🚀 feat: Agent Cache Tokens & Anthropic Reasoning Support (#6098)

* fix: handling of top_k and top_p parameters for Claude-3.7 models (allowed without reasoning)

* feat: bump @librechat/agents for Anthropic Reasoning support

* fix: update reasoning handling for OpenRouter integration

* fix: enhance agent token spending logic to include cache creation and read details

* fix: update logic for thinking status in ContentParts component

* refactor: improve agent title handling

* chore: bump @librechat/agents to version 2.1.7 for parallel tool calling for Google models

api/server/services/Endpoints/agents/title.js

@@ -20,10 +20,19 @@ const addTitle = async (req, { text, response, client }) => {
 
   const titleCache = getLogStores(CacheKeys.GEN_TITLE);
   const key = `${req.user.id}-${response.conversationId}`;
+  const responseText =
+    response?.content && Array.isArray(response?.content)
+      ? response.content.reduce((acc, block) => {
+        if (block?.type === 'text') {
+          return acc + block.text;
+        }
+        return acc;
+      }, '')
+      : (response?.content ?? response?.text ?? '');
 
   const title = await client.titleConvo({
     text,
-    responseText: response?.text ?? '',
+    responseText,
     conversationId: response.conversationId,
   });
   await titleCache.set(key, title, 120000);

api/server/services/Endpoints/anthropic/llm.js

@@ -1,6 +1,6 @@
 const { HttpsProxyAgent } = require('https-proxy-agent');
 const { anthropicSettings, removeNullishValues } = require('librechat-data-provider');
-const { checkPromptCacheSupport, getClaudeHeaders } = require('./helpers');
+const { checkPromptCacheSupport, getClaudeHeaders, configureReasoning } = require('./helpers');
 
 /**
  * Generates configuration options for creating an Anthropic language model (LLM) instance.
@@ -49,13 +49,14 @@ function getLLMConfig(apiKey, options = {}) {
     clientOptions: {},
   };
 
+  requestOptions = configureReasoning(requestOptions, systemOptions);
+
   if (!/claude-3[-.]7/.test(mergedOptions.model)) {
-    if (mergedOptions.topP !== undefined) {
-      requestOptions.topP = mergedOptions.topP;
-    }
-    if (mergedOptions.topK !== undefined) {
-      requestOptions.topK = mergedOptions.topK;
-    }
+    requestOptions.topP = mergedOptions.topP;
+    requestOptions.topK = mergedOptions.topK;
+  } else if (requestOptions.thinking == null) {
+    requestOptions.topP = mergedOptions.topP;
+    requestOptions.topK = mergedOptions.topK;
   }
 
   const supportsCacheControl =

api/server/services/Endpoints/anthropic/llm.spec.js

@@ -109,4 +109,45 @@ describe('getLLMConfig', () => {
     // Just verifying that the promptCache setting is processed
     expect(result.llmConfig).toBeDefined();
   });
+
+  it('should include topK and topP for Claude-3.7 models when thinking is not enabled', () => {
+    // Test with thinking explicitly set to null/undefined
+    const result = getLLMConfig('test-api-key', {
+      modelOptions: {
+        model: 'claude-3-7-sonnet',
+        topK: 10,
+        topP: 0.9,
+        thinking: false,
+      },
+    });
+
+    expect(result.llmConfig).toHaveProperty('topK', 10);
+    expect(result.llmConfig).toHaveProperty('topP', 0.9);
+
+    // Test with thinking explicitly set to false
+    const result2 = getLLMConfig('test-api-key', {
+      modelOptions: {
+        model: 'claude-3-7-sonnet',
+        topK: 10,
+        topP: 0.9,
+        thinking: false,
+      },
+    });
+
+    expect(result2.llmConfig).toHaveProperty('topK', 10);
+    expect(result2.llmConfig).toHaveProperty('topP', 0.9);
+
+    // Test with decimal notation as well
+    const result3 = getLLMConfig('test-api-key', {
+      modelOptions: {
+        model: 'claude-3.7-sonnet',
+        topK: 10,
+        topP: 0.9,
+        thinking: false,
+      },
+    });
+
+    expect(result3.llmConfig).toHaveProperty('topK', 10);
+    expect(result3.llmConfig).toHaveProperty('topP', 0.9);
+  });
 });

api/server/services/Endpoints/openAI/llm.js

@@ -29,7 +29,6 @@ function getLLMConfig(apiKey, options = {}) {
   const {
     modelOptions = {},
     reverseProxyUrl,
-    useOpenRouter,
     defaultQuery,
     headers,
     proxy,
@@ -56,9 +55,11 @@ function getLLMConfig(apiKey, options = {}) {
     });
   }
 
+  let useOpenRouter;
   /** @type {OpenAIClientOptions['configuration']} */
   const configOptions = {};
-  if (useOpenRouter || (reverseProxyUrl && reverseProxyUrl.includes(KnownEndpoints.openrouter))) {
+  if (reverseProxyUrl && reverseProxyUrl.includes(KnownEndpoints.openrouter)) {
+    useOpenRouter = true;
     llmConfig.include_reasoning = true;
     configOptions.baseURL = reverseProxyUrl;
     configOptions.defaultHeaders = Object.assign(
@@ -118,6 +119,13 @@ function getLLMConfig(apiKey, options = {}) {
     llmConfig.organization = process.env.OPENAI_ORGANIZATION;
   }
 
+  if (useOpenRouter && llmConfig.reasoning_effort != null) {
+    llmConfig.reasoning = {
+      effort: llmConfig.reasoning_effort,
+    };
+    delete llmConfig.reasoning_effort;
+  }
+
   return {
     /** @type {OpenAIClientOptions} */
     llmConfig,