Files
pipelines/docs/notes/storage-overview-retention.md

1.5 KiB

Storage overview and retention

Added read-only inventory and explicit preview/apply RPCs with a persisted, disabled-by-default hourly policy. The application can choose rolling days, logical-file byte target, and eligible linked-worktree inclusion. Limits are OR triggers; retained active/protected storage can exceed the target honestly.

Inventory is inode-deduplicated, does not follow symlinks, and reports partial measurements. Known workspace/repository roots survive history deletion. All operator-session directories are protected without the chat list's 128-row cap. Worktree removal verifies clean/merged state and the live registry again, retains nested/primary/locked/in-use trees, uses no force, and preserves branches.

Manual cleanup uses a ten-minute one-use preview token bound to the reviewed policy and candidate identities. Later candidates are not added. Completed-run removal reuses transactional protections and subscriber broadcasts. Automatic passes use the same selection/revalidation; there is no startup/save cleanup, chat-history deletion or live VACUUM.

The protocol runs asynchronously with one worker so scans do not block the shared socket's chat traffic. Busy requests fail promptly. Full pass context is 20 seconds; final worktree admission/removal has its own three-second budget.

No production scans, policy activation, cleanup or service restart were performed while developing this change. Backend delivery requires the operator's safe active-work deployment gate; the deployed service is never restarted for tests.