fix(llm): preserve WatsonX reasoning

pyproject.toml

@@ -18,7 +18,7 @@ dependencies = [
     # fix without it. See _invoke_middleware in
     # cuga/backend/cuga_graph/utils/context_summarizer.py for the measured comparison.
     "langchain>=1.3.9,<1.3.15",
-    "langchain-ibm>=0.3.9",
+    "langchain-ibm>=1.1.1",
     "ibm-watsonx-ai>=1.5.13; python_version>='3.14'",  # 1.5.10 StrEnum.__init__ crashes on 3.14
     "langchain-openai>=1.1.14",  # GHSA-r7w7-9xr2-qq2r: image-token-counting SSRF guard bypassed via DNS rebinding
     "langchain-core>=1.3.3",  # GHSA-pjwx-r37v-7724: unsafe deserialization via broad load() allowlists (<=1.3.2)

src/cuga/backend/cuga_graph/nodes/cuga_agent_core/graph/graph_nodes.py

@@ -139,7 +139,7 @@ def normalize_response(self, response: Any) -> Tuple[str, Optional[str]]:
         """Extract ``(content, reasoning)`` from the model response.
 
         Default passes through ``response.content`` and the
-        ``reasoning_content`` additional kwarg.
+        ``reasoning_content`` or ``reasoning`` additional kwarg.
         Lite overrides to run ``normalize_assistant_text`` and recover
         tool-call code from proxy responses.
 
@@ -151,7 +151,8 @@ def normalize_response(self, response: Any) -> Tuple[str, Optional[str]]:
         framing from a real answer when visible content is empty.
         """
         content = strip_harmony_tokens(response.content or "")
-        reasoning = (getattr(response, "additional_kwargs", None) or {}).get("reasoning_content")
+        additional_kwargs = getattr(response, "additional_kwargs", None) or {}
+        reasoning = additional_kwargs.get("reasoning_content") or additional_kwargs.get("reasoning")
         return content, reasoning
 
     def get_tools_needing_probing(self) -> frozenset[str]:

src/cuga/backend/cuga_graph/nodes/cuga_agent_core/tests/graph/test_graph_adapter_hooks.py

@@ -74,15 +74,23 @@ def test_default_prepare_system_content_returns_base_prompt():
     assert adapter.prepare_system_content(state, {}, "base prompt") == "base prompt"
 
 
-def test_default_normalize_response_extracts_content_and_reasoning():
+@pytest.mark.unit
+@pytest.mark.parametrize(
+    ("additional_kwargs", "expected"),
+    [
+        ({"reasoning_content": "legacy reasoning"}, "legacy reasoning"),
+        ({"reasoning": "WatsonX reasoning"}, "WatsonX reasoning"),
+        ({"reasoning_content": "legacy reasoning", "reasoning": "other reasoning"}, "legacy reasoning"),
+        ({"reasoning_content": "", "reasoning": "WatsonX reasoning"}, "WatsonX reasoning"),
+        ({"reasoning_content": None, "reasoning": "WatsonX reasoning"}, "WatsonX reasoning"),
+    ],
+)
+def test_default_normalize_response_extracts_content_and_reasoning(additional_kwargs, expected):
     adapter = _MinimalAdapter()
-    response = SimpleNamespace(
-        content="hello",
-        additional_kwargs={"reasoning_content": "thought"},
-    )
+    response = SimpleNamespace(content="hello", additional_kwargs=additional_kwargs)
     content, reasoning = adapter.normalize_response(response)
     assert content == "hello"
-    assert reasoning == "thought"
+    assert reasoning == expected
 
 
 def test_default_normalize_response_none_reasoning():

src/cuga/backend/cuga_graph/nodes/cuga_lite/adapter/graph_adapter.py

@@ -178,8 +178,9 @@ def normalize_response(self, response: Any) -> Tuple[str, Optional[str]]:
             if tool_code:
                 logger.warning("Empty content with tool_calls detected; recovering tool call as Python code")
                 content = tool_code
+        additional_kwargs = getattr(response, "additional_kwargs", None) or {}
         reasoning = normalize_assistant_text(
-            (getattr(response, "additional_kwargs", None) or {}).get("reasoning_content")
+            additional_kwargs.get("reasoning_content") or additional_kwargs.get("reasoning")
         )
         return content, reasoning
 

src/cuga/backend/cuga_graph/nodes/cuga_lite/tests/test_agent_graph_adapter.py

@@ -169,14 +169,52 @@ def test_normalize_response_strips_empty_content():
     assert content.strip() == "hello"
 
 
-def test_normalize_response_extracts_reasoning():
+@pytest.mark.unit
+@pytest.mark.parametrize(
+    ("additional_kwargs", "expected"),
+    [
+        ({"reasoning_content": "legacy reasoning"}, "legacy reasoning"),
+        ({"reasoning": "WatsonX reasoning"}, "WatsonX reasoning"),
+        ({"reasoning_content": "legacy reasoning", "reasoning": "other reasoning"}, "legacy reasoning"),
+        ({"reasoning_content": "", "reasoning": "WatsonX reasoning"}, "WatsonX reasoning"),
+        ({"reasoning_content": None, "reasoning": "WatsonX reasoning"}, "WatsonX reasoning"),
+    ],
+)
+def test_normalize_response_extracts_reasoning(additional_kwargs, expected):
     adapter = _make_adapter()
-    response = SimpleNamespace(
-        content="hi",
-        additional_kwargs={"reasoning_content": "I thought about it"},
+    response = SimpleNamespace(content="hi", additional_kwargs=additional_kwargs)
+    content, reasoning = adapter.normalize_response(response)
+    assert content == "hi"
+    assert reasoning == expected
+
+
+@pytest.mark.unit
+@pytest.mark.parametrize("reasoning_key", ["reasoning_content", "reasoning"])
+def test_watsonx_response_reasoning_reaches_tracker(reasoning_key):
+    from langchain_ibm import ChatWatsonx
+
+    # Exercise the installed provider's response conversion without credentials or HTTP.
+    model = ChatWatsonx.model_construct(model_id="openai/gpt-oss-120b")
+    result = model._create_chat_result(
+        {
+            "id": "watsonx-reasoning-regression",
+            "choices": [
+                {
+                    "message": {"role": "assistant", "content": "391", reasoning_key: "17 times 23 is 391."},
+                    "finish_reason": "stop",
+                }
+            ],
+        }
     )
-    _, reasoning = adapter.normalize_response(response)
-    assert reasoning == "I thought about it"
+    adapter = _make_adapter()
+    content, reasoning = adapter.normalize_response(result.generations[0].message)
+    assert content == "391"
+    assert reasoning == "17 times 23 is 391."
+
+    adapter.on_response_processed(SimpleNamespace(), code=None, content=content, reasoning=reasoning)
+    steps = [call.kwargs["step"] for call in adapter.get_tracker().collect_step.call_args_list]
+    assert [step.name for step in steps] == ["Raw_Assistant_Response", "Assistant_reasoning", "Assistant_nl"]
+    assert steps[1].data == "17 times 23 is 391."
 
 
 # ── 6. on_response_processed hook ────────────────────────────────────────