feat(agent-core): 补齐提示词编排与结构化解析

tests/conftest.py

@@ -0,0 +1,14 @@
+import pytest
+
+
+@pytest.fixture(autouse=True)
+def force_mock_llm_provider_for_tests(monkeypatch):
+    """
+    测试环境固定使用 mock Provider。
+
+    当前项目会从根目录 `.env` 自动读取真实模型配置，这对本地运行很有帮助，
+    但单元测试和页面回归测试不应该依赖外部网络或真实密钥状态。
+    因此这里统一覆盖为 mock，保证测试稳定、可重复。
+    """
+    monkeypatch.setenv("LLM_PROVIDER", "mock")
+    monkeypatch.setenv("LLM_MODEL", "mock-model")

tests/test_agent_core.py

@@ -1,24 +1,125 @@
-from agent_core.orchestrator import run_agent
+from agent_core.orchestrator import build_messages, run_agent
 from agent_core.rag.ingest import ingest_document
 from agent_core.rag.retriever import retrieve
 
 
-def test_run_agent_returns_structured_mock_result():
+def test_run_agent_returns_structured_result_from_llm_output():
     scenario = {
         "id": "knowledge_qa",
         "name": "知识库问答助手",
+        "agent": {
+            "role": "知识库助手",
+            "goal": "基于资料回答问题",
+            "instructions": ["仅根据证据回答"],
+        },
         "rag": {"enabled": True, "collection": "knowledge_qa", "top_k": 3},
         "tools": ["generate_action_items"],
         "output": {"type": "general_answer"},
     }
+    provider_response = """
+    {
+      "answer": "请先隔离异常现场，再通知负责人。",
+      "confidence": "high",
+      "references": [
+        {"source": "sop.md", "excerpt": "异常处理 SOP：先隔离现场"}
+      ]
+    }
+    """
 
-    result = run_agent(scenario, "如何处理异常？")
+    class FakeProvider:
+        def generate(self, messages, response_format=None):
+            from agent_core.llm_provider import LLMResponse
+
+            return LLMResponse(
+                content=provider_response,
+                model_name="demo-model",
+                success=True,
+            )
+
+    result = run_agent(
+        scenario,
+        "如何处理异常？",
+        options={"llm_provider": FakeProvider()},
+    )
 
     assert result.status == "success"
-    assert result.answer
+    assert result.answer == "请先隔离异常现场，再通知负责人。"
     assert result.structured_output["output_type"] == "general_answer"
+    assert result.structured_output["confidence"] == "high"
     assert isinstance(result.references, list)
     assert result.tool_calls[0]["tool_name"] == "generate_action_items"
+    assert result.model_name == "demo-model"
+
+
+def test_run_agent_falls_back_when_llm_returns_non_json():
+    scenario = {
+        "id": "document_review",
+        "name": "文档审核助手",
+        "agent": {
+            "role": "审核助手",
+            "goal": "总结审核意见",
+            "instructions": ["输出重点问题"],
+        },
+        "rag": {"enabled": False},
+        "tools": [],
+        "output": {"type": "document_review_report"},
+    }
+
+    class FakeProvider:
+        def generate(self, messages, response_format=None):
+            from agent_core.llm_provider import LLMResponse
+
+            return LLMResponse(
+                content="这是非 JSON 的普通回答",
+                model_name="demo-model",
+                success=True,
+            )
+
+    result = run_agent(
+        scenario,
+        "请检查合同风险",
+        options={"llm_provider": FakeProvider()},
+    )
+
+    assert result.status == "success"
+    assert result.answer == "这是非 JSON 的普通回答"
+    assert result.structured_output["output_type"] == "document_review_report"
+    assert result.structured_output["summary"] == "这是非 JSON 的普通回答"
+    assert result.structured_output["parse_mode"] == "fallback"
+
+
+def test_build_messages_contains_role_goal_references_and_tool_results():
+    scenario = {
+        "name": "质量异常分析助手",
+        "agent": {
+            "role": "质量管理专家",
+            "goal": "生成结构化质量分析报告",
+            "instructions": ["必须引用知识库", "缺失信息要说明"],
+        },
+        "output": {"type": "quality_report"},
+    }
+
+    messages = build_messages(
+        scenario_config=scenario,
+        user_input="分析 A 线异常",
+        references=[{"source": "sop.md", "content": "先隔离现场"}],
+        tool_calls=[
+            {
+                "tool_name": "query_demo_records",
+                "success": True,
+                "result": {"records": [{"title": "A线缺陷"}]},
+                "error": "",
+            }
+        ],
+    )
+
+    assert messages[0]["role"] == "system"
+    assert "质量管理专家" in messages[0]["content"]
+    assert "生成结构化质量分析报告" in messages[0]["content"]
+    assert "quality_report" in messages[0]["content"]
+    assert "先隔离现场" in messages[1]["content"]
+    assert "A线缺陷" in messages[1]["content"]
+    assert "分析 A 线异常" in messages[2]["content"]
 
 
 def test_rag_ingest_and_retrieve_filters_by_scenario_and_query(tmp_path):