← all posts site ↗

my smart home assistant got a brain

my home assistant voice assistant was the dumbest it has ever been on record. i said turn on my room lights and it asked which room because it could not find any lights in the bedroom. said kill them all and it claimed success while leaving two lamps on. dug in and the reasons were embarrassingly concrete, not model stupidity. my lights had no area assigned, the agent's tool matched areas by friendly name so a light called ceiling light never contains the word bedroom, and the scene it trusted to turn everything off simply omitted two lamps.

the story underneath is that the assistant is not the default model anymore. it routes through a bridge to the same agent my glasses use, which is faster and should know my home. but the bridge only forwards my words, it drops all the extra instructions the old setup had, so it knew nothing about rooms or scenes. fixed the tool to read the actual device registry instead of guessing, wrote the home layout into the prompt, and added the missing lamps to the off scene.

then speech to text failed mid test and it felt broken all over again, but it was just me restarting the system mid session. the real recurring dumbness was a fast path in the bridge, queries without trigger words skipped the agent entirely and hit a bare model with no tools, so whats the weather and now turn it off both got answered by a model that can do neither. widened the trigger list so almost everything routes to the real agent now.

the end result is it names the actual lights in the bedroom from memory, turns them on, and kills all of them correctly, including the ones the scene used to skip. the whole episode was a chain of small lies, unassigned areas, dropped context, a scene that lied about its contents, and a routing shortcut, each one easy to fix once the logs said what happened.