Thinking Level

Este conteúdo não está disponível em sua língua ainda.

The thinking_level setting is a provider-scoped preference (auto, none, low, medium, high) configured in /config > Models > Text Samplers & Parameters or /personal config. It translates abstract reasoning effort into provider-specific request parameters at runtime, leaving providers without request-side controls unaffected.

Stored in saved_provider_configs.thinking_level and user_saved_provider_configs.thinking_level, the setting defaults to auto and restores automatically when switching active providers.

Shared semantics and overrides

TomoriBot interprets thinking levels before converting them to vendor fields:

Level	Intended behavior
`auto`	Follows the provider or model default behavior.
`none`	Disables thinking if supported; otherwise selects the provider’s lowest safe setting.
`low`	Requests light reasoning effort.
`medium`	Requests balanced reasoning effort.
`high`	Requests maximum available reasoning effort.

Dynamic request overrides

When a turn sets forceReason = true (such as invoking /respond with reasoning enabled), TomoriBot upgrades an effective level of auto or none to high for that individual generation request. The stored configuration row remains unmodified.

Numeric budget defaults

For providers accepting numeric token budgets rather than named effort strings, TomoriBot maps non-empty levels using defaults from src/utils/provider/thinkingControl.ts:

low: 1,024 tokens (DEFAULT_LOW_BUDGET_TOKENS)
medium: 4,096 tokens (DEFAULT_MEDIUM_BUDGET_TOKENS)
high: 8,192 tokens (DEFAULT_HIGH_BUDGET_TOKENS)

Provider mappings and runtime constraints

Google / Vertex

Gemini behavior branches by model family:

Gemini 2.5: Maps to numeric thinkingBudget. auto sends -1. none sends 0 for Flash models and clamps to 128 for Pro models (which cannot disable thinking). Flash-Lite clamps upward to 512.
Gemini 3 / 3.1: Maps to enum thinkingLevel (MINIMAL, LOW, MEDIUM, HIGH). auto omits the configuration. none and minimal map to MINIMAL only for Flash releases known to accept it (an allow-list in thinkingControl.ts) and to LOW everywhere else, because Pro models and Gemini 3.7 Flash onward reject MINIMAL while every Gemini 3 model accepts LOW.
Assistant prefill interaction: Gemini suppresses internal thought tags when continuing assistant prefills, emitting reasoning tokens directly into visible response text. The prefill resolver (resolvePrefillBlocker in src/utils/chat/assistantPrefill.ts) permits server prefills on Gemini only when thinking resolves to thinkingBudget: 0 or thinkingLevel: MINIMAL.

Anthropic

Claude models listed in the profile table in thinkingControl.ts map to adaptive thinking. Each family rejects another family’s off switch with HTTP 400, so an unlisted model receives no thinking fields until its profile is added:

auto: thinking: { type: "adaptive" }
low / medium / high: thinking: { type: "adaptive" } paired with output_config: { effort }
none: thinking: { type: "disabled" } on most models. Sonnet 5.5 uses { type: "between_tools" } instead, and Opus 5.5 and Fable cannot stop thinking, so they get adaptive thinking at effort: "low".
temperature and top_p are omitted whenever thinking is on. Opus 4.7 and later, Sonnet 5 and later, and Haiku 5.5 reject them in every mode, so those models never receive them.

OpenRouter

Maps directly to OpenRouter’s reasoning effort parameter:

auto: Omits reasoning.
none: reasoning: { effort: "none" }.
low / medium / high: reasoning: { effort }.

DeepSeek

Thinking toggle: V4 models (deepseek-flash and its aliased codenames, deepseek-v4-pro) think at high effort when the switch is omitted, so auto keeps that default and none sends thinking: { type: "disabled" }. low, medium, and high send thinking: { type: "enabled" } with a matching reasoning_effort (DeepSeek folds medium into high). The retired deepseek-chat and deepseek-reasoner codenames are sent as deepseek-flash and follow the same mapping.
Tool replay constraint: When thinking is enabled, DeepSeek returns HTTP 400 if a replayed assistant tool-call turn omits reasoning_content. TomoriBot enforces the presence of this key (requiresReasoningContentReplay), supplying an empty string fallback ("") when no reasoning was captured.
Degradation ladder protection: Parameter degradation lists thinking in mandatoryBodyKeys for DeepSeek, preventing the adapter from dropping thinking while retaining tools (which would produce tool responses that cannot be replayed).

Z.ai / Z.ai Coding

Maps to Z.ai’s thinking flag:

auto: Omits thinking.
none: thinking: { type: "disabled" }.
low / medium / high: thinking: { type: "enabled" }.
Temperature and penalty parameters are omitted when thinking is active.

Custom endpoints

Ollama: Automatically detected by URL or port 11434, mapping to reasoning_effort ("none", "low", "medium", "high").
Other local servers (vLLM, llama.cpp, KoboldCPP): Do not receive speculative request fields. Thinking on these engines is managed via startup arguments or chat template flags to avoid request validation errors.

NVIDIA NIM

NIM serves diverse model families behind one interface. Because different architectures read different parameters, non-auto levels send multi-switch payloads:

chat_template_kwargs: { enable_thinking: boolean, thinking: boolean }
reasoning_effort: "low" | "medium" | "high"
none sends reasoning_effort: "low" rather than "none" because NIM rejects "none" with HTTP 400 on gpt-oss and Llama.

NovelAI

The active profile sets the default: GLM 4.6 enables thinking for auto, while Xialong disables it. none disables thinking on both; low, medium, and high enable it without distinct effort budgets. The per-request forceReason override also applies.

GLM represents disabled thinking with /nothink and enabled thinking with <think></think> in its chat template. Xialong appends /nothink to user turns and closes its assistant thinking block when thinking is disabled; when enabled, it opens <think> for generation. Kayra and Erato use story templates and ignore the saved thinking level. No NovelAI text model calls tools.

Response parsing and privacy

Response parsing: For models outputting thinking tokens (such as Gemma 4 on KoboldCPP), GemmaThinkingParser strips thought blocks from delta.content and routes them to thoughts before tool parsing runs.
Thought log suppression: Thought logs are suppressed in channels governed by channel rules.

Source pointers

src/constants/thinkingLevels.ts: Thinking level values and localization keys.
src/utils/provider/thinkingControl.ts: Provider mapping implementations (buildGoogleThinkingConfig, buildAnthropicThinkingRequest).
src/utils/chat/assistantPrefill.ts: Assistant prefill gating against thinking states (resolvePrefillBlocker).
src/providers/deepseek/deepseekStreamAdapter.ts: replay policy and mandatory thinking field.
src/providers/openaiCompatible/openaiCompatibleMessageBuilder.ts: reasoning_content history replay.