Follow-up to #3591, driven by the design panel's round-2 verdict.

Scores

Both panels are deliberately grumpy (10 = ship-grade, 8 = only polish left).
Every round re-rates freshly generated screenshots; no score is carried over.

What changed, and why

Grouping. "When does it update" was split across the full card width with
the manual trigger wedged between the halves, which made "Automatic updates"
read as a caption for the button. The row now presents two groups: state and
its remedy lead (freshness word + trigger), and the switch keeps its own
readout (countdown + Skip) on the trailing rail, separated by a real step6
gap rather than leftover spaceBetween slack. Stacked, the schedule sits
tight against the switch it belongs to with the larger gap between the two
questions — evenly spaced rows let the countdown read as the manual run's ETA.

Emphasis. The trigger was outlined, which is literally the same button
as Confirm all (proposals_section_part.dart:187) — the card's hero action.
It drops to the design system's text tier, so exactly one outlined action
remains and it belongs to the proposals.

Truncation. Routes are structured, so they shed whole segments instead of
characters: Qwen 3.5 Plus · Alibaba · via Melious.ai → Qwen 3.5 Plus · Melious.ai → Qwen 3.5 Plus. An ellipsis ate the serving provider — the one
fact the row exists to disclose — and left the connective "via" behind
(This rep… GLM 5.2 · Z.ai · via OpenRou…). The fixed "This report" label
never gives ground; only its route does.

Alignment. DesignSystemButton pays its own step3 content inset, so a
button box on the leading edge put its glyph 8px inside it — invisible until
the row stacked, then a broken column, and a direct contradiction of the
footer's own docstring. The row negates that inset, and a test asserts the
column at 320px in German rather than a comment claiming it. The switch also
moves to the trailing edge of its 48px slot: centring a 40px track left it
short of the rail the TTS button and proposal cards terminate on.

Also: the countdown joins the freshness word in the state register, so it
stops reading as the same class of information as the static model route; the
attribution line's vertical padding is top-only, so the card's bottom margin
no longer depends on whether provenance happens to differ; the model row's
tooltip names its action instead of repeating a visible label; and the
schedule says nothing while a run is in flight, where "Updates when this task
changes" beside a spinner described a state the card was not in.

Copy. "Skip" → "Skip once" — the persona panel could not tell whether it
skipped one run or disabled automation. Plus a new
taskAgentChangeSetupTooltip. Both hand-authored across all 11 locales.

One recommendation I did not follow

The panel wants aiCard.footerWash changed at the token source (it argues a
black multiply cannot separate a band on a near-black card). That is a Figma
-exported token consumed by generated code, so it is not mine to edit —
flagging it rather than working around it in the widget.

Testing

Analyzer clean (zero warnings, zero infos), formatter clean, 2212 tests
green across test/features/agents/ui/ and test/features/tasks/ui/. New
coverage for the segment ladder, the stacked leading column, and the
running-state schedule suppression.

Version bumped to 0.9.1070+4256 at the maintainer's request.

Summary by CodeRabbit

• New Features

  • AI summary update controls are clearer, with separate action groups and countdowns aligned to their related settings.
  • Scheduled updates can now be canceled with a clearly labeled “Skip once” action.
  • AI model and provider details adapt to narrow spaces by removing complete segments instead of cutting words short.
  • Added a tooltip for changing the AI setup, with expanded localization coverage.

• Bug Fixes

  • Improved row alignment across screen sizes and languages.
  • Preserved consistent card spacing when report attribution is absent.
  • Hid scheduled-update details while an update is running.