Skip to content

McpTool drops structuredContent and non-text content from CallToolResult#1443

Description

@akbarkanso

馃敶 Required Information

Describe the Bug:

AbstractMcpTool.wrapCallResult(...) silently discards parts of a successful MCP CallToolResult.

When the result contains at least one TextContent, native McpTool.runAsync(...) returns only the legacy text_output map. It does not expose:

  • CallToolResult.isError on successful results
  • CallToolResult.structuredContent
  • non-text content entries such as images or embedded resources
  • the original ordered content list

This loss is irreversible for callers. An outer BaseTool decorator sees only the already-reduced map and cannot recover the discarded fields without bypassing native McpTool and invoking McpSyncClient itself.

Steps to Reproduce:

  1. Use com.google.adk:google-adk:1.7.1.
  2. Construct a native McpToolset over a streamable-HTTP MCP server.
  3. Have the server advertise a zero-argument tool.
  4. Return this successful tools/call result:
{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "content": [
      {"type": "text", "text": "first"},
      {"type": "text", "text": "second"},
      {
        "type": "image",
        "data": "aW1hZ2U=",
        "mimeType": "image/png"
      }
    ],
    "structuredContent": {"count": 2},
    "isError": false
  }
}
  1. Call the discovered native tool through BaseTool.runAsync(...).

Expected Behavior:

Native MCP tools should preserve the complete semantic CallToolResult, including structured content and non-text content, without breaking callers that consume the existing text_output field.

A compatible resolution could be either:

  1. An additive envelope that retains text_output unchanged while also exposing content, structuredContent, isError, and metadata.
  2. A supported result-mapper/adapter hook on McpToolset or McpTool that receives the raw CallToolResult before reduction.

Observed Behavior:

The returned map contains only:

{text_output=[{text=first}, {text=second}]}

structuredContent, image data and MIME type, the original content parts, and explicit non-error state are absent.

The loss is silent. No exception or warning indicates that valid MCP response data was discarded.

Environment Details:

  • ADK Library Version (see maven dependency): 1.7.1
  • MCP Java SDK resolved by ADK: 1.1.2
  • MCP transport: streamable HTTP
  • Java: 21

馃煛 Optional Information

Regression:

No

Additional Context:

PR #1078 changed wrapCallResult to serialize the complete CallToolResult, but PR #1090 reverted that change to preserve backward compatibility with text_output.

The requested fix is therefore not to remove text_output. It is to preserve compatibility without silently losing the remaining MCP result fields.

This blocks replacing a custom MCP bridge with native McpToolset: the custom implementation currently preserves text, structuredContent, non-text content parts, and error state, while native McpTool does not.

Minimal Reproduction Code:

N/A. The issue can be reproduced using the MCP response above and a native McpToolset over a streamable-HTTP MCP server.

How often has this issue occurred?:

  • Always (100%)

Metadata

Metadata

Assignees

Type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions