[BUG] MCP: structuredContent overrides content field, losing actual tool response

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

When an MCP server returns a response containing both a content array (with actual tool output) and a structuredContent field (with only metadata), Claude Code reads only structuredContent and discards the content field entirely. The model never receives the actual tool output.

For example, the MCP server returns:

{
  "content": [{"type": "text", "text": "<actual 29,335 characters of response>"}],
  "structuredContent": {"result": "Content length: 29335"},
  "isError": false
}

But the model only receives:

{"result": "Content length: 29335"}

The actual content is silently lost. The same MCP server works correctly with Amazon Q, which reads from the content field.

What Should Happen?

Claude Code should fall back to the content field when structuredContent contains only metadata (like a content length summary) rather than actual tool output.

Error Messages/Logs

Tool call: mcp__coding-mcp__get_coding_standards
Tool result received by model: {"result":"Content length: 29335"}

Actual MCP server response (verified via Amazon Q):
{
  "content": [{"type": "text", "text": "<full coding standards text - 29335 chars>"}],
  "structuredContent": {"result": "Content length: 29335"},
  "isError": false
}

Steps to Reproduce

1. Configure a remote MCP server (SSE transport) that returns both content and structuredContent in tool responses
2. The content field contains the actual text output
3. The structuredContent field contains metadata only (e.g., {"result": "Content length: 29335"})
4. Call any tool on that MCP server from Claude Code
5. Observe: the model only receives structuredContent, the full text from content is never passed through

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.142

Platform

AWS Bedrock

Operating System

Windows

Terminal/Shell

Windows Terminal

Additional Information

• MCP server uses SSE transport (remote)
• All tools on the server exhibit the same behavior: init_context (23050 chars), get_coding_standards (29335 chars), search_knowledge (8767 chars), load_code_generation_context (1225 chars), load_verb_functions_context (10851 chars), load_testing_context (3060 chars)
• The varying content lengths prove the server is generating real content — it's just never reaching the model
• No client-side setting exists to control field preference (confirmed via Claude Code docs)
• The bug is in Claude Code's MCP response handler — it unconditionally prefers structuredContent with no fallback to content