This release introduces the Health tab — a daily, plain-language read on how your home is really doing — and makes the AI architect self-correcting, so the automations it writes now fix themselves before you ever see an error. It also matures Recipes with event triggers, per-integration device mapping, and a cleaner catalog, and makes spoken announcements reliable across every speaker.
✨ Highlights
🔍 Health — a daily read on your home's wellbeing
A new Health tab answers the question you actually care about: what's worth fixing or improving right now? Once a day, Selora AI quietly audits your home and writes it up in plain language — grouped by device, with the impact explained and one-tap chips to act on each item.
- A daily home audit — the AI reviews your automations, devices, and detected patterns and surfaces concrete, prioritized recommendations, not a wall of raw warnings. Tap a suggestion's chip to carry it straight into chat and act on it.
- Local health signals — underneath the audit, a deterministic Troubleshooting section flags real problems: unavailable devices, flapping sensors, batteries running low, weak signal, devices that have gone silent, and unhealthy integrations. It's grouped per device with direct links, and it runs entirely on-device with no AI required.
- Private by design — the health layer is fully local and computed offline on your own hardware. For customers on a paid plan, a home inventory can be shared with Selora to power remote monitoring and management of their installation — this is entirely optional and can be turned off anytime from your Connect account. Everyone else runs the full Health experience locally with nothing leaving the network.
⚠️ Note: The daily written audit needs a configured LLM provider (cloud or local). Without one, the local health-signal Troubleshooting view still works on its own.
🤖 Self-correcting automations + an agent activity timeline
When you ask the architect to build an automation, it now validates and repairs its own work before handing it back. If a generated automation would fail Home Assistant's checks, the AI feeds the error back to itself and retries — so you get a working automation instead of a broken one and a cryptic message.
- Retry-on-failure — invalid triggers, unknown entities, and mis-scoped multi-phase routines are caught and corrected automatically.
- Agent activity timeline — chat now shows a step-by-step timeline of what the AI actually did (which tools it called, what it checked, how it fixed things), so its reasoning is transparent instead of a black box.
- The architect can now resolve and reuse scenes and correctly time-scope multi-phase automations (e.g. "do X, then Y an hour later") instead of firing everything at once.
🍳 Recipes, leveled up
Building on the Recipes catalog from v0.11.0, this release makes the install experience richer and more accurate:
- Event triggers as roles — recipes can now be driven by event entities (a doorbell press, a button click), not just switches and sensors.
- Integration-scoped roles — a recipe can require devices from a specific integration, so the wizard only offers the hardware that will actually work.
- A clearer catalog — recipes are grouped by install state (installed vs. available), each integration shows its logo in the wizard, and recipes that need a newer version of the integration are hidden instead of failing at install.
- Smoother flow — recipe badges and row menus were polished, and the install list refreshes immediately after the wizard finishes.
🔊 Voice Announcements
Spoken announcements now work reliably regardless of how a device or automation was configured:
tts.*_sayservice calls are canonicalized totts.speak, and unregistered legacy announcements are repaired automatically.- The TTS engine is resolved to a live entity before speaking, so announcements no longer silently fail on installs where the default speaker isn't available.
💬 Chat & Reliability
- Chat settles to a clear final state after running a command — no more lingering spinners or ambiguous endings.
- Cloud session errors are classified and localized, so a failure comes back as a readable message in your language instead of a raw stack trace.
- Leaked tool-call markup is suppressed and the tool-loop budget is steered, keeping model "thinking" syntax out of your chat.
- Better doorbell handling — the AI is guided to pick the correct doorbell button entity when wiring up automations.
- Device autocomplete no longer shows duplicates, and the menu stays on-screen.
🐛 Other Fixes
- Smarter suggestions — pattern-based suggestions are capped and quality-gated, so you see fewer, better ideas.
- Stale device references are cleared from caches when hardware is removed or replaced.
- Language detection now requires two signals before switching a reply's language, avoiding false switches on a single ambiguous word.
- Unconfigured Selora AI entries are ignored instead of causing errors.
- Header tabs collapse gracefully on narrow panels, the suggestions row no longer breaks on small screens, and Clear learned data moved to its own settings card with a tidier layout.