Playback, commands for later or part of the way, brightness amounts, and 1.16.1

[AboveColin]

Fixes #32. Fixes #35.

Playback is not on or off

The action criteria described turn_on as "... start it ..." and turn_off as "... or stop it". So "stop the music" became HassTurnOff, which fails on a music player that has no turn_off.

The criteria now describe power only, and none_of_these names playing, pausing, stopping and skipping media. Measured on hosted Jev with a playing Chromecast Audio, a TV, a light, a switch and a script exposed, one run per sentence:

Sentence	Before	After
stop the music	turn_off 0.98	none_of_these 0.99
Wyłącz muzykę.	turn_off 0.98	none_of_these 0.93
pause the music	none_of_these 0.32	none_of_these 0.99
play Metallica in the salon	turn_on 0.43	none_of_these 0.99
Puść mi metallikę w salonie.	none_of_these 0.47	none_of_these 0.92
next song	none_of_these 0.92	none_of_these 1.00
turn off the TV	turn_off 1.00	turn_off 0.98
turn on the TV	turn_on 0.53	turn_on 0.60
Wyłącz telewizor	turn_off 1.00	turn_off 0.95
turn off the kitchen light	turn_off 0.90	turn_off 0.85
switch on the fan	turn_on 0.67	turn_on 0.68
run the music voice request script	turn_on 1.00	turn_on 1.00

The TV and the fan were already on, which is why those two sit near 0.6.

A failed intent goes to the fallback agent

An IntentHandleError now goes to the fallback agent with the sentence unchanged. Home Assistant raises it only when no target succeeded (ServiceIntentHandler collects each failed call and raises when success_results is empty), so nothing changed first. With no fallback agent, the reply is an error with FAILED_TO_HANDLE. Any other intent error is also an error reply now. Before, it was spoken as action_done, and a satellite reads that as a command that went through.

Commands for later or part of the way

Home Assistant's on and off intents have no timer, no condition and no position. So "turn off the lamp in 10 minutes" scored turn_off 0.97 and would have turned it off at once, and "open the blinds halfway" scored turn_on 0.80 and would have opened them fully. Two new questions refuse these, later (another time, a duration or a condition) and part (a cover part of the way), each at 0.5.

One question for both was tried first. It refused all the timed sentences, but it also scored "set the lamp to 40 percent" at 0.64 and "dim the ceiling light to 30%" at 0.82, because a brightness level reads as a partial amount. Changing the action criteria instead dropped "turn on the lamp" from 0.60 to 0.44. So they are two separate questions.

Measured on hosted Jev with 12 entities exposed, one run per sentence:

Sentence	action	later	part	Result
turn off the lamp in 10 minutes	turn_off 0.97	0.98	0.04	refuses
turn on the coffee machine at 7	turn_on 0.98	0.98	0.04	refuses
turn on the lamp for 5 minutes	turn_on 0.71	0.96	0.04	refuses
turn on the lamp when I get home	turn_on 0.90	0.97	0.05	refuses
zet de lamp over 5 minuten uit	turn_off 0.98	0.98	0.05	refuses
wyłącz lampę za 10 minut	turn_off 0.92	0.98	0.05	refuses
schalte die Lampe um 22 Uhr aus	turn_off 0.97	0.98	0.05	refuses
open the blinds halfway	turn_on 0.80	0.03	0.98	refuses
close the blinds to 30 percent	turn_off 0.76	0.03	0.97	refuses
set the lamp to 40 percent in 10 minutes	set_brightness 1.00	0.98	0.10	refuses
turn off the lamp in the living room	turn_off 1.00	0.03	0.04	acts
open the blinds fully	turn_on 0.98	0.03	0.03	acts
set the lamp to 40 percent	set_brightness 1.00	0.03	0.10	acts
dim the ceiling light to 30%	set_brightness 0.95	0.03	0.10	acts
turn the lamp on at 10 percent	set_brightness 0.77	0.10	0.10	acts
is the lamp on?	get_state 1.00	0.03	0.03	acts
are the blinds open	get_state 1.00	0.04	0.09	acts

Of 37 sentences, all 15 that should refuse refused, and 21 of the 22 that should act acted. The other one is "turn on the lamp" at 0.56 with the lamp already on, which the agent answers with "Lamp is already on". It scored 0.61 to 0.64 with and without the new questions, three runs each, so the new questions did not cause it.

The two questions cost 127 input tokens and 32 output tokens: 1,696 to 1,823 for the same sentence. Warm latency was 261 ms without them and 263 ms with them.

Brightness: "to" and "by"

A change word anywhere in the sentence used to be enough to call a number relative, and only some words counted. So 45 of 52 relative sentences in 14 languages set the amount as the level, such as "turn the lamp down 20%" giving 20, and "sube el brillo un 20%" giving 20.

find_brightness now reads the words next to the number:

• A word for "to" in front of it (to, auf, naar, à, al, para, do, till, до, 到 and others), or the Hungarian -ra/-re suffix, makes it a level.
• A word for "by" in front of it (by, um, met, de, del, un, o, med), the Hungarian -kal/-kel suffix, or a comparative after it, makes it an amount.
• With neither, a change word anywhere (increase, lower, dim, erhöhe, 调亮 and the rest) makes it an amount.

An amount gives no brightness, so the sentence goes to the fallback agent.

Test set	Wrong before	Wrong after
52 relative sentences, 14 languages	45	0
28 level sentences, 14 languages	1	0
24 more, mostly a level with a change word, such as "turn up the lamp to 80%"	3	0

Hungarian also gets százalék and the level stems fény, halvány and világos:

Text	Before	After
vedd 20%-kal halványabbra	20	None
kapcsold fel 40 százalékra	None	40
állítsd a fényerejét 40-re	None	40

Checks

• 27 new tests and cases fail on main and pass here.
• ruff check and format, mypy --strict, pytest: 520 passed, 12 skipped, coverage 97%, config_flow 100%.
• mkdocs --strict, hassfest: 0 invalid integrations.
• On a test instance running this branch, through the Assist pipeline: "turn off the office lamp in 10 minutes" and "turn on the office lamp when I get home" refused with the reason "for another time or on a condition". "open the blinds halfway" refused with "a position part of the way". "increase the office lamp brightness by 20%" refused with no brightness. "stop the music" refused as not a house command. "turn on the office lamp", "turn off the office lamp" and "set the office lamp to 40 percent" acted.