Multi-agent review (5 confirmed) + Codex found real issues in the rework:
- [P1 security] Shared `shelly` ACL had readwrite +/rpc, so a compromised
device could inject Switch.Set into ANY other device's <prefix>/rpc. Now
read-only on +/rpc (receive own commands) + write only homeos/rpc (reply).
Broker-verified: a shelly publish to victim/rpc is denied; laravel's is not.
- [P2 security] `+` wildcard reaches reserved homeos/ring namespaces → bogus
device. Ingest now rejects RESERVED_PREFIXES (homeos/ring/$SYS).
- [P2 security] Unbounded auto-onboarding = DB-exhaustion DoS. Added a device
cap (homeos.mqtt.max_devices, default 250).
- [P2 correctness] Every Shelly `input` became a phantom window contact (wall
switches shown as windows, possibly inverted). `input` is now a generic
sensor; the user PROMOTES specific inputs to window/door contacts on the
device page (invert-aware), stored in config->input_roles and applied in the
ingest — this is the "assign contacts" flow the user asked for.
- [P3 ux] Motion pill read `active`; producer writes `on`. Now reads both.
- [P1 migrations] Dedup computed survivor keys once; 3+ duplicates could
collide on unique(device_id,key). Re-query per duplicate (ring + mqtt_prefix).
+9 tests (reserved prefix, cap, input generic/promoted/inverted, demo echo).
Live-verified: input published → onboarded as input → assigned window contact
via UI → appears on Fenster page, persists across messages. Suite 57 green,
12/12 tabs clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
71b3028a8e