feat(bidi/openai): support image input via BidiImageInputEvent

[cagataycali]

Summary

Adds image-input support to BidiOpenAIRealtimeModel, bringing it to parity with BidiGeminiLiveModel which already handles BidiImageInputEvent.

Previously, sending a BidiImageInputEvent to the OpenAI realtime model raised ValueError: content not supported. OpenAI's Realtime API on gpt-realtime models does accept image inputs as input_image content blocks on conversation.item.create - this PR wires that up.

Motivation

We're building voice agents where the user says "look at me" / "what's on my whiteboard?" and we want to inject a captured camera frame straight into the realtime model's multimodal context, then let it reply in audio.

Gemini Live works out of the box. For OpenAI we had to monkeypatch BidiOpenAIRealtimeModel.send() from outside the SDK - this PR upstreams the fix.

Changes

src/strands/experimental/bidi/models/openai_realtime.py

• Import BidiImageInputEvent.
• Dispatch it in send() to a new _send_image_content() method.
• _send_image_content() encodes the image as a data: URL (data:<mime>;base64,<b64>) and emits a single conversation.item.create event with role=user and an input_image content block.
• Does not auto-trigger response.create — caller decides when to commit (mirrors the audio/buffered semantics rather than text/auto-reply, which matches typical streaming-vision use cases where you send N frames then a follow-up text/audio prompt).

tests/strands/experimental/bidi/models/test_openai_realtime.py

• The existing case that asserted BidiImageInputEvent raises ValueError is replaced with one that verifies the new wire format and that no implicit response.create is sent.

Wire format

{
  "type": "conversation.item.create",
  "item": {
    "type": "message",
    "role": "user",
    "content": [
      {
        "type": "input_image",
        "image_url": "data:image/jpeg;base64,<...>"
      }
    ]
  }
}

This matches OpenAI's documented format for image inputs on gpt-realtime.

Tests

hatch fmt --formatter
hatch fmt --linter           # All checks passed!
hatch test tests/strands/experimental/bidi/models/test_openai_realtime.py
# 27 passed in 5.21s

The broader tests/strands/experimental/bidi/ suite passes except for unrelated pre-existing failures from optional deps (aws_sdk_bedrock_runtime for Nova Sonic, pyaudio for audio I/O) — none touched by this change.

Tenets alignment

• Composability: matches Gemini Live's behavior, same event type, same shape.
• The obvious path is the happy path: agent.send(BidiImageInputEvent(...)) Just Works on OpenAI now.
• Embrace common standards: uses OpenAI's own documented input_image block + data: URL.

Checklist

• Code follows existing patterns (mirrors _send_audio_content / _send_text_content).
• Formatter (hatch fmt --formatter) clean.
• Linter + mypy (hatch fmt --linter) clean.
• Unit tests updated and passing.
• No public API changes other than removing the previous ValueError.

[mkmeral]

/strands review

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[github-actions]

Issue: Per-field assertions on image_item could miss unexpected or regressed fields. Since the entire dict is deterministic, a single equality check would be more robust.

Suggestion:

assert image_item == {
    "type": "message",
    "role": "user",
    "content": [{"type": "input_image", "image_url": f"data:image/jpeg;base64,{image_b64}"}],
}

This catches any unexpected fields that might be added in the future and is more concise.

[github-actions]

Assessment: Approve

Clean, well-scoped addition that brings OpenAI Realtime to parity with Gemini Live for image input. The implementation follows existing patterns exactly (_send_audio_content / _send_text_content) and the behavioral choice of not auto-triggering response.create is consistent with Gemini and well-motivated for streaming vision use cases.

Minor suggestion

• Testing: One inline comment suggesting collapsing per-field assertions into a single dict equality check for robustness against unexpected fields.

Nice contribution — fills a real gap with minimal, focused code. 👍