OpenRouterProvider.generate_audio rejects format='wav' under hardcoded stream=true

[santoshkumarradha]

Summary

OpenRouterProvider.generate_audio in sdk/python/agentfield/media_providers.py hardcodes "stream": True in the chat-completions audio payload. OpenRouter rejects every format except pcm16 when stream=true, so format="wav" (and mp3, flac, opus) all 400 against models in the gpt-audio family.

The SDK signature advertises format from {"wav", "mp3", "flac", "opus", "pcm16"}, so callers reasonably expect any of those values to work.

Repro

from agentfield.media_providers import OpenRouterProvider
await OpenRouterProvider().generate_audio(
    text="Hello world",
    model="openrouter/openai/gpt-audio-mini",
    format="wav",
)

Error

RuntimeError: OpenRouter audio request failed (400):
{"error":{"message":"Unsupported value: 'audio.format' does not support
'wav' when stream=true. Supported values are: 'pcm16'."}}

Suggested fix

Either:

• auto-switch stream=false when format != "pcm16", or
• request pcm16 over the wire when the caller asks for wav and wrap the streamed pcm16 in the requested container client-side before returning.

Status

This was discovered against agentfield==0.1.84 while building the reel-af example. Looking at origin/main, the second strategy appears to already be implemented (see _wrap_pcm16_as_wav_b64 in media_providers.py and the wire_format = "pcm16" if audio_format == "wav" else audio_format branch around line 1382). Filing for tracking so a release can cut and consumers can drop their workarounds.

Found by the Python SDK consumer reel-af (an AgentField example pipeline).

[santoshkumarradha]

Code on origin/main already addresses this — see PR #579. Filing for tracking so a release cuts and consumers (agentfield==0.1.84 users like reel-af) can drop their workarounds. Once 0.1.85 is on PyPI this can likely be closed.