基于生产环境复核,理解 Crazyrouter 中不同协议暴露的 thinking/reasoning 字段
更新日期:2026-06-06
2026-03-22 已在 Crazyrouter 生产环境复核过的 thinking / reasoning 暴露方式。reasoning itemthinking blockusageMetadata.thoughtsTokenCount 判断思考是否生效| 模型 | 协议 | 已验证字段或标志 | 说明 |
|---|---|---|---|
gpt-5.5 | Responses | output[].type = "reasoning" | 可配合 summary 拿到可展示的摘要 |
claude-opus-4-8 | Anthropic Messages | content[].type = "thinking" | 与 text block 并列返回 |
gemini-3.1-pro | Gemini Native | usageMetadata.thoughtsTokenCount | 说明 thinking budget 已实际消耗 |
Note: 在本轮复核中, gpt-5.5Chat Completions 的message.reasoning_content没有稳定返回可依赖内容,因此这里不把它列为主观察字段。
reasoning item{
"output": [
{
"id": "rs_xxx",
"type": "reasoning",
"encrypted_content": "...",
"summary": [
{
"type": "summary_text",
"text": "..."
}
]
},
{
"type": "message",
"content": [
{
"type": "output_text",
"text": "最终回答"
}
]
}
]
}thinking block{
"content": [
{
"type": "thinking",
"thinking": "..."
},
{
"type": "text",
"text": "最终回答"
}
]
}thoughtsTokenCount{
"usageMetadata": {
"thoughtsTokenCount": 120
}
}reasoning、thinking 还是 thoughtsTokenCount最后校验/修改:2026-09-28(已复核,三个模型 gpt-5.5 / claude-opus-4-8 / gemini-3.1-pro 均在目录中)