diff --git a/src-tauri/src/ai_model_tools.rs b/src-tauri/src/ai_model_tools.rs
index fdc5a82..11de5fb 100644
--- a/src-tauri/src/ai_model_tools.rs
+++ b/src-tauri/src/ai_model_tools.rs
@@ -19,10 +19,6 @@ const MAGIC_BRUSH_CONTINUATION_WORDS: &[&str] = &[
     "continue", "go ahead", "do it", "try it", "继续", "接着", "开始", "执行", "操作", "试试",
     "你来", "好的", "可以", "行",
 ];
-const MAGIC_BRUSH_MUTATION_WORDS: &[&str] = &[
-    "create", "write", "edit", "update", "delete", "save", "创建", "新建", "写入", "编辑", "修改",
-    "更新", "删除", "保存",
-];
 const MAGIC_BRUSH_TOOL_JSON: &str = r#"{
   "type": "function",
   "function": {
@@ -89,14 +85,7 @@ pub(crate) fn openai_chat_payload(request: &AiModelStreamRequest) -> serde_json:
     });
     if offers_tools {
         payload["tools"] = serde_json::Value::Array(openai_vault_tools());
-        payload["tool_choice"] = serde_json::Value::String(
-            if message_continues_pending_tool_work(&request.message) {
-                "required"
-            } else {
-                "auto"
-            }
-            .into(),
-        );
+        payload["tool_choice"] = serde_json::Value::String("auto".into());
     }
     payload
 }
@@ -235,33 +224,6 @@ fn message_needs_magic_brush(message: &str) -> bool {
             .any(|word| message.to_lowercase().contains(word))
 }
 
-fn message_continues_pending_tool_work(message: &str) -> bool {
-    let latest = message
-        .rsplit("[user]:")
-        .next()
-        .unwrap_or(message)
-        .to_lowercase();
-    if !MAGIC_BRUSH_CONTINUATION_WORDS
-        .iter()
-        .any(|word| latest.contains(word))
-    {
-        return false;
-    }
-
-    let recent = message
-        .chars()
-        .rev()
-        .take(4_000)
-        .collect::<String>()
-        .chars()
-        .rev()
-        .collect::<String>()
-        .to_lowercase();
-    MAGIC_BRUSH_MUTATION_WORDS
-        .iter()
-        .any(|word| recent.contains(word))
-}
-
 fn selected_model_supports_tools(request: &AiModelStreamRequest) -> bool {
     request
         .provider
@@ -944,7 +906,7 @@ mod tests {
         assert!(payload["tools"]
             .as_array()
             .is_some_and(|tools| !tools.is_empty()));
-        assert_eq!(payload["tool_choice"], "required");
+        assert_eq!(payload["tool_choice"], "auto");
         assert_eq!(payload["stream"], false);
     }
 
diff --git a/src-tauri/src/ai_models.rs b/src-tauri/src/ai_models.rs
index c42219f..428b602 100644
--- a/src-tauri/src/ai_models.rs
+++ b/src-tauri/src/ai_models.rs
@@ -220,10 +220,38 @@ where
     }
 
     run_openai_agent_loop(request, payload, emit, |payload| {
-        send_json_request(request, endpoint.clone(), payload)
+        send_openai_payload_with_tool_choice_fallback(payload, |attempt| {
+            send_json_request(request, endpoint.clone(), attempt)
+        })
     })
 }
 
+fn send_openai_payload_with_tool_choice_fallback<S>(
+    mut payload: serde_json::Value,
+    mut send: S,
+) -> Result<serde_json::Value, String>
+where
+    S: FnMut(serde_json::Value) -> Result<serde_json::Value, String>,
+{
+    match send(payload.clone()) {
+        Err(error) if tool_choice_compatibility_error(&error) => {
+            if let Some(object) = payload.as_object_mut() {
+                object.remove("tool_choice");
+            }
+            send(payload)
+        }
+        result => result,
+    }
+}
+
+fn tool_choice_compatibility_error(error: &str) -> bool {
+    let normalized = error.to_ascii_lowercase();
+    normalized.contains("tool_choice")
+        && (normalized.contains("does not support")
+            || normalized.contains("unsupported")
+            || normalized.contains("not supported"))
+}
+
 fn run_openai_agent_loop<F, S>(
     request: &AiModelStreamRequest,
     mut payload: serde_json::Value,
@@ -907,6 +935,33 @@ mod tests {
                 && message.contains("correct the arguments")));
     }
 
+    #[test]
+    fn tool_choice_compatibility_error_retries_without_the_field() {
+        let payload = json!({
+            "messages": [{ "role": "user", "content": "Continue" }],
+            "tools": [{ "type": "function" }],
+            "tool_choice": "auto",
+            "stream": false,
+        });
+        let mut attempts = Vec::new();
+
+        let response = send_openai_payload_with_tool_choice_fallback(payload, |attempt| {
+            attempts.push(attempt);
+            if attempts.len() == 1 {
+                Err("AI provider returned 400 Bad Request: Thinking mode does not support this tool_choice".into())
+            } else {
+                Ok(json!({ "choices": [{ "message": { "content": "continued" } }] }))
+            }
+        })
+        .unwrap();
+
+        assert_eq!(response["choices"][0]["message"]["content"], "continued");
+        assert_eq!(attempts.len(), 2);
+        assert_eq!(attempts[0]["tool_choice"], "auto");
+        assert!(attempts[1].get("tool_choice").is_none());
+        assert!(attempts[1]["tools"].is_array());
+    }
+
     #[test]
     fn agent_loop_reads_then_edits_before_returning_final_text() {
         let dir = tempfile::tempdir().unwrap();