Accept replayed reasoning aliases without rejecting messages

README.md

@@ -109,6 +109,10 @@ The response contains generated text in `choices`. For other clients, use
 base URL `http://127.0.0.1:11500/v1`, model `flint`, and the API key from
 `rec auth token`.
 
+When replaying assistant messages, either `reasoning` or `reasoning_content` is
+accepted. If both are present, `reasoning` takes precedence, including when its
+value is empty or null.
+
 The Anthropic-compatible API is available at
 `http://127.0.0.1:11500/v1/messages`. Point `ANTHROPIC_BASE_URL` at
 `http://127.0.0.1:11500` and send the token from `rec auth token` as a bearer

engine/src/rec_mlx/server.py

@@ -286,8 +286,7 @@ def handle_completion(self, request, stop_words):
             try:
                 for message in request.messages:
                     if "reasoning" in message:
-                        if "reasoning_content" in message:
-                            raise ValueError("Use one reasoning field per message")
+                        # Prefer the public response carrier when clients replay both aliases.
                         message["reasoning_content"] = message.pop("reasoning")
                 process_message_content(request.messages)
                 mode = self.body["grammar"] if "grammar" in self.body else grammar_mode

engine/tests/test_serving.py

@@ -92,6 +92,29 @@ def generate_tokens(self, ids, sampling, *, should_stop, return_logprobs=False,
 
 
 class ServingTest(unittest.TestCase):
+    def test_reasoning_alias_precedence(self):
+        for stream in (False, True):
+            for fields, expected in (
+                ({"reasoning": "a"}, "a"),
+                ({"reasoning_content": "b"}, "b"),
+                ({"reasoning": "a", "reasoning_content": "a"}, "a"),
+                ({"reasoning": "a", "reasoning_content": "b"}, "a"),
+                ({"reasoning": "", "reasoning_content": "b"}, ""),
+                ({"reasoning": None, "reasoning_content": "b"}, None),
+            ):
+                with self.subTest(stream=stream, fields=fields):
+                    with patch.object(self.engine, "prompt", wraps=self.engine.prompt) as prompt:
+                        status, body = self.request(
+                            "B",
+                            max_tokens=1,
+                            stream=stream,
+                            messages=[{"role": "assistant", "content": "B", **fields}],
+                        )
+                        self.assertEqual(status, 200, body)
+                        seen = prompt.call_args.args[0][0]
+                        self.assertEqual(seen["reasoning_content"], expected)
+                        self.assertNotIn("reasoning", seen)
+
     def test_idle_preparation_follows_response_and_yields_to_queued_request(self):
         preparing = threading.Event()
         proceed = threading.Event()