Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.
Python agentaianthropicclaude-codecompressioncontext-engineeringcontext-windowcursorfastapilangchainllmmcpopenaiprompt-engineeringproxypythonragtoken-optimizationtokenstypescript
Fix Vercel AI SDK v6 format: use input/output instead of args/result
chopratejas284f6d0a099a
diff --git a/sdk/typescript/src/utils/format.ts b/sdk/typescript/src/utils/format.tsindex ef84d9a56..17e3e6a2e 100644--- a/sdk/typescript/src/utils/format.ts+++ b/sdk/typescript/src/utils/format.ts@@ -253,7 +253,8 @@ export function vercelToOpenAI(messages: any[]): OpenAIMessage[] {openaiMsg.tool_calls = toolCallParts.map((p: any): ToolCall => ({id: p.toolCallId,type: "function",- function: { name: p.toolName, arguments: JSON.stringify(p.args) },+ // AI SDK v6 uses `input`, earlier versions used `args`+ function: { name: p.toolName, arguments: JSON.stringify(p.input ?? p.args) },}));}result.push(openaiMsg);@@ -318,13 +319,13 @@ export function openAIToVercel(messages: OpenAIMessage[]): any[] {if (msg.content) parts.push({ type: "text", text: msg.content });if (msg.tool_calls) {for (const tc of msg.tool_calls) {- let args: any;- try { args = JSON.parse(tc.function.arguments); } catch { args = tc.function.arguments ?? {}; }+ let input: any;+ try { input = JSON.parse(tc.function.arguments); } catch { input = tc.function.arguments ?? {}; }parts.push({type: "tool-call",toolCallId: tc.id,toolName: tc.function.name,- args,+ input, // AI SDK v6 uses `input`, not `args`});}}diff --git a/sdk/typescript/test/mdx-examples.test.ts b/sdk/typescript/test/mdx-examples.test.tsnew file mode 100644index 000000000..74b9ec6c6--- /dev/null+++ b/sdk/typescript/test/mdx-examples.test.ts@@ -0,0 +1,253 @@+/**+ * Tests every code example from the Vercel AI SDK PR MDX files.+ *+ * If these pass, the examples in the docs are correct.+ *+ * Run: HEADROOM_INTEGRATION=1 npx vitest run test/mdx-examples.test.ts+ */+import { describe, it, expect, beforeAll } from "vitest";+import { config } from "dotenv";+import { resolve } from "path";++config({ path: resolve(__dirname, "../../../.env") });++const PROXY_URL = "http://localhost:8787";+const RUN = process.env.HEADROOM_INTEGRATION === "1";++describe.skipIf(!RUN)("MDX Examples: 51-headroom.mdx", () => {+ beforeAll(async () => {+ const res = await fetch(`${PROXY_URL}/health`);+ if (!res.ok) throw new Error("Proxy not running");+ });++ // =====================================================+ // Example 1 from 51-headroom.mdx: "Compress messages before calling the model"+ // =====================================================+ it("compress() with AI SDK format messages → generateText()", { timeout: 30000 }, async () => {+ const { compress } = await import("../src/compress.js");+ const { generateText } = await import("ai");+ const { createOpenAI } = await import("@ai-sdk/openai");++ const openai = createOpenAI({ apiKey: process.env.OPENAI_API_KEY });++ // Simulated large tool result (the MDX shows `largeLogData`)+ const largeLogData = Array.from({ length: 100 }, (_, i) => ({+ timestamp: new Date(Date.now() - i * 1000).toISOString(),+ level: i === 42 ? "FATAL" : i % 10 === 0 ? "ERROR" : "INFO",+ service: `service-${["auth", "payment", "user", "api"][i % 4]}`,+ message:+ i === 42+ ? "Connection pool exhausted — max_connections=100 reached, 47 pending requests"+ : `Request processed in ${Math.round(Math.random() * 500)}ms for /${["login", "checkout", "profile", "notify"][i % 4]}`,+ trace_id: `trace-${Math.random().toString(36).substring(2, 10)}`,+ }));++ // EXACT pattern from MDX (adapted to use AI SDK message format)+ const messages: any[] = [+ { role: "user", content: "Analyze the server logs" },+ {+ role: "assistant",+ content: [+ {+ type: "tool-call",+ toolCallId: "tc_1",+ toolName: "get_logs",+ args: { limit: 100 },+ },+ ],+ },+ {+ role: "tool",+ content: [+ {+ type: "tool-result",+ toolCallId: "tc_1",+ toolName: "get_logs",+ output: { type: "json", value: largeLogData },+ },+ ],+ },+ { role: "user", content: "What is the critical issue?" },+ ];++ const compressed = await compress(messages, {+ model: "gpt-4o",+ baseUrl: PROXY_URL,+ });++ console.log(+ ` Example 1: ${compressed.tokensBefore} → ${compressed.tokensAfter} tokens (saved ${compressed.tokensSaved})`,+ );++ // Verify compression happened+ expect(compressed.tokensBefore).toBeGreaterThan(0);+ // Messages should still be in a format generateText accepts+ expect(compressed.messages.length).toBeGreaterThan(0);++ // Now call generateText with compressed messages+ const { text } = await generateText({+ model: openai("gpt-4o-mini"),+ messages: compressed.messages,+ });++ console.log(` LLM response: "${text.substring(0, 150)}"`);+ expect(text.length).toBeGreaterThan(0);+ // Should find the FATAL connection pool issue+ expect(text.toLowerCase()).toMatch(/connection|pool|fatal|exhaust/i);+ });++ // =====================================================+ // Example 2 from 51-headroom.mdx: "Use as middleware"+ // =====================================================+ it("headroomMiddleware() with wrapLanguageModel → generateText()", { timeout: 30000 }, async () => {+ const { headroomMiddleware } = await import("../src/adapters/vercel-ai.js");+ const { wrapLanguageModel, generateText } = await import("ai");+ const { createOpenAI } = await import("@ai-sdk/openai");++ const openai = createOpenAI({ apiKey: process.env.OPENAI_API_KEY });++ // EXACT pattern from MDX+ const model = wrapLanguageModel({+ model: openai("gpt-4o-mini"),+ middleware: headroomMiddleware({ baseUrl: PROXY_URL }),+ });++ // Feed it a big prompt that will get compressed+ const serverData = Array.from({ length: 100 }, (_, i) => ({+ name: `server-${i + 1}`,+ status: i % 15 === 0 ? "critical" : "healthy",+ cpu: Math.round(Math.random() * 100),+ alert: i % 15 === 0 ? `Disk at ${90 + (i % 10)}%` : null,+ description: `Production server ${i + 1} running service-${["auth", "payment"][i % 2]}`,+ }));++ const { text } = await generateText({+ model,+ system: "List only the critical servers. One line each.",+ prompt: `Fleet status:\n${JSON.stringify(serverData)}`,+ });++ console.log(` Example 2 (middleware): "${text.substring(0, 150)}"`);+ expect(text.length).toBeGreaterThan(0);+ expect(text.toLowerCase()).toMatch(/server/i);+ });++ // =====================================================+ // Example 3 from 51-headroom.mdx: "Works with any provider" (Anthropic)+ // =====================================================+ it("compress() → Anthropic via AI SDK", { timeout: 30000 }, async () => {+ if (!process.env.ANTHROPIC_API_KEY) {+ console.log(" Skipping: ANTHROPIC_API_KEY not set");+ return;+ }++ const { compress } = await import("../src/compress.js");+ const { generateText } = await import("ai");+ const { createAnthropic } = await import("@ai-sdk/anthropic");++ const anthropic = createAnthropic({+ apiKey: process.env.ANTHROPIC_API_KEY,+ baseURL: "https://api.anthropic.com/v1",+ });++ const searchResults = Array.from({ length: 80 }, (_, i) => ({+ title: `${["API Design", "Database Tuning", "Cache Strategy", "Load Balancing"][i % 4]} Guide ${i + 1}`,+ url: `https://docs.example.com/${i + 1}`,+ snippet: `Covers ${["best practices", "pitfalls", "advanced techniques", "getting started"][i % 4]} for ${["microservices", "distributed systems", "cloud native", "serverless"][i % 4]}.`,+ score: (100 - i) / 100,+ }));++ // Simple messages (no tool calls — just user content)+ const messages: any[] = [+ {+ role: "user",+ content: `Search results:\n${JSON.stringify(searchResults)}\n\nTop 3 results? One sentence each.`,+ },+ ];++ // EXACT pattern from MDX+ const compressed = await compress(messages, {+ model: "claude-haiku-4-5-20251001",+ baseUrl: PROXY_URL,+ });++ console.log(+ ` Example 3 (Anthropic): ${compressed.tokensBefore} → ${compressed.tokensAfter} tokens`,+ );++ const { text } = await generateText({+ model: anthropic("claude-haiku-4-5-20251001"),+ messages: compressed.messages,+ maxTokens: 300,+ });++ console.log(` Anthropic response: "${text.substring(0, 150)}"`);+ expect(text.length).toBeGreaterThan(0);+ });+});++describe.skipIf(!RUN)("MDX Examples: context-compression-middleware.mdx", () => {+ beforeAll(async () => {+ const res = await fetch(`${PROXY_URL}/health`);+ if (!res.ok) throw new Error("Proxy not running");+ });++ // =====================================================+ // The cookbook's compressionMiddleware — test the EXACT code from the MDX+ // =====================================================+ it("compressionMiddleware from cookbook works end-to-end", { timeout: 30000 }, async () => {+ const { compress } = await import("../src/compress.js");+ const { wrapLanguageModel, generateText } = await import("ai");+ const { createOpenAI } = await import("@ai-sdk/openai");++ const openai = createOpenAI({ apiKey: process.env.OPENAI_API_KEY });++ // EXACT middleware from the cookbook MDX+ const compressionMiddleware = {+ transformParams: async ({ params }: { params: any }) => {+ const prompt = params.prompt;+ if (!prompt || prompt.length === 0) return params;++ const result = await compress(prompt, {+ model: params.modelId ?? "gpt-4o",+ baseUrl: PROXY_URL,+ });++ if (!result.compressed) return params;++ console.log(+ ` Compressed: ${result.tokensBefore} → ${result.tokensAfter} tokens (saved ${result.tokensSaved})`,+ );++ return { ...params, prompt: result.messages };+ },+ };++ // EXACT pattern from cookbook: wrap model with middleware+ const model = wrapLanguageModel({+ model: openai("gpt-4o-mini"),+ middleware: compressionMiddleware,+ });++ // Simulate the SRE agent scenario from the cookbook+ const serverData = Array.from({ length: 100 }, (_, i) => ({+ id: i + 1,+ name: `server-${i + 1}`,+ status: i % 15 === 0 ? "critical" : i % 5 === 0 ? "warning" : "healthy",+ cpu: Math.round(Math.random() * 100),+ memory: Math.round(Math.random() * 100),+ region: ["us-east-1", "eu-west-1", "ap-southeast-1"][i % 3],+ lastAlert: i % 15 === 0 ? `Disk usage at ${90 + (i % 10)}%` : null,+ }));++ const { text } = await generateText({+ model,+ system: "You are an SRE assistant. List only the critical servers.",+ prompt: `Fleet status:\n${JSON.stringify(serverData)}`,+ });++ console.log(` Cookbook middleware response: "${text.substring(0, 200)}"`);+ expect(text.length).toBeGreaterThan(0);+ expect(text.toLowerCase()).toMatch(/server/i);+ });+});diff --git a/sdk/typescript/test/utils/format.test.ts b/sdk/typescript/test/utils/format.test.tsindex f463b84e7..c280caa0f 100644--- a/sdk/typescript/test/utils/format.test.ts+++ b/sdk/typescript/test/utils/format.test.ts@@ -327,7 +327,7 @@ describe("openAIToVercel", () => {type: "tool-call",toolCallId: "tc_1",toolName: "search",- args: { q: "test" },+ input: { q: "test" },});});@@ -347,7 +347,7 @@ describe("openAIToVercel", () => {];const result = openAIToVercel(msgs);expect(result[0].content).toEqual([- { type: "tool-call", toolCallId: "tc_1", toolName: "fn", args: {} },+ { type: "tool-call", toolCallId: "tc_1", toolName: "fn", input: {} },]);});@@ -428,7 +428,7 @@ describe("round-trip conversion", () => {type: "tool-call",toolCallId: "tc_1",toolName: "search",- args: { q: "test" },+ input: { q: "test" },});});});