Copilot for Xcode cannot consume image responses from MCP tools (Figma get_screenshot)

在 Copilot for Xcode 的 Agent 模式中重現 Figma MCP get_screenshot 工作流程，並檢查提供的 streamChoices 記錄，以了解影像結果如何處理。在 VSCode 中比較相同的 MCP 回應；完成的標準是 agent 使用具備 Vision 能力的模型接收返回的影像並對其進行推理。

Description
When using Figma MCP inside Copilot for Xcode (Agent mode), the get_screenshot MCP tool successfully returns an image from the MCP server, but Copilot for Xcode fails to consume or interpret the image output, even when using Vision-enabled models (e.g. Claude Opus 4.5).

The same MCP workflow works correctly in VSCode, where the image returned by get_screenshot is properly read and used by the agent. This suggests an issue specific to Copilot for Xcode’s handling of MCP image responses, not the MCP server itself.

Versions

• Copilot for Xcode: 0.46.0
• Xcode: 26.2
• macOS: 26.1

Steps to reproduce

1. Configure and run Figma MCP with a tool that returns an image payload (e.g. get_screenshot).
2. Open Copilot for Xcode and start Agent mode.
3. Trigger an MCP workflow that calls get_screenshot (for example via a command like /start-design-intake with a Figma node ID).
4. Observe that:
   4.1 MCP tool call succeeds
   4.2 Image data is returned by the MCP server
   4.3 Copilot UI shows the tool execution but the agent does not consume or reason over the image
   4.4 Run the same MCP server and workflow in VSCode.
   4.5 Observe that in VSCode the image is correctly read and used by the agent.

Screenshots

Logs

[2026-01-13T18:24:06.344Z] [info] [GitHubCopilot] [29344] window/logMessage: {
  "type" : 4,
  "message" : "[streamChoices] chunk data: {\"choices\":[{\"index\":0,\"delta\":{\"content\":null,\"tool_calls\":[{\"function\":{\"name\":\"mcp_figma-mcp_get_screenshot\"},\"id\":\"toolu_01673Vz98Z4tYmYaYSkEhMDw\",\"index\":1,\"type\":\"function\"}]}}],\"created\":1768328646,\"id\":\"msg_01SK3zJNMPmyrqWFRtTi2d7y\",\"model\":\"claude-opus-4.5\"}\n\n"
}
[2026-01-13T18:24:06.344Z] [info] [GitHubCopilot] [29344] window/logMessage: {
  "message" : "[streamChoices] choice { index: 0, delta: { content: null, tool_calls: [ [Object] ] } }",
  "type" : 4
}

Additional context

• The MCP server is returning a valid image response (same payload works in VSCode).
• Vision-capable models are enabled and selected in Copilot for Xcode.
• This appears to be a Copilot for Xcode limitation or bug in MCP multimodal (image) handling, not a Figma MCP issue.