Summary

Redesign of the agents-chatbox model dropdown to match the new layout: a searchable list grouped by provider, a clear "selected" treatment with a trailing checkmark, and an Effort row at the bottom of the panel. The trigger is a subtle inline pill with a chevron that fits next to the existing chat-footer controls.

This is a draft opened on @tracyjohnsonux's behalf by the Coder Agents bot for design review. Backend was intentionally not touched.

Visual reference

Screenshots of the implementation will be added as a follow-up comment on this PR.

What changed (frontend only)

• ModelSelector.tsx: replaced Radix Select with Popover + Command (cmdk).
  • Search input at the top of the panel with a strict, substring-based filter so typing a model name hides unrelated rows cleanly.
  • Provider name renders as a section heading (e.g. "Anthropic", "OpenAI").
  • Each row shows the display name with a compact context-window chip, e.g. Opus 4.7 (1M).
  • The selected row gets a darker background plus a trailing checkmark icon.
  • Trigger button restyled as a subtle inline pill. Mouse-focus ring suppressed, keyboard-focus ring preserved (same intent as before the rewrite).
  • Tooltip on each row preserved for screen-reader/visual users (full name + provider + context window).
  • aria-label on each row is the model display name, so existing getByRole("option", { name: ... }) lookups keep matching even though the visible label now includes the context chip.
• Effort row at the bottom of the panel renders the admin-configured reasoning effort for the selected model as a read-only slider plus value chip. Hidden when the selected model has no effort configured. Important: see the open question below before merging.
• ModelSelectorOption gains an optional effort?: string field.
• modelOptions.ts: getModelOptionsFromConfigs now reads the reasoning effort from each provider's options block (provider_options.openai.reasoning_effort, provider_options.openaicompat.reasoning_effort, provider_options.anthropic.effort) and passes it through so the Effort row shows real configuration values.
• ModelSelector.stories.tsx: rewritten around the new design. New stories cover OpenSingleProvider, OpenMultipleProviders, OpenWithoutEffort, SearchFiltersOptions, plus trigger/disabled/empty states.

Open question (decision needed before this can become real)

The design shows an interactive Effort slider with low / medium / high / xhigh values. Today, the backend stores reasoning effort only on the admin-managed ChatModelConfig: CreateChatRequest and the message-send payloads have no reasoning_effort field. So a user-controllable per-chat effort slider needs a backend change.

This PR ships the row as a read-only display of the admin-configured value so the layout is real and not faking interactivity. Three options for what comes next:

1. Add backend support for a per-chat effort override (new optional field on CreateChatRequest and the chat-send websocket payload, plumbed through ChatModelConfig.model_config.provider_options.{openai,openaicompat,anthropic} at request time). Then make the slider here interactive.
2. Keep it read-only as a quick way to see the model's configured effort from the chatbox, and move "change effort" into the existing admin model-config UI.
3. Drop the Effort row from this PR and ship just the search / provider headings / checkmark redesign.

Need your call here, @tracyjohnsonux.

Tests

• pnpm exec tsc --noEmit clean.
• pnpm exec biome check src/ clean.
• Unit tests (vitest --project unit) for AgentsPage: 1370 passed.
• Storybook interaction tests for AgentsPage: 1059 passed. (The single failing test Tool.stories.tsx > MCP Tool Completed reproduces on main without these changes; it is a pre-existing xterm flake.)
• make lint/emdash clean.

_{This pull request was prepared by the Coder Agents bot at @tracyjohnsonux's direction.}