Agent safety boundaries
Automation and MCP agents must follow the same rules as human operators.
Never
| Action | Reason |
|---|---|
Delete workspace/ |
Destroys site state, historian, model |
docker compose down -v |
Wipes volumes |
docker volume prune |
Irreversible data loss |
| Print tokens or passwords | Credential leak |
| BACnet write without explicit human approval | Live equipment risk |
| Expose API on public internet | OT security |
Always
| Action | When |
|---|---|
Back up workspace/ |
Before updates or destructive changes |
| Preflight CSV import | Before /api/csv/import/execute |
| Dry-run BACnet writes | Before /api/bacnet/write |
| Validate after changes | openfdd_health_check.sh |
Assignment rule
Bind drivers → Haystack IDs → FDD/CDL via /api/model/assignments before activating rules.
LAN / OT deployment
Run on loopback or behind VPN/TLS. JWT auth is required in production configurations.
See root AGENTS.md for session scripts.