diff --git a/.changeset/ai-sdk-v7-support.md b/.changeset/ai-sdk-v7-support.md
new file mode 100644
index 0000000..d75b84b
--- /dev/null
+++ b/.changeset/ai-sdk-v7-support.md
@@ -0,0 +1,31 @@
+---
+"@context-chef/ai-sdk-middleware": major
+---
+
+Support AI SDK v7 (provider spec V4).
+
+The middleware now targets `ai@>=7` / `@ai-sdk/provider@>=4`: it implements the V4
+language-model middleware spec (`specificationVersion: 'v4'`) and all public types
+move from `LanguageModelV3*` to `LanguageModelV4*`. AI SDK v7's `wrapLanguageModel`
+rejects a v3-spec middleware, so this is a breaking change that requires AI SDK v7.
+
+**Migration**
+
+- On AI SDK v7 (`ai@7`): upgrade to `@context-chef/ai-sdk-middleware@2`.
+- Still on AI SDK v6 (`ai@6`): stay on `@context-chef/ai-sdk-middleware@1` — the 1.x
+  line continues to support the v3 spec. No code change is forced on you.
+
+**Removed** (deprecated APIs that were slated for removal in the next major):
+
+- `planCompaction`, `compactHistory`, and the `CompactionPlan` type (the provider-prompt
+  altitude variants) — use `planCompactionModelMessages` / `compactModelMessages` /
+  `CompactionPlanModelMessages` at the `ModelMessage` altitude instead.
+- `onBudgetExceeded` on `ContextChefOptions` — use `onBeforeCompress` instead.
+
+Runtime behavior is unchanged. The only V4 nuance: provider-level `FilePart.data`
+became a tagged union (`SharedV4FileData`); the prompt adapter handles it
+transparently and the binary/URL payload still round-trips losslessly.
+
+On v7, durable in-loop compaction via `compactModelMessages` inside a
+`ToolLoopAgent` `prepareStep` now persists across steps (AI SDK v7 carries
+`prepareStep`-returned messages forward into later steps — v6 did not).
diff --git a/README.md b/README.md
index 05e4fcc..919ff69 100644
--- a/README.md
+++ b/README.md
@@ -940,7 +940,7 @@ Install only what you need:
 # Core library (OpenAI / Anthropic / Gemini direct SDK usage)
 npx skills add MyPrototypeWhat/context-chef --skill context-chef-core
 
-# AI SDK middleware (Vercel AI SDK v6+)
+# AI SDK middleware (Vercel AI SDK v7+)
 npx skills add MyPrototypeWhat/context-chef --skill context-chef-middleware
 
 # TanStack AI middleware (TanStack AI v0.10+)
diff --git a/packages/ai-sdk-middleware/README.md b/packages/ai-sdk-middleware/README.md
index 96d9399..51e3878 100644
--- a/packages/ai-sdk-middleware/README.md
+++ b/packages/ai-sdk-middleware/README.md
@@ -4,7 +4,7 @@
 [![npm downloads](https://img.shields.io/npm/dm/@context-chef/ai-sdk-middleware.svg)](https://www.npmjs.com/package/@context-chef/ai-sdk-middleware)
 [![License](https://img.shields.io/npm/l/@context-chef/ai-sdk-middleware.svg)](https://github.com/MyPrototypeWhat/context-chef/blob/main/LICENSE)
 [![TypeScript](https://img.shields.io/badge/TypeScript-5.9-blue.svg)](https://www.typescriptlang.org/)
-[![AI SDK](https://img.shields.io/badge/AI%20SDK-v6-black.svg)](https://ai-sdk.dev)
+[![AI SDK](https://img.shields.io/badge/AI%20SDK-v7-black.svg)](https://ai-sdk.dev)
 
 [Vercel AI SDK](https://ai-sdk.dev) middleware powered by [context-chef](https://github.com/MyPrototypeWhat/context-chef). Transparent history compression, tool result truncation, and token budget management — zero code changes required.
 
@@ -16,6 +16,9 @@
 npm install @context-chef/ai-sdk-middleware ai
 ```
 
+> **AI SDK version.** `2.x` targets **AI SDK v7** (`ai@>=7`). Still on AI SDK v6?
+> Use `@context-chef/ai-sdk-middleware@1` — the `1.x` line supports `ai@6`.
+
 ## Quick Start
 
 ```typescript
@@ -309,54 +312,6 @@ if (toSummarize.length > 0) {
 
 ModelMessage-altitude sibling of [`summarizeMessages`](#summarizemessagesprompt-model-options): summarize a `ModelMessage[]` slice into a single summary string via the same pipeline (role-flattening + core `summarizeHistory`). System messages are dropped. An empty slice returns `''` without a model call; it throws if the model call fails. Use it with `planCompactionModelMessages` when you want to drive summarization yourself instead of the one-shot `compactModelMessages`.
 
-### `compactHistory(prompt, model, options)`
-
-> **Deprecated.** `compactHistory` / `planCompaction` take and return
-> `LanguageModelV3Prompt` — the provider-protocol altitude, which nobody
-> persists. Use [`compactModelMessages`](#compactmodelmessagesmessages-model-options)
-> / `planCompactionModelMessages` instead. Still exported and working; removed in the next major.
-
-The V3-prompt variant of `compactModelMessages`. It splits the history on turn boundaries, summarizes the old slice, and returns a new prompt ready to persist — `[...system, <summary>, ...recent turns]`:
-
-```typescript
-import { compactHistory } from '@context-chef/ai-sdk-middleware';
-
-// Between model calls, when you own `messages`:
-messages = await compactHistory(messages, summarizerModel, {
-  keepRecentTurns: 4,           // keep the last 4 atomic turns verbatim
-  toolResultStubThreshold: 5000,
-});
-// Persist the result — history actually shrinks and stays shrunk.
-```
-
-- The cut lands only on **turn boundaries** (an assistant + its tool results stay together), so it never orphans a tool result or splits a multi-block assistant message.
-- System messages are preserved verbatim and never summarized.
-- Returns the prompt **unchanged** when there is nothing old enough to compact (no more turns than `keepRecentTurns`) or the summarizer yields no text — safe to call unconditionally. Throws only if the model call throws.
-- Accepts the same `SummarizeMessagesOptions` (`customCompressionInstructions`, `toolResultStubThreshold`) as `summarizeMessages`.
-
-### `planCompaction(prompt, options)`
-
-> **Deprecated.** Use [`planCompactionModelMessages`](#plancompactionmodelmessagesmessages-options).
-> This V3-prompt variant is the provider-protocol altitude — a type you never
-> persist. Still exported and working; removed in the next major.
-
-The synchronous split behind `compactHistory`, for when you want the boundary without summarizing (persist your own marker, or use a different summarizer). Returns `{ system, toSummarize, toKeep }` (all `LanguageModelV3Prompt`), cut on turn boundaries:
-
-```typescript
-import { planCompaction, summarizeMessages } from '@context-chef/ai-sdk-middleware';
-import { Prompts } from '@context-chef/core';
-
-const { system, toSummarize, toKeep } = planCompaction(messages, { keepRecentTurns: 4 });
-if (toSummarize.length > 0) {
-  const summary = await summarizeMessages(toSummarize, model);
-  messages = [
-    ...system,
-    { role: 'user', content: [{ type: 'text', text: Prompts.getCompactSummaryWrapper(summary) }] },
-    ...toKeep,
-  ];
-}
-```
-
 ## How It Works
 
 ```
diff --git a/packages/ai-sdk-middleware/package.json b/packages/ai-sdk-middleware/package.json
index f210829..89aad6a 100644
--- a/packages/ai-sdk-middleware/package.json
+++ b/packages/ai-sdk-middleware/package.json
@@ -47,13 +47,13 @@
     "@context-chef/core": "workspace:*"
   },
   "peerDependencies": {
-    "@ai-sdk/provider": ">=3",
-    "ai": ">=6"
+    "@ai-sdk/provider": ">=4",
+    "ai": ">=7"
   },
   "devDependencies": {
-    "@ai-sdk/provider": "^3.0.8",
+    "@ai-sdk/provider": "^4.0.0",
     "@types/node": "^25.3.0",
-    "ai": "^6.0.140",
+    "ai": "^7.0.0",
     "tsdown": "^0.20.3",
     "typescript": "^5.9.3",
     "vitest": "^4.0.18"
diff --git a/packages/ai-sdk-middleware/src/adapter.ts b/packages/ai-sdk-middleware/src/adapter.ts
index 80b29b8..b35c7fc 100644
--- a/packages/ai-sdk-middleware/src/adapter.ts
+++ b/packages/ai-sdk-middleware/src/adapter.ts
@@ -1,9 +1,10 @@
 import type {
-  LanguageModelV3Message,
-  LanguageModelV3Prompt,
-  LanguageModelV3ToolResultOutput,
-  LanguageModelV3ToolResultPart,
-  SharedV3ProviderOptions,
+  LanguageModelV4Message,
+  LanguageModelV4Prompt,
+  LanguageModelV4ToolResultOutput,
+  LanguageModelV4ToolResultPart,
+  SharedV4FileData,
+  SharedV4ProviderOptions,
 } from '@ai-sdk/provider';
 import {
   type Attachment,
@@ -12,10 +13,24 @@ import {
   type ToolCall,
 } from '@context-chef/core';
 
+/**
+ * Extracts the Attachment presence/metadata signal from a V4 file part's `data`.
+ *
+ * V4 restructured `FilePart.data` from a bare value into a tagged union
+ * (`SharedV4FileData`). Only inline string data (`{ type: 'data', data: string }`)
+ * yields a recorded signal — bytes, URLs, and provider references become `''`,
+ * exactly as the pre-v4 adapter only recorded already-string `data`. The real
+ * payload always round-trips losslessly through `_userContent`/`_assistantContent`;
+ * this value is only a presence signal Janitor reads via `attachments?.length`.
+ */
+function fileDataSignal(data: SharedV4FileData): string {
+  return data.type === 'data' && typeof data.data === 'string' ? data.data : '';
+}
+
 /** Content types for each AI SDK message role */
-type UserContent = Extract<LanguageModelV3Message, { role: 'user' }>['content'];
-type AssistantContent = Extract<LanguageModelV3Message, { role: 'assistant' }>['content'];
-type ToolContent = Extract<LanguageModelV3Message, { role: 'tool' }>['content'];
+type UserContent = Extract<LanguageModelV4Message, { role: 'user' }>['content'];
+type AssistantContent = Extract<LanguageModelV4Message, { role: 'assistant' }>['content'];
+type ToolContent = Extract<LanguageModelV4Message, { role: 'tool' }>['content'];
 
 /**
  * Extended IR message with typed pass-through fields for lossless AI SDK round-trip.
@@ -26,12 +41,12 @@ export interface AISDKMessage extends Message {
   _assistantContent?: AssistantContent;
   _toolContent?: ToolContent;
   _originalText?: string;
-  _providerOptions?: SharedV3ProviderOptions;
+  _providerOptions?: SharedV4ProviderOptions;
   _toolName?: string;
 }
 
 /**
- * Converts an AI SDK V3 prompt to context-chef IR messages.
+ * Converts an AI SDK V4 prompt to context-chef IR messages.
  *
  * Original AI SDK content is stored in per-role fields for lossless round-trip.
  * `_originalText` caches the extracted text so `toAISDK` can detect Janitor modifications.
@@ -42,7 +57,7 @@ export interface AISDKMessage extends Message {
  * non-system message is a user message. This is a system boundary — IR
  * downstream is trusted to satisfy invariants.
  */
-export function fromAISDK(prompt: LanguageModelV3Prompt): AISDKMessage[] {
+export function fromAISDK(prompt: LanguageModelV4Prompt): AISDKMessage[] {
   const messages: AISDKMessage[] = [];
 
   for (const msg of prompt) {
@@ -73,7 +88,7 @@ export function fromAISDK(prompt: LanguageModelV3Prompt): AISDKMessage[] {
           // so we never invent a fake encoding for non-string inputs.
           attachments.push({
             mediaType: part.mediaType,
-            data: typeof part.data === 'string' ? part.data : '',
+            data: fileDataSignal(part.data),
             ...(part.filename ? { filename: part.filename } : {}),
           });
         }
@@ -127,7 +142,7 @@ export function fromAISDK(prompt: LanguageModelV3Prompt): AISDKMessage[] {
           // _assistantContent carries the actual payload through round-trip.
           attachments.push({
             mediaType: part.mediaType,
-            data: typeof part.data === 'string' ? part.data : '',
+            data: fileDataSignal(part.data),
             ...(part.filename ? { filename: part.filename } : {}),
           });
         }
@@ -189,14 +204,14 @@ function asAISDK(msg: Message): AISDKMessage {
 }
 
 /**
- * Converts context-chef IR messages back to AI SDK V3 prompt format.
+ * Converts context-chef IR messages back to AI SDK V4 prompt format.
  *
  * Uses per-role original content when unmodified (detected via `_originalText`).
  * Falls back to constructing from IR fields when content was modified by Janitor
  * (e.g. compact() cleared tool results) or for new messages (e.g. compression summaries).
  */
-export function toAISDK(messages: Message[]): LanguageModelV3Prompt {
-  const prompt: LanguageModelV3Prompt = [];
+export function toAISDK(messages: Message[]): LanguageModelV4Prompt {
+  const prompt: LanguageModelV4Prompt = [];
 
   let i = 0;
   while (i < messages.length) {
@@ -240,11 +255,11 @@ export function toAISDK(messages: Message[]): LanguageModelV3Prompt {
     }
 
     if (msg.role === 'tool') {
-      const toolResults: LanguageModelV3ToolResultPart[] = [];
+      const toolResults: LanguageModelV4ToolResultPart[] = [];
       // Re-attach the message-level providerOptions captured on the first IR
       // message of the original tool turn (see fromAISDK). Take the first
       // non-undefined across the coalesced group.
-      let providerOptions: SharedV3ProviderOptions | undefined;
+      let providerOptions: SharedV4ProviderOptions | undefined;
       while (i < messages.length && messages[i].role === 'tool') {
         const toolMsg = asAISDK(messages[i]);
         const toolModified =
@@ -289,7 +304,7 @@ export function toAISDK(messages: Message[]): LanguageModelV3Prompt {
   return prompt;
 }
 
-export function stringifyToolOutput(output: LanguageModelV3ToolResultOutput): string {
+export function stringifyToolOutput(output: LanguageModelV4ToolResultOutput): string {
   switch (output.type) {
     case 'text':
     case 'error-text':
diff --git a/packages/ai-sdk-middleware/src/compaction.ts b/packages/ai-sdk-middleware/src/compaction.ts
index 7da2bb0..05ab4c2 100644
--- a/packages/ai-sdk-middleware/src/compaction.ts
+++ b/packages/ai-sdk-middleware/src/compaction.ts
@@ -1,4 +1,3 @@
-import type { LanguageModelV3, LanguageModelV3Prompt } from '@ai-sdk/provider';
 import {
   compactHistory as coreCompactHistory,
   planCompaction as corePlanCompaction,
@@ -6,101 +5,11 @@ import {
 } from '@context-chef/core';
 import type { LanguageModel, ModelMessage } from 'ai';
 
-import { fromAISDK, toAISDK } from './adapter';
 import { createCompressionAdapter, type SummarizeMessagesOptions } from './middleware';
 import { fromModelMessages, toModelMessages } from './modelMessageAdapter';
 
 export type { PlanCompactionOptions } from '@context-chef/core';
 
-export interface CompactionPlan {
-  /** System messages, preserved verbatim — standing instructions are never summarized. */
-  system: LanguageModelV3Prompt;
-  /**
-   * The old conversation slice to summarize (system excluded). Feed this to
-   * `summarizeMessages`. Empty when there is nothing old enough to compact.
-   */
-  toSummarize: LanguageModelV3Prompt;
-  /** The recent conversation turns to keep verbatim. */
-  toKeep: LanguageModelV3Prompt;
-}
-
-/**
- * @deprecated Use {@link planCompactionModelMessages}. This V3-prompt variant is
- * the provider-protocol altitude — a type you never persist. Removed in the next
- * major.
- *
- * Splits an AI SDK prompt into `{ system, toSummarize, toKeep }` on **turn
- * boundaries**, for durable (caller-owned) compaction.
- *
- * Unlike the in-flight middleware `compress` — which only rewrites the outgoing
- * request and is discarded each call — this is a pure, synchronous split you run
- * against your *own* message store. Summarize `toSummarize`, then persist
- * `[...system, <summary>, ...toKeep]` back to your store so the history actually
- * shrinks. See {@link compactHistory} for the one-shot version.
- *
- * The AI-SDK-typed wrapper around core's provider-agnostic `planCompaction`:
- * converts the prompt to IR via {@link fromAISDK}, splits on turn boundaries
- * (assistant + its tool results stay together), and converts each slice back via
- * {@link toAISDK}.
- */
-export function planCompaction(
-  prompt: LanguageModelV3Prompt,
-  options: PlanCompactionOptions,
-): CompactionPlan {
-  const plan = corePlanCompaction(fromAISDK(prompt), options);
-  return {
-    system: toAISDK(plan.system),
-    toSummarize: toAISDK(plan.toSummarize),
-    toKeep: toAISDK(plan.toKeep),
-  };
-}
-
-/**
- * @deprecated Use {@link compactModelMessages}. `LanguageModelV3Prompt` is the
- * provider-protocol altitude (ephemeral, never persisted); durable compaction
- * belongs at the ModelMessage altitude. Removed in the next major.
- *
- * One-shot durable compaction: plan a turn-safe split, summarize the old slice,
- * and return a new prompt ready to persist — `[...system, <summary>, ...toKeep]`.
- *
- * This is the recommended way to keep a long conversation lean when you own the
- * message store (a long agent loop, or a chat past the budget). Run it between
- * model calls and replace your stored messages with the result; the summary is
- * a real `user` message wrapped with the "continued conversation" framing.
- *
- * Returns the prompt **unchanged** (same reference) when there is nothing old
- * enough to compact (no more turns than `keepRecentTurns`) or when the summarizer
- * yields no text — so it is safe to call unconditionally, and callers can skip
- * persistence on a no-op via `result === prompt`. Throws only if the model call
- * throws.
- *
- * The AI-SDK-typed wrapper around core's `compactHistory`: it binds `model` into
- * a compression callback via {@link createCompressionAdapter} (core never calls a
- * model directly). Do NOT also configure middleware `compress` (with a `model`)
- * on the same path — that compresses twice. Use this OR in-flight `compress`,
- * not both.
- *
- * @example
- * ```ts
- * // In your loop / between turns, when you own `messages`:
- * messages = await compactHistory(messages, summarizerModel, {
- *   keepRecentTurns: 4,
- *   toolResultStubThreshold: 5000,
- * });
- * ```
- */
-export async function compactHistory(
-  prompt: LanguageModelV3Prompt,
-  model: LanguageModelV3,
-  options: PlanCompactionOptions & SummarizeMessagesOptions,
-): Promise<LanguageModelV3Prompt> {
-  const ir = fromAISDK(prompt);
-  const result = await coreCompactHistory(ir, createCompressionAdapter(model), options);
-  // core returns the input IR reference on a no-op — preserve the original
-  // prompt reference so callers can skip persistence via `result === prompt`.
-  return result === ir ? prompt : toAISDK(result);
-}
-
 export interface CompactionPlanModelMessages {
   /** System messages, preserved verbatim — standing instructions are never summarized. */
   system: ModelMessage[];
@@ -143,7 +52,7 @@ export function planCompactionModelMessages(
  * loop, or inside a `ToolLoopAgent` `prepareStep` (`return { messages: await
  * compactModelMessages(messages, model, opts) }`).
  *
- * `model` is `ai`'s `LanguageModel` (string id | V3 | V2) — exactly what
+ * `model` is `ai`'s `LanguageModel` (string id | V4) — exactly what
  * `prepareStep`/`generateText` give you. Reuses core's `compactHistory` +
  * `createCompressionAdapter` (tool-role flattening); no model is called directly.
  *
diff --git a/packages/ai-sdk-middleware/src/index.ts b/packages/ai-sdk-middleware/src/index.ts
index 4b23b7a..22e94e5 100644
--- a/packages/ai-sdk-middleware/src/index.ts
+++ b/packages/ai-sdk-middleware/src/index.ts
@@ -1,4 +1,4 @@
-import type { LanguageModelV3 } from '@ai-sdk/provider';
+import type { LanguageModelV4 } from '@ai-sdk/provider';
 import { wrapLanguageModel } from 'ai';
 
 import { createMiddleware } from './middleware';
@@ -7,12 +7,9 @@ import type { ContextChefOptions } from './types';
 export type { ClearTarget } from '@context-chef/core';
 export { type AISDKMessage, fromAISDK, toAISDK } from './adapter';
 export {
-  type CompactionPlan,
   type CompactionPlanModelMessages,
-  compactHistory,
   compactModelMessages,
   type PlanCompactionOptions,
-  planCompaction,
   planCompactionModelMessages,
 } from './compaction';
 export {
@@ -50,9 +47,9 @@ export type {
  * ```
  */
 export function withContextChef(
-  model: LanguageModelV3,
+  model: LanguageModelV4,
   options: ContextChefOptions,
-): LanguageModelV3 {
+): LanguageModelV4 {
   const middleware = createMiddleware(options);
   return wrapLanguageModel({ model, middleware });
 }
diff --git a/packages/ai-sdk-middleware/src/middleware.ts b/packages/ai-sdk-middleware/src/middleware.ts
index 1dc34ba..8d75220 100644
--- a/packages/ai-sdk-middleware/src/middleware.ts
+++ b/packages/ai-sdk-middleware/src/middleware.ts
@@ -1,8 +1,8 @@
 import type {
-  LanguageModelV3,
-  LanguageModelV3Message,
-  LanguageModelV3Prompt,
-  LanguageModelV3StreamPart,
+  LanguageModelV4,
+  LanguageModelV4Message,
+  LanguageModelV4Prompt,
+  LanguageModelV4StreamPart,
 } from '@ai-sdk/provider';
 import {
   type ChefLogger,
@@ -56,14 +56,12 @@ export function createMiddleware(options: ContextChefOptions): LanguageModelMidd
   // `contextWindow`. Truncate/compact/skill/dynamicState-only
   // configurations get no Janitor at all: no budget checks, no token-usage
   // capture, and none of the Janitor's missing-tokenizer warnings.
-  const budgeting = Boolean(
-    options.compress || options.onCompress || options.onBeforeCompress || options.onBudgetExceeded,
-  );
+  const budgeting = Boolean(options.compress || options.onCompress || options.onBeforeCompress);
 
   if (budgeting && options.contextWindow == null) {
     throw new Error(
       '[context-chef] `contextWindow` is required when a compression option (`compress`, ' +
-        '`onCompress`, `onBeforeCompress`, `onBudgetExceeded`) is configured — the budget ' +
+        '`onCompress`, `onBeforeCompress`) is configured — the budget ' +
         'check has nothing to compare against without it.',
     );
   }
@@ -115,7 +113,7 @@ export function createMiddleware(options: ContextChefOptions): LanguageModelMidd
   }
 
   return {
-    specificationVersion: 'v3',
+    specificationVersion: 'v4',
 
     transformParams: async ({ params }) => {
       let { prompt } = params;
@@ -202,7 +200,7 @@ export function createMiddleware(options: ContextChefOptions): LanguageModelMidd
 
       const { stream, ...rest } = await doStream();
 
-      const transform = new TransformStream<LanguageModelV3StreamPart, LanguageModelV3StreamPart>({
+      const transform = new TransformStream<LanguageModelV4StreamPart, LanguageModelV4StreamPart>({
         transform(chunk, controller) {
           if (chunk.type === 'finish') {
             if (chunk.usage?.inputTokens?.total != null) {
@@ -254,7 +252,7 @@ function createJanitor(
         compressedMessages: toAISDK(details.compressedMessages),
       });
     },
-    onBeforeCompress: options.onBeforeCompress ?? options.onBudgetExceeded,
+    onBeforeCompress: options.onBeforeCompress,
     logger,
   };
 
@@ -283,18 +281,18 @@ function createJanitor(
 }
 
 /**
- * Prunes a LanguageModelV3Prompt via AI SDK's pruneMessages.
+ * Prunes a LanguageModelV4Prompt via AI SDK's pruneMessages.
  *
- * LanguageModelV3Message (from @ai-sdk/provider) and ModelMessage
+ * LanguageModelV4Message (from @ai-sdk/provider) and ModelMessage
  * (from @ai-sdk/provider-utils) share identical runtime structure but
  * differ at the TypeScript level (e.g. ImagePart, FilePart.data).
  * Since pruneMessages only filters — never transforms — every content
- * part in the output is an original V3 part, making the casts safe.
+ * part in the output is an original V4 part, making the casts safe.
  */
 function compactPrompt(
-  prompt: LanguageModelV3Prompt,
+  prompt: LanguageModelV4Prompt,
   config: Omit<Parameters<typeof pruneMessages>[0], 'messages'>,
-): LanguageModelV3Prompt {
+): LanguageModelV4Prompt {
   const messages = prompt.map(
     (msg) =>
       ({
@@ -310,7 +308,7 @@ function compactPrompt(
         role: msg.role,
         content: msg.content,
         providerOptions: msg.providerOptions,
-      }) as LanguageModelV3Message,
+      }) as LanguageModelV4Message,
   );
 }
 
@@ -339,9 +337,9 @@ async function resolveSkillMessages(skill: ContextChefOptions['skill']): Promise
  * - `system`: Adds as a standalone system message at the end.
  */
 async function injectDynamicState(
-  prompt: LanguageModelV3Prompt,
+  prompt: LanguageModelV4Prompt,
   config: DynamicStateConfig,
-): Promise<LanguageModelV3Prompt> {
+): Promise<LanguageModelV4Prompt> {
   const state = await config.getState();
   const xml = XmlGenerator.objectToXml(state, 'dynamic_state');
   const placement = config.placement ?? 'last_user';
@@ -383,7 +381,7 @@ function toCompressRole(role: string): CompressRole {
 }
 
 /**
- * Adapts an AI SDK LanguageModelV3 into the compressionModel callback
+ * Adapts an AI SDK LanguageModelV4 into the compressionModel callback
  * that Janitor expects: (messages: Message[]) => Promise<string>
  *
  * Tool messages are converted to user messages describing the tool interaction,
@@ -450,8 +448,8 @@ export type SummarizeMessagesOptions = SummarizeHistoryOptions;
  * `truncate`, `clear`, and `dynamicState`, remain safe to use alongside.
  */
 export async function summarizeMessages(
-  prompt: LanguageModelV3Prompt,
-  model: LanguageModelV3,
+  prompt: LanguageModelV4Prompt,
+  model: LanguageModelV4,
   opts: SummarizeMessagesOptions = {},
 ): Promise<string> {
   const ir = fromAISDK(prompt).filter((m) => m.role !== 'system');
diff --git a/packages/ai-sdk-middleware/src/modelMessageAdapter.ts b/packages/ai-sdk-middleware/src/modelMessageAdapter.ts
index 3b5f39f..bfb114f 100644
--- a/packages/ai-sdk-middleware/src/modelMessageAdapter.ts
+++ b/packages/ai-sdk-middleware/src/modelMessageAdapter.ts
@@ -1,4 +1,4 @@
-import type { LanguageModelV3ToolResultOutput } from '@ai-sdk/provider';
+import type { LanguageModelV4ToolResultOutput } from '@ai-sdk/provider';
 import {
   type Attachment,
   ensureValidHistory,
@@ -9,10 +9,11 @@ import type { ModelMessage } from 'ai';
 
 import { stringifyToolOutput } from './adapter';
 
-// NOTE: This adapter intentionally parallels src/adapter.ts (the V3 adapter). The
-// user/assistant/file extraction logic is duplicated by design; keep the two in
-// sync. The deltas here are deliberate: string-shorthand content, ImagePart, and
-// approval parts — none of which exist at the V3 (LanguageModelV3Prompt) altitude.
+// NOTE: This adapter intentionally parallels src/adapter.ts (the provider-prompt
+// adapter). The user/assistant/file extraction logic is duplicated by design; keep
+// the two in sync. The deltas here are deliberate: string-shorthand content,
+// ImagePart, and approval parts — none of which exist at the provider-prompt
+// (LanguageModelV4Prompt) altitude.
 
 // Content/part types derived from ModelMessage — no part-type imports needed
 // (provider-utils does not export them all stably). Same trick as adapter.ts.
@@ -20,10 +21,34 @@ type UserContent = Extract<ModelMessage, { role: 'user' }>['content'];
 type AssistantContent = Extract<ModelMessage, { role: 'assistant' }>['content'];
 type ToolContent = Extract<ModelMessage, { role: 'tool' }>['content'];
 type ProviderOptions = Extract<ModelMessage, { role: 'system' }>['providerOptions'];
+type MMFilePart = Extract<Exclude<UserContent, string>[number], { type: 'file' }>;
+
+/**
+ * Extracts the Attachment presence/metadata signal from an app-altitude file
+ * part's `data`. The app `ModelMessage` accepts both the bare shorthand
+ * (`string | Uint8Array | URL`) and V4's tagged shapes (`{ type: 'data', data }`,
+ * `{ type: 'url', url }`, …). Only inline string data yields a signal — bytes,
+ * URLs, and references become `''`, mirroring the provider-prompt adapter's
+ * `fileDataSignal`. The real payload round-trips losslessly via `_mm*Content`;
+ * this value is only read by Janitor via `attachments?.length`.
+ */
+function fileDataSignal(data: MMFilePart['data']): string {
+  if (typeof data === 'string') return data;
+  if (
+    typeof data === 'object' &&
+    data !== null &&
+    'type' in data &&
+    data.type === 'data' &&
+    typeof data.data === 'string'
+  ) {
+    return data.data;
+  }
+  return '';
+}
 
 /**
  * IR message carrying the original ModelMessage content for lossless round-trip.
- * Parallel to AISDKMessage (the V3 adapter's carrier) but typed to the
+ * Parallel to AISDKMessage (the provider-prompt adapter's carrier) but typed to the
  * application-layer ModelMessage shapes, and on distinct `_mm*` fields so the two
  * adapters can never read each other's pass-through by accident.
  */
@@ -79,7 +104,7 @@ export function fromModelMessages(messages: ModelMessage[]): ModelMessageIR[] {
           if (part.type === 'file') {
             attachments.push({
               mediaType: part.mediaType,
-              data: typeof part.data === 'string' ? part.data : '',
+              data: fileDataSignal(part.data),
               ...(part.filename ? { filename: part.filename } : {}),
             });
           } else if (part.type === 'image') {
@@ -141,7 +166,7 @@ export function fromModelMessages(messages: ModelMessage[]): ModelMessageIR[] {
           } else if (part.type === 'file') {
             attachments.push({
               mediaType: part.mediaType,
-              data: typeof part.data === 'string' ? part.data : '',
+              data: fileDataSignal(part.data),
               ...(part.filename ? { filename: part.filename } : {}),
             });
           }
@@ -174,8 +199,8 @@ export function fromModelMessages(messages: ModelMessage[]): ModelMessageIR[] {
       let firstOfMessage = true;
       for (const part of msg.content) {
         if (part.type === 'tool-result') {
-          // ModelMessage's ToolResultOutput is a structural superset of V3's (extra content[] members); stringifyToolOutput only reads .type/.value, so the cast is safe and matches the V3 adapter's projection.
-          const text = stringifyToolOutput(part.output as LanguageModelV3ToolResultOutput);
+          // ModelMessage's ToolResultOutput is a structural superset of the provider output's (extra content[] members); stringifyToolOutput only reads .type/.value, so the cast is safe and matches the provider-prompt adapter's projection.
+          const text = stringifyToolOutput(part.output as LanguageModelV4ToolResultOutput);
           anchor = {
             role: 'tool',
             content: text,
diff --git a/packages/ai-sdk-middleware/src/truncator.ts b/packages/ai-sdk-middleware/src/truncator.ts
index fdaf125..d0c0218 100644
--- a/packages/ai-sdk-middleware/src/truncator.ts
+++ b/packages/ai-sdk-middleware/src/truncator.ts
@@ -1,7 +1,7 @@
 import type {
-  LanguageModelV3Prompt,
-  LanguageModelV3ToolResultOutput,
-  LanguageModelV3ToolResultPart,
+  LanguageModelV4Prompt,
+  LanguageModelV4ToolResultOutput,
+  LanguageModelV4ToolResultPart,
 } from '@ai-sdk/provider';
 import { type ChefLogger, Offloader } from '@context-chef/core';
 import type { TruncateOptions } from './types';
@@ -11,10 +11,10 @@ import type { TruncateOptions } from './types';
  * When a storage adapter is provided, original content is persisted and a URI is included in the output.
  */
 export async function truncateToolResults(
-  prompt: LanguageModelV3Prompt,
+  prompt: LanguageModelV4Prompt,
   options: TruncateOptions,
   logger: ChefLogger = console,
-): Promise<LanguageModelV3Prompt> {
+): Promise<LanguageModelV4Prompt> {
   const { threshold, headChars = 0, tailChars = 1000, storage } = options;
 
   const offloader = storage
@@ -22,7 +22,7 @@ export async function truncateToolResults(
     : null;
   const policy = buildPolicyMap(options.perTool);
 
-  const result: LanguageModelV3Prompt = [];
+  const result: LanguageModelV4Prompt = [];
 
   for (const msg of prompt) {
     if (msg.role !== 'tool') {
@@ -68,8 +68,8 @@ export async function truncateToolResults(
             output: {
               type: 'text',
               value: vfsResult.content,
-            } satisfies LanguageModelV3ToolResultOutput,
-          } satisfies LanguageModelV3ToolResultPart);
+            } satisfies LanguageModelV4ToolResultOutput,
+          } satisfies LanguageModelV4ToolResultPart);
           continue;
         } catch (error) {
           logger.warn(
@@ -96,8 +96,8 @@ export async function truncateToolResults(
 
       newContent.push({
         ...part,
-        output: { type: 'text', value: truncated } satisfies LanguageModelV3ToolResultOutput,
-      } satisfies LanguageModelV3ToolResultPart);
+        output: { type: 'text', value: truncated } satisfies LanguageModelV4ToolResultOutput,
+      } satisfies LanguageModelV4ToolResultPart);
     }
 
     result.push({ ...msg, content: newContent });
@@ -137,7 +137,7 @@ function buildPolicyMap(perTool: TruncateOptions['perTool']): Map<string, ToolPo
   return map;
 }
 
-function extractText(output: LanguageModelV3ToolResultOutput): string {
+function extractText(output: LanguageModelV4ToolResultOutput): string {
   switch (output.type) {
     case 'text':
     case 'error-text':
diff --git a/packages/ai-sdk-middleware/src/types.ts b/packages/ai-sdk-middleware/src/types.ts
index b04cc28..7ed27d6 100644
--- a/packages/ai-sdk-middleware/src/types.ts
+++ b/packages/ai-sdk-middleware/src/types.ts
@@ -1,4 +1,4 @@
-import type { LanguageModelV3, LanguageModelV3Prompt } from '@ai-sdk/provider';
+import type { LanguageModelV4, LanguageModelV4Prompt } from '@ai-sdk/provider';
 import type {
   ChefLogger,
   ClearTarget,
@@ -67,7 +67,7 @@ export interface TruncateOptions {
  */
 export interface CompressOptions {
   /** A cheap model used for summarization (e.g. openai('gpt-4o-mini')). */
-  model: LanguageModelV3;
+  model: LanguageModelV4;
   /** Ratio of context window to preserve for recent messages. Default: 0.8 */
   preserveRatio?: number;
   /**
@@ -159,7 +159,7 @@ export interface ContextChefOptions {
    * The model's context window size in tokens.
    *
    * Required when a compression option is configured (`compress`,
-   * `onCompress`, `onBeforeCompress`, `onBudgetExceeded`) —
+   * `onCompress`, `onBeforeCompress`) —
    * `createMiddleware` throws otherwise. Optional (and unused) for
    * truncate / compact / skill / dynamicState-only configurations,
    * which involve no budget check.
@@ -233,7 +233,7 @@ export interface ContextChefOptions {
   onCompress?: (
     summary: string,
     truncatedCount: number,
-    details: { compressedMessages: LanguageModelV3Prompt },
+    details: { compressedMessages: LanguageModelV4Prompt },
   ) => void;
   /**
    * Called when token budget is exceeded, before LLM compression.
@@ -244,20 +244,13 @@ export interface ContextChefOptions {
     history: Message[],
     tokenInfo: { currentTokens: number; limit: number },
   ) => Message[] | null | undefined | Promise<Message[] | null | undefined>;
-  /**
-   * @deprecated Use `onBeforeCompress` instead. Will be removed in the next major version.
-   */
-  onBudgetExceeded?: (
-    history: Message[],
-    tokenInfo: { currentTokens: number; limit: number },
-  ) => Message[] | null | undefined | Promise<Message[] | null | undefined>;
   /**
    * Transform the AI SDK prompt after compression, before sending to the model.
    * Use for custom prompt manipulation, RAG injection, etc.
    */
   transformContext?: (
-    prompt: LanguageModelV3Prompt,
-  ) => LanguageModelV3Prompt | Promise<LanguageModelV3Prompt>;
+    prompt: LanguageModelV4Prompt,
+  ) => LanguageModelV4Prompt | Promise<LanguageModelV4Prompt>;
   /**
    * Sink for degradation warnings (storage write failures, missing usage
    * data, misconfiguration). Defaults to `console`. Forwarded to the
diff --git a/packages/ai-sdk-middleware/tests/adapter.test.ts b/packages/ai-sdk-middleware/tests/adapter.test.ts
index 5604902..d762eff 100644
--- a/packages/ai-sdk-middleware/tests/adapter.test.ts
+++ b/packages/ai-sdk-middleware/tests/adapter.test.ts
@@ -1,17 +1,17 @@
-import type { LanguageModelV3Prompt } from '@ai-sdk/provider';
+import type { LanguageModelV4Prompt } from '@ai-sdk/provider';
 import type { Message } from '@context-chef/core';
 import { describe, expect, it } from 'vitest';
 import { fromAISDK, toAISDK } from '../src/adapter';
 
 describe('fromAISDK', () => {
   it('converts system messages', () => {
-    const prompt: LanguageModelV3Prompt = [{ role: 'system', content: 'You are helpful.' }];
+    const prompt: LanguageModelV4Prompt = [{ role: 'system', content: 'You are helpful.' }];
     const result = fromAISDK(prompt);
     expect(result).toEqual([{ role: 'system', content: 'You are helpful.' }]);
   });
 
   it('converts user messages with text parts', () => {
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       {
         role: 'user',
         content: [
@@ -28,11 +28,11 @@ describe('fromAISDK', () => {
   it('stores original content including file parts', () => {
     const filePart = {
       type: 'file' as const,
-      data: 'base64data',
+      data: { type: 'data' as const, data: 'base64data' },
       mediaType: 'image/png',
     };
     const content = [{ type: 'text' as const, text: 'Look at this' }, filePart];
-    const prompt: LanguageModelV3Prompt = [{ role: 'user', content }];
+    const prompt: LanguageModelV4Prompt = [{ role: 'user', content }];
     const result = fromAISDK(prompt);
     expect(result[0].content).toBe('Look at this');
     expect(result[0]._userContent).toEqual(content);
@@ -40,13 +40,18 @@ describe('fromAISDK', () => {
   });
 
   it('maps multiple file parts to attachments', () => {
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       {
         role: 'user',
         content: [
           { type: 'text', text: 'Check these' },
-          { type: 'file', data: 'img1', mediaType: 'image/png' },
-          { type: 'file', data: 'doc1', mediaType: 'application/pdf', filename: 'report.pdf' },
+          { type: 'file', data: { type: 'data', data: 'img1' }, mediaType: 'image/png' },
+          {
+            type: 'file',
+            data: { type: 'data', data: 'doc1' },
+            mediaType: 'application/pdf',
+            filename: 'report.pdf',
+          },
         ],
       },
     ];
@@ -58,7 +63,7 @@ describe('fromAISDK', () => {
   });
 
   it('does not set attachments when no file parts exist', () => {
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'Just text' }] },
     ];
     const result = fromAISDK(prompt);
@@ -67,12 +72,12 @@ describe('fromAISDK', () => {
 
   it('preserves Uint8Array file data verbatim through _userContent (no encoding into Attachment.data)', () => {
     const bytes = new Uint8Array([72, 101, 108, 108, 111]); // "Hello"
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       {
         role: 'user',
         content: [
           { type: 'text', text: 'Binary please' },
-          { type: 'file', data: bytes, mediaType: 'image/png' },
+          { type: 'file', data: { type: 'data', data: bytes }, mediaType: 'image/png' },
         ],
       },
     ];
@@ -81,34 +86,34 @@ describe('fromAISDK', () => {
     // The real binary lives on _userContent for the AI SDK round-trip.
     expect(result[0].attachments).toEqual([{ mediaType: 'image/png', data: '' }]);
     const filePart = result[0]._userContent?.find((p) => p.type === 'file');
-    expect(filePart?.data).toBe(bytes); // same reference, not a copy
+    expect((filePart?.data as { data: unknown }).data).toBe(bytes); // same reference, not a copy
   });
 
   it('preserves URL file data verbatim through _userContent (no toString into Attachment.data)', () => {
     const url = new URL('https://example.com/img.png');
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       {
         role: 'user',
         content: [
           { type: 'text', text: 'Remote image' },
-          { type: 'file', data: url, mediaType: 'image/png' },
+          { type: 'file', data: { type: 'url', url }, mediaType: 'image/png' },
         ],
       },
     ];
     const result = fromAISDK(prompt);
     expect(result[0].attachments).toEqual([{ mediaType: 'image/png', data: '' }]);
     const filePart = result[0]._userContent?.find((p) => p.type === 'file');
-    expect(filePart?.data).toBe(url); // same URL instance, not toString'd
+    expect((filePart?.data as { url: unknown }).url).toBe(url); // same URL instance, not toString'd
   });
 
   it('toAISDK round-trips Uint8Array binary back to the AI SDK provider verbatim', () => {
     const bytes = new Uint8Array([1, 2, 3, 4, 5]);
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       {
         role: 'user',
         content: [
           { type: 'text', text: 'binary' },
-          { type: 'file', data: bytes, mediaType: 'image/png' },
+          { type: 'file', data: { type: 'data', data: bytes }, mediaType: 'image/png' },
         ],
       },
     ];
@@ -119,11 +124,11 @@ describe('fromAISDK', () => {
     const filePart = (userMsg?.content as Array<{ type: string; data?: unknown }>).find(
       (p) => p.type === 'file',
     );
-    expect(filePart?.data).toBe(bytes); // same reference
+    expect((filePart?.data as { data: unknown }).data).toBe(bytes); // same reference
   });
 
   it('converts assistant messages with text + tool calls', () => {
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'weather?' }] },
       {
         role: 'assistant',
@@ -162,7 +167,7 @@ describe('fromAISDK', () => {
   });
 
   it('converts assistant reasoning to thinking', () => {
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'reason about this' }] },
       {
         role: 'assistant',
@@ -178,13 +183,13 @@ describe('fromAISDK', () => {
   });
 
   it('maps assistant file parts to attachments', () => {
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'generate an image' }] },
       {
         role: 'assistant',
         content: [
           { type: 'text', text: 'Here is the image.' },
-          { type: 'file', data: 'generated_img', mediaType: 'image/png' },
+          { type: 'file', data: { type: 'data', data: 'generated_img' }, mediaType: 'image/png' },
         ],
       },
     ];
@@ -194,7 +199,7 @@ describe('fromAISDK', () => {
   });
 
   it('converts tool messages to individual IR messages', () => {
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'q' }] },
       {
         role: 'assistant',
@@ -248,7 +253,7 @@ describe('fromAISDK', () => {
   });
 
   it('handles json tool result output', () => {
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'query' }] },
       {
         role: 'assistant',
@@ -280,7 +285,7 @@ describe('fromAISDK', () => {
 
   // ─── Boundary sanitization ───
   it('injects placeholder for missing tool result at boundary', () => {
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'do it' }] },
       {
         role: 'assistant',
@@ -307,7 +312,7 @@ describe('fromAISDK', () => {
 
   // ─── FIX #3: tool-call with input:undefined yields a string '{}' ───
   it('serializes a tool-call with undefined input to "{}" (a string)', () => {
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'go' }] },
       {
         role: 'assistant',
@@ -342,7 +347,7 @@ describe('fromAISDK', () => {
   it('round-trips sanitized placeholder with original toolName (not "unknown")', () => {
     // Regression guard for the boundary-sanitize bug where injected placeholders
     // round-tripped as `toolName: 'unknown'` because toAISDK only read _toolName.
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'do it' }] },
       {
         role: 'assistant',
@@ -464,7 +469,7 @@ describe('toAISDK', () => {
 
 describe('round-trip', () => {
   it('preserves a full conversation through fromAISDK → toAISDK', () => {
-    const original: LanguageModelV3Prompt = [
+    const original: LanguageModelV4Prompt = [
       { role: 'system', content: 'You are a helpful assistant.' },
       { role: 'user', content: [{ type: 'text', text: 'What is 2+2?' }] },
       { role: 'assistant', content: [{ type: 'text', text: '4' }] },
@@ -477,7 +482,7 @@ describe('round-trip', () => {
   });
 
   it('preserves tool call + result through round-trip', () => {
-    const original: LanguageModelV3Prompt = [
+    const original: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'find something' }] },
       {
         role: 'assistant',
@@ -510,12 +515,12 @@ describe('round-trip', () => {
   });
 
   it('preserves file parts through round-trip', () => {
-    const original: LanguageModelV3Prompt = [
+    const original: LanguageModelV4Prompt = [
       {
         role: 'user',
         content: [
           { type: 'text', text: 'Analyze this' },
-          { type: 'file', data: 'base64data', mediaType: 'image/png' },
+          { type: 'file', data: { type: 'data', data: 'base64data' }, mediaType: 'image/png' },
         ],
       },
     ];
@@ -526,7 +531,7 @@ describe('round-trip', () => {
   });
 
   it('uses modified content when Janitor changes it (e.g. compact)', () => {
-    const original: LanguageModelV3Prompt = [
+    const original: LanguageModelV4Prompt = [
       {
         role: 'tool',
         content: [
@@ -555,7 +560,7 @@ describe('round-trip', () => {
   });
 
   it('preserves providerOptions through round-trip', () => {
-    const original: LanguageModelV3Prompt = [
+    const original: LanguageModelV4Prompt = [
       {
         role: 'user',
         content: [{ type: 'text', text: 'Hello' }],
@@ -570,7 +575,7 @@ describe('round-trip', () => {
 
   // ─── FIX #1: inline (provider-executed) tool-result must not trigger a placeholder ───
   it('does not inject a placeholder for an inline (provider-executed) tool-result', () => {
-    const original: LanguageModelV3Prompt = [
+    const original: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'search the web' }] },
       {
         role: 'assistant',
@@ -604,7 +609,7 @@ describe('round-trip', () => {
   });
 
   it('pairs a normal tool-call and skips an inline one in the same assistant message', () => {
-    const original: LanguageModelV3Prompt = [
+    const original: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'do two things' }] },
       {
         role: 'assistant',
@@ -651,7 +656,7 @@ describe('round-trip', () => {
 
   // ─── FIX #2: tool-message-level providerOptions survives round-trip ───
   it('preserves tool-message-level providerOptions through round-trip', () => {
-    const original: LanguageModelV3Prompt = [
+    const original: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'run it' }] },
       {
         role: 'assistant',
diff --git a/packages/ai-sdk-middleware/tests/compaction.test.ts b/packages/ai-sdk-middleware/tests/compaction.test.ts
deleted file mode 100644
index fd52e5b..0000000
--- a/packages/ai-sdk-middleware/tests/compaction.test.ts
+++ /dev/null
@@ -1,216 +0,0 @@
-import type {
-  LanguageModelV3,
-  LanguageModelV3CallOptions,
-  LanguageModelV3Content,
-  LanguageModelV3FinishReason,
-  LanguageModelV3GenerateResult,
-  LanguageModelV3Prompt,
-} from '@ai-sdk/provider';
-import { describe, expect, it } from 'vitest';
-import { compactHistory, planCompaction } from '../src/compaction';
-
-// These are the AI-SDK boundary functions — thin wrappers over core's
-// provider-agnostic `planCompaction` / `compactHistory`. The turn-split algorithm
-// itself is exhaustively tested in core (durableCompaction.test.ts); the tests
-// here cover what the boundary adds: fromAISDK/toAISDK round-tripping, the
-// no-op reference short-circuit, and the summary's AI SDK message shape.
-
-/** Minimal V3 model whose summarization call returns a fixed string. */
-function createSummarizerModel(summaryText = 'SUMMARY'): LanguageModelV3 {
-  return {
-    specificationVersion: 'v3',
-    provider: 'test',
-    modelId: 'test-model',
-    supportedUrls: {},
-    async doGenerate(_opts: LanguageModelV3CallOptions): Promise<LanguageModelV3GenerateResult> {
-      const content: LanguageModelV3Content[] = [{ type: 'text', text: summaryText }];
-      const finishReason: LanguageModelV3FinishReason = { unified: 'stop', raw: undefined };
-      return {
-        content,
-        finishReason,
-        warnings: [],
-        usage: {
-          inputTokens: {
-            total: 50,
-            noCache: undefined,
-            cacheRead: undefined,
-            cacheWrite: undefined,
-          },
-          outputTokens: { total: 10, text: undefined, reasoning: undefined },
-        },
-        response: { id: 'id', timestamp: new Date(), modelId: 'test-model' },
-      };
-    },
-    async doStream() {
-      throw new Error('not used');
-    },
-  };
-}
-
-/** N plain user/assistant turns, optionally with a leading system message. */
-function plainTurns(n: number, withSystem = true): LanguageModelV3Prompt {
-  const prompt: LanguageModelV3Prompt = withSystem
-    ? [{ role: 'system', content: 'You are helpful.' }]
-    : [];
-  for (let i = 0; i < n; i++) {
-    prompt.push({ role: 'user', content: [{ type: 'text', text: `q${i}` }] });
-    prompt.push({ role: 'assistant', content: [{ type: 'text', text: `a${i}` }] });
-  }
-  return prompt;
-}
-
-describe('planCompaction', () => {
-  it('delegates the split to core and round-trips each slice through the adapter', () => {
-    // 3 user/assistant pairs = 6 turns (each message is its own turn).
-    const plan = planCompaction(plainTurns(3), { keepRecentTurns: 2 });
-
-    expect(plan.system.map((m) => m.role)).toEqual(['system']);
-    // 6 turns, keep 2 → summarize the first 4 turns, keep the last 2.
-    expect(plan.toSummarize).toHaveLength(4);
-    expect(plan.toKeep).toHaveLength(2);
-    // Kept tail is the most recent turn's messages.
-    const lastUser = plan.toKeep[0];
-    expect(lastUser.role).toBe('user');
-  });
-
-  it('round-trips a prompt with no system message', () => {
-    const plan = planCompaction(plainTurns(3, false), { keepRecentTurns: 2 });
-
-    expect(plan.system).toEqual([]);
-    expect(plan.toSummarize).toHaveLength(4);
-    expect(plan.toKeep).toHaveLength(2);
-  });
-
-  it('never splits an assistant tool-call from its tool result', () => {
-    // turns: [user q1] [assistant+tool_call, tool_result] [user q2] [assistant a2]
-    const prompt: LanguageModelV3Prompt = [
-      { role: 'user', content: [{ type: 'text', text: 'q1' }] },
-      {
-        role: 'assistant',
-        content: [{ type: 'tool-call', toolCallId: 'c1', toolName: 'foo', input: { a: 1 } }],
-      },
-      {
-        role: 'tool',
-        content: [
-          {
-            type: 'tool-result',
-            toolCallId: 'c1',
-            toolName: 'foo',
-            output: { type: 'text', value: 'ok' },
-          },
-        ],
-      },
-      { role: 'user', content: [{ type: 'text', text: 'q2' }] },
-      { role: 'assistant', content: [{ type: 'text', text: 'a2' }] },
-    ];
-
-    // keepRecentTurns: 3 → split at the start of the tool turn. A naive
-    // "keep last 3 messages" cut would orphan the tool result; turn-based does not.
-    const plan = planCompaction(prompt, { keepRecentTurns: 3 });
-
-    expect(plan.toSummarize.map((m) => m.role)).toEqual(['user']); // just q1
-    // The assistant (tool-call) and its tool result stay together at the head of toKeep.
-    expect(plan.toKeep.map((m) => m.role)).toEqual(['assistant', 'tool', 'user', 'assistant']);
-  });
-
-  it('preserves a multi-block assistant message verbatim in toKeep', () => {
-    const assistantContent = [
-      { type: 'text' as const, text: 'let me check' },
-      { type: 'reasoning' as const, text: 'because reasons' },
-      { type: 'tool-call' as const, toolCallId: 'c1', toolName: 'foo', input: { a: 1 } },
-    ];
-    const prompt: LanguageModelV3Prompt = [
-      { role: 'user', content: [{ type: 'text', text: 'q' }] },
-      { role: 'assistant', content: assistantContent },
-      {
-        role: 'tool',
-        content: [
-          {
-            type: 'tool-result',
-            toolCallId: 'c1',
-            toolName: 'foo',
-            output: { type: 'text', value: 'ok' },
-          },
-        ],
-      },
-    ];
-
-    const plan = planCompaction(prompt, { keepRecentTurns: 99 });
-    const keptAssistant = plan.toKeep.find((m) => m.role === 'assistant');
-    expect(keptAssistant?.content).toEqual(assistantContent);
-  });
-});
-
-describe('compactHistory', () => {
-  it('returns [...system, summary, ...toKeep] with the summary as a wrapped user message', async () => {
-    const prompt = plainTurns(4); // system + 8 messages (8 turns)
-    const result = await compactHistory(prompt, createSummarizerModel('Hello'), {
-      keepRecentTurns: 2,
-    });
-
-    expect(result[0]).toEqual(prompt[0]); // system preserved at front
-    const summaryMsg = result[1];
-    expect(summaryMsg.role).toBe('user');
-    const text =
-      summaryMsg.role === 'user' && summaryMsg.content[0].type === 'text'
-        ? summaryMsg.content[0].text
-        : '';
-    expect(text).toContain('Hello'); // summary survived
-    expect(text).toContain('continued from a previous conversation'); // wrapper framing
-    // Compacted: system + summary + last 2 turns, shorter than the original.
-    expect(result.length).toBeLessThan(prompt.length);
-    expect(result.length).toBe(1 /* system */ + 1 /* summary */ + 2 /* kept turns */);
-  });
-
-  it('returns the prompt unchanged when nothing is old enough to compact', async () => {
-    const prompt = plainTurns(2);
-    const result = await compactHistory(prompt, createSummarizerModel(), { keepRecentTurns: 99 });
-    expect(result).toBe(prompt); // same reference — untouched, no model call
-  });
-
-  it('returns the prompt unchanged when the summarizer yields no text', async () => {
-    const prompt = plainTurns(4);
-    const result = await compactHistory(prompt, createSummarizerModel('   '), {
-      keepRecentTurns: 1,
-    });
-    expect(result).toBe(prompt);
-  });
-
-  it('produces a valid prompt when the kept tail starts with a tool turn', async () => {
-    // turns: [user q1] [assistant+tool_call, tool_result] [user q2] [assistant a2]
-    const prompt: LanguageModelV3Prompt = [
-      { role: 'system', content: 'sys' },
-      { role: 'user', content: [{ type: 'text', text: 'q1' }] },
-      {
-        role: 'assistant',
-        content: [{ type: 'tool-call', toolCallId: 'c1', toolName: 'foo', input: { a: 1 } }],
-      },
-      {
-        role: 'tool',
-        content: [
-          {
-            type: 'tool-result',
-            toolCallId: 'c1',
-            toolName: 'foo',
-            output: { type: 'text', value: 'ok' },
-          },
-        ],
-      },
-      { role: 'user', content: [{ type: 'text', text: 'q2' }] },
-      { role: 'assistant', content: [{ type: 'text', text: 'a2' }] },
-    ];
-
-    // 4 conversation turns; keep 3 → summarize just [user q1], keep from the tool
-    // turn onward. The summary user message must be followed by the assistant
-    // tool-call and its result, in order — a valid, non-orphaned prompt.
-    const result = await compactHistory(prompt, createSummarizerModel('S'), { keepRecentTurns: 3 });
-    expect(result.map((m) => m.role)).toEqual([
-      'system',
-      'user',
-      'assistant',
-      'tool',
-      'user',
-      'assistant',
-    ]);
-  });
-});
diff --git a/packages/ai-sdk-middleware/tests/compactionModelMessages.test.ts b/packages/ai-sdk-middleware/tests/compactionModelMessages.test.ts
index 441eb51..ab9f394 100644
--- a/packages/ai-sdk-middleware/tests/compactionModelMessages.test.ts
+++ b/packages/ai-sdk-middleware/tests/compactionModelMessages.test.ts
@@ -1,9 +1,9 @@
 import type {
-  LanguageModelV3,
-  LanguageModelV3CallOptions,
-  LanguageModelV3Content,
-  LanguageModelV3FinishReason,
-  LanguageModelV3GenerateResult,
+  LanguageModelV4,
+  LanguageModelV4CallOptions,
+  LanguageModelV4Content,
+  LanguageModelV4FinishReason,
+  LanguageModelV4GenerateResult,
 } from '@ai-sdk/provider';
 import type { ModelMessage } from 'ai';
 import { describe, expect, it } from 'vitest';
@@ -11,15 +11,15 @@ import { compactModelMessages, planCompactionModelMessages } from '../src/compac
 
 /** Minimal V3 model whose summarization call returns a fixed string. A V3 model
  *  is a valid `LanguageModel`, so it exercises the widened model param too. */
-function createSummarizerModel(summaryText = 'SUMMARY'): LanguageModelV3 {
+function createSummarizerModel(summaryText = 'SUMMARY'): LanguageModelV4 {
   return {
-    specificationVersion: 'v3',
+    specificationVersion: 'v4',
     provider: 'test',
     modelId: 'test-model',
     supportedUrls: {},
-    async doGenerate(_opts: LanguageModelV3CallOptions): Promise<LanguageModelV3GenerateResult> {
-      const content: LanguageModelV3Content[] = [{ type: 'text', text: summaryText }];
-      const finishReason: LanguageModelV3FinishReason = { unified: 'stop', raw: undefined };
+    async doGenerate(_opts: LanguageModelV4CallOptions): Promise<LanguageModelV4GenerateResult> {
+      const content: LanguageModelV4Content[] = [{ type: 'text', text: summaryText }];
+      const finishReason: LanguageModelV4FinishReason = { unified: 'stop', raw: undefined };
       return {
         content,
         finishReason,
diff --git a/packages/ai-sdk-middleware/tests/middleware.test.ts b/packages/ai-sdk-middleware/tests/middleware.test.ts
index b65d8a3..05d69fa 100644
--- a/packages/ai-sdk-middleware/tests/middleware.test.ts
+++ b/packages/ai-sdk-middleware/tests/middleware.test.ts
@@ -1,31 +1,31 @@
 import type {
-  LanguageModelV3,
-  LanguageModelV3CallOptions,
-  LanguageModelV3Content,
-  LanguageModelV3FinishReason,
-  LanguageModelV3GenerateResult,
-  LanguageModelV3Prompt,
-  LanguageModelV3StreamPart,
-  LanguageModelV3StreamResult,
+  LanguageModelV4,
+  LanguageModelV4CallOptions,
+  LanguageModelV4Content,
+  LanguageModelV4FinishReason,
+  LanguageModelV4GenerateResult,
+  LanguageModelV4Prompt,
+  LanguageModelV4StreamPart,
+  LanguageModelV4StreamResult,
 } from '@ai-sdk/provider';
 import type { Skill } from '@context-chef/core';
 import { describe, expect, it, vi } from 'vitest';
 import { withContextChef } from '../src/index';
 import { createMiddleware } from '../src/middleware';
 
-function createMockModel(options?: { inputTokens?: number; outputText?: string }): LanguageModelV3 {
+function createMockModel(options?: { inputTokens?: number; outputText?: string }): LanguageModelV4 {
   const inputTokens = options?.inputTokens ?? 100;
   const outputText = options?.outputText ?? 'Hello';
 
-  const model: LanguageModelV3 = {
-    specificationVersion: 'v3',
+  const model: LanguageModelV4 = {
+    specificationVersion: 'v4',
     provider: 'test',
     modelId: 'test-model',
     supportedUrls: {},
 
-    async doGenerate(_opts: LanguageModelV3CallOptions): Promise<LanguageModelV3GenerateResult> {
-      const content: LanguageModelV3Content[] = [{ type: 'text', text: outputText }];
-      const finishReason: LanguageModelV3FinishReason = { unified: 'stop', raw: undefined };
+    async doGenerate(_opts: LanguageModelV4CallOptions): Promise<LanguageModelV4GenerateResult> {
+      const content: LanguageModelV4Content[] = [{ type: 'text', text: outputText }];
+      const finishReason: LanguageModelV4FinishReason = { unified: 'stop', raw: undefined };
       return {
         content,
         finishReason,
@@ -47,8 +47,8 @@ function createMockModel(options?: { inputTokens?: number; outputText?: string }
       };
     },
 
-    async doStream(_opts: LanguageModelV3CallOptions) {
-      const parts: LanguageModelV3StreamPart[] = [
+    async doStream(_opts: LanguageModelV4CallOptions) {
+      const parts: LanguageModelV4StreamPart[] = [
         { type: 'text-start', id: '1' },
         { type: 'text-delta', id: '1', delta: outputText },
         { type: 'text-end', id: '1' },
@@ -67,7 +67,7 @@ function createMockModel(options?: { inputTokens?: number; outputText?: string }
         },
       ];
 
-      const stream = new ReadableStream<LanguageModelV3StreamPart>({
+      const stream = new ReadableStream<LanguageModelV4StreamPart>({
         start(controller) {
           for (const part of parts) {
             controller.enqueue(part);
@@ -82,8 +82,8 @@ function createMockModel(options?: { inputTokens?: number; outputText?: string }
   return model;
 }
 
-function makeConversation(messageCount: number): LanguageModelV3Prompt {
-  const prompt: LanguageModelV3Prompt = [{ role: 'system', content: 'You are helpful.' }];
+function makeConversation(messageCount: number): LanguageModelV4Prompt {
+  const prompt: LanguageModelV4Prompt = [{ role: 'system', content: 'You are helpful.' }];
   for (let i = 0; i < messageCount; i++) {
     prompt.push({
       role: 'user',
@@ -109,11 +109,11 @@ describe('createMiddleware', () => {
       contextWindow: 1_000_000,
     });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'Hello' }] },
     ];
 
-    const params: LanguageModelV3CallOptions = { prompt };
+    const params: LanguageModelV4CallOptions = { prompt };
     const result = await assertDefined(
       middleware.transformParams,
       'transformParams',
@@ -133,7 +133,7 @@ describe('createMiddleware', () => {
     });
 
     const longOutput = 'x'.repeat(200);
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'Run command' }] },
       {
         role: 'assistant',
@@ -159,7 +159,7 @@ describe('createMiddleware', () => {
       },
     ];
 
-    const params: LanguageModelV3CallOptions = { prompt };
+    const params: LanguageModelV4CallOptions = { prompt };
     const result = await assertDefined(
       middleware.transformParams,
       'transformParams',
@@ -189,9 +189,9 @@ describe('createMiddleware', () => {
 
     const model = createMockModel({ inputTokens: 200 });
 
-    const doGenerate = (): PromiseLike<LanguageModelV3GenerateResult> =>
+    const doGenerate = (): PromiseLike<LanguageModelV4GenerateResult> =>
       model.doGenerate({ prompt: [] });
-    const doStream = (): PromiseLike<LanguageModelV3StreamResult> => model.doStream({ prompt: [] });
+    const doStream = (): PromiseLike<LanguageModelV4StreamResult> => model.doStream({ prompt: [] });
 
     const result = await assertDefined(
       middleware.wrapGenerate,
@@ -215,9 +215,9 @@ describe('createMiddleware', () => {
 
     const model = createMockModel({ inputTokens: 300 });
 
-    const doGenerate = (): PromiseLike<LanguageModelV3GenerateResult> =>
+    const doGenerate = (): PromiseLike<LanguageModelV4GenerateResult> =>
       model.doGenerate({ prompt: [] });
-    const doStream = (): PromiseLike<LanguageModelV3StreamResult> => model.doStream({ prompt: [] });
+    const doStream = (): PromiseLike<LanguageModelV4StreamResult> => model.doStream({ prompt: [] });
 
     const streamResult = await assertDefined(
       middleware.wrapStream,
@@ -230,7 +230,7 @@ describe('createMiddleware', () => {
     });
 
     const reader = streamResult.stream.getReader();
-    const chunks: LanguageModelV3StreamPart[] = [];
+    const chunks: LanguageModelV4StreamPart[] = [];
     while (true) {
       const { done, value } = await reader.read();
       if (done) break;
@@ -253,9 +253,9 @@ describe('createMiddleware', () => {
 
     const model = createMockModel({ inputTokens: 200 });
 
-    const doGenerate = (): PromiseLike<LanguageModelV3GenerateResult> =>
+    const doGenerate = (): PromiseLike<LanguageModelV4GenerateResult> =>
       model.doGenerate({ prompt: [] });
-    const doStream = (): PromiseLike<LanguageModelV3StreamResult> => model.doStream({ prompt: [] });
+    const doStream = (): PromiseLike<LanguageModelV4StreamResult> => model.doStream({ prompt: [] });
 
     await assertDefined(
       middleware.wrapGenerate,
@@ -289,9 +289,9 @@ describe('createMiddleware', () => {
 
     const model = createMockModel({ inputTokens: 200 });
 
-    const doGenerate = (): PromiseLike<LanguageModelV3GenerateResult> =>
+    const doGenerate = (): PromiseLike<LanguageModelV4GenerateResult> =>
       model.doGenerate({ prompt: [] });
-    const doStream = (): PromiseLike<LanguageModelV3StreamResult> => model.doStream({ prompt: [] });
+    const doStream = (): PromiseLike<LanguageModelV4StreamResult> => model.doStream({ prompt: [] });
 
     // Feed token usage over the budget so compression fires on next transformParams
     await assertDefined(
@@ -333,7 +333,7 @@ describe('compress opt-in (no budgeting configured)', () => {
         truncate: { threshold: 50, tailChars: 10 },
       });
 
-      const prompt: LanguageModelV3Prompt = [
+      const prompt: LanguageModelV4Prompt = [
         { role: 'user', content: [{ type: 'text', text: 'Hello' }] },
       ];
       const result = await assertDefined(
@@ -356,9 +356,9 @@ describe('compress opt-in (no budgeting configured)', () => {
     const middleware = createMiddleware({});
     const model = createMockModel({ inputTokens: 42 });
 
-    const doGenerate = (): PromiseLike<LanguageModelV3GenerateResult> =>
+    const doGenerate = (): PromiseLike<LanguageModelV4GenerateResult> =>
       model.doGenerate({ prompt: [] });
-    const doStream = (): PromiseLike<LanguageModelV3StreamResult> => model.doStream({ prompt: [] });
+    const doStream = (): PromiseLike<LanguageModelV4StreamResult> => model.doStream({ prompt: [] });
 
     const result = await assertDefined(
       middleware.wrapGenerate,
@@ -395,11 +395,11 @@ describe('compress opt-in (no budgeting configured)', () => {
 });
 
 describe('withContextChef wrapper', () => {
-  it('returns a LanguageModelV3', () => {
+  it('returns a LanguageModelV4', () => {
     const model = createMockModel();
     const wrapped = withContextChef(model, { contextWindow: 128_000 });
 
-    expect(wrapped.specificationVersion).toBe('v3');
+    expect(wrapped.specificationVersion).toBe('v4');
     expect(wrapped.provider).toBeDefined();
     expect(wrapped.modelId).toBeDefined();
     expect(typeof wrapped.doGenerate).toBe('function');
@@ -417,7 +417,7 @@ describe('withContextChef wrapper', () => {
       prompt: [{ role: 'user', content: [{ type: 'text', text: 'Hi' }] }],
     });
 
-    const textContent = result.content.find((c: LanguageModelV3Content) => c.type === 'text');
+    const textContent = result.content.find((c: LanguageModelV4Content) => c.type === 'text');
     expect(textContent).toBeDefined();
     if (textContent?.type === 'text') {
       expect(textContent.text).toBe('Hello from wrapped model');
@@ -432,7 +432,7 @@ describe('compact', () => {
       compact: { toolCalls: 'all' },
     });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'Run command' }] },
       {
         role: 'assistant',
@@ -488,7 +488,7 @@ describe('compact', () => {
       compact: { reasoning: 'all' },
     });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       {
         role: 'assistant',
         content: [
@@ -516,20 +516,20 @@ describe('compact', () => {
   });
 });
 
-describe('onBudgetExceeded', () => {
+describe('onBeforeCompress', () => {
   it('calls hook when budget is exceeded', async () => {
-    const onBudgetExceeded = vi.fn().mockReturnValue(null);
+    const onBeforeCompress = vi.fn().mockReturnValue(null);
     const middleware = createMiddleware({
       contextWindow: 100,
-      onBudgetExceeded,
+      onBeforeCompress,
     });
 
     const model = createMockModel({ inputTokens: 200 });
 
     // Feed high token usage to trigger budget exceeded
-    const doGenerate = (): PromiseLike<LanguageModelV3GenerateResult> =>
+    const doGenerate = (): PromiseLike<LanguageModelV4GenerateResult> =>
       model.doGenerate({ prompt: [] });
-    const doStream = (): PromiseLike<LanguageModelV3StreamResult> => model.doStream({ prompt: [] });
+    const doStream = (): PromiseLike<LanguageModelV4StreamResult> => model.doStream({ prompt: [] });
     await assertDefined(
       middleware.wrapGenerate,
       'wrapGenerate',
@@ -550,7 +550,7 @@ describe('onBudgetExceeded', () => {
       model,
     });
 
-    expect(onBudgetExceeded).toHaveBeenCalled();
+    expect(onBeforeCompress).toHaveBeenCalled();
   });
 });
 
@@ -564,7 +564,7 @@ describe('dynamicState', () => {
       },
     });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'What next?' }] },
     ];
 
@@ -599,7 +599,7 @@ describe('dynamicState', () => {
       },
     });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'Hi' }] },
     ];
 
@@ -628,7 +628,7 @@ describe('dynamicState', () => {
       },
     });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'Hello' }] },
     ];
 
@@ -664,7 +664,7 @@ describe('dynamicState', () => {
       },
     });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'Hi' }] },
     ];
 
@@ -698,7 +698,7 @@ describe('dynamicState', () => {
       },
     });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'Hi' }] },
     ];
 
@@ -727,7 +727,7 @@ describe('transformContext', () => {
       },
     });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'Hello' }] },
     ];
 
@@ -754,7 +754,7 @@ describe('transformContext', () => {
       },
     });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'Hello' }] },
     ];
 
@@ -775,7 +775,7 @@ describe('transformContext', () => {
   });
 
   it('runs after dynamicState injection', async () => {
-    const transformContext = vi.fn((prompt: LanguageModelV3Prompt) => prompt);
+    const transformContext = vi.fn((prompt: LanguageModelV4Prompt) => prompt);
     const middleware = createMiddleware({
       contextWindow: 1_000_000,
       dynamicState: {
@@ -785,7 +785,7 @@ describe('transformContext', () => {
       transformContext,
     });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'Hi' }] },
     ];
 
@@ -801,9 +801,9 @@ describe('transformContext', () => {
     expect(transformContext).toHaveBeenCalledTimes(1);
     const received = transformContext.mock.calls[0][0];
     // transformContext should see the dynamic state system message
-    const systemMsgs = received.filter((m: LanguageModelV3Prompt[number]) => m.role === 'system');
+    const systemMsgs = received.filter((m: LanguageModelV4Prompt[number]) => m.role === 'system');
     const hasState = systemMsgs.some(
-      (m: LanguageModelV3Prompt[number]) =>
+      (m: LanguageModelV4Prompt[number]) =>
         m.role === 'system' && m.content.includes('<injected>true</injected>'),
     );
     expect(hasState).toBe(true);
@@ -823,7 +823,7 @@ describe('skill', () => {
       skill: planningSkill,
     });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'system', content: 'You are helpful.' },
       { role: 'user', content: [{ type: 'text', text: 'Hi' }] },
       { role: 'assistant', content: [{ type: 'text', text: 'Hello!' }] },
@@ -865,7 +865,7 @@ describe('skill', () => {
       skill: skillFn,
     });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'Hi' }] },
     ];
 
@@ -888,7 +888,7 @@ describe('skill', () => {
       skill: () => null,
     });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'system', content: 'You are helpful.' },
       { role: 'user', content: [{ type: 'text', text: 'Hi' }] },
     ];
@@ -909,7 +909,7 @@ describe('skill', () => {
       skill: () => active,
     });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'system', content: 'You are helpful.' },
       { role: 'user', content: [{ type: 'text', text: 'Hi' }] },
     ];
@@ -934,7 +934,7 @@ describe('skill', () => {
       skill: async () => planningSkill,
     });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'user', content: [{ type: 'text', text: 'Hi' }] },
     ];
 
@@ -955,7 +955,7 @@ describe('skill', () => {
       contextWindow: 1_000_000,
     });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'system', content: 'You are helpful.' },
       { role: 'user', content: [{ type: 'text', text: 'Hi' }] },
     ];
@@ -974,7 +974,7 @@ describe('skill', () => {
       skill: { name: 'noop', description: 'noop', instructions: '' },
     });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'system', content: 'You are helpful.' },
       { role: 'user', content: [{ type: 'text', text: 'Hi' }] },
     ];
@@ -994,7 +994,7 @@ describe('skill', () => {
       skill: { name: 'blank', description: 'blank', instructions: '   \n\t  ' },
     });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'system', content: 'You are helpful.' },
       { role: 'user', content: [{ type: 'text', text: 'Hi' }] },
     ];
@@ -1017,7 +1017,7 @@ describe('skill', () => {
       },
     });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'system', content: 'You are helpful.' },
       { role: 'user', content: [{ type: 'text', text: 'Hi' }] },
     ];
@@ -1049,7 +1049,7 @@ describe('skill', () => {
       clear: [{ target: 'tool-result', keepRecent: 1 }],
     });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'system', content: 'You are helpful.' },
       { role: 'user', content: [{ type: 'text', text: 'Run both commands.' }] },
       {
@@ -1108,9 +1108,9 @@ describe('skill', () => {
     });
 
     const model = createMockModel({ inputTokens: 200 });
-    const doGenerate = (): PromiseLike<LanguageModelV3GenerateResult> =>
+    const doGenerate = (): PromiseLike<LanguageModelV4GenerateResult> =>
       model.doGenerate({ prompt: [] });
-    const doStream = (): PromiseLike<LanguageModelV3StreamResult> => model.doStream({ prompt: [] });
+    const doStream = (): PromiseLike<LanguageModelV4StreamResult> => model.doStream({ prompt: [] });
 
     // Push token usage so the next transform triggers compression.
     await assertDefined(
@@ -1135,7 +1135,7 @@ describe('skill', () => {
 
 /** Build a prompt with two tool-call/result exchanges plus a trailing user message.
  * Starts with a user message so ensureValidHistory doesn't insert a placeholder. */
-function makeToolPrompt(): LanguageModelV3Prompt {
+function makeToolPrompt(): LanguageModelV4Prompt {
   return [
     { role: 'system', content: 'You are helpful.' },
     { role: 'user', content: [{ type: 'text', text: 'Run both commands.' }] },
@@ -1229,7 +1229,7 @@ describe('clear', () => {
   it('clear without tool-result targets does not inject the explainer', async () => {
     const middleware = createMiddleware({ clear: ['thinking'] });
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'system', content: 'You are helpful.' },
       {
         role: 'assistant',
@@ -1258,7 +1258,7 @@ describe('clear', () => {
     expect(logger.warn).toHaveBeenCalledTimes(1);
     expect(logger.warn.mock.calls[0][0]).toContain("clear: ['thinking']");
 
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       {
         role: 'assistant',
         content: [
@@ -1333,8 +1333,8 @@ describe('compress persistence warning', () => {
   // write-back) — the scenario the warning is meant to surface.
   async function runCycle(
     middleware: ReturnType<typeof createMiddleware>,
-    model: LanguageModelV3,
-    prompt: LanguageModelV3Prompt,
+    model: LanguageModelV4,
+    prompt: LanguageModelV4Prompt,
   ): Promise<void> {
     await assertDefined(
       middleware.wrapGenerate,
diff --git a/packages/ai-sdk-middleware/tests/modelMessageAdapter.test.ts b/packages/ai-sdk-middleware/tests/modelMessageAdapter.test.ts
index f1fbbd4..af83f77 100644
--- a/packages/ai-sdk-middleware/tests/modelMessageAdapter.test.ts
+++ b/packages/ai-sdk-middleware/tests/modelMessageAdapter.test.ts
@@ -29,6 +29,23 @@ describe('fromModelMessages', () => {
     ]);
   });
 
+  it('records V4 tagged file data ({ type: "data" }) as the attachment signal', () => {
+    const messages: ModelMessage[] = [
+      {
+        role: 'user',
+        content: [
+          { type: 'text', text: 'tagged' },
+          { type: 'file', data: { type: 'data', data: 'base64str' }, mediaType: 'image/png' },
+        ],
+      },
+    ];
+    const ir = fromModelMessages(messages);
+    // Tagged inline string data yields the same presence signal as the bare shorthand.
+    expect(ir[0].attachments).toEqual([{ mediaType: 'image/png', data: 'base64str' }]);
+    // The real (tagged) payload round-trips verbatim via the pass-through field.
+    expect(toModelMessages(ir)).toEqual(messages);
+  });
+
   it('extracts assistant tool calls and reasoning', () => {
     const messages: ModelMessage[] = [
       { role: 'user', content: 'use a tool' },
diff --git a/packages/ai-sdk-middleware/tests/summarizeMessages.test.ts b/packages/ai-sdk-middleware/tests/summarizeMessages.test.ts
index e1b56f6..80b9502 100644
--- a/packages/ai-sdk-middleware/tests/summarizeMessages.test.ts
+++ b/packages/ai-sdk-middleware/tests/summarizeMessages.test.ts
@@ -1,11 +1,11 @@
 import type {
-  LanguageModelV3,
-  LanguageModelV3CallOptions,
-  LanguageModelV3Content,
-  LanguageModelV3FinishReason,
-  LanguageModelV3GenerateResult,
-  LanguageModelV3Prompt,
-  LanguageModelV3StreamPart,
+  LanguageModelV4,
+  LanguageModelV4CallOptions,
+  LanguageModelV4Content,
+  LanguageModelV4FinishReason,
+  LanguageModelV4GenerateResult,
+  LanguageModelV4Prompt,
+  LanguageModelV4StreamPart,
 } from '@ai-sdk/provider';
 import type { ModelMessage } from 'ai';
 import { describe, expect, it, vi } from 'vitest';
@@ -15,15 +15,15 @@ import { summarizeModelMessages } from '../src/middleware';
 
 /** Minimal V3 model whose summarization call returns a fixed string. A V3 model
  *  is a valid `LanguageModel`, so it exercises the widened model param too. */
-function createSummarizerModel(summaryText = 'SUMMARY'): LanguageModelV3 {
+function createSummarizerModel(summaryText = 'SUMMARY'): LanguageModelV4 {
   return {
-    specificationVersion: 'v3',
+    specificationVersion: 'v4',
     provider: 'test',
     modelId: 'test-model',
     supportedUrls: {},
-    async doGenerate(_opts: LanguageModelV3CallOptions): Promise<LanguageModelV3GenerateResult> {
-      const content: LanguageModelV3Content[] = [{ type: 'text', text: summaryText }];
-      const finishReason: LanguageModelV3FinishReason = { unified: 'stop', raw: undefined };
+    async doGenerate(_opts: LanguageModelV4CallOptions): Promise<LanguageModelV4GenerateResult> {
+      const content: LanguageModelV4Content[] = [{ type: 'text', text: summaryText }];
+      const finishReason: LanguageModelV4FinishReason = { unified: 'stop', raw: undefined };
       return {
         content,
         finishReason,
@@ -46,17 +46,17 @@ function createSummarizerModel(summaryText = 'SUMMARY'): LanguageModelV3 {
   };
 }
 
-function mockModel(text: string): LanguageModelV3 {
-  const model: LanguageModelV3 = {
-    specificationVersion: 'v3',
+function mockModel(text: string): LanguageModelV4 {
+  const model: LanguageModelV4 = {
+    specificationVersion: 'v4',
     provider: 'test',
     modelId: 'test-model',
     supportedUrls: {},
 
     doGenerate: vi.fn(
-      async (_opts: LanguageModelV3CallOptions): Promise<LanguageModelV3GenerateResult> => {
-        const content: LanguageModelV3Content[] = [{ type: 'text', text }];
-        const finishReason: LanguageModelV3FinishReason = { unified: 'stop', raw: undefined };
+      async (_opts: LanguageModelV4CallOptions): Promise<LanguageModelV4GenerateResult> => {
+        const content: LanguageModelV4Content[] = [{ type: 'text', text }];
+        const finishReason: LanguageModelV4FinishReason = { unified: 'stop', raw: undefined };
         return {
           content,
           finishReason,
@@ -79,8 +79,8 @@ function mockModel(text: string): LanguageModelV3 {
       },
     ),
 
-    async doStream(_opts: LanguageModelV3CallOptions) {
-      const parts: LanguageModelV3StreamPart[] = [
+    async doStream(_opts: LanguageModelV4CallOptions) {
+      const parts: LanguageModelV4StreamPart[] = [
         { type: 'text-start', id: '1' },
         { type: 'text-delta', id: '1', delta: text },
         { type: 'text-end', id: '1' },
@@ -99,7 +99,7 @@ function mockModel(text: string): LanguageModelV3 {
         },
       ];
 
-      const stream = new ReadableStream<LanguageModelV3StreamPart>({
+      const stream = new ReadableStream<LanguageModelV4StreamPart>({
         start(controller) {
           for (const part of parts) {
             controller.enqueue(part);
@@ -148,7 +148,7 @@ describe('summarizeMessages', () => {
     //   - tool message becomes a user message with "[Tool result(call_1): RAW_TOOL_OUTPUT]"
     //   - assistant tool-call becomes "[Called tool: search({"q":"x"})]" text
     //   - no raw role:'tool' message should reach the model
-    const toolPrompt: LanguageModelV3Prompt = [
+    const toolPrompt: LanguageModelV4Prompt = [
       { role: 'system', content: 'SYSTEM_MARKER_DROP_ME' },
       { role: 'user', content: [{ type: 'text', text: 'do the thing' }] },
       {
diff --git a/packages/ai-sdk-middleware/tests/truncator.test.ts b/packages/ai-sdk-middleware/tests/truncator.test.ts
index 82dc04c..150d50b 100644
--- a/packages/ai-sdk-middleware/tests/truncator.test.ts
+++ b/packages/ai-sdk-middleware/tests/truncator.test.ts
@@ -1,10 +1,10 @@
-import type { LanguageModelV3Prompt } from '@ai-sdk/provider';
+import type { LanguageModelV4Prompt } from '@ai-sdk/provider';
 import type { VFSStorageAdapter } from '@context-chef/core';
 import { describe, expect, it, vi } from 'vitest';
 import { truncateToolResults } from '../src/truncator';
 
 describe('truncateToolResults', () => {
-  const makeToolPrompt = (output: string): LanguageModelV3Prompt => [
+  const makeToolPrompt = (output: string): LanguageModelV4Prompt => [
     {
       role: 'tool',
       content: [
@@ -54,7 +54,7 @@ describe('truncateToolResults', () => {
   });
 
   it('does not affect non-tool messages', async () => {
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       { role: 'system', content: 'x'.repeat(200) },
       { role: 'user', content: [{ type: 'text', text: 'x'.repeat(200) }] },
     ];
@@ -64,7 +64,7 @@ describe('truncateToolResults', () => {
 
   it('handles json tool output', async () => {
     const bigJson = JSON.stringify({ data: 'x'.repeat(500) });
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       {
         role: 'tool',
         content: [
@@ -252,7 +252,7 @@ describe('truncateToolResults', () => {
   it('filters per-part: preserves one tool while truncating another in the same message', async () => {
     const keepOutput = 'k'.repeat(500);
     const truncOutput = 't'.repeat(500);
-    const prompt: LanguageModelV3Prompt = [
+    const prompt: LanguageModelV4Prompt = [
       {
         role: 'tool',
         content: [
diff --git a/packages/core/README.md b/packages/core/README.md
index 7b8572f..7ab3ef7 100644
--- a/packages/core/README.md
+++ b/packages/core/README.md
@@ -340,7 +340,7 @@ Contracts: `history` is a flat `Message[]` with any **system messages inline** (
 
 `keepRecentTurns: 0` is **full compaction** (Claude Code style): the whole conversation collapses into `[...system, <summary>]` with no verbatim tail. This is the simplest result to persist — there is no kept tail to reconcile against your store's unit boundaries. The trade-off is that no verbatim recent context survives, so steer the summary toward a structured handoff via `customCompressionInstructions`. Use a small `keepRecentTurns` instead when the in-flight turn should stay verbatim.
 
-> **ai-sdk users:** prefer `compactHistory` / `planCompaction` from [`@context-chef/ai-sdk-middleware`](https://www.npmjs.com/package/@context-chef/ai-sdk-middleware) — they take an `AI SDK` prompt and a model directly, wiring the adapter and role-flattening for you.
+> **ai-sdk users:** prefer `compactModelMessages` / `planCompactionModelMessages` from [`@context-chef/ai-sdk-middleware`](https://www.npmjs.com/package/@context-chef/ai-sdk-middleware) — they take an `AI SDK` prompt and a model directly, wiring the adapter and role-flattening for you.
 
 #### Compression circuit breaker
 
diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml
index 3a3ba14..e4c0ebd 100644
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -25,14 +25,14 @@ importers:
         version: link:../core
     devDependencies:
       '@ai-sdk/provider':
-        specifier: ^3.0.8
-        version: 3.0.8
+        specifier: ^4.0.0
+        version: 4.0.0
       '@types/node':
         specifier: ^25.3.0
         version: 25.3.0
       ai:
-        specifier: ^6.0.140
-        version: 6.0.140(zod@4.3.6)
+        specifier: ^7.0.0
+        version: 7.0.0(zod@4.3.6)
       tsdown:
         specifier: ^0.20.3
         version: 0.20.3(synckit@0.11.12)(typescript@5.9.3)
@@ -98,21 +98,21 @@ importers:
 
 packages:
 
-  '@ai-sdk/gateway@3.0.82':
-    resolution: {integrity: sha512-ddB9FrkHZank1zyx13vypU0RrPjsXWj3NvlsrJ4yFQnrpR+xh48W4wO9ijndUueBsaADjlvMz2Ghv8yq5tZGCQ==}
-    engines: {node: '>=18'}
+  '@ai-sdk/gateway@4.0.0':
+    resolution: {integrity: sha512-rcKukspbM4h511ot2E8TsPl7rXjRK1zHKrMCP7w4+XF55UKqQHaDzo2kKbGv5rp8Bjb1yQatIHJZE1E2yrOOMw==}
+    engines: {node: '>=22'}
     peerDependencies:
       zod: ^3.25.76 || ^4.1.8
 
-  '@ai-sdk/provider-utils@4.0.21':
-    resolution: {integrity: sha512-MtFUYI1/8mgDvRmaBDjbLJPFFrMG777AvSgyIFQtZHIMzm88R/12vYBBpnk7pfiWLFE1DSZzY4WDYzGbKAcmiw==}
-    engines: {node: '>=18'}
+  '@ai-sdk/provider-utils@5.0.0':
+    resolution: {integrity: sha512-zj66M02jc6ASYwIgWZowsooDUwaVngeNZQ3H10GwcPMZ+KR6gHMhcUuKl6tkai+JPXTKDyHY1pnszuxRtw2D4A==}
+    engines: {node: '>=22'}
     peerDependencies:
       zod: ^3.25.76 || ^4.1.8
 
-  '@ai-sdk/provider@3.0.8':
-    resolution: {integrity: sha512-oGMAgGoQdBXbZqNG0Ze56CHjDZ1IDYOwGYxYjO5KLSlz5HiNQ9udIXsPZ61VWaHGZ5XW/jyjmr6t2xz2jGVwbQ==}
-    engines: {node: '>=18'}
+  '@ai-sdk/provider@4.0.0':
+    resolution: {integrity: sha512-fr9Gs89prDWiuox/T+kCA+i2cJkHpxU5S+tr4megjTzRC27ZsvFhwjU/+XrqqMbvBUlfmXxTOYWy8ng45dsjIg==}
+    engines: {node: '>=22'}
 
   '@anthropic-ai/sdk@0.78.0':
     resolution: {integrity: sha512-PzQhR715td/m1UaaN5hHXjYB8Gl2lF9UVhrrGrZeysiF6Rb74Wc9GCB8hzLdzmQtBd1qe89F9OptgB9Za1Ib5w==}
@@ -170,28 +170,24 @@ packages:
     engines: {node: '>=14.21.3'}
     cpu: [arm64]
     os: [linux]
-    libc: [musl]
 
   '@biomejs/cli-linux-arm64@2.4.15':
     resolution: {integrity: sha512-owaAMZD/T4LrD0ELNCk0Km3qrRHuM0X6EAyVE1FSqGY0rbLoiDLrO4Us2tllm6cAeB2Ioa9C2C08NZPdr8+0Ug==}
     engines: {node: '>=14.21.3'}
     cpu: [arm64]
     os: [linux]
-    libc: [glibc]
 
   '@biomejs/cli-linux-x64-musl@2.4.15':
     resolution: {integrity: sha512-CNq/9W38SYSH023lfcQ4KKU8K0YX8T//FZUhcgtMMRABDojx5XsMV7jlweAvGSl389wJQB29Qo6Zb/a+jdvt+w==}
     engines: {node: '>=14.21.3'}
     cpu: [x64]
     os: [linux]
-    libc: [musl]
 
   '@biomejs/cli-linux-x64@2.4.15':
     resolution: {integrity: sha512-0jj7THz12GbUOLmMibktK6DZjqz2zV64KFxyBtcFTKPiiOIY0a7vns1elpO1dERvxpsZ5ik0oFfz0oGwFde1+g==}
     engines: {node: '>=14.21.3'}
     cpu: [x64]
     os: [linux]
-    libc: [glibc]
 
   '@biomejs/cli-win32-arm64@2.4.15':
     resolution: {integrity: sha512-ouhkYdlhp/1GghEJPdWwD/Vi3gQ1nFxuSpMolWsbq3Lsq3QUR4jl6UdhhscdCugKU5vOEuMiJhvKj66O0OCq+w==}
@@ -534,28 +530,24 @@ packages:
     engines: {node: ^20.19.0 || >=22.12.0}
     cpu: [arm64]
     os: [linux]
-    libc: [glibc]
 
   '@rolldown/binding-linux-arm64-musl@1.0.0-rc.3':
     resolution: {integrity: sha512-Z03/wrqau9Bicfgb3Dbs6SYTHliELk2PM2LpG2nFd+cGupTMF5kanLEcj2vuuJLLhptNyS61rtk7SOZ+lPsTUA==}
     engines: {node: ^20.19.0 || >=22.12.0}
     cpu: [arm64]
     os: [linux]
-    libc: [musl]
 
   '@rolldown/binding-linux-x64-gnu@1.0.0-rc.3':
     resolution: {integrity: sha512-iSXXZsQp08CSilff/DCTFZHSVEpEwdicV3W8idHyrByrcsRDVh9sGC3sev6d8BygSGj3vt8GvUKBPCoyMA4tgQ==}
     engines: {node: ^20.19.0 || >=22.12.0}
     cpu: [x64]
     os: [linux]
-    libc: [glibc]
 
   '@rolldown/binding-linux-x64-musl@1.0.0-rc.3':
     resolution: {integrity: sha512-qaj+MFudtdCv9xZo9znFvkgoajLdc+vwf0Kz5N44g+LU5XMe+IsACgn3UG7uTRlCCvhMAGXm1XlpEA5bZBrOcw==}
     engines: {node: ^20.19.0 || >=22.12.0}
     cpu: [x64]
     os: [linux]
-    libc: [musl]
 
   '@rolldown/binding-openharmony-arm64@1.0.0-rc.3':
     resolution: {integrity: sha512-U662UnMETyjT65gFmG9ma+XziENrs7BBnENi/27swZPYagubfHRirXHG2oMl+pEax2WvO7Kb9gHZmMakpYqBHQ==}
@@ -617,79 +609,66 @@ packages:
     resolution: {integrity: sha512-t4ONHboXi/3E0rT6OZl1pKbl2Vgxf9vJfWgmUoCEVQVxhW6Cw/c8I6hbbu7DAvgp82RKiH7TpLwxnJeKv2pbsw==}
     cpu: [arm]
     os: [linux]
-    libc: [glibc]
 
   '@rollup/rollup-linux-arm-musleabihf@4.59.0':
     resolution: {integrity: sha512-CikFT7aYPA2ufMD086cVORBYGHffBo4K8MQ4uPS/ZnY54GKj36i196u8U+aDVT2LX4eSMbyHtyOh7D7Zvk2VvA==}
     cpu: [arm]
     os: [linux]
-    libc: [musl]
 
   '@rollup/rollup-linux-arm64-gnu@4.59.0':
     resolution: {integrity: sha512-jYgUGk5aLd1nUb1CtQ8E+t5JhLc9x5WdBKew9ZgAXg7DBk0ZHErLHdXM24rfX+bKrFe+Xp5YuJo54I5HFjGDAA==}
     cpu: [arm64]
     os: [linux]
-    libc: [glibc]
 
   '@rollup/rollup-linux-arm64-musl@4.59.0':
     resolution: {integrity: sha512-peZRVEdnFWZ5Bh2KeumKG9ty7aCXzzEsHShOZEFiCQlDEepP1dpUl/SrUNXNg13UmZl+gzVDPsiCwnV1uI0RUA==}
     cpu: [arm64]
     os: [linux]
-    libc: [musl]
 
   '@rollup/rollup-linux-loong64-gnu@4.59.0':
     resolution: {integrity: sha512-gbUSW/97f7+r4gHy3Jlup8zDG190AuodsWnNiXErp9mT90iCy9NKKU0Xwx5k8VlRAIV2uU9CsMnEFg/xXaOfXg==}
     cpu: [loong64]
     os: [linux]
-    libc: [glibc]
 
   '@rollup/rollup-linux-loong64-musl@4.59.0':
     resolution: {integrity: sha512-yTRONe79E+o0FWFijasoTjtzG9EBedFXJMl888NBEDCDV9I2wGbFFfJQQe63OijbFCUZqxpHz1GzpbtSFikJ4Q==}
     cpu: [loong64]
     os: [linux]
-    libc: [musl]
 
   '@rollup/rollup-linux-ppc64-gnu@4.59.0':
     resolution: {integrity: sha512-sw1o3tfyk12k3OEpRddF68a1unZ5VCN7zoTNtSn2KndUE+ea3m3ROOKRCZxEpmT9nsGnogpFP9x6mnLTCaoLkA==}
     cpu: [ppc64]
     os: [linux]
-    libc: [glibc]
 
   '@rollup/rollup-linux-ppc64-musl@4.59.0':
     resolution: {integrity: sha512-+2kLtQ4xT3AiIxkzFVFXfsmlZiG5FXYW7ZyIIvGA7Bdeuh9Z0aN4hVyXS/G1E9bTP/vqszNIN/pUKCk/BTHsKA==}
     cpu: [ppc64]
     os: [linux]
-    libc: [musl]
 
   '@rollup/rollup-linux-riscv64-gnu@4.59.0':
     resolution: {integrity: sha512-NDYMpsXYJJaj+I7UdwIuHHNxXZ/b/N2hR15NyH3m2qAtb/hHPA4g4SuuvrdxetTdndfj9b1WOmy73kcPRoERUg==}
     cpu: [riscv64]
     os: [linux]
-    libc: [glibc]
 
   '@rollup/rollup-linux-riscv64-musl@4.59.0':
     resolution: {integrity: sha512-nLckB8WOqHIf1bhymk+oHxvM9D3tyPndZH8i8+35p/1YiVoVswPid2yLzgX7ZJP0KQvnkhM4H6QZ5m0LzbyIAg==}
     cpu: [riscv64]
     os: [linux]
-    libc: [musl]
 
   '@rollup/rollup-linux-s390x-gnu@4.59.0':
     resolution: {integrity: sha512-oF87Ie3uAIvORFBpwnCvUzdeYUqi2wY6jRFWJAy1qus/udHFYIkplYRW+wo+GRUP4sKzYdmE1Y3+rY5Gc4ZO+w==}
     cpu: [s390x]
     os: [linux]
-    libc: [glibc]
 
   '@rollup/rollup-linux-x64-gnu@4.59.0':
     resolution: {integrity: sha512-3AHmtQq/ppNuUspKAlvA8HtLybkDflkMuLK4DPo77DfthRb71V84/c4MlWJXixZz4uruIH4uaa07IqoAkG64fg==}
     cpu: [x64]
     os: [linux]
-    libc: [glibc]
 
   '@rollup/rollup-linux-x64-musl@4.59.0':
     resolution: {integrity: sha512-2UdiwS/9cTAx7qIUZB/fWtToJwvt0Vbo0zmnYt7ED35KPg13Q0ym1g442THLC7VyI6JfYTP4PiSOWyoMdV2/xg==}
     cpu: [x64]
     os: [linux]
-    libc: [musl]
 
   '@rollup/rollup-openbsd-x64@4.59.0':
     resolution: {integrity: sha512-M3bLRAVk6GOwFlPTIxVBSYKUaqfLrn8l0psKinkCFxl4lQvOSz8ZrKDz2gxcBwHFpci0B6rttydI4IpS4IS/jQ==}
@@ -771,8 +750,8 @@ packages:
   '@types/node@25.3.0':
     resolution: {integrity: sha512-4K3bqJpXpqfg2XKGK9bpDTc6xO/xoUP/RBWS7AtRMug6zZFaRekiLzjVtAoZMquxoAbzBvy5nxQ7veS5eYzf8A==}
 
-  '@vercel/oidc@3.1.0':
-    resolution: {integrity: sha512-Fw28YZpRnA3cAHHDlkt7xQHiJ0fcL+NRcIqsocZQUSmbzeIKRpwttJjik5ZGanXP+vlA4SbTg+AbA3bP363l+w==}
+  '@vercel/oidc@3.2.0':
+    resolution: {integrity: sha512-UycprH3T6n3jH0k44NHMa7pnFHGu/N05MjojYr+Mc6I7obkoLIJujSWwin1pCvdy/eOxrI/l3uDLQsmcrOb4ug==}
     engines: {node: '>= 20'}
 
   '@vitest/expect@4.0.18':
@@ -804,6 +783,9 @@ packages:
   '@vitest/utils@4.0.18':
     resolution: {integrity: sha512-msMRKLMVLWygpK3u2Hybgi4MNjcYJvwTb0Ru09+fOyCXIgT5raYP041DRRdiJiI3k/2U6SEbAETB3YtBrUkCFA==}
 
+  '@workflow/serde@4.1.0':
+    resolution: {integrity: sha512-pav4F2BoirECWR7Nf1TKt+2eETcBj7jj4cBefQ8VXQCA6NPkaKeLfj/zMgi+3zYV5ZIBT4GuUiphsj0/b9hPQQ==}
+
   acorn-walk@8.3.5:
     resolution: {integrity: sha512-HEHNfbars9v4pgpW6SO1KSPkfoS0xVOM/9UzkJltjlsHZmJasxg8aXkuZa7SMf8vKGIBhpUsPluQSqhJFCqebw==}
     engines: {node: '>=0.4.0'}
@@ -813,9 +795,9 @@ packages:
     engines: {node: '>=0.4.0'}
     hasBin: true
 
-  ai@6.0.140:
-    resolution: {integrity: sha512-+jf6fQDPZe+gQlzPoP5mzy+DdfYOpE0cgUm99U8OxVTIPv19gCDuNzlKTZSntcQzLbo6LFXyNhJdV7XTWQ+5vA==}
-    engines: {node: '>=18'}
+  ai@7.0.0:
+    resolution: {integrity: sha512-hncs+jamJh8r36K6G8xky7oF4Ai/RLU5TF85FMzI2vElyMJGGnLoHihpdmuDiuY2BsktDWHevKaJM1l0VcRLGw==}
+    engines: {node: '>=22'}
     peerDependencies:
       zod: ^3.25.76 || ^4.1.8
 
@@ -936,8 +918,8 @@ packages:
   estree-walker@3.0.3:
     resolution: {integrity: sha512-7RUKfXgSMMkzt6ZuXmqapOurLGPPfgj6l9uRZ7lRGolvk0y2yocc35LdcxKC5PQZdn2DMqioAQ2NoWcrTKmm6g==}
 
-  eventsource-parser@3.0.6:
-    resolution: {integrity: sha512-Vo1ab+QXPzZ4tCa8SwIHJFaSzy4R6SHf7BY79rFBDf0idraZWAkYrDjDj8uWaSm3S2TK+hJ7/t1CEmZ7jXw+pg==}
+  eventsource-parser@3.1.0:
+    resolution: {integrity: sha512-kJezFj9YFAMLeORyi7aCLxLbD5/qWMQnoMVlVPyHIll7lgRJCc3JVln9Vgl9nwQi0YkMnhdGTMNn7CkRRAptMg==}
     engines: {node: '>=18.0.0'}
 
   expect-type@1.3.0:
@@ -1498,21 +1480,22 @@ packages:
 
 snapshots:
 
-  '@ai-sdk/gateway@3.0.82(zod@4.3.6)':
+  '@ai-sdk/gateway@4.0.0(zod@4.3.6)':
     dependencies:
-      '@ai-sdk/provider': 3.0.8
-      '@ai-sdk/provider-utils': 4.0.21(zod@4.3.6)
-      '@vercel/oidc': 3.1.0
+      '@ai-sdk/provider': 4.0.0
+      '@ai-sdk/provider-utils': 5.0.0(zod@4.3.6)
+      '@vercel/oidc': 3.2.0
       zod: 4.3.6
 
-  '@ai-sdk/provider-utils@4.0.21(zod@4.3.6)':
+  '@ai-sdk/provider-utils@5.0.0(zod@4.3.6)':
     dependencies:
-      '@ai-sdk/provider': 3.0.8
+      '@ai-sdk/provider': 4.0.0
       '@standard-schema/spec': 1.1.0
-      eventsource-parser: 3.0.6
+      '@workflow/serde': 4.1.0
+      eventsource-parser: 3.1.0
       zod: 4.3.6
 
-  '@ai-sdk/provider@3.0.8':
+  '@ai-sdk/provider@4.0.0':
     dependencies:
       json-schema: 0.4.0
 
@@ -1900,7 +1883,8 @@ snapshots:
       '@nodelib/fs.scandir': 2.1.5
       fastq: 1.20.1
 
-  '@opentelemetry/api@1.9.0': {}
+  '@opentelemetry/api@1.9.0':
+    optional: true
 
   '@oxc-project/types@0.112.0': {}
 
@@ -2073,7 +2057,7 @@ snapshots:
     dependencies:
       undici-types: 7.18.2
 
-  '@vercel/oidc@3.1.0': {}
+  '@vercel/oidc@3.2.0': {}
 
   '@vitest/expect@4.0.18':
     dependencies:
@@ -2114,18 +2098,19 @@ snapshots:
       '@vitest/pretty-format': 4.0.18
       tinyrainbow: 3.0.3
 
+  '@workflow/serde@4.1.0': {}
+
   acorn-walk@8.3.5:
     dependencies:
       acorn: 8.16.0
 
   acorn@8.16.0: {}
 
-  ai@6.0.140(zod@4.3.6):
+  ai@7.0.0(zod@4.3.6):
     dependencies:
-      '@ai-sdk/gateway': 3.0.82(zod@4.3.6)
-      '@ai-sdk/provider': 3.0.8
-      '@ai-sdk/provider-utils': 4.0.21(zod@4.3.6)
-      '@opentelemetry/api': 1.9.0
+      '@ai-sdk/gateway': 4.0.0(zod@4.3.6)
+      '@ai-sdk/provider': 4.0.0
+      '@ai-sdk/provider-utils': 5.0.0(zod@4.3.6)
       zod: 4.3.6
 
   ansi-colors@4.1.3: {}
@@ -2236,7 +2221,7 @@ snapshots:
     dependencies:
       '@types/estree': 1.0.8
 
-  eventsource-parser@3.0.6: {}
+  eventsource-parser@3.1.0: {}
 
   expect-type@1.3.0: {}
 
diff --git a/skills/context-chef-middleware/SKILL.md b/skills/context-chef-middleware/SKILL.md
index 6bedb37..9a67ed8 100644
--- a/skills/context-chef-middleware/SKILL.md
+++ b/skills/context-chef-middleware/SKILL.md
@@ -1,6 +1,6 @@
 ---
 name: context-chef-middleware
-description: "Helps developers integrate @context-chef/ai-sdk-middleware into their Vercel AI SDK (v6+) projects. Use this skill when the user wants to add transparent context management to an AI SDK app, wrap a model with automatic history compression, truncate large tool results, manage token budgets, or inject dynamic state into AI SDK prompts. Also trigger when the user mentions 'context-chef middleware', 'AI SDK middleware', 'ai-sdk context', or asks about compressing history / truncating tool results / managing tokens in a Vercel AI SDK project."
+description: "Helps developers integrate @context-chef/ai-sdk-middleware into their Vercel AI SDK (v7+) projects. Use this skill when the user wants to add transparent context management to an AI SDK app, wrap a model with automatic history compression, truncate large tool results, manage token budgets, or inject dynamic state into AI SDK prompts. Also trigger when the user mentions 'context-chef middleware', 'AI SDK middleware', 'ai-sdk context', or asks about compressing history / truncating tool results / managing tokens in a Vercel AI SDK project."
 argument-hint: "[feature-focus]"
 allowed-tools: Read, Grep, Glob, Bash, Write, Edit
 ---
@@ -16,7 +16,7 @@ The key selling point: **zero code changes** to existing `generateText` / `strea
 Before asking questions, silently inspect the project:
 
 ```
-1. package.json → confirm they use `ai` (v6+) and an AI SDK provider (@ai-sdk/openai, @ai-sdk/anthropic, @ai-sdk/google, etc.)
+1. package.json → confirm they use `ai` (v7+) and an AI SDK provider (@ai-sdk/openai, @ai-sdk/anthropic, @ai-sdk/google, etc.)
 2. Lock file → detect package manager (pnpm-lock.yaml / yarn.lock / package-lock.json / bun.lockb)
 3. tsconfig.json → TypeScript or JavaScript?
 4. Existing AI SDK usage → look for patterns like:
@@ -40,7 +40,7 @@ Based on what you found, present a brief summary of their setup and ask which fe
 | Need to inject task state for LLM attention | Dynamic state as XML in prompt | `dynamicState` |
 | Want custom prompt manipulation (RAG, metadata) | Post-compression transform hook | `transformContext` |
 | Want to know when compression happens | Compression callback | `onCompress` |
-| Need control over what happens at budget limit | Budget exceeded hook | `onBudgetExceeded` |
+| Need control over what happens at budget limit | Budget exceeded hook | `onBeforeCompress` |
 
 If the developer is unsure, recommend starting with: **compression + truncation** — these solve the most common problems with minimal setup.
 
@@ -55,7 +55,7 @@ yarn add @context-chef/ai-sdk-middleware
 bun add @context-chef/ai-sdk-middleware
 ```
 
-The middleware depends on `ai` (v6+) and `@ai-sdk/provider` (v3+) as peer dependencies — the developer should already have these.
+The middleware depends on `ai` (v7+) and `@ai-sdk/provider` (v4+) as peer dependencies — the developer should already have these.
 
 ## Step 4: Generate integration code
 
@@ -154,9 +154,9 @@ transformContext: (prompt) => {
 }
 ```
 
-**Budget exceeded hook (`onBudgetExceeded`):**
+**Budget exceeded hook (`onBeforeCompress`):**
 ```typescript
-onBudgetExceeded: (history, { currentTokens, limit }) => {
+onBeforeCompress: (history, { currentTokens, limit }) => {
   // Return modified messages, or null to let default compression handle it
   console.log(`Budget exceeded: ${currentTokens}/${limit} tokens`);
   return null;
@@ -240,7 +240,7 @@ const aiSdkPrompt = toAISDK(irMessages);
 
 After generating the code:
 
-1. Verify `ai` (v6+) is in their dependencies — the middleware requires AI SDK v6
+1. Verify `ai` (v7+) is in their dependencies — the middleware requires AI SDK v7
 2. Verify they have at least one `@ai-sdk/*` provider installed
 3. Explain the processing pipeline briefly:
    - Truncate large tool results (if configured)
@@ -255,7 +255,7 @@ After generating the code:
 - Don't create a new wrapped model per LLM call — reuse it across the conversation (it tracks token usage)
 - Don't manually call `reportTokenUsage()` — the middleware extracts it automatically from `generateText` / `streamText` responses
 - The `compress.model` should be a cheap, fast model (e.g. `gpt-4o-mini`, `claude-haiku`) — it's used for summarization, not the main task
-- `withContextChef()` returns a standard `LanguageModelV3` — it works anywhere the original model works
+- `withContextChef()` returns a standard `LanguageModelV4` — it works anywhere the original model works
 - If using `compact` with `truncate` together, `truncate` runs first (on the raw AI SDK prompt), then `compact` runs on the IR (after conversion)
 
 ## When to recommend @context-chef/core instead
diff --git a/skills/context-chef-middleware/references/api-reference.md b/skills/context-chef-middleware/references/api-reference.md
index 9136585..ac4b035 100644
--- a/skills/context-chef-middleware/references/api-reference.md
+++ b/skills/context-chef-middleware/references/api-reference.md
@@ -2,9 +2,9 @@
 
 ## Exports
 
-### `withContextChef(model, options): LanguageModelV3`
+### `withContextChef(model, options): LanguageModelV4`
 
-Wraps an AI SDK language model with context-chef middleware. Returns a standard `LanguageModelV3` that can be used anywhere the original model was used.
+Wraps an AI SDK language model with context-chef middleware. Returns a standard `LanguageModelV4` that can be used anywhere the original model was used.
 
 ```typescript
 import { withContextChef } from '@context-chef/ai-sdk-middleware';
@@ -24,11 +24,11 @@ const model = wrapLanguageModel({ model: openai('gpt-4o'), middleware });
 
 ### `fromAISDK(prompt): AISDKMessage[]`
 
-Converts an AI SDK `LanguageModelV3Prompt` to context-chef `Message[]` IR. Original AI SDK content is stored in per-role fields for lossless round-trip.
+Converts an AI SDK `LanguageModelV4Prompt` to context-chef `Message[]` IR. Original AI SDK content is stored in per-role fields for lossless round-trip.
 
-### `toAISDK(messages): LanguageModelV3Prompt`
+### `toAISDK(messages): LanguageModelV4Prompt`
 
-Converts context-chef `Message[]` IR back to AI SDK `LanguageModelV3Prompt`. Uses original content when unmodified; falls back to constructing from IR fields when content was modified by Janitor.
+Converts context-chef `Message[]` IR back to AI SDK `LanguageModelV4Prompt`. Uses original content when unmodified; falls back to constructing from IR fields when content was modified by Janitor.
 
 ### `summarizeMessages(prompt, model, opts?): Promise<string>`
 
@@ -54,9 +54,9 @@ The main configuration object passed to `withContextChef()` or `createMiddleware
 | `compact` | `CompactConfig` | No | Mechanical compaction before LLM compression |
 | `dynamicState` | `DynamicStateConfig` | No | Dynamic state injection into prompt |
 | `tokenizer` | `(msgs: unknown[]) => number` | No | Custom tokenizer for precise token counting |
-| `onCompress` | `(summary: string, count: number, details: { compressedMessages: LanguageModelV3Prompt }) => void` | No | Hook called after compression occurs. `details.compressedMessages` is the compressed slice in AI SDK format. |
-| `onBudgetExceeded` | `(history, tokenInfo) => Message[] \| null \| Promise<...>` | No | Hook called when token budget is exceeded |
-| `transformContext` | `(prompt) => LanguageModelV3Prompt \| Promise<...>` | No | Transform prompt after compression, before model |
+| `onCompress` | `(summary: string, count: number, details: { compressedMessages: LanguageModelV4Prompt }) => void` | No | Hook called after compression occurs. `details.compressedMessages` is the compressed slice in AI SDK format. |
+| `onBeforeCompress` | `(history, tokenInfo) => Message[] \| null \| Promise<...>` | No | Hook called when token budget is exceeded |
+| `transformContext` | `(prompt) => LanguageModelV4Prompt \| Promise<...>` | No | Transform prompt after compression, before model |
 | `logger` | `ChefLogger` | No | Sink for degradation warnings (storage/compaction); defaults to `console`. `ChefLogger = { warn(message: string, ...args: unknown[]): void }` |
 | `clear` | `ClearTarget[]` | No | Placeholder-style tool-result clearing, runs after compression, auto-injects an explainer. Only `'tool-result'` takes effect; `'thinking'` is a no-op that warns — use `compact` for reasoning. |
 
@@ -66,7 +66,7 @@ The main configuration object passed to `withContextChef()` or `createMiddleware
 
 | Field | Type | Required | Description |
 |---|---|---|---|
-| `model` | `LanguageModelV3` | Yes | A cheap model for summarization (e.g. `openai('gpt-4o-mini')`) |
+| `model` | `LanguageModelV4` | Yes | A cheap model for summarization (e.g. `openai('gpt-4o-mini')`) |
 | `preserveRatio` | `number` | No | Ratio of context window to preserve for recent messages. Default: `0.8` |
 | `toolResultStubThreshold` | `number` | No | Replace tool-result content longer than this many chars with a one-line metadata stub (`[Tool name returned N chars; omitted before summarization]`) before the to-be-summarized history is sent to the compression model. Recent (preserved) tool results are untouched. Default: undefined (disabled). |