fix: remove flush() call and disable write queues for SQLite to prevent deadlock

The flush() call was causing deadlocks because:
- WAL mode allows only ONE concurrent writer despite multiple readers
- Write queues with multiple workers create writer contention
- flush() waits for all queued writes while workers block on DB locks

Solution: Disable write queues for SQLite and let WAL mode handle concurrency
directly. This makes writes synchronous, eliminating visibility issues without
deadlocks. Keep queues enabled for other backends like Postgres.

src/llama_stack/providers/inline/agents/meta_reference/responses/openai_responses.py

@@ -316,9 +316,6 @@ class OpenAIResponsesImpl:
 
             if final_response is None:
                 raise ValueError("The response stream never reached a terminal state")
-
-            # Flush any queued writes to ensure immediate visibility
-            await self.responses_store.flush()
             return final_response
 
     async def _create_streaming_response(

src/llama_stack/providers/utils/inference/inference_store.py

@@ -16,12 +16,12 @@ from llama_stack.apis.inference import (
     Order,
 )
 from llama_stack.core.datatypes import AccessRule
-from llama_stack.core.storage.datatypes import InferenceStoreReference
+from llama_stack.core.storage.datatypes import InferenceStoreReference, StorageBackendType
 from llama_stack.log import get_logger
 
 from ..sqlstore.api import ColumnDefinition, ColumnType
 from ..sqlstore.authorized_sqlstore import AuthorizedSqlStore
-from ..sqlstore.sqlstore import sqlstore_impl
+from ..sqlstore.sqlstore import _SQLSTORE_BACKENDS, sqlstore_impl
 
 logger = get_logger(name=__name__, category="inference")
 
@@ -48,6 +48,13 @@ class InferenceStore:
         base_store = sqlstore_impl(self.reference)
         self.sql_store = AuthorizedSqlStore(base_store, self.policy)
 
+        # Disable write queue for SQLite since WAL mode handles concurrency
+        # Keep it enabled for other backends (like Postgres) for performance
+        backend_config = _SQLSTORE_BACKENDS.get(self.reference.backend)
+        if backend_config and backend_config.type == StorageBackendType.SQL_SQLITE:
+            self.enable_write_queue = False
+            logger.debug("Write queue disabled for SQLite (WAL mode handles concurrency)")
+
         await self.sql_store.create_table(
             "chat_completions",
             {

src/llama_stack/providers/utils/responses/responses_store.py

@@ -19,12 +19,12 @@ from llama_stack.apis.agents.openai_responses import (
 )
 from llama_stack.apis.inference import OpenAIMessageParam
 from llama_stack.core.datatypes import AccessRule
-from llama_stack.core.storage.datatypes import ResponsesStoreReference, SqlStoreReference
+from llama_stack.core.storage.datatypes import ResponsesStoreReference, SqlStoreReference, StorageBackendType
 from llama_stack.log import get_logger
 
 from ..sqlstore.api import ColumnDefinition, ColumnType
 from ..sqlstore.authorized_sqlstore import AuthorizedSqlStore
-from ..sqlstore.sqlstore import sqlstore_impl
+from ..sqlstore.sqlstore import _SQLSTORE_BACKENDS, sqlstore_impl
 
 logger = get_logger(name=__name__, category="openai_responses")
 
@@ -70,6 +70,13 @@ class ResponsesStore:
         base_store = sqlstore_impl(self.reference)
         self.sql_store = AuthorizedSqlStore(base_store, self.policy)
 
+        # Disable write queue for SQLite since WAL mode handles concurrency
+        # Keep it enabled for other backends (like Postgres) for performance
+        backend_config = _SQLSTORE_BACKENDS.get(self.reference.backend)
+        if backend_config and backend_config.type == StorageBackendType.SQL_SQLITE:
+            self.enable_write_queue = False
+            logger.debug("Write queue disabled for SQLite (WAL mode handles concurrency)")
+
         await self.sql_store.create_table(
             "openai_responses",
             {