`AnthropicBedrock` handles streaming errors differently from Anthropic client

[deepankarm]

When a message sequence has a tool_use without a following tool_result:

1. Anthropic client fails fast with a BadRequestError before stream creation.
2. AnthropicBedrock client allows stream creation but raises a ValueError during streaming.

The bedrock client should also fail fast before creating the stream.

# Define list of messages with a `tool_use` not following a `tool_result`
msgs = [
    {"role": "user", "content": "draw a line"},
    {
        "role": "assistant",
        "content": [
            {"text": "Sure, I can help you draw a line", "type": "text"},
            {"id": "123", "input": {"color": "red"}, "name": "draw_line", "type": "tool_use"},
        ],
    },
    {"role": "user", "content": "draw a circle"},
]

tools = [...] # non-empty list of tools

Anthropic client doesn't allow stream creation (BadRequestError)

from anthropic import AsyncAnthropic

async with AsyncAnthropic(api_key=anthropic_api_key) as client:
    stream = await client.messages.create(
        max_tokens=1024,
        messages=msgs,
        model="claude-3-5-haiku",
        tools=tools,
        stream=True,
    )
    print("stream created!")
    async for message in stream:
        print(message)

"""
BadRequestError: Error code: 400 - {'type': 'error', 'error': {'type': 'invalid_request_error', 'message': 'messages.2: Did not find 1 `tool_result` block(s) at the beginning of this message. Messages following `tool_use` blocks must begin with a matching number of `tool_result` blocks.'}}
"""

AnthropicBedrock client allows stream creation and raises ValueError during streaming

from anthropic import AsyncAnthropicBedrock

async with AsyncAnthropicBedrock() as client:
    stream = await client.messages.create(
        max_tokens=1024,
        messages=msgs,
        model="us.anthropic.claude-3-5-haiku-20241022-v1:0",
        tools=tools,
        stream=True,
    )
    print("stream created!")
    async for message in stream:
        print(message)

"""
stream created!

ValueError: Bad response code, expected 200: {'status_code': 400, 'headers': {':exception-type': 'validationException', ':content-type': 'application/json', ':message-type': 'exception'}, 'body': b'{"message":"messages.2: Did not find 1 `tool_result` block(s) at the beginning of this message. Messages following `tool_use` blocks must begin with a matching number of `tool_result` blocks."}'}
"""

[deepankarm]

More examples

budget_tokens < 1024

from anthropic import AsyncAnthropicBedrock

async with AsyncAnthropicBedrock() as client:
    stream = await client.messages.create(
        max_tokens=2048,
        messages=[{"content": "What is the weather in London?", "role": "user"}],
        model="us.anthropic.claude-3-7-sonnet-20250219-v1:0",
        thinking={"budget_tokens": 512, "type": "enabled"},
        stream=True,
    )
    print("stream created!")
    async for message in stream:
        print(message)

"""
stream created!

ValueError: Bad response code, expected 200: {'status_code': 400, 'headers': {':exception-type': 'validationException', ':content-type': 'application/json', ':message-type': 'exception'}, 'body': b'{"message":"thinking.enabled.budget_tokens: Input should be greater than or equal to 1024"}'}
"""

budget_tokens == max_tokens

from anthropic import AsyncAnthropicBedrock

async with AsyncAnthropicBedrock() as client:
    stream = await client.messages.create(
        max_tokens=2048,
        messages=[{"content": "What is the weather in London?", "role": "user"}],
        model="us.anthropic.claude-3-7-sonnet-20250219-v1:0",
        thinking={"budget_tokens": 2048, "type": "enabled"},
        stream=True,
    )
    print("stream created!")
    async for message in stream:
        print(message)

"""
stream created!

ValueError: Bad response code, expected 200: {'status_code': 400, 'headers': {':exception-type': 'validationException', ':content-type': 'application/json', ':message-type': 'exception'}, 'body': b'{"message":"max_tokens must be greater than thinking.budget_tokens"}'}
"""

[RobertCraigie]

Thanks for the report but unfortunately I don't think we can do anything to fix these differences.

The Anthropic case fails earlier because the API responds with a 400 but in the AnthropicBedrock case, we don't get the error until we start parsing the response (as far as I can tell).