@ljoukov/llm 7.0.2 → 7.0.3

package/dist/index.d.cts

@@ -282,6 +282,12 @@ type LlmBaseRequest = {
     readonly imageAspectRatio?: string;
     readonly imageSize?: LlmImageSize;
     readonly thinkingLevel?: LlmThinkingLevel;
+    /**
+     * Gemini-only explicit thinking budget. When provided, this overrides
+     * `thinkingLevel` and the model default for Gemini requests. Use `0` to
+     * disable thinking where the selected Gemini model supports it.
+     */
+    readonly thinkingBudget?: number;
     readonly mediaResolution?: LlmMediaResolution;
     readonly openAiTextFormat?: ResponseTextConfig["format"];
     readonly telemetry?: TelemetrySelection;
@@ -429,6 +435,7 @@ type LlmToolLoopRequest = LlmInput & {
     readonly modelTools?: readonly LlmToolConfig[];
     readonly maxSteps?: number;
     readonly thinkingLevel?: LlmThinkingLevel;
+    readonly thinkingBudget?: number;
     readonly mediaResolution?: LlmMediaResolution;
     readonly steering?: LlmToolLoopSteeringChannel;
     readonly onEvent?: (event: LlmStreamEvent) => void;

package/dist/index.d.ts

@@ -282,6 +282,12 @@ type LlmBaseRequest = {
     readonly imageAspectRatio?: string;
     readonly imageSize?: LlmImageSize;
     readonly thinkingLevel?: LlmThinkingLevel;
+    /**
+     * Gemini-only explicit thinking budget. When provided, this overrides
+     * `thinkingLevel` and the model default for Gemini requests. Use `0` to
+     * disable thinking where the selected Gemini model supports it.
+     */
+    readonly thinkingBudget?: number;
     readonly mediaResolution?: LlmMediaResolution;
     readonly openAiTextFormat?: ResponseTextConfig["format"];
     readonly telemetry?: TelemetrySelection;
@@ -429,6 +435,7 @@ type LlmToolLoopRequest = LlmInput & {
     readonly modelTools?: readonly LlmToolConfig[];
     readonly maxSteps?: number;
     readonly thinkingLevel?: LlmThinkingLevel;
+    readonly thinkingBudget?: number;
     readonly mediaResolution?: LlmMediaResolution;
     readonly steering?: LlmToolLoopSteeringChannel;
     readonly onEvent?: (event: LlmStreamEvent) => void;

package/dist/index.js

@@ -7145,10 +7145,14 @@ function toGemini25ProThinkingBudget(thinkingLevel) {
       return 32768;
   }
 }
-function resolveGeminiThinkingConfig(modelId, thinkingLevel) {
+function resolveGeminiThinkingConfig(modelId, thinkingLevel, thinkingBudget) {
   if (isGeminiImageModelId(modelId) || modelId === "gemini-flash-lite-latest") {
     return void 0;
   }
+  if (thinkingBudget !== void 0) {
+    const normalizedBudget = Math.max(0, Math.floor(thinkingBudget));
+    return normalizedBudget === 0 ? { thinkingBudget: 0 } : { includeThoughts: true, thinkingBudget: normalizedBudget };
+  }
   if (thinkingLevel) {
     if (modelId === "gemini-2.5-pro") {
       return {
@@ -7524,6 +7528,7 @@ function startLlmCallLoggerFromContents(options) {
       ...options.request.imageAspectRatio ? { imageAspectRatio: options.request.imageAspectRatio } : {},
       ...options.request.imageSize ? { imageSize: options.request.imageSize } : {},
       ...options.request.thinkingLevel ? { thinkingLevel: options.request.thinkingLevel } : {},
+      ...options.request.thinkingBudget !== void 0 ? { thinkingBudget: options.request.thinkingBudget } : {},
       ...options.request.mediaResolution ? { mediaResolution: options.request.mediaResolution } : {},
       ...options.request.openAiTextFormat ? { openAiTextFormat: sanitiseLogValue(options.request.openAiTextFormat) } : {},
       ...getCurrentToolCallContext() ? { toolContext: getCurrentToolCallContext() } : {}
@@ -7825,7 +7830,11 @@ async function runTextCall(params) {
             })
           )
         );
-        const thinkingConfig = resolveGeminiThinkingConfig(modelForProvider, request.thinkingLevel);
+        const thinkingConfig = resolveGeminiThinkingConfig(
+          modelForProvider,
+          request.thinkingLevel,
+          request.thinkingBudget
+        );
         const mediaResolution = toGeminiMediaResolution(request.mediaResolution);
         const config = {
           maxOutputTokens: 32e3,
@@ -8132,6 +8141,7 @@ function startJsonStream(request, operation) {
                 responseMimeType: request.responseMimeType ?? "application/json",
                 responseJsonSchema,
                 thinkingLevel: request.thinkingLevel,
+                thinkingBudget: request.thinkingBudget,
                 ...openAiTextFormatForAttempt ? { openAiTextFormat: openAiTextFormatForAttempt } : {},
                 telemetry: false,
                 signal
@@ -9536,7 +9546,11 @@ async function runToolLoop(request) {
           firstModelEventAtMs = Date.now();
         }
       };
-      const thinkingConfig = resolveGeminiThinkingConfig(request.model, request.thinkingLevel);
+      const thinkingConfig = resolveGeminiThinkingConfig(
+        request.model,
+        request.thinkingLevel,
+        request.thinkingBudget
+      );
       const mediaResolution = toGeminiMediaResolution(request.mediaResolution);
       const config = {
         maxOutputTokens: 32e3,