Repository navigation
v3.46.0-beta.4
Pre-release
Pre-release
Pre-release of v3.46.0, for field validation of the bundle before the stable cut. HACS offers it only to users who opted into betas. It contains everything on main as of 2f704aa, so it includes all of v3.46.0-beta.3.
New since beta.3
- A long conversation now remembers the turns that were trimmed from it (#1048). When a conversation outgrows the model's context, its oldest turns are replaced by a summary. Since v2.0.0 that summary was wiped at the start of every turn, so the agent never saw it and the trimmed turns were simply forgotten. The summary is now kept, and later summaries add notes for the newly trimmed turns instead of rewriting the whole thing. Past 4,000 characters it is shortened.
- On a voice satellite the agent stops ending replies with a question it does not need answered (#954). A reply that ends in a question makes the satellite listen again, so a closing "Is there anything else I can help with?" cost another turn, and a declined offer was often followed by a second one. This lowers how often it happens and does not remove it; typed chats are unchanged.
What to try
- A conversation long enough to be trimmed (many tool-using turns), then a question about something said early on.
- On a satellite: ask about a device, decline whatever the agent offers, and note whether it asks again. A question the agent needs answered ("which bedroom light?") should still be asked.
Still being validated from earlier betas: summarizing a conversation that used tools on OpenAI or Anthropic (#1045), a second Ollama server as a fallback running the feature's model (#1022), the 8-second failover from an Ollama server that never answers (#1026), speaker recognition on voice satellites, memory management, and Top P per feature.
Tested against Home Assistant 2026.11.2.
The full notes are in the [Unreleased] section of CHANGELOG.md.