fix(llm): gate MiniMax reasoning_split by model capability (#826)

MinimaxChatOpenAI unconditionally set reasoning_split=True, but the
kwarg is only valid on M2.x reasoning models. The openai SDK's strict
kwarg validation raised TypeError for Coding Plan and any other non-
reasoning MiniMax model.

Adds requires_reasoning_split to ModelCapabilities, gates the payload
injection on it, and only sets True for _MINIMAX_THINKING (M2.x exact
IDs and the ^MiniMax-M\d forward-compat pattern). Same shape as the
existing supports_tool_choice gate.

Regression tests cover both halves: M2.x models still receive the flag,
non-reasoning MiniMax models do not.

tests/test_capabilities.py

@@ -75,6 +75,22 @@ class TestMinimaxExactMatches:
     def test_m2_base_rejects_tool_choice(self):
         assert get_capabilities("MiniMax-M2").supports_tool_choice is False
 
+    def test_m2_x_requires_reasoning_split(self):
+        # M2.x reasoning models need reasoning_split=True so <think> blocks
+        # land in reasoning_details instead of content (#826).
+        for model in ("MiniMax-M2.7", "MiniMax-M2.5-highspeed", "MiniMax-M2"):
+            assert get_capabilities(model).requires_reasoning_split is True
+
+    def test_future_m3_inherits_reasoning_split(self):
+        assert get_capabilities("MiniMax-M3-highspeed").requires_reasoning_split is True
+
+    def test_non_reasoning_minimax_does_not_get_reasoning_split(self):
+        # Coding Plan, MiniMax-Text-01, and any non-M2-prefixed MiniMax model
+        # reject the reasoning_split kwarg via the openai SDK's strict
+        # validation (#826). Default capability has it disabled.
+        for model in ("minimax-text-01", "MiniMax-Coding-Plan", "abab6.5-chat"):
+            assert get_capabilities(model).requires_reasoning_split is False
+
 
 @pytest.mark.unit
 class TestDefault:

tests/test_minimax.py

@@ -42,6 +42,18 @@ class TestMinimaxReasoningSplit:
         # the caller passed. setdefault leaves an existing value alone.
         assert payload.get("reasoning_split") in (False, True)
 
+    def test_non_reasoning_minimax_does_not_inject_reasoning_split(self):
+        """Coding Plan / MiniMax-Text-01 / any non-M2-prefixed model must NOT
+        receive reasoning_split — the openai SDK rejects unknown kwargs with
+        TypeError (#826)."""
+        for model in ("minimax-text-01", "MiniMax-Coding-Plan"):
+            payload = _client(model)._get_request_payload(
+                [HumanMessage(content="hi")]
+            )
+            assert "reasoning_split" not in payload, (
+                f"{model!r} payload unexpectedly contains reasoning_split"
+            )
+
 
 @pytest.mark.unit
 class TestMinimaxStructuredOutputDispatch:

tradingagents/llm_clients/capabilities.py

@@ -38,6 +38,12 @@ class ModelCapabilities:
     # DeepSeek thinking-mode models 400 if reasoning_content from prior
     # assistant turns is not echoed back on the next request.
     requires_reasoning_content_roundtrip: bool = False
+    # MiniMax M2.x reasoning models need ``reasoning_split=True`` so the
+    # <think> block lands in ``reasoning_details`` instead of polluting
+    # ``content``. The flag is rejected by non-reasoning MiniMax models
+    # (Coding Plan, MiniMax-Text-01, etc.), so we only set it where the
+    # model actually consumes it. (#826)
+    requires_reasoning_split: bool = False
 
 
 # DeepSeek's thinking models accept the ``tools`` array but reject the
@@ -74,6 +80,7 @@ _MINIMAX_THINKING = ModelCapabilities(
     supports_json_mode=False,
     supports_json_schema=False,
     preferred_structured_method="function_calling",
+    requires_reasoning_split=True,
 )
 
 _DEFAULT = ModelCapabilities(

tradingagents/llm_clients/openai_client.py

@@ -118,6 +118,11 @@ class MinimaxChatOpenAI(NormalizedChatOpenAI):
     ``reasoning_split=True`` in the request body redirects the thinking
     block into ``reasoning_details`` so ``content`` stays clean.
 
+    The flag is gated by ``ModelCapabilities.requires_reasoning_split``
+    because non-reasoning MiniMax endpoints (Coding Plan, MiniMax-Text-01)
+    reject the parameter via the openai SDK's strict kwarg validation
+    (#826).
+
     Tool-choice handling for M2.x — those models accept only the string
     enum ``{"none", "auto"}`` and reject langchain's function-spec dict —
     is handled by the capability dispatch in
@@ -126,6 +131,7 @@ class MinimaxChatOpenAI(NormalizedChatOpenAI):
 
     def _get_request_payload(self, input_, *, stop=None, **kwargs):
         payload = super()._get_request_payload(input_, stop=stop, **kwargs)
+        if get_capabilities(self.model_name).requires_reasoning_split:
             payload.setdefault("reasoning_split", True)
         return payload