@ljoukov/llm 3.0.15 → 4.0.0

package/dist/index.cjs

@@ -3154,9 +3154,16 @@ function resolveProvider(model) {
 function isOpenAiCodexModel(modelId) {
   return modelId.includes("codex");
 }
-function resolveOpenAiReasoningEffort(modelId, override) {
-  if (override) {
-    return override;
+function resolveOpenAiReasoningEffort(modelId, thinkingLevel) {
+  if (thinkingLevel) {
+    switch (thinkingLevel) {
+      case "low":
+        return "low";
+      case "medium":
+        return "medium";
+      case "high":
+        return "xhigh";
+    }
   }
   if (isOpenAiCodexModel(modelId)) {
     return "medium";
@@ -4441,10 +4448,42 @@ function extractFireworksToolCalls(message) {
   }
   return calls;
 }
-function resolveGeminiThinkingConfig(modelId) {
+function toGeminiThinkingLevel(thinkingLevel) {
+  switch (thinkingLevel) {
+    case "low":
+      return import_genai2.ThinkingLevel.LOW;
+    case "medium":
+      return import_genai2.ThinkingLevel.MEDIUM;
+    case "high":
+      return import_genai2.ThinkingLevel.HIGH;
+  }
+}
+function toGemini25ProThinkingBudget(thinkingLevel) {
+  switch (thinkingLevel) {
+    case "low":
+      return 256;
+    case "medium":
+      return 4096;
+    case "high":
+      return 32768;
+  }
+}
+function resolveGeminiThinkingConfig(modelId, thinkingLevel) {
   if (isGeminiImageModelId(modelId)) {
     return void 0;
   }
+  if (thinkingLevel) {
+    if (modelId === "gemini-2.5-pro") {
+      return {
+        includeThoughts: true,
+        thinkingBudget: toGemini25ProThinkingBudget(thinkingLevel)
+      };
+    }
+    return {
+      includeThoughts: true,
+      thinkingLevel: toGeminiThinkingLevel(thinkingLevel)
+    };
+  }
   switch (modelId) {
     case "gemini-3.1-pro-preview":
       return { includeThoughts: true };
@@ -4531,10 +4570,7 @@ async function runTextCall(params) {
   if (provider === "openai") {
     const openAiInput = toOpenAiInput(contents);
     const openAiTools = toOpenAiTools(request.tools);
-    const reasoningEffort = resolveOpenAiReasoningEffort(
-      modelForProvider,
-      request.openAiReasoningEffort
-    );
+    const reasoningEffort = resolveOpenAiReasoningEffort(modelForProvider, request.thinkingLevel);
     const openAiTextConfig = {
       format: request.openAiTextFormat ?? { type: "text" },
       verbosity: resolveOpenAiVerbosity(modelForProvider)
@@ -4604,10 +4640,7 @@ async function runTextCall(params) {
     }, modelForProvider);
   } else if (provider === "chatgpt") {
     const chatGptInput = toChatGptInput(contents);
-    const reasoningEffort = resolveOpenAiReasoningEffort(
-      request.model,
-      request.openAiReasoningEffort
-    );
+    const reasoningEffort = resolveOpenAiReasoningEffort(request.model, request.thinkingLevel);
     const openAiTools = toOpenAiTools(request.tools);
     const requestPayload = {
       model: modelForProvider,
@@ -4699,7 +4732,7 @@ async function runTextCall(params) {
     }, modelForProvider);
   } else {
     const geminiContents = contents.map(convertLlmContentToGeminiContent);
-    const thinkingConfig = resolveGeminiThinkingConfig(modelForProvider);
+    const thinkingConfig = resolveGeminiThinkingConfig(modelForProvider, request.thinkingLevel);
     const config = {
       maxOutputTokens: 32e3,
       ...thinkingConfig ? { thinkingConfig } : {},
@@ -4877,7 +4910,7 @@ function streamJson(request) {
           tools: request.tools,
           responseMimeType: request.responseMimeType ?? "application/json",
           responseJsonSchema,
-          openAiReasoningEffort: request.openAiReasoningEffort,
+          thinkingLevel: request.thinkingLevel,
           ...openAiTextFormatForAttempt ? { openAiTextFormat: openAiTextFormatForAttempt } : {},
           signal
         });
@@ -4954,7 +4987,7 @@ async function generateJson(request) {
         tools: request.tools,
         responseMimeType: request.responseMimeType ?? "application/json",
         responseJsonSchema,
-        openAiReasoningEffort: request.openAiReasoningEffort,
+        thinkingLevel: request.thinkingLevel,
         ...openAiTextFormatForAttempt ? { openAiTextFormat: openAiTextFormatForAttempt } : {},
         signal: request.signal
       });
@@ -5220,7 +5253,7 @@ async function runToolLoop(request) {
       const openAiTools = openAiNativeTools ? [...openAiNativeTools, ...openAiAgentTools] : [...openAiAgentTools];
       const reasoningEffort = resolveOpenAiReasoningEffort(
         providerInfo.model,
-        request.openAiReasoningEffort
+        request.thinkingLevel
       );
       const textConfig = {
         format: { type: "text" },
@@ -5488,10 +5521,7 @@ async function runToolLoop(request) {
       const openAiAgentTools = buildOpenAiToolsFromToolSet(request.tools);
       const openAiNativeTools = toOpenAiTools(request.modelTools);
       const openAiTools = openAiNativeTools ? [...openAiNativeTools, ...openAiAgentTools] : [...openAiAgentTools];
-      const reasoningEffort = resolveOpenAiReasoningEffort(
-        request.model,
-        request.openAiReasoningEffort
-      );
+      const reasoningEffort = resolveOpenAiReasoningEffort(request.model, request.thinkingLevel);
       const toolLoopInput = toChatGptInput(contents);
       const conversationId = `tool-loop-${(0, import_node_crypto.randomBytes)(8).toString("hex")}`;
       const promptCacheKey = conversationId;
@@ -5930,7 +5960,7 @@ async function runToolLoop(request) {
           firstModelEventAtMs = Date.now();
         }
       };
-      const thinkingConfig = resolveGeminiThinkingConfig(request.model);
+      const thinkingConfig = resolveGeminiThinkingConfig(request.model, request.thinkingLevel);
       const config = {
         maxOutputTokens: 32e3,
         tools: geminiTools,
@@ -9306,7 +9336,7 @@ function createSubagentController(params) {
           subagentTool: params.subagentSelection,
           modelTools: params.toolLoopRequest.modelTools,
           maxSteps: subagentRequest.maxSteps,
-          openAiReasoningEffort: params.toolLoopRequest.openAiReasoningEffort,
+          thinkingLevel: params.toolLoopRequest.thinkingLevel,
           signal: subagentRequest.signal
         },
         {