fix(chat): simplify header cards and llm fallback

tests/test_llm_provider.py

@@ -5,6 +5,8 @@ from agent_core.llm_provider import (
     create_llm_provider,
     get_runtime_llm_config,
 )
+from urllib.error import HTTPError
+from io import BytesIO
 
 
 def test_create_llm_provider_requires_api_key_for_openai_compatible():
@@ -72,6 +74,53 @@ def test_openai_compatible_provider_posts_chat_completion(monkeypatch):
     assert captured["headers"]["Authorization"] == "Bearer sk-test"
 
 
+def test_openai_compatible_provider_falls_back_when_response_format_is_rejected(monkeypatch):
+    captured_bodies = []
+
+    class FakeResponse:
+        def __enter__(self):
+            return self
+
+        def __exit__(self, exc_type, exc, traceback):
+            return False
+
+        def read(self):
+            return b'{"choices":[{"message":{"content":"fallback ok"}}],"model":"demo-model"}'
+
+    def fake_urlopen(request, timeout):
+        body = request.data.decode("utf-8")
+        captured_bodies.append(body)
+        if len(captured_bodies) == 1:
+            raise HTTPError(
+                request.full_url,
+                400,
+                "Bad Request",
+                hdrs=None,
+                fp=BytesIO(b'{"error":{"message":"response_format is not supported"}}'),
+            )
+        return FakeResponse()
+
+    monkeypatch.setattr("agent_core.llm_provider.urlopen", fake_urlopen)
+    provider = create_llm_provider(
+        {
+            "LLM_PROVIDER": "openai_compatible",
+            "LLM_API_KEY": "sk-test",
+            "LLM_BASE_URL": "https://api.siliconflow.cn/v1",
+            "LLM_MODEL": "demo-model",
+        }
+    )
+
+    response = provider.generate(
+        [{"role": "user", "content": "hello"}],
+        response_format={"type": "json_object"},
+    )
+
+    assert response.success is True
+    assert response.content == "fallback ok"
+    assert '"response_format"' in captured_bodies[0]
+    assert '"response_format"' not in captured_bodies[1]
+
+
 def test_embedding_provider_uses_openai_compatible_embeddings(monkeypatch):
     class FakeResponse:
         def __enter__(self):