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):