模型支持协议扩展参数，用于关闭思考模式

DeepSeek 现役的 v4 系列全部默认开启思考模式，没有单独的非思考型模型
可换；官方给的关闭方式是在请求里带 thinking={"type":"disabled"}。

ai_models 表早有 extra_json 字段但从未接线，这里接通：模型配置里的
JSON 对象会合并进每次调用的请求体，因此任何协议级参数都能在管理端
配置，不必改代码。合并不会覆盖 model/messages/max_tokens/temperature/
stream —— 预算和结构由代码掌握，重试逻辑才不会被配置绕过。
写入前校验 JSON，避免非法值存进 JSON 列后每次调用才报错。

线上模型已配置为关闭思考：同一会话下由 10133 ms / 1024 输出 tokens、
回复被截断，变为 669 ms / 10 tokens、回复完整。max_tokens 相应回调至 512。

管理端模型表单新增「协议参数」输入框（该目录尚未纳入版本库，仅已发布）。

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

im/backend/internal/app/ai_providers.go

@@ -68,6 +68,10 @@ type aiModel struct {
 	IsDefault   bool    `json:"isDefault"`
 	APIKey      string  `json:"-"`
 	HasAPIKey   bool    `json:"hasApiKey"`
+	// Provider-specific request fields, merged into the call body. This is how a
+	// model that thinks by default is asked not to, without a code change:
+	// {"thinking":{"type":"disabled"}}.
+	Extra map[string]any `json:"extra,omitempty"`
 	CreatedAt   string  `json:"createdAt"`
 	UpdatedAt   string  `json:"updatedAt"`
 }
@@ -169,6 +173,19 @@ func aiRequestMaxTokens(req aiChatRequest, model aiModel) int {
 	return 256
 }
 
+// aiWithExtra adds the model's configured protocol fields to a request body.
+// It never overwrites what the caller set: the structural fields and the budget
+// this code reasons about stay under this code's control.
+func aiWithExtra(model aiModel, body map[string]any) map[string]any {
+	for key, value := range model.Extra {
+		if _, taken := body[key]; taken {
+			continue
+		}
+		body[key] = value
+	}
+	return body
+}
+
 // One retry, with enough room for the thinking plus a short answer, and a cap
 // so a misconfigured model cannot bill an unbounded completion.
 func aiRetryMaxTokens(current int) int {
@@ -242,13 +259,13 @@ func (p openAIProvider) Chat(ctx context.Context, req aiChatRequest) (aiChatResu
 	}
 	raw, err := aiPostJSON(ctx, p.client, aiEndpoint(p.base, "/chat/completions"),
 		map[string]string{"Authorization": "Bearer " + p.model.APIKey},
-		map[string]any{
+		aiWithExtra(p.model, map[string]any{
 			"model":       p.model.ModelName,
 			"messages":    messages,
 			"max_tokens":  aiRequestMaxTokens(req, p.model),
 			"temperature": aiRequestTemperature(req, p.model),
 			"stream":      false,
-		})
+		}))
 	if err != nil {
 		return aiChatResult{}, err
 	}