diff --git a/roboco/services/messaging.py b/roboco/services/messaging.py
index 20d10e31..07158193 100644
--- a/roboco/services/messaging.py
+++ b/roboco/services/messaging.py
@@ -386,6 +386,27 @@ class MessagingService(BaseService):
         )
         return result.scalar_one_or_none()
 
+    async def _lock_group(self, group_id: UUID) -> GroupTable | None:
+        """Lock the group row for the rest of this transaction and re-read its
+        ``active_session_id`` under the lock.
+
+        ``SELECT ... FOR UPDATE`` serializes concurrent session creation for the
+        same group: a check-then-create on ``active_session_id`` otherwise races
+        (two concurrent posts both miss the active session, both INSERT a new
+        ACTIVE session, and the second ``flush`` overwrites the group's
+        ``active_session_id`` — orphaning the first session as a forever-ACTIVE,
+        unreferenced row). ``populate_existing`` refreshes the identity-map-cached
+        instance's columns under the lock so the re-check sees the winner's link,
+        not the stale pre-lock value.
+        """
+        result = await self.session.execute(
+            select(GroupTable)
+            .where(GroupTable.id == group_id)
+            .with_for_update()
+            .execution_options(populate_existing=True)
+        )
+        return result.scalar_one_or_none()
+
     async def list_groups_in_channel(self, channel_id: UUID) -> list[GroupTable]:
         """List all groups in a channel."""
         result = await self.session.execute(
@@ -429,6 +450,16 @@ class MessagingService(BaseService):
         # sessions (the smoke run showed ~one session per message, and the CEO
         # could not hold a conversation). Only open a fresh session when none is
         # currently active (the prior one closed via timeout / boundary / merge).
+        #
+        # Serialize the check-then-create per group: lock the row and re-read
+        # ``active_session_id`` under the lock so a concurrent creator that won
+        # the race is visible here (we reuse its session) instead of both
+        # inserting and orphaning the loser's session. ``get_or_create_active_
+        # session`` and the channel-post adapter (L1868) route through here, so
+        # the lock covers every active-session creation entry point.
+        group = await self._lock_group(req.group_id)
+        if group is None:
+            raise ValueError(f"Group {req.group_id} not found")
         if group.active_session_id:
             existing = await self.get_session(cast("UUID", group.active_session_id))
             if existing is not None and existing.status == SessionStatus.ACTIVE:
diff --git a/tests/unit/services/test_messaging_session_race.py b/tests/unit/services/test_messaging_session_race.py
new file mode 100644
index 00000000..2504c166
--- /dev/null
+++ b/tests/unit/services/test_messaging_session_race.py
@@ -0,0 +1,110 @@
+"""F056: ``create_session`` (and its delegate ``get_or_create_active_session``,
+plus the L1868 channel-post adapter that routes through it) must not orphan an
+ACTIVE session under concurrent posts.
+
+``create_session`` does a plain check-then-create: read ``group.active_session_id``,
+reuse if ACTIVE, else INSERT a new ACTIVE session and point the group at it.
+Two concurrent posts can both miss the active session, both INSERT, and the
+second ``flush`` overwrites ``group.active_session_id`` — the first session
+stays ACTIVE but unreferenced (orphaned) forever. There is no DB uniqueness on
+``(group_id, status='active')`` (tables.py:1121-1125 only carries indexes), so
+nothing stops the double-insert.
+
+The fix: lock the group row (``SELECT ... FOR UPDATE``) and re-read
+``active_session_id`` under the lock before deciding to create, so concurrent
+callers serialize per group and the loser reuses the winner's session.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+from roboco.db.tables import SessionTable
+from roboco.models.base import SessionStatus
+from roboco.models.messaging import SessionCreateRequest
+from roboco.services.messaging import MessagingService
+from sqlalchemy.dialects import postgresql
+
+_GROUP_ID = MagicMock(name="group-id")
+_WINNER_SESSION_ID = MagicMock(name="winner-session-id")
+
+
+@pytest.mark.asyncio
+async def test_lock_group_emits_for_update() -> None:
+    """``_lock_group`` must issue ``SELECT ... FOR UPDATE`` (the row lock that
+    serializes concurrent session creation per group)."""
+    session = AsyncMock()
+    captured: list[Any] = []
+    result_mock = MagicMock()
+    result_mock.scalar_one_or_none.return_value = MagicMock(active_session_id=None)
+
+    async def _exec(stmt: Any) -> Any:
+        captured.append(stmt)
+        return result_mock
+
+    session.execute = AsyncMock(side_effect=_exec)
+    svc = MessagingService(session)
+
+    await svc._lock_group(_GROUP_ID)
+
+    sql = str(
+        captured[0].compile(
+            dialect=postgresql.dialect(), compile_kwargs={"literal_binds": True}
+        )
+    )
+    assert "FOR UPDATE" in sql
+
+
+@pytest.mark.asyncio
+async def test_create_session_race_loser_reuses_winner_under_lock() -> None:
+    """Concurrent posts: caller A wins the race and links its session while
+    caller B is between the check and the create. Caller B locks the group,
+    re-reads ``active_session_id`` (now A's session), and reuses it — no second
+    ACTIVE session is created (no orphan)."""
+    session = AsyncMock()
+    session.add = MagicMock()
+    session.flush = AsyncMock()
+    svc = MessagingService(session)
+
+    winner = MagicMock(name="winner-session", status=SessionStatus.ACTIVE)
+    svc.get_group = AsyncMock(return_value=MagicMock(active_session_id=None))
+    # Under the lock, the group now reflects the winner's link.
+    svc._lock_group = AsyncMock(
+        return_value=MagicMock(active_session_id=_WINNER_SESSION_ID)
+    )
+    svc.get_session = AsyncMock(return_value=winner)
+
+    result = await svc.create_session(SessionCreateRequest(group_id=_GROUP_ID))
+
+    assert result is winner
+    svc._lock_group.assert_awaited_once()
+    svc.get_session.assert_awaited_once()
+    session.add.assert_not_called()  # no orphaning INSERT
+    session.flush.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_create_session_creates_when_no_active_under_lock() -> None:
+    """No race: under the lock there is still no active session, so create a
+    new ACTIVE session and link it on the group (regression guard — the lock
+    must not break the happy path)."""
+    session = AsyncMock()
+    session.add = MagicMock()
+    session.flush = AsyncMock()
+    svc = MessagingService(session)
+
+    locked_group = MagicMock(active_session_id=None)
+    svc.get_group = AsyncMock(return_value=MagicMock(active_session_id=None))
+    svc._lock_group = AsyncMock(return_value=locked_group)
+    svc.get_session = AsyncMock()  # should NOT be called (no active id)
+
+    result = await svc.create_session(SessionCreateRequest(group_id=_GROUP_ID))
+
+    assert isinstance(result, SessionTable)
+    assert result.status == SessionStatus.ACTIVE
+    svc._lock_group.assert_awaited_once()
+    svc.get_session.assert_not_awaited()
+    session.add.assert_called_once()
+    assert session.flush.await_count >= 1