth a follow-up to unify the three inline wrappers.

No coverage for response_format="text" — when the caller passes response_format="text", OpenAI returns a plain str rather than a JSON object. The isinstance(response, str) branch in _extract_text handles this, but there's no test for it. A cassette for this path would be a good addition.

---

Summary

The core logic is correct and all existing + new tests pass. I'd address the _extract_text fallback (issue 1) and strengthen the transcription/translation output assertions (issue 2) before merge. Everything else is minor.