ui: Add Thinking mode toggle with reasoning effort levels + improvements for Chat Form Add Action UI

[allozaur]

Overview

• (wip) Adds Thinking capability detection by doing heuristic checks for thinking/reasoning tags in chat template
• Adds Thinking controls with reasoning effort levels (Off/Low/Medium/High/Max)
• Redesigns the "Add" button panel with collapsible sections and nested submenus
• Thinking status icon next to the model name in Models Selector

Closes #23444
Closes #18405
Closes #22895
Closes #20557
Closes #20001
Supersedes #21657

Demo

Nagranie.z.Ekranu_converted.mp4

Additional information

Storage & Settings

• New localStorage keys: thinkingEnabledDefault, reasoningEffortDefault
• New settings: ENABLE_THINKING, SHOW_MODEL_QUANTIZATION, SHOW_MODEL_TAGS
• Settings registry entries for display and developer sections

Backend API Integration (chat.service.ts)

  // Adds to request body:
  chat_template_kwargs: { enable_thinking: boolean }
  thinking_budget_tokens: number  // from effort level mapping

Database Schema Changes

New optional fields on DatabaseConversation:

• thinkingEnabled?: boolean
• reasoningEffort?: ReasoningEffort

UX Redesign of "Add" Panel

Desktop (Dropdown):

• Reorganized into nested submenus
• "Add files" grouped under a submenu
• New "Thinking" submenu with effort levels
• Better MCP tool and server toggles

Mobile (Sheet):

• Collapsible sections: Files, MCP Servers, Tools, Thinking
• Toggle switches instead of checkboxes
• Visual feedback with enabled tool/server counters

Bug Fixes

• Fixed PDF Files duplicate entry (Misc. UX bug: WebUI, simplify the logic of the "Tools" submenu and deduplicate "PDF Files" #22895)
• Fixed tool selection being "forgotten" (using toggleGroupByLabel to avoid stale references)
• Removed duplicate reasoning_effort commented-out code from common/arg.cpp
• Fixed E2E test to wait for greeting h1 to be visible

Data Flow: Thinking Toggle → API Request

  User selects thinking setting (On/Off + Effort Level)
           ↓
  conversationsStore.setThinkingEnabled() / setReasoningEffort()
           ↓
  Per-conversation stored in DatabaseConversation
           ↓
  chat.svelte.ts → ChatService.sendMessage()
           ↓
  Sets: chat_template_kwargs.enable_thinking + thinking_budget_tokens
           ↓
  Server receives: POST /chat/completions

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, refactoring

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[coder543]

I have wanted a thinking toggle here for a long time, I'm just not sure how it can automatically detect the right kind of toggle to show. Some models only support one mode, some models support turning thinking on/off, some models support different reasoning efforts: low, medium, high (like GPT-OSS). How does this PR handle that?

[allozaur]

I have wanted a thinking toggle here for a long time, I'm just not sure how it can automatically detect the right kind of toggle to show. Some models only support one mode, some models support turning thinking on/off, some models support different reasoning efforts: low, medium, high (like GPT-OSS). How does this PR handle that?

my assumption was to release this as is, in an experimental form as most popular models support reasoning and follow up with further implementation on the server-side with some kind of support for detecting the thinking support and passing it to /props endpoint.

[engrtipusultan]

Thank you for this. I requested it some time back. #18405 I think many people will be happy.

Can you kindly confirm what will be required on server end for it to work as of now. If server is either started with reasoning on or off.

[allozaur]

Thank you for this. I requested it some time back. #18405 I think many people will be happy.

Can you kindly confirm what will be required on server end for it to work as of now. If server is either started with reasoning on or off.

Yes, actually I might devote a bit more time to this PR to also address #18405

[allozaur]

Okay, so after seeing multiple feature requests I've updated the description of the PR and changed it back to draft as I'd like to do a bit more in-depth implementation of Thinking controls feature from the start.

I've created a parent issue ##23444

[shredderique]

Hi, despite being late to the party....
I've vibecoded a simple toggle button into svelte-ui.
I suggest to have an easily accessible toggle button in favor or in addition to the fine-grained settings. This avoids complicated navigation through the menu.
Our customers love it so far.

Can't wait to see thinking mode toggle in main one way or another!

[allozaur]

@shredderique here's a refined toggle with dropdown choice :)

[allozaur]

@ggerganov @julien-c please review/test and lemme know :)

[allozaur]

Last tweaks

[ggerganov]

Very useful.

Note that this is compatible with the CLI argument --reasoning-budget-message. For example:

llama serve ... --reasoning-budget-message " ... I am thinking too long. Will proceed with the final answer now."

[julien-c]

just tried it, very cool!