Pre-release of v3.45.0, for staging validation of #783 (Sentinel autonomy level). HACS offers it only to users who opted into betas. It contains everything on main as of 2a4ecdd, including v3.44.4.
Added
- Sentinel's autonomy level can now be set in Sentinel's Advanced setup (#783).
- Before, Sentinel ran at level 1 (notify only) unless
sentinel_set_autonomy_levelraised it, which lasted only an hour. The setting the engine read for its normal level could not be set anywhere. - Autonomy level (0–3, default 1) is now kept across restarts and Basic setup re-runs.
- Level 0 now silences finding alerts and the daily digest, as documented, and alerts already waiting in a batch are dropped (a warning lists their kinds); change alerts are never held in a batch, so none is dropped this way. Before, it still sent every notification and only skipped triage. Findings are still recorded in the audit, and nothing seen at level 0 is lost: a change (a new device or client, a new admin or token, a public IP or port-mapping change) alerts once the level is raised, one alert per kind of change, even one undone while at 0, such as a port mapping added and removed again or a device paired and removed (#799). That alert starts with when the change was seen and says it may have been undone since, lists every device, client or account seen and not already reported (skipping any you trusted meanwhile) at the highest severity seen, and points to the trust device service instead of a Trust device button. Each public IP or port-mapping change seen is told. A change a cooldown, quiet hours or presence grace stopped at 0 is owed too, and LLM triage never drops an owed alert. A condition that is still true alerts again once its cooldown ends, without muting other devices meanwhile. An alert being prepared when level 0 is set is not sent. Level 0 in force when Sentinel starts is logged as a warning. Baseline and one-time setup notices still appear.
- The
sentinel_healthsensor shows the level in force and when an override ends (autonomy_level,autonomy_override_expires), updated as soon as the service sets a level. A service call that sets level 0 is logged as a warning. - Autonomy override duration (default 60 minutes, up to 30 days) sets how long a level from the service lasts, and a Basic setup re-run keeps it. An override now also ends when the configured level changes, so a raised one cannot outlive a lowered setting. A level or duration an older install kept in a different form (a level of
0.0, a duration of"5") is shown and saved as it runs, so saving the form unchanged changes nothing. - Levels 2 and 3 are reserved for automatic actions, which are not available yet (#789), and now notify exactly like level 1. Before, they could drop a finding whose device was unavailable. The docs described them as acting on their own, which no setup could do, and now say so.
- A level an older install stored in its main options is kept on upgrade, and shown in the form.
- When automatic actions arrive, Sentinel will never run a critical action on its own, whatever is allowed: an unlock, an alarm panel or cover action (also through
homeassistant.*), anything a Critical Action PIN rule covers (reaching entities by device class, name and group, as the PIN gate does since v3.44.3), or anything that starts a script, scene or automation. That guard is already in place, and automatic actions stay off even if an older setting enabled them.
- Before, Sentinel ran at level 1 (notify only) unless
Fixed
- Opening or closing a cover by voice no longer fails when other device tools rank higher (#795). "Open the study shade" could fail with "I couldn't find a suitable tool": the agent offers the model only the few tools that best match the request, and device tools such as a vacuum's start or a media player's unpause could outscore the general "turn on" tool, which is the only one Home Assistant gives an assistant for opening a cover. The model then guessed a tool that does not exist and the call was refused. A request with an English action word ("turn", "open", "close", ...), or a "yes" to the agent offering one ("Want me to open the Study Shade?"), now always offers the turn on and turn off tools, alongside what the ranking chose. Requests in other languages, and replies such as "the left one", still rely on the ranking. A tool you excluded is still never offered.
Changed
- The model-provider libraries are no longer pinned to one exact version each:
langchain-openai,langchain-anthropicandlangchain-ollamaaccept any 1.x release (langchain-ollamafrom 0.3.10),langchain-google-genaiany 3.x from 3.1.0, and theanthropicSDK anything from 0.125.0 below 2. When a Home Assistant release moved a library these depend on, such asopenaiin 2026.10, an exact pin could not be installed next to it and the integration failed to set up for everyone until a new release shipped (#735). Home Assistant can now pick a compatible release on its own. A fresh install, or the reinstall that follows a Home Assistant update, gets the newest release in each range: todaylangchain-openai1.7.2,langchain-anthropic1.7.6 withanthropic1.15.0,langchain-ollama1.1.0 andlangchain-google-genai3.2.0. The test suite runs against the newest releases every day.