SOVA Database Patterns

When working on sova/db/, ORM models, or migrations, follow these conventions. Reference: docs/database-guidelines.md.

Session Management

Always use the async context manager:

async with await get_session() as session:
    result = await session.execute(select(TaskRun))

Never acquire sessions without async with: it leaks on exception. All factories use expire_on_commit=False.

ORM Model Conventions (sova/db/models.py)

• Type-annotated columns: Mapped[str] with mapped_column(String(50))
• Money: Numeric(10, 6) with Decimal("0") default
• Timestamps: DateTime(timezone=True) with lambda: datetime.now(timezone.utc)
• FK constants: _FK_TASK_RUNS_ID = "task_runs.id"
• Indexes in __table_args__: Index("ix_{table}_{column}", "column")
• Input normalization: @validates decorator (e.g., strip # from issue numbers)

JSON Column NULL Gotcha

SQLAlchemy JSON defaults to none_as_null=False. Python None becomes JSON "null", not SQL NULL.

• TaskRun.handoff_json.isnot(None) only catches SQL NULLs
• Use if not handoff: to catch both None and {}

Terminal Status Sets

_TERMINAL = frozenset({TaskStatus.DONE, TaskStatus.FAILED, TaskStatus.REJECTED})  # state machine
TASK_RUN_TERMINAL = frozenset(
    {"done", "failed", "rejected", "interrupted", "paused", "awaiting_approval"}
)  # DB queries, sova/core/state.py

Use TASK_RUN_TERMINAL for DB queries. Always guard finalization:

if task_run.status in TASK_RUN_TERMINAL:
    return  # don't overwrite, but still update cost

Migrations

• Always use batch_alter_table (SQLite requires it)
• Use idempotent helpers: _column_exists(), _table_exists(), _index_exists()
• Sequential numbering (currently up to 034), not Alembic UUIDs
• Engine disposal after migration for file-backed SQLite (stale schema cache)
• Self-healing: if Alembic fails, drops corrupted alembic_version, runs create_all + stamps head

Multi-Project

get_session(project_dir=...) routes to per-project engine. Each project gets its own .claude/sova.db.