diff --git a/packages/api/alembic/versions/092e2aa153ce_add_chat_messages_langfuse_trace_id.py b/packages/api/alembic/versions/092e2aa153ce_add_chat_messages_langfuse_trace_id.py
new file mode 100644
index 0000000..d14b8d0
--- /dev/null
+++ b/packages/api/alembic/versions/092e2aa153ce_add_chat_messages_langfuse_trace_id.py
@@ -0,0 +1,36 @@
+"""add chat_messages.langfuse_trace_id
+
+Revision ID: 092e2aa153ce
+Revises: a15a753f44c8
+Create Date: 2026-05-21 09:00:00.000000
+
+"""
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+
+
+revision: str = '092e2aa153ce'
+down_revision: Union[str, Sequence[str], None] = 'a15a753f44c8'
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    """Upgrade schema."""
+    with op.batch_alter_table('chat_messages') as batch_op:
+        batch_op.add_column(
+            sa.Column('langfuse_trace_id', sa.String(), nullable=True)
+        )
+        batch_op.create_index(
+            'ix_chat_messages_langfuse_trace_id',
+            ['langfuse_trace_id'],
+        )
+
+
+def downgrade() -> None:
+    """Downgrade schema."""
+    with op.batch_alter_table('chat_messages') as batch_op:
+        batch_op.drop_index('ix_chat_messages_langfuse_trace_id')
+        batch_op.drop_column('langfuse_trace_id')
diff --git a/packages/api/src/cell_explorer_api/db/models.py b/packages/api/src/cell_explorer_api/db/models.py
index f3c09b2..2e31a51 100644
--- a/packages/api/src/cell_explorer_api/db/models.py
+++ b/packages/api/src/cell_explorer_api/db/models.py
@@ -101,6 +101,9 @@ class ChatMessageRow(SQLModel, table=True):
     role: str  # "user" | "assistant"
     content: str
     created_at: datetime = Field(default_factory=_utcnow)
+    # Set on assistant rows when Langfuse tracing is enabled. Used by the
+    # feedback PUT route to forward thumbs ratings to Langfuse Scores.
+    langfuse_trace_id: str | None = Field(default=None, index=True)
 
 
 class ChatFeedback(SQLModel, table=True):
diff --git a/packages/api/src/cell_explorer_api/routes/chat.py b/packages/api/src/cell_explorer_api/routes/chat.py
index e5c3c4c..28452a4 100644
--- a/packages/api/src/cell_explorer_api/routes/chat.py
+++ b/packages/api/src/cell_explorer_api/routes/chat.py
@@ -1,6 +1,7 @@
 """HTTP routes for the chat agent (Plan 2a)."""
 
 import asyncio
+import logging
 import uuid as _uuid
 from collections.abc import AsyncIterator
 from datetime import datetime
@@ -13,8 +14,10 @@
 from sqlmodel import select
 from sqlmodel.ext.asyncio.session import AsyncSession
 
+from cell_explorer_agent.config import AgentConfig
 from cell_explorer_agent.events import Error, ThreadOpen
 from cell_explorer_agent.messages import AssistantMessage, UserMessage
+from cell_explorer_agent.telemetry import langfuse_client
 from cell_explorer_api.auth.dependencies import require_auth
 from cell_explorer_api.auth.models import User
 from cell_explorer_api.auth.optional import optional_auth
@@ -48,6 +51,8 @@
     load_thread,
 )
 
+logger = logging.getLogger(__name__)
+
 router = APIRouter(tags=["chat"])
 
 # Note: this router is included by routes/__init__.py whose own router has
@@ -55,6 +60,39 @@
 # matching the convention in datasets.py.
 
 
+def _forward_feedback_to_langfuse(
+    *,
+    trace_id: str,
+    user_sub: str,
+    message_id: _uuid.UUID,
+    rating: str,
+    comment: str | None,
+) -> None:
+    """Best-effort forward of a thumbs rating to Langfuse Scores.
+
+    Idempotent on (user_sub, message_id) via deterministic score_id so
+    re-clicking updates the score in place (assuming Langfuse upserts;
+    duplicate rows are at worst dedupable later via score_id prefix).
+
+    Any failure is logged and swallowed — feedback writes to our DB are
+    the source of truth; Langfuse is an observability sink.
+    """
+    try:
+        client = langfuse_client.get(AgentConfig())
+        if client is None:
+            return
+        client.create_score(
+            name="user_feedback",
+            value=1.0 if rating == "up" else 0.0,
+            data_type="NUMERIC",
+            trace_id=trace_id,
+            score_id=f"feedback-{user_sub}-{message_id}",
+            comment=comment,
+        )
+    except Exception:
+        logger.exception("Failed to forward feedback to Langfuse")
+
+
 class ChatMessage(BaseModel):
     role: Literal["user", "assistant"]
     content: str = Field(min_length=1)
@@ -183,6 +221,7 @@ async def _ndjson_event_stream(
                         msg = await append_message(
                             stream_db, thread, role="assistant",
                             content="".join(assistant_buffer),
+                            langfuse_trace_id=event.trace_id,
                         )
                         # Capture id before commit() expires ORM attributes.
                         message_id = str(msg.id)
@@ -534,6 +573,8 @@ async def put_message_feedback(
             status_code=422, detail="Only assistant messages can be rated"
         )
 
+    # Capture trace_id before commit() expires ORM attributes.
+    langfuse_trace_id = msg.langfuse_trace_id
     # Upsert by (message_id, user_sub).
     existing = (
         await db.exec(
@@ -559,6 +600,15 @@ async def put_message_feedback(
         fb = existing
     await db.commit()
     await db.refresh(fb)
+    # Forward to Langfuse Scores when we know the trace id. Best-effort.
+    if langfuse_trace_id:
+        _forward_feedback_to_langfuse(
+            trace_id=langfuse_trace_id,
+            user_sub=user.sub,
+            message_id=message_id,
+            rating=fb.rating,
+            comment=fb.comment,
+        )
     return FeedbackResponse(
         rating=fb.rating,  # type: ignore[arg-type]
         comment=fb.comment,
diff --git a/packages/api/src/cell_explorer_api/services/threads.py b/packages/api/src/cell_explorer_api/services/threads.py
index 57890fc..8a51997 100644
--- a/packages/api/src/cell_explorer_api/services/threads.py
+++ b/packages/api/src/cell_explorer_api/services/threads.py
@@ -144,9 +144,15 @@ async def append_message(
     *,
     role: str,
     content: str,
+    langfuse_trace_id: str | None = None,
 ) -> ChatMessageRow:
     """Append a message to the thread and bump its updated_at."""
-    msg = ChatMessageRow(thread_id=thread.id, role=role, content=content)
+    msg = ChatMessageRow(
+        thread_id=thread.id,
+        role=role,
+        content=content,
+        langfuse_trace_id=langfuse_trace_id,
+    )
     session.add(msg)
     thread.updated_at = datetime.now(timezone.utc).replace(tzinfo=None)
     session.add(thread)
diff --git a/packages/api/tests/routes/test_chat_feedback.py b/packages/api/tests/routes/test_chat_feedback.py
index d3cb9f2..aead3e5 100644
--- a/packages/api/tests/routes/test_chat_feedback.py
+++ b/packages/api/tests/routes/test_chat_feedback.py
@@ -284,3 +284,152 @@ def test_thread_detail_includes_message_id(seeded_app):
     res = client.get(f"/api/chat/public-atlas/threads/{thread_id}")
     assert res.status_code == 200
     assert res.json()["messages"][0]["id"] == msg_id
+
+
+def _create_assistant_msg_with_trace_id(seeded_app, *, user_sub: str, trace_id: str) -> str:
+    """Seed one assistant ChatMessageRow with langfuse_trace_id set."""
+    from cell_explorer_api.db.models import ChatMessageRow, ChatThread, Dataset
+    from sqlmodel import select
+
+    async def _seed() -> str:
+        engine = seeded_app.state.db_engine
+        async with AsyncSession(engine) as session:
+            dataset = (await session.exec(select(Dataset))).first()
+            t = ChatThread(user_sub=user_sub, dataset_id=dataset.id, title="t")
+            session.add(t)
+            await session.flush()
+            m = ChatMessageRow(
+                thread_id=t.id,
+                role="assistant",
+                content="hello",
+                langfuse_trace_id=trace_id,
+            )
+            session.add(m)
+            await session.commit()
+            await session.refresh(m)
+            return str(m.id)
+
+    return asyncio.run(_seed())
+
+
+def test_put_feedback_forwards_to_langfuse_when_trace_id_present(
+    seeded_app, monkeypatch
+):
+    """PUT 👍 on a message with langfuse_trace_id calls Langfuse create_score."""
+    from cell_explorer_agent.telemetry import langfuse_client
+    from cell_explorer_agent.telemetry.fake import FakeLangfuseClient
+
+    fake = FakeLangfuseClient()
+    monkeypatch.setattr(langfuse_client, "get", lambda _config: fake)
+
+    client = TestClient(seeded_app)
+    _set_auth_cookie(client, seeded_app, sub="user-1")
+    msg_id = _create_assistant_msg_with_trace_id(
+        seeded_app, user_sub="user-1", trace_id="trace-xyz",
+    )
+
+    res = client.put(
+        f"/api/chat/public-atlas/messages/{msg_id}/feedback",
+        json={"rating": "up", "comment": "great answer"},
+    )
+    assert res.status_code == 200, res.text
+
+    assert len(fake.scores) == 1
+    s = fake.scores[0]
+    assert s["name"] == "user_feedback"
+    assert s["value"] == 1.0
+    assert s["data_type"] == "NUMERIC"
+    assert s["trace_id"] == "trace-xyz"
+    assert s["score_id"] == f"feedback-user-1-{msg_id}"
+    assert s["comment"] == "great answer"
+
+
+def test_put_feedback_forwards_down_as_zero(seeded_app, monkeypatch):
+    """👎 maps to value=0.0."""
+    from cell_explorer_agent.telemetry import langfuse_client
+    from cell_explorer_agent.telemetry.fake import FakeLangfuseClient
+
+    fake = FakeLangfuseClient()
+    monkeypatch.setattr(langfuse_client, "get", lambda _config: fake)
+
+    client = TestClient(seeded_app)
+    _set_auth_cookie(client, seeded_app, sub="user-1")
+    msg_id = _create_assistant_msg_with_trace_id(
+        seeded_app, user_sub="user-1", trace_id="trace-xyz",
+    )
+
+    res = client.put(
+        f"/api/chat/public-atlas/messages/{msg_id}/feedback",
+        json={"rating": "down"},
+    )
+    assert res.status_code == 200
+    assert fake.scores[0]["value"] == 0.0
+
+
+def test_put_feedback_skips_langfuse_when_no_trace_id(seeded_app, monkeypatch):
+    """When the message has no langfuse_trace_id, no Langfuse call is made."""
+    from cell_explorer_agent.telemetry import langfuse_client
+    from cell_explorer_agent.telemetry.fake import FakeLangfuseClient
+
+    fake = FakeLangfuseClient()
+    monkeypatch.setattr(langfuse_client, "get", lambda _config: fake)
+
+    client = TestClient(seeded_app)
+    _set_auth_cookie(client, seeded_app, sub="user-1")
+    # Message created via _create_thread_with_assistant_msg has no trace_id.
+    _thread_id, msg_id = _create_thread_with_assistant_msg(seeded_app, user_sub="user-1")
+
+    res = client.put(
+        f"/api/chat/public-atlas/messages/{msg_id}/feedback",
+        json={"rating": "up"},
+    )
+    assert res.status_code == 200
+    assert fake.scores == []
+
+
+def test_put_feedback_swallows_langfuse_errors(seeded_app, monkeypatch):
+    """A Langfuse failure must not break the user-facing feedback PUT."""
+    from cell_explorer_agent.telemetry import langfuse_client
+
+    class _BrokenClient:
+        def create_score(self, **_kwargs):
+            raise RuntimeError("langfuse down")
+
+    monkeypatch.setattr(langfuse_client, "get", lambda _config: _BrokenClient())
+
+    client = TestClient(seeded_app)
+    _set_auth_cookie(client, seeded_app, sub="user-1")
+    msg_id = _create_assistant_msg_with_trace_id(
+        seeded_app, user_sub="user-1", trace_id="trace-xyz",
+    )
+
+    res = client.put(
+        f"/api/chat/public-atlas/messages/{msg_id}/feedback",
+        json={"rating": "up"},
+    )
+    assert res.status_code == 200, res.text
+
+
+def test_delete_feedback_does_not_call_langfuse(seeded_app, monkeypatch):
+    """DELETE clears local feedback only; Langfuse history is preserved."""
+    from cell_explorer_agent.telemetry import langfuse_client
+    from cell_explorer_agent.telemetry.fake import FakeLangfuseClient
+
+    fake = FakeLangfuseClient()
+    monkeypatch.setattr(langfuse_client, "get", lambda _config: fake)
+
+    client = TestClient(seeded_app)
+    _set_auth_cookie(client, seeded_app, sub="user-1")
+    msg_id = _create_assistant_msg_with_trace_id(
+        seeded_app, user_sub="user-1", trace_id="trace-xyz",
+    )
+    # First PUT creates feedback + one score; then DELETE clears the row.
+    client.put(
+        f"/api/chat/public-atlas/messages/{msg_id}/feedback",
+        json={"rating": "up"},
+    )
+    pre_delete_scores = list(fake.scores)
+    res = client.delete(f"/api/chat/public-atlas/messages/{msg_id}/feedback")
+    assert res.status_code == 204
+    # DELETE did not produce any additional Langfuse calls.
+    assert fake.scores == pre_delete_scores
diff --git a/packages/cell-explorer-agent/src/cell_explorer_agent/agent.py b/packages/cell-explorer-agent/src/cell_explorer_agent/agent.py
index e9f34d9..9b9446b 100644
--- a/packages/cell-explorer-agent/src/cell_explorer_agent/agent.py
+++ b/packages/cell-explorer-agent/src/cell_explorer_agent/agent.py
@@ -176,7 +176,7 @@ async def run(
 
                 if stop_reason == "end_turn" or not pending_calls:
                     trace.set_output("".join(final_assistant_text_buffer))
-                    yield Done(usage=total_usage)
+                    yield Done(usage=total_usage, trace_id=trace.trace_id)
                     return
 
                 if tool_calls_this_turn + len(pending_calls) > max_calls:
@@ -187,7 +187,7 @@ async def run(
                         retryable=False,
                     )
                     trace.set_output("".join(final_assistant_text_buffer))
-                    yield Done(usage=total_usage)
+                    yield Done(usage=total_usage, trace_id=trace.trace_id)
                     return
 
                 results: list[ToolResult] = []
diff --git a/packages/cell-explorer-agent/src/cell_explorer_agent/events.py b/packages/cell-explorer-agent/src/cell_explorer_agent/events.py
index 907dd1a..3f0a119 100644
--- a/packages/cell-explorer-agent/src/cell_explorer_agent/events.py
+++ b/packages/cell-explorer-agent/src/cell_explorer_agent/events.py
@@ -58,6 +58,11 @@ class Done(BaseModel):
     # does not persist, so this is None at agent emit time; the API layer fills
     # it in just before yielding the event to the client.
     message_id: str | None = None
+    # Langfuse trace id for this turn. Set by the agent when telemetry is
+    # enabled; the API layer persists it on the assistant row so later
+    # feedback (PUT /feedback) can forward to Langfuse Scores. Wire-level
+    # field — the frontend ignores it.
+    trace_id: str | None = None
 
 
 class ThreadOpen(BaseModel):
diff --git a/packages/cell-explorer-agent/src/cell_explorer_agent/telemetry/fake.py b/packages/cell-explorer-agent/src/cell_explorer_agent/telemetry/fake.py
index 9905861..9e5efb1 100644
--- a/packages/cell-explorer-agent/src/cell_explorer_agent/telemetry/fake.py
+++ b/packages/cell-explorer-agent/src/cell_explorer_agent/telemetry/fake.py
@@ -93,6 +93,11 @@ def __init__(self) -> None:
         self.trace_updates: list[dict[str, Any]] = []
         self.flushed_count: int = 0
         self._current_stack: list[FakeObservation] = []
+        # Tests can override; defaults to a synthetic value returned by
+        # get_current_trace_id() while a span context is active.
+        self.fake_trace_id: str = "fake-trace-id"
+        # Records every create_score(**kwargs) call for assertion in tests.
+        self.scores: list[dict[str, Any]] = []
 
     # ---- observation creation ----
 
@@ -135,6 +140,19 @@ def update_current_span(self, **kwargs: Any) -> None:
                 obs.update(**kwargs)
                 return
 
+    # ---- trace id resolution ----
+
+    def get_current_trace_id(self) -> str | None:
+        """Mirrors v3 SDK: returns the active trace id while a span context
+        is open; None otherwise."""
+        return self.fake_trace_id if self._current_stack else None
+
+    # ---- scores ----
+
+    def create_score(self, **kwargs: Any) -> None:
+        """Record a create_score call. Tests inspect self.scores."""
+        self.scores.append(dict(kwargs))
+
     # ---- lifecycle ----
 
     def flush(self) -> None:
diff --git a/packages/cell-explorer-agent/src/cell_explorer_agent/telemetry/trace_context.py b/packages/cell-explorer-agent/src/cell_explorer_agent/telemetry/trace_context.py
index c687e79..d3b7346 100644
--- a/packages/cell-explorer-agent/src/cell_explorer_agent/telemetry/trace_context.py
+++ b/packages/cell-explorer-agent/src/cell_explorer_agent/telemetry/trace_context.py
@@ -63,6 +63,10 @@ def __init__(
         # v3 internals
         self._root_cm: Any | None = None
         self._root_obs: Any | None = None
+        # Captured inside __aenter__ once the root span is active. Used by
+        # callers (route layer) to persist a message → trace association so
+        # later feedback can be forwarded to Langfuse Scores.
+        self.trace_id: str | None = None
         # Per-turn counter so a turn with multiple LLM rounds shows
         # llm-call-1, llm-call-2, ... in the Langfuse trace tree
         # instead of three identical 'llm-call' rows.
@@ -94,6 +98,12 @@ async def __aenter__(self) -> "TurnTrace":
                 },
             )
             self._root_obs = self._root_cm.__enter__()
+            # Capture trace id from the now-active span context. None if the
+            # SDK can't resolve one (e.g. fake clients without OTel context).
+            try:
+                self.trace_id = self._client.get_current_trace_id()
+            except Exception:
+                self.trace_id = None
             # Set trace-level fields (user/session/tags) once the root span
             # is active — these belong on the implicit trace, not the span.
             self._client.update_current_trace(
diff --git a/packages/cell-explorer-agent/tests/test_telemetry_trace_context.py b/packages/cell-explorer-agent/tests/test_telemetry_trace_context.py
index 4fa880d..e70c301 100644
--- a/packages/cell-explorer-agent/tests/test_telemetry_trace_context.py
+++ b/packages/cell-explorer-agent/tests/test_telemetry_trace_context.py
@@ -110,6 +110,40 @@ async def test_private_dataset_trace_redacts_content(fake):
     assert fake.tool_spans[0].input == {"_redacted": "tool_args", "tool": "filter_by_ids"}
 
 
+async def test_turn_trace_captures_trace_id_from_active_span(fake):
+    """Once the root span is open, get_current_trace_id() returns a value;
+    TurnTrace caches it on self.trace_id so the route layer can persist it."""
+    fake.fake_trace_id = "abc-trace-123"
+    async with TurnTrace(
+        client=fake,
+        user_id="u",
+        thread_id="t",
+        dataset_slug="ds",
+        is_public=True,
+        model="m",
+        environment="test",
+        user_input="hi",
+        view_state=None,
+    ) as trace:
+        assert trace.trace_id == "abc-trace-123"
+
+
+async def test_turn_trace_trace_id_none_when_no_client():
+    """No client → no span context → no trace_id."""
+    async with TurnTrace(
+        client=None,
+        user_id="u",
+        thread_id="t",
+        dataset_slug="ds",
+        is_public=True,
+        model="m",
+        environment="test",
+        user_input="hi",
+        view_state=None,
+    ) as trace:
+        assert trace.trace_id is None
+
+
 async def test_no_client_is_noop():
     """When client=None, the context manager records nothing and does not raise."""
     async with TurnTrace(