Skip to content

Voice-to-text dictation app with local (Nvidia Parakeet/Whisper) and cloud models (BYOK). Privacy-first and available cross-platform.

feat(reasoning): pick a reasoning effort per scope (#2694)

void48cb1001debf
diff --git a/CLAUDE.md b/CLAUDE.md
index 707af8152..472163d02 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -445,7 +445,7 @@ Non-secret env vars persisted to `.env` (via `saveAllKeysToEnvFile()`):
- Gemini 3.5 Flash (`gemini-3.5-flash`) - Latest fast, high-capability Gemini model
- Gemini 3.5 Flash Lite (`gemini-3.5-flash-lite`) - Fastest, most cost-effective 3.5 model
- Gemini 3.1 Pro (`gemini-3.1-pro-preview`) - Most capable Gemini model
- - Gemini 3.1 Flash Lite (`gemini-3.1-flash-lite`) - Frontier-class performance at low cost (no thinking support, so no `supportsThinking` flag)
+ - Gemini 3.1 Flash Lite (`gemini-3.1-flash-lite`) - Frontier-class performance at low cost. It thinks, and requests that aren't streamed pin it to `minimal` unless a reasoning effort is picked; it has no `supportsThinking` flag, so Settings shows the effort picker rather than the thinking toggle
- Gemini 3 Flash (`gemini-3-flash-preview`) - Ultra-fast, high-capability next-gen model
- Gemini 2.5 Flash Lite (`gemini-2.5-flash-lite`) - Lowest latency and cost (retired for new API keys; kept for existing ones)
- Gemma 4 (`gemma-4-31b-it`, `gemma-4-26b-a4b-it`) - Google's open Gemma 4 models served through the Gemini API
diff --git a/src/components/ControlPanel.tsx b/src/components/ControlPanel.tsx
index a543941b2..63cb1bb77 100644
--- a/src/components/ControlPanel.tsx
+++ b/src/components/ControlPanel.tsx
@@ -767,6 +767,7 @@ export default function ControlPanel({ initialSettingsSection }: ControlPanelPro
const agentName = getAgentName();
const reasonedText = await ReasoningService.processText(rawText, model, agentName, {
disableThinking: getSettings().cleanupDisableThinking,
+ reasoningEffort: getSettings().cleanupReasoningEffort,
requireCompleteOutput: true,
});
if (hasTextContent(reasonedText) && reasonedText !== rawText) {
diff --git a/src/components/chat/useChatStreaming.ts b/src/components/chat/useChatStreaming.ts
index a5b312881..a3e3d6989 100644
--- a/src/components/chat/useChatStreaming.ts
+++ b/src/components/chat/useChatStreaming.ts
@@ -692,6 +692,7 @@ export function useChatStreaming({
customApiKey:
isCustomAgent || isLanAgent ? llmConfig.customApiKey || undefined : undefined,
disableThinking: llmConfig.disableThinking,
+ reasoningEffort: llmConfig.reasoningEffort,
},
aiTools
);
diff --git a/src/components/settings/InferenceConfigEditor.tsx b/src/components/settings/InferenceConfigEditor.tsx
index 239df3444..261c94924 100644
--- a/src/components/settings/InferenceConfigEditor.tsx
+++ b/src/components/settings/InferenceConfigEditor.tsx
@@ -18,6 +18,7 @@ import ReasoningModelSelector from "../ReasoningModelSelector";
import EnterpriseSection from "../EnterpriseSection";
import OpenAICompatiblePanel from "../OpenAICompatiblePanel";
import { Toggle } from "../ui/toggle";
+import { Select, SelectContent, SelectItem, SelectTrigger, SelectValue } from "../ui/select";
import type { InferenceMode } from "../../types/electron";
import {
INFERENCE_SCOPES,
@@ -30,12 +31,17 @@ import {
getLocalModel,
enterpriseProviderName,
} from "../../models/ModelRegistry";
+import { AUTO_REASONING_EFFORT, isReasoningEffortSupported } from "../../models/reasoningEffort";
+import { selectableReasoningEffort } from "../../services/ai/chatRequestBody";
import { useManagedScopeResolution } from "../../stores/enterpriseIdentityStore";
import { requestSignIn } from "../../utils/requestSignIn";
import TestConnectionButton from "../TestConnectionButton";
import { getEnterpriseCallSettings } from "../../services/ai/enterpriseSettings";
import { Button } from "../ui/button";
+// Radix Select can't hold an empty value, so Auto ("") gets a sentinel.
+const AUTO_REASONING_EFFORT_VALUE = "__auto__";
+
const MODE_LABEL_PREFIX: Record<InferenceScope, string> = {
dictationCleanup: "settingsPage.aiModels.modes",
noteFormatting: "settingsPage.aiModels.modes",
@@ -182,6 +188,11 @@ export default function InferenceConfigEditor({
!!getCloudModel(config.model)?.supportsThinking)) ||
(effectiveMode === "local" && !!getLocalModel(config.model)?.supportsThinking);
+ const effortCapability =
+ effectiveMode === "providers"
+ ? selectableReasoningEffort(config.provider, config.model, config, config.cloudBaseUrl)
+ : null;
+
if (managed.kind === "error") {
return (
<div className="rounded-lg border border-destructive/30 bg-destructive/5 p-3" role="alert">
@@ -308,6 +319,54 @@ export default function InferenceConfigEditor({
</div>
)}
+ {effortCapability && (
+ <div className="flex items-start justify-between gap-3 pt-1">
+ <div className="flex-1 min-w-0">
+ <h4 className="text-sm font-medium text-foreground">{t("reasoning.effort.title")}</h4>
+ <p className="text-xs text-muted-foreground">
+ {effortCapability.defaultValue
+ ? t("reasoning.effort.descriptionWithDefault", {
+ value: t(`reasoning.effort.options.${effortCapability.defaultValue}`),
+ })
+ : t("reasoning.effort.description")}
+ </p>
+ </div>
+ <Select
+ value={
+ isReasoningEffortSupported(effortCapability, config.reasoningEffort)
+ ? config.reasoningEffort
+ : AUTO_REASONING_EFFORT_VALUE
+ }
+ onValueChange={(value) =>
+ setField("reasoningEffort")(
+ value === AUTO_REASONING_EFFORT_VALUE ? AUTO_REASONING_EFFORT : value
+ )
+ }
+ >
+ <SelectTrigger className="h-7 w-44 shrink-0 text-xs rounded-lg px-2.5 [&>svg]:h-3 [&>svg]:w-3">
+ <SelectValue />
+ </SelectTrigger>
+ <SelectContent>
+ <SelectItem
+ value={AUTO_REASONING_EFFORT_VALUE}
+ className="text-xs py-1.5 ps-2.5 pe-7 rounded-md"
+ >
+ {t("reasoning.effort.auto")}
+ </SelectItem>
+ {effortCapability.options.map((effort) => (
+ <SelectItem
+ key={effort}
+ value={effort}
+ className="text-xs py-1.5 ps-2.5 pe-7 rounded-md"
+ >
+ {t(`reasoning.effort.options.${effort}`)}
+ </SelectItem>
+ ))}
+ </SelectContent>
+ </Select>
+ </div>
+ )}
+
{effectiveMode === "local" && (
<div className="flex items-start justify-between gap-3 pt-1">
<div className="flex-1 min-w-0">
diff --git a/src/components/ui/PromptStudio.tsx b/src/components/ui/PromptStudio.tsx
index 62339e043..3aaf8386f 100644
--- a/src/components/ui/PromptStudio.tsx
+++ b/src/components/ui/PromptStudio.tsx
@@ -272,6 +272,7 @@ export default function PromptStudio({ className = "", kind = "cleanup" }: Promp
const result = await ReasoningService.processText(testText, modelToUse, agentName, {
inferenceScope: "dictationCleanup",
disableThinking: effectiveSettings.cleanupDisableThinking,
+ reasoningEffort: effectiveSettings.cleanupReasoningEffort,
});
setTestResult(result);
} finally {
diff --git a/src/config/inferenceScopes.ts b/src/config/inferenceScopes.ts
index fbaacdcbd..fdd64af3f 100644
--- a/src/config/inferenceScopes.ts
+++ b/src/config/inferenceScopes.ts
@@ -10,6 +10,7 @@ export interface InferenceScopeStoreKeys {
remoteUrl?: keyof SettingsState;
customApiKey?: keyof SettingsState;
disableThinking?: keyof SettingsState;
+ reasoningEffort?: keyof SettingsState;
}
export interface InferenceScopeDefinition {
@@ -40,6 +41,7 @@ export const INFERENCE_SCOPES = {
remoteUrl: "cleanupRemoteUrl",
customApiKey: "cleanupCustomApiKey",
disableThinking: "cleanupDisableThinking",
+ reasoningEffort: "cleanupReasoningEffort",
},
},
dictationAgent: {
@@ -53,6 +55,7 @@ export const INFERENCE_SCOPES = {
remoteUrl: "dictationAgentRemoteUrl",
customApiKey: "dictationAgentCustomApiKey",
disableThinking: "dictationAgentDisableThinking",
+ reasoningEffort: "dictationAgentReasoningEffort",
},
},
// Optional override used only when a voice-agent request carries a screen
@@ -68,6 +71,7 @@ export const INFERENCE_SCOPES = {
cloudBaseUrl: "dictationAgentVisionCloudBaseUrl",
customApiKey: "dictationAgentVisionCustomApiKey",
disableThinking: "dictationAgentVisionDisableThinking",
+ reasoningEffort: "dictationAgentVisionReasoningEffort",
},
fallbackScope: "dictationAgent",
optional: true,
@@ -83,6 +87,7 @@ export const INFERENCE_SCOPES = {
remoteUrl: "noteFormattingRemoteUrl",
customApiKey: "noteFormattingCustomApiKey",
disableThinking: "noteFormattingDisableThinking",
+ reasoningEffort: "noteFormattingReasoningEffort",
},
fallbackScope: "dictationCleanup",
},
@@ -100,6 +105,7 @@ export const INFERENCE_SCOPES = {
remoteUrl: "chatAgentRemoteUrl",
customApiKey: "chatAgentCustomApiKey",
disableThinking: "chatAgentDisableThinking",
+ reasoningEffort: "chatAgentReasoningEffort",
},
},
dictationTranslation: {
@@ -113,6 +119,7 @@ export const INFERENCE_SCOPES = {
remoteUrl: "translationRemoteUrl",
customApiKey: "translationCustomApiKey",
disableThinking: "translationDisableThinking",
+ reasoningEffort: "translationReasoningEffort",
},
},
} as const satisfies Record<string, InferenceScopeDefinition>;
diff --git a/src/helpers/audioManager.js b/src/helpers/audioManager.js
index 230c3f477..a6e8726df 100644
--- a/src/helpers/audioManager.js
+++ b/src/helpers/audioManager.js
@@ -333,6 +333,7 @@ function resolveReasoningRoute(
const cleanupConfig = {
inferenceScope: /** @type {const} */ ("dictationCleanup"),
disableThinking: settings.cleanupDisableThinking,
+ reasoningEffort: settings.cleanupReasoningEffort,
temperature: cleanupTemperature,
requireCompleteOutput: true,
};
diff --git a/src/helpers/dictationAgentInference.js b/src/helpers/dictationAgentInference.js
index ea2ba0812..5cacac078 100644
--- a/src/helpers/dictationAgentInference.js
+++ b/src/helpers/dictationAgentInference.js
@@ -29,6 +29,7 @@ export function resolveDictationAgentInference(settings, { isCloudAgent = false
inferenceScope: /** @type {const} */ ("dictationAgent"),
provider: managed.provider,
disableThinking: settings.dictationAgentDisableThinking,
+ reasoningEffort: settings.dictationAgentReasoningEffort,
},
};
}
@@ -68,6 +69,7 @@ export function resolveDictationAgentInference(settings, { isCloudAgent = false
customApiKey:
isCustom || isSelfHosted ? settings.dictationAgentCustomApiKey || undefined : undefined,
disableThinking: settings.dictationAgentDisableThinking,
+ reasoningEffort: settings.dictationAgentReasoningEffort,
},
};
}
@@ -118,6 +120,7 @@ export function resolveDictationAgentVisionInference(settings, { isSignedIn = fa
baseUrl: isCustom ? resolved.cloudBaseUrl || undefined : undefined,
customApiKey: isCustom ? customApiKey || undefined : undefined,
disableThinking: resolved.disableThinking,
+ reasoningEffort: resolved.reasoningEffort,
},
};
}
@@ -197,6 +200,7 @@ export function resolveChatStreamingInference(
cloudBaseUrl: vision.config.baseUrl,
customApiKey: vision.config.customApiKey,
disableThinking: vision.config.disableThinking,
+ reasoningEffort: vision.config.reasoningEffort,
},
attachScreenContext: true,
};
diff --git a/src/helpers/dictationTranslationInference.js b/src/helpers/dictationTranslationInference.js
index e2e5c558c..24af05622 100644
--- a/src/helpers/dictationTranslationInference.js
+++ b/src/helpers/dictationTranslationInference.js
@@ -25,6 +25,7 @@ export function resolveDictationTranslationInference(
provider: managed.provider,
language: settings.translationTargetLanguage,
disableThinking: settings.translationDisableThinking,
+ reasoningEffort: settings.translationReasoningEffort,
},
};
}
@@ -65,6 +66,7 @@ export function resolveDictationTranslationInference(
customApiKey:
isCustom || isSelfHosted ? settings.translationCustomApiKey || undefined : undefined,
disableThinking: settings.translationDisableThinking,
+ reasoningEffort: settings.translationReasoningEffort,
},
};
}
diff --git a/src/locales/ar/translation.json b/src/locales/ar/translation.json
index 91e0e5d63..c4a76320f 100644
--- a/src/locales/ar/translation.json
+++ b/src/locales/ar/translation.json
@@ -410,6 +410,21 @@
"selectModelHint": "لم يتم تحديد أي نموذج بعد — اختر واحدًا من القائمة أدناه.",
"keyRequiredHint": "مفتاح API مطلوب للاستدلال — أضف واحدًا لاستخدام هذا المزود."
},
+ "effort": {
+ "title": "مستوى الاستدلال",
+ "description": "مقدار تفكير النموذج قبل أن يجيب.",
+ "descriptionWithDefault": "مقدار تفكير النموذج قبل أن يجيب. يستخدم الوضع التلقائي الإعداد الافتراضي للمزوّد ({{value}}).",
+ "auto": "تلقائي",
+ "options": {
+ "none": "بلا",
+ "minimal": "أدنى",
+ "low": "منخفض",
+ "medium": "متوسط",
+ "high": "مرتفع",
+ "xhigh": "مرتفع جدًا",
+ "max": "أقصى"
+ }
+ },
"getApiKey": "احصل على مفتاح API الخاص بك ←",
"selectModel": "تحديد النموذج",
"availableModels": "النماذج المتاحة",
diff --git a/src/locales/de/translation.json b/src/locales/de/translation.json
index 9583a664e..9782dd2e3 100644
--- a/src/locales/de/translation.json
+++ b/src/locales/de/translation.json
@@ -773,6 +773,21 @@
"selectModelHint": "Noch kein Modell ausgewählt — wählen Sie eines aus der Liste unten.",
"keyRequiredHint": "Für die Inferenz ist ein API-Schlüssel erforderlich — fügen Sie einen hinzu, um diesen Anbieter zu nutzen."
},
+ "effort": {
+ "title": "Denkaufwand",
+ "description": "Wie viel das Modell nachdenkt, bevor es antwortet.",
+ "descriptionWithDefault": "Wie viel das Modell nachdenkt, bevor es antwortet. Auto verwendet die Standardeinstellung des Anbieters ({{value}}).",
+ "auto": "Auto",
+ "options": {
+ "none": "Keiner",
+ "minimal": "Minimal",
+ "low": "Niedrig",
+ "medium": "Mittel",
+ "high": "Hoch",
+ "xhigh": "Sehr hoch",
+ "max": "Maximal"
+ }
+ },
"getApiKey": "API-Schlüssel holen ->",
"selectModel": "Modell auswählen",
"availableModels": "Verfügbare Modelle",
diff --git a/src/locales/en/translation.json b/src/locales/en/translation.json
index 5ebea019d..7b3d321dd 100644
--- a/src/locales/en/translation.json
+++ b/src/locales/en/translation.json
@@ -773,6 +773,21 @@
"selectModelHint": "No model selected yet — choose one from the list below.",
"keyRequiredHint": "An API key is required for inference — add one to use this provider."
},
+ "effort": {
+ "title": "Reasoning effort",
+ "description": "How much the model thinks before it answers.",
+ "descriptionWithDefault": "How much the model thinks before it answers. Auto uses the provider's default ({{value}}).",
+ "auto": "Auto",
+ "options": {
+ "none": "None",
+ "minimal": "Minimal",
+ "low": "Low",
+ "medium": "Medium",
+ "high": "High",
+ "xhigh": "Extra high",
+ "max": "Max"
+ }
+ },
"getApiKey": "Get your API key ->",
"selectModel": "Select Model",
"availableModels": "Available Models",
diff --git a/src/locales/es/translation.json b/src/locales/es/translation.json
index 807f4bfa2..163890de6 100644
--- a/src/locales/es/translation.json
+++ b/src/locales/es/translation.json
@@ -773,6 +773,21 @@
"selectModelHint": "Aún no hay un modelo seleccionado — elige uno de la lista de abajo.",
"keyRequiredHint": "Se requiere una clave API para la inferencia — agrega una para usar este proveedor."
},
+ "effort": {
+ "title": "Esfuerzo de razonamiento",
+ "description": "Cuánto piensa el modelo antes de responder.",
+ "descriptionWithDefault": "Cuánto piensa el modelo antes de responder. Auto usa el valor predeterminado del proveedor ({{value}}).",
+ "auto": "Auto",
+ "options": {
+ "none": "Ninguno",
+ "minimal": "Mínimo",
+ "low": "Bajo",
+ "medium": "Medio",
+ "high": "Alto",
+ "xhigh": "Muy alto",
+ "max": "Máximo"
+ }
+ },
"getApiKey": "Obtener clave API ->",
"selectModel": "Seleccionar modelo",
"availableModels": "Modelos disponibles",
diff --git a/src/locales/fr/translation.json b/src/locales/fr/translation.json
index d2f3b8adb..8fd20867a 100644
--- a/src/locales/fr/translation.json
+++ b/src/locales/fr/translation.json
@@ -773,6 +773,21 @@
"selectModelHint": "Aucun modèle sélectionné — choisissez-en un dans la liste ci-dessous.",
"keyRequiredHint": "Une clé API est requise pour l'inférence — ajoutez-en une pour utiliser ce fournisseur."
},
+ "effort": {
+ "title": "Effort de raisonnement",
+ "description": "Le temps de réflexion du modèle avant de répondre.",
+ "descriptionWithDefault": "Le temps de réflexion du modèle avant de répondre. Auto utilise la valeur par défaut du fournisseur ({{value}}).",
+ "auto": "Auto",
+ "options": {
+ "none": "Aucun",
+ "minimal": "Minimal",
+ "low": "Faible",
+ "medium": "Moyen",
+ "high": "Élevé",
+ "xhigh": "Très élevé",
+ "max": "Maximum"
+ }
+ },
"getApiKey": "Obtenir la clé API ->",
"selectModel": "Sélectionner un modèle",
"availableModels": "Modèles disponibles",
diff --git a/src/locales/it/translation.json b/src/locales/it/translation.json
index 6fa5a0cd1..12998cfd6 100644
--- a/src/locales/it/translation.json
+++ b/src/locales/it/translation.json
@@ -773,6 +773,21 @@
"selectModelHint": "Nessun modello selezionato — scegline uno dall'elenco qui sotto.",
"keyRequiredHint": "Per l'inferenza è richiesta una chiave API — aggiungine una per usare questo provider."
},
+ "effort": {
+ "title": "Livello di ragionamento",
+ "description": "Quanto ragiona il modello prima di rispondere.",
+ "descriptionWithDefault": "Quanto ragiona il modello prima di rispondere. Auto usa l'impostazione predefinita del provider ({{value}}).",
+ "auto": "Auto",
+ "options": {
+ "none": "Nessuno",
+ "minimal": "Minimo",
+ "low": "Basso",
+ "medium": "Medio",
+ "high": "Alto",
+ "xhigh": "Molto alto",
+ "max": "Massimo"
+ }
+ },
"getApiKey": "Ottieni la chiave API ->",
"selectModel": "Seleziona modello",
"availableModels": "Modelli disponibili",
diff --git a/src/locales/ja/translation.json b/src/locales/ja/translation.json
index 3ee93e73c..55908555d 100644
--- a/src/locales/ja/translation.json
+++ b/src/locales/ja/translation.json
@@ -773,6 +773,21 @@
"selectModelHint": "モデルが未選択です — 下のリストから選択してください。",
"keyRequiredHint": "推論には API キーが必要です — このプロバイダーを使用するには追加してください。"
},
+ "effort": {
+ "title": "推論の深さ",
+ "description": "モデルが回答する前にどれだけ考えるかを設定します。",
+ "descriptionWithDefault": "モデルが回答する前にどれだけ考えるかを設定します。自動ではプロバイダーの既定値({{value}})を使用します。",
+ "auto": "自動",
+ "options": {
+ "none": "なし",
+ "minimal": "最小",
+ "low": "低",
+ "medium": "中",
+ "high": "高",
+ "xhigh": "非常に高い",
+ "max": "最大"
+ }
+ },
"getApiKey": "API キーを取得 ->",
"selectModel": "モデルを選択",
"availableModels": "利用可能なモデル",
diff --git a/src/locales/pt/translation.json b/src/locales/pt/translation.json
index c9b829f69..9bcb2f064 100644
--- a/src/locales/pt/translation.json
+++ b/src/locales/pt/translation.json
@@ -752,6 +752,21 @@
"selectModelHint": "Nenhum modelo selecionado ainda — escolha um na lista abaixo.",
"keyRequiredHint": "Uma chave API é necessária para a inferência — adicione uma para usar este provedor."
},
+ "effort": {
+ "title": "Esforço de raciocínio",
+ "description": "Quanto o modelo pensa antes de responder.",
+ "descriptionWithDefault": "Quanto o modelo pensa antes de responder. Auto usa o padrão do provedor ({{value}}).",
+ "auto": "Auto",
+ "options": {
+ "none": "Nenhum",
+ "minimal": "Mínimo",
+ "low": "Baixo",
+ "medium": "Médio",
+ "high": "Alto",
+ "xhigh": "Muito alto",
+ "max": "Máximo"
+ }
+ },
"getApiKey": "Obter chave API ->",
"selectModel": "Selecionar modelo",
"availableModels": "Modelos disponíveis",
diff --git a/src/locales/ru/translation.json b/src/locales/ru/translation.json
index aa38733f9..76fc112ab 100644
--- a/src/locales/ru/translation.json
+++ b/src/locales/ru/translation.json
@@ -773,6 +773,21 @@
"selectModelHint": "Модель ещё не выбрана — выберите её из списка ниже.",
"keyRequiredHint": "Для инференса требуется API-ключ — добавьте его, чтобы использовать этого провайдера."
},
+ "effort": {
+ "title": "Глубина рассуждений",
+ "description": "Сколько модель размышляет перед ответом.",
+ "descriptionWithDefault": "Сколько модель размышляет перед ответом. «Авто» использует настройку провайдера по умолчанию ({{value}}).",
+ "auto": "Авто",
+ "options": {
+ "none": "Нет",
+ "minimal": "Минимальная",
+ "low": "Низкая",
+ "medium": "Средняя",
+ "high": "Высокая",
+ "xhigh": "Очень высокая",
+ "max": "Максимальная"
+ }
+ },
"getApiKey": "Получить API-ключ ->",
"selectModel": "Выберите модель",
"availableModels": "Доступные модели",
diff --git a/src/locales/zh-CN/translation.json b/src/locales/zh-CN/translation.json
index 1fe403fb1..4e26dd7d0 100644
--- a/src/locales/zh-CN/translation.json
+++ b/src/locales/zh-CN/translation.json
@@ -773,6 +773,21 @@
"selectModelHint": "尚未选择模型 — 请从下方列表中选择一个。",
"keyRequiredHint": "推理需要 API 密钥 — 请添加密钥以使用此提供商。"
},
+ "effort": {
+ "title": "推理强度",
+ "description": "模型在回答前思考的程度。",
+ "descriptionWithDefault": "模型在回答前思考的程度。“自动”使用服务商的默认设置({{value}})。",
+ "auto": "自动",
+ "options": {
+ "none": "无",
+ "minimal": "极低",
+ "low": "低",
+ "medium": "中",
+ "high": "高",
+ "xhigh": "极高",
+ "max": "最大"
+ }
+ },
"getApiKey": "获取你的 API Key ->",
"selectModel": "选择模型",
"availableModels": "可用模型",
diff --git a/src/locales/zh-TW/translation.json b/src/locales/zh-TW/translation.json
index d805eeda3..96e7a46e9 100644
--- a/src/locales/zh-TW/translation.json
+++ b/src/locales/zh-TW/translation.json
@@ -773,6 +773,21 @@
"selectModelHint": "尚未選擇模型 — 請從下方清單中選擇一個。",
"keyRequiredHint": "推論需要 API 金鑰 — 請新增金鑰以使用此供應商。"
},
+ "effort": {
+ "title": "推理強度",
+ "description": "模型在回答前思考的程度。",
+ "descriptionWithDefault": "模型在回答前思考的程度。「自動」使用服務商的預設設定({{value}})。",
+ "auto": "自動",
+ "options": {
+ "none": "無",
+ "minimal": "極低",
+ "low": "低",
+ "medium": "中",
+ "high": "高",
+ "xhigh": "極高",
+ "max": "最大"
+ }
+ },
"getApiKey": "取得 API Key ->",
"selectModel": "選擇模型",
"availableModels": "可用模型",
diff --git a/src/models/ModelRegistry.ts b/src/models/ModelRegistry.ts
index 92eba21f9..5504bf914 100644
--- a/src/models/ModelRegistry.ts
+++ b/src/models/ModelRegistry.ts
@@ -48,6 +48,12 @@ export interface CloudModelDefinition {
supportsVision?: boolean;
tokenParam?: "max_tokens" | "max_completion_tokens";
supportsTemperature?: boolean;
+ reasoningEffort?: ReasoningEffortCapability;
+}
+
+export interface ReasoningEffortCapability {
+ options: string[];
+ defaultValue?: string;
}
export interface CloudProviderData {
diff --git a/src/models/modelRegistryData.json b/src/models/modelRegistryData.json
index a72cc641e..0afa897c4 100644
--- a/src/models/modelRegistryData.json
+++ b/src/models/modelRegistryData.json
@@ -433,7 +433,10 @@
"descriptionKey": "models.descriptions.cloud.openai_gpt_6_astra",
"tokenParam": "max_completion_tokens",
"supportsTemperature": false,
- "supportsVision": true
+ "supportsVision": true,
+ "reasoningEffort": {
+ "options": ["low", "medium", "high", "xhigh", "max"]
+ }
},
{
"id": "gpt-5.6-sol",
@@ -442,7 +445,11 @@
"descriptionKey": "models.descriptions.cloud.openai_gpt_5_6_sol",
"tokenParam": "max_completion_tokens",
"supportsTemperature": false,
- "supportsVision": true
+ "supportsVision": true,
+ "reasoningEffort": {
+ "options": ["none", "low", "medium", "high", "xhigh", "max"],
+ "defaultValue": "medium"
+ }
},
{
"id": "gpt-5.6-terra",
@@ -451,7 +458,11 @@
"descriptionKey": "models.descriptions.cloud.openai_gpt_5_6_terra",
"tokenParam": "max_completion_tokens",
"supportsTemperature": false,
- "supportsVision": true
+ "supportsVision": true,
+ "reasoningEffort": {
+ "options": ["none", "low", "medium", "high", "xhigh", "max"],
+ "defaultValue": "medium"
+ }
},
{
"id": "gpt-5.6-luna",
@@ -460,7 +471,11 @@
"descriptionKey": "models.descriptions.cloud.openai_gpt_5_6_luna",
"tokenParam": "max_completion_tokens",
"supportsTemperature": false,
- "supportsVision": true
+ "supportsVision": true,
+ "reasoningEffort": {
+ "options": ["none", "low", "medium", "high", "xhigh", "max"],
+ "defaultValue": "medium"
+ }
},
{
"id": "gpt-5.5",
@@ -469,7 +484,11 @@
"descriptionKey": "models.descriptions.cloud.openai_gpt_5_5",
"tokenParam": "max_completion_tokens",
"supportsTemperature": false,
- "supportsVision": true
+ "supportsVision": true,
+ "reasoningEffort": {
+ "options": ["none", "low", "medium", "high", "xhigh"],
+ "defaultValue": "medium"
+ }
},
{
"id": "gpt-5.2",
@@ -478,7 +497,11 @@
"descriptionKey": "models.descriptions.cloud.openai_gpt_5_2",
"tokenParam": "max_completion_tokens",
"supportsTemperature": false,
- "supportsVision": true
+ "supportsVision": true,
+ "reasoningEffort": {
+ "options": ["none", "low", "medium", "high", "xhigh"],
+ "defaultValue": "none"
+ }
},
{
"id": "gpt-5-mini",
@@ -487,7 +510,11 @@
"descriptionKey": "models.descriptions.cloud.openai_gpt_5_mini",
"tokenParam": "max_completion_tokens",
"supportsTemperature": false,
- "supportsVision": true
+ "supportsVision": true,
+ "reasoningEffort": {
+ "options": ["minimal", "low", "medium", "high"],
+ "defaultValue": "medium"
+ }
},
{
"id": "gpt-5-nano",
@@ -496,7 +523,11 @@
"descriptionKey": "models.descriptions.cloud.openai_gpt_5_nano",
"tokenParam": "max_completion_tokens",
"supportsTemperature": false,
- "supportsVision": true
+ "supportsVision": true,
+ "reasoningEffort": {
+ "options": ["minimal", "low", "medium", "high"],
+ "defaultValue": "medium"
+ }
},
{
"id": "gpt-4.1",
@@ -631,7 +662,11 @@
"description": "Latest fast, high-capability Gemini model",
"descriptionKey": "models.descriptions.cloud.gemini_gemini_3_5_flash",
"supportsThinking": true,
- "supportsVision": true
+ "supportsVision": true,
+ "reasoningEffort": {
+ "options": ["minimal", "low", "medium", "high"],
+ "defaultValue": "medium"
+ }
},
{
"id": "gemini-3.5-flash-lite",
@@ -639,7 +674,11 @@
"description": "Fastest, most cost-effective 3.5 model",
"descriptionKey": "models.descriptions.cloud.gemini_gemini_3_5_flash_lite",
"supportsThinking": true,
- "supportsVision": true
+ "supportsVision": true,
+ "reasoningEffort": {
+ "options": ["minimal", "low", "medium", "high"],
+ "defaultValue": "minimal"
+ }
},
{
"id": "gemini-3.1-pro-preview",
@@ -647,14 +686,21 @@
"description": "Next-gen flagship model for complex reasoning",
"descriptionKey": "models.descriptions.cloud.gemini_gemini_3_1_pro_preview",
"supportsThinking": true,
- "supportsVision": true
+ "supportsVision": true,
+ "reasoningEffort": {
+ "options": ["low", "medium", "high"],
+ "defaultValue": "high"
+ }
},
{
"id": "gemini-3.1-flash-lite",
"name": "Gemini 3.1 Flash Lite",
"description": "Frontier-class performance at low cost",
"descriptionKey": "models.descriptions.cloud.gemini_gemini_3_1_flash_lite",
- "supportsVision": true
+ "supportsVision": true,
+ "reasoningEffort": {
+ "options": ["minimal", "low", "medium", "high"]
+ }
},
{
"id": "gemini-3-flash-preview",
@@ -662,7 +708,11 @@
"description": "Ultra-fast, high-capability next-gen model",
"descriptionKey": "models.descriptions.cloud.gemini_gemini_3_flash_preview",
"supportsThinking": true,
- "supportsVision": true
+ "supportsVision": true,
+ "reasoningEffort": {
+ "options": ["minimal", "low", "medium", "high"],
+ "defaultValue": "high"
+ }
},
{
"id": "gemini-2.5-flash-lite",
@@ -694,14 +744,20 @@
"name": "GPT-OSS 120B",
"description": "OpenAI's open-source flagship, 500 T/sec",
"supportsThinking": true,
- "descriptionKey": "models.descriptions.cloud.groq_openai_gpt_oss_120b"
+ "descriptionKey": "models.descriptions.cloud.groq_openai_gpt_oss_120b",
+ "reasoningEffort": {
+ "options": ["low", "medium", "high"]
+ }
},
{
"id": "openai/gpt-oss-20b",
"name": "GPT-OSS 20B",
"description": "Fast open-source model, 1000 T/sec",
"supportsThinking": true,
- "descriptionKey": "models.descriptions.cloud.groq_openai_gpt_oss_20b"
+ "descriptionKey": "models.descriptions.cloud.groq_openai_gpt_oss_20b",
+ "reasoningEffort": {
+ "options": ["low", "medium", "high"]
+ }
},
{
"id": "groq/compound",
diff --git a/src/models/reasoningEffort.ts b/src/models/reasoningEffort.ts
new file mode 100644
index 000000000..15ce017b6
--- /dev/null
+++ b/src/models/reasoningEffort.ts
@@ -0,0 +1,63 @@
+import { getCloudModel, modelRegistry, type ReasoningEffortCapability } from "./ModelRegistry";
+import { getModelFamilyConstraints } from "../services/ai/modelFamilyConstraints";
+import { matchesHost } from "../utils/urlUtils";
+
+export const AUTO_REASONING_EFFORT = "";
+
+export const OPEN_ENDPOINT_PROVIDERS = new Set(["custom", "openrouter"]);
+
+const OPEN_ENDPOINT_OPTIONS = ["none", "minimal", "low", "medium", "high", "xhigh"];
+
+// Hosts whose levels are fixed by the API whatever the model, from the provider docs.
+const HOST_OPTIONS: Array<{ host: string; options: string[] }> = [
+ // grok-4.5 and later; reasoning can't be turned off, and grok-4.5 reads xhigh as high.
+ { host: "x.ai", options: ["low", "medium", "high", "xhigh"] },
+ // DeepSeek rejects "none" (#1260) and folds every other level into low, high or
+ // max. Max is left out: the AI SDK's Chat Completions options refuse it.
+ { host: "deepseek.com", options: ["low", "high"] },
+];
+
+export function getReasoningEffortCapability(
+ provider: string,
+ model: string,
+ endpoint?: string | null
+): ReasoningEffortCapability | null {
+ // Custom endpoints and OpenRouter serve models the registry may not describe,
+ // so they offer the host's levels, else the family's, else the registry's,
+ // else the OpenAI-style list. The server's default is unknown, so none is named.
+ if (OPEN_ENDPOINT_PROVIDERS.has(provider)) {
+ const hostOptions = HOST_OPTIONS.find(({ host }) => matchesHost(endpoint, host))?.options;
+ if (hostOptions) return { options: hostOptions };
+ const family = getModelFamilyConstraints(model);
+ if (family?.omitReasoningParams) return null;
+ // A Custom turn streams through the AI SDK's Chat Completions options,
+ // which take no "max"; OpenRouter's own reasoning object does.
+ const registryOptions = getCloudModel(model, provider)?.reasoningEffort?.options.filter(
+ (option) => provider === "openrouter" || OPEN_ENDPOINT_OPTIONS.includes(option)
+ );
+ return {
+ options: family?.reasoningEffort?.options ?? registryOptions ?? OPEN_ENDPOINT_OPTIONS,
+ };
+ }
+ // Only the provider's own entries, so a self-hosted server that serves a
+ // registry model id never inherits a level saved for a cloud provider.
+ return (
+ modelRegistry
+ .getCloudProviders()
+ .find((cloudProvider) => cloudProvider.id === provider)
+ ?.models.find((cloudModel) => cloudModel.id === model)?.reasoningEffort ?? null
+ );
+}
+
+export function isReasoningEffortSupported(
+ capability: ReasoningEffortCapability | null | undefined,
+ value: string | null | undefined
+): boolean {
+ return !!capability && !!value && capability.options.includes(value);
+}
+
+export function hasSelectableReasoningEffort(
+ capability: ReasoningEffortCapability | null | undefined
+): capability is ReasoningEffortCapability {
+ return Boolean(capability && capability.options.length > 1);
+}
diff --git a/src/services/BaseReasoningService.ts b/src/services/BaseReasoningService.ts
index 2c828962d..0320066ab 100644
--- a/src/services/BaseReasoningService.ts
+++ b/src/services/BaseReasoningService.ts
@@ -15,6 +15,7 @@ export interface ReasoningConfig {
customApiKey?: string;
provider?: string;
disableThinking?: boolean;
+ reasoningEffort?: string;
/** Screenshot attached to voice-agent requests when screen context is on. */
screenContext?: ScreenContextImage;
/** Suffix-free prompt used when a screenshot-carrying request is retried text-only. */
diff --git a/src/services/ReasoningService.ts b/src/services/ReasoningService.ts
index 7af5f0cec..0b6caf7be 100644
--- a/src/services/ReasoningService.ts
+++ b/src/services/ReasoningService.ts
@@ -34,10 +34,12 @@ import {
} from "./ai/openaiBase";
import {
applyChatCompletionsParams,
+ chosenReasoningEffort,
emptyResponseError,
fetchWithParamFallback,
isTruncatedFinishReason,
truncatedOutputError,
+ type ParamFallbackDetails,
} from "./ai/chatRequestBody";
import { getModelFamilyConstraints } from "./ai/modelFamilyConstraints";
import { detectEndpointDialect } from "./ai/thinkingSuppressionDialects";
@@ -113,8 +115,7 @@ function assertAgentSessionAllowedByPolicy(provider: string, mode: InferenceMode
}
function logParamFallback(logEvent: string) {
- return (details: { status: number; stripped: string[] }) =>
- logger.logReasoning(logEvent, details);
+ return (details: ParamFallbackDetails) => logger.logReasoning(logEvent, details);
}
class ReasoningService extends BaseReasoningService {
@@ -364,6 +365,7 @@ class ReasoningService extends BaseReasoningService {
}
const res = await fetchWithParamFallback(
+ endpoint,
() =>
fetch(endpoint, {
method: "POST",
@@ -695,6 +697,7 @@ class ReasoningService extends BaseReasoningService {
let response: Response;
try {
response = await fetchWithParamFallback(
+ endpoint,
() =>
fetch(endpoint, {
method: "POST",
@@ -901,11 +904,14 @@ class ReasoningService extends BaseReasoningService {
// OpenRouter ids are never in the local registry, so the supportsThinking
// exemption below can't apply — honor the toggle directly.
const openrouterDisableThinking = provider === "openrouter" && config.disableThinking === true;
+ // No Tinfoil model offers a level, so this needn't wait for the registry refresh below.
+ const reasoningEffort = chosenReasoningEffort(aiProvider, model, config, baseURL);
// Resolving a Tinfoil model refreshes the registry, so read model config after it.
const aiModel = isEnterprise
? createEnterpriseChatModel(provider as EnterpriseProvider, model, config.inferenceScope)
: await getAIModel(aiProvider, model, apiKey, baseURL, {
disableThinking: openrouterDisableThinking,
+ reasoningEffort: provider === "openrouter" ? reasoningEffort : undefined,
});
if (abortController.signal.aborted) {
@@ -932,6 +938,16 @@ class ReasoningService extends BaseReasoningService {
...(needsGeminiMinimalThinking
? { google: { thinkingConfig: { thinkingLevel: "minimal", includeThoughts: false } } }
: {}),
+ ...(reasoningEffort && provider === "groq" ? { groq: { reasoningEffort } } : {}),
+ ...(reasoningEffort && provider === "gemini"
+ ? { google: { thinkingConfig: { thinkingLevel: reasoningEffort, includeThoughts: false } } }
+ : {}),
+ // The SDK sends a Responses effort only for model ids it knows as reasoning
+ // models; forceReasoning covers newer OpenAI ids such as gpt-6-astra.
+ ...(reasoningEffort && provider === "openai"
+ ? { openai: { reasoningEffort, forceReasoning: true } }
+ : {}),
+ ...(reasoningEffort && provider === "custom" ? { openai: { reasoningEffort } } : {}),
};
const hasProviderOptions = Object.keys(providerOptions).length > 0;
diff --git a/src/services/ai/chatRequestBody.ts b/src/services/ai/chatRequestBody.ts
index 7fd3bc6c2..d63546632 100644
--- a/src/services/ai/chatRequestBody.ts
+++ b/src/services/ai/chatRequestBody.ts
@@ -1,5 +1,12 @@
import type { ReasoningConfig } from "../BaseReasoningService";
-import { getOpenAiApiConfig } from "../../models/ModelRegistry";
+import { getOpenAiApiConfig, type ReasoningEffortCapability } from "../../models/ModelRegistry";
+import {
+ getReasoningEffortCapability,
+ hasSelectableReasoningEffort,
+ isReasoningEffortSupported,
+ OPEN_ENDPOINT_PROVIDERS,
+} from "../../models/reasoningEffort";
+import { matchesHost } from "../../utils/urlUtils";
import { detectEndpointDialect } from "./thinkingSuppressionDialects";
import { getModelFamilyConstraints } from "./modelFamilyConstraints";
import { applyThinkingSuppression, isThinkingSuppressed } from "./thinkingSuppression";
@@ -16,11 +23,11 @@ const LEGACY_CHAT_COMPLETIONS_PROVIDERS = new Set(["local", "lan", "groq", "tinf
/**
* Single place that turns (model, provider, endpoint, config) into the
* parameter set of an OpenAI-compatible chat-completions body: token-limit
- * param, temperature, family reasoning effort, and thinking suppression.
+ * param, temperature, reasoning effort, and thinking suppression.
* Callers own `model`/`messages`/`stream`; this owns every tunable param, so
* a family or provider fact changed in the data tables reaches all transports
* at once instead of one call site at a time (#1611). Responses bodies get the
- * same family effort from familyResponsesEffort; keep the two in step.
+ * same effort from responsesReasoningEffort; keep the two in step.
*/
export function applyChatCompletionsParams(
requestBody: Record<string, unknown>,
@@ -55,11 +62,17 @@ export function applyChatCompletionsParams(
}
}
- // Deterministic transforms (cleanup, selection edits) pin the family's
- // preferred effort — see modelFamilyConstraints for the gpt-oss rationale.
- // applyThinkingSuppression still wins when thinking is disabled by the user.
+ // The user's chosen effort wins; otherwise deterministic transforms
+ // (cleanup, selection edits) pin the family's preferred effort — see
+ // modelFamilyConstraints for the gpt-oss rationale. A suppressed request has
+ // no chosen effort, so applyThinkingSuppression still decides it.
const familyEffort = getModelFamilyConstraints(model)?.reasoningEffort;
- if (familyEffort?.cleanupValue && isDeterministicTransform(config)) {
+ const effort = chosenReasoningEffort(providerKey, model, config, endpoint);
+ if (effort) {
+ // OpenRouter takes effort only inside its reasoning object.
+ if (providerKey === "openrouter") requestBody.reasoning = { effort };
+ else requestBody.reasoning_effort = effort;
+ } else if (familyEffort?.cleanupValue && isDeterministicTransform(config)) {
requestBody.reasoning_effort = familyEffort.cleanupValue;
}
@@ -72,24 +85,75 @@ export function applyChatCompletionsParams(
);
// Some dialects suppress through other fields, so a default effort beside
// them would contradict the suppression.
- if (!suppressed && familyEffort?.defaultValue) {
+ if (!suppressed && !effort && familyEffort?.defaultValue) {
requestBody.reasoning_effort ??= familyEffort.defaultValue;
}
}
/**
- * The family's effort for a Responses API body (`reasoning: { effort }`), by
- * the same rules as the chat body above. A model outside the family table
- * sends no effort, leaving the server's default.
+ * The effort for a Responses API body (`reasoning: { effort }`), by the same
+ * rules as the chat body above. Suppression sends the family's floor, or
+ * "none" like the generic chat dialect; with no choice, a model outside the
+ * family table sends no effort, leaving the server's default.
*/
-export function familyResponsesEffort(model: string, config: ReasoningConfig): string | undefined {
+export function responsesReasoningEffort(
+ model: string,
+ provider: string,
+ config: ReasoningConfig,
+ endpoint?: string | null
+): string | undefined {
const familyEffort = getModelFamilyConstraints(model)?.reasoningEffort;
- if (!familyEffort) return undefined;
- if (isThinkingSuppressed(model, config)) return familyEffort.suppressValue;
- if (familyEffort.cleanupValue && isDeterministicTransform(config)) {
+ if (isThinkingSuppressed(model, config)) {
+ // Native OpenAI shows no toggle, so its default-on value is not the user's
+ // choice: a model id the registry no longer lists keeps the server's default.
+ return (
+ familyEffort?.suppressValue ?? (provider.toLowerCase() === "openai" ? undefined : "none")
+ );
+ }
+ const effort = chosenReasoningEffort(provider, model, config, endpoint);
+ if (effort) return effort;
+ if (familyEffort?.cleanupValue && isDeterministicTransform(config)) {
return familyEffort.cleanupValue;
}
- return familyEffort.defaultValue;
+ return familyEffort?.defaultValue;
+}
+
+/**
+ * The levels Settings offers for this provider, model and endpoint, or null
+ * when it shows no picker: thinking is suppressed, the toggle Custom and
+ * OpenRouter show for every model is on, or the model has fewer than two.
+ */
+export function selectableReasoningEffort(
+ provider: string,
+ model: string,
+ config: ReasoningConfig,
+ endpoint?: string | null
+): ReasoningEffortCapability | null {
+ const providerKey = provider.toLowerCase();
+ if (!model || isThinkingSuppressed(model, config)) return null;
+ if (OPEN_ENDPOINT_PROVIDERS.has(providerKey) && config.disableThinking === true) return null;
+ const capability = getReasoningEffortCapability(providerKey, model, endpoint);
+ return hasSelectableReasoningEffort(capability) ? capability : null;
+}
+
+/**
+ * The effort the user picked for this scope, when Settings offers it for this
+ * provider and model. A stored level it doesn't offer (the model, provider or
+ * toggle changed since) is ignored.
+ */
+export function chosenReasoningEffort(
+ provider: string,
+ model: string,
+ config: ReasoningConfig,
+ endpoint?: string | null
+): string | undefined {
+ const effort = config.reasoningEffort;
+ return isReasoningEffortSupported(
+ selectableReasoningEffort(provider, model, config, endpoint),
+ effort
+ )
+ ? effort
+ : undefined;
}
function isDeterministicTransform(config: ReasoningConfig): boolean {
@@ -147,7 +211,7 @@ export function emptyResponseError(
* module's shaping layer added are strippable — never messages or model — so a
* retry degrades the request (model reasons when asked not to, default
* sampling) instead of failing it outright. The strip is logged loudly: a
- * fallback that fires on every request is a dialect bug to fix, not a rescue.
+ * fallback that keeps firing is a dialect bug to fix, not a rescue.
*/
const STRIPPABLE_SHAPED_PARAMS = [
"reasoning_effort",
@@ -157,39 +221,79 @@ const STRIPPABLE_SHAPED_PARAMS = [
"temperature",
] as const;
+/**
+ * Values a server rejected, as `endpoint, model, param=value` → when the
+ * memory expires. Keyed by value, so another effort level is still sent; the
+ * expiry retries a value once the model behind the endpoint may have changed.
+ */
+const rejectedParams = new Map<string, number>();
+const REJECTION_MEMORY_MS = 10 * 60 * 1000;
+
+/** `remembered`: left out up front because an earlier retry proved it rejected. */
+export type ParamFallbackDetails = { status?: number; stripped: string[]; remembered?: true };
+
/**
* Fetch with a bounded param-stripping ladder for 400/422 rejections.
* Rung 1 (blind): old Ollama/strict proxies reject the `reasoning` object
* without naming it — drop it and retry once. Rung 2 (named): strip exactly
* the shaped params the error body names and retry once. At most two retries;
* the caller must build the body inside `doFetch` so retries re-serialize.
+ * A value the error named whose removal let the retry through is left out of
+ * later requests to the same endpoint and model for REJECTION_MEMORY_MS, so
+ * each pays the rejection once rather than on every request. OpenRouter is
+ * never remembered: it routes each request to one of several upstreams.
*/
export async function fetchWithParamFallback(
+ endpoint: string,
doFetch: () => Promise<Response>,
requestBody: Record<string, unknown>,
- logRejection: (details: { status: number; stripped: string[] }) => void
+ logRejection: (details: ParamFallbackDetails) => void
): Promise<Response> {
+ const rejectionKey = (param: string) =>
+ `${endpoint}\n${requestBody.model}\n${param}=${JSON.stringify(requestBody[param])}`;
+ const remembered = ["reasoning", ...STRIPPABLE_SHAPED_PARAMS].filter(
+ (param) => param in requestBody && (rejectedParams.get(rejectionKey(param)) ?? 0) > Date.now()
+ );
+ const canRemember = !matchesHost(endpoint, "openrouter.ai");
+ if (remembered.length > 0) {
+ logRejection({ stripped: remembered, remembered: true });
+ for (const param of remembered) delete requestBody[param];
+ }
+ const readError = (response: Response) =>
+ response
+ .clone()
+ .text()
+ .catch(() => "");
+ const retryWithout = async (params: readonly string[], remember: boolean) => {
+ const keys = params.map(rejectionKey);
+ for (const param of params) delete requestBody[param];
+ const retried = await doFetch();
+ if (remember && canRemember && retried.ok) {
+ const expiresAt = Date.now() + REJECTION_MEMORY_MS;
+ for (const key of keys) rejectedParams.set(key, expiresAt);
+ }
+ return retried;
+ };
+
let res = await doFetch();
if (res.ok || (res.status !== 400 && res.status !== 422)) return res;
if (requestBody.reasoning) {
+ // Remembered only when the error names it: a blind 400 may have come from
+ // one of several backends behind the same endpoint.
+ const reasoningNamed = (await readError(res)).toLowerCase().includes("reasoning");
logRejection({ status: res.status, stripped: ["reasoning"] });
- delete requestBody.reasoning;
void res.body?.cancel();
- res = await doFetch();
+ res = await retryWithout(["reasoning"], reasoningNamed);
if (res.ok || (res.status !== 400 && res.status !== 422)) return res;
}
- const errorText = await res
- .clone()
- .text()
- .catch(() => "");
+ const errorText = await readError(res);
const named = STRIPPABLE_SHAPED_PARAMS.filter((p) => p in requestBody && errorText.includes(p));
if (named.length > 0) {
logRejection({ status: res.status, stripped: named });
- for (const param of named) delete requestBody[param];
void res.body?.cancel();
- res = await doFetch();
+ res = await retryWithout(named, true);
}
return res;
}
diff --git a/src/services/ai/inferenceProviders/gemini.ts b/src/services/ai/inferenceProviders/gemini.ts
index c22c65d09..fbe5f8e88 100644
--- a/src/services/ai/inferenceProviders/gemini.ts
+++ b/src/services/ai/inferenceProviders/gemini.ts
@@ -7,7 +7,7 @@ import {
} from "../../../helpers/llmRequestTimeout.js";
import { extractGeminiText } from "../../../helpers/geminiResponse.js";
import { wrapCleanupTranscript } from "../../../config/prompts";
-import { emptyOutputError, truncatedOutputError } from "../chatRequestBody";
+import { chosenReasoningEffort, emptyOutputError, truncatedOutputError } from "../chatRequestBody";
import logger from "../../../utils/logger";
import {
asProviderError,
@@ -31,7 +31,7 @@ interface GeminiResponse {
interface GeminiGenerationConfig {
temperature: number;
thinkingConfig?: {
- thinkingLevel?: "minimal" | "low";
+ thinkingLevel?: string;
thinkingBudget?: number;
includeThoughts: boolean;
};
@@ -73,7 +73,12 @@ export const geminiProvider: InferenceProvider = {
// limit and the request timeout bound the reply.
};
- if (config.disableThinking === true && minimalThinking[model]) {
+ // A chosen level is only offered while thinking isn't suppressed, so it wins
+ // over the minimal floor that a model without the toggle gets by default.
+ const thinkingLevel = chosenReasoningEffort("gemini", model, config);
+ if (thinkingLevel) {
+ generationConfig.thinkingConfig = { thinkingLevel, includeThoughts: false };
+ } else if (config.disableThinking === true && minimalThinking[model]) {
generationConfig.thinkingConfig = minimalThinking[model];
}
diff --git a/src/services/ai/inferenceProviders/openai.ts b/src/services/ai/inferenceProviders/openai.ts
index f0f79df66..5b592a19e 100644
--- a/src/services/ai/inferenceProviders/openai.ts
+++ b/src/services/ai/inferenceProviders/openai.ts
@@ -8,9 +8,9 @@ import { canBorrowCleanupCustomKey, resolveConfiguredOpenAIBase } from "../opena
import {
applyChatCompletionsParams,
emptyResponseError,
- familyResponsesEffort,
fetchWithParamFallback,
isTruncatedFinishReason,
+ responsesReasoningEffort,
truncatedOutputError,
} from "../chatRequestBody";
import { detectEndpointDialect } from "../thinkingSuppressionDialects";
@@ -275,7 +275,7 @@ export const openaiProvider: InferenceProvider = {
if (apiConfig.supportsTemperature) {
requestBody.temperature = config.temperature ?? (config.systemPrompt ? 0.3 : 0);
}
- const effort = familyResponsesEffort(model, config);
+ const effort = responsesReasoningEffort(model, resolvedProvider, config, openAiBase);
if (effort) requestBody.reasoning = { effort };
} else {
requestBody.messages = buildMessages(type);
@@ -289,6 +289,7 @@ export const openaiProvider: InferenceProvider = {
}
const res = await fetchWithParamFallback(
+ endpoint,
() =>
fetch(endpoint, {
method: "POST",
diff --git a/src/services/ai/modelFamilyConstraints.ts b/src/services/ai/modelFamilyConstraints.ts
index 924ecafcd..54ad4a68f 100644
--- a/src/services/ai/modelFamilyConstraints.ts
+++ b/src/services/ai/modelFamilyConstraints.ts
@@ -13,6 +13,8 @@ export interface ModelFamilyConstraints {
reasoningEffort?: {
/** Value that best approximates "thinking off" for reasoning_effort. */
suppressValue: string;
+ /** Levels a Custom or OpenRouter endpoint offers for this family. */
+ options?: string[];
/**
* Effort for deterministic transforms (cleanup, selection edits). gpt-oss
* defaults to medium; low cuts hidden reasoning tokens (latency) and the
@@ -37,22 +39,27 @@ const FAMILIES: Array<ModelFamilyConstraints & { match: RegExp }> = [
match: /gpt-oss/,
// gpt-oss accepts low|medium|high only; it has no off switch. Confirmed
// live on Groq and Tinfoil (#1611): "none" is a 400 on both.
- reasoningEffort: { suppressValue: "low", cleanupValue: "low" },
+ reasoningEffort: {
+ suppressValue: "low",
+ cleanupValue: "low",
+ options: ["low", "medium", "high"],
+ },
},
{
family: "qwen",
match: /qwen/,
// qwen3 accepts none|default only (Groq's enum; "none" is also what the
// generic dialect sends everywhere, so the fact is provider-agnostic).
- reasoningEffort: { suppressValue: "none" },
+ reasoningEffort: { suppressValue: "none", options: ["none"] },
},
{
family: "glm-5.3",
match: /glm[-_]?5[-.]?3/,
// GLM-5.3/Flash always reason: low|high|max only, no off switch.
// "low" is the closest to off. At the server default, max, a long meeting
- // spent the whole 4,096-token notes budget thinking (#2008); at high it fit.
- reasoningEffort: { suppressValue: "low", defaultValue: "high" },
+ // spent the whole 4,096-token notes budget thinking (#2008); at high it fit,
+ // so max isn't offered either.
+ reasoningEffort: { suppressValue: "low", defaultValue: "high", options: ["low", "high"] },
},
{
family: "magistral",
diff --git a/src/services/ai/providers.ts b/src/services/ai/providers.ts
index 015720591..48b9aab9c 100644
--- a/src/services/ai/providers.ts
+++ b/src/services/ai/providers.ts
@@ -15,23 +15,25 @@ import { openCodeSessionHeaders } from "./openCodeSession";
// OpenRouter's reasoning control is a top-level request field the AI SDK
// can't emit — inject it at the fetch boundary.
-const withDisabledReasoning: typeof fetch = (input, init) => {
- if (typeof init?.body === "string") {
- try {
- const body = JSON.parse(init.body);
- body.reasoning = { enabled: false };
- init = { ...init, body: JSON.stringify(body) };
- } catch {}
- }
- return fetch(input, init);
-};
+const withReasoning =
+ (reasoning: Record<string, unknown>): typeof fetch =>
+ (input, init) => {
+ if (typeof init?.body === "string") {
+ try {
+ const body = JSON.parse(init.body);
+ body.reasoning = reasoning;
+ init = { ...init, body: JSON.stringify(body) };
+ } catch {}
+ }
+ return fetch(input, init);
+ };
export async function getAIModel(
provider: string,
model: string,
apiKey: string,
baseURL?: string,
- opts?: { disableThinking?: boolean }
+ opts?: { disableThinking?: boolean; reasoningEffort?: string }
): Promise<LanguageModel> {
switch (provider) {
case "openai":
@@ -62,13 +64,19 @@ export async function getAIModel(
baseURL,
headers: openCodeSessionHeaders(baseURL),
}).chat(model);
- case "openrouter":
+ case "openrouter": {
+ const reasoning = opts?.disableThinking
+ ? { enabled: false }
+ : opts?.reasoningEffort
+ ? { effort: opts.reasoningEffort }
+ : null;
// OpenRouter implements Chat Completions, not the OpenAI Responses API.
return createOpenAI({
apiKey,
baseURL,
- ...(opts?.disableThinking ? { fetch: withDisabledReasoning } : {}),
+ ...(reasoning ? { fetch: withReasoning(reasoning) } : {}),
}).chat(model);
+ }
case "local":
return createOpenAI({ apiKey: apiKey || "no-key", baseURL }).chat(model);
default:
diff --git a/src/stores/actionProcessingStore.ts b/src/stores/actionProcessingStore.ts
index 63d04207e..fa8191ac7 100644
--- a/src/stores/actionProcessingStore.ts
+++ b/src/stores/actionProcessingStore.ts
@@ -499,6 +499,7 @@ export function runBackgroundAction(
maxTokens: rewritesSummary ? 2 * NOTE_OUTPUT_MAX_TOKENS : NOTE_OUTPUT_MAX_TOKENS,
temperature: 0.3,
disableThinking: settings.noteFormattingDisableThinking,
+ reasoningEffort: settings.noteFormattingReasoningEffort,
// A local model that shrinks the reply to fit the prompt refuses a reply
// that fills the shrunken allowance, so a recording is summarised in
// parts rather than saved clipped. A plain note has no parts route, so
diff --git a/src/stores/settingsStore.ts b/src/stores/settingsStore.ts
index eac85f35e..dec4a3994 100644
--- a/src/stores/settingsStore.ts
+++ b/src/stores/settingsStore.ts
@@ -978,6 +978,7 @@ export interface SettingsState
translationRemoteUrl: string;
translationCustomApiKey: string;
translationDisableThinking: boolean;
+ translationReasoningEffort: string;
useDictationTranslation: boolean;
translationSourceLanguage: string;
translationTargetLanguage: string;
@@ -1008,6 +1009,11 @@ export interface SettingsState
dictationAgentVisionDisableThinking: boolean;
noteFormattingDisableThinking: boolean;
chatAgentDisableThinking: boolean;
+ cleanupReasoningEffort: string;
+ dictationAgentReasoningEffort: string;
+ dictationAgentVisionReasoningEffort: string;
+ noteFormattingReasoningEffort: string;
+ chatAgentReasoningEffort: string;
customPrompts: Record<PromptKind, string>;
setCustomPrompt: (kind: PromptKind, value: string) => void;
@@ -1806,6 +1812,7 @@ export const useSettingsStore = create<SettingsState>()((set, get) => ({
translationRemoteUrl: readString("translationRemoteUrl", ""),
translationCustomApiKey: readString("translationCustomApiKey", ""),
translationDisableThinking: readBoolean("translationDisableThinking", true),
+ translationReasoningEffort: readString("translationReasoningEffort", ""),
useDictationTranslation: readBoolean("useDictationTranslation", false),
translationSourceLanguage: readString("translationSourceLanguage", "auto"),
translationTargetLanguage: readString("translationTargetLanguage", ""),
@@ -1954,6 +1961,11 @@ export const useSettingsStore = create<SettingsState>()((set, get) => ({
dictationAgentVisionDisableThinking: readBoolean("dictationAgentVisionDisableThinking", true),
noteFormattingDisableThinking: readBoolean("noteFormattingDisableThinking", true),
chatAgentDisableThinking: readBoolean("chatAgentDisableThinking", true),
+ cleanupReasoningEffort: readString("cleanupReasoningEffort", ""),
+ dictationAgentReasoningEffort: readString("dictationAgentReasoningEffort", ""),
+ dictationAgentVisionReasoningEffort: readString("dictationAgentVisionReasoningEffort", ""),
+ noteFormattingReasoningEffort: readString("noteFormattingReasoningEffort", ""),
+ chatAgentReasoningEffort: readString("chatAgentReasoningEffort", ""),
customPrompts: PROMPT_KIND_LIST.reduce(
(acc, kind) => ({ ...acc, [kind]: readString(`customPrompt.${kind}`, "") }),
@@ -2881,6 +2893,7 @@ export interface ResolvedLLMConfig {
remoteUrl?: string;
customApiKey?: string;
disableThinking: boolean;
+ reasoningEffort?: string;
}
export const selectResolvedLLMConfig = (
@@ -2914,6 +2927,7 @@ export const selectResolvedLLMConfig = (
// belongs on the request path — see selectResolvedNoteFormatting.
customApiKey: read("customApiKey"),
disableThinking,
+ reasoningEffort: read("reasoningEffort"),
};
const managed = getManagedScopeResolution(scope, state.enterpriseSetupMode);
if (managed.kind === "error") {
diff --git a/src/utils/generateTitle.ts b/src/utils/generateTitle.ts
index 70b45f12b..4460c43ad 100644
--- a/src/utils/generateTitle.ts
+++ b/src/utils/generateTitle.ts
@@ -17,6 +17,7 @@ export async function generateNoteTitle(
inferenceScope: "noteFormatting",
temperature: 0.3,
disableThinking: getSettings().noteFormattingDisableThinking,
+ reasoningEffort: getSettings().noteFormattingReasoningEffort,
...config,
});
return sanitizeGeneratedTitle(raw);
diff --git a/test/helpers/cleanupRouteConfig.test.js b/test/helpers/cleanupRouteConfig.test.js
index e0258e427..2e9b5419c 100644
--- a/test/helpers/cleanupRouteConfig.test.js
+++ b/test/helpers/cleanupRouteConfig.test.js
@@ -46,7 +46,11 @@ async function loadRouteResolver(t, provider = "test", mode = "providers") {
`,
},
});
- const settings = { useCleanupModel: true, cleanupDisableThinking: true };
+ const settings = {
+ useCleanupModel: true,
+ cleanupDisableThinking: true,
+ cleanupReasoningEffort: "medium",
+ };
const resolveReasoningRoute = (await vite.ssrLoadModule("/helpers/audioManager.js"))
.resolveReasoningRoute;
return (text, { voiceAgentRequested = false, translationRequested = false } = {}) =>
@@ -62,6 +66,7 @@ test("the cleanup route pins temperature 0 and requires complete output", async
assert.equal(route.config.inferenceScope, "dictationCleanup");
assert.equal(route.config.temperature, 0);
assert.equal(route.config.requireCompleteOutput, true);
+ assert.equal(route.config.reasoningEffort, "medium");
});
test("the translation chain's cleanup step pins the same config", async (t) => {
diff --git a/test/helpers/dictationAgentInference.test.js b/test/helpers/dictationAgentInference.test.js
index a80320713..da4326018 100644
--- a/test/helpers/dictationAgentInference.test.js
+++ b/test/helpers/dictationAgentInference.test.js
@@ -12,6 +12,7 @@ const baseSettings = {
dictationAgentCloudBaseUrl: "",
dictationAgentCustomApiKey: "",
dictationAgentDisableThinking: true,
+ dictationAgentReasoningEffort: "high",
};
test("uses the dictation agent scope, not the cleanup scope", async () => {
@@ -22,10 +23,12 @@ test("uses the dictation agent scope, not the cleanup scope", async () => {
// Cleanup is configured completely differently; it must not leak through.
cleanupProvider: "openai",
cleanupModel: "gpt-4.1-mini",
+ cleanupReasoningEffort: "low",
});
assert.equal(result.model, "gpt-5-mini");
assert.equal(result.config.provider, "openai");
+ assert.equal(result.config.reasoningEffort, "high");
});
test("a model-required agent is unreachable without an explicit model", async () => {
@@ -242,9 +245,11 @@ test("vision override runs as the dictation agent scope and inherits key with en
dictationAgentVisionModel: "vision-model",
dictationAgentVisionCloudBaseUrl: "",
dictationAgentVisionCustomApiKey: "",
+ dictationAgentVisionReasoningEffort: "low",
});
assert.equal(result.config.inferenceScope, "dictationAgent");
+ assert.equal(result.config.reasoningEffort, "low", "the vision scope keeps its own effort");
assert.equal(result.config.baseUrl, "https://agent.example.com/v1");
assert.equal(
result.config.customApiKey,
diff --git a/test/helpers/dictationTranslationInference.test.js b/test/helpers/dictationTranslationInference.test.js
index b4c787b35..a106a6faa 100644
--- a/test/helpers/dictationTranslationInference.test.js
+++ b/test/helpers/dictationTranslationInference.test.js
@@ -13,6 +13,7 @@ const baseSettings = {
translationCloudBaseUrl: "",
translationCustomApiKey: "",
translationDisableThinking: true,
+ translationReasoningEffort: "low",
};
test("providers mode keeps its explicit provider and translation settings", async () => {
@@ -25,6 +26,7 @@ test("providers mode keeps its explicit provider and translation settings", asyn
assert.equal(result.config.provider, "openai");
assert.equal(result.config.language, "es");
assert.equal(result.config.disableThinking, true);
+ assert.equal(result.config.reasoningEffort, "low");
});
test("local mode routes to llama.cpp with an empty stored provider", async () => {
@@ -171,9 +173,8 @@ test("enterprise mode without a provider fails closed", async () => {
test("managed enterprise access supplies the provider and model, ignoring manual settings", async (t) => {
const { resolveDictationTranslationInference } = await load();
- const { useEnterpriseIdentityStore } = await import(
- "../../src/stores/enterpriseIdentityStore.ts"
- );
+ const { useEnterpriseIdentityStore } =
+ await import("../../src/stores/enterpriseIdentityStore.ts");
const scopes = [
"dictationCleanup",
diff --git a/test/services/fetchWithParamFallback.test.js b/test/services/fetchWithParamFallback.test.js
index 8f243c3b5..c3db2e368 100644
--- a/test/services/fetchWithParamFallback.test.js
+++ b/test/services/fetchWithParamFallback.test.js
@@ -20,12 +20,19 @@ function fetcher(...specs) {
return doFetch;
}
+// Rejections are remembered per endpoint for the module's lifetime, so each
+// test talks to its own endpoint.
+let endpoints = 0;
+const nextEndpoint = () => `https://host-${++endpoints}.example/v1/chat/completions`;
+
test("2xx passes through with a single attempt", async () => {
const { fetchWithParamFallback } = await load();
const doFetch = fetcher(jsonResponse(200, { ok: true }));
const body = { reasoning: { effort: "none" } };
- const res = await fetchWithParamFallback(doFetch, body, () => assert.fail("no fallback"));
+ const res = await fetchWithParamFallback(nextEndpoint(), doFetch, body, () =>
+ assert.fail("no fallback")
+ );
assert.equal(res.status, 200);
assert.equal(doFetch.count(), 1);
assert.ok(body.reasoning, "body untouched on success");
@@ -37,7 +44,7 @@ test("400 with a reasoning object strips it blind and retries (old Ollama proxie
const body = { reasoning: { effort: "none" }, max_tokens: 10 };
const logged = [];
- const res = await fetchWithParamFallback(doFetch, body, (d) => logged.push(d));
+ const res = await fetchWithParamFallback(nextEndpoint(), doFetch, body, (d) => logged.push(d));
assert.equal(res.status, 200);
assert.equal(doFetch.count(), 2);
assert.ok(!("reasoning" in body));
@@ -53,7 +60,7 @@ test("400 naming a shaped param strips exactly that param and retries (#1611 cla
const body = { reasoning_effort: "none", chat_template_kwargs: {}, max_tokens: 10 };
const logged = [];
- const res = await fetchWithParamFallback(doFetch, body, (d) => logged.push(d));
+ const res = await fetchWithParamFallback(nextEndpoint(), doFetch, body, (d) => logged.push(d));
assert.equal(res.status, 200);
assert.ok(!("reasoning_effort" in body));
assert.ok("chat_template_kwargs" in body, "unnamed params stay");
@@ -69,7 +76,7 @@ test("422 naming temperature strips it (#1417 class)", async () => {
);
const body = { temperature: 0.3, max_tokens: 10 };
- const res = await fetchWithParamFallback(doFetch, body, () => {});
+ const res = await fetchWithParamFallback(nextEndpoint(), doFetch, body, () => {});
assert.equal(res.status, 200);
assert.ok(!("temperature" in body));
});
@@ -79,7 +86,9 @@ test("400 naming nothing strippable returns the failure without retrying", async
const doFetch = fetcher(jsonResponse(400, { error: "model not found" }));
const body = { max_tokens: 10 };
- const res = await fetchWithParamFallback(doFetch, body, () => assert.fail("no fallback"));
+ const res = await fetchWithParamFallback(nextEndpoint(), doFetch, body, () =>
+ assert.fail("no fallback")
+ );
assert.equal(res.status, 400);
assert.equal(doFetch.count(), 1);
});
@@ -89,7 +98,7 @@ test("the ladder is bounded: at most two retries even when everything fails", as
const doFetch = fetcher(jsonResponse(400, { error: "reasoning_effort and thinking rejected" }));
const body = { reasoning: {}, reasoning_effort: "low", thinking: { type: "disabled" } };
- const res = await fetchWithParamFallback(doFetch, body, () => {});
+ const res = await fetchWithParamFallback(nextEndpoint(), doFetch, body, () => {});
assert.equal(res.status, 400);
assert.equal(doFetch.count(), 3);
});
@@ -99,8 +108,169 @@ test("non-4xx server errors never trigger stripping", async () => {
const doFetch = fetcher(jsonResponse(500, { error: "reasoning_effort exploded" }));
const body = { reasoning_effort: "low" };
- const res = await fetchWithParamFallback(doFetch, body, () => assert.fail("no fallback"));
+ const res = await fetchWithParamFallback(nextEndpoint(), doFetch, body, () =>
+ assert.fail("no fallback")
+ );
assert.equal(res.status, 500);
assert.equal(doFetch.count(), 1);
assert.ok("reasoning_effort" in body);
});
+
+const REASONING_REJECTED = { error: { message: "Unsupported parameter: 'reasoning.effort'" } };
+
+test("a value the retry proved rejected is left out of later requests", async () => {
+ const { fetchWithParamFallback } = await load();
+ const endpoint = nextEndpoint();
+ const rejecting = fetcher(jsonResponse(400, REASONING_REJECTED), jsonResponse(200, {}));
+ await fetchWithParamFallback(
+ endpoint,
+ rejecting,
+ { model: "m", reasoning: { effort: "none" } },
+ () => {}
+ );
+
+ const doFetch = fetcher(jsonResponse(200, {}));
+ const body = { model: "m", reasoning: { effort: "none" }, max_tokens: 10 };
+ const logged = [];
+ const res = await fetchWithParamFallback(endpoint, doFetch, body, (d) => logged.push(d));
+ assert.equal(res.status, 200);
+ assert.equal(doFetch.count(), 1);
+ assert.ok(!("reasoning" in body));
+ assert.ok("max_tokens" in body);
+ assert.deepEqual(logged, [{ stripped: ["reasoning"], remembered: true }]);
+});
+
+test("another value, model or endpoint is still sent after a rejection", async () => {
+ const { fetchWithParamFallback } = await load();
+ const endpoint = nextEndpoint();
+ await fetchWithParamFallback(
+ endpoint,
+ fetcher(jsonResponse(400, REASONING_REJECTED), jsonResponse(200, {})),
+ { model: "m", reasoning: { effort: "none" } },
+ () => {}
+ );
+
+ for (const [target, body] of [
+ [endpoint, { model: "m", reasoning: { effort: "low" } }],
+ [endpoint, { model: "other", reasoning: { effort: "none" } }],
+ [nextEndpoint(), { model: "m", reasoning: { effort: "none" } }],
+ ]) {
+ await fetchWithParamFallback(target, fetcher(jsonResponse(200, {})), body, () => {});
+ assert.ok("reasoning" in body, JSON.stringify(body));
+ }
+});
+
+test("a named param is remembered once the retry without it succeeds", async () => {
+ const { fetchWithParamFallback } = await load();
+ const endpoint = nextEndpoint();
+ await fetchWithParamFallback(
+ endpoint,
+ fetcher(
+ jsonResponse(400, { error: "Unrecognized request argument: chat_template_kwargs" }),
+ jsonResponse(200, {})
+ ),
+ { model: "m", reasoning_effort: "none", chat_template_kwargs: { enable_thinking: false } },
+ () => {}
+ );
+
+ const body = {
+ model: "m",
+ reasoning_effort: "none",
+ chat_template_kwargs: { enable_thinking: false },
+ };
+ const doFetch = fetcher(jsonResponse(200, {}));
+ const logged = [];
+ await fetchWithParamFallback(endpoint, doFetch, body, (d) => logged.push(d));
+ assert.equal(doFetch.count(), 1);
+ assert.deepEqual(logged, [{ stripped: ["chat_template_kwargs"], remembered: true }]);
+ assert.ok(!("chat_template_kwargs" in body));
+ assert.equal(body.reasoning_effort, "none", "only the rejected param is left out");
+});
+
+test("a rejection that doesn't name reasoning is not remembered", async () => {
+ // It may have come from one of several backends behind the same endpoint.
+ const { fetchWithParamFallback } = await load();
+ const endpoint = nextEndpoint();
+ await fetchWithParamFallback(
+ endpoint,
+ fetcher(jsonResponse(400, { error: "bad request" }), jsonResponse(200, {})),
+ { model: "m", reasoning: { effort: "none" } },
+ () => {}
+ );
+
+ const body = { model: "m", reasoning: { effort: "none" } };
+ await fetchWithParamFallback(endpoint, fetcher(jsonResponse(200, {})), body, () => {});
+ assert.ok("reasoning" in body);
+});
+
+test("an error naming Reasoning in any case is remembered", async () => {
+ const { fetchWithParamFallback } = await load();
+ const endpoint = nextEndpoint();
+ await fetchWithParamFallback(
+ endpoint,
+ fetcher(
+ jsonResponse(400, { error: "Reasoning is not supported for this model" }),
+ jsonResponse(200, {})
+ ),
+ { model: "m", reasoning: { effort: "none" } },
+ () => {}
+ );
+
+ const body = { model: "m", reasoning: { effort: "none" } };
+ const doFetch = fetcher(jsonResponse(200, {}));
+ await fetchWithParamFallback(endpoint, doFetch, body, () => {});
+ assert.ok(!("reasoning" in body));
+ assert.equal(doFetch.count(), 1);
+});
+
+test("OpenRouter rejections are never remembered: each request may reach another upstream", async () => {
+ const { fetchWithParamFallback } = await load();
+ const endpoint = "https://openrouter.ai/api/v1/chat/completions";
+ await fetchWithParamFallback(
+ endpoint,
+ fetcher(jsonResponse(400, REASONING_REJECTED), jsonResponse(200, {})),
+ { model: "m", reasoning: { effort: "high" } },
+ () => {}
+ );
+
+ const body = { model: "m", reasoning: { effort: "high" } };
+ await fetchWithParamFallback(endpoint, fetcher(jsonResponse(200, {})), body, () => {});
+ assert.deepEqual(body.reasoning, { effort: "high" });
+});
+
+test("a remembered rejection expires after ten minutes", async (t) => {
+ const { fetchWithParamFallback } = await load();
+ t.mock.timers.enable({ apis: ["Date"], now: 1_000_000 });
+ const endpoint = nextEndpoint();
+ await fetchWithParamFallback(
+ endpoint,
+ fetcher(jsonResponse(400, REASONING_REJECTED), jsonResponse(200, {})),
+ { model: "m", reasoning: { effort: "none" } },
+ () => {}
+ );
+
+ t.mock.timers.tick(10 * 60 * 1000 - 1);
+ const stillRemembered = { model: "m", reasoning: { effort: "none" } };
+ await fetchWithParamFallback(endpoint, fetcher(jsonResponse(200, {})), stillRemembered, () => {});
+ assert.ok(!("reasoning" in stillRemembered));
+
+ t.mock.timers.tick(1);
+ const expired = { model: "m", reasoning: { effort: "none" } };
+ await fetchWithParamFallback(endpoint, fetcher(jsonResponse(200, {})), expired, () => {});
+ assert.ok("reasoning" in expired);
+});
+
+test("nothing is remembered when the retry fails too", async () => {
+ const { fetchWithParamFallback } = await load();
+ const endpoint = nextEndpoint();
+ await fetchWithParamFallback(
+ endpoint,
+ fetcher(jsonResponse(400, REASONING_REJECTED)),
+ { model: "m", reasoning: { effort: "none" } },
+ () => {}
+ );
+
+ const body = { model: "m", reasoning: { effort: "none" } };
+ await fetchWithParamFallback(endpoint, fetcher(jsonResponse(200, {})), body, () => {});
+ assert.ok("reasoning" in body);
+});
diff --git a/test/services/llmRequestShapeMatrix.test.js b/test/services/llmRequestShapeMatrix.test.js
index 8e7323d38..37f18de9a 100644
--- a/test/services/llmRequestShapeMatrix.test.js
+++ b/test/services/llmRequestShapeMatrix.test.js
@@ -89,6 +89,30 @@ const MATRIX = [
["corti keeps the legacy chat-completions shape", "Corti", "corti-s1", null, AGENT,
{ max_tokens: MAX_TOKENS, temperature: 0.3 }],
+ // --- Chosen reasoning effort (the per-scope picker) ---
+ ["groq gpt-oss: a chosen effort replaces the cleanup pin", "Groq", "openai/gpt-oss-120b", null, { disableThinking: false, reasoningEffort: "high" },
+ { max_tokens: MAX_TOKENS, temperature: 0, reasoning_effort: "high" }],
+ ["groq gpt-oss: suppression beats a chosen effort", "Groq", "openai/gpt-oss-120b", null, { disableThinking: true, reasoningEffort: "high" },
+ { max_tokens: MAX_TOKENS, temperature: 0, reasoning_effort: "low" }],
+ ["groq gpt-oss: a level the model doesn't offer is ignored", "Groq", "openai/gpt-oss-120b", null, { ...AGENT, disableThinking: false, reasoningEffort: "none" },
+ { max_tokens: MAX_TOKENS, temperature: 0.3 }],
+ ["custom endpoint sends a chosen effort as reasoning_effort (#791)", "custom", "grok-4.3", "https://api.x.ai/v1", { ...AGENT, disableThinking: false, reasoningEffort: "low" },
+ { max_completion_tokens: MAX_TOKENS, reasoning_effort: "low" }],
+ ["custom x.ai: a level its host doesn't take is never sent", "custom", "grok-4.7", "https://api.x.ai/v1", { ...AGENT, disableThinking: false, reasoningEffort: "minimal" },
+ { max_completion_tokens: MAX_TOKENS }],
+ ["custom deepseek: none is never sent (#1260)", "custom", "deepseek-chat", "https://api.deepseek.com", { ...AGENT, disableThinking: false, reasoningEffort: "none" },
+ { max_tokens: MAX_TOKENS, temperature: 0.3 }],
+ ["custom with a registry model: a level the registry doesn't list is never sent", "custom", "gpt-5.5", "https://custom.example/v1", { ...AGENT, disableThinking: false, reasoningEffort: "minimal" },
+ { max_completion_tokens: MAX_TOKENS }],
+ ["openrouter puts a chosen effort in its reasoning object", "openrouter", "openai/gpt-4o", null, { ...AGENT, disableThinking: false, reasoningEffort: "high" },
+ { max_tokens: MAX_TOKENS, temperature: 0.3, reasoning: { effort: "high" } }],
+ ["openrouter glm-5.3: a chosen effort sends no family default beside it", "openrouter", "z-ai/glm-5.3", null, { ...AGENT, disableThinking: false, reasoningEffort: "low" },
+ { max_tokens: MAX_TOKENS, temperature: 0.3, reasoning: { effort: "low" } }],
+ ["custom gpt-oss: a level its family rejects is never sent (#1611)", "custom", "gpt-oss-120b", "https://custom.example/v1", { ...AGENT, disableThinking: false, reasoningEffort: "none" },
+ { max_tokens: MAX_TOKENS, temperature: 0.3 }],
+ ["self-hosted never sends a level saved for a cloud provider", "lan", "openai/gpt-oss-120b", null, { disableThinking: false, reasoningEffort: "high" },
+ { max_tokens: MAX_TOKENS, temperature: 0, reasoning_effort: "low" }],
+
// --- Gemini through its OpenAI-compat endpoint (agent streaming route) ---
["gemini openai-compat: fallback config, no temperature param", "gemini", "gemini-3-flash-preview", null, AGENT,
{ max_completion_tokens: MAX_TOKENS }],
diff --git a/test/services/modelFamilyConstraints.test.js b/test/services/modelFamilyConstraints.test.js
index d05d0163d..77899fdf9 100644
--- a/test/services/modelFamilyConstraints.test.js
+++ b/test/services/modelFamilyConstraints.test.js
@@ -21,7 +21,11 @@ test("unknown, empty, and missing ids resolve to no constraints", async () => {
test("gpt-oss has no reasoning off switch: suppress and cleanup both floor at low", async () => {
const { getModelFamilyConstraints } = await load();
const effort = getModelFamilyConstraints("gpt-oss-120b")?.reasoningEffort;
- assert.deepEqual(effort, { suppressValue: "low", cleanupValue: "low" });
+ assert.deepEqual(effort, {
+ suppressValue: "low",
+ cleanupValue: "low",
+ options: ["low", "medium", "high"],
+ });
});
test("glm-5.3 always reasons: suppress floors at low, high stands in for the server's max", async () => {
@@ -29,6 +33,7 @@ test("glm-5.3 always reasons: suppress floors at low, high stands in for the ser
assert.deepEqual(getModelFamilyConstraints("glm-5.3")?.reasoningEffort, {
suppressValue: "low",
defaultValue: "high",
+ options: ["low", "high"],
});
assert.equal(getModelFamilyConstraints("glm-5.3-flash")?.family, "glm-5.3");
assert.equal(getModelFamilyConstraints("glm-5-3")?.family, "glm-5.3");
diff --git a/test/services/openaiReasoningOutput.test.js b/test/services/openaiReasoningOutput.test.js
index 526a0730b..09542389b 100644
--- a/test/services/openaiReasoningOutput.test.js
+++ b/test/services/openaiReasoningOutput.test.js
@@ -197,13 +197,72 @@ test("Responses requests carry the other families' effort facts too", async (t)
assert.deepEqual(bodies[1].reasoning, { effort: "none" });
});
-test("a model outside the family table gets no reasoning param on a Responses endpoint", async (t) => {
+test("a model outside the family table gets the generic off value on a Responses endpoint", async (t) => {
const bodies = installFetch(t, () => completeMessage("# Notes"));
const { openaiProvider } = await load();
await openaiProvider.call({ ...glmNotesCall({ disableThinking: true }), model: "glm-5.2" });
+ await openaiProvider.call({ ...glmNotesCall({ disableThinking: false }), model: "glm-5.2" });
- assert.equal(bodies[0].reasoning, undefined);
+ // "none", as the chat body's generic dialect sends (#791); thinking on leaves the default.
+ assert.deepEqual(bodies[0].reasoning, { effort: "none" });
+ assert.equal(bodies[1].reasoning, undefined);
+});
+
+test("a chosen level on a Responses endpoint is checked against the endpoint's host", async (t) => {
+ const bodies = installFetch(t, () => completeMessage("Cleaned."));
+ const { openaiProvider } = await load();
+ const xaiCall = (reasoningEffort) => ({
+ text: "raw transcript",
+ model: "grok-4.7",
+ agentName: null,
+ config: {
+ provider: "custom",
+ baseUrl: "https://api.x.ai/v1",
+ customApiKey: "test-key",
+ disableThinking: false,
+ reasoningEffort,
+ },
+ ctx: makeCtx(),
+ });
+
+ await openaiProvider.call(xaiCall("xhigh"));
+ // x.ai takes no "minimal", so it's left at the server's default.
+ await openaiProvider.call(xaiCall("minimal"));
+
+ assert.deepEqual(bodies[0].reasoning, { effort: "xhigh" });
+ assert.equal(bodies[1].reasoning, undefined);
+});
+
+test("a Custom endpoint that rejects the off value pays the 400 once, not on every request", async (t) => {
+ // e.g. Azure OpenAI with a non-reasoning deployment, whose name we can't recognise.
+ const bodies = installFetch(t, (body) =>
+ body.reasoning
+ ? jsonResponse({ error: { message: "Unsupported parameter: 'reasoning.effort'" } }, 400)
+ : completeMessage("Cleaned.")
+ );
+ const { openaiProvider } = await load();
+ const cleanup = () =>
+ openaiProvider.call({
+ text: "raw transcript",
+ model: "gpt-4o-prod",
+ agentName: null,
+ config: {
+ provider: "custom",
+ baseUrl: "https://rejecting-host.example/v1",
+ customApiKey: "test-key",
+ disableThinking: true,
+ },
+ ctx: makeCtx(),
+ });
+
+ assert.equal(await cleanup(), "Cleaned.");
+ assert.equal(await cleanup(), "Cleaned.");
+
+ assert.deepEqual(
+ bodies.map((body) => body.reasoning),
+ [{ effort: "none" }, undefined, undefined]
+ );
});
test("a Responses result that ran out of tokens before any text is an error, not the input echoed back", async (t) => {
diff --git a/test/services/reasoningEffortRequests.test.js b/test/services/reasoningEffortRequests.test.js
new file mode 100644
index 000000000..e81dc2428
--- /dev/null
+++ b/test/services/reasoningEffortRequests.test.js
@@ -0,0 +1,344 @@
+const test = require("node:test");
+const assert = require("node:assert/strict");
+const path = require("node:path");
+const { createRendererServer, installBrowserGlobals } = require("../lib/rendererTestHarness");
+
+// The per-scope reasoning effort must reach every transport a scope can use:
+// the Responses API body, Gemini's native thinkingConfig, and the AI SDK
+// streaming requests behind the assistant panel and typed chat.
+
+const CHOSEN = (reasoningEffort, extra = {}) => ({
+ disableThinking: false,
+ reasoningEffort,
+ ...extra,
+});
+
+test("picker: offered for a registry model even with the toggle at its default", async () => {
+ const { selectableReasoningEffort } = await import("../../src/services/ai/chatRequestBody.ts");
+
+ assert.deepEqual(selectableReasoningEffort("openai", "gpt-5.5", { disableThinking: true }), {
+ options: ["none", "low", "medium", "high", "xhigh"],
+ defaultValue: "medium",
+ });
+ assert.equal(selectableReasoningEffort("openai", "gpt-4.1", {}), null);
+});
+
+test("picker: hidden for a Custom model while thinking is disabled", async () => {
+ const { selectableReasoningEffort } = await import("../../src/services/ai/chatRequestBody.ts");
+
+ assert.equal(selectableReasoningEffort("custom", "grok-4.3", { disableThinking: true }), null);
+ assert.deepEqual(selectableReasoningEffort("custom", "grok-4.3", { disableThinking: false }), {
+ options: ["none", "minimal", "low", "medium", "high", "xhigh"],
+ });
+});
+
+test("picker: Custom and OpenRouter offer only what the model family accepts", async () => {
+ const { selectableReasoningEffort } = await import("../../src/services/ai/chatRequestBody.ts");
+ const on = { disableThinking: false };
+
+ assert.deepEqual(selectableReasoningEffort("custom", "gpt-oss-20b", on).options, [
+ "low",
+ "medium",
+ "high",
+ ]);
+ assert.deepEqual(selectableReasoningEffort("openrouter", "z-ai/glm-5.3", on).options, [
+ "low",
+ "high",
+ ]);
+ assert.equal(selectableReasoningEffort("custom", "qwen3:32b", on), null);
+ assert.equal(selectableReasoningEffort("custom", "magistral-small-latest", on), null);
+});
+
+test("picker: Custom offers the levels its endpoint's host takes", async () => {
+ const { selectableReasoningEffort } = await import("../../src/services/ai/chatRequestBody.ts");
+ const on = { disableThinking: false };
+
+ assert.deepEqual(
+ selectableReasoningEffort("custom", "grok-4.7", on, "https://api.x.ai/v1").options,
+ ["low", "medium", "high", "xhigh"]
+ );
+ // DeepSeek's API rejects "none" (#1260).
+ assert.deepEqual(
+ selectableReasoningEffort("custom", "deepseek-chat", on, "https://api.deepseek.com").options,
+ ["low", "high"]
+ );
+});
+
+test("picker: Custom and OpenRouter take a registry model's own levels", async () => {
+ const { selectableReasoningEffort } = await import("../../src/services/ai/chatRequestBody.ts");
+ const on = { disableThinking: false };
+
+ // No "minimal" for gpt-5.5, and no server default named for an unknown server.
+ assert.deepEqual(selectableReasoningEffort("custom", "gpt-5.5", on), {
+ options: ["none", "low", "medium", "high", "xhigh"],
+ });
+ // A streamed Custom turn can't carry "max"; OpenRouter's reasoning object can.
+ assert.deepEqual(selectableReasoningEffort("custom", "gpt-6-astra", on).options, [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ ]);
+ assert.deepEqual(selectableReasoningEffort("openrouter", "openai/gpt-6-astra", on).options, [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max",
+ ]);
+});
+
+test("picker: the toggle Custom always shows hides the picker, registry model or not", async () => {
+ const { chosenReasoningEffort, selectableReasoningEffort } =
+ await import("../../src/services/ai/chatRequestBody.ts");
+
+ assert.equal(selectableReasoningEffort("custom", "gpt-5.5", { disableThinking: true }), null);
+ assert.equal(
+ chosenReasoningEffort("custom", "gpt-5.5", { disableThinking: true, reasoningEffort: "high" }),
+ undefined
+ );
+});
+
+test("a level saved for a cloud provider never reaches a self-hosted server", async () => {
+ const { chosenReasoningEffort, selectableReasoningEffort } =
+ await import("../../src/services/ai/chatRequestBody.ts");
+ const config = { disableThinking: false, reasoningEffort: "high" };
+
+ assert.equal(chosenReasoningEffort("groq", "openai/gpt-oss-120b", config), "high");
+ assert.equal(selectableReasoningEffort("lan", "openai/gpt-oss-120b", config), null);
+ assert.equal(chosenReasoningEffort("lan", "openai/gpt-oss-120b", config), undefined);
+});
+
+test("Responses effort: OpenAI sends the chosen level the model offers", async () => {
+ const { responsesReasoningEffort } = await import("../../src/services/ai/chatRequestBody.ts");
+
+ assert.equal(responsesReasoningEffort("gpt-5.5", "openai", CHOSEN("high")), "high");
+ assert.equal(responsesReasoningEffort("gpt-6-astra", "openai", CHOSEN("max")), "max");
+ // GPT-6 Astra rejects "none", so a stored "none" from another model is dropped.
+ assert.equal(responsesReasoningEffort("gpt-6-astra", "openai", CHOSEN("none")), undefined);
+});
+
+test("Responses effort: Auto leaves OpenAI at the server default, toggle or not", async () => {
+ const { responsesReasoningEffort } = await import("../../src/services/ai/chatRequestBody.ts");
+
+ assert.equal(responsesReasoningEffort("gpt-5.5", "openai", {}), undefined);
+ assert.equal(responsesReasoningEffort("gpt-5.5", "openai", { disableThinking: true }), undefined);
+ // An id the registry no longer lists: OpenAI hides the toggle, so its
+ // default-on value never turns into "none".
+ assert.equal(responsesReasoningEffort("gpt-5.4", "openai", { disableThinking: true }), undefined);
+ assert.equal(responsesReasoningEffort("o4-mini", "openai", { disableThinking: true }), undefined);
+});
+
+test("Responses effort: a Custom endpoint sends any listed level (#791)", async () => {
+ const { responsesReasoningEffort } = await import("../../src/services/ai/chatRequestBody.ts");
+
+ assert.equal(responsesReasoningEffort("grok-4.3", "custom", CHOSEN("none")), "none");
+ assert.equal(responsesReasoningEffort("grok-4.3", "custom", CHOSEN("max")), undefined);
+});
+
+test("Responses effort: the thinking toggle reaches a Custom endpoint's Responses API (#791)", async () => {
+ const { responsesReasoningEffort } = await import("../../src/services/ai/chatRequestBody.ts");
+
+ assert.equal(responsesReasoningEffort("grok-4.3", "custom", { disableThinking: true }), "none");
+ // A family without an off switch keeps its floor.
+ assert.equal(
+ responsesReasoningEffort("gpt-oss-120b", "custom", { disableThinking: true }),
+ "low"
+ );
+});
+
+test("Gemini native: a chosen level becomes thinkingLevel; the toggle still wins", async (t) => {
+ const { geminiProvider } = await import("../../src/services/ai/inferenceProviders/gemini.ts");
+ const requests = [];
+ t.mock.method(globalThis, "fetch", async (_url, init) => {
+ requests.push(JSON.parse(init.body));
+ return Response.json({
+ candidates: [{ finishReason: "STOP", content: { parts: [{ text: "Clean text." }] } }],
+ });
+ });
+ const call = (model, config) =>
+ geminiProvider.call({
+ text: "transcript",
+ model,
+ agentName: null,
+ config,
+ ctx: { getApiKey: async () => "test-key", getSystemPrompt: () => "Clean." },
+ });
+
+ await call("gemini-3.1-pro-preview", CHOSEN("medium"));
+ assert.deepEqual(requests[0].generationConfig.thinkingConfig, {
+ thinkingLevel: "medium",
+ includeThoughts: false,
+ });
+
+ await call("gemini-3-flash-preview", { disableThinking: true, reasoningEffort: "high" });
+ assert.deepEqual(requests[1].generationConfig.thinkingConfig, {
+ thinkingLevel: "minimal",
+ includeThoughts: false,
+ });
+});
+
+test("Gemini 3.1 Flash Lite: no toggle, so a chosen level beats the minimal floor", async (t) => {
+ const { selectableReasoningEffort } = await import("../../src/services/ai/chatRequestBody.ts");
+ const { geminiProvider } = await import("../../src/services/ai/inferenceProviders/gemini.ts");
+ const requests = [];
+ t.mock.method(globalThis, "fetch", async (_url, init) => {
+ requests.push(JSON.parse(init.body));
+ return Response.json({
+ candidates: [{ finishReason: "STOP", content: { parts: [{ text: "Clean text." }] } }],
+ });
+ });
+ const call = (config) =>
+ geminiProvider.call({
+ text: "transcript",
+ model: "gemini-3.1-flash-lite",
+ agentName: null,
+ config,
+ ctx: { getApiKey: async () => "test-key", getSystemPrompt: () => "Clean." },
+ });
+
+ // The toggle is hidden for this model and stays at its default, on.
+ assert.deepEqual(
+ selectableReasoningEffort("gemini", "gemini-3.1-flash-lite", { disableThinking: true }),
+ { options: ["minimal", "low", "medium", "high"] }
+ );
+
+ await call({ disableThinking: true, reasoningEffort: "low" });
+ assert.equal(requests[0].generationConfig.thinkingConfig.thinkingLevel, "low");
+
+ await call({ disableThinking: true });
+ assert.equal(requests[1].generationConfig.thinkingConfig.thinkingLevel, "minimal");
+});
+
+async function streamRequestBody(t, { provider, model, config, tools, keys = {} }) {
+ installBrowserGlobals(t, {
+ window: {
+ electronAPI: {
+ getOpenAIKey: async () => keys.openai ?? "test-key",
+ getGroqKey: async () => "test-key",
+ getGeminiKey: async () => "test-key",
+ getOpenrouterKey: async () => "test-key",
+ },
+ },
+ });
+ const vite = await createRendererServer(t, { cachePrefix: `openwhispr-effort-${provider}-` });
+ const reasoningService = (await vite.ssrLoadModule("/services/ReasoningService.ts")).default;
+ const { usePolicyStore } = await vite.ssrLoadModule("/stores/policyStore.ts");
+ usePolicyStore.setState({ status: "unmanaged", appVersion: "1.8.3", policy: null });
+ t.after(() => reasoningService.destroy());
+
+ const bodies = [];
+ const originalFetch = globalThis.fetch;
+ t.after(() => {
+ globalThis.fetch = originalFetch;
+ });
+ globalThis.fetch = async (_url, init) => {
+ bodies.push(JSON.parse(init.body));
+ return new Response(JSON.stringify({ error: { message: "stop here" } }), {
+ status: 400,
+ headers: { "content-type": "application/json" },
+ });
+ };
+
+ const stream = reasoningService.processTextStreamingAI(
+ [{ role: "user", content: "hello" }],
+ model,
+ provider,
+ { systemPrompt: "Answer the user.", ...config },
+ tools
+ );
+ await assert.rejects(async () => {
+ for await (const _chunk of stream);
+ });
+ return bodies[0];
+}
+
+test("streaming: a Custom endpoint gets reasoning_effort", async (t) => {
+ const body = await streamRequestBody(t, {
+ provider: "custom",
+ model: "grok-4.3",
+ config: CHOSEN("low", { baseUrl: "https://api.x.ai/v1", customApiKey: "test-key" }),
+ });
+ assert.equal(body.reasoning_effort, "low");
+});
+
+test("streaming: a Custom level is checked against the endpoint's host", async (t) => {
+ const body = await streamRequestBody(t, {
+ provider: "custom",
+ model: "deepseek-chat",
+ // DeepSeek's API rejects "none" (#1260), and a streamed turn has no retry.
+ config: CHOSEN("none", { baseUrl: "https://api.deepseek.com", customApiKey: "test-key" }),
+ });
+ assert.equal(body.reasoning_effort, undefined);
+});
+
+test("streaming: a self-hosted turn never sends the level saved for the scope's provider", async (t) => {
+ const body = await streamRequestBody(t, {
+ provider: "custom",
+ model: "grok-4.3",
+ config: CHOSEN("low", { lanUrl: "http://127.0.0.1:11434/v1" }),
+ tools: {},
+ });
+ assert.equal(body.reasoning_effort, undefined);
+});
+
+test("streaming: native OpenAI sends the effort even for ids the SDK doesn't know", async (t) => {
+ const body = await streamRequestBody(t, {
+ provider: "openai",
+ model: "gpt-6-astra",
+ config: CHOSEN("high"),
+ });
+ assert.deepEqual(body.reasoning, { effort: "high" });
+});
+
+test("streaming: OpenRouter gets its reasoning object", async (t) => {
+ const body = await streamRequestBody(t, {
+ provider: "openrouter",
+ model: "openai/gpt-5.5",
+ config: CHOSEN("high"),
+ });
+ assert.deepEqual(body.reasoning, { effort: "high" });
+});
+
+test("streaming: Groq gets the chosen level for gpt-oss", async (t) => {
+ const body = await streamRequestBody(t, {
+ provider: "groq",
+ model: "openai/gpt-oss-120b",
+ config: CHOSEN("high"),
+ });
+ assert.equal(body.reasoning_effort, "high");
+});
+
+test("streaming: Gemini gets the chosen thinkingLevel", async (t) => {
+ const body = await streamRequestBody(t, {
+ provider: "gemini",
+ model: "gemini-3.1-pro-preview",
+ config: CHOSEN("low"),
+ });
+ assert.deepEqual(body.generationConfig.thinkingConfig, {
+ thinkingLevel: "low",
+ includeThoughts: false,
+ });
+});
+
+test("settings: each scope keeps its own effort", async (t) => {
+ installBrowserGlobals(t, {
+ initialStorage: { _llmScopeKeysMigrated: "1", _dictationAgentSeeded: "1" },
+ });
+ const vite = await createRendererServer(t, {
+ cachePrefix: "openwhispr-effort-settings-",
+ resolveAlias: { "@": path.resolve(__dirname, "../../src") },
+ });
+ await vite.ssrLoadModule("/models/ModelRegistry.ts");
+ const s = await vite.ssrLoadModule("/stores/settingsStore.ts");
+
+ s.setResolvedLLMConfig("dictationCleanup", { reasoningEffort: "low" });
+ s.setResolvedLLMConfig("chatIntelligence", { reasoningEffort: "high" });
+ const state = s.useSettingsStore.getState();
+
+ assert.equal(localStorage.getItem("cleanupReasoningEffort"), "low");
+ assert.equal(s.selectResolvedLLMConfig(state, "dictationCleanup").reasoningEffort, "low");
+ assert.equal(s.selectResolvedLLMConfig(state, "chatIntelligence").reasoningEffort, "high");
+ // Note formatting inherits cleanup's model, not its effort.
+ assert.equal(s.selectResolvedLLMConfig(state, "noteFormatting").reasoningEffort, undefined);
+});
diff --git a/test/stores/actionProcessingChunked.test.js b/test/stores/actionProcessingChunked.test.js
index fe0d32578..9dbce6ef2 100644
--- a/test/stores/actionProcessingChunked.test.js
+++ b/test/stores/actionProcessingChunked.test.js
@@ -711,12 +711,13 @@ test("a transcript without speaker labels is packed without a prose prefix", asy
test("parts never spend their output budget on thinking", async (t) => {
const { store, calls, updates } = await loadStore(t, {
failFirst: true,
- storage: { noteFormattingDisableThinking: "false" },
+ storage: { noteFormattingDisableThinking: "false", noteFormattingReasoningEffort: "high" },
});
run(store, 22, longMaterial(400));
await waitForResult(store, updates);
assert.equal(updates.length, 1);
assert.equal(calls[0].config.disableThinking, false);
+ assert.equal(calls[0].config.reasoningEffort, "high");
assert.equal(calls.at(-1).config.disableThinking, false);
for (const call of calls.slice(1, -1)) assert.equal(call.config.disableThinking, true);
});