fix(core): degrade gracefully when sqlite-vec cannot load on init (#774)

Signed-off-by: phernandez <paul@basicmachines.co>

src/basic_memory/repository/sqlite_search_repository.py

@@ -94,9 +94,23 @@ class SQLiteSearchRepository(SearchRepositoryBase):
             raise e
 
         # Fail fast: create vector tables at startup so missing sqlite-vec
-        # or embedding provider errors surface immediately
+        # or embedding provider errors surface immediately.
+        # Trigger: the runtime semantic stack (sqlite-vec extension or embedding
+        #          provider) is unavailable at startup.
+        # Why: failing the whole MCP boot for a search-only feature blocks
+        #      Claude Desktop's handshake (#711). Keyword-only search is a
+        #      reasonable fallback while the user resolves the dependency.
+        # Outcome: log the cause, mark this repository as semantic-disabled so
+        #      downstream calls short-circuit cleanly, and let init complete.
         if self._semantic_enabled:
-            await self._ensure_vector_tables()
+            try:
+                await self._ensure_vector_tables()
+            except SemanticDependenciesMissingError as exc:
+                logger.warning(
+                    f"Semantic search disabled: {exc}. "
+                    "Falling back to keyword-only search."
+                )
+                self._semantic_enabled = False
 
     # ------------------------------------------------------------------
     # FTS5 query preparation (backend-specific)
@@ -374,6 +388,25 @@ class SQLiteSearchRepository(SearchRepositoryBase):
             async_connection = await session.connection()
             raw_connection = await async_connection.get_raw_connection()
             driver_connection = raw_connection.driver_connection
+
+            # Trigger: the underlying CPython was built without sqlite extension support.
+            # Why: python.org's macOS installer ships a stripped sqlite3 module with no
+            #      enable_load_extension; when uvx happens to pick that interpreter (#711),
+            #      the AttributeError surfaces here and previously crashed startup before
+            #      Claude Desktop could complete its MCP handshake.
+            # Outcome: convert to SemanticDependenciesMissingError so the init-time
+            #      handler can degrade gracefully to keyword search instead of dying.
+            if not hasattr(driver_connection, "enable_load_extension"):
+                raise SemanticDependenciesMissingError(
+                    "This Python build does not support SQLite extension loading "
+                    "(no enable_load_extension on sqlite3.Connection). "
+                    "Common cause: python.org Python on macOS. "
+                    "Reinstall basic-memory under a Python that ships extension "
+                    "support (uv-managed CPython, Homebrew Python, or the official "
+                    "Docker image), or set semantic_search_enabled=false in config "
+                    "to silence this and use keyword-only search."
+                )
+
             await driver_connection.enable_load_extension(True)
             await driver_connection.load_extension(sqlite_vec.loadable_path())
             await driver_connection.enable_load_extension(False)

tests/repository/test_search_repository.py

@@ -107,6 +107,78 @@ async def test_init_search_index(search_repository, app_config):
         assert table_name == "search_index"
 
 
+@pytest.mark.asyncio
+async def test_init_search_index_degrades_when_extension_loading_unavailable(
+    search_repository, monkeypatch
+):
+    """Regression for #711: when sqlite-vec cannot be loaded (e.g. python.org Python
+    3.12 ships sqlite3 without enable_load_extension), init must NOT crash. It should
+    log a warning, mark the repository as semantic-disabled, and let the rest of the
+    process come up so Claude Desktop's MCP handshake completes."""
+    if is_postgres_backend(search_repository):
+        pytest.skip("python.org enable_load_extension issue is SQLite-specific")
+
+    from basic_memory.repository.semantic_errors import SemanticDependenciesMissingError
+
+    # Force the codepath even if semantic_search wasn't enabled by default.
+    search_repository._semantic_enabled = True
+
+    async def _raise_missing():
+        raise SemanticDependenciesMissingError("simulated: enable_load_extension missing")
+
+    monkeypatch.setattr(search_repository, "_ensure_vector_tables", _raise_missing)
+
+    # Must not raise — startup needs to complete even when the semantic stack is dead.
+    await search_repository.init_search_index()
+
+    assert search_repository._semantic_enabled is False, (
+        "Repository should mark itself semantic-disabled after a missing-deps error "
+        "so downstream calls short-circuit cleanly instead of re-attempting load."
+    )
+
+
+@pytest.mark.asyncio
+async def test_ensure_sqlite_vec_loaded_raises_typed_error_without_extension_support(
+    search_repository, monkeypatch
+):
+    """Regression for #711: AttributeError from a sqlite3.Connection that lacks
+    enable_load_extension must surface as SemanticDependenciesMissingError so the
+    init-time handler can degrade. Otherwise the AttributeError bubbles through and
+    crashes startup before Claude Desktop completes its handshake."""
+    if is_postgres_backend(search_repository):
+        pytest.skip("enable_load_extension is SQLite-specific")
+
+    from basic_memory.repository.semantic_errors import SemanticDependenciesMissingError
+    from sqlalchemy.exc import OperationalError as SAOperationalError
+
+    # Stub session that always reports vec missing on probe, then yields a connection
+    # whose driver_connection has no enable_load_extension attribute (mirroring the
+    # python.org sqlite3 build).
+    class _StubDriverConnection:
+        # Deliberately omit enable_load_extension to mimic the python.org build.
+        pass
+
+    class _StubRawConnection:
+        driver_connection = _StubDriverConnection()
+
+    class _StubAsyncConnection:
+        async def get_raw_connection(self):
+            return _StubRawConnection()
+
+    class _StubSession:
+        async def execute(self, _stmt):
+            # First (and any) probe call reports vec missing.
+            raise SAOperationalError("SELECT vec_version()", {}, Exception("no vec"))
+
+        async def connection(self):
+            return _StubAsyncConnection()
+
+    with pytest.raises(SemanticDependenciesMissingError) as exc_info:
+        await search_repository._ensure_sqlite_vec_loaded(_StubSession())
+
+    assert "enable_load_extension" in str(exc_info.value)
+
+
 @pytest.mark.asyncio
 async def test_init_search_index_preserves_data(search_repository, search_entity):
     """Regression test: calling init_search_index() twice should preserve indexed data.