Python: [Bug]: BedrockChatClient drops images from user messages, so GPT-6 Sol/Luna/Astra answer without seeing them

[kimnamu]

Description

Thanks for the Bedrock connector and for the careful tool-result handling in _convert_content_to_bedrock_block. As far as I can tell, image content in a user message (Content.from_data(..., media_type="image/png")) is dropped before the Converse call, so the model answers without seeing the image. I noticed it while trying the new OpenAI GPT-6 Sol/Luna/Astra models, which accept image input on Bedrock.

Red 64x64 PNG + "What color is this image? One word.", us-east-1	BedrockChatClient (main @ f42ff01)	Raw boto3 converse, same bytes
us.openai.gpt-6-sol	❌ 'Unknown.', user blocks sent: ['text']	'Red'
us.openai.gpt-6-luna	❌ 'Gray', user blocks sent: ['text']	'Red'
us.openai.gpt-5.6-sol	❌ 'Unknown', user blocks sent: ['text']	'Red'
us.openai.gpt-6-astra	❌ 'Unknown.', user blocks sent: ['text']	'Red'
global.openai.gpt-6-astra	❌ 'Unknown', user blocks sent: ['text']	'Red'

The guesses vary between runs (Luna also said 'Black.'); the image is never sent.

Where: _convert_content_to_bedrock_block in python/packages/bedrock/agent_framework_bedrock/_chat_client.py (L549-585) matches text, function_call and function_result; everything else hits case _ ("Bedrock does not support other content types at this time") and is skipped with a debug log (L544). Converse accepts image blocks with raw bytes. The Anthropic client already maps image data content via has_top_level_media_type("image"), and Gemini uses _get_data_bytes for the bytes.

Impact: vision use cases silently send a text-only request and the model still answers, so it looks like a model quality problem. The drop happens before any model is called, so it isn't specific to GPT-6.

Suggested fix: for user messages, add a case "data" if content.has_top_level_media_type("image") branch that returns {"image": {"format": <subtype>, "source": {"bytes": _get_data_bytes(content)}}}. Images in assistant messages and other data types stay skipped as today. I have a small PR with regression tests and will link it here.

Code Sample

import asyncio, pathlib
from agent_framework import Content, Message
from agent_framework.amazon import BedrockChatClient

async def main():
    client = BedrockChatClient(model="us.openai.gpt-6-sol", region="us-east-1")
    png = pathlib.Path("red.png").read_bytes()  # 64x64 red
    msg = Message(role="user", contents=[
        Content.from_text("What color is this image? One word."),
        Content.from_data(data=png, media_type="image/png"),
    ])
    print((await client.get_response([msg])).text)

asyncio.run(main())

Error Messages / Stack Traces

# BedrockChatClient, main @ f42ff01 (request captured with a botocore provide-client-params hook)
us.openai.gpt-6-sol -> 'Unknown.' | user blocks sent ['text']
us.openai.gpt-6-luna -> 'Gray' | user blocks sent ['text']
us.openai.gpt-5.6-sol -> 'Unknown' | user blocks sent ['text']
us.openai.gpt-6-astra -> 'Unknown.' | user blocks sent ['text']
global.openai.gpt-6-astra -> 'Unknown' | user blocks sent ['text']

# raw boto3 converse, same PNG
us.openai.gpt-6-sol raw image -> ['Red']
us.openai.gpt-6-luna raw image -> ['Red']
us.openai.gpt-5.6-sol raw image -> ['Red']
us.openai.gpt-6-astra raw image -> ['Red']
global.openai.gpt-6-astra raw image -> ['Red']

No error is raised; only a debug log line.

Package Versions

agent-framework-core: 1.19.0, agent-framework-bedrock: 1.0.0b260918 (main @ f42ff01), boto3: 1.43.100

Python Version

Python 3.12

Additional Context

Why only user messages, and what is out of scope

An image block in an assistant message is rejected by GPT-6 Sol, Luna and Astra on Converse: ValidationException: This model doesn't support the image content block that you provided. Main skips those today and the request succeeds, so the fix keeps skipping them.

Images inside tool results are a separate path (dropped with a warning on purpose). HTTP image URLs are also out of scope, because Converse takes only bytes or an S3 location.

[github-actions]

🤖 Automated triage reproduction notes (agent-authored — trust but verify)

Agent analysis

The bug reproduces in python/packages/bedrock/agent_framework_bedrock/_chat_client.py::_convert_content_to_bedrock_block at lines 549-585 when a user message contains Content.from_data(..., media_type="image/png"); the conversion returns None, and _convert_message_to_content_blocks drops it. Minimal repro: call BedrockChatClient.get_response with text and PNG contents using a stub runtime, then assert the captured Converse request includes an image block.

• Failing test: python/packages/bedrock/tests/test_bedrock_client.py
• Files examined: python/AGENTS.md, python/packages/bedrock/AGENTS.md, python/packages/bedrock/pyproject.toml, python/packages/bedrock/agent_framework_bedrock/_chat_client.py, python/packages/bedrock/tests/test_bedrock_client.py
• Tests run: python/packages/bedrock/tests/test_bedrock_client.py::test_get_response_preserves_user_image_data_in_converse_request
• Reported version: 1.0.0b260918
• Current version: 1.0.0b260918