chore(ai): track Spring AI #5466 — Gemini thinking + tool call broken

[samzhu]

問題描述

候選人使用 Gemini thinking 模型（如 gemini-3.1-flash-lite-preview）搭配 AI tool calling 時，後端回傳 400 錯誤： Function call is missing a thought_signature.

根本原因

上游 bug：spring-projects/spring-ai#5466 GoogleGenAiChatModel.responseCandidateToGeneration()（line 655）使用 allMatch 判斷是否為 function call：

boolean isFunctionCall = parts.stream().allMatch(part -> part.functionCall().isPresent());

Gemini 3 thinking 模型回傳混合型 parts：[thought_part, functionCall_part]。

• allMatch = false → tool call 被整個丟棄
• 下一次請求中 function call 缺少 thought_signature → Gemini API 回 400

目前解法

▎ 已實作於 PatchedGoogleGenAiChatModel.java，參考 spring-projects/spring-ai#5489

繼承 GoogleGenAiChatModel，override responseCandidateToGeneration()，修改判斷邏輯：

// 修復：anyMatch 取代 allMatch
// 混合回應（thinking + function call）時，只要有任意 functionCall part 就視為 tool call
boolean hasFunctionCall = parts.stream().anyMatch(part -> part.functionCall().isPresent());

同時：

• 提取非 function call 的 text parts（thinking summary）合併為 AssistantMessage content
• 保留父類別的 thoughtSignatures 提取與重新附加邏輯，確保 thought_signature 正確傳遞

AiModelRegistry 的 google-genai case 改用 PatchedGoogleGenAiChatModel constructor 直接建立（因 GoogleGenAiChatModel.Builder.build() 固定回傳父類別實例）：

PatchedGoogleGenAiChatModel chatModel = new PatchedGoogleGenAiChatModel(
    genAiClient,
    GoogleGenAiChatOptions.builder().model(entry.id()).build()
);

移除條件
Spring AI 發布包含 spring-projects/spring-ai#5489 修復的版本（預計 2.0.0-M3 或更高）後：

• 升級 Spring AI 版本
• 刪除 PatchedGoogleGenAiChatModel.java
• AiModelRegistry 改回使用 GoogleGenAiChatModel.builder()
• 驗證 Gemini thinking 模型 + tool call 正常運作

參考

• GoogleGenAiChatModel fails to recognize tool calls when combined with text or thinking parts spring-projects/spring-ai#5466
• fix: detect tool calls in mixed content responses for GoogleGenAiChatModel spring-projects/spring-ai#5489
• https://ai.google.dev/gemini-api/docs/thinking
• https://ai.google.dev/gemini-api/docs/thought-signatures

[weijiafu14]

Gemini thinking + tool call is broken across most frameworks right now. The core issue is that thoughtSignature handling has strict requirements that aren't well documented:

1. Signed and unsigned parts cannot coexist in the same message
2. Signatures must be persisted and echoed back across turns
3. Missing signatures need dummy fallback (for abort recovery)

I've documented the full solution and open-sourced it: gemini-claude-bridge — see the architecture in README. Also submitted fixes to LiteLLM (#23809).