feat: critic agent in observability

packages/narada-core/src/narada_core/actions/critic.py

@@ -74,10 +74,39 @@ async def run_critic(
         if critic_action_trace_raw is not None
         else None
     )
+    critic_workflow_trace = critic_content.get("workflowTrace")
 
     return CriticResult(
         validation_passed=validation_passed,
         structured_output=structured_output,
         usage=AgentUsage.model_validate(critic_dispatch_response["usage"]),
         action_trace=critic_action_trace,
+        workflow_trace=critic_workflow_trace,
     )
+
+
+def merge_critic_workflow_trace(
+    *,
+    workflow_trace: dict[str, Any] | None,
+    critic_result: CriticResult | None,
+) -> dict[str, Any] | None:
+    critic_workflow_trace = (
+        critic_result.workflow_trace if critic_result is not None else None
+    )
+    if critic_workflow_trace is None:
+        return workflow_trace
+
+    if workflow_trace is None:
+        return critic_workflow_trace
+
+    children = workflow_trace.get("children")
+    if not isinstance(children, list):
+        return workflow_trace
+
+    return {
+        **workflow_trace,
+        "children": [
+            *children,
+            {"kind": "sub_workflow", "trace": critic_workflow_trace},
+        ],
+    }

packages/narada-core/src/narada_core/actions/models.py

@@ -40,10 +40,13 @@ class StructuredOutput(BaseModel, Generic[_StructuredOutputT]):
 
 
 class CriticResult(BaseModel):
+    model_config = ConfigDict(populate_by_name=True)
+
     validation_passed: bool
     structured_output: Any
     usage: AgentUsage
     action_trace: tracing_model.ActionTrace | None = None
+    workflow_trace: dict[str, Any] | None = Field(default=None, alias="workflowTrace")
 
 
 class AgentResponse(BaseModel, Generic[_StructuredOutputT]):

packages/narada-pyodide/src/narada/window.py

@@ -21,7 +21,7 @@
 from urllib.parse import urlencode
 
 from js import AbortController, setTimeout  # type: ignore
-from narada_core.actions.critic import run_critic
+from narada_core.actions.critic import merge_critic_workflow_trace, run_critic
 from narada_core.actions.models import (
     DEFAULT_HITL_TIMEOUT_SECONDS,
     AgenticMatchingSelectorsFinderRequest,
@@ -709,10 +709,6 @@ async def agent(
         )
         workflow_trace = response_content.get("workflowTrace")
         parent_request_id = self._current_parent_request_id()
-        # Preserve the response contract for direct callers, but avoid adding a second
-        # child node when the backend will stitch the child request into the parent row.
-        if workflow_trace is not None and parent_request_id is None:
-            _trace.emit_sub_workflow(workflow_trace=workflow_trace)
 
         critic_result: CriticResult | None = None
         if critic is not None:
@@ -725,6 +721,15 @@ async def agent(
                 time_zone=time_zone,
                 timeout=timeout,
             )
+            workflow_trace = merge_critic_workflow_trace(
+                workflow_trace=workflow_trace,
+                critic_result=critic_result,
+            )
+
+        # Preserve the response contract for direct callers, but avoid adding a second
+        # child node when the backend will stitch the child request into the parent row.
+        if workflow_trace is not None and parent_request_id is None:
+            _trace.emit_sub_workflow(workflow_trace=workflow_trace)
 
         return AgentResponse(
             request_id=remote_dispatch_response["requestId"],

packages/narada-pyodide/tests/test_cloud_browser.py

@@ -450,6 +450,90 @@ async def test_window_agent_exposes_workflow_trace_alias(
     ]
 
 
+@pytest.mark.asyncio
+async def test_window_agent_emits_combined_critic_workflow_trace(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    workflow_trace = {
+        "workflowId": "main-workflow",
+        "workflowName": "Main Workflow",
+        "runtime": "gui",
+        "status": "success",
+        "startTs": 100,
+        "children": [],
+    }
+    critic_workflow_trace = {
+        "workflowId": "critic-workflow",
+        "workflowName": "Critic Workflow",
+        "runtime": "gui",
+        "status": "success",
+        "startTs": 200,
+        "children": [],
+    }
+    pyfetch = AsyncMock(
+        side_effect=[
+            _FakeResponse(json_data={"requestId": "main-request-123"}),
+            _FakeResponse(
+                json_data={
+                    "status": "success",
+                    "response": {
+                        "text": "done",
+                        "output": {"type": "text", "content": "done"},
+                        "workflowTrace": workflow_trace,
+                    },
+                    "usage": {"actions": 0, "credits": 0},
+                }
+            ),
+            _FakeResponse(json_data={"requestId": "critic-request-123"}),
+            _FakeResponse(
+                json_data={
+                    "status": "success",
+                    "response": {
+                        "text": '{"narada_validation_passed":true}',
+                        "output": {
+                            "type": "structured",
+                            "content": {"narada_validation_passed": True},
+                        },
+                        "workflowTrace": critic_workflow_trace,
+                    },
+                    "usage": {"actions": 0, "credits": 0},
+                }
+            ),
+        ]
+    )
+    _, _, window_module = _import_pyodide_narada(monkeypatch, pyfetch=pyfetch)
+    emitted_events: list[str] = []
+    monkeypatch.setattr(
+        sys.modules["narada._trace"],
+        "_narada_emit_trace_event",
+        emitted_events.append,
+        raising=False,
+    )
+
+    window = window_module.CloudBrowserWindow(
+        browser_window_id="browser-window-123",
+        session_id="session-123",
+        api_key="test-api-key",
+    )
+    response = await window.agent(prompt="return a trace", critic={})
+
+    combined_workflow_trace = {
+        **workflow_trace,
+        "children": [{"kind": "sub_workflow", "trace": critic_workflow_trace}],
+    }
+    assert response.critic_result is not None
+    assert response.critic_result.workflow_trace == critic_workflow_trace
+    assert response.workflow_trace == combined_workflow_trace
+    sub_workflow_events = [
+        json.loads(event)
+        for event in emitted_events
+        if json.loads(event)["kind"] == "subWorkflow"
+    ]
+    assert sub_workflow_events == [
+        {"kind": "subWorkflow", "workflowTrace": combined_workflow_trace}
+    ]
+
+
 @pytest.mark.asyncio
 async def test_cloud_browser_window_dispatch_request_retries_poll_fetch_failures(
     monkeypatch: pytest.MonkeyPatch,

packages/narada/src/narada/window.py

@@ -21,7 +21,7 @@
 )
 
 import aiohttp
-from narada_core.actions.critic import run_critic
+from narada_core.actions.critic import merge_critic_workflow_trace, run_critic
 from narada_core.actions.models import (
     DEFAULT_HITL_TIMEOUT_SECONDS,
     AgenticMatchingSelectorsFinderRequest,
@@ -656,6 +656,10 @@ async def agent(
                 time_zone=time_zone,
                 timeout=timeout,
             )
+            workflow_trace = merge_critic_workflow_trace(
+                workflow_trace=workflow_trace,
+                critic_result=critic_result,
+            )
 
         return AgentResponse(
             request_id=remote_dispatch_response["requestId"],

packages/narada/tests/test_cloud_browser.py

@@ -205,3 +205,73 @@ async def test_window_agent_exposes_workflow_trace_alias(
 
     assert response.workflow_trace == workflow_trace
     assert response.model_dump(by_alias=True)["workflowTrace"] == workflow_trace
+
+
+@pytest.mark.asyncio
+async def test_window_agent_appends_critic_workflow_trace(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    workflow_trace = {
+        "workflowId": "main-workflow",
+        "workflowName": "Main Workflow",
+        "runtime": "gui",
+        "status": "success",
+        "startTs": 100,
+        "children": [],
+    }
+    critic_workflow_trace = {
+        "workflowId": "critic-workflow",
+        "workflowName": "Critic Workflow",
+        "runtime": "gui",
+        "status": "success",
+        "startTs": 200,
+        "children": [],
+    }
+    window = CloudBrowserWindow(
+        browser_window_id="browser-window-123",
+        session_id="session-123",
+        auth_headers={"x-api-key": "test-key"},
+    )
+    monkeypatch.setattr(
+        window,
+        "dispatch_request",
+        AsyncMock(
+            side_effect=[
+                {
+                    "requestId": "request-123",
+                    "status": "success",
+                    "response": {
+                        "text": "done",
+                        "output": {"type": "text", "content": "done"},
+                        "workflowTrace": workflow_trace,
+                    },
+                    "usage": {"actions": 0, "credits": 0},
+                },
+                {
+                    "requestId": "critic-request-123",
+                    "status": "success",
+                    "response": {
+                        "text": '{"narada_validation_passed":true}',
+                        "output": {
+                            "type": "structured",
+                            "content": {"narada_validation_passed": True},
+                        },
+                        "structuredOutput": SimpleNamespace(
+                            narada_validation_passed=True
+                        ),
+                        "workflowTrace": critic_workflow_trace,
+                    },
+                    "usage": {"actions": 0, "credits": 0},
+                },
+            ]
+        ),
+    )
+
+    response = await window.agent(prompt="return a trace", critic={})
+
+    assert response.critic_result is not None
+    assert response.critic_result.workflow_trace == critic_workflow_trace
+    assert response.workflow_trace == {
+        **workflow_trace,
+        "children": [{"kind": "sub_workflow", "trace": critic_workflow_trace}],
+    }