sync update from LAN repo (source dhk-1 @ 3b1ec5b4)

offline-backend-20260512/backend/tests/test_roundtable_model_fallback.py

@@ -20,13 +20,20 @@ from types import SimpleNamespace
 import pytest
 
 import app.gateway.roundtable_model_fallback as mf
+import app.gateway.routers.multi_agent as multi_agent
 import app.gateway.routers.parallel_agents as parallel_agents
 from app.gateway.roundtable_model_fallback import (
     fallback_model_chain,
     is_llm_error_message,
     reason_text,
 )
-from app.gateway.routers._strategy_selection_seed import default_strategy_agent_roles
+from app.gateway.routers._strategy_selection_seed import (
+    STRATEGY_INTENT_AGENT_ID,
+    STRATEGY_JUDGE_AGENT_ID,
+    STRATEGY_SECURITY_AGENT_ID,
+    STRATEGY_SOLUTION_AGENT_ID,
+    default_strategy_agent_roles,
+)
 from deerflow.agents.middlewares.llm_error_handling_middleware import (
     LLM_ERROR_MARKER,
     LLMErrorHandlingMiddleware,
@@ -248,6 +255,38 @@ def test_parallel_agent_substantive_content_accepts_chinese_and_english() -> Non
     assert not parallel_agents._has_substantive_content("<think>internal reasoning only</think>!!!")
 
 
+def test_strategy_stage_tool_surface_allows_clarification_only_for_intent() -> None:
+    assert "ask_clarification" not in parallel_agents._excluded_tools_for_agent(STRATEGY_INTENT_AGENT_ID)
+
+    for agent_id in (
+        STRATEGY_SECURITY_AGENT_ID,
+        STRATEGY_JUDGE_AGENT_ID,
+        STRATEGY_SOLUTION_AGENT_ID,
+    ):
+        assert "ask_clarification" in parallel_agents._excluded_tools_for_agent(agent_id)
+
+
+def test_strategy_task_message_matches_stage_clarification_contract() -> None:
+    judge_message = parallel_agents._build_task_message(
+        "策略评判",
+        "比较候选策略并输出结构化评分。",
+        agent_id=STRATEGY_JUDGE_AGENT_ID,
+    )
+    assert "【本阶段禁止反问】" in judge_message
+    assert "不得调用 `ask_clarification`" in judge_message
+    assert "数据限制与待核验" in judge_message
+    assert "不要重复输出同一正文或多个版本" in judge_message
+
+    intent_message = parallel_agents._build_task_message(
+        "意图理解",
+        "整理用户需求。",
+        agent_id=STRATEGY_INTENT_AGENT_ID,
+    )
+    assert "【需要向用户确认时】" in intent_message
+    assert "调用 `ask_clarification` 工具" in intent_message
+    assert "【本阶段禁止反问】" not in intent_message
+
+
 def test_parallel_agent_extracts_provider_reasoning_without_mixing_answer() -> None:
     answer, reasoning = parallel_agents._extract_ai_stream_parts(
         {
@@ -779,6 +818,149 @@ def test_reason_text() -> None:
     assert reason_text("不存在的码") == "调用失败"
 
 
+@pytest.mark.asyncio
+async def test_thinking_only_seat_does_not_switch_model(monkeypatch) -> None:
+    """Reasoning-only output is an incomplete delivery, not a model failure."""
+    calls: list[str] = []
+
+    async def fake_stream(_client, _base, _headers, _thread_id, body):
+        calls.append(body["context"]["model_name"])
+        yield None, [
+            'data: {"messages":[{"type":"ai","content":"<think>仍在分析约束</think>"}]}'
+        ]
+
+    monkeypatch.setattr(multi_agent, "_stream_upstream", fake_stream)
+    monkeypatch.setattr(
+        multi_agent,
+        "fallback_model_chain",
+        lambda _model: ["slow-thinking-model", "backup-model"],
+    )
+    monkeypatch.setattr(
+        multi_agent,
+        "append_seat_skill_directive",
+        lambda _agent_id, task, **_kwargs: task,
+    )
+    monkeypatch.setattr(multi_agent, "_spawn_detached_broadcast", lambda *_args, **_kwargs: None)
+
+    frames = [
+        frame
+        async for frame in multi_agent._special_run(
+            client=object(),
+            base="http://gateway",
+            headers={},
+            agent_threads={"seat-a": "seat-thread"},
+            agent_name="seat-a",
+            new_message="完成分析",
+            skill_stop_names=[],
+            model_name="slow-thinking-model",
+        )
+    ]
+
+    assert calls == ["slow-thinking-model"]
+    assert not any(
+        isinstance(frame, dict) and frame.get("status") == "model_switch"
+        for frame in frames
+    )
+    assert frames[-1]["status"] == "invalid_delivery"
+    assert frames[-1]["reason"] == "thinking_only"
+
+
+@pytest.mark.asyncio
+async def test_real_model_error_still_switches_to_next_model(monkeypatch) -> None:
+    calls: list[str] = []
+
+    async def fake_stream(_client, _base, _headers, _thread_id, body):
+        model = body["context"]["model_name"]
+        calls.append(model)
+        message = (
+            {
+                "type": "ai",
+                "content": "大模型服务繁忙，请稍后再试。",
+                "additional_kwargs": {LLM_ERROR_MARKER: "busy"},
+            }
+            if model == "broken-model"
+            else {"type": "ai", "content": "这是备用模型完成的有效交付正文。"}
+        )
+        yield None, [f"data: {json.dumps({'messages': [message]}, ensure_ascii=False)}"]
+
+    monkeypatch.setattr(multi_agent, "_stream_upstream", fake_stream)
+    monkeypatch.setattr(
+        multi_agent,
+        "fallback_model_chain",
+        lambda _model: ["broken-model", "backup-model"],
+    )
+    monkeypatch.setattr(
+        multi_agent,
+        "append_seat_skill_directive",
+        lambda _agent_id, task, **_kwargs: task,
+    )
+    monkeypatch.setattr(multi_agent, "_spawn_detached_broadcast", lambda *_args, **_kwargs: None)
+
+    frames = [
+        frame
+        async for frame in multi_agent._special_run(
+            client=object(),
+            base="http://gateway",
+            headers={},
+            agent_threads={"seat-a": "seat-thread"},
+            agent_name="seat-a",
+            new_message="完成分析",
+            skill_stop_names=[],
+            model_name="broken-model",
+        )
+    ]
+
+    assert calls == ["broken-model", "backup-model"]
+    assert any(
+        isinstance(frame, dict) and frame.get("status") == "model_switch"
+        for frame in frames
+    )
+    assert frames[-1]["model_used"] == "backup-model"
+    assert frames[-1]["content"] == "这是备用模型完成的有效交付正文。"
+
+
+@pytest.mark.asyncio
+async def test_background_thinking_only_does_not_switch_model(monkeypatch) -> None:
+    import app.gateway.roundtable_inprocess_gateway as gateway_module
+
+    calls: list[str] = []
+
+    async def fake_run_agent_turn(*_args, **kwargs):
+        calls.append(kwargs["model"])
+        return [{"type": "ai", "content": "<think>继续进行长思考</think>"}]
+
+    monkeypatch.setattr(gateway_module, "_run_agent_turn", fake_run_agent_turn)
+    monkeypatch.setattr(
+        gateway_module,
+        "fallback_model_chain",
+        lambda _model: ["slow-thinking-model", "backup-model"],
+    )
+    gateway = gateway_module.InProcessRoundtableGateway(
+        SimpleNamespace(state=SimpleNamespace()),
+        user_id="u1",
+        agents=[],
+    )
+
+    messages, switches, used_model = await gateway._run_turn_with_fallback(
+        thread_id="seat-thread",
+        agent_name="seat-a",
+        message="完成分析",
+        requested_model="slow-thinking-model",
+        excluded_tools=[],
+        allowed_tools=None,
+        skill_stop_names=[],
+        thinking_enabled=True,
+        reasoning_effort="medium",
+        recursion_limit=1000,
+        stage="step2_seat",
+    )
+
+    assert calls == ["slow-thinking-model"]
+    assert switches == []
+    assert used_model == "slow-thinking-model"
+    assert messages[-1]["content"] == "<think>继续进行长思考</think>"
+
+
 # ── 引擎：单席位异常隔离 + 模型容错记录 → 系统提示气泡 ──────────────────────────
 
 _AGENTS = [SeatRef("seat-a", "情报"), SeatRef("seat-b", "环境"), SeatRef("seat-c", "方案")]
@@ -800,7 +982,15 @@ class _ChainGateway:
         self.switch_agent = switch_agent
         self.seen: list[str] = []
 
-    async def run_leader(self, *, thread_ids, coordinator_name, message, model) -> LeaderTurn:
+    async def run_leader(
+        self,
+        *,
+        thread_ids,
+        coordinator_name,
+        message,
+        model,
+        finalize_only: bool = False,
+    ) -> LeaderTurn:
         return LeaderTurn(final_text="总控说明", dispatched=[])
 
     async def run_seat(self, *, thread_ids, agent_id, task, model) -> SeatTurn: