feat(ai-sdk-middleware)!: support AI SDK v7 (provider spec V4)

.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).

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+)

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
 
 ```

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"

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':

packages/ai-sdk-middleware/src/compaction.ts

@@ -1,106 +1,15 @@
-import type { LanguageModelV3, LanguageModelV3Prompt } from '@ai-sdk/provider';
 import {
   compactHistory as coreCompactHistory,
   planCompaction as corePlanCompaction,
   type PlanCompactionOptions,
 } 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.
  *