Description

When using Claude models through Anthropic's Vertex AI integration with the beta messages API and streaming enabled, tool input parameters are not populated correctly. The SDK returns a plain Stream object instead of BetaMessageStreamManager, which bypasses the event accumulation logic entirely. This causes all tool calls to fail with validation errors.

Environment

• Library: anthropic[vertex] v0.64.0
• Python: 3.8+
• Vertex AI Region: us-east5 (or any region)
• Models affected: All Claude models via Vertex (e.g., claude-3-5-sonnet@20240620)

Steps to Reproduce

from anthropic import AnthropicVertex

# Initialize client
client = AnthropicVertex(
    project_id="your-project",  # Replace with your GCP project ID
    region="us-east5"
)

# Define a tool
tools = [{
    "name": "get_weather",
    "description": "Get the current weather for a location",
    "input_schema": {
        "type": "object",
        "properties": {
            "location": {
                "type": "string",
                "description": "The city to get weather for"
            }
        },
        "required": ["location"]
    }
}]

# This fails - tool inputs are empty
response = client.beta.messages.create(
    model="claude-3-5-sonnet@20240620",
    max_tokens=1024,
    messages=[{"role": "user", "content": "What's the weather in Paris?"}],
    tools=tools,
    stream=True  # Issue only occurs with streaming
)

# Process the stream
for event in response:
    if hasattr(event, 'type') and event.type == 'content_block_start':
        if hasattr(event, 'content_block'):
            if event.content_block.type == 'tool_use':
                print(f"Tool: {event.content_block.name}")
                print(f"Input: {event.content_block.input}")  # Will be {} instead of {"location": "Paris"}

Expected Behavior

Tool inputs should be properly populated from the input_json_delta events:

# Expected
{"location": "Paris"}

Actual Behavior

Tool inputs remain empty:

# Actual
{}

This causes validation errors:

Input validation error: 'location' is a required property

Root Cause Analysis

The issue is in /src/anthropic/lib/vertex/_beta_messages.py:

1. The Vertex beta messages implementation delegates directly to FirstPartyMessagesAPI.create
2. This returns a plain Stream object instead of BetaMessageStream with accumulation logic
3. The accumulation logic in BetaMessageStream is never reached
4. Vertex DOES send proper input_json_delta events, but they're not accumulated

Debug output showing the issue:

# With current implementation
type(client.beta.messages.create(..., stream=True))
# <class 'anthropic._streaming.Stream'>  ❌ Wrong!

# Should be wrapped in BetaMessageStreamManager
# which returns BetaMessageStream that accumulates deltas

Impact

• Severity: High - Makes streaming completely unusable with tools in Vertex AI beta API
• Scope: All Vertex AI users using beta messages with tools and streaming
• Workaround: Disable streaming (stream=False) but loses streaming benefits

Workaround

Use non-streaming mode (but this limits the thinking budget to 4096 in Opus 4):

response = client.beta.messages.create(
    # ... same parameters ...
    stream=False  # Works correctly without streaming
)

Solution Implemented

The Vertex beta messages implementation now properly wraps streaming responses:

Changes Made

1. create() method: When stream=True, wraps the raw Stream in BetaMessageStream
2. stream() method: Returns BetaMessageStreamManager as expected
3. Async versions: Same fixes applied to AsyncMessages

# In /src/anthropic/lib/vertex/_beta_messages.py
if stream is True:
    raw_stream = FirstPartyMessagesAPI.create(..., stream=True)
    # Wrap in BetaMessageStream for accumulation
    return BetaMessageStream(raw_stream)

Additional Context

• The regular (non-beta) Vertex messages API has a similar issue
• Vertex correctly sends input_json_delta events - they're just not being processed
• The accumulation logic exists and works, it's just not being used due to wrong stream class
• Issue confirmed with extensive testing and debugging

Related Issues

• Not the same as Server tool input reconstruction missing in streaming #954 (server tool use) - this affects regular tool_use blocks
• Not related to Empty args/inputs when turning streaming on and setting tool choice to any #658 (Sentry interference) - occurs in clean environments