fix(tracing): attach Langfuse trace metadata to the goal evaluator (#4202)

The goal evaluator (runtime/goal.py) runs from runtime/runs/worker.py after
the main graph run has already completed, so there is no graph root for it
to inherit tracing from. create_goal_evaluator_model was built with
attach_tracing=False, and evaluate_goal_completion invoked the model with a
bare config={"run_name": "goal_evaluator"} — no tracing callbacks, no
Langfuse session/user attribution. Every goal-evaluator LLM call went
untraced.

Same class of gap fixed by #2944 for the main agent graph and by #3902 for
memory_agent/suggest_agent: a standalone call site that invokes a model
directly instead of through a traced graph root must attach its own tracing
callbacks and inject Langfuse trace-attribute metadata itself.

- create_goal_evaluator_model: attach_tracing=False -> True, matching the
  other standalone non-graph callers (oneshot_llm.run_oneshot_llm,
  MemoryUpdater).
- evaluate_goal_completion: accept optional thread_id/user_id/
  deerflow_trace_id and inject Langfuse trace metadata onto the ainvoke
  config via the shared inject_langfuse_metadata() helper, mirroring
  oneshot_llm.py's pattern.
- worker.py: thread user_id (resolve_runtime_user_id(runtime)) and
  deerflow_trace_id through _prepare_goal_continuation_input into
  evaluate_goal_completion so the evaluator's trace groups under the
  triggering run's thread/session.

Updates the existing test that pinned attach_tracing=False as expected
behavior, and adds a regression test asserting the ainvoke config carries
Langfuse trace metadata when enabled.

backend/AGENTS.md

@@ -347,7 +347,7 @@ metadata only.
 - `POST /wait` (both thread-scoped and `/api/runs/wait`) drains the stream bridge via `wait_for_run_completion()` instead of bare `await record.task`, so it honours the run's `on_disconnect` setting and cancels the background run on real client disconnect rather than returning a stale checkpoint (issue #3265).
 - Redis `StreamBridge` keys use a rolling retained-buffer TTL (`stream_bridge.stream_ttl_seconds`, refreshed on `publish()` / `publish_end()`) as a leak safety net, not as a run timeout. Startup orphan recovery publishes `END_SENTINEL` and schedules stream cleanup for recovered runs; malformed `Last-Event-ID` reconnect values live-tail new Redis events rather than replaying the retained buffer. Do not broaden this into a shared-database multi-pod reaper without adding worker ownership/liveness first.
 - Thread-scoped run creation accepts `checkpoint` / `checkpoint_id`; Gateway validates the checkpoint belongs to the request thread before writing `checkpoint_id` / `checkpoint_ns` into `config.configurable` for LangGraph branching.
-- Thread-scoped Gateway runs evaluate an active `ThreadState.goal` after the visible turn completes. `runtime/goal.py` asks a non-thinking evaluator model to judge only visible conversation evidence and return a typed blocker; the evaluator model is created once per run and reused across hidden continuation checks. Satisfied goals are cleared; every non-satisfied evaluation — continuable or stand-down — is persisted with `last_evaluation` (the blocker, reason, and evidence summary; outcomes that stop the loop additionally record a `stand_down_reason` for observability), but only `goal_not_met_yet` evaluations are streamed as hidden `HumanMessage` continuations, and only when a durable assistant end-of-turn checkpoint exists, the run has not been aborted, the thread did not change during evaluation, and the no-progress breaker has not fired. The continuation cap is 8 — a hard maximum in the `0`–`8` range; callers requesting more are clamped (`set_goal`/TUI) or rejected with 422 (`PUT /goal`). The no-progress breaker keys on the latest visible assistant evidence (not the evaluator's free-text reason, which an LLM rewords every turn), so two consecutive continuations that add no new visible assistant output stop the loop after 2 attempts. Model-response cleanup helpers such as think-block stripping and code-fence stripping live in `deerflow.utils.llm_text` so `runtime/goal.py` and Gateway suggestion parsing share the same JSON-prep behavior.
+- Thread-scoped Gateway runs evaluate an active `ThreadState.goal` after the visible turn completes. `runtime/goal.py` asks a non-thinking evaluator model to judge only visible conversation evidence and return a typed blocker; the evaluator model is created once per run and reused across hidden continuation checks. The evaluator runs after the graph root's tracing scope has already closed, so `create_goal_evaluator_model`/`evaluate_goal_completion` attach their own model-level tracing callbacks (`attach_tracing=True`) and inject Langfuse trace metadata (`thread_id`/`user_id`/`deerflow_trace_id`) directly onto the `ainvoke` call — the same standalone-caller pattern as `oneshot_llm.run_oneshot_llm` and `MemoryUpdater` (see Tracing System below). Satisfied goals are cleared; every non-satisfied evaluation — continuable or stand-down — is persisted with `last_evaluation` (the blocker, reason, and evidence summary; outcomes that stop the loop additionally record a `stand_down_reason` for observability), but only `goal_not_met_yet` evaluations are streamed as hidden `HumanMessage` continuations, and only when a durable assistant end-of-turn checkpoint exists, the run has not been aborted, the thread did not change during evaluation, and the no-progress breaker has not fired. The continuation cap is 8 — a hard maximum in the `0`–`8` range; callers requesting more are clamped (`set_goal`/TUI) or rejected with 422 (`PUT /goal`). The no-progress breaker keys on the latest visible assistant evidence (not the evaluator's free-text reason, which an LLM rewords every turn), so two consecutive continuations that add no new visible assistant output stop the loop after 2 attempts. Model-response cleanup helpers such as think-block stripping and code-fence stripping live in `deerflow.utils.llm_text` so `runtime/goal.py` and Gateway suggestion parsing share the same JSON-prep behavior.
 
 Proxied through nginx: `/api/langgraph/*` → Gateway LangGraph-compatible runtime, all other `/api/*` → Gateway REST APIs.
 

backend/packages/harness/deerflow/runtime/goal.py

@@ -13,6 +13,7 @@ import hashlib
 import inspect
 import json
 import logging
+import os
 import threading
 import weakref
 from collections.abc import AsyncIterator
@@ -25,6 +26,7 @@ from langgraph.checkpoint.base import empty_checkpoint, uuid6
 import deerflow.utils.llm_text as llm_text
 from deerflow.agents.goal_state import GoalBlocker, GoalEvaluation, GoalState
 from deerflow.models import create_chat_model
+from deerflow.tracing import inject_langfuse_metadata
 from deerflow.utils.messages import message_to_text
 from deerflow.utils.time import now_iso
 
@@ -242,15 +244,29 @@ def create_goal_evaluator_model(
     model_name: str | None = None,
     app_config: Any | None = None,
 ) -> Any:
-    """Create the non-thinking chat model used by the goal evaluator."""
+    """Create the non-thinking chat model used by the goal evaluator.
+
+    The evaluator runs from ``runtime/runs/worker.py`` after the main graph
+    run has already completed, so — unlike ``make_lead_agent``/
+    ``DeerFlowClient.stream``, which attach ``build_tracing_callbacks()`` at
+    the graph root and correctly pass ``attach_tracing=False`` to avoid
+    double-attaching — there is no graph root here for the evaluator's model
+    call to inherit tracing from. It must attach its own model-level tracing
+    callbacks, same as the other standalone, non-graph callers
+    (``oneshot_llm.run_oneshot_llm``, ``MemoryUpdater``).
+    """
     return create_chat_model(
         name=model_name,
         thinking_enabled=False,
         app_config=app_config,
-        attach_tracing=False,
+        attach_tracing=True,
     )
 
 
+def _resolve_environment() -> str | None:
+    return os.environ.get("DEER_FLOW_ENV") or os.environ.get("ENVIRONMENT")
+
+
 async def evaluate_goal_completion(
     goal: GoalState,
     messages: list[Any],
@@ -258,8 +274,19 @@ async def evaluate_goal_completion(
     model: Any | None = None,
     model_name: str | None = None,
     app_config: Any | None = None,
+    thread_id: str | None = None,
+    user_id: str | None = None,
+    deerflow_trace_id: str | None = None,
 ) -> GoalEvaluation:
-    """Ask a small non-thinking model whether the active goal is satisfied."""
+    """Ask a small non-thinking model whether the active goal is satisfied.
+
+    ``thread_id``/``user_id``/``deerflow_trace_id`` are forwarded to Langfuse
+    trace metadata only (mirrors ``oneshot_llm.run_oneshot_llm``): this is a
+    standalone model call outside the main graph, so it must inject its own
+    Langfuse session/user attribution instead of relying on graph-root
+    callbacks to lift it — same fix as PR #2944 (main graph) and PR #3902
+    (memory_agent/suggest_agent).
+    """
     conversation = format_visible_conversation(messages)
     if not conversation or not has_visible_assistant_evidence(messages):
         return GoalEvaluation(
@@ -283,9 +310,19 @@ async def evaluate_goal_completion(
 
     if model is None:
         model = create_goal_evaluator_model(model_name=model_name, app_config=app_config)
+    invoke_config: dict[str, Any] = {"run_name": "goal_evaluator"}
+    inject_langfuse_metadata(
+        invoke_config,
+        thread_id=thread_id,
+        user_id=user_id,
+        assistant_id="goal_evaluator",
+        model_name=model_name,
+        environment=_resolve_environment(),
+        deerflow_trace_id=deerflow_trace_id,
+    )
     response = await model.ainvoke(
         [SystemMessage(content=system_instruction), HumanMessage(content=user_content)],
-        config={"run_name": "goal_evaluator"},
+        config=invoke_config,
     )
     return parse_goal_evaluation_response(_extract_response_text(response.content))
 

backend/packages/harness/deerflow/runtime/runs/worker.py

@@ -528,6 +528,8 @@ async def run_agent(
                 app_config=ctx.app_config,
                 evaluator_model_factory=_get_goal_evaluator_model,
                 abort_event=record.abort_event,
+                user_id=resolve_runtime_user_id(runtime),
+                deerflow_trace_id=deerflow_trace_id,
             )
             if continuation_input is None or record.abort_event.is_set():
                 break
@@ -842,6 +844,8 @@ async def _prepare_goal_continuation_input(
     app_config: AppConfig | None,
     evaluator_model_factory: Any | None = None,
     abort_event: asyncio.Event | None = None,
+    user_id: str | None = None,
+    deerflow_trace_id: str | None = None,
 ) -> dict[str, Any] | None:
     """Evaluate the active goal and return a hidden continuation input if needed.
 
@@ -919,6 +923,9 @@ async def _prepare_goal_continuation_input(
             model=evaluator_model,
             model_name=model_name,
             app_config=app_config,
+            thread_id=thread_id,
+            user_id=user_id,
+            deerflow_trace_id=deerflow_trace_id,
         )
         if abort_event is not None and abort_event.is_set():
             return None

backend/tests/test_goal_runtime.py

@@ -116,11 +116,75 @@ def test_evaluate_goal_completion_uses_non_thinking_model(monkeypatch):
     assert result["satisfied"] is True
     assert result["blocker"] == "none"
     assert captured["thinking_enabled"] is False
-    assert captured["attach_tracing"] is False
+    # The goal evaluator runs from runtime/runs/worker.py after the main graph
+    # run has already finished, so there is no graph root for it to inherit
+    # tracing callbacks from (unlike make_lead_agent/DeerFlowClient.stream,
+    # which attach build_tracing_callbacks() at the graph root and correctly
+    # pass attach_tracing=False to avoid double-attaching). It must attach its
+    # own model-level tracing callbacks, same as the other standalone,
+    # non-graph callers (oneshot_llm.run_oneshot_llm, MemoryUpdater).
+    assert captured["attach_tracing"] is True
     fake_model.ainvoke.assert_awaited_once()
+    # No thread_id/user_id supplied here, and Langfuse is not enabled in the
+    # ambient test env, so inject_langfuse_metadata() is a no-op and the
+    # config is unchanged from the plain run_name — see
+    # test_evaluate_goal_completion_injects_langfuse_metadata below for the
+    # Langfuse-enabled case.
     assert fake_model.ainvoke.await_args.kwargs["config"] == {"run_name": "goal_evaluator"}
 
 
+def test_evaluate_goal_completion_injects_langfuse_metadata(monkeypatch):
+    """Regression test for the goal evaluator's Langfuse tracing gap.
+
+    Mirrors PR #2944 (main graph) and PR #3902 (memory_agent/suggest_agent):
+    a standalone, non-graph model call must inject Langfuse trace-attribute
+    metadata itself since there is no graph root to lift it from.
+    """
+    from deerflow.config.tracing_config import reset_tracing_config
+
+    for name in ("LANGFUSE_TRACING", "LANGFUSE_PUBLIC_KEY", "LANGFUSE_SECRET_KEY", "LANGFUSE_BASE_URL"):
+        monkeypatch.delenv(name, raising=False)
+    monkeypatch.setenv("LANGFUSE_TRACING", "true")
+    monkeypatch.setenv("LANGFUSE_PUBLIC_KEY", "pk-lf-test")
+    monkeypatch.setenv("LANGFUSE_SECRET_KEY", "sk-lf-test")
+    reset_tracing_config()
+
+    fake_model = MagicMock()
+    fake_model.ainvoke = AsyncMock(return_value=SimpleNamespace(content='{"satisfied": true, "reason": "Done", "evidence_summary": "Done"}'))
+    state = goal.build_goal_state("Finish")
+
+    try:
+        result = asyncio.run(
+            goal.evaluate_goal_completion(
+                state,
+                [
+                    HumanMessage(content="Please finish this."),
+                    AIMessage(content="Done."),
+                ],
+                model=fake_model,
+                model_name="gpt-4o",
+                app_config=object(),
+                thread_id="thread-xyz",
+                user_id="alice",
+                deerflow_trace_id="gateway-trace-1",
+            )
+        )
+    finally:
+        reset_tracing_config()
+
+    assert result["satisfied"] is True
+    fake_model.ainvoke.assert_awaited_once()
+    config = fake_model.ainvoke.await_args.kwargs["config"]
+    assert config["run_name"] == "goal_evaluator"
+    metadata = config.get("metadata") or {}
+    assert metadata.get("langfuse_session_id") == "thread-xyz", "goal evaluator trace must group under the thread's session"
+    assert metadata.get("langfuse_user_id") == "alice"
+    assert metadata.get("langfuse_trace_name") == "goal_evaluator"
+    assert metadata.get("deerflow_trace_id") == "gateway-trace-1"
+    tags = metadata.get("langfuse_tags") or []
+    assert "model:gpt-4o" in tags
+
+
 def test_evaluate_goal_completion_uses_injected_model(monkeypatch):
     fake_model = MagicMock()
     fake_model.ainvoke = AsyncMock(return_value=SimpleNamespace(content='{"satisfied": true, "reason": "Done", "evidence_summary": "Done"}'))