Selora AI v0.15.0: Retire Dead Devices and Preview Refinements Before You Accept
Selora AI v0.15.0 lets you delete devices that have been offline for good, shows a diff of what a refined proposal will actually change, and fixes provider compatibility across OpenAI reasoning models, Ollama, Gemini, and Anthropic.
Selora AI v0.15.0 is out. This one is about closing loops: retiring devices that are never coming back, and seeing exactly what a refinement will change before you accept it. It also carries a substantial round of provider compatibility fixes, so if you run OpenAI reasoning models, Ollama, Gemini, or Anthropic, there is something here for you.
Delete a device that’s gone for good
Every home accumulates them. A sensor whose battery you never replaced, a plug you took out of the wall, a bulb that finally died. Home Assistant keeps them around, and the Health tab keeps reporting them as offline forever.
A device that has been offline for seven days or more can now be deleted straight from the Health tab. The confirmation names what breaks if you go ahead, so you are not guessing whether an automation still depends on it.
Seven days is deliberate. It is long enough that a holiday power cut or a router replacement will not tempt you into deleting something you still want.
See what a refinement will change
When you ask Selora AI to adjust a proposal, the refined version now carries a View changes diff: what is currently on disk against what accepting the proposal would write.
The point is that “make it 15 minutes earlier” is easy to say and hard to verify. Now you read the change instead of trusting it.
Provider compatibility
Four separate problems, all of them the same shape: a provider that behaves slightly differently from the rest, and Selora AI not accounting for it.
- OpenAI reasoning models work again. The o-series and GPT-5 reject
max_tokensoutright. They no longer receive it. - Ollama and Selora AI Local size prompts correctly. Prompts are now built against the context window the server actually serves, not an assumed one.
- Ollama models without tool support no longer fail every turn. Some chat templates cannot accept tool definitions at all. Those models are now handled rather than broken.
- Gemini reports thinking tokens. They are counted in usage, and an output cap that could return a completely empty reply is gone.
- Anthropic replies arrive whole. A reply split across several content blocks was being truncated to the first one.
Smaller fixes
- Refined automations keep their card instead of collapsing into plain text.
- The delete-device dialog is translated across all 13 supported languages, with correct German and Russian plurals.
- Row icons line up with their titles.
Compatibility
The minimum Home Assistant version is unchanged at 2025.1.0.
See the full changelog for the complete list.
What’s Next
Dashboards. Asking for one, reshaping one, and moving cards around by describing what you want.
Update via HACS, or if you already have Selora AI installed, the update will appear in your Home Assistant settings.