Interactions: stateless replay of an `mcp_server_tool_result` whose `structured_content` holds a `$ref` returns 400

未关闭
  • 哪里不对？
#3,313 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
3/5
预计耗时
1-2 天
新手友好度
57/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
活跃
技术栈
python
领域
api

调研方向

Start by tracing how the SDK models and serializes mcp_server_tool_result steps for stateless Interactions API requests; the issue does not name specific files or tests. Compare the returned structured_content reference with the replayed function response, then add a regression test using the reproduction shape. Done means replaying the emitted steps succeeds when structured_content contains a $ref.

由索引模型根据 Issue 内容生成。

描述

This issue was posted by Claude Code using claude-opus-5-5 on behalf of David.

Interactions API rejects its own mcp_server_tool_result step on stateless replay when structured_content contains a $ref.

With store: false, sending back the steps a previous interaction returned fails with HTTP 400 when an MCP image result came from a server that also sets structuredContent:

400 {"error":{"message":"The referenced name `structured_content.result[0]` in function_response.response does not match to a display_name in the function_response.parts.","code":"invalid_request"}}

The step is exactly what the API emitted. Image results without structured_content, and audio and embedded-resource results, replay fine (200).

Reproduce (raw HTTP, no SDK)

  1. Run a public MCP server with a tool returning an ImageContent block and matching structuredContent. FastMCP does this for a tool annotated -> list[ImageContent]:

    from mcp.server.fastmcp import FastMCP
from mcp.types import ImageContent

PNG = 'iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR4nGNg+M/A8B8ABf4C/gzLSmsAAAAASUVORK5CYII='
mcp = FastMCP('probe', stateless_http=True)

@mcp.tool()
def get_image() -> list[ImageContent]:
    return [ImageContent(type='image', data=PNG, mimeType='image/png')]

mcp.run(transport='streamable-http')

  2. POST /v1beta/interactions with model: gemini-3.8-flash, store: false, tools: [{"type": "mcp_server", "name": "probe", "url": "<server>/mcp"}], and a user_input asking it to call get_image. The response contains:

    {"type": "mcp_server_tool_result", "call_id": "...", "signature": "...", "name": "probe:get_image", "server_name": "probe",
 "result": {"call_tool_result_json": "{\"content\":[{\"type\":\"image\",...}],\"structuredContent\":{...}}",
            "structured_content": {"result": [{"$ref": "structured_content.result[0]"}]}}}

  3. Send a second request whose input is the first user_input, every returned step unchanged, and a new user_input. It returns the 400 above. Dropping only that tool's call and result steps returns 200.

Related: the live result shape differs from the reference and SDK types
  • The API reference documents mcp_server_tool_result.result as (ImageContent or TextContent) array, object or string. Live results for text, image, audio and embedded-resource tools all arrive as an undocumented object: call_tool_result_json (the MCP CallToolResult as a JSON string), optional structured_content, a plain-text result, and for blob resources a key named after the MIME type ("application/pdf": "<base64>").
  • MCPServerToolResultStepResult in this SDK is an empty model, so none of those fields are typed. The list branch types only text and image items; audio, video and document items parse as UnknownFunctionResultSubcontent.

Observed with google-genai 2.30.0 on the Gemini API (gemini-3.8-flash), November 2026.

主要语言
Python
星标
4.1k
派生
1.1k
平均合并
1 天 21 小时
30 天内合并 PR
103

环境准备

  • 没有 Dockerfile 或 Docker Compose 文件
  • 没有 Pull Request 模板
  • 阅读贡献指南

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

googleapis/python-genai 的其他 Issue

查看 googleapis/python-genai 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。