Skip to content

Copilot CLI exposes both MCP content and structuredContent #4515

Description

@rroesch1

Describe the bug

Summary

When an MCP tool result contains both content and structuredContent,
Copilot CLI appears to add both fields to the conversation context.

When structuredContent is present, the client should use it as the
structured tool result instead of exposing an equivalent or unrelated
content representation to the model.

For backward compatibility MCP tools which return structured content, should also set the unstructured content

For backwards compatibility, a tool that returns structured content SHOULD also return the serialized JSON in a TextContent block.
see: MCP spec about structured content

Example

server.py

Here is a simple MCP server (based on the MCP Python SDK) to demonstrate the issue:

from pydantic import BaseModel

from mcp.server.mcpserver import MCPServer


class HelloOutput(BaseModel):
    message: str


server = MCPServer(name="hello", version="0.1.0")


@server.tool(structured_output=True)
def hello() -> HelloOutput:
    """Return a simple JSON hello object."""
    return HelloOutput(message="Hello World")


if __name__ == "__main__":
    server.run("stdio")

This is what the MCP server returns for the hello tool call.

{
  "_meta": null,
  "content": [
    {
      "type": "text",
      "text": "{\n  \"message\": \"Hello World\"\n}",
      "annotations": null,
      "_meta": null
    }
  ],
  "structuredContent": {
    "message": "Hello World"
  },
  "isError": false,
  "resultType": "complete"
}

Actual Copilot CLI context

Copilot CLI exposes both representations:

{
  "message": "Hello World"
}

{"message":"Hello World"}

Expected result

Only the structured result should be provided to the model:

{"message":"Hello World"}

The content field may remain available for UI rendering or backwards
compatibility, but it should not be concatenated into the model context when it
duplicates structuredContent.

Impact

The behavior unnecessarily increases context size and may cause the model to
interpret one tool result as two JSON objects.

Related issue

The reporter of issue #3258 expects the opposite behavior.
Anyway, this contradicts the intention of the MCP spec:

Tool results may contain structured or unstructured content.

For backwards compatibility, a tool that returns structured content SHOULD also return the serialized JSON in a TextContent block.

That means:

  • a MCP tool must either return structured or unstructured content (but not both)
  • but they should return a serialized representation in the unstructured content block for backward compatibility (i.e. for legacy clients which don't support structured content)

Clients which support structured content must therefore silently ignore the unstructured content if there is structured content.

Affected version

GitHub Copilot CLI 1.0.80.

Metadata

Metadata

Assignees

No one assigned

    Labels

    area:mcpMCP server configuration, discovery, connectivity, OAuth, policy, and registryarea:toolsBuilt-in tools: file editing, shell, search, LSP, git, and tool call behavior

    Type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions