Now we have server tool and client tool

engine/tests/test_tool_call_flow.py

@@ -136,6 +136,7 @@ async def test_turn_with_tool_call_then_results(monkeypatch):
                     type="tool_call",
                     tool_call={
                         "id": "call_ok",
+                        "executor": "client",
                         "type": "function",
                         "function": {"name": "weather", "arguments": "{\"city\":\"hz\"}"},
                     },
@@ -183,6 +184,7 @@ async def test_turn_with_tool_call_timeout(monkeypatch):
                     type="tool_call",
                     tool_call={
                         "id": "call_timeout",
+                        "executor": "client",
                         "type": "function",
                         "function": {"name": "search", "arguments": "{\"query\":\"x\"}"},
                     },
@@ -217,3 +219,36 @@ async def test_duplicate_tool_results_are_ignored(monkeypatch):
     result = await pipeline._wait_for_single_tool_result("call_dup")
 
     assert result.get("output", {}).get("value") == 1
+
+
+@pytest.mark.asyncio
+async def test_server_calculator_emits_tool_result(monkeypatch):
+    pipeline, events = _build_pipeline(
+        monkeypatch,
+        [
+            [
+                LLMStreamEvent(
+                    type="tool_call",
+                    tool_call={
+                        "id": "call_calc",
+                        "executor": "server",
+                        "type": "function",
+                        "function": {"name": "calculator", "arguments": "{\"expression\":\"1+2\"}"},
+                    },
+                ),
+                LLMStreamEvent(type="done"),
+            ],
+            [
+                LLMStreamEvent(type="text_delta", text="done."),
+                LLMStreamEvent(type="done"),
+            ],
+        ],
+    )
+
+    await pipeline._handle_turn("calc")
+
+    tool_results = [e for e in events if e.get("type") == "assistant.tool_result"]
+    assert tool_results
+    payload = tool_results[-1].get("result", {})
+    assert payload.get("status", {}).get("code") == 200
+    assert payload.get("output", {}).get("result") == 3