fix: désactiver thinking mode DeepSeek pour fiabiliser le tool calling

Co-authored-by: Cursor <cursoragent@cursor.com>

memento-note/lib/ai/providers/deepseek.ts

@@ -9,9 +9,21 @@ export class DeepSeekProvider implements AIProvider {
 
   constructor(apiKey: string, modelName: string = 'deepseek-chat', embeddingModelName: string = 'deepseek-embedding') {
     // Create OpenAI-compatible client for DeepSeek
+    // Disable extended thinking to ensure reliable tool/function calling
     const deepseek = createOpenAI({
       baseURL: 'https://api.deepseek.com/v1',
       apiKey: apiKey,
+      fetch: async (url, options) => {
+        if (options?.body) {
+          try {
+            const body = JSON.parse(options.body as string)
+            // Disable thinking mode — tool calling is unreliable with it enabled
+            body.thinking = { type: 'disabled' }
+            return fetch(url, { ...options, body: JSON.stringify(body) })
+          } catch { /* ignore parse errors */ }
+        }
+        return fetch(url, options)
+      },
     });
 
     this.model = deepseek.chat(modelName);