fix: local inference provider loses thinking and fragments tool calls across turns (#8415)

Signed-off-by: jh-block <jhugo@block.xyz>

ui/desktop/openapi.json

@@ -6486,6 +6486,9 @@
             "nullable": true,
             "minimum": 0
           },
+          "enable_thinking": {
+            "type": "boolean"
+          },
           "flash_attention": {
             "type": "boolean",
             "nullable": true

ui/desktop/src/api/types.gen.ts

@@ -818,6 +818,7 @@ export type ModelInfoResponse = {
 
 export type ModelSettings = {
     context_size?: number | null;
+    enable_thinking?: boolean;
     flash_attention?: boolean | null;
     frequency_penalty?: number;
     max_output_tokens?: number | null;