[BUG] Empty toolResult content array causes ValidationException with Nemotron (and potentially other strict models) via Bedrock Converse API

[ghhamel]

Checks

• I have updated to the lastest minor and patch version of Strands
• I have checked the documentation and this is not expected behavior
• I have searched ./issues and there are no duplicates of my issue

Strands Version

1.35.0

Python Version

3.13 (Debian) & 3.14.2 (macOS)

Operating System

Debian (python:3.13-slim via ECS Fargate) & macOS 26

Installation Method

pip

Steps to Reproduce

When a tool returns no content (e.g., an MCP tool returns an empty result), the SDK passes toolResult.content: [] (empty array) to the Bedrock Converse API. This causes a ValidationException with Nemotron models.

Minimal reproducer using the Strands SDK with a tool that returns empty content:

from strands import Agent, tool
from strands.models import BedrockModel

@tool
def list_tables() -> dict:
    """List all tables in the database.

    Returns:
        A dict with status and content containing table names.
    """
    # Returns a pre-formatted tool result with empty content array.
    # This is the same format that MCP tools produce when they return
    # a successful result with no content items (e.g., postgres-mcp-server
    # run_query returning an empty result set).
    #
    # The SDK's _wrap_tool_result passes dicts with "status" and "content"
    # keys through directly without normalization.
    return {
        "status": "success",
        "content": [],
    }

# FAILS with Nemotron:
agent = Agent(
    model=BedrockModel(model_id="nvidia.nemotron-super-3-120b", streaming=False),
    tools=[list_tables],
)
result = agent("What tables are in the database?")

# SUCCEEDS with Claude (same tool, same code):
agent = Agent(
    model=BedrockModel(model_id="us.anthropic.claude-sonnet-4-5-20250929-v1:0", streaming=False),
    tools=[list_tables],
)
result = agent("What tables are in the database?")

Output with Nemotron (strands-agents 1.35.0):

Tool #1: list_tables
cycle failed
...
botocore.errorfactory.ValidationException: An error occurred (ValidationException) when calling the
Converse operation: The model returned the following errors:
{"error":{"code":"validation_error","message":"Failed to deserialize the JSON body into the target type:
?[2]: Invalid 'messages': missing field `content` at line 1 column 308","param":null,
"type":"invalid_request_error"}}
└ Bedrock region: us-east-1
└ Model id: nvidia.nemotron-super-3-120b

Output with Claude Sonnet 4.5 (same code, same tool):

Tool #1: list_tables
SUCCESS: Agent responded without error.

In a real Strands agent scenario, this happens when:

1. Agent calls an MCP tool (e.g., postgres-mcp-server run_query)
2. The MCP tool returns a successful result but with no content items
3. _handle_tool_result in mcp_client.py maps the empty content to content: []
4. The event loop constructs a toolResult message with content: []
5. _format_bedrock_messages passes it through to the Converse API unchanged
6. Nemotron's endpoint rejects it; Claude accepts it

Expected Behavior

The SDK should normalize empty toolResult.content arrays before sending to the Converse API, ensuring cross-model compatibility. A toolResult with content: [] should be replaced with content: [{"text": ""}].

Actual Behavior

Nemotron returns a ValidationException:

botocore.errorfactory.ValidationException: An error occurred (ValidationException) when calling the
Converse operation: The model returned the following errors:
{"error":{"code":"validation_error","message":"Failed to deserialize the JSON body into the target type:
?[2]: Invalid 'messages': missing field `content` at line 1 column 308","param":null,
"type":"invalid_request_error"}}
└ Bedrock region: us-east-1
└ Model id: nvidia.nemotron-super-3-120b

Key observations from the error:

• The error format {"error":{"code":"validation_error","type":"invalid_request_error"}} is OpenAI-style, originating from Nemotron's inference endpoint — not from Bedrock's own validation layer.
• Bedrock accepts the request (no client-side validation error), but the Converse-to-Nemotron translation layer produces a message missing the required content field when it encounters an empty content array.
• Claude Sonnet 4.5 processes the identical payload successfully.
• The AWS ToolResultBlock docs mark content as Required: Yes but do not specify a minimum array length, so [] is technically valid per the spec.

Additional Context

This was discovered with a Strands agent using nvidia.nemotron-super-3-120b via the Bedrock Converse API. The agent used the awslabs/postgres-mcp-server MCP tool, which returned empty content for a successful run_query call. The error was 100% reproducible — every tool use cycle that produced an empty tool result triggered the failure.

Possible Solution

Normalize empty toolResult.content arrays in _format_bedrock_messages in src/strands/models/bedrock.py. This follows existing normalization patterns in the same function (filtering SDK_UNKNOWN_MEMBER, dropping DeepSeek reasoningContent) and matches the TypeScript SDK's approach in bedrock.ts → _formatMessages.

# In _format_bedrock_messages, before formatting content blocks:
if "toolResult" in content_block:
    if not content_block["toolResult"].get("content"):
        content_block["toolResult"]["content"] = [{"text": ""}]

This is:

• Model-agnostic and defensive — protects against all strict model providers, not just Nemotron
• Consistent with existing SDK normalization patterns in _format_bedrock_messages
• Consistent with the TypeScript SDK's _formatMessages which already filters empty content
• A single-point fix at the model-provider boundary

Related Issues

• TypeScript SDK fix pattern: sdk-typescript/src/models/bedrock.ts → _formatMessages - Python SDK normalization: sdk-python/src/strands/models/bedrock.py → _format_bedrock_messages - Empty content test pattern: sdk-python/tests/strands/experimental/bidi/models/test_nova_sonic.py