feat(desktop): declare one thinking level across a relay's models at once

[Joob1n]

Summary

设置 → 模型 → a custom relay → 能力 declared thinking levels one model at a time. A relay normally fronts one model family sharing a reasoning_effort vocabulary, so high across eight enabled models was eight menus opened and eight ticks — for one fact about the relay.

The section now opens with one control that writes into every enabled model's row.

Per level, not per level set. Ticking high adds it everywhere; unticking removes it everywhere; every other level on every row is left exactly as the user left it. Replacing the whole set would have been one fewer concept and would silently discard what a row already declared — the destructive reading of a control whose only job is to save clicks.

The box ticks only at full coverage. A level 3 of 5 models declare shows unticked with 3/5 个模型 beside it, so the next click means "give it to everyone" rather than "take it from the rows that have it". The count is the only thing separating partial coverage from none, since a checkbox has no third state here.

The control shares the 思考档位 row's relay gate — reasoning_effort is a wire field only the OpenAI-compatible relays accept — and appears only from two rows up, where there is repetition to save. Edits land in the same local draft the rows use and commit through the same 保存能力声明. Vision, context window, and fast mode stay per-row: those are per-model facts.

Fixes #3447

Review focus

The bulk menu offers only the declarable vocabulary. off is a disable wire (reasoning_effort: 'none'), not an intensity tier, and no generic relay is presumed to honor it. It cannot reach a draft anyway — the seed sanitizes through normalizeRelayModelProfiles — but spreading one across a whole table is the specific thing worth not being able to do by accident.

One state update, not a loop of per-model setters. A bulk tick is a single gesture; committing it in N steps would let a re-render land mid-way and paint a half-applied table. The rule for what an emptied declaration collapses to now lives in one place (relayProfileWithThinkingLevels) so the row setter and the bulk path cannot drift on it.

Also in this PR: an accessibility fix the new story exposed

上下文窗口（tokens） was the one capability control without a model in its accessible name — vision, thinking, and fast mode all carry — ${modelId}. Every row therefore exposed an identically-named spinbutton, and the AX audit's ambiguous_actionable rule fires on it:

[product-settings-providers--relay-connection-detail] AX audit failed:
  ambiguous_actionable — role=spinbutton name=上下文窗口（tokens）
  scope=region:能力 > region:House Relay count=4

This is a defect on main, reachable by anyone with a relay and two or more enabled models — the ordinary case. It had no story: the only existing multi-model capability story is a non-relay whose rows are filtered down to one. Fixed by giving the field the same per-model name as its three siblings.

Verification

Behaviour, on the new story (product-settings-providers--relay-connection-detail, a relay with four enabled models, one of which already declares low + high):

step	result
open 批量设置思考档位	low 1/4 个模型, high 1/4 个模型, rest 全部未声明
tick high	high → 4/4 个模型, ticked; low still 1/4; rows read 已选择 2/1/1/1 个
保存能力声明	commits, button disarms; reopening reads high 4/4 from the saved table
untick high	high → 全部未声明 everywhere; deepseek-r2 drops to 已选择 1 个, keeping its low

The last row is the non-destructive claim: the model that had two levels kept the one the bulk edit was not about.

Checks run locally:

• @maka/desktop typecheck (preload + main + renderer + storybook) — clean
• biome lint apps/desktop/src apps/desktop/stories — 719 files, clean
• @maka/desktop suite — 1048/1048 pass, including 14 new
• build-storybook + smoke:storybook — 161 stories pass, AX audit included. It failed before the context-window fix, on exactly the story this PR adds.

Not run: the packaged-artifact and release checks, and the other workspaces' suites — nothing outside apps/desktop is touched.

AI use

Select exactly one:

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: Claude Opus 5 (Claude Code) — wrote the implementation, tests, and story; ran the checks above. Reviewed and directed by me. Generated-by trailer is on the commit.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — described under Summary above
• No

[Astro-Han]

[P2] Expose partial model coverage to assistive technology

Thanks for showing the partial n/total state visually—it makes the batch behavior much easier to understand. Because this item also supplies an aria-label, however, the screen reader does not receive that visible description and cannot distinguish "none of the models declare this level" from "some models already declare it."

Could we expose thinkingBulkCoverage(...) through aria-description or an aria-describedby relationship? A focused accessibility assertion for the partial state would be enough; the existing pure-function coverage does not need to be expanded further.

[Astro-Han]

Thanks for making the batch operation reuse the existing relay-profile draft and save path. The per-level add/remove semantics are clear, partial coverage is preserved, other thinking levels are not overwritten, and the focused tests cover the important state transitions and edge cases.

I'm approving this exact head with one non-blocking accessibility P2 below.

AI-assisted review disclosure: OpenAI Codex coordinated two independent exact-head review passes. I verified the form-state and persistence path, batch semantics, accessibility behavior, tests, exact head, and CI state, and I independently made the approval decision.

[Joob1n]

Addressed in bb06b8c99. This one was a real defect and I had not tested for it.

Read straight off the rendered story before the fix:

low      aria-label="批量设置思考档位 low"      aria-checked=false
minimal  aria-label="批量设置思考档位 minimal"  aria-checked=false

low is declared by one of the four models, minimal by none — and to a screen reader they were the same unchecked box with the same name. The n/total I added specifically to make partial coverage legible was the one part that did not survive, because the item's own aria-label replaces the name the visible description would have joined, and the component does not wire description to aria-describedby (checked: aria-describedby is null on every item).

The item now carries aria-description. Confirmed through CDP against the real Chromium accessibility tree, not just the DOM attribute:

批量设置思考档位 low      description="1/4 个模型"
批量设置思考档位 minimal  description="全部未声明"

The assertion. It lives in the story's own play — the storybook smoke already runs play before reading the AX tree, so this is checked on every CI run that builds the catalog, without adding an e2e that would need a configured relay. It opens the batch menu and pins both states:

await expect(partial).toHaveAttribute('aria-checked', 'false');
await expect(none).toHaveAttribute('aria-checked', 'false');
await expect(partial).toHaveAttribute('aria-description', '1/4 个模型');
await expect(none).toHaveAttribute('aria-description', '全部未声明');

Both unchecked is the point: coverage is the only thing separating them. Removing aria-description fails that story — verified by removing it:

✗ product-settings-providers--relay-connection-detail
  expect(element).toHaveAttribute("aria-description", "1/4 个模型")

Per your note, the pure-function coverage is unchanged.

Verification

• @maka/desktop suite: 1048/1048
• typecheck (preload + main + renderer + storybook): clean
• biome lint apps/desktop/src apps/desktop/stories: 719 files, clean
• build-storybook + smoke:storybook: 161 stories, including the new assertion