feat(llm): 统一思考模式配置，支持显式禁用状态

src/llm/anthropic.rs

@@ -46,7 +46,8 @@ struct SystemContent {
 struct ThinkingConfig {
     #[serde(rename = "type")]
     thinking_type: String,
-    budget_tokens: u32,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    budget_tokens: Option<u32>,
 }
 
 #[derive(Debug, Serialize, Deserialize, Clone)]
@@ -354,7 +355,10 @@ impl AnthropicClient {
             top_p: self.top_p,
             messages,
             system,
-            thinking: None,
+            thinking: Some(ThinkingConfig {
+                thinking_type: "disabled".to_string(),
+                budget_tokens: None,
+            }),
             stream: false,
         };
 
@@ -409,7 +413,7 @@ impl AnthropicClient {
 
         let thinking = ThinkingConfig {
             thinking_type: "enabled".to_string(),
-            budget_tokens: self.thinking_budget_tokens,
+            budget_tokens: Some(self.thinking_budget_tokens),
         };
 
         // max_tokens must exceed budget_tokens
@@ -596,13 +600,23 @@ mod tests {
     fn test_thinking_config_serialization() {
         let config = ThinkingConfig {
             thinking_type: "enabled".to_string(),
-            budget_tokens: 2048,
+            budget_tokens: Some(2048),
         };
         let json = serde_json::to_string(&config).unwrap();
         assert!(json.contains(r#""type":"enabled""#));
         assert!(json.contains(r#""budget_tokens":2048"#));
     }
 
+    #[test]
+    fn test_thinking_config_disabled_serialization() {
+        let config = ThinkingConfig {
+            thinking_type: "disabled".to_string(),
+            budget_tokens: None,
+        };
+        let json = serde_json::to_string(&config).unwrap();
+        assert_eq!(json, r#"{"type":"disabled"}"#);
+    }
+
     #[test]
     fn test_system_content_serialization() {
         let content = SystemContent {

src/llm/deepseek.rs

@@ -291,13 +291,13 @@ impl DeepSeekClient {
     async fn chat_completion(&self, messages: Vec<Message>) -> Result<String> {
         let url = format!("{}/chat/completions", self.base_url);
 
-        let thinking = if self.thinking_enabled {
-            Some(ThinkingConfig {
-                thinking_type: "enabled".to_string(),
-            })
+        let thinking = Some(ThinkingConfig {
+            thinking_type: if self.thinking_enabled {
+                "enabled".to_string()
             } else {
-            None
-        };
+                "disabled".to_string()
+            },
+        });
 
         // 思考模式下，temperature/top_p 等参数不应传递
         // 非思考模式下可以正常传递

src/llm/kimi.rs

@@ -273,13 +273,13 @@ impl KimiClient {
     async fn chat_completion(&self, messages: Vec<Message>) -> Result<String> {
         let url = format!("{}/chat/completions", self.base_url);
 
-        let thinking = if self.thinking_enabled {
-            Some(ThinkingConfig {
-                thinking_type: "enabled".to_string(),
-            })
+        let thinking = Some(ThinkingConfig {
+            thinking_type: if self.thinking_enabled {
+                "enabled".to_string()
             } else {
-            None
-        };
+                "disabled".to_string()
+            },
+        });
 
         // 对于 kimi-k2.6 等支持思考模式的模型，使用默认 temperature 即可
         // 思考模式下不显式指定 temperature

src/llm/openai.rs

@@ -275,7 +275,11 @@ impl OpenAiClient {
             max_tokens: Some(self.max_tokens),
             temperature: Some(self.temperature),
             top_p: self.top_p,
-            reasoning_effort: None,
+            reasoning_effort: if is_reasoning_model(&self.model) {
+                Some("none".to_string())
+            } else {
+                None
+            },
             stream: false,
         };
 
@@ -619,6 +623,10 @@ pub fn is_valid_model(model: &str) -> bool {
     OPENAI_MODELS.contains(&model)
 }
 
+fn is_reasoning_model(model: &str) -> bool {
+    model.starts_with("o")
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;