fix: NullPool for Celery workers + skip pgvector on first message

- Celery workers use NullPool to avoid "Future attached to a different
  loop" errors from stale pooled async connections across asyncio.run()
  calls. FastAPI keeps regular pool (single event loop, safe to reuse).
- Skip pgvector similarity search when no conversation history exists
  (first message) — saves ~3s embedding + query overhead.
- Wrap pgvector retrieval in try/except to prevent DB errors from
  blocking the LLM response.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

packages/orchestrator/orchestrator/tasks.py

@@ -493,21 +493,26 @@ async def _process_message(
         )
 
         # 2. Long-term: pgvector similarity search
+        # Skip if no conversation history exists yet (first message optimization —
+        # embedding + pgvector query adds ~3s before the first token appears)
         relevant_context: list[str] = []
-        if user_text:
-            query_embedding = embed_text(user_text)
-            rls_token = current_tenant_id.set(tenant_uuid)
+        if user_text and recent_messages:
             try:
-                async with async_session_factory() as session:
-                    relevant_context = await retrieve_relevant(
-                        session,
-                        tenant_uuid,
-                        agent.id,
-                        user_id,
-                        query_embedding,
-                    )
-            finally:
-                current_tenant_id.reset(rls_token)
+                query_embedding = embed_text(user_text)
+                rls_token = current_tenant_id.set(tenant_uuid)
+                try:
+                    async with async_session_factory() as session:
+                        relevant_context = await retrieve_relevant(
+                            session,
+                            tenant_uuid,
+                            agent.id,
+                            user_id,
+                            query_embedding,
+                        )
+                finally:
+                    current_tenant_id.reset(rls_token)
+            except Exception:
+                logger.warning("pgvector retrieval failed — continuing without long-term memory")
     finally:
         await redis_client2.aclose()
 

packages/shared/shared/db.py

@@ -14,19 +14,32 @@ from __future__ import annotations
 
 from collections.abc import AsyncGenerator
 
+import os
+
 from sqlalchemy.ext.asyncio import AsyncEngine, AsyncSession, async_sessionmaker, create_async_engine
+from sqlalchemy.pool import NullPool
 
 from shared.config import settings
 
 # ---------------------------------------------------------------------------
 # Engine — one per process; shared across all requests
+#
+# Celery workers use asyncio.run() per task, creating a new event loop each
+# time. Connection pools hold connections bound to the previous (closed) loop,
+# causing "Future attached to a different loop" errors. NullPool avoids this
+# by never reusing connections. FastAPI (single event loop) can safely use a
+# regular pool, but NullPool works fine there too with minimal overhead.
 # ---------------------------------------------------------------------------
+_is_celery_worker = "celery" in os.environ.get("_", "") or "celery" in " ".join(os.sys.argv)
+
 engine: AsyncEngine = create_async_engine(
     settings.database_url,
     echo=settings.debug,
-    pool_pre_ping=True,
-    pool_size=10,
-    max_overflow=20,
+    **({"poolclass": NullPool} if _is_celery_worker else {
+        "pool_pre_ping": True,
+        "pool_size": 10,
+        "max_overflow": 20,
+    }),
 )
 
 # ---------------------------------------------------------------------------