Local persistence¶
Persistence owns SQL/file operations, not trading decisions or experiment scheduling. Platform injects these capabilities and controls their lifetime. ORM sessions stay within short units of work; detached values cross into callers.
| Location | Responsibility |
|---|---|
sql/database.py, sql/models.py |
Database/session lifetime and ORM entities |
sql/repositories/, sql/migrations/ |
Queries, atomic writes and preserved migration history |
agents/ |
Private source snapshots, Python/dependency environments, readiness/lease records and SDK artifacts |
prepared/ |
Prepared artifact inventory and publication |
datasets/ |
Saved local source references and managed copies |
workspace/ |
App temporary workspaces, source cache files and cleanup |
replay/ |
Isolated recorded-run inputs/destination handling |
agent_log_files.py |
Ordered JSON activity files later archived into the existing run record |
locking.py |
Shared POSIX/Windows exclusive file locks; callers own streams and lock lifetime |
paths.py |
OS-specific application data location |
Financial events/projections/checkpoints are saved through the existing transaction path. Tool/control records use ordinary JSON during execution and are archived at terminal output; there is no additional tool-receipt SQL table. Preserve migration identifiers and historical rows when changing code. Decision order references reconstruct their original ledger prefix; current projection rows cannot substitute for past observations.
Temporary workspaces are separate from persistent account storage. Filesystem locks and cleanup have platform-specific limits; do not infer full Windows App support from Agent host CI. Reset/backup semantics belong to Platform lifecycle, not an arbitrary repository.
See current support, architecture and tests.