I think the problem, at least in my case, is: the command I gave is all local, both the timer part and the command that be executed later, but just the timer is processed as local. The real command, when executed, is using the AI even should be processed local.
Maybe we could check if the command is local and act accordingly when we executed it (after the timer period)?
Yes, the code must implement the logic for selecting the executing agent. If local command priority is enabled and/or “no control” mode is selected, conversation.home_assistant should be triggered. Otherwise, for full Assist mode, the current agent should be selected.
The issue is almost a year old.
Now that I understand the problem, I would open an issue in the main repository; the edits need to be made there. I hope someone will do this, as I’m too lazy.
2 Likes
Thanks. Please post the link to the issue so I can subscribe the notifications.
It’s definitely something within home assistant. It works for a while. Then I only get ‘unable to get response’. When I go into the settings of the Google Generative AI conversation agent in devices and only press submit to save the default values it all works again. I think I will raise a GitHub issue.
The issue with thinking in 2.5 and 3.0 may finally be fixed in 2026.2. The release notes mention upgrading the SDK to deal with an issue that sounds related and so far I haven’t hit any of these errors after upgrading.
I think we have a couple problems happening here: first, the one opened this topic, is this problem about generating response which just worked with Gemini 2.0, which will be discontinued soon. According @ghazm, the problem was fixed on 2026.2 and now Gemini 2.5 and 3.0 should work.
Another issue is the Home Assistant itself not using local processing for local commands when using a timer.
As said, the first one should be solved now, I will test this as soon as possible, and the second one has an issue opened by @mchk.
Hello! In my experience with 2026.2, I’m still getting the “Unable to get response”. Although I feel is not as constant as before, but it seems to fix itself when I change from voice queries to written ones or the other way around. For example, when I get that error while I’m talking with the assist by voice, I try 3 times the same query by voice and get the same error, but then I type the same query on the same chat window and it responds successfully. It acts the same when I get the error typing the query and then I change to voice.
Also, in 2026.2, I’ve been sometimes having an Unexpected error during intent recognition (intent-failed) error when Gemini tries to google search for a response using the other assistant meant for google searching. It responds an incomplete response and in the trace it shows that error. I don’t know what might be the reason.
Gemini 2.0 seems to not be a good solution for HA anymore because it’s not smart enough for some HA functionalities, like adding items to To-Do lists. That only works for me when I use versions 2.5+.