fix(structured): harden structured output for local servers and thinking models

- Local servers (LM Studio, vLLM) reject the object-form tool_choice langchain
  sends for function calling. The generic openai_compatible provider now binds
  the schema as a tool without forcing tool_choice.
- A structured call can return no parsed result (a thinking model answering in
  plain text); fall back to free text with a clear reason instead of an opaque
  render error.

tests/test_openai_compatible_provider.py

@@ -36,7 +36,7 @@ def test_keyless_local_uses_placeholder_and_chat_completions(monkeypatch):
     llm = create_llm_client(
         provider="openai_compatible", model="qwen2.5", base_url="http://localhost:8000/v1"
     ).get_llm()
-    assert type(llm).__name__ == "NormalizedChatOpenAI"
+    assert type(llm).__name__ == "LocalCompatibleChatOpenAI"
     assert str(llm.openai_api_base) == "http://localhost:8000/v1"
     # keyless local servers: a placeholder key is sent
     key = llm.openai_api_key.get_secret_value() if hasattr(llm.openai_api_key, "get_secret_value") else llm.openai_api_key
@@ -72,3 +72,29 @@ def test_env_backend_url_precedence():
     assert resolve_backend_url("openai", "https://api.openai.com/v1", env_url="http://proxy/v1") == "http://proxy/v1"
     assert resolve_backend_url("openai", "https://api.openai.com/v1", env_url=None) == "https://api.openai.com/v1"
     assert resolve_backend_url("deepseek", None, None) == "https://api.deepseek.com"
+
+
+@pytest.mark.unit
+def test_structured_output_suppresses_object_tool_choice(monkeypatch):
+    # LM Studio / vLLM reject the object-form tool_choice langchain sends for
+    # function-calling structured output (#1057). The generic provider binds the
+    # schema as a tool but must not force tool_choice.
+    from langchain_openai import ChatOpenAI
+    from pydantic import BaseModel
+
+    class Schema(BaseModel):
+        x: int
+
+    captured = {}
+    monkeypatch.setattr(
+        ChatOpenAI,
+        "with_structured_output",
+        lambda self, schema, method=None, **kw: captured.update({"method": method, **kw}) or "BOUND",
+    )
+    llm = create_llm_client(
+        provider="openai_compatible", model="local-llm-30b", base_url="http://localhost:1234/v1"
+    ).get_llm()
+    out = llm.with_structured_output(Schema)
+    assert out == "BOUND"
+    assert captured["method"] == "function_calling"
+    assert captured["tool_choice"] is None  # not the object form

tests/test_structured_agents.py

@@ -121,6 +121,24 @@ def _structured_trader_llm(captured: dict, proposal: TraderProposal | None = Non
     return llm
 
 
+@pytest.mark.unit
+def test_invoke_structured_falls_back_when_result_is_none():
+    # A thinking model can answer in plain text, leaving the parser with None.
+    # That must fall back to free text, not crash on render(None) (#1051).
+    from tradingagents.agents.utils.structured import invoke_structured_or_freetext
+
+    structured = MagicMock()
+    structured.invoke.return_value = None
+    plain = MagicMock()
+    plain.invoke.return_value = MagicMock(content="FREETEXT")
+
+    out = invoke_structured_or_freetext(
+        structured, plain, "prompt", render=lambda r: r.rating, agent_name="t"
+    )
+    assert out == "FREETEXT"
+    plain.invoke.assert_called_once()
+
+
 @pytest.mark.unit
 class TestTraderAgent:
     def test_structured_path_produces_rendered_markdown(self):

tradingagents/agents/utils/structured.py

@@ -63,6 +63,11 @@ def invoke_structured_or_freetext(
     if structured_llm is not None:
         try:
             result = structured_llm.invoke(prompt)
+            if result is None:
+                # A thinking model can answer in plain text instead of calling
+                # the tool, leaving the parser with nothing to return. Treat it
+                # as a structured miss and fall back, with a clear reason.
+                raise ValueError("structured output returned no parsed result")
             return render(result)
         except Exception as exc:
             logger.warning(

tradingagents/llm_clients/openai_client.py

@@ -51,6 +51,23 @@ class NormalizedChatOpenAI(ChatOpenAI):
         return super().with_structured_output(schema, method=method, **kwargs)
 
 
+class LocalCompatibleChatOpenAI(NormalizedChatOpenAI):
+    """OpenAI-compatible client for arbitrary local servers (LM Studio, vLLM,
+    llama.cpp via the generic ``openai_compatible`` provider).
+
+    Their tool-calling support varies, and many reject the object-form
+    ``tool_choice`` langchain sends for function-calling structured output. Bind
+    the schema as a tool but don't force tool_choice, so structured output works
+    across local servers regardless of the model ID's capabilities (#1057).
+    """
+
+    def with_structured_output(self, schema, *, method=None, **kwargs):
+        resolved = method or get_capabilities(self.model_name).preferred_structured_method
+        if resolved == "function_calling":
+            kwargs.setdefault("tool_choice", None)
+        return super().with_structured_output(schema, method=method, **kwargs)
+
+
 def _input_to_messages(input_: Any) -> list:
     """Normalise a langchain LLM input to a list of message objects.
 
@@ -210,7 +227,9 @@ OPENAI_COMPATIBLE_PROVIDERS: dict[str, ProviderSpec] = {
     "ollama":     ProviderSpec(base_url="http://localhost:11434/v1", base_url_env="OLLAMA_BASE_URL",
                                key_optional=True, placeholder_key="ollama"),
     # Generic endpoint: user supplies base_url; key optional (keyless local).
-    "openai_compatible": ProviderSpec(require_base_url=True, key_optional=True),
+    "openai_compatible": ProviderSpec(
+        require_base_url=True, key_optional=True, chat_class=LocalCompatibleChatOpenAI
+    ),
 }