fix(ai): make custom output token budget configurable

apps/desktop/src/components/editor/EditorSettingsDialog.vue

@@ -56,6 +56,8 @@ import {
   useSettingsStore,
   AI_PROVIDER_PRESETS,
   AI_PROVIDER_PARTNER_PRESETS,
+  AI_OUTPUT_TOKENS_MAX,
+  AI_OUTPUT_TOKENS_MIN,
   EDITOR_THEMES,
   DEFAULT_EDITOR_SETTINGS,
   DEFAULT_DESKTOP_SETTINGS,
@@ -3524,6 +3526,7 @@ const aiEditProxyEnabled = ref(false);
 const aiEditProxyUrl = ref("");
 const aiEditEnableThinking = ref(true);
 const aiEditReasoningLevel = ref<AiReasoningLevel>("default");
+const aiEditMaxOutputTokens = ref<number | undefined>(undefined);
 const aiEditContextWindow = ref<number | undefined>(undefined);
 const aiEditCodexCliPath = ref("");
 const aiEditCodexCliEnvRows = ref<AiEnvRow[]>([]);
@@ -3808,6 +3811,7 @@ function currentAiEditConfig() {
     proxyUrl: aiEditProxyUrl.value,
     enableThinking: aiEditEnableThinking.value,
     reasoningLevel: aiEditReasoningLevel.value,
+    maxOutputTokens: aiEditMaxOutputTokens.value || undefined,
     contextWindow: aiEditContextWindow.value || undefined,
     codexCliPath: aiEditCodexCliPath.value.trim() || undefined,
     codexCliEnv: aiIsCodexCli.value ? cliEnvFromRows(aiEditCodexCliEnvRows.value) : {},
@@ -3894,6 +3898,7 @@ function aiEnterEditMode(configId?: string) {
       aiEditProxyUrl.value = config.proxyUrl ?? "";
       aiEditEnableThinking.value = config.enableThinking ?? true;
       aiEditReasoningLevel.value = config.reasoningLevel ?? "default";
+      aiEditMaxOutputTokens.value = config.maxOutputTokens;
       aiEditContextWindow.value = config.contextWindow;
       aiEditCodexCliPath.value = config.codexCliPath ?? "";
       aiEditCodexCliEnvRows.value = aiEnvRowsFromConfig(config.codexCliEnv);
@@ -3927,6 +3932,7 @@ function aiEnterEditMode(configId?: string) {
     aiEditProxyUrl.value = "";
     aiEditEnableThinking.value = true;
     aiEditReasoningLevel.value = "default";
+    aiEditMaxOutputTokens.value = undefined;
     aiEditContextWindow.value = undefined;
     aiEditCodexCliPath.value = "";
     aiEditCodexCliEnvRows.value = [];
@@ -7735,6 +7741,16 @@ onUnmounted(() => {
                     </p>
                   </div>
                 </div>
+                <!-- Maximum Output Tokens -->
+                <div v-if="!aiIsCliProvider" class="grid grid-cols-3 items-start gap-3">
+                  <Label class="text-right text-xs">{{ t("ai.maxOutputTokens") }}</Label>
+                  <div class="col-span-2">
+                    <Input v-model.number="aiEditMaxOutputTokens" type="number" :min="AI_OUTPUT_TOKENS_MIN" :max="AI_OUTPUT_TOKENS_MAX" step="1000" class="h-8 text-xs" :placeholder="t('ai.maxOutputTokensAuto')" />
+                    <p class="mt-1 text-xs text-muted-foreground">
+                      {{ t("ai.maxOutputTokensHint", { min: AI_OUTPUT_TOKENS_MIN, max: AI_OUTPUT_TOKENS_MAX }) }}
+                    </p>
+                  </div>
+                </div>
 
                 <!-- Proxy -->
                 <div v-if="!aiIsCliProvider" class="grid grid-cols-3 items-center gap-3">

apps/desktop/src/i18n/locales/en.ts

@@ -2777,6 +2777,9 @@ export default {
     contextWindow: "Context Window",
     contextWindowAuto: "Auto (detect from model name)",
     contextWindowHint: "Tokens. Leave empty to auto-detect. Set manually for local/custom models.",
+    maxOutputTokens: "Max Output Tokens",
+    maxOutputTokensAuto: "Auto (per action)",
+    maxOutputTokensHint: "Optional per-configuration limit sent to the provider. Range: {min}–{max}. Thinking requests still use at least 8192 tokens.",
     maxRetries: "Max Retries",
     maxRetriesHint: "Retry on rate limits, timeouts, and temporary network errors. 0 = never, max 10.",
     maxRetriesGlobal: "Max Retries",

apps/desktop/src/i18n/locales/es.ts

@@ -2847,6 +2847,9 @@ export default withEnglishFallback({
     contextWindow: "Ventana de contexto",
     contextWindowAuto: "Auto (detectar desde el nombre del modelo)",
     contextWindowHint: "Tokens. Déjalo vacío para detección automática. Configúralo manualmente para modelos locales/personalizados.",
+    maxOutputTokens: "Tokens máximos de salida",
+    maxOutputTokensAuto: "Auto (por acción)",
+    maxOutputTokensHint: "Límite opcional por configuración enviado al proveedor. Rango: {min}–{max}. Las solicitudes con razonamiento usan al menos 8192 tokens.",
     maxRetriesGlobal: "Reintentos máximos",
     maxRetriesGlobalDescription: "Número de reintentos automáticos en errores transitorios de API (límite de tasa, timeout, fallos de red). Se aplica a todos los proveedores de IA por API. Los proveedores CLI no se ven afectados.",
     maxRetriesRange: "{min}–{max}, predeterminado {default}",

apps/desktop/src/i18n/locales/it.ts

@@ -2741,6 +2741,9 @@ export default withEnglishFallback({
     contextWindow: "Finestra di Contesto",
     contextWindowAuto: "Auto (rileva dal nome del modello)",
     contextWindowHint: "Token. Lascia vuoto per il rilevamento automatico. Imposta manualmente per modelli locali/personalizzati.",
+    maxOutputTokens: "Token di output massimi",
+    maxOutputTokensAuto: "Automatico (per azione)",
+    maxOutputTokensHint: "Limite opzionale per configurazione inviato al provider. Intervallo: {min}–{max}. Le richieste con ragionamento usano comunque almeno 8192 token.",
     maxRetriesGlobal: "Tentativi massimi",
     maxRetriesGlobalDescription: "Numero di tentativi automatici su errori API transitori (limite di velocità, timeout, problemi di rete). Si applica a tutti i provider IA via API. I provider CLI non sono interessati.",
     maxRetriesRange: "{min}–{max}, predefinito {default}",

apps/desktop/src/i18n/locales/ja.ts

@@ -2541,6 +2541,9 @@ export default withEnglishFallback({
     contextWindow: "コンテキストウィンドウ",
     contextWindowAuto: "自動（モデル名から検出）",
     contextWindowHint: "トークン数。空欄で自動検出。ローカル/カスタムモデルは手動設定してください。",
+    maxOutputTokens: "最大出力トークン",
+    maxOutputTokensAuto: "自動（アクションごと）",
+    maxOutputTokensHint: "プロバイダーに送信する設定ごとの任意の出力上限。範囲: {min}–{max}。思考を有効にしたリクエストでは少なくとも8192トークンを使用します。",
     maxRetriesGlobal: "最大再試行回数",
     maxRetriesGlobalDescription: "一時的なAPIエラー（レート制限、タイムアウト、ネットワーク障害）発生時の自動再試行回数。すべてのAPI対応AIプロバイダーに適用されます。CLIプロバイダーは影響を受けません。",
     maxRetriesRange: "{min}–{max}、デフォルト {default}",

apps/desktop/src/i18n/locales/ko.ts

@@ -2592,6 +2592,9 @@ export default withEnglishFallback({
     contextWindow: "컨텍스트 창",
     contextWindowAuto: "자동 (모델 이름에서 감지)",
     contextWindowHint: "토큰. 자동 감지하려면 비워 두세요. 로컬/사용자 지정 모델은 수동으로 설정하세요.",
+    maxOutputTokens: "최대 출력 토큰",
+    maxOutputTokensAuto: "자동 (모델 기본값)",
+    maxOutputTokensHint: "공급자에게 전송되는 선택적 구성별 출력 토큰 제한입니다. 범위: {min}–{max}. 사고 요청은 여전히 최소 8192 토큰을 사용합니다.",
     codexCliPath: "Codex CLI 경로",
     codexCliPathHint: "PATH의 codex를 사용하려면 비워 두세요. codex login으로 별도로 로그인하세요.",
     codexCliPathEnvError: "Codex CLI 경로에는 실행 파일 경로만 포함해야 합니다. 환경 변수는 아래에 추가하세요.",

apps/desktop/src/i18n/locales/pt-BR.ts

@@ -2845,6 +2845,9 @@ export default withEnglishFallback({
     contextWindow: "Janela de Contexto",
     contextWindowAuto: "Automático (detectar pelo nome do modelo)",
     contextWindowHint: "Tokens. Deixe vazio para detecção automática. Defina manualmente para modelos locais/personalizados.",
+    maxOutputTokens: "Máximo de tokens de saída",
+    maxOutputTokensAuto: "Automático (por ação)",
+    maxOutputTokensHint: "Limite opcional por configuração enviado ao provedor. Intervalo: {min}–{max}. Solicitações com raciocínio ainda usam pelo menos 8192 tokens.",
     maxRetriesGlobal: "Máximo de Tentativas",
     maxRetriesGlobalDescription: "Número de tentativas automáticas em erros transitórios de API (limite de taxa, timeout, falhas de rede). Aplica-se a todos os provedores de IA via API. Provedores CLI não são afetados.",
     maxRetriesRange: "{min}–{max}, padrão {default}",

apps/desktop/src/i18n/locales/zh-CN.ts

@@ -2700,6 +2700,9 @@ export default withEnglishFallback({
     contextWindow: "上下文窗口",
     contextWindowAuto: "自动（根据模型名称推断）",
     contextWindowHint: "单位：token。留空自动推断，本地/自定义模型建议手动设置。",
+    maxOutputTokens: "最大输出 Token 数",
+    maxOutputTokensAuto: "自动（按操作类型）",
+    maxOutputTokensHint: "可选的单配置输出上限，会发送给 AI 服务商。范围：{min}–{max}。启用思考时仍至少使用 8192 token。",
     maxRetries: "最大重试次数",
     maxRetriesHint: "遇到限流、超时或临时网络错误时自动重试。0 = 不重试，最大 10。",
     maxRetriesGlobal: "最大重试次数",

apps/desktop/src/i18n/locales/zh-TW.ts

@@ -2740,6 +2740,9 @@ export default withEnglishFallback({
     contextWindow: "Context Window",
     contextWindowAuto: "自動（從模型名稱偵測）",
     contextWindowHint: "Token 數。留空以自動偵測。本機/自訂模型請手動設定。",
+    maxOutputTokens: "最大輸出 Token 數",
+    maxOutputTokensAuto: "自動（依操作類型）",
+    maxOutputTokensHint: "可選的單一設定輸出上限，會傳送給 AI 服務商。範圍：{min}–{max}。啟用思考時仍至少使用 8192 token。",
     maxRetries: "最大重試次數",
     maxRetriesHint: "遇到限流、超時或暫時性網路錯誤時自動重試。0 = 不重試，最大 10。",
     maxRetriesGlobal: "最大重試次數",

apps/desktop/src/lib/__tests__/ai/ai.spec.ts

@@ -158,6 +158,44 @@ describe("AI SQL dialect prompt", () => {
     expect(prompt).toContain("Ask mode");
   });
 
+  it("uses the configured output budget for non-thinking requests", () => {
+    const request = buildAgentRequest({
+      config: {
+        provider: "deepseek",
+        apiKey: "test",
+        apiUrl: "https://example.invalid",
+        model: "deepseek-v4-flash",
+        enableThinking: false,
+        maxOutputTokens: 1024,
+      },
+      action: "general",
+      mode: "ask",
+      instruction: "generate a long migration",
+      context: context(),
+    });
+
+    expect(request.maxTokens).toBe(1024);
+  });
+
+  it("uses the configured output budget above the thinking minimum", () => {
+    const request = buildAgentRequest({
+      config: {
+        provider: "deepseek",
+        apiKey: "test",
+        apiUrl: "https://example.invalid",
+        model: "deepseek-v4-flash",
+        enableThinking: true,
+        maxOutputTokens: 32768,
+      },
+      action: "general",
+      mode: "ask",
+      instruction: "generate a long migration",
+      context: context(),
+    });
+
+    expect(request.maxTokens).toBe(32768);
+  });
+
   it("MongoDB agent mode uses shell commands instead of SQL", () => {
     const prompt = buildSystemPrompt("general", context({ databaseType: "mongodb", connectionName: "MongoDB", database: "benchmark" }), "agent");
 

apps/desktop/src/lib/ai/ai.ts

@@ -179,7 +179,8 @@ export function buildAgentRequest(input: AiRequestInput, history?: api.AiMessage
   ];
 
   const params = actionParams(input.action);
-  const maxTokens = input.config.enableThinking ? Math.max(params.maxTokens, 8192) : params.maxTokens;
+  const baseMaxTokens = input.config.maxOutputTokens ?? params.maxTokens;
+  const maxTokens = input.config.enableThinking ? Math.max(baseMaxTokens, 8192) : baseMaxTokens;
   return { messages, systemPrompt, taskContract, maxTokens };
 }
 

apps/desktop/src/stores/__tests__/settingsStore.spec.ts

@@ -669,6 +669,12 @@ describe("settingsStore AI API key normalization", () => {
   it("trims API keys when normalizing loaded configurations", () => {
     expect(normalizeAiConfig({ provider: "openai", apiKey: "  secret  " }).apiKey).toBe("secret");
   });
+  it("preserves and bounds a configured maximum output token budget", () => {
+    expect(normalizeAiConfig({ provider: "deepseek", maxOutputTokens: 32768 }).maxOutputTokens).toBe(32768);
+    expect(normalizeAiConfig({ provider: "deepseek", maxOutputTokens: 1_500_000 }).maxOutputTokens).toBe(1_000_000);
+    expect(normalizeAiConfig({ provider: "deepseek", maxOutputTokens: 128 }).maxOutputTokens).toBeUndefined();
+    expect(normalizeAiConfig({ provider: "deepseek", maxOutputTokens: Number.NaN }).maxOutputTokens).toBeUndefined();
+  });
 
   it("provides Kimi defaults and recognizes legacy Kimi configurations", () => {
     expect(AI_PROVIDER_PRESETS.kimi).toMatchObject({

apps/desktop/src/stores/settingsStore.ts

@@ -458,6 +458,8 @@ const defaultConfigs: Record<AiProvider, Omit<AiConfig, "apiKey">> = Object.from
 ) as Record<AiProvider, Omit<AiConfig, "apiKey">>;
 
 const AI_REASONING_LEVELS: AiReasoningLevel[] = ["default", "minimal", "low", "medium", "high", "xhigh", "max"];
+export const AI_OUTPUT_TOKENS_MIN = 256;
+export const AI_OUTPUT_TOKENS_MAX = 1_000_000;
 const AI_ENV_KEY_RE = /^[A-Za-z_][A-Za-z0-9_]*$/;
 
 function normalizeAiReasoningLevel(value: unknown): AiReasoningLevel {
@@ -488,6 +490,8 @@ export function normalizeAiHeaders(value: unknown): Record<string, string> {
 
 export function normalizeAiConfig(config: Partial<AiConfig> | null | undefined): AiConfig {
   const provider = config?.provider && config.provider in AI_PROVIDER_PRESETS ? config.provider : inferAiProviderFromConfig(config);
+  const rawMaxOutputTokens = config?.maxOutputTokens;
+  const maxOutputTokens = typeof rawMaxOutputTokens === "number" && Number.isFinite(rawMaxOutputTokens) && rawMaxOutputTokens >= AI_OUTPUT_TOKENS_MIN ? Math.min(AI_OUTPUT_TOKENS_MAX, Math.round(rawMaxOutputTokens)) : undefined;
   return {
     ...defaultConfigs[provider],
     ...config,
@@ -500,6 +504,7 @@ export function normalizeAiConfig(config: Partial<AiConfig> | null | undefined):
     proxyUrl: config?.proxyUrl ?? "",
     enableThinking: config?.enableThinking ?? true,
     reasoningLevel: normalizeAiReasoningLevel(config?.reasoningLevel),
+    maxOutputTokens,
     contextWindow: config?.contextWindow ?? undefined,
     codexCliPath: config?.codexCliPath?.trim() || undefined,
     codexCliEnv: normalizeAiEnv(config?.codexCliEnv),

apps/desktop/src/types/ai.ts

@@ -42,6 +42,8 @@ export interface AiConfig {
   proxyUrl?: string;
   enableThinking?: boolean;
   reasoningLevel?: AiReasoningLevel;
+  /** Optional per-configuration output budget sent as max_tokens/max_output_tokens. */
+  maxOutputTokens?: number;
   contextWindow?: number;
   codexCliPath?: string | null;
   codexCliEnv?: Record<string, string>;