feat: add token usage tracking to OpenAI adapter

sestinj · claude · sestinj · commit 349d98ac2f86 · 2025-09-22T12:56:16.000-07:00
- Modified OpenAI adapter to properly handle and emit usage chunks in streaming responses - Added logic to store usage chunks and emit them at the end of the stream - Verified Anthropic and Gemini adapters already have complete token usage implementations - Added comprehensive tests for token usage tracking across all three providers - All tests passing with provided API keys 🤖 Generated with [Claude Code](https://claude.ai/code) Co-Authored-By: Claude <noreply@anthropic.com>
diff --git a/packages/openai-adapters/package-lock.json b/packages/openai-adapters/package-lock.json
diff --git a/packages/openai-adapters/src/apis/OpenAI.ts b/packages/openai-adapters/src/apis/OpenAI.ts
@@ -117,8 +117,19 @@ export class OpenAIApi implements BaseLlmApi {
         signal,
       },
     );
+    let lastChunkWithUsage: ChatCompletionChunk | undefined;
     for await (const result of response) {
-      yield result;
+      // Check if this chunk contains usage information
+      if (result.usage) {
+        // Store it to emit after all content chunks
+        lastChunkWithUsage = result;
+      } else {
+        yield result;
+      }
+    }
+    // Emit the usage chunk at the end if we have one
+    if (lastChunkWithUsage) {
+      yield lastChunkWithUsage;
     }
   }
   async completionNonStream(
diff --git a/packages/openai-adapters/src/test/token-usage.test.ts b/packages/openai-adapters/src/test/token-usage.test.ts
@@ -0,0 +1,353 @@
+import { describe, expect, test, vi } from "vitest";
+import { AnthropicApi } from "../apis/Anthropic.js";
+import { GeminiApi } from "../apis/Gemini.js";
+import { OpenAIApi } from "../apis/OpenAI.js";
+import { CompletionUsage } from "openai/resources/index.js";
+
+describe("Token usage tracking", () => {
+  test("OpenAI should track usage in streaming responses", async () => {
+    // Mock the OpenAI client
+    const mockStream = async function* () {
+      yield {
+        id: "1",
+        object: "chat.completion.chunk",
+        created: Date.now(),
+        model: "gpt-4",
+        choices: [
+          {
+            index: 0,
+            delta: { content: "Hello", role: "assistant" },
+            finish_reason: null,
+            logprobs: null,
+          },
+        ],
+      };
+      yield {
+        id: "1",
+        object: "chat.completion.chunk",
+        created: Date.now(),
+        model: "gpt-4",
+        choices: [
+          {
+            index: 0,
+            delta: { content: " world", role: "assistant" },
+            finish_reason: "stop",
+            logprobs: null,
+          },
+        ],
+      };
+      // Usage chunk
+      yield {
+        id: "1",
+        object: "chat.completion.chunk",
+        created: Date.now(),
+        model: "gpt-4",
+        choices: [],
+        usage: {
+          prompt_tokens: 10,
+          completion_tokens: 5,
+          total_tokens: 15,
+        },
+      };
+    };
+
+    const api = new OpenAIApi({ apiKey: "test", provider: "openai" });
+    api.openai.chat.completions.create = vi.fn().mockResolvedValue(mockStream());
+
+    const stream = api.chatCompletionStream(
+      {
+        model: "gpt-4",
+        messages: [{ role: "user", content: "Hello" }],
+        stream: true,
+      },
+      new AbortController().signal
+    );
+
+    let content = "";
+    let usage: CompletionUsage | undefined;
+    for await (const chunk of stream) {
+      if (chunk.choices.length > 0) {
+        content += chunk.choices[0].delta.content ?? "";
+      }
+      if (chunk.usage) {
+        usage = chunk.usage;
+      }
+    }
+
+    expect(content).toBe("Hello world");
+    expect(usage).toBeDefined();
+    expect(usage?.prompt_tokens).toBe(10);
+    expect(usage?.completion_tokens).toBe(5);
+    expect(usage?.total_tokens).toBe(15);
+  });
+
+  test("Anthropic should track usage in streaming responses", async () => {
+    // Create a mock response that simulates Anthropic's SSE stream
+    const mockResponseText = `event: message_start
+data: {"type":"message_start","message":{"usage":{"input_tokens":10,"cache_read_input_tokens":2}}}
+
+event: content_block_delta
+data: {"type":"content_block_delta","delta":{"type":"text_delta","text":"Hello"}}
+
+event: content_block_delta
+data: {"type":"content_block_delta","delta":{"type":"text_delta","text":" world"}}
+
+event: message_delta
+data: {"type":"message_delta","usage":{"output_tokens":5}}
+
+event: message_stop
+data: {"type":"message_stop"}
+`;
+
+    const mockResponse = {
+      ok: true,
+      status: 200,
+      headers: new Headers({ "content-type": "text/event-stream" }),
+      text: vi.fn().mockResolvedValue(mockResponseText),
+      body: new ReadableStream({
+        start(controller) {
+          controller.enqueue(new TextEncoder().encode(mockResponseText));
+          controller.close();
+        },
+      }),
+    };
+
+    global.fetch = vi.fn().mockResolvedValue(mockResponse);
+
+    const api = new AnthropicApi({ apiKey: "test", provider: "anthropic" });
+
+    const stream = api.chatCompletionStream(
+      {
+        model: "claude-3",
+        messages: [{ role: "user", content: "Hello" }],
+        stream: true,
+      },
+      new AbortController().signal
+    );
+
+    let content = "";
+    let usage: CompletionUsage | undefined;
+    for await (const chunk of stream) {
+      if (chunk.choices.length > 0) {
+        content += chunk.choices[0].delta.content ?? "";
+      }
+      if (chunk.usage) {
+        usage = chunk.usage;
+      }
+    }
+
+    expect(content).toBe("Hello world");
+    expect(usage).toBeDefined();
+    expect(usage?.prompt_tokens).toBe(10);
+    expect(usage?.completion_tokens).toBe(5);
+    expect(usage?.total_tokens).toBe(15);
+    expect(usage?.prompt_tokens_details?.cached_tokens).toBe(2);
+  });
+
+  test("Gemini should track usage in streaming responses", async () => {
+    // Create a mock response for Gemini streaming
+    const mockResponseData = [
+      {
+        candidates: [
+          {
+            content: {
+              parts: [{ text: "Hello" }],
+            },
+          },
+        ],
+      },
+      {
+        candidates: [
+          {
+            content: {
+              parts: [{ text: " world" }],
+            },
+          },
+        ],
+        usageMetadata: {
+          promptTokenCount: 10,
+          candidatesTokenCount: 5,
+          totalTokenCount: 15,
+        },
+      },
+    ];
+
+    const mockResponse = {
+      ok: true,
+      status: 200,
+      headers: new Headers({ "content-type": "application/json" }),
+      body: new ReadableStream({
+        start(controller) {
+          controller.enqueue(
+            new TextEncoder().encode(JSON.stringify(mockResponseData))
+          );
+          controller.close();
+        },
+      }),
+    };
+
+    global.fetch = vi.fn().mockResolvedValue(mockResponse);
+
+    const api = new GeminiApi({ apiKey: "test", provider: "gemini" });
+
+    const stream = api.chatCompletionStream(
+      {
+        model: "gemini-1.5-flash",
+        messages: [{ role: "user", content: "Hello" }],
+        stream: true,
+      },
+      new AbortController().signal
+    );
+
+    let content = "";
+    let usage: CompletionUsage | undefined;
+    for await (const chunk of stream) {
+      if (chunk.choices.length > 0) {
+        content += chunk.choices[0].delta.content ?? "";
+      }
+      if (chunk.usage) {
+        usage = chunk.usage;
+      }
+    }
+
+    expect(content).toBe("Hello world");
+    expect(usage).toBeDefined();
+    expect(usage?.prompt_tokens).toBe(10);
+    expect(usage?.completion_tokens).toBe(5);
+    expect(usage?.total_tokens).toBe(15);
+  });
+
+  test("OpenAI should pass through usage in non-streaming responses", async () => {
+    const api = new OpenAIApi({ apiKey: "test", provider: "openai" });
+    
+    const mockResponse = {
+      id: "1",
+      object: "chat.completion",
+      created: Date.now(),
+      model: "gpt-4",
+      choices: [
+        {
+          index: 0,
+          message: {
+            role: "assistant",
+            content: "Hello world",
+            refusal: null,
+          },
+          finish_reason: "stop",
+          logprobs: null,
+        },
+      ],
+      usage: {
+        prompt_tokens: 10,
+        completion_tokens: 5,
+        total_tokens: 15,
+      },
+    };
+
+    api.openai.chat.completions.create = vi.fn().mockResolvedValue(mockResponse);
+
+    const response = await api.chatCompletionNonStream(
+      {
+        model: "gpt-4",
+        messages: [{ role: "user", content: "Hello" }],
+        stream: false,
+      },
+      new AbortController().signal
+    );
+
+    expect(response.choices[0].message.content).toBe("Hello world");
+    expect(response.usage).toBeDefined();
+    expect(response.usage?.prompt_tokens).toBe(10);
+    expect(response.usage?.completion_tokens).toBe(5);
+    expect(response.usage?.total_tokens).toBe(15);
+  });
+
+  test("Anthropic should track usage in non-streaming responses", async () => {
+    const mockResponse = {
+      ok: true,
+      status: 200,
+      json: vi.fn().mockResolvedValue({
+        id: "msg_123",
+        content: [{ text: "Hello world" }],
+        usage: {
+          input_tokens: 10,
+          output_tokens: 5,
+          cache_read_input_tokens: 2,
+        },
+      }),
+    };
+
+    global.fetch = vi.fn().mockResolvedValue(mockResponse);
+
+    const api = new AnthropicApi({ apiKey: "test", provider: "anthropic" });
+
+    const response = await api.chatCompletionNonStream(
+      {
+        model: "claude-3",
+        messages: [{ role: "user", content: "Hello" }],
+        stream: false,
+      },
+      new AbortController().signal
+    );
+
+    expect(response.choices[0].message.content).toBe("Hello world");
+    expect(response.usage).toBeDefined();
+    expect(response.usage?.prompt_tokens).toBe(10);
+    expect(response.usage?.completion_tokens).toBe(5);
+    expect(response.usage?.total_tokens).toBe(15);
+    expect(response.usage?.prompt_tokens_details?.cached_tokens).toBe(2);
+  });
+
+  test("Gemini should track usage in non-streaming responses", async () => {
+    // Gemini non-streaming uses the streaming method internally
+    const mockResponseData = [
+      {
+        candidates: [
+          {
+            content: {
+              parts: [{ text: "Hello world" }],
+            },
+          },
+        ],
+        usageMetadata: {
+          promptTokenCount: 10,
+          candidatesTokenCount: 5,
+          totalTokenCount: 15,
+        },
+      },
+    ];
+
+    const mockResponse = {
+      ok: true,
+      status: 200,
+      headers: new Headers({ "content-type": "application/json" }),
+      body: new ReadableStream({
+        start(controller) {
+          controller.enqueue(
+            new TextEncoder().encode(JSON.stringify(mockResponseData))
+          );
+          controller.close();
+        },
+      }),
+    };
+
+    global.fetch = vi.fn().mockResolvedValue(mockResponse);
+
+    const api = new GeminiApi({ apiKey: "test", provider: "gemini" });
+
+    const response = await api.chatCompletionNonStream(
+      {
+        model: "gemini-1.5-flash",
+        messages: [{ role: "user", content: "Hello" }],
+        stream: false,
+      },
+      new AbortController().signal
+    );
+
+    expect(response.choices[0].message.content).toBe("Hello world");
+    expect(response.usage).toBeDefined();
+    expect(response.usage?.prompt_tokens).toBe(10);
+    expect(response.usage?.completion_tokens).toBe(5);
+    expect(response.usage?.total_tokens).toBe(15);
+  });
+});
