The module-level _conn was being hit from FastAPI handlers, the retention
daemon thread, and asyncio.to_thread workers simultaneously. Sharing a
single sqlite3.Connection across threads is unsafe (cursors collide)
even with check_same_thread=False and WAL. The writer _lock didn't cover
readers, so a reader cursor could race a writer mid-statement.
Switch to threading.local(): each thread gets its own Connection via
_get_conn(). WAL handles concurrent readers/writer at the DB level;
busy_timeout=5000 absorbs short-lived "database is locked" when two
threads both try to BEGIN IMMEDIATE. The write-serialising _lock stays
— it keeps multi-statement writer blocks atomic and avoids busy-loop on
concurrent writers.
External access via db._conn replaced with a new db.has_running_application
helper (the only caller outside db.py).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
EiSiMo
617c76cb54