feat: 实现AI拒答机制和语言控制优化

- 添加 RunStatus.REFUSED 状态，AI可明确表达拒答意图
- 优化 System Prompt 使用视觉强调提高AI对拒答规则的遵守
- User Prompt 添加 SCOPE CHECK 和语言约束
- Worker Rules 添加多语言版本角色扮演规则
- Runner 传递 language 参数到 worker stage
- json_finalize 添加语言约束参数
- 更新单元测试匹配新的 prompt 结构

backend/src/core/agentscope/runtime/runner.py

@@ -13,7 +13,10 @@ from agentscope.message import Msg
 from agentscope.tool import Toolkit
 from agentscope.model import OpenAIChatModel
 from core.agentscope.prompts.system_prompt import build_system_prompt
-from core.agentscope.prompts.user_prompt import build_divination_user_prompt
+from core.agentscope.prompts.user_prompt import (
+    build_divination_user_prompt,
+    build_follow_up_user_prompt,
+)
 from core.agentscope.schemas.agui_input import extract_latest_user_payload
 from core.divination import derive_divination
 from core.agentscope.runtime.json_react_agent import JsonReActAgent
@@ -209,6 +212,12 @@ class AgentScopeRunner:
         worker_output_model = resolve_worker_output_model(
             runtime_mode=runtime_mode.value
         )
+        language = "zh-CN"
+        if user_context.settings is not None:
+            prefs = getattr(user_context.settings, "preferences", None)
+            if prefs is not None:
+                language = getattr(prefs, "language", "zh-CN") or "zh-CN"
+
         await self._emit_step_event(
             pipeline=pipeline,
             run_input=run_input,
@@ -222,6 +231,7 @@ class AgentScopeRunner:
                 context_messages=context_messages,
                 run_input=run_input,
                 derived_divination=derived_divination,
+                language=language,
             ),
             toolkit=toolkit,
             run_input=run_input,
@@ -230,6 +240,7 @@ class AgentScopeRunner:
             pipeline=pipeline,
             runtime_mode=runtime_mode,
             derived_divination=derived_divination,
+            language=language,
         )
         worker_output = worker_output_model.model_validate(worker_result.payload)
         await self._emit_step_event(
@@ -253,6 +264,7 @@ class AgentScopeRunner:
         pipeline: PipelineLike,
         runtime_mode: RuntimeMode,
         derived_divination: DerivedDivinationData | None,
+        language: str,
     ) -> StageExecutionResult:
         tracking_model = self._build_model(stage_config=stage_config)
         formatter = OpenAIChatFormatter()
@@ -265,11 +277,6 @@ class AgentScopeRunner:
             emit_text_events=True,
             emit_tool_events=False,
         )
-        language = "zh-CN"
-        if user_context.settings is not None:
-            prefs = getattr(user_context.settings, "preferences", None)
-            if prefs is not None:
-                language = getattr(prefs, "language", "zh-CN") or "zh-CN"
 
         system_prompt = build_system_prompt(
             agent_type=stage_config.agent_type,
@@ -285,6 +292,7 @@ class AgentScopeRunner:
             base_messages=[Msg("system", system_prompt, "system"), *input_messages],
             output_model=worker_output_model,
             retries=2,
+            language=language,
         )
         worker_payload = worker_output_model.model_validate(worker_payload_raw)
         response_metadata = self._llm_pricing_service.build_usage_metadata(
@@ -316,11 +324,17 @@ class AgentScopeRunner:
         context_messages: list[Msg],
         run_input: RunAgentInput,
         derived_divination: DerivedDivinationData | None,
+        language: str = "zh-CN",
     ) -> list[Msg]:
         if derived_divination is not None:
-            user_text = build_divination_user_prompt(derived=derived_divination)
+            user_text = build_divination_user_prompt(
+                derived=derived_divination, language=language
+            )
         else:
-            user_text, _ = extract_latest_user_payload(run_input)
+            raw_user_text, _ = extract_latest_user_payload(run_input)
+            user_text = build_follow_up_user_prompt(
+                question=raw_user_text, language=language
+            )
 
         if derived_divination is not None and context_messages:
             last = context_messages[-1]