KIA e-Niro / Niro Electric 2021 (Hyundai): Car's DAW (Driver Attention Warning) rapidly reducing, leading to "Consider taking a break" coffee-cup.

_(Preamble: This is a 2021 KIA e-Niro / Niro Electric, "variant" C5E11, "version" E11DY1 (picking some data from registration). I just bought Comma 3X with Hyundai C harness, and the installation went great, including the OBDII. AFAICS everything works as expected, which evidently for this make/model means that speed/brake is still from car)_

On the driver attention display of the car, I've never seen the "attention-meter" at anything else than fully alert ("High"), or at most one level down (of the 5 levels).

However, after having run with openpilot for some max 20 minutes, I suddenly got a coffee-cup warning that I've never seen before, stating "Ta en pause", i.e. "take a break" in Norwegian. (AFAIU, the warning reads "consider taking a break" in English).

I then displayed the attention meter (on the car), and it showed "Low". After restarting, it shows high, and then less than a minute of openpilot engaging, it drops down one level.

When this happens, I am fully alert (and as mentioned have pretty much never seen anything other than "high" before), but I do let openpilot do the driving and don't touch the steering wheel.

~From limited research, it doesn't really seem to matter if I actually do nudge a bit on the steering wheel either.~
Edit: After new test: Continuously holding on the steering wheel, somewhat "actively resisting" openpilot's steering, can very slowly get the attention level back up from coffee-cup and the lowest "orange" level (1).

Googled, but not seen this discussed, so filing this bug. Is it possible for openpilot to send some commands to override this to its own evaluated attention - or at least send something that makes the car believe that I'm alert?

It might correlate with #1108, which is a Hyundai (although Ioniq, whereas Kona is the sibling for e-Niro).
This PR references "Consider taking a break": https://github.com/commaai/openpilot/pull/30679
This PR seems relevant wrt. mitigating the problem? https://github.com/commaai/openpilot/pull/24711

This doesn't seem to affect anything, but it is obviously unfair and annoying! (..and I worry a tad if there is black-box recorder, and could affect an insurance situation).

"Take a break"

Attention level showing "low" (cruise here disengaged)

The meter drops 1 level after few minutes (Cruise disengaged - it went back up to full after 4 minutes of manual driving)

Start by reviewing the related issues and pull requests linked in the report, especially #1108, PR #30679, and PR #24711. Reproduce the attention-meter drop on the specified 2021 KIA e-Niro variant and determine whether opendbc exposes a supported command that affects the car's driver-attention state; done requires a confirmed cause and a defined, tested change if one is possible.

Written by the indexing model from the issue text.