ui : add the models manager

[allozaur]

Overview

Demo

Screen.Recording.2026-10-09.at.10.13.23.mp4

Screen.Recording.2026-10-09.at.10.15.33.mp4

Features

• New Manage models dialog: one table of every model the server can serve — one row per repo with expandable quants, sections (loaded → downloading → favorites → local → hidden), sortable columns (name/context/status, sleep-aware), search and context/capability/modality filters (only modalities present are offered).
• Row actions in a drawer: favorite, load/unload, delete from disk, hide/unhide, download pause/resume; downloads get their own section with a progress bar; a partly hidden repo splits its quants across the hidden and local blocks.
• Model configuration pane as a side drawer: full model info and chat template, plus Start a new chat (loading the model first) / Use in this chat and load controls; on a phone the model information dialog stands in for the pane.
• Reworked model selector (ModelsSelector): owns the shared picker state and renders the desktop dropdown and the phone drawer as equal surfaces — provider-aware grouping, favorites and loaded pinned on top (listed once), hidden models filtered out, windowed lists with show more, provider mark + org avatar in the trigger; falls back to the manager on single-model servers, and a phone with one model still gets the drawer.
• Reasoning effort becomes its own control beside the selector with a fully transparent trigger (no background, no hover fill, any state); MCP servers move from the add menu to the sidebar rail.
• Mobile experience: add actions in a drawer (sharing its labels with the desktop dropdown via constants), floating model action bar, context gauge in a drawer.
• New settings: Hugging Face Hub metadata (off by default — no fetches and no avatars when off), group-models-by-family, capabilities-in-selector, org-name display (renamed + migrated).

Architecture / logic

• New shared data layer: modelsStore composes list, selection, loading, recents (manual picks only), favorites, hidden and group-open state, all persisted to localStorage; status store feeds download progress (and the delete-from-disk action resolves through its cancel endpoint).
• HF Hub details cached per repo with request dedup (404s cached, failures retried; failed base-model lookups not cached), warmed for a capped slice (12) and fetched per row only near the viewport; local file paths never query the Hub; avatars load in CORS mode with a per-org retry cooldown, search icons ride the server's CORS proxy.
• Selector state derived in one hook, with the chat-form model concerns (conversation sync, modality tracking) extracted into useChatFormModel; both surfaces and the assistant-message badge share it.
• Shared model components extracted (avatar, capabilities, context, load control, row actions) reused by both surfaces; new UI primitives (drawer/sheet via vaul-svelte, toggles, collapsible sections, grouped list, near-viewport helper).
• ModelOption enriched with contextLength, draftSidecars (from --model-draft args) and chat-template-based capability fallback; family grouping, list windowing and manager helpers (groupModelFamilies, groupModelQuants, splitHiddenQuants, hasActiveFilters, statusRank) live in tested utils.
• Settings gained a versioned migration for the renamed org-name key; tools/ui/README.md documents the new structure and the drawer dependency.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, pi:llama.cpp/DeepSeek-V4.1-Flash / pi:zai-org/GLM-5.3-Flash

[allozaur]

Besides the comments for the files, there's one regression vs what we used to have some time ago — i can't see the draft sidecar badges in the model id in models selector and models manager