fix(providers): resolve reasoning through the catalog; add appearance preferences

[Zongwei9888]

Description

Addresses #153, #151, #152 and #155.

#153 reports that DeepSeek's thinking selector has no levels. That turned out to be one symptom of a resolver that had rotted, and chasing it surfaced two more gaps. The Desktop half of the other three reports is a separate, self-contained change in the same PR.

#153 — reasoning capabilities resolved through the model catalog

infer_reasoning_capabilities() was a chain of substring checks kept apart from the model catalog. It recognised claude-sonnet-4-6 while the catalog had moved on to claude-sonnet-5, so on main today all of these advertise no reasoning at all:

claude-sonnet-5   claude-opus-4-8   claude-sonnet-4-5   claude-haiku-4-5
deepseek-r1       deepseek-reasoner  deepseek-chat
qwen3-max         grok-4

The Desktop picker builds its options from model.reasoning, so an empty capability renders a control with nothing in it — exactly what #153 describes. The provider side was never at fault: deepseek already carried thinking_style="thinking_type", so the wire format had been wired all along.

Both catalog.py and model_compat.py open by naming model_id.includes(...) as the anti-pattern to avoid, and catalog.py already resolves every other per-model fact — context window, output cap, pricing — through a normalize → snapshot → seed → family → default cascade. Reasoning now joins it as a ModelInfo field. profiles.py had been calling resolve_model_info() and the name chain on adjacent lines for the same model.

Consequences beyond the report:

• An unseen deepseek-r2 or gpt-5.9 inherits its family row instead of silently losing its controls. Shipping a model is a table row, not a code edit.
• Family fallback deliberately advertises less than an exact seed row: claude-sonnet-4-<date> keeps the effort ladder but not summarised thinking, which it predates. Claiming it would route raw trace text through the summary channel.
• Dotted aliases (claude-sonnet-4.6) are listed rather than normalised away — the dot is load-bearing elsewhere in the table, where gpt-5.4 and kimi-k2.5 are distinct models.

Two gaps this exposed

GLM and MiniMax were resolving to the 128K default — a 200K model trimmed to 128K, and a thinking-capable one advertising nothing. Limits taken from the vendors, not guessed: GLM-4.6 is 200K/128K (docs.z.ai); the MiniMax M2 line is 204,800 with M3 at 1M (platform.minimax.io).

Zhipu never declared a thinking_style. GLM enables reasoning with thinking: {"type": "enabled"} — the same body DeepSeek takes, and the shape _THINKING_STYLE_BUILDERS["thinking_type"] already builds. Every GLM request had been going out with reasoning silently omitted whatever effort the user picked. One field; the mechanism was already there.

What this does not fix

DeepSeek still shows Auto/Off rather than low/medium/high, because its API is a binary switch — there are no named levels to expose, and inventing three would put dead options in the picker. deepseek-reasoner has no user-facing choice at all: reasoning is intrinsic and cannot be turned off.

The dramatic change is elsewhere: Claude's current line-up goes from a single "Auto" back to the full six-option ladder.

#151 / #152 / #155 — appearance preferences

One mechanism, not three. desktop/src/app/appearance.ts holds the settings as a table; each row owns its label, range, sanitiser and CSS custom property, and the settings UI renders from the same table.

• Width ([Feature Request]: Add adjustable conversation pane width (percentage) in Desktop GUI #151) — .conversation, its composer and the goal rail each hardcoded min(820px, 100%). They now read one --conversation-width. 100% restores the built-in cap, so the default is byte-identical to today.
• Theme ([Feature Request]: Add theme support (dark mode + VS Code theme import) in Desktop GUI #152) — the report says there is no dark theme. There is; tokens.css has carried a full dark palette behind prefers-color-scheme all along. What was missing is overriding the OS. CSS cannot share a declaration block between a media query and a selector, so the dark palette is now written twice — and styles/tokens.test.ts fails if the copies ever disagree, because two tables drifting apart is the bug this PR's first half exists to fix. VS Code theme import is not here.
• Typography ([Feature Request]: Add font size & font family options (English + CJK) in Desktop GUI #155) — a free-text family list rather than a dropdown: a curated menu is a hardcoded guess about fonts the user's machine may not have. The value is a prefix of the built-in stack, so anything missing falls through.

Testing

• Python: 1189 passed, 3 skipped
• Desktop: 138 passed across 23 files; tsc --noEmit, eslint and npm run build all clean
• Every new suite verified in reverse — reverting the fix fails the tests (13/23 for the catalog move, 4/6 for the Zhipu wire, and the palette-drift test catches a single changed value)

Not verified: the Desktop changes have not been seen running in Tauri. Unit tests and the build cover syntax and logic, not appearance. Worth a look at the dark toggle and a narrow conversation width before release.

tests/application/test_automation_goal_runs.py::test_legacy_unreserved_turn_is_never_adopted_as_automation_initial_turn fails intermittently — measured at 2/20 on clean main, unrelated to this work, and deserving its own issue.

Known gaps, deliberately left

thinking_style is applied per provider without consulting the model, so gpt-4o routed through a thinking-capable gateway still receives a thinking body. Pre-existing, but this PR widens it by giving Zhipu a style. Both halves of the information now exist in one place — the catalog knows the model, the spec knows the wire — so intersecting them is a focused follow-up rather than more surface here.