Overview cold start:
- InstanceResourceService persists every live fetch into status_snapshots
(category 'resources') and serves a snapshot younger than 2 minutes
before calling upstream, so the first fleet overview after a restart
costs zero device requests; instance deletion cleans up via FK cascade
Bounded retention for the four fastest-growing tables (15-minute sweep):
- connection_logs: one row per probe per beat had no automatic cleanup
- audit_events: new pruneBefore (90d)
- operation_preparations: every prepare/retry attempt inserted a row,
terminal rows now expire after 7 days
- notification_queue: terminal rows pruned after 30 days
SSE events now cover instance lifecycle:
- create/update emit instance envelopes from the routes; the two-phase
delete emits a job envelope on every terminal transition plus an
instance envelope when the node actually disappears, so fleet and
instance views invalidate in real time
Linux ops:
- 'install.sh unit' writes systemd user units for API and gateway with
Restart=on-failure, 0600 secret injection, and
MULTI_SIMADMIN_SYSTEMD_UNIT wired so console self-update restarts via
systemctl -- closing the self-update loop on Linux
The production composition hardwired the macOS Keychain (/usr/bin/security),
which made Linux deployment impossible. Now:
- SecretStore gains a stable provider identity persisted in
secret_references.provider; services stop hardcoding 'macos-keychain'
- shared reference codec (secret-reference.ts) understands both
keychain:// and secret-file:// references
- FileSecretStore: single 0600 JSON map under the data root, atomic
temp+rename writes, serialized in-process, same validation envelope
- production composition picks the backend via
MULTI_SIMADMIN_SECRET_BACKEND (default: Keychain on darwin, file store
elsewhere) and readiness probes the matching backend
chick