[hsusul]

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

Changes being requested

LocalAudioPlayer (in src/openai/helpers/local_audio_player.py) mishandles 1-D float32 NumPy arrays.

Both play() and play_stream() reshape int16 input to (-1, channels), but pass float32 input through without reshaping. A 1-D float32 array — e.g. a mono waveform loaded via soundfile/librosa, and an explicitly declared input type of play() (npt.NDArray[np.float32]) — then breaks:

• play() raises IndexError: tuple index out of range, because the still-1-D array is used as channels=audio_content.shape[1] when constructing the sounddevice.OutputStream.
• play_stream() raises a broadcasting error inside the audio callback when assigning a 1-D slice into the 2-D outdata buffer.

int16 input already works because it is reshaped; only the float32 branch was missing the reshape.

Repro (play())

import asyncio
import numpy as np
from openai.helpers import LocalAudioPlayer

# a mono float32 waveform, e.g. from soundfile / librosa
samples = np.array([0.1, 0.2, 0.3, -0.4, -0.5], dtype=np.float32)
asyncio.run(LocalAudioPlayer().play(samples))
# IndexError: tuple index out of range   (at channels=audio_content.shape[1])

Fix

Reshape float32 input to (-1, self.channels) in both code paths, mirroring the existing int16 handling. Already-2-D input (e.g. the (-1, 1) buffer produced from a streamed TTS response) is unaffected — the reshape is a no-op.

             elif input.dtype == np.float32:
-                audio_content = cast("npt.NDArray[np.float32]", input)
+                audio_content = cast("npt.NDArray[np.float32]", input).reshape(-1, self.channels)

                         if current_buffer.dtype == np.int16 and self.dtype == np.float32:
                             current_buffer = (current_buffer.astype(np.float32) / 32767.0).reshape(-1, self.channels)
+                        elif current_buffer.dtype == np.float32:
+                            current_buffer = current_buffer.reshape(-1, self.channels)

Why this is a hand-maintained file (not generated)

src/openai/helpers/local_audio_player.py has no File generated from our OpenAPI spec by Stainless header and is not part of the OpenAPI surface — it is one of the hand-written helpers under helpers/. The change is a self-contained correctness fix in that helper; it touches no generated code.

Tests

Added tests/lib/test_local_audio_player.py with a lightweight, hardware-free fake for sounddevice.OutputStream that drives the real callback flow from a background thread:

• test_play_accepts_one_dimensional_float32 — reproduces the IndexError (fails before the fix) and asserts the played frames match the input.
• test_play_accepts_one_dimensional_int16 — regression guard for the existing int16 path.
• test_play_stream_accepts_one_dimensional_float32 — covers the streaming path.

Each test fails on main for the exact reason described above and passes with the fix. No network, no API key, no audio device.

Validation (run locally)

• pytest tests/lib/test_local_audio_player.py — 3 passed (and confirmed red on main: IndexError for play, broadcast/hang for play_stream).
• pytest tests/lib/ — 259 passed.
• ruff format --check / ruff check . — clean.
• scripts/run-pyright -p pyproject.toml — 0 errors.
• mypy . — Success: no issues.

Additional context & links

• Compatibility: no public API change; only fixes a previously-crashing input shape. int16 and already-2-D float32 behavior is unchanged.
• local_audio_player.py carries # mypy: ignore-errors, so mypy does not type-check it; the change is pyright-clean.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7603230061

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".

[chatgpt-codex-connector]

Preserve multi-channel float32 playback

When play() receives a 2-D float32 array such as (frames, 2) stereo audio, self.channels is hardcoded to 1, so this reshape flattens interleaved left/right samples into a single channel with twice as many frames. The subsequent OutputStream therefore plays corrupted mono audio instead of the stereo stream supported by the previous audio_content.shape[1] behavior. Reshape only 1-D input, while preserving the channel dimension of existing 2-D arrays.

AGENTS.md reference: AGENTS.md:L5-L8

Useful? React with 👍 / 👎.

[hsusul]

Fixed in 472919a. The float32 reshape now only applies to 1-D input, so a 2-D array keeps its channel dimension and play() still opens the stream with shape[1] channels. The same guard is in play_stream. Added test_play_preserves_two_dimensional_float32_channels, which fails without the guard.

[jnohclee-rgb]

AI-assisted offline comparison at head 7603230061873c3c692069d0b7807246f4a5c3cf against parent 0c09a3fe815184f0a46fbf18b1aba84a467c854e, using Python 3.14.7 and NumPy 2.5.3. I replaced the module's sounddevice binding with a fake OutputStream that invokes the real callback on zero-filled output arrays; no audio devices or network were accessed.

Four play() cases observed:

Input / player setting	Parent	PR head
float32 [1,2,3,4], default mono	IndexError before opening stream	mono samples preserved, tail zero-filled
float32 [[1,10],[2,20]], explicitly set player.channels = 2	two-channel frames preserved	same frames preserved
same two-dimensional array, default LocalAudioPlayer()	two-channel frames preserved	one channel containing [1,10,2,20]
float32 [1,2,3], explicitly set player.channels = 2	IndexError before opening stream	ValueError before opening stream

The one-dimensional fix is confirmed in this harness. The third case may deserve an explicit test or documented contract: the new reshape also flattens already two-dimensional float32 input when its channel dimension differs from self.channels. If existing array shape is meant to determine playback channels, limiting the reshape to one-dimensional inputs would preserve that behavior; if the player setting should be authoritative, the changed behavior should be intentional.

Minimal array for the third case:

player = LocalAudioPlayer()  # channels defaults to 1
samples = np.array([[1, 10], [2, 20]], dtype=np.float32)
await player.play(samples)

Values are fictional markers, not recorded audio. Setting player.channels above is an explicit attribute mutation, not a constructor option. This checks play() only, not threaded native playback, play_stream(), device behavior, or the full suite.

[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 2026-10-19T01:47:00.727013Z	472919a	New commits

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.