feat: save account model and thinking settings for Hermes

tests/unit/test_hermes_llm.py

@@ -99,7 +99,7 @@ class HermesLLMTests(unittest.IsolatedAsyncioTestCase):
         self.assertEqual(request.headers["x-hermes-session-id"], "stable-room")
         self.assertEqual(request.headers["x-hermes-session-key"], "hope:alice")
         self.assertEqual(json.loads(request.content), {
-            "model": "hermes-agent", "stream": True,
+            "model": "auto", "provider": "custom:hope-upstream", "stream": True,
             "messages": [
                 {"role": "system", "content": "EXACT system prompt\nDo not alter."},
                 {"role": "user", "content": "hello\nworld"},
@@ -347,6 +347,32 @@ class HermesLLMTests(unittest.IsolatedAsyncioTestCase):
         self.assertEqual(result.text, "Done")
         self.assertEqual(len(self.requests), 1)
 
+    async def test_saved_model_and_thinking_refreshed_per_turn(self):
+        selected = {'model': 'openai/gpt-example', 'thinking_level': 'high'}
+        async def settings():
+            return dict(selected)
+        bridge = self.bridge(FakeSSE([data({'content': 'Done'}), b'data: [DONE]\n\n']), get_model_settings=settings)
+        for effort in ('high', 'none', 'low'):
+            selected['thinking_level'] = effort
+            await bridge.chat(chat_ctx=self.context(), extra_kwargs={'provider': 'evil', 'model': 'evil'}).collect()
+            body = json.loads(self.requests[-1].content)
+            self.assertEqual(body['model'], 'openai/gpt-example')
+            self.assertEqual(body['provider'], 'custom:hope-upstream')
+            expected = {'enabled': False} if effort == 'none' else {'enabled': True, 'effort': effort}
+            self.assertEqual(body['model_options']['reasoning'], expected)
+        selected['model'] = 'intel-b70-1/Qwen3.8-27b'
+        await bridge.chat(chat_ctx=self.context()).collect()
+        self.assertEqual(json.loads(self.requests[-1].content)['model'], selected['model'])
+
+    async def test_settings_failure_does_not_call_hermes_or_leak(self):
+        async def broken():
+            raise RuntimeError('private-key')
+        with self.assertRaises(HermesLLMError) as caught:
+            await self.bridge(FakeSSE([]), get_model_settings=broken).chat(chat_ctx=self.context()).collect()
+        self.assertEqual(caught.exception.code, 'settings_unavailable')
+        self.assertNotIn('private-key', str(caught.exception))
+        self.assertEqual(self.requests, [])
+
     async def test_transport_exception_redacted(self):
         def broken(request):
             self.requests.append(request)