Reasoning effort controls

[shatfield4]

Pull Request Type

• ✨ feat (New feature)
• 🐛 fix (Bug fix)
• ♻️ refactor (Code refactoring without changing behavior)
• 💄 style (UI style changes)
• 🔨 chore (Build, CI, maintenance)
• 📝 docs (Documentation updates)

Relevant Issues

None

Description

Adds a reasoning effort picker (brain icon) to the chat prompt input.

Scope: per chat session, per browser

• The choice is stored in the browser's local storage, keyed by workspace and thread (or the workspace's default chat). It is sent with each chat request and is never saved to the workspace, the thread, or the chat history. In multi-user mode, one user's choice never changes another user's chats, and a choice on one thread does not carry to other threads.
• The home page's picker keeps its choice under its own draft session until the first message creates a thread. The choice then moves to that thread.
• "Provider default" clears the choice and sends no reasoning params.
• There is no workspace-level or system-wide reasoning setting.

Only verified levels are sent
Before every request, the stored level is checked against the model's current capabilities. A level the model does not list is dropped, so a value left over from a model switch can't fail the chat. A failed capability lookup also sends nothing, and is not cached so the next message retries. The picker only renders when the model lists levels.

Model router
Workspaces on the model router show no picker and never apply a stored effort, because the routed model changes per message.

Where each provider's levels come from

Provider	Supported levels	Request format
Anthropic	Models API capabilities.effort	output_config.effort
OpenAI	models.dev reasoning_options	reasoning.effort, reasoning.summary (summary dropped automatically for unverified orgs)
Gemini	models.dev reasoning_options	thinking_config: a level for 3.x models, Google's documented budgets for 2.5
DeepSeek	models.dev reasoning_options	thinking.type, reasoning_effort
Ollama	/api/show thinking.values, falling back to the model family on older Ollama versions	think
LM Studio	/api/v1/models reasoning.allowed_options	reasoning_effort ("on" is sent as medium)
Lemonade	reasoning model label, levels by model family	llama.cpp chat_template_kwargs
OpenRouter	models.dev reasoning_options (effort levels and on/off; models with only a token budget, or mandatory thinking with no levels, get no picker)	reasoning.effort / reasoning.enabled, replacing the legacy include_reasoning flag when an effort is set

Reasoning models that support it now show their reasoning (OpenAI summaries, Gemini thoughts) as a thinking block.

Paths
The effort applies to workspace and thread chats, including agent sessions, and can change mid agent session. The embed, developer API, OpenAI-compatible, and Telegram paths send no reasoning params, so the provider default applies.

Pricing cache
models.dev reasoning options are cached next to the existing pricing cache. A pricing cache from before this change keeps working offline, and the reasoning options are filled in on the next successful refresh.

Developer Validations

• I ran yarn lint from the root of the repo & committed changes
• Relevant documentation has been updated (if applicable)
• I have tested my code functionality
• Docker build succeeds locally

[shatfield4]

Tested every provider end to end from the chat UI. For each one: set the effort with the brain icon in the chat input, sent a message, and checked the header label, the metrics line under the reply, and the workspace_chats row. To confirm the effort actually reached the provider, I ran the dev server with a temporary fetch tap that logged the reasoning fields of every outbound request body (not committed, only used for testing).

Provider / model	Effort set in UI	Request body sent to provider	Reply
OpenAI gpt-5.1	high / off	reasoning: { effort: "high" } / { effort: "none" }	42 vs 13 output tokens, correct
Anthropic claude-sonnet-5	max	output_config: { effort: "max" }	correct
Gemini gemini-3.5-flash	minimal	reasoning_effort: "minimal"	correct
DeepSeek deepseek-flash	on	thinking: { type: "enabled" }	thoughts block shown, correct
Ollama qwen3:8b	off	think: false	no thinking, correct
LM Studio grayline-qwen3-8b	off	reasoning_effort: "none"	no thinking, correct
Lemonade Qwen3-4B-GGUF	on	chat_template_kwargs: { enable_thinking: true }	thoughts block shown, correct
Lemonade gpt-oss-20b-GGUF	low	chat_template_kwargs: { reasoning_effort: "low" }	short thinking, correct

In every case the header, metrics line, and DB row all show the same effort that was sent.

Provider-side confirmation that the levels are honored (direct SDK calls, same prompt):

• OpenAI gpt-5.1: reasoning tokens 0 (off) → 6 (low) → 197 (high)
• Anthropic claude-sonnet-5: output tokens 42 (low) → 280 (max)
• Gemini 3.5-flash: thinking tokens 0 (minimal) → 1429 (high); 2.5-pro 1327 (low) → 2792 (high)
• DeepSeek: reasoning_content absent (off) / present (on)
• Ollama qwen3:8b: no thinking (off) / 4693 chars (on)
• LM Studio: 0 reasoning tokens (off) / 1198 (default)
• Lemonade gpt-oss-20b: 100 thinking chars (low) → 1694 (high); Qwen3-4B 0 (off) / 5612 (on)

Also verified:

• system-wide default set on the LLM preference page for all 7 providers, picked up by workspaces with no override
• workspace override beats the global default; clearing it falls back
• a stored effort the current model does not support (eg: low on claude-sonnet-4-5, high on gpt-4.1-mini, high on llama3.2) is dropped and the chat still succeeds
• agent chats for all 7 providers

[timothycarambat]

fixes applied below:

1. Model-router workspaces couldn't undo a stored effort
GET /workspace/:slug/llm-capabilities built the connector with getLLMProvider, which throws for anythingllm-router. The picker was hidden, but the chat path still applied the stored effort to whichever routed model listed it, so a thread with "high" set before a switch to the router kept sending it with no way to clear it. There is now a usesModelRouter(workspace) check in the capabilities endpoint, streamChatWithWorkspace (which also covers the effort handed to agents), and AgentHandler#reasoningEffortForRoute (covers mid-session updates). Routed workspaces show no picker and never apply an effort.

2. An offline upgrade wiped the pricing cache
#loadFromDisk read the pricing and reasoning files in one try. After an upgrade the reasoning file doesn't exist, so pricing was nulled too, and air-gapped installs lost cost data. The reasoning file is now read on its own. Pricing stays loaded, and the etag is dropped so the next refresh is a full GET. A new test fails on the previous code.

3. Ollama levels come from Ollama's own API
/api/show returns thinking.values per model, so that list is used now. The gpt-oss name check is kept as a fallback for Ollama versions that don't return it.

4. Sources
Each provider's case in reasoningParams now links to its docs. Two are still unconfirmed:

• LM Studio's per-model allowed_options is documented, but sending "on" as reasoning_effort: "medium" on the OpenAI-compatible endpoint isn't.
• Lemonade's chat_template_kwargs depends on the llama.cpp backend.

Both fail safely, since "Provider default" clears the choice.

5. Smaller fixes

• The home page now keeps its choice under its own draft session, so it no longer writes to the workspace's default chat session.
• Moved getProviderForConfig's JSDoc back above it.
• Fixed the "system default" wording in #updateReasoningEffort.