output_audio crashes with TypeError when Vertex AI returns legacy Lyria response (steps=None); output_text has the fix but output_audio doesn't

[zJvco]

Is this a client library issue or a product issue?

Client library issue. Two related bugs in google-genai that together cause interaction.output_audio to raise TypeError when using Lyria models (lyria-3-clip-preview, lyria-3-pro-preview) via Vertex AI with ADC credentials.

Environment details

• Programming language: Python
• OS: Linux (Ubuntu / WSL2)
• Language runtime version: Python 3.12
• Package version: google-genai==2.7.0
• Auth mode: Vertex AI with ADC (vertexai=True, project, location="global")

Steps to reproduce

• Set up Vertex AI ADC credentials with location="global" (required for Lyria)
• Call client.interactions.create() with a Lyria model (async client)
• Access interaction.output_audio on the returned object

import asyncio
from google import genai

async def main():
    client = genai.Client(vertexai=True, project="YOUR_PROJECT_ID", location="global").aio

    interaction = await client.interactions.create(
        model="lyria-3-clip-preview",
        input="A short upbeat lo-fi hip hop piece with piano.",
    )

    # Crashes here:
    audio = interaction.output_audio

asyncio.run(main())

Error:

File ".../google/genai/_interactions/types/interaction.py", line 281, in output_audio
    for step in reversed(self.steps):
TypeError: 'NoneType' object is not reversible

Root cause analysis

Bug 1 - output_audio missing or [] guard (line 281 of interaction.py)
output_text already handles steps=None safely:

# output_text — line 248 ✓
for step in reversed(self.steps or []):

# output_audio — line 281 ✗
for step in reversed(self.steps):

# output_image — line 269 ✗
for step in reversed(self.steps):

output_audio and output_image are missing the or [] guard that output_text already has. A one-line fix in each would make them consistent.

Bug 2 - is_legacy_lyria_response_body doesn't match Vertex AI resource paths

_maybe_coerce_outputs (which converts the legacy outputs field into steps) only triggers when the response body's model field matches:

# _legacy_lyria.py
LEGACY_LYRIA_MODELS = frozenset({"lyria-3-pro-preview", "lyria-3-clip-preview"})

def is_legacy_lyria_response_body(data):
    model = typed_data.get("model")
    return isinstance(model, str) and model in LEGACY_LYRIA_MODELS

However, the Vertex AI endpoint returns model as a full resource path, e.g.:

projects/5786546123365/locations/global/publishers/google/models/lyria-3-clip-preview

This does not match "lyria-3-clip-preview" exactly, so _maybe_coerce_outputs skips the outputs → steps rewrite. As a result, steps stays None (Pydantic allows it despite the List[Step] annotation) and the outputs field is stored in pydantic_extra instead.
Combined effect: steps=None → output_audio crashes; audio data is accessible but stranded in interaction.model_extra["outputs"].

Expected behavior

interaction.output_audio should return the AudioContent object (or None) without raising TypeError, regardless of whether steps is None.

Suggested fixes

Bug 1 - add or [] to output_audio and output_image (same pattern already used in output_text):

# interaction.py line 281
for step in reversed(self.steps or []):  # was: reversed(self.steps)

# interaction.py line 269
for step in reversed(self.steps or []):  # was: reversed(self.steps)

Bug 2 - make is_legacy_lyria_response_body match the full resource path variant:

def is_legacy_lyria_response_body(data):
    model = typed_data.get("model")
    if not isinstance(model, str):
        return False
    # Match both short name and full Vertex AI resource path
    return any(model == m or model.endswith(f"/models/{m}") for m in LEGACY_LYRIA_MODELS)
Workaround (applied in our codebase)

try:
    audio = interaction.output_audio
except TypeError:
    # steps=None due to Bug 2; fall back to __pydantic_extra__["outputs"]
    from google.genai._interactions.types.audio_content import AudioContent as _AudioContent
    raw_outputs = (interaction.model_extra or {}).get("outputs") or []
    audio = None
    for item in reversed(raw_outputs):
        if isinstance(item, dict) and item.get("type") == "audio":
            audio = _AudioContent.model_validate(item)
            break
        elif hasattr(item, "type") and item.type == "audio":
            audio = item
            break

[kkorpal]

Hi @zJvco ,
Thanks for reaching out and reporting this!
We've tried to reproduce this from our end using the details provided, but we aren't getting any errors and the audio is extracting successfully.

import asyncio
 from google import genai

async def main():
    # Ensure you are on google-genai version 2.7.0
    client = genai.Client(vertexai=True, project="YOUR_PROJECT_ID", location="global").aio

    interaction = await client.interactions.create(
        model="lyria-3-clip-preview",
        input="A short upbeat lo-fi hip hop piece with piano.",
    )

    audio = interaction.output_audio
    
    if audio and audio.data:
        import base64
        audio_bytes = base64.b64decode(audio.data) if isinstance(audio.data, str) else audio.data
        
        with open("output.wav", "wb") as f:
            f.write(audio_bytes)
        print("Success! Audio saved.")

asyncio.run(main())

If this snippet still throws an error in your environment, could you please share a few more details so we can figure out exactly what is going wrong?

The full, runnable code snippet you are using.

The full stack trace / error details you are seeing.

The exact version of the SDK you are running (you can check this with pip show google-genai).

Whether you are routing this through Google AI Studio (API keys) or Vertex AI.

Once we have those details, we can dive right back in. Thanks!