fix: reasoning support for Anthropic via OpenAI provider

Author: danny-avila

package.json

@@ -51,7 +51,7 @@
     "caching": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/caching.ts --name 'Jo' --location 'New York, NY'",
     "thinking": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/thinking.ts --name 'Jo' --location 'New York, NY'",
     "memory": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/memory.ts --provider 'openAI' --name 'Jo' --location 'New York, NY'",
-    "tool-test": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/tools.ts --provider 'bedrock' --name 'Jo' --location 'New York, NY'",
+    "tool-test": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/tools.ts --provider 'anthropicLITELLM' --name 'Jo' --location 'New York, NY'",
     "search": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/search.ts --provider 'anthropic' --name 'Jo' --location 'New York, NY'",
     "ant_web_search": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/ant_web_search.ts --name 'Jo' --location 'New York, NY'",
     "abort": "node -r dotenv/config --loader ./tsconfig-paths-bootstrap.mjs --experimental-specifier-resolution=node ./src/scripts/abort.ts --provider 'openAI' --name 'Jo' --location 'New York, NY'",

src/llm/openai/index.ts

@@ -345,6 +345,10 @@ export class ChatOpenAI extends OriginalChatOpenAI<t.ChatOpenAICallOptions> {
       } else if ('reasoning' in delta) {
         chunk.additional_kwargs.reasoning_content = delta.reasoning;
       }
+      if ('provider_specific_fields' in delta) {
+        chunk.additional_kwargs.provider_specific_fields =
+          delta.provider_specific_fields;
+      }
       defaultRole = delta.role ?? defaultRole;
       const newTokenIndices = {
         prompt: options.promptIndex ?? 0,

src/llm/openai/utils/index.ts

@@ -298,10 +298,16 @@ export function _convertMessagesToOpenAIParams(
       role = 'developer';
     }
 
+    let hasAnthropicThinkingBlock: boolean = false;
+
     const content =
       typeof message.content === 'string'
         ? message.content
         : message.content.map((m) => {
+          if ('type' in m && m.type === 'thinking') {
+            hasAnthropicThinkingBlock = true;
+            return m;
+          }
           if (isDataContentBlock(m)) {
             return convertToProviderContentBlock(
               m,
@@ -326,7 +332,7 @@ export function _convertMessagesToOpenAIParams(
       completionParam.tool_calls = message.tool_calls.map(
         convertLangChainToolCallToOpenAI
       );
-      completionParam.content = '';
+      completionParam.content = hasAnthropicThinkingBlock ? content : '';
     } else {
       if (message.additional_kwargs.tool_calls != null) {
         completionParam.tool_calls = message.additional_kwargs.tool_calls;

src/messages/prune.ts

@@ -389,6 +389,14 @@ export function checkValidNumber(value: unknown): value is number {
   return typeof value === 'number' && !isNaN(value) && value > 0;
 }
 
+type ThinkingBlocks = {
+  thinking_blocks?: Array<{
+    type: 'thinking';
+    thinking: string;
+    signature: string;
+  }>;
+};
+
 export function createPruneMessages(factoryParams: PruneMessagesFactoryParams) {
   const indexTokenCountMap = { ...factoryParams.indexTokenCountMap };
   let lastTurnStartIndex = factoryParams.startIndex;
@@ -402,6 +410,49 @@ export function createPruneMessages(factoryParams: PruneMessagesFactoryParams) {
     context: BaseMessage[];
     indexTokenCountMap: Record<string, number | undefined>;
   } {
+    if (
+      factoryParams.provider === Providers.OPENAI &&
+      factoryParams.thinkingEnabled === true
+    ) {
+      for (let i = lastTurnStartIndex; i < params.messages.length; i++) {
+        const m = params.messages[i];
+        if (
+          m.getType() === 'ai' &&
+          typeof m.additional_kwargs.reasoning_content === 'string' &&
+          Array.isArray(
+            (
+              m.additional_kwargs.provider_specific_fields as
+                | ThinkingBlocks
+                | undefined
+            )?.thinking_blocks
+          ) &&
+          (m as AIMessage).tool_calls &&
+          ((m as AIMessage).tool_calls?.length ?? 0) > 0
+        ) {
+          const message = m as AIMessage;
+          const thinkingBlocks = (
+            message.additional_kwargs.provider_specific_fields as ThinkingBlocks
+          ).thinking_blocks;
+          const signature =
+            thinkingBlocks?.[thinkingBlocks.length - 1].signature;
+          const thinkingBlock: ThinkingContentText = {
+            signature,
+            type: ContentTypes.THINKING,
+            thinking: message.additional_kwargs.reasoning_content as string,
+          };
+
+          params.messages[i] = new AIMessage({
+            ...message,
+            content: [thinkingBlock],
+            additional_kwargs: {
+              ...message.additional_kwargs,
+              reasoning_content: undefined,
+            },
+          });
+        }
+      }
+    }
+
     let currentUsage: UsageMetadata | undefined;
     if (
       params.usageMetadata &&

src/scripts/tools.ts

@@ -127,7 +127,10 @@ async function testStandardStreaming(): Promise<void> {
   const inputs = {
     messages: conversationHistory,
   };
-  const finalContentParts = await run.processStream(inputs, config);
+  const finalContentParts = await run.processStream(inputs, config, {
+    indexTokenCountMap: { 0: 35 },
+    maxContextTokens: 89000,
+  });
   const finalMessages = run.getRunMessages();
   if (finalMessages) {
     conversationHistory.push(...finalMessages);

src/utils/llmConfig.ts

@@ -12,6 +12,30 @@ export const llmConfigs: Record<string, t.LLMConfig | undefined> = {
     streamUsage: true,
     // disableStreaming: true,
   },
+  anthropicLITELLM: {
+    provider: Providers.OPENAI,
+    streaming: true,
+    streamUsage: false,
+    apiKey: 'sk-1234',
+    model: 'claude-sonnet-4',
+    maxTokens: 8192,
+    modelKwargs: {
+      metadata: {
+        user_id: 'some_user_id',
+      },
+      thinking: {
+        type: 'enabled',
+        budget_tokens: 2000,
+      },
+    },
+    configuration: {
+      baseURL: 'http://host.docker.internal:4000/v1',
+      defaultHeaders: {
+        'anthropic-beta': 'prompt-caching-2024-07-31,context-1m-2025-08-07',
+      },
+    },
+    // disableStreaming: true,
+  },
   [Providers.XAI]: {
     provider: Providers.XAI,
     model: 'grok-2-latest',