Add Anthropic thinking params and responses support

.github/workflows/js.yaml

@@ -25,3 +25,7 @@ jobs:
           pnpm install
           pnpm run test
           pnpm run build
+        env:
+          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+          GEMINI_API_KEY: ${{ secrets.GEMINI_API_KEY }}

.pre-commit-config.yaml

@@ -13,7 +13,7 @@ repos:
               .*\.(json|prisma|svg)|
               .*pnpm-lock.yaml
           )$
-        args: ["-L rouge,coo,couldn,unsecure"]
+        args: ["-L rouge,coo,couldn,unsecure,afterall"]
   - repo: https://github.com/rbubley/mirrors-prettier
     rev: v3.3.2
     hooks:

packages/proxy/edge/deps.test.ts

@@ -0,0 +1,20 @@
+import skott from "skott";
+import { describe, expect, it } from "vitest";
+
+describe("proxy/edge", () => {
+  it("no circ dependencies", async () => {
+    const { useGraph } = await skott({
+      entrypoint: `${__dirname}/index.ts`,
+      tsConfigPath: `${__dirname}/../tsconfig.json`,
+      dependencyTracking: {
+        builtin: false,
+        thirdParty: true,
+        typeOnly: true,
+      },
+    });
+
+    const { findCircularDependencies } = useGraph();
+
+    expect(findCircularDependencies()).toEqual([]);
+  });
+});

packages/proxy/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@braintrust/proxy",
-  "version": "0.0.7",
+  "version": "0.0.8",
   "description": "A proxy server that load balances across AI providers.",
   "main": "./dist/index.js",
   "module": "./dist/index.mjs",
@@ -36,12 +36,19 @@
       "import": "./utils/dist/index.mjs",
       "module": "./utils/dist/index.mjs",
       "require": "./utils/dist/index.js"
+    },
+    "./types": {
+      "types": "./types/dist/index.d.ts",
+      "import": "./types/dist/index.mjs",
+      "module": "./types/dist/index.mjs",
+      "require": "./types/dist/index.js"
     }
   },
   "files": [
     "dist/**/*",
     "edge/dist/**/*",
-    "schema/dist/**/*"
+    "schema/dist/**/*",
+    "types/dist/**/*"
   ],
   "license": "MIT",
   "publishConfig": {
@@ -74,7 +81,9 @@
     "@types/yargs": "^17.0.33",
     "@typescript-eslint/eslint-plugin": "^8.21.0",
     "esbuild": "^0.19.10",
+    "msw": "^2.8.2",
     "npm-run-all": "^4.1.5",
+    "skott": "^0.35.4",
     "tsup": "^8.4.0",
     "typescript": "5.5.4",
     "vite-tsconfig-paths": "^4.3.2",
@@ -84,10 +93,10 @@
   "dependencies": {
     "@anthropic-ai/sdk": "^0.39.0",
     "@apidevtools/json-schema-ref-parser": "^11.9.1",
-    "@aws-sdk/client-bedrock-runtime": "^3.738.0",
-    "@braintrust/core": "^0.0.85",
+    "@aws-sdk/client-bedrock-runtime": "^3.806.0",
+    "@braintrust/core": "^0.0.87",
     "@breezystack/lamejs": "^1.2.7",
-    "@google/generative-ai": "^0.24.0",
+    "@google/genai": "^0.13.0",
     "@opentelemetry/api": "^1.7.0",
     "@opentelemetry/core": "^1.19.0",
     "@opentelemetry/resources": "^1.19.0",

packages/proxy/schema/deps.test.ts

@@ -0,0 +1,20 @@
+import skott from "skott";
+import { describe, expect, it } from "vitest";
+
+describe("proxy/schema", () => {
+  it("no circ dependencies", async () => {
+    const { useGraph } = await skott({
+      entrypoint: `${__dirname}/index.ts`,
+      tsConfigPath: `${__dirname}/../tsconfig.json`,
+      dependencyTracking: {
+        builtin: false,
+        thirdParty: true,
+        typeOnly: true,
+      },
+    });
+
+    const { findCircularDependencies } = useGraph();
+
+    expect(findCircularDependencies()).toEqual([]);
+  });
+});

packages/proxy/schema/index.test.ts

@@ -0,0 +1,270 @@
+import { MessageCreateParamsBase } from "@anthropic-ai/sdk/resources/messages";
+import { GenerateContentParameters } from "@google/genai";
+import { ChatCompletionCreateParams } from "openai/resources";
+import { expect, it } from "vitest";
+import { ModelFormat, translateParams } from "./index";
+
+const examples: Record<
+  string,
+  {
+    openai: ChatCompletionCreateParams;
+  } & ( // NOTE: these are not strictly the API params.
+    | { google: GenerateContentParameters }
+    | { anthropic: MessageCreateParamsBase }
+  )
+> = {
+  simple: {
+    openai: {
+      model: "gpt-4o",
+      max_tokens: 1500,
+      temperature: 0.7,
+      top_p: 0.9,
+      frequency_penalty: 0.1,
+      presence_penalty: 0.2,
+      messages: [
+        { role: "system", content: "You are a helpful assistant." },
+        { role: "user", content: "Hello, how are you?" },
+      ],
+      stream: true,
+    },
+    google: {
+      maxOutputTokens: 1500,
+      max_tokens: 1500,
+      messages: [
+        {
+          content: "You are a helpful assistant.",
+          role: "system",
+        },
+        {
+          content: "Hello, how are you?",
+          role: "user",
+        },
+      ],
+      model: "gpt-4o",
+      stream: true,
+      temperature: 0.7,
+      top_p: 0.9,
+    },
+    anthropic: {
+      max_tokens: 1500,
+      messages: [
+        {
+          content: "You are a helpful assistant.",
+          // @ts-expect-error -- TODO: shouldn't we have translated this to a non system role?
+          role: "system",
+        },
+        {
+          content: "Hello, how are you?",
+          role: "user",
+        },
+      ],
+      model: "gpt-4o",
+      stream: true,
+      temperature: 0.7,
+      top_p: 0.9,
+    },
+  },
+  reasoning_effort: {
+    openai: {
+      model: "gpt-4o",
+      messages: [
+        {
+          role: "system",
+          content: "You are a detailed reasoning assistant.",
+        },
+        {
+          role: "user",
+          content: "Explain how to solve 2x + 4 = 12 step by step.",
+        },
+      ],
+      temperature: 0,
+      max_tokens: 1000,
+      reasoning_effort: "high",
+      stream: false,
+    },
+    google: {
+      model: "gpt-4o",
+      // notice how this is still an intermediate param
+      // google's api expects a content instead of messages, for example
+      messages: [
+        {
+          role: "system",
+          content: "You are a detailed reasoning assistant.",
+        },
+        {
+          role: "user",
+          content: "Explain how to solve 2x + 4 = 12 step by step.",
+        },
+      ],
+      temperature: 0,
+      thinkingConfig: {
+        thinkingBudget: 800,
+        includeThoughts: true,
+      },
+      maxOutputTokens: 1000,
+      max_tokens: 1000,
+      stream: false,
+    },
+    anthropic: {
+      model: "gpt-4o",
+      messages: [
+        {
+          // @ts-expect-error  -- we use the role to later manipulate the request
+          role: "system",
+          content: "You are a detailed reasoning assistant.",
+        },
+        {
+          role: "user",
+          content: "Explain how to solve 2x + 4 = 12 step by step.",
+        },
+      ],
+      temperature: 1,
+      stream: false,
+      max_tokens: 1536,
+      thinking: {
+        budget_tokens: 1024,
+        type: "enabled",
+      },
+    },
+  },
+  "reasoning disable": {
+    openai: {
+      model: "gpt-4o",
+      messages: [
+        {
+          role: "system",
+          content: "You are a detailed reasoning assistant.",
+        },
+        {
+          role: "user",
+          content: "Explain how to solve 2x + 4 = 12 step by step.",
+        },
+      ],
+      temperature: 0,
+      reasoning_enabled: false,
+      reasoning_budget: 1024,
+      stream: false,
+    },
+    google: {
+      model: "gpt-4o",
+      // notice how this is still an intermediate param
+      // google's api expects a content instead of messages, for example
+      messages: [
+        {
+          role: "system",
+          content: "You are a detailed reasoning assistant.",
+        },
+        {
+          role: "user",
+          content: "Explain how to solve 2x + 4 = 12 step by step.",
+        },
+      ],
+      temperature: 0,
+      thinkingConfig: {
+        thinkingBudget: 0,
+      },
+      stream: false,
+    },
+    anthropic: {
+      model: "gpt-4o",
+      messages: [
+        {
+          // @ts-expect-error  -- we use the role to later manipulate the request
+          role: "system",
+          content: "You are a detailed reasoning assistant.",
+        },
+        {
+          role: "user",
+          content: "Explain how to solve 2x + 4 = 12 step by step.",
+        },
+      ],
+      temperature: 0,
+      stream: false,
+      max_tokens: 1024,
+      thinking: {
+        type: "disabled",
+      },
+    },
+  },
+  "reasoning budget": {
+    openai: {
+      model: "gpt-4o",
+      messages: [
+        {
+          role: "system",
+          content: "You are a detailed reasoning assistant.",
+        },
+        {
+          role: "user",
+          content: "Explain how to solve 2x + 4 = 12 step by step.",
+        },
+      ],
+      temperature: 0,
+      reasoning_enabled: true,
+      reasoning_budget: 1024,
+      stream: false,
+    },
+    google: {
+      model: "gpt-4o",
+      // notice how this is still an intermediate param
+      // google's api expects a content instead of messages, for example
+      messages: [
+        {
+          role: "system",
+          content: "You are a detailed reasoning assistant.",
+        },
+        {
+          role: "user",
+          content: "Explain how to solve 2x + 4 = 12 step by step.",
+        },
+      ],
+      temperature: 0,
+      thinkingConfig: {
+        thinkingBudget: 1024,
+        includeThoughts: true,
+      },
+      stream: false,
+    },
+    anthropic: {
+      model: "gpt-4o",
+      messages: [
+        {
+          // @ts-expect-error  -- we use the role to later manipulate the request
+          role: "system",
+          content: "You are a detailed reasoning assistant.",
+        },
+        {
+          role: "user",
+          content: "Explain how to solve 2x + 4 = 12 step by step.",
+        },
+      ],
+      temperature: 1,
+      stream: false,
+      max_tokens: 1536,
+      thinking: {
+        budget_tokens: 1024,
+        type: "enabled",
+      },
+    },
+  },
+};
+
+Object.entries(examples).forEach(([example, { openai, ...providers }]) => {
+  Object.entries(providers).forEach(([provider, expected]) => {
+    it(`[${example}] translate openai to ${provider} params`, () => {
+      const result = translateParams(
+        provider as ModelFormat,
+        openai as unknown as Record<string, unknown>,
+      );
+      try {
+        expect(result).toEqual(expected);
+      } catch (error) {
+        console.warn(
+          `Exact openai -> ${provider} translation failed. Found:`,
+          JSON.stringify(result, null, 2),
+        );
+        expect.soft(result).toEqual(expected);
+      }
+    });
+  });
+});