Merge branch 'main' into fix-2804

backend/tests/test_dangling_tool_call_middleware.py

@@ -14,6 +14,10 @@ def _ai_with_tool_calls(tool_calls):
     return AIMessage(content="", tool_calls=tool_calls)
 
 
+def _ai_with_invalid_tool_calls(invalid_tool_calls):
+    return AIMessage(content="", tool_calls=[], invalid_tool_calls=invalid_tool_calls)
+
+
 def _tool_msg(tool_call_id, name="test_tool"):
     return ToolMessage(content="result", tool_call_id=tool_call_id, name=name)
 
@@ -22,6 +26,16 @@ def _tc(name="bash", tc_id="call_1"):
     return {"name": name, "id": tc_id, "args": {}}
 
 
+def _invalid_tc(name="write_file", tc_id="write_file:36", error="Failed to parse tool arguments: malformed JSON"):
+    return {
+        "type": "invalid_tool_call",
+        "name": name,
+        "id": tc_id,
+        "args": '{"description":"write report","path":"/mnt/user-data/outputs/report.md","content":"bad {"json"}"}',
+        "error": error,
+    }
+
+
 class TestBuildPatchedMessagesNoPatch:
     def test_empty_messages(self):
         mw = DanglingToolCallMiddleware()
@@ -144,6 +158,42 @@ class TestBuildPatchedMessagesPatching:
         assert patched[1].name == "bash"
         assert patched[1].status == "error"
 
+    def test_invalid_tool_call_is_patched(self):
+        mw = DanglingToolCallMiddleware()
+        msgs = [_ai_with_invalid_tool_calls([_invalid_tc()])]
+        patched = mw._build_patched_messages(msgs)
+        assert patched is not None
+        assert len(patched) == 2
+        assert isinstance(patched[1], ToolMessage)
+        assert patched[1].tool_call_id == "write_file:36"
+        assert patched[1].name == "write_file"
+        assert patched[1].status == "error"
+        assert "arguments were invalid" in patched[1].content
+        assert "Failed to parse tool arguments" in patched[1].content
+
+    def test_valid_and_invalid_tool_calls_are_both_patched(self):
+        mw = DanglingToolCallMiddleware()
+        msgs = [
+            AIMessage(
+                content="",
+                tool_calls=[_tc("bash", "call_1")],
+                invalid_tool_calls=[_invalid_tc()],
+            )
+        ]
+        patched = mw._build_patched_messages(msgs)
+        assert patched is not None
+        tool_msgs = [m for m in patched if isinstance(m, ToolMessage)]
+        assert len(tool_msgs) == 2
+        assert {tm.tool_call_id for tm in tool_msgs} == {"call_1", "write_file:36"}
+
+    def test_invalid_tool_call_already_responded_is_not_patched(self):
+        mw = DanglingToolCallMiddleware()
+        msgs = [
+            _ai_with_invalid_tool_calls([_invalid_tc()]),
+            _tool_msg("write_file:36", "write_file"),
+        ]
+        assert mw._build_patched_messages(msgs) is None
+
 
 class TestWrapModelCall:
     def test_no_patch_passthrough(self):

backend/tests/test_gateway_docs_toggle.py

@@ -122,3 +122,45 @@ def test_health_still_works_when_docs_disabled():
         resp = client.get("/health")
         assert resp.status_code == 200
         assert resp.json()["status"] == "healthy"
+
+
+# ---------------------------------------------------------------------------
+# Runtime CORS behavior
+# ---------------------------------------------------------------------------
+
+
+def _make_gateway_client(cors_origins: str) -> TestClient:
+    with patch.dict(os.environ, {"GATEWAY_CORS_ORIGINS": cors_origins}):
+        _reset_gateway_config()
+        from app.gateway.app import create_app
+
+        return TestClient(create_app())
+
+
+def test_gateway_cors_allows_configured_origin():
+    """GATEWAY_CORS_ORIGINS should control actual browser CORS responses."""
+    client = _make_gateway_client("https://app.example")
+
+    response = client.get("/health", headers={"Origin": "https://app.example"})
+
+    assert response.status_code == 200
+    assert response.headers["access-control-allow-origin"] == "https://app.example"
+    assert response.headers["access-control-allow-credentials"] == "true"
+
+
+def test_gateway_cors_rejects_unconfigured_origin():
+    client = _make_gateway_client("https://app.example")
+
+    response = client.get("/health", headers={"Origin": "https://evil.example"})
+
+    assert response.status_code == 200
+    assert "access-control-allow-origin" not in response.headers
+
+
+def test_gateway_cors_normalizes_configured_default_port():
+    client = _make_gateway_client("https://app.example:443")
+
+    response = client.get("/health", headers={"Origin": "https://app.example"})
+
+    assert response.status_code == 200
+    assert response.headers["access-control-allow-origin"] == "https://app.example"

backend/tests/test_gateway_runtime_cleanup.py

@@ -53,6 +53,29 @@ def test_nginx_routes_official_langgraph_prefix_to_gateway_api():
         assert "proxy_pass http://gateway" in content or "proxy_pass http://$gateway_upstream" in content
 
 
+def test_nginx_defers_cors_to_gateway_allowlist():
+    for path in ("docker/nginx/nginx.local.conf", "docker/nginx/nginx.conf"):
+        content = _read(path)
+
+        assert "Access-Control-Allow-Origin" not in content
+        assert "Access-Control-Allow-Methods" not in content
+        assert "Access-Control-Allow-Headers" not in content
+        assert "Access-Control-Allow-Credentials" not in content
+        assert "proxy_hide_header 'Access-Control-Allow-" not in content
+        assert "if ($request_method = 'OPTIONS')" not in content
+
+
+def test_gateway_cors_configuration_uses_gateway_allowlist():
+    gateway_config = _read("backend/app/gateway/config.py")
+    gateway_app = _read("backend/app/gateway/app.py")
+    csrf_middleware = _read("backend/app/gateway/csrf_middleware.py")
+
+    assert not re.search(r"(?<!GATEWAY_)[\"']CORS_ORIGINS[\"']", gateway_config)
+    assert "cors_origins" not in gateway_config
+    assert "get_configured_cors_origins" in gateway_app
+    assert "GATEWAY_CORS_ORIGINS" in csrf_middleware
+
+
 def test_frontend_rewrites_langgraph_prefix_to_gateway():
     next_config = _read("frontend/next.config.js")
     api_client = _read("frontend/src/core/api/api-client.ts")

backend/tests/test_mcp_sync_wrapper.py

@@ -5,7 +5,8 @@ import pytest
 from langchain_core.tools import StructuredTool
 from pydantic import BaseModel, Field
 
-from deerflow.mcp.tools import _make_sync_tool_wrapper, get_mcp_tools
+from deerflow.mcp.tools import get_mcp_tools
+from deerflow.tools.sync import make_sync_tool_wrapper
 
 
 class MockArgs(BaseModel):
@@ -51,14 +52,13 @@ def test_mcp_tool_sync_wrapper_generation():
 
 
 def test_mcp_tool_sync_wrapper_in_running_loop():
-    """Test the actual helper function from production code (Fix for Comment 1 & 3)."""
+    """Test the shared sync wrapper from production code."""
 
     async def mock_coro(x: int):
         await asyncio.sleep(0.01)
         return f"async_result: {x}"
 
-    # Test the real helper function exported from deerflow.mcp.tools
-    sync_func = _make_sync_tool_wrapper(mock_coro, "test_tool")
+    sync_func = make_sync_tool_wrapper(mock_coro, "test_tool")
 
     async def run_in_loop():
         # This call should succeed due to ThreadPoolExecutor in the real helper
@@ -70,16 +70,16 @@ def test_mcp_tool_sync_wrapper_in_running_loop():
 
 
 def test_mcp_tool_sync_wrapper_exception_logging():
-    """Test the actual helper's error logging (Fix for Comment 3)."""
+    """Test the shared sync wrapper's error logging."""
 
     async def error_coro():
         raise ValueError("Tool failure")
 
-    sync_func = _make_sync_tool_wrapper(error_coro, "error_tool")
+    sync_func = make_sync_tool_wrapper(error_coro, "error_tool")
 
-    with patch("deerflow.mcp.tools.logger.error") as mock_log_error:
+    with patch("deerflow.tools.sync.logger.error") as mock_log_error:
         with pytest.raises(ValueError, match="Tool failure"):
             sync_func()
         mock_log_error.assert_called_once()
         # Verify the tool name is in the log message
-        assert "error_tool" in mock_log_error.call_args[0][0]
+        assert mock_log_error.call_args[0][1] == "error_tool"

backend/tests/test_run_journal.py

@@ -339,6 +339,99 @@ class TestConvenienceFields:
         data = j.get_completion_data()
         assert data["first_human_message"] == "What is AI?"
 
+    @pytest.mark.anyio
+    async def test_completion_data_counts_human_ai_and_tool_messages(self, journal_setup):
+        from langchain_core.messages import HumanMessage, ToolMessage
+
+        j, _ = journal_setup
+        j.on_chat_model_start({}, [[HumanMessage(content="Question")]], run_id=uuid4(), tags=["lead_agent"])
+        j.on_llm_end(_make_llm_response("Answer"), run_id=uuid4(), parent_run_id=None, tags=["lead_agent"])
+        j.on_tool_end(ToolMessage(content="Tool result", tool_call_id="call_1", name="search"), run_id=uuid4())
+
+        data = j.get_completion_data()
+
+        assert data["message_count"] == 3
+        assert data["first_human_message"] == "Question"
+        assert data["last_ai_message"] == "Answer"
+
+    @pytest.mark.anyio
+    async def test_tool_call_only_ai_does_not_clear_last_ai_message(self, journal_setup):
+        j, _ = journal_setup
+        j.on_llm_end(_make_llm_response("Useful answer"), run_id=uuid4(), parent_run_id=None, tags=["lead_agent"])
+        j.on_llm_end(
+            _make_llm_response("", tool_calls=[{"id": "call_1", "name": "search", "args": {}}]),
+            run_id=uuid4(),
+            parent_run_id=None,
+            tags=["lead_agent"],
+        )
+
+        data = j.get_completion_data()
+
+        assert data["message_count"] == 2
+        assert data["last_ai_message"] == "Useful answer"
+
+    @pytest.mark.anyio
+    async def test_last_ai_message_extracts_mixed_content_without_extra_newlines(self, journal_setup):
+        j, _ = journal_setup
+        j.on_llm_end(
+            _make_llm_response(
+                [
+                    {"type": "text", "text": "First "},
+                    {"type": "text", "content": "second"},
+                    " third",
+                    {"type": "image", "url": "ignored"},
+                ]
+            ),
+            run_id=uuid4(),
+            parent_run_id=None,
+            tags=["lead_agent"],
+        )
+
+        data = j.get_completion_data()
+
+        assert data["message_count"] == 1
+        assert data["last_ai_message"] == "First second third"
+
+    @pytest.mark.anyio
+    async def test_last_ai_message_extracts_mapping_content(self, journal_setup):
+        j, _ = journal_setup
+        j.on_llm_end(_make_llm_response({"content": "Nested answer"}), run_id=uuid4(), parent_run_id=None, tags=["lead_agent"])
+
+        data = j.get_completion_data()
+
+        assert data["message_count"] == 1
+        assert data["last_ai_message"] == "Nested answer"
+
+    @pytest.mark.anyio
+    async def test_duplicate_llm_run_id_does_not_double_count_message_summary(self, journal_setup):
+        j, _ = journal_setup
+        run_id = uuid4()
+
+        j.on_llm_end(_make_llm_response("Answer", usage=None), run_id=run_id, parent_run_id=None, tags=["lead_agent"])
+        j.on_llm_end(
+            _make_llm_response("Answer", usage={"input_tokens": 10, "output_tokens": 5, "total_tokens": 15}),
+            run_id=run_id,
+            parent_run_id=None,
+            tags=["lead_agent"],
+        )
+
+        data = j.get_completion_data()
+
+        assert data["message_count"] == 1
+        assert data["last_ai_message"] == "Answer"
+        assert data["total_tokens"] == 15
+
+    @pytest.mark.anyio
+    async def test_subagent_ai_does_not_overwrite_lead_last_ai_message(self, journal_setup):
+        j, _ = journal_setup
+        j.on_llm_end(_make_llm_response("Lead answer"), run_id=uuid4(), parent_run_id=None, tags=["lead_agent"])
+        j.on_llm_end(_make_llm_response("Subagent detail"), run_id=uuid4(), parent_run_id=None, tags=["subagent:research"])
+
+        data = j.get_completion_data()
+
+        assert data["message_count"] == 2
+        assert data["last_ai_message"] == "Lead answer"
+
     @pytest.mark.anyio
     async def test_get_completion_data(self, journal_setup):
         j, _ = journal_setup

backend/tests/test_run_manager.py

@@ -5,6 +5,7 @@ import re
 import pytest
 
 from deerflow.runtime import RunManager, RunStatus
+from deerflow.runtime.runs.store.memory import MemoryRunStore
 
 ISO_RE = re.compile(r"^\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}")
 
@@ -141,3 +142,53 @@ async def test_create_defaults(manager: RunManager):
     assert record.kwargs == {}
     assert record.multitask_strategy == "reject"
     assert record.assistant_id is None
+
+
+@pytest.mark.anyio
+async def test_model_name_create_or_reject():
+    """create_or_reject should accept and persist model_name."""
+    from deerflow.runtime.runs.schemas import DisconnectMode
+
+    store = MemoryRunStore()
+    mgr = RunManager(store=store)
+
+    record = await mgr.create_or_reject(
+        "thread-1",
+        assistant_id="lead_agent",
+        on_disconnect=DisconnectMode.cancel,
+        metadata={"key": "val"},
+        kwargs={"input": {}},
+        multitask_strategy="reject",
+        model_name="anthropic.claude-sonnet-4-20250514-v1:0",
+    )
+    assert record.model_name == "anthropic.claude-sonnet-4-20250514-v1:0"
+    assert record.status == RunStatus.pending
+
+    # Verify model_name was persisted to store
+    stored = await store.get(record.run_id)
+    assert stored is not None
+    assert stored["model_name"] == "anthropic.claude-sonnet-4-20250514-v1:0"
+
+    # Verify retrieval returns the model_name via in-memory record
+    fetched = mgr.get(record.run_id)
+    assert fetched is not None
+    assert fetched.model_name == "anthropic.claude-sonnet-4-20250514-v1:0"
+
+
+@pytest.mark.anyio
+async def test_model_name_default_is_none():
+    """create_or_reject without model_name should default to None."""
+    from deerflow.runtime.runs.schemas import DisconnectMode
+
+    store = MemoryRunStore()
+    mgr = RunManager(store=store)
+
+    record = await mgr.create_or_reject(
+        "thread-1",
+        on_disconnect=DisconnectMode.cancel,
+        model_name=None,
+    )
+    assert record.model_name is None
+
+    stored = await store.get(record.run_id)
+    assert stored["model_name"] is None

backend/tests/test_run_repository.py

@@ -249,3 +249,32 @@ class TestRunRepository:
         rows = await repo.list_by_thread("t1", user_id=None)
         assert len(rows) == 2
         await _cleanup()
+
+    @pytest.mark.anyio
+    async def test_model_name_persistence(self, tmp_path):
+        """RunRepository should persist, normalize, and truncate model_name correctly via SQL."""
+        from deerflow.persistence.engine import get_session_factory, init_engine
+
+        url = f"sqlite+aiosqlite:///{tmp_path / 'test.db'}"
+        await init_engine("sqlite", url=url, sqlite_dir=str(tmp_path))
+        repo = RunRepository(get_session_factory())
+
+        await repo.put("run-1", thread_id="thread-1", model_name="gpt-4o")
+        row = await repo.get("run-1")
+        assert row is not None
+        assert row["model_name"] == "gpt-4o"
+
+        long_name = "a" * 200
+        await repo.put("run-2", thread_id="thread-1", model_name=long_name)
+        row2 = await repo.get("run-2")
+        assert row2["model_name"] == "a" * 128
+
+        await repo.put("run-3", thread_id="thread-1", model_name=123)
+        row3 = await repo.get("run-3")
+        assert row3["model_name"] == "123"
+
+        await repo.put("run-4", thread_id="thread-1", model_name=None)
+        row4 = await repo.get("run-4")
+        assert row4["model_name"] is None
+
+        await _cleanup()

backend/tests/test_subagent_executor.py

@@ -291,7 +291,7 @@ class TestAgentConstruction:
         assert captured["agent"]["model"] is model
         assert captured["agent"]["middleware"] is middlewares
         assert captured["agent"]["tools"] == []
-        assert captured["agent"]["system_prompt"] == base_config.system_prompt
+        assert captured["agent"]["system_prompt"] is None  # system_prompt is merged into initial state messages
 
     @pytest.mark.anyio
     async def test_load_skill_messages_uses_explicit_app_config_for_skill_storage(
@@ -331,6 +331,124 @@ class TestAgentConstruction:
         assert len(messages) == 1
         assert "Use demo skill" in messages[0].content
 
+    @pytest.mark.anyio
+    async def test_build_initial_state_consolidates_system_prompt_and_skills(
+        self,
+        classes,
+        base_config,
+        monkeypatch: pytest.MonkeyPatch,
+        tmp_path,
+    ):
+        """_build_initial_state merges system_prompt and skills into one SystemMessage."""
+        SubagentExecutor = classes["SubagentExecutor"]
+
+        skill_dir = tmp_path / "my-skill"
+        skill_dir.mkdir()
+        skill_file = skill_dir / "SKILL.md"
+        skill_file.write_text("Skill instructions here", encoding="utf-8")
+
+        monkeypatch.setattr(
+            sys.modules["deerflow.skills.storage"],
+            "get_or_new_skill_storage",
+            lambda *, app_config=None: SimpleNamespace(load_skills=lambda *, enabled_only: [SimpleNamespace(name="my-skill", skill_file=skill_file, allowed_tools=None)]),
+        )
+
+        executor = SubagentExecutor(
+            config=base_config,
+            tools=[],
+            thread_id="test-thread",
+        )
+
+        state, _filtered_tools = await executor._build_initial_state("Do the task")
+
+        messages = state["messages"]
+        # Should have exactly 2 messages: one combined SystemMessage + one HumanMessage
+        assert len(messages) == 2
+
+        from langchain_core.messages import HumanMessage, SystemMessage
+
+        assert isinstance(messages[0], SystemMessage)
+        assert isinstance(messages[1], HumanMessage)
+        # SystemMessage should contain both the system_prompt and skill content
+        assert base_config.system_prompt in messages[0].content
+        assert "Skill instructions here" in messages[0].content
+        # HumanMessage should be the task
+        assert messages[1].content == "Do the task"
+
+    @pytest.mark.anyio
+    async def test_build_initial_state_no_skills_only_system_prompt(
+        self,
+        classes,
+        base_config,
+        monkeypatch: pytest.MonkeyPatch,
+    ):
+        """_build_initial_state works when there are no skills."""
+        SubagentExecutor = classes["SubagentExecutor"]
+
+        monkeypatch.setattr(
+            sys.modules["deerflow.skills.storage"],
+            "get_or_new_skill_storage",
+            lambda *, app_config=None: SimpleNamespace(load_skills=lambda *, enabled_only: []),
+        )
+
+        executor = SubagentExecutor(
+            config=base_config,
+            tools=[],
+            thread_id="test-thread",
+        )
+
+        state, _filtered_tools = await executor._build_initial_state("Do the task")
+
+        messages = state["messages"]
+        from langchain_core.messages import HumanMessage, SystemMessage
+
+        assert len(messages) == 2
+        assert isinstance(messages[0], SystemMessage)
+        assert base_config.system_prompt in messages[0].content
+        assert isinstance(messages[1], HumanMessage)
+
+    @pytest.mark.anyio
+    async def test_build_initial_state_no_system_prompt_with_skills(
+        self,
+        classes,
+        monkeypatch: pytest.MonkeyPatch,
+        tmp_path,
+    ):
+        """_build_initial_state works when there is no system_prompt but there are skills."""
+        SubagentConfig = classes["SubagentConfig"]
+
+        config = SubagentConfig(
+            name="test-agent",
+            description="Test agent",
+            system_prompt=None,
+            max_turns=10,
+            timeout_seconds=60,
+        )
+
+        skill_dir = tmp_path / "my-skill"
+        skill_dir.mkdir()
+        skill_file = skill_dir / "SKILL.md"
+        skill_file.write_text("Skill content", encoding="utf-8")
+
+        monkeypatch.setattr(
+            sys.modules["deerflow.skills.storage"],
+            "get_or_new_skill_storage",
+            lambda *, app_config=None: SimpleNamespace(load_skills=lambda *, enabled_only: [SimpleNamespace(name="my-skill", skill_file=skill_file, allowed_tools=None)]),
+        )
+
+        SubagentExecutor = classes["SubagentExecutor"]
+        executor = SubagentExecutor(config=config, tools=[], thread_id="test-thread")
+
+        state, _filtered_tools = await executor._build_initial_state("Do the task")
+
+        messages = state["messages"]
+        from langchain_core.messages import HumanMessage, SystemMessage
+
+        assert len(messages) == 2
+        assert isinstance(messages[0], SystemMessage)
+        assert "Skill content" in messages[0].content
+        assert isinstance(messages[1], HumanMessage)
+
 
 # -----------------------------------------------------------------------------
 # Async Execution Path Tests
@@ -514,6 +632,70 @@ class TestAsyncExecutionPath:
         assert result.status == SubagentStatus.COMPLETED
         assert "Task" in result.result
 
+    @pytest.mark.anyio
+    async def test_aexecute_passes_at_most_one_system_message_to_agent(
+        self,
+        classes,
+        base_config,
+        monkeypatch: pytest.MonkeyPatch,
+        tmp_path,
+    ):
+        """Regression: messages sent to agent.astream must contain at most one
+        SystemMessage and it must be the first message.
+
+        This catches any regression where system_prompt would be re-injected
+        via create_agent() (e.g. system_prompt not passed as None) and appear
+        as a second SystemMessage, which providers like vLLM and Xinference
+        reject with "System message must be at the beginning."
+        """
+        from langchain_core.messages import AIMessage, SystemMessage
+
+        SubagentExecutor = classes["SubagentExecutor"]
+        SubagentStatus = classes["SubagentStatus"]
+
+        # Set up a skill so both system_prompt AND skill content are present,
+        # maximising the chance of catching a double-SystemMessage regression.
+        skill_dir = tmp_path / "regression-skill"
+        skill_dir.mkdir()
+        (skill_dir / "SKILL.md").write_text("Skill instruction text", encoding="utf-8")
+
+        monkeypatch.setattr(
+            sys.modules["deerflow.skills.storage"],
+            "get_or_new_skill_storage",
+            lambda *, app_config=None: SimpleNamespace(load_skills=lambda *, enabled_only: [SimpleNamespace(name="regression-skill", skill_file=skill_dir / "SKILL.md", allowed_tools=None)]),
+        )
+
+        captured_states: list[dict] = []
+
+        async def capturing_astream(state, **kwargs):
+            captured_states.append(state)
+            yield {"messages": [AIMessage(content="Done", id="msg-1")]}
+
+        mock_agent = MagicMock()
+        mock_agent.astream = capturing_astream
+
+        executor = SubagentExecutor(
+            config=base_config,
+            tools=[],
+            thread_id="test-thread",
+        )
+
+        with patch.object(executor, "_create_agent", return_value=mock_agent):
+            result = await executor._aexecute("Do something")
+
+        assert result.status == SubagentStatus.COMPLETED
+        assert len(captured_states) == 1, "astream should be called exactly once"
+        initial_messages = captured_states[0]["messages"]
+
+        system_messages = [m for m in initial_messages if isinstance(m, SystemMessage)]
+        assert len(system_messages) <= 1, f"Expected at most 1 SystemMessage but got {len(system_messages)}: {system_messages}"
+        if system_messages:
+            assert initial_messages[0] is system_messages[0], "SystemMessage must be the first message in the conversation"
+            # The consolidated SystemMessage must carry both the system_prompt
+            # and all skill content — nothing should be split across two messages.
+            assert base_config.system_prompt in system_messages[0].content
+            assert "Skill instruction text" in system_messages[0].content
+
 
 class TestSkillAllowedTools:
     @pytest.mark.anyio

backend/tests/test_tool_deduplication.py

@@ -10,7 +10,8 @@ from __future__ import annotations
 
 from unittest.mock import MagicMock, patch
 
-from langchain_core.tools import BaseTool, tool
+from langchain_core.tools import BaseTool, StructuredTool, tool
+from pydantic import BaseModel, Field
 
 from deerflow.tools.tools import get_available_tools
 
@@ -19,6 +20,10 @@ from deerflow.tools.tools import get_available_tools
 # ---------------------------------------------------------------------------
 
 
+class AsyncToolArgs(BaseModel):
+    x: int = Field(..., description="test input")
+
+
 @tool
 def _tool_alpha(x: str) -> str:
     """Alpha tool."""
@@ -52,10 +57,45 @@ def _make_minimal_config(tools):
     config.tools = tools
     config.models = []
     config.tool_search.enabled = False
+    config.skill_evolution.enabled = False
     config.sandbox = MagicMock()
+    config.acp_agents = {}
     return config
 
 
+@patch("deerflow.tools.tools.get_app_config")
+@patch("deerflow.tools.tools.is_host_bash_allowed", return_value=True)
+@patch("deerflow.tools.tools.reset_deferred_registry")
+def test_config_loaded_async_only_tool_gets_sync_wrapper(mock_reset, mock_bash, mock_cfg):
+    """Config-loaded async-only tools can still be invoked by sync clients."""
+
+    async def async_tool_impl(x: int) -> str:
+        return f"result: {x}"
+
+    async_tool = StructuredTool(
+        name="async_tool",
+        description="Async-only test tool.",
+        args_schema=AsyncToolArgs,
+        func=None,
+        coroutine=async_tool_impl,
+    )
+    tool_cfg = MagicMock()
+    tool_cfg.name = "async_tool"
+    tool_cfg.group = "test"
+    tool_cfg.use = "tests.fake:async_tool"
+    mock_cfg.return_value = _make_minimal_config([tool_cfg])
+
+    with (
+        patch("deerflow.tools.tools.resolve_variable", return_value=async_tool),
+        patch("deerflow.tools.tools.BUILTIN_TOOLS", []),
+    ):
+        result = get_available_tools(include_mcp=False, app_config=mock_cfg.return_value)
+
+    assert async_tool in result
+    assert async_tool.func is not None
+    assert async_tool.invoke({"x": 42}) == "result: 42"
+
+
 @patch("deerflow.tools.tools.get_app_config")
 @patch("deerflow.tools.tools.is_host_bash_allowed", return_value=True)
 @patch("deerflow.tools.tools.reset_deferred_registry")