@librechat/agents 1.9.5 → 1.9.6

package/dist/cjs/llm/anthropic/llm.cjs

@@ -46,19 +46,56 @@ function cloneChunk(text, tokenType, chunk) {
 }
 class CustomAnthropic extends anthropic.ChatAnthropicMessages {
     _lc_stream_delay;
+    message_start;
+    message_delta;
+    tools_in_params;
+    emitted_usage;
     constructor(fields) {
         super(fields);
         this._lc_stream_delay = fields._lc_stream_delay ?? 25;
     }
-    createGenerationChunk({ token, chunk, shouldStreamUsage, }) {
+    /**
+     * Get stream usage as returned by this client's API response.
+     * @returns {AnthropicStreamUsage} The stream usage object.
+     */
+    getStreamUsage() {
+        if (this.emitted_usage === true) {
+            return;
+        }
+        const inputUsage = (this.message_start?.message)?.usage;
+        const outputUsage = this.message_delta?.usage;
+        if (!outputUsage) {
+            return;
+        }
+        const totalUsage = {
+            total_tokens: (inputUsage?.input_tokens ?? 0)
+                + (inputUsage?.output_tokens ?? 0)
+                + (inputUsage?.cache_creation_input_tokens ?? 0)
+                + (inputUsage?.cache_read_input_tokens ?? 0)
+                + (outputUsage.input_tokens ?? 0)
+                + (outputUsage.output_tokens ?? 0)
+                + (outputUsage.cache_creation_input_tokens ?? 0)
+                + (outputUsage.cache_read_input_tokens ?? 0),
+        };
+        this.emitted_usage = true;
+        return Object.assign(totalUsage, inputUsage, outputUsage);
+    }
+    resetTokenEvents() {
+        this.message_start = undefined;
+        this.message_delta = undefined;
+        this.emitted_usage = undefined;
+        this.tools_in_params = undefined;
+    }
+    createGenerationChunk({ token, chunk, usageMetadata, shouldStreamUsage, }) {
+        const usage_metadata = shouldStreamUsage ? usageMetadata ?? chunk.usage_metadata : undefined;
         return new outputs.ChatGenerationChunk({
             message: new messages.AIMessageChunk({
                 // Just yield chunk as it is and tool_use will be concat by BaseChatModel._generateUncached().
                 content: chunk.content,
                 additional_kwargs: chunk.additional_kwargs,
                 tool_call_chunks: chunk.tool_call_chunks,
-                usage_metadata: shouldStreamUsage ? chunk.usage_metadata : undefined,
                 response_metadata: chunk.response_metadata,
+                usage_metadata,
                 id: chunk.id,
             }),
             text: token ?? '',
@@ -67,11 +104,12 @@ class CustomAnthropic extends anthropic.ChatAnthropicMessages {
     async *_streamResponseChunks(messages, options, runManager) {
         const params = this.invocationParams(options);
         const formattedMessages = message_inputs._convertMessagesToAnthropicPayload(messages);
-        const coerceContentToString = !_toolsInParams({
+        this.tools_in_params = _toolsInParams({
             ...params,
             ...formattedMessages,
             stream: false,
         });
+        const coerceContentToString = !this.tools_in_params;
         const stream = await this.createStreamWithRetry({
             ...params,
             ...formattedMessages,
@@ -79,12 +117,23 @@ class CustomAnthropic extends anthropic.ChatAnthropicMessages {
         }, {
             headers: options.headers,
         });
+        const shouldStreamUsage = this.streamUsage ?? options.streamUsage;
         for await (const data of stream) {
             if (options.signal?.aborted === true) {
                 stream.controller.abort();
                 throw new Error('AbortError: User aborted the request.');
             }
-            const shouldStreamUsage = this.streamUsage ?? options.streamUsage;
+            const type = data.type ?? '';
+            if (type === 'message_start') {
+                this.message_start = data;
+            }
+            else if (type === 'message_delta') {
+                this.message_delta = data;
+            }
+            let usageMetadata;
+            if (this.tools_in_params !== true && this.emitted_usage !== true) {
+                usageMetadata = this.getStreamUsage();
+            }
             const result = message_outputs._makeMessageChunkFromAnthropicEvent(data, {
                 streamUsage: shouldStreamUsage,
                 coerceContentToString,
@@ -93,10 +142,11 @@ class CustomAnthropic extends anthropic.ChatAnthropicMessages {
                 continue;
             const { chunk } = result;
             const [token = '', tokenType] = extractToken(chunk);
-            if (!tokenType || tokenType === 'input') {
+            if (!tokenType || tokenType === 'input' || (token === '' && usageMetadata)) {
                 const generationChunk = this.createGenerationChunk({
                     token,
                     chunk,
+                    usageMetadata,
                     shouldStreamUsage,
                 });
                 yield generationChunk;
@@ -111,13 +161,18 @@ class CustomAnthropic extends anthropic.ChatAnthropicMessages {
             });
             const generator = textStream.generateText();
             try {
+                let emittedUsage = false;
                 for await (const currentToken of generator) {
                     const newChunk = cloneChunk(currentToken, tokenType, chunk);
                     const generationChunk = this.createGenerationChunk({
                         token: currentToken,
                         chunk: newChunk,
+                        usageMetadata: emittedUsage ? undefined : usageMetadata,
                         shouldStreamUsage,
                     });
+                    if (usageMetadata && !emittedUsage) {
+                        emittedUsage = true;
+                    }
                     yield generationChunk;
                     await runManager?.handleLLMNewToken(token, undefined, undefined, undefined, undefined, { chunk: generationChunk });
                 }
@@ -126,6 +181,7 @@ class CustomAnthropic extends anthropic.ChatAnthropicMessages {
                 await generator.return();
             }
         }
+        this.resetTokenEvents();
     }
 }
 

package/dist/cjs/llm/anthropic/llm.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"llm.cjs","sources":["../../../../src/llm/anthropic/llm.ts"],"sourcesContent":["import { AIMessageChunk } from '@langchain/core/messages';\nimport { ChatAnthropicMessages } from '@langchain/anthropic';\nimport { ChatGenerationChunk } from '@langchain/core/outputs';\nimport type { BaseMessage, MessageContentComplex } from '@langchain/core/messages';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { AnthropicInput } from '@langchain/anthropic';\nimport type { AnthropicMessageCreateParams } from '@/llm/anthropic/types';\nimport { _makeMessageChunkFromAnthropicEvent } from './utils/message_outputs';\nimport { _convertMessagesToAnthropicPayload } from './utils/message_inputs';\nimport { TextStream } from '@/llm/text';\n\nfunction _toolsInParams(params: AnthropicMessageCreateParams): boolean {\n  return !!(params.tools && params.tools.length > 0);\n}\n\nfunction extractToken(chunk: AIMessageChunk): [string, 'string' | 'input' | 'content'] | [undefined] {\n  if (typeof chunk.content === 'string') {\n    return [chunk.content, 'string'];\n  } else if (\n    Array.isArray(chunk.content) &&\n    chunk.content.length >= 1 &&\n    'input' in chunk.content[0]\n  ) {\n    return typeof chunk.content[0].input === 'string'\n      ? [chunk.content[0].input, 'input']\n      : [JSON.stringify(chunk.content[0].input), 'input'];\n  } else if (\n    Array.isArray(chunk.content) &&\n    chunk.content.length >= 1 &&\n    'text' in chunk.content[0]\n  ) {\n    return [chunk.content[0].text, 'content'];\n  }\n  return [undefined];\n}\n\nfunction cloneChunk(text: string, tokenType: string, chunk: AIMessageChunk): AIMessageChunk {\n  if (tokenType === 'string') {\n    return new AIMessageChunk(Object.assign({}, chunk, { content: text }));\n  } else if (tokenType === 'input') {\n    return chunk;\n  }\n  const content = chunk.content[0] as MessageContentComplex;\n  if (tokenType === 'content' && content.type === 'text') {\n    return new AIMessageChunk(Object.assign({}, chunk, { content: [Object.assign({}, content, { text })] }));\n  } else if (tokenType === 'content' && content.type === 'text_delta') {\n    return new AIMessageChunk(Object.assign({}, chunk, { content: [Object.assign({}, content, { text })] }));\n  }\n\n  return chunk;\n}\n\nexport type CustomAnthropicInput = AnthropicInput & { _lc_stream_delay?: number };\n\nexport class CustomAnthropic extends ChatAnthropicMessages {\n  _lc_stream_delay: number;\n  constructor(fields: CustomAnthropicInput) {\n    super(fields);\n    this._lc_stream_delay = fields._lc_stream_delay ?? 25;\n  }\n\n  private createGenerationChunk({\n    token,\n    chunk,\n    shouldStreamUsage,\n  }: {\n    token?: string,\n    chunk: AIMessageChunk,\n    shouldStreamUsage: boolean\n  }): ChatGenerationChunk {\n    return new ChatGenerationChunk({\n      message: new AIMessageChunk({\n        // Just yield chunk as it is and tool_use will be concat by BaseChatModel._generateUncached().\n        content: chunk.content,\n        additional_kwargs: chunk.additional_kwargs,\n        tool_call_chunks: chunk.tool_call_chunks,\n        usage_metadata: shouldStreamUsage ? chunk.usage_metadata : undefined,\n        response_metadata: chunk.response_metadata,\n        id: chunk.id,\n      }),\n      text: token ?? '',\n    });\n  }\n\n  async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'],\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    const params = this.invocationParams(options);\n    const formattedMessages = _convertMessagesToAnthropicPayload(messages);\n    const coerceContentToString = !_toolsInParams({\n      ...params,\n      ...formattedMessages,\n      stream: false,\n    });\n\n    const stream = await this.createStreamWithRetry(\n      {\n        ...params,\n        ...formattedMessages,\n        stream: true,\n      },\n      {\n        headers: options.headers,\n      }\n    );\n\n    for await (const data of stream) {\n      if (options.signal?.aborted === true) {\n        stream.controller.abort();\n        throw new Error('AbortError: User aborted the request.');\n      }\n\n      const shouldStreamUsage = this.streamUsage ?? options.streamUsage;\n      const result = _makeMessageChunkFromAnthropicEvent(data, {\n        streamUsage: shouldStreamUsage,\n        coerceContentToString,\n      });\n      if (!result) continue;\n\n      const { chunk } = result;\n      const [token = '', tokenType] = extractToken(chunk);\n\n      if (!tokenType || tokenType === 'input') {\n        const generationChunk = this.createGenerationChunk({\n          token,\n          chunk,\n          shouldStreamUsage,\n        });\n        yield generationChunk;\n        await runManager?.handleLLMNewToken(\n          token,\n          undefined,\n          undefined,\n          undefined,\n          undefined,\n          { chunk: generationChunk }\n        );\n        continue;\n      }\n\n      const textStream = new TextStream(token, {\n        delay: this._lc_stream_delay,\n        firstWordChunk: true,\n        minChunkSize: 4,\n        maxChunkSize: 8,\n      });\n\n      const generator = textStream.generateText();\n      try {\n        for await (const currentToken of generator) {\n          const newChunk = cloneChunk(currentToken, tokenType, chunk);\n          const generationChunk = this.createGenerationChunk({\n            token: currentToken,\n            chunk: newChunk,\n            shouldStreamUsage,\n          });\n          yield generationChunk;\n\n          await runManager?.handleLLMNewToken(\n            token,\n            undefined,\n            undefined,\n            undefined,\n            undefined,\n            { chunk: generationChunk }\n          );\n        }\n      } finally {\n        await generator.return();\n      }\n    }\n  }\n}"],"names":["AIMessageChunk","ChatAnthropicMessages","ChatGenerationChunk","_convertMessagesToAnthropicPayload","_makeMessageChunkFromAnthropicEvent","TextStream"],"mappings":";;;;;;;;;AAWA,SAAS,cAAc,CAAC,MAAoC,EAAA;AAC1D,IAAA,OAAO,CAAC,EAAE,MAAM,CAAC,KAAK,IAAI,MAAM,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;AACrD,CAAC;AAED,SAAS,YAAY,CAAC,KAAqB,EAAA;AACzC,IAAA,IAAI,OAAO,KAAK,CAAC,OAAO,KAAK,QAAQ,EAAE;AACrC,QAAA,OAAO,CAAC,KAAK,CAAC,OAAO,EAAE,QAAQ,CAAC,CAAC;KAClC;AAAM,SAAA,IACL,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,OAAO,CAAC;AAC5B,QAAA,KAAK,CAAC,OAAO,CAAC,MAAM,IAAI,CAAC;QACzB,OAAO,IAAI,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,EAC3B;QACA,OAAO,OAAO,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,KAAK,KAAK,QAAQ;AAC/C,cAAE,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,KAAK,EAAE,OAAO,CAAC;AACnC,cAAE,CAAC,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,EAAE,OAAO,CAAC,CAAC;KACvD;AAAM,SAAA,IACL,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,OAAO,CAAC;AAC5B,QAAA,KAAK,CAAC,OAAO,CAAC,MAAM,IAAI,CAAC;QACzB,MAAM,IAAI,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,EAC1B;AACA,QAAA,OAAO,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,SAAS,CAAC,CAAC;KAC3C;IACD,OAAO,CAAC,SAAS,CAAC,CAAC;AACrB,CAAC;AAED,SAAS,UAAU,CAAC,IAAY,EAAE,SAAiB,EAAE,KAAqB,EAAA;AACxE,IAAA,IAAI,SAAS,KAAK,QAAQ,EAAE;AAC1B,QAAA,OAAO,IAAIA,uBAAc,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,EAAE,KAAK,EAAE,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC;KACxE;AAAM,SAAA,IAAI,SAAS,KAAK,OAAO,EAAE;AAChC,QAAA,OAAO,KAAK,CAAC;KACd;IACD,MAAM,OAAO,GAAG,KAAK,CAAC,OAAO,CAAC,CAAC,CAA0B,CAAC;IAC1D,IAAI,SAAS,KAAK,SAAS,IAAI,OAAO,CAAC,IAAI,KAAK,MAAM,EAAE;AACtD,QAAA,OAAO,IAAIA,uBAAc,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,EAAE,KAAK,EAAE,EAAE,OAAO,EAAE,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,EAAE,OAAO,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;KAC1G;SAAM,IAAI,SAAS,KAAK,SAAS,IAAI,OAAO,CAAC,IAAI,KAAK,YAAY,EAAE;AACnE,QAAA,OAAO,IAAIA,uBAAc,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,EAAE,KAAK,EAAE,EAAE,OAAO,EAAE,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,EAAE,OAAO,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;KAC1G;AAED,IAAA,OAAO,KAAK,CAAC;AACf,CAAC;AAIK,MAAO,eAAgB,SAAQC,+BAAqB,CAAA;AACxD,IAAA,gBAAgB,CAAS;AACzB,IAAA,WAAA,CAAY,MAA4B,EAAA;QACtC,KAAK,CAAC,MAAM,CAAC,CAAC;QACd,IAAI,CAAC,gBAAgB,GAAG,MAAM,CAAC,gBAAgB,IAAI,EAAE,CAAC;KACvD;AAEO,IAAA,qBAAqB,CAAC,EAC5B,KAAK,EACL,KAAK,EACL,iBAAiB,GAKlB,EAAA;QACC,OAAO,IAAIC,2BAAmB,CAAC;YAC7B,OAAO,EAAE,IAAIF,uBAAc,CAAC;;gBAE1B,OAAO,EAAE,KAAK,CAAC,OAAO;gBACtB,iBAAiB,EAAE,KAAK,CAAC,iBAAiB;gBAC1C,gBAAgB,EAAE,KAAK,CAAC,gBAAgB;gBACxC,cAAc,EAAE,iBAAiB,GAAG,KAAK,CAAC,cAAc,GAAG,SAAS;gBACpE,iBAAiB,EAAE,KAAK,CAAC,iBAAiB;gBAC1C,EAAE,EAAE,KAAK,CAAC,EAAE;aACb,CAAC;YACF,IAAI,EAAE,KAAK,IAAI,EAAE;AAClB,SAAA,CAAC,CAAC;KACJ;IAED,OAAO,qBAAqB,CAC1B,QAAuB,EACvB,OAAkC,EAClC,UAAqC,EAAA;QAErC,MAAM,MAAM,GAAG,IAAI,CAAC,gBAAgB,CAAC,OAAO,CAAC,CAAC;AAC9C,QAAA,MAAM,iBAAiB,GAAGG,iDAAkC,CAAC,QAAQ,CAAC,CAAC;AACvE,QAAA,MAAM,qBAAqB,GAAG,CAAC,cAAc,CAAC;AAC5C,YAAA,GAAG,MAAM;AACT,YAAA,GAAG,iBAAiB;AACpB,YAAA,MAAM,EAAE,KAAK;AACd,SAAA,CAAC,CAAC;AAEH,QAAA,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,qBAAqB,CAC7C;AACE,YAAA,GAAG,MAAM;AACT,YAAA,GAAG,iBAAiB;AACpB,YAAA,MAAM,EAAE,IAAI;SACb,EACD;YACE,OAAO,EAAE,OAAO,CAAC,OAAO;AACzB,SAAA,CACF,CAAC;AAEF,QAAA,WAAW,MAAM,IAAI,IAAI,MAAM,EAAE;YAC/B,IAAI,OAAO,CAAC,MAAM,EAAE,OAAO,KAAK,IAAI,EAAE;AACpC,gBAAA,MAAM,CAAC,UAAU,CAAC,KAAK,EAAE,CAAC;AAC1B,gBAAA,MAAM,IAAI,KAAK,CAAC,uCAAuC,CAAC,CAAC;aAC1D;YAED,MAAM,iBAAiB,GAAG,IAAI,CAAC,WAAW,IAAI,OAAO,CAAC,WAAW,CAAC;AAClE,YAAA,MAAM,MAAM,GAAGC,mDAAmC,CAAC,IAAI,EAAE;AACvD,gBAAA,WAAW,EAAE,iBAAiB;gBAC9B,qBAAqB;AACtB,aAAA,CAAC,CAAC;AACH,YAAA,IAAI,CAAC,MAAM;gBAAE,SAAS;AAEtB,YAAA,MAAM,EAAE,KAAK,EAAE,GAAG,MAAM,CAAC;AACzB,YAAA,MAAM,CAAC,KAAK,GAAG,EAAE,EAAE,SAAS,CAAC,GAAG,YAAY,CAAC,KAAK,CAAC,CAAC;AAEpD,YAAA,IAAI,CAAC,SAAS,IAAI,SAAS,KAAK,OAAO,EAAE;AACvC,gBAAA,MAAM,eAAe,GAAG,IAAI,CAAC,qBAAqB,CAAC;oBACjD,KAAK;oBACL,KAAK;oBACL,iBAAiB;AAClB,iBAAA,CAAC,CAAC;AACH,gBAAA,MAAM,eAAe,CAAC;gBACtB,MAAM,UAAU,EAAE,iBAAiB,CACjC,KAAK,EACL,SAAS,EACT,SAAS,EACT,SAAS,EACT,SAAS,EACT,EAAE,KAAK,EAAE,eAAe,EAAE,CAC3B,CAAC;gBACF,SAAS;aACV;AAED,YAAA,MAAM,UAAU,GAAG,IAAIC,eAAU,CAAC,KAAK,EAAE;gBACvC,KAAK,EAAE,IAAI,CAAC,gBAAgB;AAC5B,gBAAA,cAAc,EAAE,IAAI;AACpB,gBAAA,YAAY,EAAE,CAAC;AACf,gBAAA,YAAY,EAAE,CAAC;AAChB,aAAA,CAAC,CAAC;AAEH,YAAA,MAAM,SAAS,GAAG,UAAU,CAAC,YAAY,EAAE,CAAC;AAC5C,YAAA,IAAI;AACF,gBAAA,WAAW,MAAM,YAAY,IAAI,SAAS,EAAE;oBAC1C,MAAM,QAAQ,GAAG,UAAU,CAAC,YAAY,EAAE,SAAS,EAAE,KAAK,CAAC,CAAC;AAC5D,oBAAA,MAAM,eAAe,GAAG,IAAI,CAAC,qBAAqB,CAAC;AACjD,wBAAA,KAAK,EAAE,YAAY;AACnB,wBAAA,KAAK,EAAE,QAAQ;wBACf,iBAAiB;AAClB,qBAAA,CAAC,CAAC;AACH,oBAAA,MAAM,eAAe,CAAC;oBAEtB,MAAM,UAAU,EAAE,iBAAiB,CACjC,KAAK,EACL,SAAS,EACT,SAAS,EACT,SAAS,EACT,SAAS,EACT,EAAE,KAAK,EAAE,eAAe,EAAE,CAC3B,CAAC;iBACH;aACF;oBAAS;AACR,gBAAA,MAAM,SAAS,CAAC,MAAM,EAAE,CAAC;aAC1B;SACF;KACF;AACF;;;;"}
+{"version":3,"file":"llm.cjs","sources":["../../../../src/llm/anthropic/llm.ts"],"sourcesContent":["import { AIMessageChunk } from '@langchain/core/messages';\nimport { ChatAnthropicMessages } from '@langchain/anthropic';\nimport { ChatGenerationChunk } from '@langchain/core/outputs';\nimport type { BaseMessage, MessageContentComplex } from '@langchain/core/messages';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { AnthropicInput } from '@langchain/anthropic';\nimport type { AnthropicMessageCreateParams, AnthropicStreamUsage, AnthropicMessageStartEvent, AnthropicMessageDeltaEvent } from '@/llm/anthropic/types';\nimport { _makeMessageChunkFromAnthropicEvent } from './utils/message_outputs';\nimport { _convertMessagesToAnthropicPayload } from './utils/message_inputs';\nimport { TextStream } from '@/llm/text';\n\nfunction _toolsInParams(params: AnthropicMessageCreateParams): boolean {\n  return !!(params.tools && params.tools.length > 0);\n}\n\nfunction extractToken(chunk: AIMessageChunk): [string, 'string' | 'input' | 'content'] | [undefined] {\n  if (typeof chunk.content === 'string') {\n    return [chunk.content, 'string'];\n  } else if (\n    Array.isArray(chunk.content) &&\n    chunk.content.length >= 1 &&\n    'input' in chunk.content[0]\n  ) {\n    return typeof chunk.content[0].input === 'string'\n      ? [chunk.content[0].input, 'input']\n      : [JSON.stringify(chunk.content[0].input), 'input'];\n  } else if (\n    Array.isArray(chunk.content) &&\n    chunk.content.length >= 1 &&\n    'text' in chunk.content[0]\n  ) {\n    return [chunk.content[0].text, 'content'];\n  }\n  return [undefined];\n}\n\nfunction cloneChunk(text: string, tokenType: string, chunk: AIMessageChunk): AIMessageChunk {\n  if (tokenType === 'string') {\n    return new AIMessageChunk(Object.assign({}, chunk, { content: text }));\n  } else if (tokenType === 'input') {\n    return chunk;\n  }\n  const content = chunk.content[0] as MessageContentComplex;\n  if (tokenType === 'content' && content.type === 'text') {\n    return new AIMessageChunk(Object.assign({}, chunk, { content: [Object.assign({}, content, { text })] }));\n  } else if (tokenType === 'content' && content.type === 'text_delta') {\n    return new AIMessageChunk(Object.assign({}, chunk, { content: [Object.assign({}, content, { text })] }));\n  }\n\n  return chunk;\n}\n\nexport type CustomAnthropicInput = AnthropicInput & { _lc_stream_delay?: number };\n\nexport class CustomAnthropic extends ChatAnthropicMessages {\n  _lc_stream_delay: number;\n  private message_start: AnthropicMessageStartEvent | undefined;\n  private message_delta: AnthropicMessageDeltaEvent | undefined;\n  private tools_in_params?: boolean;\n  private emitted_usage?: boolean;\n  constructor(fields: CustomAnthropicInput) {\n    super(fields);\n    this._lc_stream_delay = fields._lc_stream_delay ?? 25;\n  }\n\n  /**\n   * Get stream usage as returned by this client's API response.\n   * @returns {AnthropicStreamUsage} The stream usage object.\n   */\n  getStreamUsage(): AnthropicStreamUsage | undefined {\n    if (this.emitted_usage === true) {\n      return;\n    }\n    const inputUsage = (this.message_start?.message)?.usage as undefined | AnthropicStreamUsage;\n    const outputUsage = this.message_delta?.usage as undefined | Partial<AnthropicStreamUsage>;\n    if (!outputUsage) {\n      return;\n    }\n    const totalUsage = {\n      total_tokens: (inputUsage?.input_tokens ?? 0)\n      + (inputUsage?.output_tokens ?? 0)\n      + (inputUsage?.cache_creation_input_tokens ?? 0)\n      + (inputUsage?.cache_read_input_tokens ?? 0)\n      + (outputUsage.input_tokens ?? 0)\n      + (outputUsage.output_tokens ?? 0)\n      + (outputUsage.cache_creation_input_tokens ?? 0)\n      + (outputUsage.cache_read_input_tokens ?? 0),\n    };\n\n    this.emitted_usage = true;\n    return Object.assign(totalUsage, inputUsage, outputUsage);\n  }\n\n  resetTokenEvents(): void {\n    this.message_start = undefined;\n    this.message_delta = undefined;\n    this.emitted_usage = undefined;\n    this.tools_in_params = undefined;\n  }\n\n  private createGenerationChunk({\n    token,\n    chunk,\n    usageMetadata,\n    shouldStreamUsage,\n  }: {\n    token?: string,\n    chunk: AIMessageChunk,\n    shouldStreamUsage: boolean\n    usageMetadata?: AnthropicStreamUsage,\n  }): ChatGenerationChunk {\n    const usage_metadata = shouldStreamUsage ? usageMetadata ?? chunk.usage_metadata : undefined;\n    return new ChatGenerationChunk({\n      message: new AIMessageChunk({\n        // Just yield chunk as it is and tool_use will be concat by BaseChatModel._generateUncached().\n        content: chunk.content,\n        additional_kwargs: chunk.additional_kwargs,\n        tool_call_chunks: chunk.tool_call_chunks,\n        response_metadata: chunk.response_metadata,\n        usage_metadata,\n        id: chunk.id,\n      }),\n      text: token ?? '',\n    });\n  }\n\n  async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'],\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    const params = this.invocationParams(options);\n    const formattedMessages = _convertMessagesToAnthropicPayload(messages);\n    this.tools_in_params = _toolsInParams({\n      ...params,\n      ...formattedMessages,\n      stream: false,\n    });\n    const coerceContentToString = !this.tools_in_params;\n\n    const stream = await this.createStreamWithRetry(\n      {\n        ...params,\n        ...formattedMessages,\n        stream: true,\n      },\n      {\n        headers: options.headers,\n      }\n    );\n\n    const shouldStreamUsage = this.streamUsage ?? options.streamUsage;\n\n    for await (const data of stream) {\n      if (options.signal?.aborted === true) {\n        stream.controller.abort();\n        throw new Error('AbortError: User aborted the request.');\n      }\n\n      const type = data.type ?? '';\n      if (type === 'message_start') {\n        this.message_start = data as AnthropicMessageStartEvent;\n      } else if (type === 'message_delta') {\n        this.message_delta = data as AnthropicMessageDeltaEvent;\n      }\n\n      let usageMetadata: AnthropicStreamUsage | undefined;\n      if (this.tools_in_params !== true && this.emitted_usage !== true) {\n        usageMetadata = this.getStreamUsage();\n      }\n\n      const result = _makeMessageChunkFromAnthropicEvent(data, {\n        streamUsage: shouldStreamUsage,\n        coerceContentToString,\n      });\n      if (!result) continue;\n\n      const { chunk } = result;\n      const [token = '', tokenType] = extractToken(chunk);\n\n      if (!tokenType || tokenType === 'input' || (token === '' && usageMetadata)) {\n        const generationChunk = this.createGenerationChunk({\n          token,\n          chunk,\n          usageMetadata,\n          shouldStreamUsage,\n        });\n        yield generationChunk;\n        await runManager?.handleLLMNewToken(\n          token,\n          undefined,\n          undefined,\n          undefined,\n          undefined,\n          { chunk: generationChunk }\n        );\n        continue;\n      }\n\n      const textStream = new TextStream(token, {\n        delay: this._lc_stream_delay,\n        firstWordChunk: true,\n        minChunkSize: 4,\n        maxChunkSize: 8,\n      });\n\n      const generator = textStream.generateText();\n      try {\n        let emittedUsage = false;\n        for await (const currentToken of generator) {\n          const newChunk = cloneChunk(currentToken, tokenType, chunk);\n\n          const generationChunk = this.createGenerationChunk({\n            token: currentToken,\n            chunk: newChunk,\n            usageMetadata: emittedUsage ? undefined : usageMetadata,\n            shouldStreamUsage,\n          });\n\n          if (usageMetadata && !emittedUsage) {\n            emittedUsage = true;\n          }\n          yield generationChunk;\n\n          await runManager?.handleLLMNewToken(\n            token,\n            undefined,\n            undefined,\n            undefined,\n            undefined,\n            { chunk: generationChunk }\n          );\n        }\n      } finally {\n        await generator.return();\n      }\n    }\n\n    this.resetTokenEvents();\n  }\n}"],"names":["AIMessageChunk","ChatAnthropicMessages","ChatGenerationChunk","_convertMessagesToAnthropicPayload","_makeMessageChunkFromAnthropicEvent","TextStream"],"mappings":";;;;;;;;;AAWA,SAAS,cAAc,CAAC,MAAoC,EAAA;AAC1D,IAAA,OAAO,CAAC,EAAE,MAAM,CAAC,KAAK,IAAI,MAAM,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;AACrD,CAAC;AAED,SAAS,YAAY,CAAC,KAAqB,EAAA;AACzC,IAAA,IAAI,OAAO,KAAK,CAAC,OAAO,KAAK,QAAQ,EAAE;AACrC,QAAA,OAAO,CAAC,KAAK,CAAC,OAAO,EAAE,QAAQ,CAAC,CAAC;KAClC;AAAM,SAAA,IACL,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,OAAO,CAAC;AAC5B,QAAA,KAAK,CAAC,OAAO,CAAC,MAAM,IAAI,CAAC;QACzB,OAAO,IAAI,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,EAC3B;QACA,OAAO,OAAO,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,KAAK,KAAK,QAAQ;AAC/C,cAAE,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,KAAK,EAAE,OAAO,CAAC;AACnC,cAAE,CAAC,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,EAAE,OAAO,CAAC,CAAC;KACvD;AAAM,SAAA,IACL,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,OAAO,CAAC;AAC5B,QAAA,KAAK,CAAC,OAAO,CAAC,MAAM,IAAI,CAAC;QACzB,MAAM,IAAI,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,EAC1B;AACA,QAAA,OAAO,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,SAAS,CAAC,CAAC;KAC3C;IACD,OAAO,CAAC,SAAS,CAAC,CAAC;AACrB,CAAC;AAED,SAAS,UAAU,CAAC,IAAY,EAAE,SAAiB,EAAE,KAAqB,EAAA;AACxE,IAAA,IAAI,SAAS,KAAK,QAAQ,EAAE;AAC1B,QAAA,OAAO,IAAIA,uBAAc,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,EAAE,KAAK,EAAE,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC;KACxE;AAAM,SAAA,IAAI,SAAS,KAAK,OAAO,EAAE;AAChC,QAAA,OAAO,KAAK,CAAC;KACd;IACD,MAAM,OAAO,GAAG,KAAK,CAAC,OAAO,CAAC,CAAC,CAA0B,CAAC;IAC1D,IAAI,SAAS,KAAK,SAAS,IAAI,OAAO,CAAC,IAAI,KAAK,MAAM,EAAE;AACtD,QAAA,OAAO,IAAIA,uBAAc,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,EAAE,KAAK,EAAE,EAAE,OAAO,EAAE,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,EAAE,OAAO,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;KAC1G;SAAM,IAAI,SAAS,KAAK,SAAS,IAAI,OAAO,CAAC,IAAI,KAAK,YAAY,EAAE;AACnE,QAAA,OAAO,IAAIA,uBAAc,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,EAAE,KAAK,EAAE,EAAE,OAAO,EAAE,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,EAAE,OAAO,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;KAC1G;AAED,IAAA,OAAO,KAAK,CAAC;AACf,CAAC;AAIK,MAAO,eAAgB,SAAQC,+BAAqB,CAAA;AACxD,IAAA,gBAAgB,CAAS;AACjB,IAAA,aAAa,CAAyC;AACtD,IAAA,aAAa,CAAyC;AACtD,IAAA,eAAe,CAAW;AAC1B,IAAA,aAAa,CAAW;AAChC,IAAA,WAAA,CAAY,MAA4B,EAAA;QACtC,KAAK,CAAC,MAAM,CAAC,CAAC;QACd,IAAI,CAAC,gBAAgB,GAAG,MAAM,CAAC,gBAAgB,IAAI,EAAE,CAAC;KACvD;AAED;;;AAGG;IACH,cAAc,GAAA;AACZ,QAAA,IAAI,IAAI,CAAC,aAAa,KAAK,IAAI,EAAE;YAC/B,OAAO;SACR;QACD,MAAM,UAAU,GAAG,CAAC,IAAI,CAAC,aAAa,EAAE,OAAO,GAAG,KAAyC,CAAC;AAC5F,QAAA,MAAM,WAAW,GAAG,IAAI,CAAC,aAAa,EAAE,KAAkD,CAAC;QAC3F,IAAI,CAAC,WAAW,EAAE;YAChB,OAAO;SACR;AACD,QAAA,MAAM,UAAU,GAAG;AACjB,YAAA,YAAY,EAAE,CAAC,UAAU,EAAE,YAAY,IAAI,CAAC;AAC1C,mBAAC,UAAU,EAAE,aAAa,IAAI,CAAC,CAAC;AAChC,mBAAC,UAAU,EAAE,2BAA2B,IAAI,CAAC,CAAC;AAC9C,mBAAC,UAAU,EAAE,uBAAuB,IAAI,CAAC,CAAC;AAC1C,mBAAC,WAAW,CAAC,YAAY,IAAI,CAAC,CAAC;AAC/B,mBAAC,WAAW,CAAC,aAAa,IAAI,CAAC,CAAC;AAChC,mBAAC,WAAW,CAAC,2BAA2B,IAAI,CAAC,CAAC;AAC9C,mBAAC,WAAW,CAAC,uBAAuB,IAAI,CAAC,CAAC;SAC7C,CAAC;AAEF,QAAA,IAAI,CAAC,aAAa,GAAG,IAAI,CAAC;QAC1B,OAAO,MAAM,CAAC,MAAM,CAAC,UAAU,EAAE,UAAU,EAAE,WAAW,CAAC,CAAC;KAC3D;IAED,gBAAgB,GAAA;AACd,QAAA,IAAI,CAAC,aAAa,GAAG,SAAS,CAAC;AAC/B,QAAA,IAAI,CAAC,aAAa,GAAG,SAAS,CAAC;AAC/B,QAAA,IAAI,CAAC,aAAa,GAAG,SAAS,CAAC;AAC/B,QAAA,IAAI,CAAC,eAAe,GAAG,SAAS,CAAC;KAClC;IAEO,qBAAqB,CAAC,EAC5B,KAAK,EACL,KAAK,EACL,aAAa,EACb,iBAAiB,GAMlB,EAAA;AACC,QAAA,MAAM,cAAc,GAAG,iBAAiB,GAAG,aAAa,IAAI,KAAK,CAAC,cAAc,GAAG,SAAS,CAAC;QAC7F,OAAO,IAAIC,2BAAmB,CAAC;YAC7B,OAAO,EAAE,IAAIF,uBAAc,CAAC;;gBAE1B,OAAO,EAAE,KAAK,CAAC,OAAO;gBACtB,iBAAiB,EAAE,KAAK,CAAC,iBAAiB;gBAC1C,gBAAgB,EAAE,KAAK,CAAC,gBAAgB;gBACxC,iBAAiB,EAAE,KAAK,CAAC,iBAAiB;gBAC1C,cAAc;gBACd,EAAE,EAAE,KAAK,CAAC,EAAE;aACb,CAAC;YACF,IAAI,EAAE,KAAK,IAAI,EAAE;AAClB,SAAA,CAAC,CAAC;KACJ;IAED,OAAO,qBAAqB,CAC1B,QAAuB,EACvB,OAAkC,EAClC,UAAqC,EAAA;QAErC,MAAM,MAAM,GAAG,IAAI,CAAC,gBAAgB,CAAC,OAAO,CAAC,CAAC;AAC9C,QAAA,MAAM,iBAAiB,GAAGG,iDAAkC,CAAC,QAAQ,CAAC,CAAC;AACvE,QAAA,IAAI,CAAC,eAAe,GAAG,cAAc,CAAC;AACpC,YAAA,GAAG,MAAM;AACT,YAAA,GAAG,iBAAiB;AACpB,YAAA,MAAM,EAAE,KAAK;AACd,SAAA,CAAC,CAAC;AACH,QAAA,MAAM,qBAAqB,GAAG,CAAC,IAAI,CAAC,eAAe,CAAC;AAEpD,QAAA,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,qBAAqB,CAC7C;AACE,YAAA,GAAG,MAAM;AACT,YAAA,GAAG,iBAAiB;AACpB,YAAA,MAAM,EAAE,IAAI;SACb,EACD;YACE,OAAO,EAAE,OAAO,CAAC,OAAO;AACzB,SAAA,CACF,CAAC;QAEF,MAAM,iBAAiB,GAAG,IAAI,CAAC,WAAW,IAAI,OAAO,CAAC,WAAW,CAAC;AAElE,QAAA,WAAW,MAAM,IAAI,IAAI,MAAM,EAAE;YAC/B,IAAI,OAAO,CAAC,MAAM,EAAE,OAAO,KAAK,IAAI,EAAE;AACpC,gBAAA,MAAM,CAAC,UAAU,CAAC,KAAK,EAAE,CAAC;AAC1B,gBAAA,MAAM,IAAI,KAAK,CAAC,uCAAuC,CAAC,CAAC;aAC1D;AAED,YAAA,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,IAAI,EAAE,CAAC;AAC7B,YAAA,IAAI,IAAI,KAAK,eAAe,EAAE;AAC5B,gBAAA,IAAI,CAAC,aAAa,GAAG,IAAkC,CAAC;aACzD;AAAM,iBAAA,IAAI,IAAI,KAAK,eAAe,EAAE;AACnC,gBAAA,IAAI,CAAC,aAAa,GAAG,IAAkC,CAAC;aACzD;AAED,YAAA,IAAI,aAA+C,CAAC;AACpD,YAAA,IAAI,IAAI,CAAC,eAAe,KAAK,IAAI,IAAI,IAAI,CAAC,aAAa,KAAK,IAAI,EAAE;AAChE,gBAAA,aAAa,GAAG,IAAI,CAAC,cAAc,EAAE,CAAC;aACvC;AAED,YAAA,MAAM,MAAM,GAAGC,mDAAmC,CAAC,IAAI,EAAE;AACvD,gBAAA,WAAW,EAAE,iBAAiB;gBAC9B,qBAAqB;AACtB,aAAA,CAAC,CAAC;AACH,YAAA,IAAI,CAAC,MAAM;gBAAE,SAAS;AAEtB,YAAA,MAAM,EAAE,KAAK,EAAE,GAAG,MAAM,CAAC;AACzB,YAAA,MAAM,CAAC,KAAK,GAAG,EAAE,EAAE,SAAS,CAAC,GAAG,YAAY,CAAC,KAAK,CAAC,CAAC;AAEpD,YAAA,IAAI,CAAC,SAAS,IAAI,SAAS,KAAK,OAAO,KAAK,KAAK,KAAK,EAAE,IAAI,aAAa,CAAC,EAAE;AAC1E,gBAAA,MAAM,eAAe,GAAG,IAAI,CAAC,qBAAqB,CAAC;oBACjD,KAAK;oBACL,KAAK;oBACL,aAAa;oBACb,iBAAiB;AAClB,iBAAA,CAAC,CAAC;AACH,gBAAA,MAAM,eAAe,CAAC;gBACtB,MAAM,UAAU,EAAE,iBAAiB,CACjC,KAAK,EACL,SAAS,EACT,SAAS,EACT,SAAS,EACT,SAAS,EACT,EAAE,KAAK,EAAE,eAAe,EAAE,CAC3B,CAAC;gBACF,SAAS;aACV;AAED,YAAA,MAAM,UAAU,GAAG,IAAIC,eAAU,CAAC,KAAK,EAAE;gBACvC,KAAK,EAAE,IAAI,CAAC,gBAAgB;AAC5B,gBAAA,cAAc,EAAE,IAAI;AACpB,gBAAA,YAAY,EAAE,CAAC;AACf,gBAAA,YAAY,EAAE,CAAC;AAChB,aAAA,CAAC,CAAC;AAEH,YAAA,MAAM,SAAS,GAAG,UAAU,CAAC,YAAY,EAAE,CAAC;AAC5C,YAAA,IAAI;gBACF,IAAI,YAAY,GAAG,KAAK,CAAC;AACzB,gBAAA,WAAW,MAAM,YAAY,IAAI,SAAS,EAAE;oBAC1C,MAAM,QAAQ,GAAG,UAAU,CAAC,YAAY,EAAE,SAAS,EAAE,KAAK,CAAC,CAAC;AAE5D,oBAAA,MAAM,eAAe,GAAG,IAAI,CAAC,qBAAqB,CAAC;AACjD,wBAAA,KAAK,EAAE,YAAY;AACnB,wBAAA,KAAK,EAAE,QAAQ;wBACf,aAAa,EAAE,YAAY,GAAG,SAAS,GAAG,aAAa;wBACvD,iBAAiB;AAClB,qBAAA,CAAC,CAAC;AAEH,oBAAA,IAAI,aAAa,IAAI,CAAC,YAAY,EAAE;wBAClC,YAAY,GAAG,IAAI,CAAC;qBACrB;AACD,oBAAA,MAAM,eAAe,CAAC;oBAEtB,MAAM,UAAU,EAAE,iBAAiB,CACjC,KAAK,EACL,SAAS,EACT,SAAS,EACT,SAAS,EACT,SAAS,EACT,EAAE,KAAK,EAAE,eAAe,EAAE,CAC3B,CAAC;iBACH;aACF;oBAAS;AACR,gBAAA,MAAM,SAAS,CAAC,MAAM,EAAE,CAAC;aAC1B;SACF;QAED,IAAI,CAAC,gBAAgB,EAAE,CAAC;KACzB;AACF;;;;"}

package/dist/esm/llm/anthropic/llm.mjs

@@ -44,19 +44,56 @@ function cloneChunk(text, tokenType, chunk) {
 }
 class CustomAnthropic extends ChatAnthropicMessages {
     _lc_stream_delay;
+    message_start;
+    message_delta;
+    tools_in_params;
+    emitted_usage;
     constructor(fields) {
         super(fields);
         this._lc_stream_delay = fields._lc_stream_delay ?? 25;
     }
-    createGenerationChunk({ token, chunk, shouldStreamUsage, }) {
+    /**
+     * Get stream usage as returned by this client's API response.
+     * @returns {AnthropicStreamUsage} The stream usage object.
+     */
+    getStreamUsage() {
+        if (this.emitted_usage === true) {
+            return;
+        }
+        const inputUsage = (this.message_start?.message)?.usage;
+        const outputUsage = this.message_delta?.usage;
+        if (!outputUsage) {
+            return;
+        }
+        const totalUsage = {
+            total_tokens: (inputUsage?.input_tokens ?? 0)
+                + (inputUsage?.output_tokens ?? 0)
+                + (inputUsage?.cache_creation_input_tokens ?? 0)
+                + (inputUsage?.cache_read_input_tokens ?? 0)
+                + (outputUsage.input_tokens ?? 0)
+                + (outputUsage.output_tokens ?? 0)
+                + (outputUsage.cache_creation_input_tokens ?? 0)
+                + (outputUsage.cache_read_input_tokens ?? 0),
+        };
+        this.emitted_usage = true;
+        return Object.assign(totalUsage, inputUsage, outputUsage);
+    }
+    resetTokenEvents() {
+        this.message_start = undefined;
+        this.message_delta = undefined;
+        this.emitted_usage = undefined;
+        this.tools_in_params = undefined;
+    }
+    createGenerationChunk({ token, chunk, usageMetadata, shouldStreamUsage, }) {
+        const usage_metadata = shouldStreamUsage ? usageMetadata ?? chunk.usage_metadata : undefined;
         return new ChatGenerationChunk({
             message: new AIMessageChunk({
                 // Just yield chunk as it is and tool_use will be concat by BaseChatModel._generateUncached().
                 content: chunk.content,
                 additional_kwargs: chunk.additional_kwargs,
                 tool_call_chunks: chunk.tool_call_chunks,
-                usage_metadata: shouldStreamUsage ? chunk.usage_metadata : undefined,
                 response_metadata: chunk.response_metadata,
+                usage_metadata,
                 id: chunk.id,
             }),
             text: token ?? '',
@@ -65,11 +102,12 @@ class CustomAnthropic extends ChatAnthropicMessages {
     async *_streamResponseChunks(messages, options, runManager) {
         const params = this.invocationParams(options);
         const formattedMessages = _convertMessagesToAnthropicPayload(messages);
-        const coerceContentToString = !_toolsInParams({
+        this.tools_in_params = _toolsInParams({
             ...params,
             ...formattedMessages,
             stream: false,
         });
+        const coerceContentToString = !this.tools_in_params;
         const stream = await this.createStreamWithRetry({
             ...params,
             ...formattedMessages,
@@ -77,12 +115,23 @@ class CustomAnthropic extends ChatAnthropicMessages {
         }, {
             headers: options.headers,
         });
+        const shouldStreamUsage = this.streamUsage ?? options.streamUsage;
         for await (const data of stream) {
             if (options.signal?.aborted === true) {
                 stream.controller.abort();
                 throw new Error('AbortError: User aborted the request.');
             }
-            const shouldStreamUsage = this.streamUsage ?? options.streamUsage;
+            const type = data.type ?? '';
+            if (type === 'message_start') {
+                this.message_start = data;
+            }
+            else if (type === 'message_delta') {
+                this.message_delta = data;
+            }
+            let usageMetadata;
+            if (this.tools_in_params !== true && this.emitted_usage !== true) {
+                usageMetadata = this.getStreamUsage();
+            }
             const result = _makeMessageChunkFromAnthropicEvent(data, {
                 streamUsage: shouldStreamUsage,
                 coerceContentToString,
@@ -91,10 +140,11 @@ class CustomAnthropic extends ChatAnthropicMessages {
                 continue;
             const { chunk } = result;
             const [token = '', tokenType] = extractToken(chunk);
-            if (!tokenType || tokenType === 'input') {
+            if (!tokenType || tokenType === 'input' || (token === '' && usageMetadata)) {
                 const generationChunk = this.createGenerationChunk({
                     token,
                     chunk,
+                    usageMetadata,
                     shouldStreamUsage,
                 });
                 yield generationChunk;
@@ -109,13 +159,18 @@ class CustomAnthropic extends ChatAnthropicMessages {
             });
             const generator = textStream.generateText();
             try {
+                let emittedUsage = false;
                 for await (const currentToken of generator) {
                     const newChunk = cloneChunk(currentToken, tokenType, chunk);
                     const generationChunk = this.createGenerationChunk({
                         token: currentToken,
                         chunk: newChunk,
+                        usageMetadata: emittedUsage ? undefined : usageMetadata,
                         shouldStreamUsage,
                     });
+                    if (usageMetadata && !emittedUsage) {
+                        emittedUsage = true;
+                    }
                     yield generationChunk;
                     await runManager?.handleLLMNewToken(token, undefined, undefined, undefined, undefined, { chunk: generationChunk });
                 }
@@ -124,6 +179,7 @@ class CustomAnthropic extends ChatAnthropicMessages {
                 await generator.return();
             }
         }
+        this.resetTokenEvents();
     }
 }
 

package/dist/esm/llm/anthropic/llm.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"llm.mjs","sources":["../../../../src/llm/anthropic/llm.ts"],"sourcesContent":["import { AIMessageChunk } from '@langchain/core/messages';\nimport { ChatAnthropicMessages } from '@langchain/anthropic';\nimport { ChatGenerationChunk } from '@langchain/core/outputs';\nimport type { BaseMessage, MessageContentComplex } from '@langchain/core/messages';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { AnthropicInput } from '@langchain/anthropic';\nimport type { AnthropicMessageCreateParams } from '@/llm/anthropic/types';\nimport { _makeMessageChunkFromAnthropicEvent } from './utils/message_outputs';\nimport { _convertMessagesToAnthropicPayload } from './utils/message_inputs';\nimport { TextStream } from '@/llm/text';\n\nfunction _toolsInParams(params: AnthropicMessageCreateParams): boolean {\n  return !!(params.tools && params.tools.length > 0);\n}\n\nfunction extractToken(chunk: AIMessageChunk): [string, 'string' | 'input' | 'content'] | [undefined] {\n  if (typeof chunk.content === 'string') {\n    return [chunk.content, 'string'];\n  } else if (\n    Array.isArray(chunk.content) &&\n    chunk.content.length >= 1 &&\n    'input' in chunk.content[0]\n  ) {\n    return typeof chunk.content[0].input === 'string'\n      ? [chunk.content[0].input, 'input']\n      : [JSON.stringify(chunk.content[0].input), 'input'];\n  } else if (\n    Array.isArray(chunk.content) &&\n    chunk.content.length >= 1 &&\n    'text' in chunk.content[0]\n  ) {\n    return [chunk.content[0].text, 'content'];\n  }\n  return [undefined];\n}\n\nfunction cloneChunk(text: string, tokenType: string, chunk: AIMessageChunk): AIMessageChunk {\n  if (tokenType === 'string') {\n    return new AIMessageChunk(Object.assign({}, chunk, { content: text }));\n  } else if (tokenType === 'input') {\n    return chunk;\n  }\n  const content = chunk.content[0] as MessageContentComplex;\n  if (tokenType === 'content' && content.type === 'text') {\n    return new AIMessageChunk(Object.assign({}, chunk, { content: [Object.assign({}, content, { text })] }));\n  } else if (tokenType === 'content' && content.type === 'text_delta') {\n    return new AIMessageChunk(Object.assign({}, chunk, { content: [Object.assign({}, content, { text })] }));\n  }\n\n  return chunk;\n}\n\nexport type CustomAnthropicInput = AnthropicInput & { _lc_stream_delay?: number };\n\nexport class CustomAnthropic extends ChatAnthropicMessages {\n  _lc_stream_delay: number;\n  constructor(fields: CustomAnthropicInput) {\n    super(fields);\n    this._lc_stream_delay = fields._lc_stream_delay ?? 25;\n  }\n\n  private createGenerationChunk({\n    token,\n    chunk,\n    shouldStreamUsage,\n  }: {\n    token?: string,\n    chunk: AIMessageChunk,\n    shouldStreamUsage: boolean\n  }): ChatGenerationChunk {\n    return new ChatGenerationChunk({\n      message: new AIMessageChunk({\n        // Just yield chunk as it is and tool_use will be concat by BaseChatModel._generateUncached().\n        content: chunk.content,\n        additional_kwargs: chunk.additional_kwargs,\n        tool_call_chunks: chunk.tool_call_chunks,\n        usage_metadata: shouldStreamUsage ? chunk.usage_metadata : undefined,\n        response_metadata: chunk.response_metadata,\n        id: chunk.id,\n      }),\n      text: token ?? '',\n    });\n  }\n\n  async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'],\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    const params = this.invocationParams(options);\n    const formattedMessages = _convertMessagesToAnthropicPayload(messages);\n    const coerceContentToString = !_toolsInParams({\n      ...params,\n      ...formattedMessages,\n      stream: false,\n    });\n\n    const stream = await this.createStreamWithRetry(\n      {\n        ...params,\n        ...formattedMessages,\n        stream: true,\n      },\n      {\n        headers: options.headers,\n      }\n    );\n\n    for await (const data of stream) {\n      if (options.signal?.aborted === true) {\n        stream.controller.abort();\n        throw new Error('AbortError: User aborted the request.');\n      }\n\n      const shouldStreamUsage = this.streamUsage ?? options.streamUsage;\n      const result = _makeMessageChunkFromAnthropicEvent(data, {\n        streamUsage: shouldStreamUsage,\n        coerceContentToString,\n      });\n      if (!result) continue;\n\n      const { chunk } = result;\n      const [token = '', tokenType] = extractToken(chunk);\n\n      if (!tokenType || tokenType === 'input') {\n        const generationChunk = this.createGenerationChunk({\n          token,\n          chunk,\n          shouldStreamUsage,\n        });\n        yield generationChunk;\n        await runManager?.handleLLMNewToken(\n          token,\n          undefined,\n          undefined,\n          undefined,\n          undefined,\n          { chunk: generationChunk }\n        );\n        continue;\n      }\n\n      const textStream = new TextStream(token, {\n        delay: this._lc_stream_delay,\n        firstWordChunk: true,\n        minChunkSize: 4,\n        maxChunkSize: 8,\n      });\n\n      const generator = textStream.generateText();\n      try {\n        for await (const currentToken of generator) {\n          const newChunk = cloneChunk(currentToken, tokenType, chunk);\n          const generationChunk = this.createGenerationChunk({\n            token: currentToken,\n            chunk: newChunk,\n            shouldStreamUsage,\n          });\n          yield generationChunk;\n\n          await runManager?.handleLLMNewToken(\n            token,\n            undefined,\n            undefined,\n            undefined,\n            undefined,\n            { chunk: generationChunk }\n          );\n        }\n      } finally {\n        await generator.return();\n      }\n    }\n  }\n}"],"names":[],"mappings":";;;;;;;AAWA,SAAS,cAAc,CAAC,MAAoC,EAAA;AAC1D,IAAA,OAAO,CAAC,EAAE,MAAM,CAAC,KAAK,IAAI,MAAM,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;AACrD,CAAC;AAED,SAAS,YAAY,CAAC,KAAqB,EAAA;AACzC,IAAA,IAAI,OAAO,KAAK,CAAC,OAAO,KAAK,QAAQ,EAAE;AACrC,QAAA,OAAO,CAAC,KAAK,CAAC,OAAO,EAAE,QAAQ,CAAC,CAAC;KAClC;AAAM,SAAA,IACL,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,OAAO,CAAC;AAC5B,QAAA,KAAK,CAAC,OAAO,CAAC,MAAM,IAAI,CAAC;QACzB,OAAO,IAAI,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,EAC3B;QACA,OAAO,OAAO,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,KAAK,KAAK,QAAQ;AAC/C,cAAE,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,KAAK,EAAE,OAAO,CAAC;AACnC,cAAE,CAAC,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,EAAE,OAAO,CAAC,CAAC;KACvD;AAAM,SAAA,IACL,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,OAAO,CAAC;AAC5B,QAAA,KAAK,CAAC,OAAO,CAAC,MAAM,IAAI,CAAC;QACzB,MAAM,IAAI,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,EAC1B;AACA,QAAA,OAAO,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,SAAS,CAAC,CAAC;KAC3C;IACD,OAAO,CAAC,SAAS,CAAC,CAAC;AACrB,CAAC;AAED,SAAS,UAAU,CAAC,IAAY,EAAE,SAAiB,EAAE,KAAqB,EAAA;AACxE,IAAA,IAAI,SAAS,KAAK,QAAQ,EAAE;AAC1B,QAAA,OAAO,IAAI,cAAc,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,EAAE,KAAK,EAAE,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC;KACxE;AAAM,SAAA,IAAI,SAAS,KAAK,OAAO,EAAE;AAChC,QAAA,OAAO,KAAK,CAAC;KACd;IACD,MAAM,OAAO,GAAG,KAAK,CAAC,OAAO,CAAC,CAAC,CAA0B,CAAC;IAC1D,IAAI,SAAS,KAAK,SAAS,IAAI,OAAO,CAAC,IAAI,KAAK,MAAM,EAAE;AACtD,QAAA,OAAO,IAAI,cAAc,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,EAAE,KAAK,EAAE,EAAE,OAAO,EAAE,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,EAAE,OAAO,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;KAC1G;SAAM,IAAI,SAAS,KAAK,SAAS,IAAI,OAAO,CAAC,IAAI,KAAK,YAAY,EAAE;AACnE,QAAA,OAAO,IAAI,cAAc,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,EAAE,KAAK,EAAE,EAAE,OAAO,EAAE,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,EAAE,OAAO,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;KAC1G;AAED,IAAA,OAAO,KAAK,CAAC;AACf,CAAC;AAIK,MAAO,eAAgB,SAAQ,qBAAqB,CAAA;AACxD,IAAA,gBAAgB,CAAS;AACzB,IAAA,WAAA,CAAY,MAA4B,EAAA;QACtC,KAAK,CAAC,MAAM,CAAC,CAAC;QACd,IAAI,CAAC,gBAAgB,GAAG,MAAM,CAAC,gBAAgB,IAAI,EAAE,CAAC;KACvD;AAEO,IAAA,qBAAqB,CAAC,EAC5B,KAAK,EACL,KAAK,EACL,iBAAiB,GAKlB,EAAA;QACC,OAAO,IAAI,mBAAmB,CAAC;YAC7B,OAAO,EAAE,IAAI,cAAc,CAAC;;gBAE1B,OAAO,EAAE,KAAK,CAAC,OAAO;gBACtB,iBAAiB,EAAE,KAAK,CAAC,iBAAiB;gBAC1C,gBAAgB,EAAE,KAAK,CAAC,gBAAgB;gBACxC,cAAc,EAAE,iBAAiB,GAAG,KAAK,CAAC,cAAc,GAAG,SAAS;gBACpE,iBAAiB,EAAE,KAAK,CAAC,iBAAiB;gBAC1C,EAAE,EAAE,KAAK,CAAC,EAAE;aACb,CAAC;YACF,IAAI,EAAE,KAAK,IAAI,EAAE;AAClB,SAAA,CAAC,CAAC;KACJ;IAED,OAAO,qBAAqB,CAC1B,QAAuB,EACvB,OAAkC,EAClC,UAAqC,EAAA;QAErC,MAAM,MAAM,GAAG,IAAI,CAAC,gBAAgB,CAAC,OAAO,CAAC,CAAC;AAC9C,QAAA,MAAM,iBAAiB,GAAG,kCAAkC,CAAC,QAAQ,CAAC,CAAC;AACvE,QAAA,MAAM,qBAAqB,GAAG,CAAC,cAAc,CAAC;AAC5C,YAAA,GAAG,MAAM;AACT,YAAA,GAAG,iBAAiB;AACpB,YAAA,MAAM,EAAE,KAAK;AACd,SAAA,CAAC,CAAC;AAEH,QAAA,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,qBAAqB,CAC7C;AACE,YAAA,GAAG,MAAM;AACT,YAAA,GAAG,iBAAiB;AACpB,YAAA,MAAM,EAAE,IAAI;SACb,EACD;YACE,OAAO,EAAE,OAAO,CAAC,OAAO;AACzB,SAAA,CACF,CAAC;AAEF,QAAA,WAAW,MAAM,IAAI,IAAI,MAAM,EAAE;YAC/B,IAAI,OAAO,CAAC,MAAM,EAAE,OAAO,KAAK,IAAI,EAAE;AACpC,gBAAA,MAAM,CAAC,UAAU,CAAC,KAAK,EAAE,CAAC;AAC1B,gBAAA,MAAM,IAAI,KAAK,CAAC,uCAAuC,CAAC,CAAC;aAC1D;YAED,MAAM,iBAAiB,GAAG,IAAI,CAAC,WAAW,IAAI,OAAO,CAAC,WAAW,CAAC;AAClE,YAAA,MAAM,MAAM,GAAG,mCAAmC,CAAC,IAAI,EAAE;AACvD,gBAAA,WAAW,EAAE,iBAAiB;gBAC9B,qBAAqB;AACtB,aAAA,CAAC,CAAC;AACH,YAAA,IAAI,CAAC,MAAM;gBAAE,SAAS;AAEtB,YAAA,MAAM,EAAE,KAAK,EAAE,GAAG,MAAM,CAAC;AACzB,YAAA,MAAM,CAAC,KAAK,GAAG,EAAE,EAAE,SAAS,CAAC,GAAG,YAAY,CAAC,KAAK,CAAC,CAAC;AAEpD,YAAA,IAAI,CAAC,SAAS,IAAI,SAAS,KAAK,OAAO,EAAE;AACvC,gBAAA,MAAM,eAAe,GAAG,IAAI,CAAC,qBAAqB,CAAC;oBACjD,KAAK;oBACL,KAAK;oBACL,iBAAiB;AAClB,iBAAA,CAAC,CAAC;AACH,gBAAA,MAAM,eAAe,CAAC;gBACtB,MAAM,UAAU,EAAE,iBAAiB,CACjC,KAAK,EACL,SAAS,EACT,SAAS,EACT,SAAS,EACT,SAAS,EACT,EAAE,KAAK,EAAE,eAAe,EAAE,CAC3B,CAAC;gBACF,SAAS;aACV;AAED,YAAA,MAAM,UAAU,GAAG,IAAI,UAAU,CAAC,KAAK,EAAE;gBACvC,KAAK,EAAE,IAAI,CAAC,gBAAgB;AAC5B,gBAAA,cAAc,EAAE,IAAI;AACpB,gBAAA,YAAY,EAAE,CAAC;AACf,gBAAA,YAAY,EAAE,CAAC;AAChB,aAAA,CAAC,CAAC;AAEH,YAAA,MAAM,SAAS,GAAG,UAAU,CAAC,YAAY,EAAE,CAAC;AAC5C,YAAA,IAAI;AACF,gBAAA,WAAW,MAAM,YAAY,IAAI,SAAS,EAAE;oBAC1C,MAAM,QAAQ,GAAG,UAAU,CAAC,YAAY,EAAE,SAAS,EAAE,KAAK,CAAC,CAAC;AAC5D,oBAAA,MAAM,eAAe,GAAG,IAAI,CAAC,qBAAqB,CAAC;AACjD,wBAAA,KAAK,EAAE,YAAY;AACnB,wBAAA,KAAK,EAAE,QAAQ;wBACf,iBAAiB;AAClB,qBAAA,CAAC,CAAC;AACH,oBAAA,MAAM,eAAe,CAAC;oBAEtB,MAAM,UAAU,EAAE,iBAAiB,CACjC,KAAK,EACL,SAAS,EACT,SAAS,EACT,SAAS,EACT,SAAS,EACT,EAAE,KAAK,EAAE,eAAe,EAAE,CAC3B,CAAC;iBACH;aACF;oBAAS;AACR,gBAAA,MAAM,SAAS,CAAC,MAAM,EAAE,CAAC;aAC1B;SACF;KACF;AACF;;;;"}
+{"version":3,"file":"llm.mjs","sources":["../../../../src/llm/anthropic/llm.ts"],"sourcesContent":["import { AIMessageChunk } from '@langchain/core/messages';\nimport { ChatAnthropicMessages } from '@langchain/anthropic';\nimport { ChatGenerationChunk } from '@langchain/core/outputs';\nimport type { BaseMessage, MessageContentComplex } from '@langchain/core/messages';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { AnthropicInput } from '@langchain/anthropic';\nimport type { AnthropicMessageCreateParams, AnthropicStreamUsage, AnthropicMessageStartEvent, AnthropicMessageDeltaEvent } from '@/llm/anthropic/types';\nimport { _makeMessageChunkFromAnthropicEvent } from './utils/message_outputs';\nimport { _convertMessagesToAnthropicPayload } from './utils/message_inputs';\nimport { TextStream } from '@/llm/text';\n\nfunction _toolsInParams(params: AnthropicMessageCreateParams): boolean {\n  return !!(params.tools && params.tools.length > 0);\n}\n\nfunction extractToken(chunk: AIMessageChunk): [string, 'string' | 'input' | 'content'] | [undefined] {\n  if (typeof chunk.content === 'string') {\n    return [chunk.content, 'string'];\n  } else if (\n    Array.isArray(chunk.content) &&\n    chunk.content.length >= 1 &&\n    'input' in chunk.content[0]\n  ) {\n    return typeof chunk.content[0].input === 'string'\n      ? [chunk.content[0].input, 'input']\n      : [JSON.stringify(chunk.content[0].input), 'input'];\n  } else if (\n    Array.isArray(chunk.content) &&\n    chunk.content.length >= 1 &&\n    'text' in chunk.content[0]\n  ) {\n    return [chunk.content[0].text, 'content'];\n  }\n  return [undefined];\n}\n\nfunction cloneChunk(text: string, tokenType: string, chunk: AIMessageChunk): AIMessageChunk {\n  if (tokenType === 'string') {\n    return new AIMessageChunk(Object.assign({}, chunk, { content: text }));\n  } else if (tokenType === 'input') {\n    return chunk;\n  }\n  const content = chunk.content[0] as MessageContentComplex;\n  if (tokenType === 'content' && content.type === 'text') {\n    return new AIMessageChunk(Object.assign({}, chunk, { content: [Object.assign({}, content, { text })] }));\n  } else if (tokenType === 'content' && content.type === 'text_delta') {\n    return new AIMessageChunk(Object.assign({}, chunk, { content: [Object.assign({}, content, { text })] }));\n  }\n\n  return chunk;\n}\n\nexport type CustomAnthropicInput = AnthropicInput & { _lc_stream_delay?: number };\n\nexport class CustomAnthropic extends ChatAnthropicMessages {\n  _lc_stream_delay: number;\n  private message_start: AnthropicMessageStartEvent | undefined;\n  private message_delta: AnthropicMessageDeltaEvent | undefined;\n  private tools_in_params?: boolean;\n  private emitted_usage?: boolean;\n  constructor(fields: CustomAnthropicInput) {\n    super(fields);\n    this._lc_stream_delay = fields._lc_stream_delay ?? 25;\n  }\n\n  /**\n   * Get stream usage as returned by this client's API response.\n   * @returns {AnthropicStreamUsage} The stream usage object.\n   */\n  getStreamUsage(): AnthropicStreamUsage | undefined {\n    if (this.emitted_usage === true) {\n      return;\n    }\n    const inputUsage = (this.message_start?.message)?.usage as undefined | AnthropicStreamUsage;\n    const outputUsage = this.message_delta?.usage as undefined | Partial<AnthropicStreamUsage>;\n    if (!outputUsage) {\n      return;\n    }\n    const totalUsage = {\n      total_tokens: (inputUsage?.input_tokens ?? 0)\n      + (inputUsage?.output_tokens ?? 0)\n      + (inputUsage?.cache_creation_input_tokens ?? 0)\n      + (inputUsage?.cache_read_input_tokens ?? 0)\n      + (outputUsage.input_tokens ?? 0)\n      + (outputUsage.output_tokens ?? 0)\n      + (outputUsage.cache_creation_input_tokens ?? 0)\n      + (outputUsage.cache_read_input_tokens ?? 0),\n    };\n\n    this.emitted_usage = true;\n    return Object.assign(totalUsage, inputUsage, outputUsage);\n  }\n\n  resetTokenEvents(): void {\n    this.message_start = undefined;\n    this.message_delta = undefined;\n    this.emitted_usage = undefined;\n    this.tools_in_params = undefined;\n  }\n\n  private createGenerationChunk({\n    token,\n    chunk,\n    usageMetadata,\n    shouldStreamUsage,\n  }: {\n    token?: string,\n    chunk: AIMessageChunk,\n    shouldStreamUsage: boolean\n    usageMetadata?: AnthropicStreamUsage,\n  }): ChatGenerationChunk {\n    const usage_metadata = shouldStreamUsage ? usageMetadata ?? chunk.usage_metadata : undefined;\n    return new ChatGenerationChunk({\n      message: new AIMessageChunk({\n        // Just yield chunk as it is and tool_use will be concat by BaseChatModel._generateUncached().\n        content: chunk.content,\n        additional_kwargs: chunk.additional_kwargs,\n        tool_call_chunks: chunk.tool_call_chunks,\n        response_metadata: chunk.response_metadata,\n        usage_metadata,\n        id: chunk.id,\n      }),\n      text: token ?? '',\n    });\n  }\n\n  async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'],\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    const params = this.invocationParams(options);\n    const formattedMessages = _convertMessagesToAnthropicPayload(messages);\n    this.tools_in_params = _toolsInParams({\n      ...params,\n      ...formattedMessages,\n      stream: false,\n    });\n    const coerceContentToString = !this.tools_in_params;\n\n    const stream = await this.createStreamWithRetry(\n      {\n        ...params,\n        ...formattedMessages,\n        stream: true,\n      },\n      {\n        headers: options.headers,\n      }\n    );\n\n    const shouldStreamUsage = this.streamUsage ?? options.streamUsage;\n\n    for await (const data of stream) {\n      if (options.signal?.aborted === true) {\n        stream.controller.abort();\n        throw new Error('AbortError: User aborted the request.');\n      }\n\n      const type = data.type ?? '';\n      if (type === 'message_start') {\n        this.message_start = data as AnthropicMessageStartEvent;\n      } else if (type === 'message_delta') {\n        this.message_delta = data as AnthropicMessageDeltaEvent;\n      }\n\n      let usageMetadata: AnthropicStreamUsage | undefined;\n      if (this.tools_in_params !== true && this.emitted_usage !== true) {\n        usageMetadata = this.getStreamUsage();\n      }\n\n      const result = _makeMessageChunkFromAnthropicEvent(data, {\n        streamUsage: shouldStreamUsage,\n        coerceContentToString,\n      });\n      if (!result) continue;\n\n      const { chunk } = result;\n      const [token = '', tokenType] = extractToken(chunk);\n\n      if (!tokenType || tokenType === 'input' || (token === '' && usageMetadata)) {\n        const generationChunk = this.createGenerationChunk({\n          token,\n          chunk,\n          usageMetadata,\n          shouldStreamUsage,\n        });\n        yield generationChunk;\n        await runManager?.handleLLMNewToken(\n          token,\n          undefined,\n          undefined,\n          undefined,\n          undefined,\n          { chunk: generationChunk }\n        );\n        continue;\n      }\n\n      const textStream = new TextStream(token, {\n        delay: this._lc_stream_delay,\n        firstWordChunk: true,\n        minChunkSize: 4,\n        maxChunkSize: 8,\n      });\n\n      const generator = textStream.generateText();\n      try {\n        let emittedUsage = false;\n        for await (const currentToken of generator) {\n          const newChunk = cloneChunk(currentToken, tokenType, chunk);\n\n          const generationChunk = this.createGenerationChunk({\n            token: currentToken,\n            chunk: newChunk,\n            usageMetadata: emittedUsage ? undefined : usageMetadata,\n            shouldStreamUsage,\n          });\n\n          if (usageMetadata && !emittedUsage) {\n            emittedUsage = true;\n          }\n          yield generationChunk;\n\n          await runManager?.handleLLMNewToken(\n            token,\n            undefined,\n            undefined,\n            undefined,\n            undefined,\n            { chunk: generationChunk }\n          );\n        }\n      } finally {\n        await generator.return();\n      }\n    }\n\n    this.resetTokenEvents();\n  }\n}"],"names":[],"mappings":";;;;;;;AAWA,SAAS,cAAc,CAAC,MAAoC,EAAA;AAC1D,IAAA,OAAO,CAAC,EAAE,MAAM,CAAC,KAAK,IAAI,MAAM,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;AACrD,CAAC;AAED,SAAS,YAAY,CAAC,KAAqB,EAAA;AACzC,IAAA,IAAI,OAAO,KAAK,CAAC,OAAO,KAAK,QAAQ,EAAE;AACrC,QAAA,OAAO,CAAC,KAAK,CAAC,OAAO,EAAE,QAAQ,CAAC,CAAC;KAClC;AAAM,SAAA,IACL,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,OAAO,CAAC;AAC5B,QAAA,KAAK,CAAC,OAAO,CAAC,MAAM,IAAI,CAAC;QACzB,OAAO,IAAI,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,EAC3B;QACA,OAAO,OAAO,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,KAAK,KAAK,QAAQ;AAC/C,cAAE,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,KAAK,EAAE,OAAO,CAAC;AACnC,cAAE,CAAC,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,EAAE,OAAO,CAAC,CAAC;KACvD;AAAM,SAAA,IACL,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,OAAO,CAAC;AAC5B,QAAA,KAAK,CAAC,OAAO,CAAC,MAAM,IAAI,CAAC;QACzB,MAAM,IAAI,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,EAC1B;AACA,QAAA,OAAO,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,SAAS,CAAC,CAAC;KAC3C;IACD,OAAO,CAAC,SAAS,CAAC,CAAC;AACrB,CAAC;AAED,SAAS,UAAU,CAAC,IAAY,EAAE,SAAiB,EAAE,KAAqB,EAAA;AACxE,IAAA,IAAI,SAAS,KAAK,QAAQ,EAAE;AAC1B,QAAA,OAAO,IAAI,cAAc,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,EAAE,KAAK,EAAE,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC;KACxE;AAAM,SAAA,IAAI,SAAS,KAAK,OAAO,EAAE;AAChC,QAAA,OAAO,KAAK,CAAC;KACd;IACD,MAAM,OAAO,GAAG,KAAK,CAAC,OAAO,CAAC,CAAC,CAA0B,CAAC;IAC1D,IAAI,SAAS,KAAK,SAAS,IAAI,OAAO,CAAC,IAAI,KAAK,MAAM,EAAE;AACtD,QAAA,OAAO,IAAI,cAAc,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,EAAE,KAAK,EAAE,EAAE,OAAO,EAAE,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,EAAE,OAAO,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;KAC1G;SAAM,IAAI,SAAS,KAAK,SAAS,IAAI,OAAO,CAAC,IAAI,KAAK,YAAY,EAAE;AACnE,QAAA,OAAO,IAAI,cAAc,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,EAAE,KAAK,EAAE,EAAE,OAAO,EAAE,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,EAAE,OAAO,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;KAC1G;AAED,IAAA,OAAO,KAAK,CAAC;AACf,CAAC;AAIK,MAAO,eAAgB,SAAQ,qBAAqB,CAAA;AACxD,IAAA,gBAAgB,CAAS;AACjB,IAAA,aAAa,CAAyC;AACtD,IAAA,aAAa,CAAyC;AACtD,IAAA,eAAe,CAAW;AAC1B,IAAA,aAAa,CAAW;AAChC,IAAA,WAAA,CAAY,MAA4B,EAAA;QACtC,KAAK,CAAC,MAAM,CAAC,CAAC;QACd,IAAI,CAAC,gBAAgB,GAAG,MAAM,CAAC,gBAAgB,IAAI,EAAE,CAAC;KACvD;AAED;;;AAGG;IACH,cAAc,GAAA;AACZ,QAAA,IAAI,IAAI,CAAC,aAAa,KAAK,IAAI,EAAE;YAC/B,OAAO;SACR;QACD,MAAM,UAAU,GAAG,CAAC,IAAI,CAAC,aAAa,EAAE,OAAO,GAAG,KAAyC,CAAC;AAC5F,QAAA,MAAM,WAAW,GAAG,IAAI,CAAC,aAAa,EAAE,KAAkD,CAAC;QAC3F,IAAI,CAAC,WAAW,EAAE;YAChB,OAAO;SACR;AACD,QAAA,MAAM,UAAU,GAAG;AACjB,YAAA,YAAY,EAAE,CAAC,UAAU,EAAE,YAAY,IAAI,CAAC;AAC1C,mBAAC,UAAU,EAAE,aAAa,IAAI,CAAC,CAAC;AAChC,mBAAC,UAAU,EAAE,2BAA2B,IAAI,CAAC,CAAC;AAC9C,mBAAC,UAAU,EAAE,uBAAuB,IAAI,CAAC,CAAC;AAC1C,mBAAC,WAAW,CAAC,YAAY,IAAI,CAAC,CAAC;AAC/B,mBAAC,WAAW,CAAC,aAAa,IAAI,CAAC,CAAC;AAChC,mBAAC,WAAW,CAAC,2BAA2B,IAAI,CAAC,CAAC;AAC9C,mBAAC,WAAW,CAAC,uBAAuB,IAAI,CAAC,CAAC;SAC7C,CAAC;AAEF,QAAA,IAAI,CAAC,aAAa,GAAG,IAAI,CAAC;QAC1B,OAAO,MAAM,CAAC,MAAM,CAAC,UAAU,EAAE,UAAU,EAAE,WAAW,CAAC,CAAC;KAC3D;IAED,gBAAgB,GAAA;AACd,QAAA,IAAI,CAAC,aAAa,GAAG,SAAS,CAAC;AAC/B,QAAA,IAAI,CAAC,aAAa,GAAG,SAAS,CAAC;AAC/B,QAAA,IAAI,CAAC,aAAa,GAAG,SAAS,CAAC;AAC/B,QAAA,IAAI,CAAC,eAAe,GAAG,SAAS,CAAC;KAClC;IAEO,qBAAqB,CAAC,EAC5B,KAAK,EACL,KAAK,EACL,aAAa,EACb,iBAAiB,GAMlB,EAAA;AACC,QAAA,MAAM,cAAc,GAAG,iBAAiB,GAAG,aAAa,IAAI,KAAK,CAAC,cAAc,GAAG,SAAS,CAAC;QAC7F,OAAO,IAAI,mBAAmB,CAAC;YAC7B,OAAO,EAAE,IAAI,cAAc,CAAC;;gBAE1B,OAAO,EAAE,KAAK,CAAC,OAAO;gBACtB,iBAAiB,EAAE,KAAK,CAAC,iBAAiB;gBAC1C,gBAAgB,EAAE,KAAK,CAAC,gBAAgB;gBACxC,iBAAiB,EAAE,KAAK,CAAC,iBAAiB;gBAC1C,cAAc;gBACd,EAAE,EAAE,KAAK,CAAC,EAAE;aACb,CAAC;YACF,IAAI,EAAE,KAAK,IAAI,EAAE;AAClB,SAAA,CAAC,CAAC;KACJ;IAED,OAAO,qBAAqB,CAC1B,QAAuB,EACvB,OAAkC,EAClC,UAAqC,EAAA;QAErC,MAAM,MAAM,GAAG,IAAI,CAAC,gBAAgB,CAAC,OAAO,CAAC,CAAC;AAC9C,QAAA,MAAM,iBAAiB,GAAG,kCAAkC,CAAC,QAAQ,CAAC,CAAC;AACvE,QAAA,IAAI,CAAC,eAAe,GAAG,cAAc,CAAC;AACpC,YAAA,GAAG,MAAM;AACT,YAAA,GAAG,iBAAiB;AACpB,YAAA,MAAM,EAAE,KAAK;AACd,SAAA,CAAC,CAAC;AACH,QAAA,MAAM,qBAAqB,GAAG,CAAC,IAAI,CAAC,eAAe,CAAC;AAEpD,QAAA,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,qBAAqB,CAC7C;AACE,YAAA,GAAG,MAAM;AACT,YAAA,GAAG,iBAAiB;AACpB,YAAA,MAAM,EAAE,IAAI;SACb,EACD;YACE,OAAO,EAAE,OAAO,CAAC,OAAO;AACzB,SAAA,CACF,CAAC;QAEF,MAAM,iBAAiB,GAAG,IAAI,CAAC,WAAW,IAAI,OAAO,CAAC,WAAW,CAAC;AAElE,QAAA,WAAW,MAAM,IAAI,IAAI,MAAM,EAAE;YAC/B,IAAI,OAAO,CAAC,MAAM,EAAE,OAAO,KAAK,IAAI,EAAE;AACpC,gBAAA,MAAM,CAAC,UAAU,CAAC,KAAK,EAAE,CAAC;AAC1B,gBAAA,MAAM,IAAI,KAAK,CAAC,uCAAuC,CAAC,CAAC;aAC1D;AAED,YAAA,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,IAAI,EAAE,CAAC;AAC7B,YAAA,IAAI,IAAI,KAAK,eAAe,EAAE;AAC5B,gBAAA,IAAI,CAAC,aAAa,GAAG,IAAkC,CAAC;aACzD;AAAM,iBAAA,IAAI,IAAI,KAAK,eAAe,EAAE;AACnC,gBAAA,IAAI,CAAC,aAAa,GAAG,IAAkC,CAAC;aACzD;AAED,YAAA,IAAI,aAA+C,CAAC;AACpD,YAAA,IAAI,IAAI,CAAC,eAAe,KAAK,IAAI,IAAI,IAAI,CAAC,aAAa,KAAK,IAAI,EAAE;AAChE,gBAAA,aAAa,GAAG,IAAI,CAAC,cAAc,EAAE,CAAC;aACvC;AAED,YAAA,MAAM,MAAM,GAAG,mCAAmC,CAAC,IAAI,EAAE;AACvD,gBAAA,WAAW,EAAE,iBAAiB;gBAC9B,qBAAqB;AACtB,aAAA,CAAC,CAAC;AACH,YAAA,IAAI,CAAC,MAAM;gBAAE,SAAS;AAEtB,YAAA,MAAM,EAAE,KAAK,EAAE,GAAG,MAAM,CAAC;AACzB,YAAA,MAAM,CAAC,KAAK,GAAG,EAAE,EAAE,SAAS,CAAC,GAAG,YAAY,CAAC,KAAK,CAAC,CAAC;AAEpD,YAAA,IAAI,CAAC,SAAS,IAAI,SAAS,KAAK,OAAO,KAAK,KAAK,KAAK,EAAE,IAAI,aAAa,CAAC,EAAE;AAC1E,gBAAA,MAAM,eAAe,GAAG,IAAI,CAAC,qBAAqB,CAAC;oBACjD,KAAK;oBACL,KAAK;oBACL,aAAa;oBACb,iBAAiB;AAClB,iBAAA,CAAC,CAAC;AACH,gBAAA,MAAM,eAAe,CAAC;gBACtB,MAAM,UAAU,EAAE,iBAAiB,CACjC,KAAK,EACL,SAAS,EACT,SAAS,EACT,SAAS,EACT,SAAS,EACT,EAAE,KAAK,EAAE,eAAe,EAAE,CAC3B,CAAC;gBACF,SAAS;aACV;AAED,YAAA,MAAM,UAAU,GAAG,IAAI,UAAU,CAAC,KAAK,EAAE;gBACvC,KAAK,EAAE,IAAI,CAAC,gBAAgB;AAC5B,gBAAA,cAAc,EAAE,IAAI;AACpB,gBAAA,YAAY,EAAE,CAAC;AACf,gBAAA,YAAY,EAAE,CAAC;AAChB,aAAA,CAAC,CAAC;AAEH,YAAA,MAAM,SAAS,GAAG,UAAU,CAAC,YAAY,EAAE,CAAC;AAC5C,YAAA,IAAI;gBACF,IAAI,YAAY,GAAG,KAAK,CAAC;AACzB,gBAAA,WAAW,MAAM,YAAY,IAAI,SAAS,EAAE;oBAC1C,MAAM,QAAQ,GAAG,UAAU,CAAC,YAAY,EAAE,SAAS,EAAE,KAAK,CAAC,CAAC;AAE5D,oBAAA,MAAM,eAAe,GAAG,IAAI,CAAC,qBAAqB,CAAC;AACjD,wBAAA,KAAK,EAAE,YAAY;AACnB,wBAAA,KAAK,EAAE,QAAQ;wBACf,aAAa,EAAE,YAAY,GAAG,SAAS,GAAG,aAAa;wBACvD,iBAAiB;AAClB,qBAAA,CAAC,CAAC;AAEH,oBAAA,IAAI,aAAa,IAAI,CAAC,YAAY,EAAE;wBAClC,YAAY,GAAG,IAAI,CAAC;qBACrB;AACD,oBAAA,MAAM,eAAe,CAAC;oBAEtB,MAAM,UAAU,EAAE,iBAAiB,CACjC,KAAK,EACL,SAAS,EACT,SAAS,EACT,SAAS,EACT,SAAS,EACT,EAAE,KAAK,EAAE,eAAe,EAAE,CAC3B,CAAC;iBACH;aACF;oBAAS;AACR,gBAAA,MAAM,SAAS,CAAC,MAAM,EAAE,CAAC;aAC1B;SACF;QAED,IAAI,CAAC,gBAAgB,EAAE,CAAC;KACzB;AACF;;;;"}

package/dist/types/llm/anthropic/llm.d.ts

@@ -3,12 +3,23 @@ import { ChatGenerationChunk } from '@langchain/core/outputs';
 import type { BaseMessage } from '@langchain/core/messages';
 import type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';
 import type { AnthropicInput } from '@langchain/anthropic';
+import type { AnthropicStreamUsage } from '@/llm/anthropic/types';
 export type CustomAnthropicInput = AnthropicInput & {
     _lc_stream_delay?: number;
 };
 export declare class CustomAnthropic extends ChatAnthropicMessages {
     _lc_stream_delay: number;
+    private message_start;
+    private message_delta;
+    private tools_in_params?;
+    private emitted_usage?;
     constructor(fields: CustomAnthropicInput);
+    /**
+     * Get stream usage as returned by this client's API response.
+     * @returns {AnthropicStreamUsage} The stream usage object.
+     */
+    getStreamUsage(): AnthropicStreamUsage | undefined;
+    resetTokenEvents(): void;
     private createGenerationChunk;
     _streamResponseChunks(messages: BaseMessage[], options: this['ParsedCallOptions'], runManager?: CallbackManagerForLLMRun): AsyncGenerator<ChatGenerationChunk>;
 }

package/dist/types/llm/anthropic/types.d.ts

@@ -8,6 +8,8 @@ export type AnthropicToolResponse = {
     input: Record<string, any>;
 };
 export type AnthropicMessageParam = Anthropic.MessageParam;
+export type AnthropicMessageDeltaEvent = Anthropic.MessageDeltaEvent;
+export type AnthropicMessageStartEvent = Anthropic.MessageStartEvent;
 export type AnthropicMessageResponse = Anthropic.ContentBlock | AnthropicToolResponse;
 export type AnthropicMessageCreateParams = Anthropic.MessageCreateParamsNonStreaming;
 export type AnthropicStreamingMessageCreateParams = Anthropic.MessageCreateParamsStreaming;
@@ -22,3 +24,29 @@ export type AnthropicTextBlockParam = Anthropic.Messages.TextBlockParam;
 export type AnthropicImageBlockParam = Anthropic.Messages.ImageBlockParam;
 export type AnthropicToolUseBlockParam = Anthropic.Messages.ToolUseBlockParam;
 export type AnthropicToolResultBlockParam = Anthropic.Messages.ToolResultBlockParam;
+/**
+ * Stream usage information for Anthropic API calls
+ * @see https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching#pricing
+ */
+export interface AnthropicStreamUsage {
+    /**
+     * The number of input tokens used in the request
+     */
+    input_tokens: number;
+    /**
+     * The number of cache creation input tokens used (write operations)
+     */
+    cache_creation_input_tokens?: number;
+    /**
+     * The number of cache input tokens used (read operations)
+     */
+    cache_read_input_tokens?: number;
+    /**
+     * The number of output tokens generated in the response
+     */
+    output_tokens: number;
+    /**
+     * The total number of tokens generated in the response
+     */
+    total_tokens: number;
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@librechat/agents",
-  "version": "1.9.5",
+  "version": "1.9.6",
   "main": "./dist/cjs/main.cjs",
   "module": "./dist/esm/main.mjs",
   "types": "./dist/types/index.d.ts",

package/src/llm/anthropic/llm.ts

@@ -4,7 +4,7 @@ import { ChatGenerationChunk } from '@langchain/core/outputs';
 import type { BaseMessage, MessageContentComplex } from '@langchain/core/messages';
 import type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';
 import type { AnthropicInput } from '@langchain/anthropic';
-import type { AnthropicMessageCreateParams } from '@/llm/anthropic/types';
+import type { AnthropicMessageCreateParams, AnthropicStreamUsage, AnthropicMessageStartEvent, AnthropicMessageDeltaEvent } from '@/llm/anthropic/types';
 import { _makeMessageChunkFromAnthropicEvent } from './utils/message_outputs';
 import { _convertMessagesToAnthropicPayload } from './utils/message_inputs';
 import { TextStream } from '@/llm/text';
@@ -54,28 +54,70 @@ export type CustomAnthropicInput = AnthropicInput & { _lc_stream_delay?: number
 
 export class CustomAnthropic extends ChatAnthropicMessages {
   _lc_stream_delay: number;
+  private message_start: AnthropicMessageStartEvent | undefined;
+  private message_delta: AnthropicMessageDeltaEvent | undefined;
+  private tools_in_params?: boolean;
+  private emitted_usage?: boolean;
   constructor(fields: CustomAnthropicInput) {
     super(fields);
     this._lc_stream_delay = fields._lc_stream_delay ?? 25;
   }
 
+  /**
+   * Get stream usage as returned by this client's API response.
+   * @returns {AnthropicStreamUsage} The stream usage object.
+   */
+  getStreamUsage(): AnthropicStreamUsage | undefined {
+    if (this.emitted_usage === true) {
+      return;
+    }
+    const inputUsage = (this.message_start?.message)?.usage as undefined | AnthropicStreamUsage;
+    const outputUsage = this.message_delta?.usage as undefined | Partial<AnthropicStreamUsage>;
+    if (!outputUsage) {
+      return;
+    }
+    const totalUsage = {
+      total_tokens: (inputUsage?.input_tokens ?? 0)
+      + (inputUsage?.output_tokens ?? 0)
+      + (inputUsage?.cache_creation_input_tokens ?? 0)
+      + (inputUsage?.cache_read_input_tokens ?? 0)
+      + (outputUsage.input_tokens ?? 0)
+      + (outputUsage.output_tokens ?? 0)
+      + (outputUsage.cache_creation_input_tokens ?? 0)
+      + (outputUsage.cache_read_input_tokens ?? 0),
+    };
+
+    this.emitted_usage = true;
+    return Object.assign(totalUsage, inputUsage, outputUsage);
+  }
+
+  resetTokenEvents(): void {
+    this.message_start = undefined;
+    this.message_delta = undefined;
+    this.emitted_usage = undefined;
+    this.tools_in_params = undefined;
+  }
+
   private createGenerationChunk({
     token,
     chunk,
+    usageMetadata,
     shouldStreamUsage,
   }: {
     token?: string,
     chunk: AIMessageChunk,
     shouldStreamUsage: boolean
+    usageMetadata?: AnthropicStreamUsage,
   }): ChatGenerationChunk {
+    const usage_metadata = shouldStreamUsage ? usageMetadata ?? chunk.usage_metadata : undefined;
     return new ChatGenerationChunk({
       message: new AIMessageChunk({
         // Just yield chunk as it is and tool_use will be concat by BaseChatModel._generateUncached().
         content: chunk.content,
         additional_kwargs: chunk.additional_kwargs,
         tool_call_chunks: chunk.tool_call_chunks,
-        usage_metadata: shouldStreamUsage ? chunk.usage_metadata : undefined,
         response_metadata: chunk.response_metadata,
+        usage_metadata,
         id: chunk.id,
       }),
       text: token ?? '',
@@ -89,11 +131,12 @@ export class CustomAnthropic extends ChatAnthropicMessages {
   ): AsyncGenerator<ChatGenerationChunk> {
     const params = this.invocationParams(options);
     const formattedMessages = _convertMessagesToAnthropicPayload(messages);
-    const coerceContentToString = !_toolsInParams({
+    this.tools_in_params = _toolsInParams({
       ...params,
       ...formattedMessages,
       stream: false,
     });
+    const coerceContentToString = !this.tools_in_params;
 
     const stream = await this.createStreamWithRetry(
       {
@@ -106,13 +149,26 @@ export class CustomAnthropic extends ChatAnthropicMessages {
       }
     );
 
+    const shouldStreamUsage = this.streamUsage ?? options.streamUsage;
+
     for await (const data of stream) {
       if (options.signal?.aborted === true) {
         stream.controller.abort();
         throw new Error('AbortError: User aborted the request.');
       }
 
-      const shouldStreamUsage = this.streamUsage ?? options.streamUsage;
+      const type = data.type ?? '';
+      if (type === 'message_start') {
+        this.message_start = data as AnthropicMessageStartEvent;
+      } else if (type === 'message_delta') {
+        this.message_delta = data as AnthropicMessageDeltaEvent;
+      }
+
+      let usageMetadata: AnthropicStreamUsage | undefined;
+      if (this.tools_in_params !== true && this.emitted_usage !== true) {
+        usageMetadata = this.getStreamUsage();
+      }
+
       const result = _makeMessageChunkFromAnthropicEvent(data, {
         streamUsage: shouldStreamUsage,
         coerceContentToString,
@@ -122,10 +178,11 @@ export class CustomAnthropic extends ChatAnthropicMessages {
       const { chunk } = result;
       const [token = '', tokenType] = extractToken(chunk);
 
-      if (!tokenType || tokenType === 'input') {
+      if (!tokenType || tokenType === 'input' || (token === '' && usageMetadata)) {
         const generationChunk = this.createGenerationChunk({
           token,
           chunk,
+          usageMetadata,
           shouldStreamUsage,
         });
         yield generationChunk;
@@ -149,13 +206,20 @@ export class CustomAnthropic extends ChatAnthropicMessages {
 
       const generator = textStream.generateText();
       try {
+        let emittedUsage = false;
         for await (const currentToken of generator) {
           const newChunk = cloneChunk(currentToken, tokenType, chunk);
+
           const generationChunk = this.createGenerationChunk({
             token: currentToken,
             chunk: newChunk,
+            usageMetadata: emittedUsage ? undefined : usageMetadata,
             shouldStreamUsage,
           });
+
+          if (usageMetadata && !emittedUsage) {
+            emittedUsage = true;
+          }
           yield generationChunk;
 
           await runManager?.handleLLMNewToken(
@@ -171,5 +235,7 @@ export class CustomAnthropic extends ChatAnthropicMessages {
         await generator.return();
       }
     }
+
+    this.resetTokenEvents();
   }
 }

package/src/llm/anthropic/types.ts

@@ -12,6 +12,8 @@ export type AnthropicToolResponse = {
 };
 
 export type AnthropicMessageParam = Anthropic.MessageParam;
+export type AnthropicMessageDeltaEvent= Anthropic.MessageDeltaEvent;
+export type AnthropicMessageStartEvent= Anthropic.MessageStartEvent;
 export type AnthropicMessageResponse =
   | Anthropic.ContentBlock
   | AnthropicToolResponse;
@@ -35,4 +37,34 @@ export type AnthropicTextBlockParam = Anthropic.Messages.TextBlockParam;
 export type AnthropicImageBlockParam = Anthropic.Messages.ImageBlockParam;
 export type AnthropicToolUseBlockParam = Anthropic.Messages.ToolUseBlockParam;
 export type AnthropicToolResultBlockParam =
-  Anthropic.Messages.ToolResultBlockParam;
+  Anthropic.Messages.ToolResultBlockParam;
+
+/**
+ * Stream usage information for Anthropic API calls
+ * @see https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching#pricing
+ */
+export interface AnthropicStreamUsage {
+  /**
+   * The number of input tokens used in the request
+   */
+  input_tokens: number;
+
+  /**
+   * The number of cache creation input tokens used (write operations)
+   */
+  cache_creation_input_tokens?: number;
+
+  /**
+   * The number of cache input tokens used (read operations)
+   */
+  cache_read_input_tokens?: number;
+
+  /**
+   * The number of output tokens generated in the response
+   */
+  output_tokens: number;
+  /**
+   * The total number of tokens generated in the response
+   */
+  total_tokens: number;
+}