[BUG] Gemini 3+ 工具结果包含图片时,后续请求失败("Requests ending with a model turn are not supported")

What version of Kimi Code is running?

0.38.0

Which open platform/subscription were you using?

kimi code

Which model were you using?

gemini-3.7-flash

What platform is your computer?

windows 11

What issue are you seeing?

问题描述

使用 google-genai 协议、gemini-3.7-flash(思考型模型)时,会话在调用工具后(尤其是工具结果包含图片,例如 ReadMediaFile 读取多帧图像),下一个请求会被上游 Gemini API 以 HTTP 400 拒绝:

Requests ending with a model turn are not supported.

该错误对 finishReason 重试无影响,直接终止本轮对话;长会话中只要工具结果带媒体就稳定复现。

原因分析

上游 Gemini 3+ 对 contents 中 functionResponse 的结构有明确规定(官方文档 "Multimodal function responses":媒体部分(图片/PDF 等)必须嵌套在 functionResponse.parts[] 内部,不能在 functionResponse 之后作为兄弟 part 出现)。

当前转换逻辑(messagesToGoogleGenAIContents 中的 toolMessageToFunctionResponseParts)生成的是:

[functionResponse, inlineData, functionResponse, inlineData, ...]

媒体作为 functionResponse 的兄弟 part 紧跟其后。经验证(直接对上游构造请求):

结构	结果
[functionResponse] 纯 functionResponse	✅ 200
[functionResponse, text/media...]	❌ 400 "Requests ending with a model turn are not supported."
functionResponse.parts 内嵌媒体	✅ 200

修复方案

将工具结果中的媒体 part 移入 functionResponse 自身的 parts 数组,即生成:

[{ functionResponse: { name, response, parts: [inlineData/fileData...] } }]

附带问题

部分会话中,并行函数调用只有第一个 functionCall 带 thought_signature,其余调用为 null——如果上游要求每个调用都回传签名,可能还需要在流式解析中补充签名提取逻辑。

What steps can reproduce the bug?

1. 配置 google-genai provider,模型为 gemini-3.7-flash,开启 thinking(参考配置):

   [providers.google]
   type = "google-genai"
   base_url = "https://example.com"
   api_key = "YOUR_API_KEY"
   
   [models."google/gemini-3.7-flash"]
   provider = "google"
   model = "gemini-3.7-flash"
   

2. 在任意工作目录(非 git 仓库亦可)中给 agent 发送一条要求读取图像的 prompt,例如:"读取 gif_frames/f0_i0.png ~ f5_i37.png 这 6 张图片并描述内容"。
3. 模型会并行调用 ReadMediaFile 6 次,每次返回 [text, image_url(data: URI), text]。
4. 工具全部执行成功后,携带图片的工具结果被回传给模型 —— 此时请求 400,错误信息如上,会话终止。

What is the expected behavior?

模型收到工具结果并继续回答。实际:HTTP 400,turnStep= 停留在工具调用后的下一步。

Additional information

No response

Contribution

• I am willing to submit a PR for this bug fix myself (please wait for maintainer approval in this issue first)

Start in the messagesToGoogleGenAIContents conversion, focusing on toolMessageToFunctionResponseParts and how media-bearing tool results are assembled. Reproduce the google-genai request with Gemini 3.7 Flash and image results, then verify that media is nested in functionResponse.parts and the follow-up request no longer returns HTTP 400; the optional thought_signature concern is a separate scope question.