🚀 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/app/clients/AnthropicClient.js

@@ -746,15 +746,6 @@ class AnthropicClient extends BaseClient {
       metadata,
     };
 
-    if (!/claude-3[-.]7/.test(model)) {
-      if (top_p !== undefined) {
-        requestOptions.top_p = top_p;
-      }
-      if (top_k !== undefined) {
-        requestOptions.top_k = top_k;
-      }
-    }
-
     if (this.useMessages) {
       requestOptions.messages = payload;
       requestOptions.max_tokens =
@@ -769,6 +760,14 @@ class AnthropicClient extends BaseClient {
       thinkingBudget: this.options.thinkingBudget,
     });
 
+    if (!/claude-3[-.]7/.test(model)) {
+      requestOptions.top_p = top_p;
+      requestOptions.top_k = top_k;
+    } else if (requestOptions.thinking == null) {
+      requestOptions.topP = top_p;
+      requestOptions.topK = top_k;
+    }
+
     if (this.systemMessage && this.supportsCacheControl === true) {
       requestOptions.system = [
         {

api/app/clients/OpenAIClient.js

@@ -1309,6 +1309,12 @@ ${convo}
         modelOptions.include_reasoning = true;
         reasoningKey = 'reasoning';
       }
+      if (this.useOpenRouter && modelOptions.reasoning_effort != null) {
+        modelOptions.reasoning = {
+          effort: modelOptions.reasoning_effort,
+        };
+        delete modelOptions.reasoning_effort;
+      }
 
       this.streamHandler = new SplitStreamHandler({
         reasoningKey,

api/app/clients/specs/AnthropicClient.test.js

@@ -680,4 +680,53 @@ describe('AnthropicClient', () => {
       expect(capturedOptions).not.toHaveProperty('top_p');
     });
   });
+
+  it('should include top_k and top_p parameters for Claude-3.7 models when thinking is explicitly disabled', async () => {
+    const client = new AnthropicClient('test-api-key', {
+      modelOptions: {
+        model: 'claude-3-7-sonnet',
+        temperature: 0.7,
+        topK: 10,
+        topP: 0.9,
+      },
+      thinking: false,
+    });
+
+    async function* mockAsyncGenerator() {
+      yield { type: 'message_start', message: { usage: {} } };
+      yield { delta: { text: 'Test response' } };
+      yield { type: 'message_delta', usage: {} };
+    }
+
+    jest.spyOn(client, 'createResponse').mockImplementation(() => {
+      return mockAsyncGenerator();
+    });
+
+    let capturedOptions = null;
+    jest.spyOn(client, 'getClient').mockImplementation((options) => {
+      capturedOptions = options;
+      return {};
+    });
+
+    const payload = [{ role: 'user', content: 'Test message' }];
+    await client.sendCompletion(payload, {});
+
+    expect(capturedOptions).toHaveProperty('topK', 10);
+    expect(capturedOptions).toHaveProperty('topP', 0.9);
+
+    client.setOptions({
+      modelOptions: {
+        model: 'claude-3.7-sonnet',
+        temperature: 0.7,
+        topK: 10,
+        topP: 0.9,
+      },
+      thinking: false,
+    });
+
+    await client.sendCompletion(payload, {});
+
+    expect(capturedOptions).toHaveProperty('topK', 10);
+    expect(capturedOptions).toHaveProperty('topP', 0.9);
+  });
 });