fix(voice): handle odd-length audio buffers in StreamedAudioResult

[OiPunk]

Summary

Fixes odd-length audio-buffer crashes in voice streaming when converting PCM chunks with np.frombuffer(..., dtype=np.int16).

Problem

StreamedAudioResult._transform_audio_buffer() assumed a 2-byte aligned buffer. If a provider emits an odd number of bytes in a chunk (observed with some MP3 streaming providers), np.frombuffer raises:

• ValueError: buffer size must be a multiple of element size

Changes

• In StreamedAudioResult._transform_audio_buffer, join buffered chunks and pad with one zero byte when the combined buffer length is odd.
• Added regression tests:
  • odd-length buffer with dtype=np.int16
  • odd-length buffer with dtype=np.float32

Why this is safe

• Padding only happens for odd-length payloads.
• Existing even-length behavior is unchanged.
• Output path for both int16 and float32 remains the same after alignment.

Validation

• uv run ruff check src/agents/voice/result.py tests/voice/test_pipeline.py
• uv run --extra voice pytest tests/voice/test_pipeline.py -q
• uv run --extra voice python -m trace --count --coverdir /tmp/openai_agents_1824_trace --module pytest tests/voice/test_pipeline.py -k odd_length -q
  • confirms new alignment logic is executed in tests

Fixes #1824

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 039083d702

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".

[OiPunk]

Thanks for the careful catch. You were right: zero-padding every odd flush can corrupt sample boundaries when a single PCM sample is split across chunks.

I pushed commit 6b10362 to address this:

• Preserve a trailing byte across intermediate flushes in _stream_audio.
• Only apply zero-padding on the final flush if one dangling byte still remains.
• Added a regression test that reproduces split-sample chunks with buffer_size=1 and verifies we emit a single correct sample instead of two corrupted samples.

Local validation:

• uv run ruff check src/agents/voice/result.py tests/voice/test_pipeline.py
• uv run --extra voice pytest tests/voice/test_pipeline.py -q
• uv run --extra voice python -m trace --count --coverdir /tmp/openai_agents_2456_cr_trace --module pytest tests/voice/test_pipeline.py -q