[V6] Migrating to LanguageModelV3 for AI SDK v6 support

.changeset/late-ducks-think.md

@@ -0,0 +1,5 @@
+---
+"@ai-sdk-tool/parser": major
+---
+
+upgrade language model interfaces to V3

.changeset/pre.json

@@ -0,0 +1,16 @@
+{
+  "mode": "pre",
+  "tag": "canary",
+  "initialVersions": {
+    "@ai-sdk-tool/examples-eval-core": "0.0.0",
+    "@ai-sdk-tool/examples-parser-core": "0.0.0",
+    "@ai-sdk-tool/examples-rxml-core": "0.0.0",
+    "@ai-sdk-tool/eval": "0.1.8",
+    "@ai-sdk-tool/parser": "2.1.7",
+    "@ai-sdk-tool/rxml": "0.1.1"
+  },
+  "changesets": [
+    "late-ducks-think",
+    "spicy-ants-scream"
+  ]
+}

.changeset/spicy-ants-scream.md

@@ -0,0 +1,6 @@
+---
+"@ai-sdk-tool/parser": major
+"@ai-sdk-tool/eval": major
+---
+
+bump ai v6 (middleware v3 not yet)

.github/dependabot.yml

@@ -0,0 +1,43 @@
+version: 2
+updates:
+  # Root package dependencies (pnpm workspace)
+  # This covers all workspace packages in apps/*, packages/*, and internals/*
+  - package-ecosystem: "npm"
+    directory: "/"
+    schedule:
+      interval: "daily"
+      time: "09:00"
+      timezone: "Asia/Seoul"
+    open-pull-requests-limit: 5
+    versioning-strategy: increase
+    groups:
+      npm-all-updates:
+        patterns:
+          - "*"
+    labels:
+      - "dependencies"
+      - "automated"
+      - "npm"
+    commit-message:
+      prefix: "dependabot:"
+      include: "scope"
+
+  # GitHub Actions
+  - package-ecosystem: "github-actions"
+    directory: "/"
+    schedule:
+      interval: "daily"
+      time: "09:00"
+      timezone: "Asia/Seoul"
+    open-pull-requests-limit: 5
+    groups:
+      github-actions:
+        patterns:
+          - "*"
+    labels:
+      - "dependencies"
+      - "automated"
+      - "github-actions"
+    commit-message:
+      prefix: "dependabot:"
+      include: "scope"

.github/workflows/release-changeset.yml

@@ -58,7 +58,7 @@ jobs:
         id: changesets
         uses: changesets/action@v1
         with:
-          branch: main
+          branch: ${{ github.ref_name }}
           # This expects you to have a script called release which does a build for your packages and calls changeset publish
           version: pnpm ci:version
           publish: pnpm ci:release

docs/concepts/middleware.md

@@ -1,6 +1,6 @@
 # [dev] Middleware Architecture
 
-The middleware composes with AI SDK `LanguageModelV2Middleware` to provide tool calling for models without native support.
+The middleware composes with AI SDK `LanguageModelV3Middleware` to provide tool calling for models without native support.
 
 ## Responsibilities
 
@@ -22,6 +22,7 @@ See `packages/parser/src/tool-call-middleware.ts` and `packages/parser/src/proto
 ## How it works (end-to-end)
 
 1. `transformParams`
+
    - Extracts custom function tools (`type: "function"`) and renders a system prompt via `protocol.formatTools` and a provided `toolSystemPromptTemplate`.
    - Normalizes the existing prompt:
      - Assistant tool-call parts are converted to provider-friendly text with `protocol.formatToolCall`.
@@ -34,6 +35,7 @@ See `packages/parser/src/tool-call-middleware.ts` and `packages/parser/src/proto
      - `toolChoice: { type: "none" }`: not supported (throws). Use `auto` (default) instead.
 
 2. `wrapStream`
+
    - If tool-choice fast-path is active, performs a single `generate` call and emits a synthetic `tool-call` followed by `finish`.
    - Otherwise, pipes provider stream through `protocol.createStreamParser`, emitting normalized `tool-call` parts as they arrive.
 

docs/evaluation.md

@@ -40,16 +40,16 @@ See runnable examples in `examples/eval-core/src/*`.
 
 ## [dev] Create a Custom Benchmark
 
-Implement `LanguageModelV2Benchmark` and pass it to `evaluate`.
+Implement `LanguageModelV3Benchmark` and pass it to `evaluate`.
 
 ```ts
 import { generateText } from "ai";
 import type {
-  LanguageModelV2Benchmark,
+  LanguageModelV3Benchmark,
   BenchmarkResult,
 } from "@ai-sdk-tool/eval";
 
-export const myBenchmark: LanguageModelV2Benchmark = {
+export const myBenchmark: LanguageModelV3Benchmark = {
   name: "my-benchmark",
   version: "1.0.0",
   description: "Minimal example",

examples/eval-core/package.json

@@ -13,12 +13,12 @@
   "dependencies": {
     "@ai-sdk-tool/eval": "workspace:*",
     "@ai-sdk-tool/parser": "workspace:*",
-    "@ai-sdk/openai": "2.0.56",
-    "@ai-sdk/openai-compatible": "1.0.23",
-    "ai": "5.0.81",
+    "@ai-sdk/openai": "3.0.0-beta.48",
+    "@ai-sdk/openai-compatible": "2.0.0-beta.31",
+    "ai": "6.0.0-beta.92",
     "tsx": "^4.20.6"
   },
   "devDependencies": {
-    "@types/node": "^24.9.1"
+    "@types/node": "^24.10.0"
   }
 }

examples/parser-core/package.json

@@ -13,14 +13,14 @@
   "license": "Apache-2.0",
   "dependencies": {
     "@ai-sdk-tool/parser": "workspace:*",
-    "@ai-sdk/openai": "2.0.56",
-    "@ai-sdk/openai-compatible": "1.0.23",
-    "ai": "5.0.81",
+    "@ai-sdk/openai": "3.0.0-beta.48",
+    "@ai-sdk/openai-compatible": "2.0.0-beta.31",
+    "ai": "6.0.0-beta.92",
     "tsx": "^4.20.6",
     "zod": "^4.1.12"
   },
   "devDependencies": {
     "@ai-sdk/provider": "2.0.0",
-    "@types/node": "^24.9.1"
+    "@types/node": "^24.10.0"
   }
 }

examples/parser-core/src/00-stream-tool-call.ts

@@ -7,16 +7,28 @@ import { z } from "zod";
 const MAX_STEPS = 4;
 const MAX_TEMPERATURE = 100;
 
-const openrouter = createOpenAICompatible({
-  name: "openrouter",
-  apiKey: process.env.OPENROUTER_API_KEY,
-  baseURL: "https://openrouter.ai/api/v1",
+// const openrouter = createOpenAICompatible({
+//   name: "openrouter",
+//   apiKey: process.env.OPENROUTER_API_KEY,
+//   baseURL: "https://openrouter.ai/api/v1",
+// });
+
+const friendli = createOpenAICompatible({
+  name: "friendli",
+  apiKey: process.env.FRIENDLI_TOKEN,
+  baseURL: "https://api.friendli.ai/serverless/v1",
+  includeUsage: true,
+  fetch: (url, options) => {
+    const body = options?.body ? JSON.parse(options.body as string) : {};
+    body.parse_reasoning = true;
+    return fetch(url, { ...options, body: JSON.stringify(body) });
+  },
 });
 
 async function main() {
   const result = streamText({
     model: wrapLanguageModel({
-      model: openrouter("z-ai/glm-4.5-air"),
+      model: friendli("zai-org/GLM-4.6"),
       middleware: sijawaraDetailedXmlToolMiddleware,
     }),
 
@@ -62,6 +74,9 @@ async function main() {
   for await (const part of result.fullStream) {
     if (part.type === "text-delta") {
       process.stdout.write(part.text);
+    } else if (part.type === "reasoning-delta") {
+      // Print reasoning text in a different color (e.g., yellow)
+      process.stdout.write(`\x1b[33m${part.text}\x1b[0m`);
     } else if (part.type === "tool-result") {
       console.log({
         name: part.toolName,

examples/parser-core/src/logging-middleware.ts

@@ -1,6 +1,6 @@
 import type {
-  LanguageModelV2Middleware,
-  LanguageModelV2StreamPart,
+  LanguageModelV3Middleware,
+  LanguageModelV3StreamPart,
 } from "@ai-sdk/provider";
 
 const INV = "\x1b[7m"; // ANSI SGR: reverse video
@@ -39,7 +39,7 @@ function invLog(...args: unknown[]) {
   }
 }
 
-export const loggingMiddleware: LanguageModelV2Middleware = {
+export const loggingMiddleware: LanguageModelV3Middleware = {
   wrapGenerate: async ({ doGenerate, params }) => {
     invLog("doGenerate called");
     invLog(`params: ${JSON.stringify(params, null, 2)}`);
@@ -62,8 +62,8 @@ export const loggingMiddleware: LanguageModelV2Middleware = {
     const textBlocks = new Map<string, string>();
 
     const transformStream = new TransformStream<
-      LanguageModelV2StreamPart,
-      LanguageModelV2StreamPart
+      LanguageModelV3StreamPart,
+      LanguageModelV3StreamPart
     >({
       transform(chunk, controller) {
         switch (chunk.type) {

examples/parser-core/src/xx-debug-file-write.ts

@@ -150,7 +150,7 @@ async function main() {
     detail: string;
     first?: FirstToolCall;
   }>;
-  for (let i = 1; i <= MAX_RUNS; i++) {
+  for (let i = 1; i <= MAX_RUNS; i += 1) {
     console.log(`\n--- Run ${i} ---`);
     const outcome = await runOnce(i);
     results.push(outcome);

examples/rxml-core/package.json

@@ -16,7 +16,7 @@
     "tsx": "^4.20.6"
   },
   "devDependencies": {
-    "@types/node": "^24.9.1",
+    "@types/node": "^24.10.0",
     "zod": "^4.1.12"
   }
 }

package.json

@@ -15,17 +15,17 @@
     "ci:version": "changeset version"
   },
   "devDependencies": {
-    "@biomejs/biome": "2.3.1",
+    "@biomejs/biome": "2.3.3",
     "@changesets/cli": "2.29.7",
-    "@types/node": "^24.9.1",
-    "@vitest/coverage-v8": "^4.0.4",
-    "vitest": "^4.0.4",
-    "globals": "^16.4.0",
+    "@types/node": "^24.10.0",
+    "@vitest/coverage-v8": "^4.0.6",
+    "globals": "^16.5.0",
     "jiti": "^2.6.1",
     "npm-run-all": "^4.1.5",
-    "turbo": "^2.5.8",
+    "turbo": "^2.6.0",
     "typescript": "5.9.3",
-    "ultracite": "6.0.4"
+    "ultracite": "6.2.0",
+    "vitest": "^4.0.6"
   },
   "packageManager": "[email protected]",
   "engines": {

packages/eval/CHANGELOG.md

@@ -1,5 +1,11 @@
 # @ai-sdk-tool/eval
 
+## 1.0.0-canary.0
+
+### Major Changes
+
+- df62ec5: bump ai v6 (middleware v3 not yet)
+
 ## 0.1.8
 
 ### Patch Changes

packages/eval/README.md

@@ -14,7 +14,7 @@ It allows developers to:
 
 ## Core Concepts
 
-- **Benchmark (`LanguageModelV2Benchmark`)**: A standardized interface for creating an evaluation task. It has a `run` method that takes a `LanguageModel` and returns a `BenchmarkResult`.
+- **Benchmark (`LanguageModelV3Benchmark`)**: A standardized interface for creating an evaluation task. It has a `run` method that takes a `LanguageModel` and returns a `BenchmarkResult`.
 - **`evaluate` function**: The core function that runs a set of benchmarks against one or more models and provides a report on the results.
 - **Reporter**: Formats the evaluation results into different outputs, such as a human-readable console report or a machine-readable JSON object.
 
@@ -77,20 +77,20 @@ cd examples/eval-core && pnpm dlx tsx src/json-generation.ts
 
 ## Creating a Custom Benchmark
 
-You can easily create your own benchmark by implementing the `LanguageModelV2Benchmark` interface. This is useful for testing model performance on tasks specific to your application.
+You can easily create your own benchmark by implementing the `LanguageModelV3Benchmark` interface. This is useful for testing model performance on tasks specific to your application.
 
 **Example: A custom benchmark to test politeness.**
 
 ```typescript
 import {
-  LanguageModelV2Benchmark,
+  LanguageModelV3Benchmark,
   BenchmarkResult,
   EvaluateOptions,
 } from "@ai-sdk-tool/eval";
 import { LanguageModel, generateText } from "ai";
 
 // Define the benchmark object
-export const politenessBenchmark: LanguageModelV2Benchmark = {
+export const politenessBenchmark: LanguageModelV3Benchmark = {
   name: "politeness-check",
   version: "1.0.0",
   description: "Checks if the model's response is polite.",

packages/eval/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ai-sdk-tool/eval",
-  "version": "0.1.8",
+  "version": "1.0.0-canary.0",
   "license": "Apache-2.0",
   "type": "module",
   "main": "./dist/index.js",
@@ -28,12 +28,12 @@
     "test:e2e": "tsup src/run-test.ts --format esm --clean --out-dir dist --no-config && node dist/run-test.js"
   },
   "dependencies": {
-    "@ai-sdk/provider": "2.0.0",
+    "@ai-sdk/provider": "3.0.0-beta.14",
     "ajv": "^8.17.1",
     "zod": "^4.1.12"
   },
   "devDependencies": {
-    "@types/node": "^24.9.1",
+    "@types/node": "^24.10.0",
     "tsup": "^8.5.0",
     "typescript": "^5.9.3"
   },

packages/eval/src/benchmarks/bfcl.ts

@@ -8,7 +8,7 @@ import {
   tool,
 } from "ai";
 
-import type { BenchmarkResult, LanguageModelV2Benchmark } from "@/interfaces";
+import type { BenchmarkResult, LanguageModelV3Benchmark } from "@/interfaces";
 import { resolveDataDir } from "@/utils/paths";
 
 import {
@@ -124,7 +124,7 @@ function createBfclBenchmark(
   description: string,
   testDataFile: string,
   answerDataFile: string
-): LanguageModelV2Benchmark {
+): LanguageModelV3Benchmark {
   return {
     name,
     version: "1.0.0",
@@ -563,7 +563,7 @@ function createBfclBenchmark(
           restoredCalls: Record<string, unknown>[],
           usedActual: Set<number>
         ): number => {
-          for (let i = 0; i < restoredCalls.length; i++) {
+          for (let i = 0; i < restoredCalls.length; i += 1) {
             if (usedActual.has(i)) {
               continue;
             }
@@ -1149,7 +1149,8 @@ function createBfclBenchmark(
             .fill(0)
             .map(async () => {
               while (true) {
-                const current = idx++;
+                const current = idx;
+                idx += 1;
                 if (current >= items.length) {
                   break;
                 }

packages/eval/src/benchmarks/bfcl/ast-checker.ts

@@ -379,7 +379,7 @@ export function parallelFunctionCheckerNoOrder(
     }
 
     let foundMatch = false;
-    for (let i = 0; i < modelToolCalls.length; i++) {
+    for (let i = 0; i < modelToolCalls.length; i += 1) {
       if (matchedModelCallIndices.has(i)) {
         continue;
       }