Thanks to the Python team for the Bedrock connector (#9100) and for dropping the model info check in #12395. Because of that, the Claude Opus 5.5 profile IDs load with no code change.
Describe the bug
With no thinking settings, Claude Opus 5.5 on Bedrock sends a reasoningContent block before the text or toolUse blocks whenever the model decides to think, and BedrockChatCompletion fails on it. This depends on the prompt (raw converse: "Weather in Paris? brief" 5/5, "Hello" 0/5), so a trivial prompt usually works. Opus 5 fails the same way (there even "Hello" had the block 3/3).
case (main ca40aa7, live us-east-1) |
result |
us.anthropic.claude-opus-5-5, "Weather in Paris? brief", non-streaming |
❌ 0/5 ServiceInvalidResponseError: Unsupported content type in the response: {'reasoningContent': ...} |
| same, streaming |
❌ 0/5 KeyError: 'toolUse' |
global.anthropic.claude-opus-5-5, both modes |
❌ 0/3 each, same errors |
us.anthropic.claude-opus-5, both modes |
❌ 0/3 each, same errors |
Opus 5.5, FunctionChoiceBehavior.Auto, model thinks before calling tools |
❌ 2/2 ServiceInvalidResponseError (auto_invoke False and True) |
| Opus 5.5, auto-invoke, model thinks on the turn after the tool result |
❌ 3/3 ServiceInvalidResponseError |
| Opus 5.5, "Hello" |
✅ 3/3 (no reasoning block) |
| Claude Haiku 4.5, Nova Lite |
✅ text returned |
Cause, in python/semantic_kernel/connectors/ai/bedrock/services/bedrock_chat_completion.py:
_create_chat_message_content handles text, image and toolUse, and raises for anything else (L257).
_parse_content_block_delta_event treats every delta without text as toolUse (L346-348). The reasoning delta is {"reasoningContent": {"text": ""}}, so it fails with a KeyError.
To Reproduce
# AWS_DEFAULT_REGION=us-east-1 python repro.py
import asyncio
from semantic_kernel.connectors.ai.bedrock import BedrockChatCompletion, BedrockChatPromptExecutionSettings
from semantic_kernel.contents import ChatHistory
async def main():
svc = BedrockChatCompletion(model_id="us.anthropic.claude-opus-5-5")
history = ChatHistory()
history.add_user_message("Weather in Paris? brief")
print(await svc.get_chat_message_content(history, BedrockChatPromptExecutionSettings(max_tokens=300)))
asyncio.run(main())
semantic_kernel.exceptions.service_exceptions.ServiceInvalidResponseError: Unsupported content type in the response: {'reasoningContent': {'reasoningText': {'text': '', 'signature': 'CA...'}}}
Failed 3/3 runs; raw converse returns ['reasoningContent', 'text'].
Tool-call rows: prompts and output
One get_weather(city) kernel function, FunctionChoiceBehavior.Auto, max_tokens=1000:
prompt A: "I am deciding between Paris, Lyon and Nice for a picnic tomorrow; think about which cities matter, check the weather where needed, and recommend one."
prompt B: "Think about whether tomorrow suits a picnic in Paris: check the weather with the tool, then recommend yes or no."
A Auto(auto_invoke=False) runs 1-2: ServiceInvalidResponseError: Unsupported content type in the response: {'reasoningContent': ...
A Auto(auto_invoke=True) runs 1-2: ServiceInvalidResponseError: Unsupported content type in the response: {'reasoningContent': ...
B Auto() runs 1-3: ServiceInvalidResponseError: Unsupported content type in the response: {'reasoningContent': ...
Expected behavior
The text reply or the tool calls come back. The connector doesn't surface reasoning today, so skipping these blocks would be enough, as the Anthropic connector does (it reads only TextBlock and tool-use blocks, anthropic_chat_completion.py L267-269). Locally that makes the ❌ rows pass, except that auto-invoke with parallel tool calls then hits the separate #14073. I have a small patch with unit tests and can open a PR.
Platform
- Language: Python
- Source: main branch (ca40aa7, 1.44.1)
- AI model: Bedrock
us.anthropic.claude-opus-5-5, global.anthropic.claude-opus-5-5, us.anthropic.claude-opus-5
- IDE: none (script)
- OS: macOS
Additional context
Thanks to the Python team for the Bedrock connector (#9100) and for dropping the model info check in #12395. Because of that, the Claude Opus 5.5 profile IDs load with no code change.
Describe the bug
With no thinking settings, Claude Opus 5.5 on Bedrock sends a
reasoningContentblock before the text ortoolUseblocks whenever the model decides to think, andBedrockChatCompletionfails on it. This depends on the prompt (rawconverse: "Weather in Paris? brief" 5/5, "Hello" 0/5), so a trivial prompt usually works. Opus 5 fails the same way (there even "Hello" had the block 3/3).mainca40aa7, live us-east-1)us.anthropic.claude-opus-5-5, "Weather in Paris? brief", non-streamingServiceInvalidResponseError: Unsupported content type in the response: {'reasoningContent': ...}KeyError: 'toolUse'global.anthropic.claude-opus-5-5, both modesus.anthropic.claude-opus-5, both modesFunctionChoiceBehavior.Auto, model thinks before calling toolsServiceInvalidResponseError(auto_invokeFalse and True)ServiceInvalidResponseErrorCause, in
python/semantic_kernel/connectors/ai/bedrock/services/bedrock_chat_completion.py:_create_chat_message_contenthandlestext,imageandtoolUse, and raises for anything else (L257)._parse_content_block_delta_eventtreats every delta withouttextastoolUse(L346-348). The reasoning delta is{"reasoningContent": {"text": ""}}, so it fails with aKeyError.To Reproduce
Failed 3/3 runs; raw
conversereturns['reasoningContent', 'text'].Tool-call rows: prompts and output
One
get_weather(city)kernel function,FunctionChoiceBehavior.Auto,max_tokens=1000:Expected behavior
The text reply or the tool calls come back. The connector doesn't surface reasoning today, so skipping these blocks would be enough, as the Anthropic connector does (it reads only
TextBlockand tool-use blocks,anthropic_chat_completion.pyL267-269). Locally that makes the ❌ rows pass, except that auto-invoke with parallel tool calls then hits the separate #14073. I have a small patch with unit tests and can open a PR.Platform
us.anthropic.claude-opus-5-5,global.anthropic.claude-opus-5-5,us.anthropic.claude-opus-5Additional context