@context-chef/ai-sdk-middleware 0.1.0

package/README.md

@@ -0,0 +1,164 @@
+# @context-chef/ai-sdk-middleware
+
+[![npm version](https://img.shields.io/npm/v/@context-chef/ai-sdk-middleware.svg)](https://www.npmjs.com/package/@context-chef/ai-sdk-middleware)
+
+[Vercel AI SDK](https://sdk.vercel.ai) 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.
+
+## Installation
+
+```bash
+npm install @context-chef/ai-sdk-middleware @context-chef/core ai
+```
+
+## Quick Start
+
+```typescript
+import { withContextChef } from '@context-chef/ai-sdk-middleware';
+import { openai } from '@ai-sdk/openai';
+import { generateText } from 'ai';
+
+const model = withContextChef(openai('gpt-4o'), {
+  contextWindow: 128_000,
+});
+
+// Use exactly like normal — everything below is unchanged
+const result = await generateText({
+  model,
+  messages: conversationHistory,
+  tools: myTools,
+});
+```
+
+That's it. History compression and token budget tracking happen automatically behind the scenes.
+
+## Features
+
+### History Compression
+
+When the conversation exceeds the token budget, the middleware compresses older messages to make room. Two modes:
+
+**Without a compression model** (default) — old messages are discarded, only recent messages are kept:
+
+```typescript
+const model = withContextChef(openai('gpt-4o'), {
+  contextWindow: 128_000,
+});
+```
+
+**With a compression model** — old messages are summarized by a cheap model before being replaced:
+
+```typescript
+const model = withContextChef(openai('gpt-4o'), {
+  contextWindow: 128_000,
+  compress: {
+    model: openai('gpt-4o-mini'),  // cheap model for summarization
+    preserveRatio: 0.8,             // keep 80% of context for recent messages
+  },
+});
+```
+
+### Tool Result Truncation
+
+Large tool outputs (terminal logs, API responses) are automatically truncated while preserving the head and tail:
+
+```typescript
+const model = withContextChef(openai('gpt-4o'), {
+  contextWindow: 128_000,
+  truncate: {
+    threshold: 5000,   // truncate tool results over 5000 chars
+    headChars: 500,    // preserve first 500 chars
+    tailChars: 1000,   // preserve last 1000 chars
+  },
+});
+```
+
+### Token Budget Tracking
+
+The middleware automatically extracts token usage from `generateText` and `streamText` responses and feeds it back to the compression engine. No manual `reportTokenUsage()` calls needed.
+
+## API
+
+### `withContextChef(model, options)`
+
+Wraps an AI SDK language model with context-chef middleware.
+
+```typescript
+import { withContextChef } from '@context-chef/ai-sdk-middleware';
+
+const wrappedModel = withContextChef(model, options);
+```
+
+**Parameters:**
+
+| Option | Type | Required | Description |
+|---|---|---|---|
+| `contextWindow` | `number` | Yes | Model's context window size in tokens |
+| `compress` | `CompressOptions` | No | Enable LLM-based compression |
+| `compress.model` | `LanguageModelV3` | Yes (if compress) | Cheap model for summarization |
+| `compress.preserveRatio` | `number` | No | Ratio of context to preserve (default: `0.8`) |
+| `truncate` | `TruncateOptions` | No | Enable tool result truncation |
+| `truncate.threshold` | `number` | Yes (if truncate) | Character count to trigger truncation |
+| `truncate.headChars` | `number` | No | Characters to preserve from start (default: `0`) |
+| `truncate.tailChars` | `number` | No | Characters to preserve from end (default: `1000`) |
+| `tokenizer` | `(msgs) => number` | No | Custom tokenizer for precise counting |
+| `onCompress` | `(summary, count) => void` | No | Hook called after compression |
+
+**Returns:** `LanguageModelV3` — a wrapped model that can be used anywhere the original model was used.
+
+### `createMiddleware(options)`
+
+Creates a raw `LanguageModelMiddleware` if you want to apply it yourself via `wrapLanguageModel`:
+
+```typescript
+import { createMiddleware } from '@context-chef/ai-sdk-middleware';
+import { wrapLanguageModel } from 'ai';
+
+const middleware = createMiddleware({ contextWindow: 128_000 });
+const model = wrapLanguageModel({ model: openai('gpt-4o'), middleware });
+```
+
+### `fromAISDK(prompt)` / `toAISDK(messages)`
+
+Low-level converters between AI SDK `LanguageModelV3Prompt` and context-chef `Message[]` IR. Useful if you want to use context-chef modules directly with AI SDK message formats.
+
+```typescript
+import { fromAISDK, toAISDK } from '@context-chef/ai-sdk-middleware';
+
+const irMessages = fromAISDK(aiSdkPrompt);
+// ... process with context-chef modules ...
+const aiSdkPrompt = toAISDK(irMessages);
+```
+
+## How It Works
+
+```
+generateText({ model: wrappedModel, messages })
+  |
+  v
+transformParams (before LLM call)
+  1. Truncate large tool results (if configured)
+  2. Convert AI SDK messages -> context-chef IR
+  3. Run Janitor compression (if over token budget)
+  4. Convert back to AI SDK messages
+  |
+  v
+LLM call executes normally
+  |
+  v
+wrapGenerate / wrapStream (after LLM call)
+  5. Extract token usage from response
+  6. Feed back to Janitor for next call's budget check
+  |
+  v
+Result returned unchanged
+```
+
+The middleware is **stateful** — it tracks token usage across calls to know when compression is needed. Create one wrapped model per conversation/session.
+
+## Need More Control?
+
+The middleware covers the most common use case: transparent compression and truncation. For advanced features like dynamic state injection, tool namespaces, memory, or snapshot/restore, use [`@context-chef/core`](../core) directly.
+
+## License
+
+ISC

package/dist/index.cjs

@@ -0,0 +1,371 @@
+Object.defineProperty(exports, Symbol.toStringTag, { value: 'Module' });
+let ai = require("ai");
+let _context_chef_core = require("@context-chef/core");
+
+//#region src/adapter.ts
+/**
+* Converts an AI SDK V3 prompt to context-chef IR messages.
+*
+* Original AI SDK content is stored in `_originalContent` for lossless round-trip.
+* `_originalText` caches the extracted text so `toAISDK` can detect Janitor modifications.
+* `_providerOptions` preserves message-level provider options (e.g. Anthropic cache control).
+*/
+function fromAISDK(prompt) {
+	const messages = [];
+	for (const msg of prompt) {
+		if (msg.role === "system") {
+			messages.push({
+				role: "system",
+				content: msg.content,
+				...msg.providerOptions ? { _providerOptions: msg.providerOptions } : {}
+			});
+			continue;
+		}
+		if (msg.role === "user") {
+			const text = msg.content.filter((p) => p.type === "text").map((p) => p.text).join("\n");
+			messages.push({
+				role: "user",
+				content: text,
+				_originalContent: msg.content,
+				_originalText: text,
+				...msg.providerOptions ? { _providerOptions: msg.providerOptions } : {}
+			});
+			continue;
+		}
+		if (msg.role === "assistant") {
+			const text = [];
+			const toolCalls = [];
+			let thinking;
+			for (const part of msg.content) if (part.type === "text") text.push(part.text);
+			else if (part.type === "tool-call") toolCalls.push({
+				id: part.toolCallId,
+				type: "function",
+				function: {
+					name: part.toolName,
+					arguments: typeof part.input === "string" ? part.input : JSON.stringify(part.input)
+				}
+			});
+			else if (part.type === "reasoning") thinking = { thinking: part.text };
+			const joinedText = text.join("\n");
+			const m = {
+				role: "assistant",
+				content: joinedText,
+				_originalContent: msg.content,
+				_originalText: joinedText,
+				...msg.providerOptions ? { _providerOptions: msg.providerOptions } : {}
+			};
+			if (toolCalls.length > 0) m.tool_calls = toolCalls;
+			if (thinking) m.thinking = thinking;
+			messages.push(m);
+			continue;
+		}
+		if (msg.role === "tool") {
+			for (const part of msg.content) if (part.type === "tool-result") {
+				const text = stringifyToolOutput(part.output);
+				messages.push({
+					role: "tool",
+					content: text,
+					tool_call_id: part.toolCallId,
+					_originalContent: [part],
+					_originalText: text,
+					_toolName: part.toolName
+				});
+			}
+		}
+	}
+	return messages;
+}
+/**
+* Converts context-chef IR messages back to AI SDK V3 prompt format.
+*
+* Uses `_originalContent` when content is 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).
+*/
+function toAISDK(messages) {
+	const prompt = [];
+	let i = 0;
+	while (i < messages.length) {
+		const msg = messages[i];
+		const providerOptions = msg._providerOptions;
+		const contentModified = msg._originalText !== void 0 && msg._originalText !== msg.content;
+		if (msg.role === "system") {
+			prompt.push({
+				role: "system",
+				content: msg.content,
+				...providerOptions ? { providerOptions } : {}
+			});
+			i++;
+			continue;
+		}
+		if (msg.role === "user") {
+			const content = !contentModified && Array.isArray(msg._originalContent) ? msg._originalContent : [{
+				type: "text",
+				text: msg.content
+			}];
+			prompt.push({
+				role: "user",
+				content,
+				...providerOptions ? { providerOptions } : {}
+			});
+			i++;
+			continue;
+		}
+		if (msg.role === "assistant") {
+			const content = !contentModified && Array.isArray(msg._originalContent) ? msg._originalContent : [{
+				type: "text",
+				text: msg.content
+			}];
+			prompt.push({
+				role: "assistant",
+				content,
+				...providerOptions ? { providerOptions } : {}
+			});
+			i++;
+			continue;
+		}
+		if (msg.role === "tool") {
+			const toolResults = [];
+			while (i < messages.length && messages[i].role === "tool") {
+				const toolMsg = messages[i];
+				if (!(toolMsg._originalText !== void 0 && toolMsg._originalText !== toolMsg.content) && toolMsg._originalContent) toolResults.push(...toolMsg._originalContent);
+				else toolResults.push({
+					type: "tool-result",
+					toolCallId: toolMsg.tool_call_id ?? "",
+					toolName: toolMsg._toolName ?? "unknown",
+					output: {
+						type: "text",
+						value: toolMsg.content
+					}
+				});
+				i++;
+			}
+			prompt.push({
+				role: "tool",
+				content: toolResults
+			});
+			continue;
+		}
+		i++;
+	}
+	return prompt;
+}
+function stringifyToolOutput(output) {
+	switch (output.type) {
+		case "text":
+		case "error-text": return output.value;
+		case "json":
+		case "error-json": return JSON.stringify(output.value);
+		case "content": return output.value.map((v) => v.type === "text" ? v.text ?? "" : "").filter(Boolean).join("\n");
+		default: return JSON.stringify(output);
+	}
+}
+
+//#endregion
+//#region src/truncator.ts
+/**
+* Truncates tool-result content within an AI SDK prompt when it exceeds the configured threshold.
+* When a storage adapter is provided, original content is persisted and a URI is included in the output.
+*/
+async function truncateToolResults(prompt, options) {
+	const { threshold, headChars = 0, tailChars = 1e3, storage } = options;
+	const offloader = storage ? new _context_chef_core.Offloader({
+		threshold,
+		adapter: storage,
+		storageDir: ""
+	}) : null;
+	const result = [];
+	for (const msg of prompt) {
+		if (msg.role !== "tool") {
+			result.push(msg);
+			continue;
+		}
+		const newContent = [];
+		for (const part of msg.content) {
+			if (part.type !== "tool-result") {
+				newContent.push(part);
+				continue;
+			}
+			const text = extractText(part.output);
+			if (text.length <= threshold || headChars + tailChars >= text.length) {
+				newContent.push(part);
+				continue;
+			}
+			if (offloader) try {
+				const vfsResult = await offloader.offloadAsync(text, {
+					threshold,
+					headChars,
+					tailChars
+				});
+				newContent.push({
+					...part,
+					output: {
+						type: "text",
+						value: vfsResult.content
+					}
+				});
+				continue;
+			} catch (error) {
+				console.warn(`[context-chef] Storage adapter write failed for tool result (${part.toolCallId}). Falling back to simple truncation. Error: ${error instanceof Error ? error.message : String(error)}`);
+			}
+			const head = text.slice(0, headChars);
+			const tail = text.slice(text.length - tailChars);
+			const truncated = [
+				head,
+				`\n--- truncated (${text.split("\n").length} lines, ${text.length} chars total) ---\n`,
+				tail
+			].filter(Boolean).join("").trim();
+			newContent.push({
+				...part,
+				output: {
+					type: "text",
+					value: truncated
+				}
+			});
+		}
+		result.push({
+			...msg,
+			content: newContent
+		});
+	}
+	return result;
+}
+function extractText(output) {
+	switch (output.type) {
+		case "text":
+		case "error-text": return output.value;
+		case "json":
+		case "error-json": return JSON.stringify(output.value);
+		case "content": return output.value.map((v) => v.type === "text" ? v.text ?? "" : "").filter(Boolean).join("\n");
+		default: return "";
+	}
+}
+
+//#endregion
+//#region src/middleware.ts
+/**
+* Creates a LanguageModelMiddleware that transparently applies
+* context-chef compression and truncation to AI SDK model calls.
+*
+* The middleware holds a stateful Janitor instance that tracks
+* token usage across calls for compression decisions.
+*/
+function createMiddleware(options) {
+	let usageWarned = false;
+	const janitor = new _context_chef_core.Janitor({
+		contextWindow: options.contextWindow,
+		tokenizer: options.tokenizer ? (msgs) => options.tokenizer?.(msgs) ?? 0 : void 0,
+		preserveRatio: options.compress?.preserveRatio ?? .8,
+		compressionModel: options.compress?.model ? createCompressionAdapter(options.compress.model) : void 0,
+		onCompress: options.onCompress ? (summary, count) => options.onCompress?.(summary.content, count) : void 0
+	});
+	return {
+		specificationVersion: "v3",
+		transformParams: async ({ params }) => {
+			let { prompt } = params;
+			if (options.truncate) prompt = await truncateToolResults(prompt, options.truncate);
+			const irMessages = fromAISDK(prompt);
+			const compressed = await janitor.compress(irMessages);
+			if (compressed !== irMessages) prompt = toAISDK(compressed);
+			return {
+				...params,
+				prompt
+			};
+		},
+		wrapGenerate: async ({ doGenerate }) => {
+			const result = await doGenerate();
+			if (result.usage?.inputTokens?.total != null) janitor.feedTokenUsage(result.usage.inputTokens.total);
+			else if (!usageWarned && !options.tokenizer) {
+				usageWarned = true;
+				console.warn("[context-chef] Model response did not include usage.inputTokens.total. Token-based compression may not trigger accurately. Consider providing a tokenizer for precise token counting.");
+			}
+			return result;
+		},
+		wrapStream: async ({ doStream }) => {
+			const { stream, ...rest } = await doStream();
+			const transform = new TransformStream({ transform(chunk, controller) {
+				if (chunk.type === "finish") {
+					if (chunk.usage?.inputTokens?.total != null) janitor.feedTokenUsage(chunk.usage.inputTokens.total);
+					else if (!usageWarned && !options.tokenizer) {
+						usageWarned = true;
+						console.warn("[context-chef] Stream finish did not include usage.inputTokens.total. Token-based compression may not trigger accurately. Consider providing a tokenizer for precise token counting.");
+					}
+				}
+				controller.enqueue(chunk);
+			} });
+			return {
+				...rest,
+				stream: stream.pipeThrough(transform)
+			};
+		}
+	};
+}
+/**
+* Adapts an AI SDK LanguageModelV3 into the compressionModel callback
+* that Janitor expects: (messages: Message[]) => Promise<string>
+*
+* Tool messages are converted to user messages describing the tool interaction,
+* since generateText only accepts system/user/assistant roles.
+*/
+function createCompressionAdapter(model) {
+	return async (messages) => {
+		const { text } = await (0, ai.generateText)({
+			model,
+			messages: messages.map((m) => {
+				if (m.role === "tool") return {
+					role: "user",
+					content: `[Tool result${m.tool_call_id ? ` (${m.tool_call_id})` : ""}: ${m.content}]`
+				};
+				if (m.role === "assistant" && m.tool_calls?.length) {
+					const toolCallsDesc = m.tool_calls.map((tc) => `[Called tool: ${tc.function.name}(${tc.function.arguments})]`).join("\n");
+					return {
+						role: "assistant",
+						content: m.content ? `${m.content}\n${toolCallsDesc}` : toolCallsDesc
+					};
+				}
+				return {
+					role: m.role,
+					content: m.content
+				};
+			}),
+			maxOutputTokens: 2048
+		});
+		return text || "[Compression produced no output]";
+	};
+}
+
+//#endregion
+//#region src/index.ts
+/**
+* Wraps an AI SDK language model with context-chef middleware for
+* transparent history compression, tool result truncation, and token budget management.
+*
+* @example
+* ```typescript
+* import { withContextChef } from '@context-chef/ai-sdk-middleware';
+* import { openai } from '@ai-sdk/openai';
+* import { generateText } from 'ai';
+*
+* const model = withContextChef(openai('gpt-4o'), {
+*   contextWindow: 128_000,
+*   compress: { model: openai('gpt-4o-mini') },
+*   truncate: { threshold: 5000, headChars: 500, tailChars: 1000 },
+* });
+*
+* // Use exactly like normal — zero other code changes
+* const result = await generateText({ model, messages, tools });
+* ```
+*/
+function withContextChef(model, options) {
+	return (0, ai.wrapLanguageModel)({
+		model,
+		middleware: createMiddleware(options)
+	});
+}
+
+//#endregion
+exports.createMiddleware = createMiddleware;
+exports.fromAISDK = fromAISDK;
+exports.toAISDK = toAISDK;
+exports.withContextChef = withContextChef;
+//# sourceMappingURL=index.cjs.map

package/dist/index.cjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.cjs","names":["Offloader","Janitor"],"sources":["../src/adapter.ts","../src/truncator.ts","../src/middleware.ts","../src/index.ts"],"sourcesContent":["import type {\n  LanguageModelV3Prompt,\n  LanguageModelV3ToolResultOutput,\n  LanguageModelV3ToolResultPart,\n  SharedV3ProviderOptions,\n} from '@ai-sdk/provider';\nimport type { Message, ToolCall } from '@context-chef/core';\n\n/**\n * Converts an AI SDK V3 prompt to context-chef IR messages.\n *\n * Original AI SDK content is stored in `_originalContent` for lossless round-trip.\n * `_originalText` caches the extracted text so `toAISDK` can detect Janitor modifications.\n * `_providerOptions` preserves message-level provider options (e.g. Anthropic cache control).\n */\nexport function fromAISDK(prompt: LanguageModelV3Prompt): Message[] {\n  const messages: Message[] = [];\n\n  for (const msg of prompt) {\n    if (msg.role === 'system') {\n      messages.push({\n        role: 'system',\n        content: msg.content,\n        ...(msg.providerOptions ? { _providerOptions: msg.providerOptions } : {}),\n      });\n      continue;\n    }\n\n    if (msg.role === 'user') {\n      const text = msg.content\n        .filter((p) => p.type === 'text')\n        .map((p) => p.text)\n        .join('\\n');\n      messages.push({\n        role: 'user',\n        content: text,\n        _originalContent: msg.content,\n        _originalText: text,\n        ...(msg.providerOptions ? { _providerOptions: msg.providerOptions } : {}),\n      });\n      continue;\n    }\n\n    if (msg.role === 'assistant') {\n      const text: string[] = [];\n      const toolCalls: ToolCall[] = [];\n      let thinking: { thinking: string } | undefined;\n\n      for (const part of msg.content) {\n        if (part.type === 'text') text.push(part.text);\n        else if (part.type === 'tool-call') {\n          toolCalls.push({\n            id: part.toolCallId,\n            type: 'function',\n            function: {\n              name: part.toolName,\n              arguments: typeof part.input === 'string' ? part.input : JSON.stringify(part.input),\n            },\n          });\n        } else if (part.type === 'reasoning') {\n          thinking = { thinking: part.text };\n        }\n      }\n\n      const joinedText = text.join('\\n');\n      const m: Message = {\n        role: 'assistant',\n        content: joinedText,\n        _originalContent: msg.content,\n        _originalText: joinedText,\n        ...(msg.providerOptions ? { _providerOptions: msg.providerOptions } : {}),\n      };\n      if (toolCalls.length > 0) m.tool_calls = toolCalls;\n      if (thinking) m.thinking = thinking;\n      messages.push(m);\n      continue;\n    }\n\n    if (msg.role === 'tool') {\n      for (const part of msg.content) {\n        if (part.type === 'tool-result') {\n          const text = stringifyToolOutput(part.output);\n          messages.push({\n            role: 'tool',\n            content: text,\n            tool_call_id: part.toolCallId,\n            _originalContent: [part],\n            _originalText: text,\n            _toolName: part.toolName,\n          });\n        }\n      }\n    }\n  }\n\n  return messages;\n}\n\n/**\n * Converts context-chef IR messages back to AI SDK V3 prompt format.\n *\n * Uses `_originalContent` when content is unmodified (detected via `_originalText`).\n * Falls back to constructing from IR fields when content was modified by Janitor\n * (e.g. compact() cleared tool results) or for new messages (e.g. compression summaries).\n */\nexport function toAISDK(messages: Message[]): LanguageModelV3Prompt {\n  const prompt: LanguageModelV3Prompt = [];\n\n  let i = 0;\n  while (i < messages.length) {\n    const msg = messages[i];\n    const providerOptions = msg._providerOptions as SharedV3ProviderOptions | undefined;\n    const contentModified = msg._originalText !== undefined && msg._originalText !== msg.content;\n\n    if (msg.role === 'system') {\n      prompt.push({\n        role: 'system',\n        content: msg.content,\n        ...(providerOptions ? { providerOptions } : {}),\n      });\n      i++;\n      continue;\n    }\n\n    if (msg.role === 'user') {\n      const content =\n        !contentModified && Array.isArray(msg._originalContent)\n          ? (msg._originalContent as any)\n          : [{ type: 'text' as const, text: msg.content }];\n      prompt.push({\n        role: 'user',\n        content,\n        ...(providerOptions ? { providerOptions } : {}),\n      });\n      i++;\n      continue;\n    }\n\n    if (msg.role === 'assistant') {\n      const content =\n        !contentModified && Array.isArray(msg._originalContent)\n          ? (msg._originalContent as any)\n          : [{ type: 'text' as const, text: msg.content }];\n      prompt.push({\n        role: 'assistant',\n        content,\n        ...(providerOptions ? { providerOptions } : {}),\n      });\n      i++;\n      continue;\n    }\n\n    if (msg.role === 'tool') {\n      const toolResults: LanguageModelV3ToolResultPart[] = [];\n      while (i < messages.length && messages[i].role === 'tool') {\n        const toolMsg = messages[i];\n        const toolModified =\n          toolMsg._originalText !== undefined && toolMsg._originalText !== toolMsg.content;\n\n        if (!toolModified && toolMsg._originalContent) {\n          toolResults.push(...(toolMsg._originalContent as LanguageModelV3ToolResultPart[]));\n        } else {\n          toolResults.push({\n            type: 'tool-result',\n            toolCallId: toolMsg.tool_call_id ?? '',\n            toolName: (toolMsg._toolName as string) ?? 'unknown',\n            output: { type: 'text', value: toolMsg.content },\n          });\n        }\n        i++;\n      }\n      prompt.push({ role: 'tool', content: toolResults });\n      continue;\n    }\n\n    i++;\n  }\n\n  return prompt;\n}\n\nfunction stringifyToolOutput(output: LanguageModelV3ToolResultOutput): string {\n  switch (output.type) {\n    case 'text':\n    case 'error-text':\n      return output.value;\n    case 'json':\n    case 'error-json':\n      return JSON.stringify(output.value);\n    case 'content':\n      return output.value\n        .map((v: { type: string; text?: string }) => (v.type === 'text' ? (v.text ?? '') : ''))\n        .filter(Boolean)\n        .join('\\n');\n    default:\n      return JSON.stringify(output);\n  }\n}\n","import type {\n  LanguageModelV3Prompt,\n  LanguageModelV3ToolResultOutput,\n  LanguageModelV3ToolResultPart,\n} from '@ai-sdk/provider';\nimport { Offloader } from '@context-chef/core';\nimport type { TruncateOptions } from './types';\n\n/**\n * Truncates tool-result content within an AI SDK prompt when it exceeds the configured threshold.\n * When a storage adapter is provided, original content is persisted and a URI is included in the output.\n */\nexport async function truncateToolResults(\n  prompt: LanguageModelV3Prompt,\n  options: TruncateOptions,\n): Promise<LanguageModelV3Prompt> {\n  const { threshold, headChars = 0, tailChars = 1000, storage } = options;\n\n  const offloader = storage\n    ? new Offloader({ threshold, adapter: storage, storageDir: '' })\n    : null;\n\n  const result: LanguageModelV3Prompt = [];\n\n  for (const msg of prompt) {\n    if (msg.role !== 'tool') {\n      result.push(msg);\n      continue;\n    }\n\n    const newContent: typeof msg.content = [];\n\n    for (const part of msg.content) {\n      if (part.type !== 'tool-result') {\n        newContent.push(part);\n        continue;\n      }\n\n      const text = extractText(part.output);\n      if (text.length <= threshold || headChars + tailChars >= text.length) {\n        newContent.push(part);\n        continue;\n      }\n\n      // With storage: use Offloader to persist original and get a URI-annotated truncation\n      if (offloader) {\n        try {\n          const vfsResult = await offloader.offloadAsync(text, { threshold, headChars, tailChars });\n          newContent.push({\n            ...part,\n            output: { type: 'text', value: vfsResult.content } satisfies LanguageModelV3ToolResultOutput,\n          } satisfies LanguageModelV3ToolResultPart);\n          continue;\n        } catch (error) {\n          console.warn(\n            `[context-chef] Storage adapter write failed for tool result (${part.toolCallId}). ` +\n              `Falling back to simple truncation. Error: ${error instanceof Error ? error.message : String(error)}`,\n          );\n          // Fall through to simple truncation below\n        }\n      }\n\n      // Without storage: simple truncation, original is discarded\n      const head = text.slice(0, headChars);\n      const tail = text.slice(text.length - tailChars);\n      const totalLines = text.split('\\n').length;\n\n      const truncated = [\n        head,\n        `\\n--- truncated (${totalLines} lines, ${text.length} chars total) ---\\n`,\n        tail,\n      ]\n        .filter(Boolean)\n        .join('')\n        .trim();\n\n      newContent.push({\n        ...part,\n        output: { type: 'text', value: truncated } satisfies LanguageModelV3ToolResultOutput,\n      } satisfies LanguageModelV3ToolResultPart);\n    }\n\n    result.push({ ...msg, content: newContent });\n  }\n\n  return result;\n}\n\nfunction extractText(output: LanguageModelV3ToolResultOutput): string {\n  switch (output.type) {\n    case 'text':\n    case 'error-text':\n      return output.value;\n    case 'json':\n    case 'error-json':\n      return JSON.stringify(output.value);\n    case 'content':\n      return output.value\n        .map((v: { type: string; text?: string }) => (v.type === 'text' ? (v.text ?? '') : ''))\n        .filter(Boolean)\n        .join('\\n');\n    default:\n      return '';\n  }\n}\n","import type { LanguageModelV3, LanguageModelV3StreamPart } from '@ai-sdk/provider';\nimport { generateText, type LanguageModelMiddleware } from 'ai';\nimport { Janitor, type Message } from '@context-chef/core';\n\nimport { fromAISDK, toAISDK } from './adapter';\nimport { truncateToolResults } from './truncator';\nimport type { ContextChefOptions } from './types';\n\n/**\n * Creates a LanguageModelMiddleware that transparently applies\n * context-chef compression and truncation to AI SDK model calls.\n *\n * The middleware holds a stateful Janitor instance that tracks\n * token usage across calls for compression decisions.\n */\nexport function createMiddleware(options: ContextChefOptions): LanguageModelMiddleware {\n  let usageWarned = false;\n\n  const janitor = new Janitor({\n    contextWindow: options.contextWindow,\n    tokenizer: options.tokenizer ? (msgs: Message[]) => options.tokenizer?.(msgs) ?? 0 : undefined,\n    preserveRatio: options.compress?.preserveRatio ?? 0.8,\n    compressionModel: options.compress?.model\n      ? createCompressionAdapter(options.compress.model)\n      : undefined,\n    onCompress: options.onCompress\n      ? (summary, count) => options.onCompress?.(summary.content, count)\n      : undefined,\n  });\n\n  return {\n    specificationVersion: 'v3',\n\n    transformParams: async ({ params }) => {\n      let { prompt } = params;\n\n      // 1. Truncate large tool results\n      if (options.truncate) {\n        prompt = await truncateToolResults(prompt, options.truncate);\n      }\n\n      // 2. Compress history if over token budget\n      const irMessages = fromAISDK(prompt);\n      const compressed = await janitor.compress(irMessages);\n\n      // Only convert back if compression actually changed something\n      if (compressed !== irMessages) {\n        prompt = toAISDK(compressed);\n      }\n\n      return { ...params, prompt };\n    },\n\n    wrapGenerate: async ({ doGenerate }) => {\n      const result = await doGenerate();\n\n      if (result.usage?.inputTokens?.total != null) {\n        janitor.feedTokenUsage(result.usage.inputTokens.total);\n      } else if (!usageWarned && !options.tokenizer) {\n        usageWarned = true;\n        console.warn(\n          '[context-chef] Model response did not include usage.inputTokens.total. ' +\n            'Token-based compression may not trigger accurately. ' +\n            'Consider providing a tokenizer for precise token counting.',\n        );\n      }\n\n      return result;\n    },\n\n    wrapStream: async ({ doStream }) => {\n      const { stream, ...rest } = await doStream();\n\n      const transform = new TransformStream<LanguageModelV3StreamPart, LanguageModelV3StreamPart>({\n        transform(chunk, controller) {\n          if (chunk.type === 'finish') {\n            if (chunk.usage?.inputTokens?.total != null) {\n              janitor.feedTokenUsage(chunk.usage.inputTokens.total);\n            } else if (!usageWarned && !options.tokenizer) {\n              usageWarned = true;\n              console.warn(\n                '[context-chef] Stream finish did not include usage.inputTokens.total. ' +\n                  'Token-based compression may not trigger accurately. ' +\n                  'Consider providing a tokenizer for precise token counting.',\n              );\n            }\n          }\n          controller.enqueue(chunk);\n        },\n      });\n\n      return { ...rest, stream: stream.pipeThrough(transform) };\n    },\n  };\n}\n\n/**\n * Adapts an AI SDK LanguageModelV3 into the compressionModel callback\n * that Janitor expects: (messages: Message[]) => Promise<string>\n *\n * Tool messages are converted to user messages describing the tool interaction,\n * since generateText only accepts system/user/assistant roles.\n */\nfunction createCompressionAdapter(model: LanguageModelV3): (messages: Message[]) => Promise<string> {\n  return async (messages: Message[]): Promise<string> => {\n    const formatted = messages.map((m) => {\n      if (m.role === 'tool') {\n        return {\n          role: 'user' as const,\n          content: `[Tool result${m.tool_call_id ? ` (${m.tool_call_id})` : ''}: ${m.content}]`,\n        };\n      }\n      // assistant messages with tool_calls: include tool call info in content\n      if (m.role === 'assistant' && m.tool_calls?.length) {\n        const toolCallsDesc = m.tool_calls\n          .map((tc) => `[Called tool: ${tc.function.name}(${tc.function.arguments})]`)\n          .join('\\n');\n        return {\n          role: 'assistant' as const,\n          content: m.content ? `${m.content}\\n${toolCallsDesc}` : toolCallsDesc,\n        };\n      }\n      return {\n        role: m.role as 'system' | 'user' | 'assistant',\n        content: m.content,\n      };\n    });\n\n    const { text } = await generateText({\n      model,\n      messages: formatted,\n      maxOutputTokens: 2048,\n    });\n\n    return text || '[Compression produced no output]';\n  };\n}\n","import type { LanguageModelV3 } from '@ai-sdk/provider';\nimport { wrapLanguageModel } from 'ai';\n\nimport { createMiddleware } from './middleware';\nimport type { ContextChefOptions } from './types';\n\nexport { fromAISDK, toAISDK } from './adapter';\nexport { createMiddleware } from './middleware';\nexport type { CompressOptions, ContextChefOptions, TruncateOptions } from './types';\n\n/**\n * Wraps an AI SDK language model with context-chef middleware for\n * transparent history compression, tool result truncation, and token budget management.\n *\n * @example\n * ```typescript\n * import { withContextChef } from '@context-chef/ai-sdk-middleware';\n * import { openai } from '@ai-sdk/openai';\n * import { generateText } from 'ai';\n *\n * const model = withContextChef(openai('gpt-4o'), {\n *   contextWindow: 128_000,\n *   compress: { model: openai('gpt-4o-mini') },\n *   truncate: { threshold: 5000, headChars: 500, tailChars: 1000 },\n * });\n *\n * // Use exactly like normal — zero other code changes\n * const result = await generateText({ model, messages, tools });\n * ```\n */\nexport function withContextChef(model: LanguageModelV3, options: ContextChefOptions): LanguageModelV3 {\n  const middleware = createMiddleware(options);\n  return wrapLanguageModel({ model, middleware });\n}\n"],"mappings":";;;;;;;;;;;;AAeA,SAAgB,UAAU,QAA0C;CAClE,MAAM,WAAsB,EAAE;AAE9B,MAAK,MAAM,OAAO,QAAQ;AACxB,MAAI,IAAI,SAAS,UAAU;AACzB,YAAS,KAAK;IACZ,MAAM;IACN,SAAS,IAAI;IACb,GAAI,IAAI,kBAAkB,EAAE,kBAAkB,IAAI,iBAAiB,GAAG,EAAE;IACzE,CAAC;AACF;;AAGF,MAAI,IAAI,SAAS,QAAQ;GACvB,MAAM,OAAO,IAAI,QACd,QAAQ,MAAM,EAAE,SAAS,OAAO,CAChC,KAAK,MAAM,EAAE,KAAK,CAClB,KAAK,KAAK;AACb,YAAS,KAAK;IACZ,MAAM;IACN,SAAS;IACT,kBAAkB,IAAI;IACtB,eAAe;IACf,GAAI,IAAI,kBAAkB,EAAE,kBAAkB,IAAI,iBAAiB,GAAG,EAAE;IACzE,CAAC;AACF;;AAGF,MAAI,IAAI,SAAS,aAAa;GAC5B,MAAM,OAAiB,EAAE;GACzB,MAAM,YAAwB,EAAE;GAChC,IAAI;AAEJ,QAAK,MAAM,QAAQ,IAAI,QACrB,KAAI,KAAK,SAAS,OAAQ,MAAK,KAAK,KAAK,KAAK;YACrC,KAAK,SAAS,YACrB,WAAU,KAAK;IACb,IAAI,KAAK;IACT,MAAM;IACN,UAAU;KACR,MAAM,KAAK;KACX,WAAW,OAAO,KAAK,UAAU,WAAW,KAAK,QAAQ,KAAK,UAAU,KAAK,MAAM;KACpF;IACF,CAAC;YACO,KAAK,SAAS,YACvB,YAAW,EAAE,UAAU,KAAK,MAAM;GAItC,MAAM,aAAa,KAAK,KAAK,KAAK;GAClC,MAAM,IAAa;IACjB,MAAM;IACN,SAAS;IACT,kBAAkB,IAAI;IACtB,eAAe;IACf,GAAI,IAAI,kBAAkB,EAAE,kBAAkB,IAAI,iBAAiB,GAAG,EAAE;IACzE;AACD,OAAI,UAAU,SAAS,EAAG,GAAE,aAAa;AACzC,OAAI,SAAU,GAAE,WAAW;AAC3B,YAAS,KAAK,EAAE;AAChB;;AAGF,MAAI,IAAI,SAAS,QACf;QAAK,MAAM,QAAQ,IAAI,QACrB,KAAI,KAAK,SAAS,eAAe;IAC/B,MAAM,OAAO,oBAAoB,KAAK,OAAO;AAC7C,aAAS,KAAK;KACZ,MAAM;KACN,SAAS;KACT,cAAc,KAAK;KACnB,kBAAkB,CAAC,KAAK;KACxB,eAAe;KACf,WAAW,KAAK;KACjB,CAAC;;;;AAMV,QAAO;;;;;;;;;AAUT,SAAgB,QAAQ,UAA4C;CAClE,MAAM,SAAgC,EAAE;CAExC,IAAI,IAAI;AACR,QAAO,IAAI,SAAS,QAAQ;EAC1B,MAAM,MAAM,SAAS;EACrB,MAAM,kBAAkB,IAAI;EAC5B,MAAM,kBAAkB,IAAI,kBAAkB,UAAa,IAAI,kBAAkB,IAAI;AAErF,MAAI,IAAI,SAAS,UAAU;AACzB,UAAO,KAAK;IACV,MAAM;IACN,SAAS,IAAI;IACb,GAAI,kBAAkB,EAAE,iBAAiB,GAAG,EAAE;IAC/C,CAAC;AACF;AACA;;AAGF,MAAI,IAAI,SAAS,QAAQ;GACvB,MAAM,UACJ,CAAC,mBAAmB,MAAM,QAAQ,IAAI,iBAAiB,GAClD,IAAI,mBACL,CAAC;IAAE,MAAM;IAAiB,MAAM,IAAI;IAAS,CAAC;AACpD,UAAO,KAAK;IACV,MAAM;IACN;IACA,GAAI,kBAAkB,EAAE,iBAAiB,GAAG,EAAE;IAC/C,CAAC;AACF;AACA;;AAGF,MAAI,IAAI,SAAS,aAAa;GAC5B,MAAM,UACJ,CAAC,mBAAmB,MAAM,QAAQ,IAAI,iBAAiB,GAClD,IAAI,mBACL,CAAC;IAAE,MAAM;IAAiB,MAAM,IAAI;IAAS,CAAC;AACpD,UAAO,KAAK;IACV,MAAM;IACN;IACA,GAAI,kBAAkB,EAAE,iBAAiB,GAAG,EAAE;IAC/C,CAAC;AACF;AACA;;AAGF,MAAI,IAAI,SAAS,QAAQ;GACvB,MAAM,cAA+C,EAAE;AACvD,UAAO,IAAI,SAAS,UAAU,SAAS,GAAG,SAAS,QAAQ;IACzD,MAAM,UAAU,SAAS;AAIzB,QAAI,EAFF,QAAQ,kBAAkB,UAAa,QAAQ,kBAAkB,QAAQ,YAEtD,QAAQ,iBAC3B,aAAY,KAAK,GAAI,QAAQ,iBAAqD;QAElF,aAAY,KAAK;KACf,MAAM;KACN,YAAY,QAAQ,gBAAgB;KACpC,UAAW,QAAQ,aAAwB;KAC3C,QAAQ;MAAE,MAAM;MAAQ,OAAO,QAAQ;MAAS;KACjD,CAAC;AAEJ;;AAEF,UAAO,KAAK;IAAE,MAAM;IAAQ,SAAS;IAAa,CAAC;AACnD;;AAGF;;AAGF,QAAO;;AAGT,SAAS,oBAAoB,QAAiD;AAC5E,SAAQ,OAAO,MAAf;EACE,KAAK;EACL,KAAK,aACH,QAAO,OAAO;EAChB,KAAK;EACL,KAAK,aACH,QAAO,KAAK,UAAU,OAAO,MAAM;EACrC,KAAK,UACH,QAAO,OAAO,MACX,KAAK,MAAwC,EAAE,SAAS,SAAU,EAAE,QAAQ,KAAM,GAAI,CACtF,OAAO,QAAQ,CACf,KAAK,KAAK;EACf,QACE,QAAO,KAAK,UAAU,OAAO;;;;;;;;;;ACvLnC,eAAsB,oBACpB,QACA,SACgC;CAChC,MAAM,EAAE,WAAW,YAAY,GAAG,YAAY,KAAM,YAAY;CAEhE,MAAM,YAAY,UACd,IAAIA,6BAAU;EAAE;EAAW,SAAS;EAAS,YAAY;EAAI,CAAC,GAC9D;CAEJ,MAAM,SAAgC,EAAE;AAExC,MAAK,MAAM,OAAO,QAAQ;AACxB,MAAI,IAAI,SAAS,QAAQ;AACvB,UAAO,KAAK,IAAI;AAChB;;EAGF,MAAM,aAAiC,EAAE;AAEzC,OAAK,MAAM,QAAQ,IAAI,SAAS;AAC9B,OAAI,KAAK,SAAS,eAAe;AAC/B,eAAW,KAAK,KAAK;AACrB;;GAGF,MAAM,OAAO,YAAY,KAAK,OAAO;AACrC,OAAI,KAAK,UAAU,aAAa,YAAY,aAAa,KAAK,QAAQ;AACpE,eAAW,KAAK,KAAK;AACrB;;AAIF,OAAI,UACF,KAAI;IACF,MAAM,YAAY,MAAM,UAAU,aAAa,MAAM;KAAE;KAAW;KAAW;KAAW,CAAC;AACzF,eAAW,KAAK;KACd,GAAG;KACH,QAAQ;MAAE,MAAM;MAAQ,OAAO,UAAU;MAAS;KACnD,CAAyC;AAC1C;YACO,OAAO;AACd,YAAQ,KACN,gEAAgE,KAAK,WAAW,+CACjC,iBAAiB,QAAQ,MAAM,UAAU,OAAO,MAAM,GACtG;;GAML,MAAM,OAAO,KAAK,MAAM,GAAG,UAAU;GACrC,MAAM,OAAO,KAAK,MAAM,KAAK,SAAS,UAAU;GAGhD,MAAM,YAAY;IAChB;IACA,oBAJiB,KAAK,MAAM,KAAK,CAAC,OAIH,UAAU,KAAK,OAAO;IACrD;IACD,CACE,OAAO,QAAQ,CACf,KAAK,GAAG,CACR,MAAM;AAET,cAAW,KAAK;IACd,GAAG;IACH,QAAQ;KAAE,MAAM;KAAQ,OAAO;KAAW;IAC3C,CAAyC;;AAG5C,SAAO,KAAK;GAAE,GAAG;GAAK,SAAS;GAAY,CAAC;;AAG9C,QAAO;;AAGT,SAAS,YAAY,QAAiD;AACpE,SAAQ,OAAO,MAAf;EACE,KAAK;EACL,KAAK,aACH,QAAO,OAAO;EAChB,KAAK;EACL,KAAK,aACH,QAAO,KAAK,UAAU,OAAO,MAAM;EACrC,KAAK,UACH,QAAO,OAAO,MACX,KAAK,MAAwC,EAAE,SAAS,SAAU,EAAE,QAAQ,KAAM,GAAI,CACtF,OAAO,QAAQ,CACf,KAAK,KAAK;EACf,QACE,QAAO;;;;;;;;;;;;;ACvFb,SAAgB,iBAAiB,SAAsD;CACrF,IAAI,cAAc;CAElB,MAAM,UAAU,IAAIC,2BAAQ;EAC1B,eAAe,QAAQ;EACvB,WAAW,QAAQ,aAAa,SAAoB,QAAQ,YAAY,KAAK,IAAI,IAAI;EACrF,eAAe,QAAQ,UAAU,iBAAiB;EAClD,kBAAkB,QAAQ,UAAU,QAChC,yBAAyB,QAAQ,SAAS,MAAM,GAChD;EACJ,YAAY,QAAQ,cACf,SAAS,UAAU,QAAQ,aAAa,QAAQ,SAAS,MAAM,GAChE;EACL,CAAC;AAEF,QAAO;EACL,sBAAsB;EAEtB,iBAAiB,OAAO,EAAE,aAAa;GACrC,IAAI,EAAE,WAAW;AAGjB,OAAI,QAAQ,SACV,UAAS,MAAM,oBAAoB,QAAQ,QAAQ,SAAS;GAI9D,MAAM,aAAa,UAAU,OAAO;GACpC,MAAM,aAAa,MAAM,QAAQ,SAAS,WAAW;AAGrD,OAAI,eAAe,WACjB,UAAS,QAAQ,WAAW;AAG9B,UAAO;IAAE,GAAG;IAAQ;IAAQ;;EAG9B,cAAc,OAAO,EAAE,iBAAiB;GACtC,MAAM,SAAS,MAAM,YAAY;AAEjC,OAAI,OAAO,OAAO,aAAa,SAAS,KACtC,SAAQ,eAAe,OAAO,MAAM,YAAY,MAAM;YAC7C,CAAC,eAAe,CAAC,QAAQ,WAAW;AAC7C,kBAAc;AACd,YAAQ,KACN,wLAGD;;AAGH,UAAO;;EAGT,YAAY,OAAO,EAAE,eAAe;GAClC,MAAM,EAAE,QAAQ,GAAG,SAAS,MAAM,UAAU;GAE5C,MAAM,YAAY,IAAI,gBAAsE,EAC1F,UAAU,OAAO,YAAY;AAC3B,QAAI,MAAM,SAAS,UACjB;SAAI,MAAM,OAAO,aAAa,SAAS,KACrC,SAAQ,eAAe,MAAM,MAAM,YAAY,MAAM;cAC5C,CAAC,eAAe,CAAC,QAAQ,WAAW;AAC7C,oBAAc;AACd,cAAQ,KACN,uLAGD;;;AAGL,eAAW,QAAQ,MAAM;MAE5B,CAAC;AAEF,UAAO;IAAE,GAAG;IAAM,QAAQ,OAAO,YAAY,UAAU;IAAE;;EAE5D;;;;;;;;;AAUH,SAAS,yBAAyB,OAAkE;AAClG,QAAO,OAAO,aAAyC;EAwBrD,MAAM,EAAE,SAAS,2BAAmB;GAClC;GACA,UAzBgB,SAAS,KAAK,MAAM;AACpC,QAAI,EAAE,SAAS,OACb,QAAO;KACL,MAAM;KACN,SAAS,eAAe,EAAE,eAAe,KAAK,EAAE,aAAa,KAAK,GAAG,IAAI,EAAE,QAAQ;KACpF;AAGH,QAAI,EAAE,SAAS,eAAe,EAAE,YAAY,QAAQ;KAClD,MAAM,gBAAgB,EAAE,WACrB,KAAK,OAAO,iBAAiB,GAAG,SAAS,KAAK,GAAG,GAAG,SAAS,UAAU,IAAI,CAC3E,KAAK,KAAK;AACb,YAAO;MACL,MAAM;MACN,SAAS,EAAE,UAAU,GAAG,EAAE,QAAQ,IAAI,kBAAkB;MACzD;;AAEH,WAAO;KACL,MAAM,EAAE;KACR,SAAS,EAAE;KACZ;KACD;GAKA,iBAAiB;GAClB,CAAC;AAEF,SAAO,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;ACxGnB,SAAgB,gBAAgB,OAAwB,SAA8C;AAEpG,kCAAyB;EAAE;EAAO,YADf,iBAAiB,QAAQ;EACE,CAAC"}

package/dist/index.d.cts

@@ -0,0 +1,92 @@
+import { LanguageModelV3, LanguageModelV3Prompt } from "@ai-sdk/provider";
+import { Message, VFSStorageAdapter } from "@context-chef/core";
+import { LanguageModelMiddleware } from "ai";
+
+//#region src/types.d.ts
+interface TruncateOptions {
+  /** Character count threshold to trigger truncation. */
+  threshold: number;
+  /** Characters to preserve from the start. Default: 0 */
+  headChars?: number;
+  /** Characters to preserve from the end. Default: 1000 */
+  tailChars?: number;
+  /**
+   * Storage adapter for persisting original content before truncation.
+   * Can be a FileSystemAdapter, database adapter, or any custom implementation.
+   * When provided, truncated output includes a `context://vfs/` URI for retrieval.
+   * When omitted, original content is discarded after truncation.
+   */
+  storage?: VFSStorageAdapter;
+}
+interface CompressOptions {
+  /** A cheap model used for summarization (e.g. openai('gpt-4o-mini')). */
+  model: LanguageModelV3;
+  /** Ratio of context window to preserve for recent messages. Default: 0.8 */
+  preserveRatio?: number;
+}
+interface ContextChefOptions {
+  /** The model's context window size in tokens. */
+  contextWindow: number;
+  /** Enable history compression. Omit for no compression. */
+  compress?: CompressOptions;
+  /** Enable tool result truncation. Omit for no truncation. */
+  truncate?: TruncateOptions;
+  /** Optional tokenizer for precise per-message token counting. */
+  tokenizer?: (messages: unknown[]) => number;
+  /** Hook called after compression occurs. */
+  onCompress?: (summary: string, truncatedCount: number) => void;
+}
+//#endregion
+//#region src/adapter.d.ts
+/**
+ * Converts an AI SDK V3 prompt to context-chef IR messages.
+ *
+ * Original AI SDK content is stored in `_originalContent` for lossless round-trip.
+ * `_originalText` caches the extracted text so `toAISDK` can detect Janitor modifications.
+ * `_providerOptions` preserves message-level provider options (e.g. Anthropic cache control).
+ */
+declare function fromAISDK(prompt: LanguageModelV3Prompt): Message[];
+/**
+ * Converts context-chef IR messages back to AI SDK V3 prompt format.
+ *
+ * Uses `_originalContent` when content is 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).
+ */
+declare function toAISDK(messages: Message[]): LanguageModelV3Prompt;
+//#endregion
+//#region src/middleware.d.ts
+/**
+ * Creates a LanguageModelMiddleware that transparently applies
+ * context-chef compression and truncation to AI SDK model calls.
+ *
+ * The middleware holds a stateful Janitor instance that tracks
+ * token usage across calls for compression decisions.
+ */
+declare function createMiddleware(options: ContextChefOptions): LanguageModelMiddleware;
+//#endregion
+//#region src/index.d.ts
+/**
+ * Wraps an AI SDK language model with context-chef middleware for
+ * transparent history compression, tool result truncation, and token budget management.
+ *
+ * @example
+ * ```typescript
+ * import { withContextChef } from '@context-chef/ai-sdk-middleware';
+ * import { openai } from '@ai-sdk/openai';
+ * import { generateText } from 'ai';
+ *
+ * const model = withContextChef(openai('gpt-4o'), {
+ *   contextWindow: 128_000,
+ *   compress: { model: openai('gpt-4o-mini') },
+ *   truncate: { threshold: 5000, headChars: 500, tailChars: 1000 },
+ * });
+ *
+ * // Use exactly like normal — zero other code changes
+ * const result = await generateText({ model, messages, tools });
+ * ```
+ */
+declare function withContextChef(model: LanguageModelV3, options: ContextChefOptions): LanguageModelV3;
+//#endregion
+export { type CompressOptions, type ContextChefOptions, type TruncateOptions, createMiddleware, fromAISDK, toAISDK, withContextChef };
+//# sourceMappingURL=index.d.cts.map

package/dist/index.d.mts

@@ -0,0 +1,92 @@
+import { LanguageModelMiddleware } from "ai";
+import { Message, VFSStorageAdapter } from "@context-chef/core";
+import { LanguageModelV3, LanguageModelV3Prompt } from "@ai-sdk/provider";
+
+//#region src/types.d.ts
+interface TruncateOptions {
+  /** Character count threshold to trigger truncation. */
+  threshold: number;
+  /** Characters to preserve from the start. Default: 0 */
+  headChars?: number;
+  /** Characters to preserve from the end. Default: 1000 */
+  tailChars?: number;
+  /**
+   * Storage adapter for persisting original content before truncation.
+   * Can be a FileSystemAdapter, database adapter, or any custom implementation.
+   * When provided, truncated output includes a `context://vfs/` URI for retrieval.
+   * When omitted, original content is discarded after truncation.
+   */
+  storage?: VFSStorageAdapter;
+}
+interface CompressOptions {
+  /** A cheap model used for summarization (e.g. openai('gpt-4o-mini')). */
+  model: LanguageModelV3;
+  /** Ratio of context window to preserve for recent messages. Default: 0.8 */
+  preserveRatio?: number;
+}
+interface ContextChefOptions {
+  /** The model's context window size in tokens. */
+  contextWindow: number;
+  /** Enable history compression. Omit for no compression. */
+  compress?: CompressOptions;
+  /** Enable tool result truncation. Omit for no truncation. */
+  truncate?: TruncateOptions;
+  /** Optional tokenizer for precise per-message token counting. */
+  tokenizer?: (messages: unknown[]) => number;
+  /** Hook called after compression occurs. */
+  onCompress?: (summary: string, truncatedCount: number) => void;
+}
+//#endregion
+//#region src/adapter.d.ts
+/**
+ * Converts an AI SDK V3 prompt to context-chef IR messages.
+ *
+ * Original AI SDK content is stored in `_originalContent` for lossless round-trip.
+ * `_originalText` caches the extracted text so `toAISDK` can detect Janitor modifications.
+ * `_providerOptions` preserves message-level provider options (e.g. Anthropic cache control).
+ */
+declare function fromAISDK(prompt: LanguageModelV3Prompt): Message[];
+/**
+ * Converts context-chef IR messages back to AI SDK V3 prompt format.
+ *
+ * Uses `_originalContent` when content is 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).
+ */
+declare function toAISDK(messages: Message[]): LanguageModelV3Prompt;
+//#endregion
+//#region src/middleware.d.ts
+/**
+ * Creates a LanguageModelMiddleware that transparently applies
+ * context-chef compression and truncation to AI SDK model calls.
+ *
+ * The middleware holds a stateful Janitor instance that tracks
+ * token usage across calls for compression decisions.
+ */
+declare function createMiddleware(options: ContextChefOptions): LanguageModelMiddleware;
+//#endregion
+//#region src/index.d.ts
+/**
+ * Wraps an AI SDK language model with context-chef middleware for
+ * transparent history compression, tool result truncation, and token budget management.
+ *
+ * @example
+ * ```typescript
+ * import { withContextChef } from '@context-chef/ai-sdk-middleware';
+ * import { openai } from '@ai-sdk/openai';
+ * import { generateText } from 'ai';
+ *
+ * const model = withContextChef(openai('gpt-4o'), {
+ *   contextWindow: 128_000,
+ *   compress: { model: openai('gpt-4o-mini') },
+ *   truncate: { threshold: 5000, headChars: 500, tailChars: 1000 },
+ * });
+ *
+ * // Use exactly like normal — zero other code changes
+ * const result = await generateText({ model, messages, tools });
+ * ```
+ */
+declare function withContextChef(model: LanguageModelV3, options: ContextChefOptions): LanguageModelV3;
+//#endregion
+export { type CompressOptions, type ContextChefOptions, type TruncateOptions, createMiddleware, fromAISDK, toAISDK, withContextChef };
+//# sourceMappingURL=index.d.mts.map

package/dist/index.mjs

@@ -0,0 +1,367 @@
+import { generateText, wrapLanguageModel } from "ai";
+import { Janitor, Offloader } from "@context-chef/core";
+
+//#region src/adapter.ts
+/**
+* Converts an AI SDK V3 prompt to context-chef IR messages.
+*
+* Original AI SDK content is stored in `_originalContent` for lossless round-trip.
+* `_originalText` caches the extracted text so `toAISDK` can detect Janitor modifications.
+* `_providerOptions` preserves message-level provider options (e.g. Anthropic cache control).
+*/
+function fromAISDK(prompt) {
+	const messages = [];
+	for (const msg of prompt) {
+		if (msg.role === "system") {
+			messages.push({
+				role: "system",
+				content: msg.content,
+				...msg.providerOptions ? { _providerOptions: msg.providerOptions } : {}
+			});
+			continue;
+		}
+		if (msg.role === "user") {
+			const text = msg.content.filter((p) => p.type === "text").map((p) => p.text).join("\n");
+			messages.push({
+				role: "user",
+				content: text,
+				_originalContent: msg.content,
+				_originalText: text,
+				...msg.providerOptions ? { _providerOptions: msg.providerOptions } : {}
+			});
+			continue;
+		}
+		if (msg.role === "assistant") {
+			const text = [];
+			const toolCalls = [];
+			let thinking;
+			for (const part of msg.content) if (part.type === "text") text.push(part.text);
+			else if (part.type === "tool-call") toolCalls.push({
+				id: part.toolCallId,
+				type: "function",
+				function: {
+					name: part.toolName,
+					arguments: typeof part.input === "string" ? part.input : JSON.stringify(part.input)
+				}
+			});
+			else if (part.type === "reasoning") thinking = { thinking: part.text };
+			const joinedText = text.join("\n");
+			const m = {
+				role: "assistant",
+				content: joinedText,
+				_originalContent: msg.content,
+				_originalText: joinedText,
+				...msg.providerOptions ? { _providerOptions: msg.providerOptions } : {}
+			};
+			if (toolCalls.length > 0) m.tool_calls = toolCalls;
+			if (thinking) m.thinking = thinking;
+			messages.push(m);
+			continue;
+		}
+		if (msg.role === "tool") {
+			for (const part of msg.content) if (part.type === "tool-result") {
+				const text = stringifyToolOutput(part.output);
+				messages.push({
+					role: "tool",
+					content: text,
+					tool_call_id: part.toolCallId,
+					_originalContent: [part],
+					_originalText: text,
+					_toolName: part.toolName
+				});
+			}
+		}
+	}
+	return messages;
+}
+/**
+* Converts context-chef IR messages back to AI SDK V3 prompt format.
+*
+* Uses `_originalContent` when content is 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).
+*/
+function toAISDK(messages) {
+	const prompt = [];
+	let i = 0;
+	while (i < messages.length) {
+		const msg = messages[i];
+		const providerOptions = msg._providerOptions;
+		const contentModified = msg._originalText !== void 0 && msg._originalText !== msg.content;
+		if (msg.role === "system") {
+			prompt.push({
+				role: "system",
+				content: msg.content,
+				...providerOptions ? { providerOptions } : {}
+			});
+			i++;
+			continue;
+		}
+		if (msg.role === "user") {
+			const content = !contentModified && Array.isArray(msg._originalContent) ? msg._originalContent : [{
+				type: "text",
+				text: msg.content
+			}];
+			prompt.push({
+				role: "user",
+				content,
+				...providerOptions ? { providerOptions } : {}
+			});
+			i++;
+			continue;
+		}
+		if (msg.role === "assistant") {
+			const content = !contentModified && Array.isArray(msg._originalContent) ? msg._originalContent : [{
+				type: "text",
+				text: msg.content
+			}];
+			prompt.push({
+				role: "assistant",
+				content,
+				...providerOptions ? { providerOptions } : {}
+			});
+			i++;
+			continue;
+		}
+		if (msg.role === "tool") {
+			const toolResults = [];
+			while (i < messages.length && messages[i].role === "tool") {
+				const toolMsg = messages[i];
+				if (!(toolMsg._originalText !== void 0 && toolMsg._originalText !== toolMsg.content) && toolMsg._originalContent) toolResults.push(...toolMsg._originalContent);
+				else toolResults.push({
+					type: "tool-result",
+					toolCallId: toolMsg.tool_call_id ?? "",
+					toolName: toolMsg._toolName ?? "unknown",
+					output: {
+						type: "text",
+						value: toolMsg.content
+					}
+				});
+				i++;
+			}
+			prompt.push({
+				role: "tool",
+				content: toolResults
+			});
+			continue;
+		}
+		i++;
+	}
+	return prompt;
+}
+function stringifyToolOutput(output) {
+	switch (output.type) {
+		case "text":
+		case "error-text": return output.value;
+		case "json":
+		case "error-json": return JSON.stringify(output.value);
+		case "content": return output.value.map((v) => v.type === "text" ? v.text ?? "" : "").filter(Boolean).join("\n");
+		default: return JSON.stringify(output);
+	}
+}
+
+//#endregion
+//#region src/truncator.ts
+/**
+* Truncates tool-result content within an AI SDK prompt when it exceeds the configured threshold.
+* When a storage adapter is provided, original content is persisted and a URI is included in the output.
+*/
+async function truncateToolResults(prompt, options) {
+	const { threshold, headChars = 0, tailChars = 1e3, storage } = options;
+	const offloader = storage ? new Offloader({
+		threshold,
+		adapter: storage,
+		storageDir: ""
+	}) : null;
+	const result = [];
+	for (const msg of prompt) {
+		if (msg.role !== "tool") {
+			result.push(msg);
+			continue;
+		}
+		const newContent = [];
+		for (const part of msg.content) {
+			if (part.type !== "tool-result") {
+				newContent.push(part);
+				continue;
+			}
+			const text = extractText(part.output);
+			if (text.length <= threshold || headChars + tailChars >= text.length) {
+				newContent.push(part);
+				continue;
+			}
+			if (offloader) try {
+				const vfsResult = await offloader.offloadAsync(text, {
+					threshold,
+					headChars,
+					tailChars
+				});
+				newContent.push({
+					...part,
+					output: {
+						type: "text",
+						value: vfsResult.content
+					}
+				});
+				continue;
+			} catch (error) {
+				console.warn(`[context-chef] Storage adapter write failed for tool result (${part.toolCallId}). Falling back to simple truncation. Error: ${error instanceof Error ? error.message : String(error)}`);
+			}
+			const head = text.slice(0, headChars);
+			const tail = text.slice(text.length - tailChars);
+			const truncated = [
+				head,
+				`\n--- truncated (${text.split("\n").length} lines, ${text.length} chars total) ---\n`,
+				tail
+			].filter(Boolean).join("").trim();
+			newContent.push({
+				...part,
+				output: {
+					type: "text",
+					value: truncated
+				}
+			});
+		}
+		result.push({
+			...msg,
+			content: newContent
+		});
+	}
+	return result;
+}
+function extractText(output) {
+	switch (output.type) {
+		case "text":
+		case "error-text": return output.value;
+		case "json":
+		case "error-json": return JSON.stringify(output.value);
+		case "content": return output.value.map((v) => v.type === "text" ? v.text ?? "" : "").filter(Boolean).join("\n");
+		default: return "";
+	}
+}
+
+//#endregion
+//#region src/middleware.ts
+/**
+* Creates a LanguageModelMiddleware that transparently applies
+* context-chef compression and truncation to AI SDK model calls.
+*
+* The middleware holds a stateful Janitor instance that tracks
+* token usage across calls for compression decisions.
+*/
+function createMiddleware(options) {
+	let usageWarned = false;
+	const janitor = new Janitor({
+		contextWindow: options.contextWindow,
+		tokenizer: options.tokenizer ? (msgs) => options.tokenizer?.(msgs) ?? 0 : void 0,
+		preserveRatio: options.compress?.preserveRatio ?? .8,
+		compressionModel: options.compress?.model ? createCompressionAdapter(options.compress.model) : void 0,
+		onCompress: options.onCompress ? (summary, count) => options.onCompress?.(summary.content, count) : void 0
+	});
+	return {
+		specificationVersion: "v3",
+		transformParams: async ({ params }) => {
+			let { prompt } = params;
+			if (options.truncate) prompt = await truncateToolResults(prompt, options.truncate);
+			const irMessages = fromAISDK(prompt);
+			const compressed = await janitor.compress(irMessages);
+			if (compressed !== irMessages) prompt = toAISDK(compressed);
+			return {
+				...params,
+				prompt
+			};
+		},
+		wrapGenerate: async ({ doGenerate }) => {
+			const result = await doGenerate();
+			if (result.usage?.inputTokens?.total != null) janitor.feedTokenUsage(result.usage.inputTokens.total);
+			else if (!usageWarned && !options.tokenizer) {
+				usageWarned = true;
+				console.warn("[context-chef] Model response did not include usage.inputTokens.total. Token-based compression may not trigger accurately. Consider providing a tokenizer for precise token counting.");
+			}
+			return result;
+		},
+		wrapStream: async ({ doStream }) => {
+			const { stream, ...rest } = await doStream();
+			const transform = new TransformStream({ transform(chunk, controller) {
+				if (chunk.type === "finish") {
+					if (chunk.usage?.inputTokens?.total != null) janitor.feedTokenUsage(chunk.usage.inputTokens.total);
+					else if (!usageWarned && !options.tokenizer) {
+						usageWarned = true;
+						console.warn("[context-chef] Stream finish did not include usage.inputTokens.total. Token-based compression may not trigger accurately. Consider providing a tokenizer for precise token counting.");
+					}
+				}
+				controller.enqueue(chunk);
+			} });
+			return {
+				...rest,
+				stream: stream.pipeThrough(transform)
+			};
+		}
+	};
+}
+/**
+* Adapts an AI SDK LanguageModelV3 into the compressionModel callback
+* that Janitor expects: (messages: Message[]) => Promise<string>
+*
+* Tool messages are converted to user messages describing the tool interaction,
+* since generateText only accepts system/user/assistant roles.
+*/
+function createCompressionAdapter(model) {
+	return async (messages) => {
+		const { text } = await generateText({
+			model,
+			messages: messages.map((m) => {
+				if (m.role === "tool") return {
+					role: "user",
+					content: `[Tool result${m.tool_call_id ? ` (${m.tool_call_id})` : ""}: ${m.content}]`
+				};
+				if (m.role === "assistant" && m.tool_calls?.length) {
+					const toolCallsDesc = m.tool_calls.map((tc) => `[Called tool: ${tc.function.name}(${tc.function.arguments})]`).join("\n");
+					return {
+						role: "assistant",
+						content: m.content ? `${m.content}\n${toolCallsDesc}` : toolCallsDesc
+					};
+				}
+				return {
+					role: m.role,
+					content: m.content
+				};
+			}),
+			maxOutputTokens: 2048
+		});
+		return text || "[Compression produced no output]";
+	};
+}
+
+//#endregion
+//#region src/index.ts
+/**
+* Wraps an AI SDK language model with context-chef middleware for
+* transparent history compression, tool result truncation, and token budget management.
+*
+* @example
+* ```typescript
+* import { withContextChef } from '@context-chef/ai-sdk-middleware';
+* import { openai } from '@ai-sdk/openai';
+* import { generateText } from 'ai';
+*
+* const model = withContextChef(openai('gpt-4o'), {
+*   contextWindow: 128_000,
+*   compress: { model: openai('gpt-4o-mini') },
+*   truncate: { threshold: 5000, headChars: 500, tailChars: 1000 },
+* });
+*
+* // Use exactly like normal — zero other code changes
+* const result = await generateText({ model, messages, tools });
+* ```
+*/
+function withContextChef(model, options) {
+	return wrapLanguageModel({
+		model,
+		middleware: createMiddleware(options)
+	});
+}
+
+//#endregion
+export { createMiddleware, fromAISDK, toAISDK, withContextChef };
+//# sourceMappingURL=index.mjs.map

package/dist/index.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.mjs","names":[],"sources":["../src/adapter.ts","../src/truncator.ts","../src/middleware.ts","../src/index.ts"],"sourcesContent":["import type {\n  LanguageModelV3Prompt,\n  LanguageModelV3ToolResultOutput,\n  LanguageModelV3ToolResultPart,\n  SharedV3ProviderOptions,\n} from '@ai-sdk/provider';\nimport type { Message, ToolCall } from '@context-chef/core';\n\n/**\n * Converts an AI SDK V3 prompt to context-chef IR messages.\n *\n * Original AI SDK content is stored in `_originalContent` for lossless round-trip.\n * `_originalText` caches the extracted text so `toAISDK` can detect Janitor modifications.\n * `_providerOptions` preserves message-level provider options (e.g. Anthropic cache control).\n */\nexport function fromAISDK(prompt: LanguageModelV3Prompt): Message[] {\n  const messages: Message[] = [];\n\n  for (const msg of prompt) {\n    if (msg.role === 'system') {\n      messages.push({\n        role: 'system',\n        content: msg.content,\n        ...(msg.providerOptions ? { _providerOptions: msg.providerOptions } : {}),\n      });\n      continue;\n    }\n\n    if (msg.role === 'user') {\n      const text = msg.content\n        .filter((p) => p.type === 'text')\n        .map((p) => p.text)\n        .join('\\n');\n      messages.push({\n        role: 'user',\n        content: text,\n        _originalContent: msg.content,\n        _originalText: text,\n        ...(msg.providerOptions ? { _providerOptions: msg.providerOptions } : {}),\n      });\n      continue;\n    }\n\n    if (msg.role === 'assistant') {\n      const text: string[] = [];\n      const toolCalls: ToolCall[] = [];\n      let thinking: { thinking: string } | undefined;\n\n      for (const part of msg.content) {\n        if (part.type === 'text') text.push(part.text);\n        else if (part.type === 'tool-call') {\n          toolCalls.push({\n            id: part.toolCallId,\n            type: 'function',\n            function: {\n              name: part.toolName,\n              arguments: typeof part.input === 'string' ? part.input : JSON.stringify(part.input),\n            },\n          });\n        } else if (part.type === 'reasoning') {\n          thinking = { thinking: part.text };\n        }\n      }\n\n      const joinedText = text.join('\\n');\n      const m: Message = {\n        role: 'assistant',\n        content: joinedText,\n        _originalContent: msg.content,\n        _originalText: joinedText,\n        ...(msg.providerOptions ? { _providerOptions: msg.providerOptions } : {}),\n      };\n      if (toolCalls.length > 0) m.tool_calls = toolCalls;\n      if (thinking) m.thinking = thinking;\n      messages.push(m);\n      continue;\n    }\n\n    if (msg.role === 'tool') {\n      for (const part of msg.content) {\n        if (part.type === 'tool-result') {\n          const text = stringifyToolOutput(part.output);\n          messages.push({\n            role: 'tool',\n            content: text,\n            tool_call_id: part.toolCallId,\n            _originalContent: [part],\n            _originalText: text,\n            _toolName: part.toolName,\n          });\n        }\n      }\n    }\n  }\n\n  return messages;\n}\n\n/**\n * Converts context-chef IR messages back to AI SDK V3 prompt format.\n *\n * Uses `_originalContent` when content is unmodified (detected via `_originalText`).\n * Falls back to constructing from IR fields when content was modified by Janitor\n * (e.g. compact() cleared tool results) or for new messages (e.g. compression summaries).\n */\nexport function toAISDK(messages: Message[]): LanguageModelV3Prompt {\n  const prompt: LanguageModelV3Prompt = [];\n\n  let i = 0;\n  while (i < messages.length) {\n    const msg = messages[i];\n    const providerOptions = msg._providerOptions as SharedV3ProviderOptions | undefined;\n    const contentModified = msg._originalText !== undefined && msg._originalText !== msg.content;\n\n    if (msg.role === 'system') {\n      prompt.push({\n        role: 'system',\n        content: msg.content,\n        ...(providerOptions ? { providerOptions } : {}),\n      });\n      i++;\n      continue;\n    }\n\n    if (msg.role === 'user') {\n      const content =\n        !contentModified && Array.isArray(msg._originalContent)\n          ? (msg._originalContent as any)\n          : [{ type: 'text' as const, text: msg.content }];\n      prompt.push({\n        role: 'user',\n        content,\n        ...(providerOptions ? { providerOptions } : {}),\n      });\n      i++;\n      continue;\n    }\n\n    if (msg.role === 'assistant') {\n      const content =\n        !contentModified && Array.isArray(msg._originalContent)\n          ? (msg._originalContent as any)\n          : [{ type: 'text' as const, text: msg.content }];\n      prompt.push({\n        role: 'assistant',\n        content,\n        ...(providerOptions ? { providerOptions } : {}),\n      });\n      i++;\n      continue;\n    }\n\n    if (msg.role === 'tool') {\n      const toolResults: LanguageModelV3ToolResultPart[] = [];\n      while (i < messages.length && messages[i].role === 'tool') {\n        const toolMsg = messages[i];\n        const toolModified =\n          toolMsg._originalText !== undefined && toolMsg._originalText !== toolMsg.content;\n\n        if (!toolModified && toolMsg._originalContent) {\n          toolResults.push(...(toolMsg._originalContent as LanguageModelV3ToolResultPart[]));\n        } else {\n          toolResults.push({\n            type: 'tool-result',\n            toolCallId: toolMsg.tool_call_id ?? '',\n            toolName: (toolMsg._toolName as string) ?? 'unknown',\n            output: { type: 'text', value: toolMsg.content },\n          });\n        }\n        i++;\n      }\n      prompt.push({ role: 'tool', content: toolResults });\n      continue;\n    }\n\n    i++;\n  }\n\n  return prompt;\n}\n\nfunction stringifyToolOutput(output: LanguageModelV3ToolResultOutput): string {\n  switch (output.type) {\n    case 'text':\n    case 'error-text':\n      return output.value;\n    case 'json':\n    case 'error-json':\n      return JSON.stringify(output.value);\n    case 'content':\n      return output.value\n        .map((v: { type: string; text?: string }) => (v.type === 'text' ? (v.text ?? '') : ''))\n        .filter(Boolean)\n        .join('\\n');\n    default:\n      return JSON.stringify(output);\n  }\n}\n","import type {\n  LanguageModelV3Prompt,\n  LanguageModelV3ToolResultOutput,\n  LanguageModelV3ToolResultPart,\n} from '@ai-sdk/provider';\nimport { Offloader } from '@context-chef/core';\nimport type { TruncateOptions } from './types';\n\n/**\n * Truncates tool-result content within an AI SDK prompt when it exceeds the configured threshold.\n * When a storage adapter is provided, original content is persisted and a URI is included in the output.\n */\nexport async function truncateToolResults(\n  prompt: LanguageModelV3Prompt,\n  options: TruncateOptions,\n): Promise<LanguageModelV3Prompt> {\n  const { threshold, headChars = 0, tailChars = 1000, storage } = options;\n\n  const offloader = storage\n    ? new Offloader({ threshold, adapter: storage, storageDir: '' })\n    : null;\n\n  const result: LanguageModelV3Prompt = [];\n\n  for (const msg of prompt) {\n    if (msg.role !== 'tool') {\n      result.push(msg);\n      continue;\n    }\n\n    const newContent: typeof msg.content = [];\n\n    for (const part of msg.content) {\n      if (part.type !== 'tool-result') {\n        newContent.push(part);\n        continue;\n      }\n\n      const text = extractText(part.output);\n      if (text.length <= threshold || headChars + tailChars >= text.length) {\n        newContent.push(part);\n        continue;\n      }\n\n      // With storage: use Offloader to persist original and get a URI-annotated truncation\n      if (offloader) {\n        try {\n          const vfsResult = await offloader.offloadAsync(text, { threshold, headChars, tailChars });\n          newContent.push({\n            ...part,\n            output: { type: 'text', value: vfsResult.content } satisfies LanguageModelV3ToolResultOutput,\n          } satisfies LanguageModelV3ToolResultPart);\n          continue;\n        } catch (error) {\n          console.warn(\n            `[context-chef] Storage adapter write failed for tool result (${part.toolCallId}). ` +\n              `Falling back to simple truncation. Error: ${error instanceof Error ? error.message : String(error)}`,\n          );\n          // Fall through to simple truncation below\n        }\n      }\n\n      // Without storage: simple truncation, original is discarded\n      const head = text.slice(0, headChars);\n      const tail = text.slice(text.length - tailChars);\n      const totalLines = text.split('\\n').length;\n\n      const truncated = [\n        head,\n        `\\n--- truncated (${totalLines} lines, ${text.length} chars total) ---\\n`,\n        tail,\n      ]\n        .filter(Boolean)\n        .join('')\n        .trim();\n\n      newContent.push({\n        ...part,\n        output: { type: 'text', value: truncated } satisfies LanguageModelV3ToolResultOutput,\n      } satisfies LanguageModelV3ToolResultPart);\n    }\n\n    result.push({ ...msg, content: newContent });\n  }\n\n  return result;\n}\n\nfunction extractText(output: LanguageModelV3ToolResultOutput): string {\n  switch (output.type) {\n    case 'text':\n    case 'error-text':\n      return output.value;\n    case 'json':\n    case 'error-json':\n      return JSON.stringify(output.value);\n    case 'content':\n      return output.value\n        .map((v: { type: string; text?: string }) => (v.type === 'text' ? (v.text ?? '') : ''))\n        .filter(Boolean)\n        .join('\\n');\n    default:\n      return '';\n  }\n}\n","import type { LanguageModelV3, LanguageModelV3StreamPart } from '@ai-sdk/provider';\nimport { generateText, type LanguageModelMiddleware } from 'ai';\nimport { Janitor, type Message } from '@context-chef/core';\n\nimport { fromAISDK, toAISDK } from './adapter';\nimport { truncateToolResults } from './truncator';\nimport type { ContextChefOptions } from './types';\n\n/**\n * Creates a LanguageModelMiddleware that transparently applies\n * context-chef compression and truncation to AI SDK model calls.\n *\n * The middleware holds a stateful Janitor instance that tracks\n * token usage across calls for compression decisions.\n */\nexport function createMiddleware(options: ContextChefOptions): LanguageModelMiddleware {\n  let usageWarned = false;\n\n  const janitor = new Janitor({\n    contextWindow: options.contextWindow,\n    tokenizer: options.tokenizer ? (msgs: Message[]) => options.tokenizer?.(msgs) ?? 0 : undefined,\n    preserveRatio: options.compress?.preserveRatio ?? 0.8,\n    compressionModel: options.compress?.model\n      ? createCompressionAdapter(options.compress.model)\n      : undefined,\n    onCompress: options.onCompress\n      ? (summary, count) => options.onCompress?.(summary.content, count)\n      : undefined,\n  });\n\n  return {\n    specificationVersion: 'v3',\n\n    transformParams: async ({ params }) => {\n      let { prompt } = params;\n\n      // 1. Truncate large tool results\n      if (options.truncate) {\n        prompt = await truncateToolResults(prompt, options.truncate);\n      }\n\n      // 2. Compress history if over token budget\n      const irMessages = fromAISDK(prompt);\n      const compressed = await janitor.compress(irMessages);\n\n      // Only convert back if compression actually changed something\n      if (compressed !== irMessages) {\n        prompt = toAISDK(compressed);\n      }\n\n      return { ...params, prompt };\n    },\n\n    wrapGenerate: async ({ doGenerate }) => {\n      const result = await doGenerate();\n\n      if (result.usage?.inputTokens?.total != null) {\n        janitor.feedTokenUsage(result.usage.inputTokens.total);\n      } else if (!usageWarned && !options.tokenizer) {\n        usageWarned = true;\n        console.warn(\n          '[context-chef] Model response did not include usage.inputTokens.total. ' +\n            'Token-based compression may not trigger accurately. ' +\n            'Consider providing a tokenizer for precise token counting.',\n        );\n      }\n\n      return result;\n    },\n\n    wrapStream: async ({ doStream }) => {\n      const { stream, ...rest } = await doStream();\n\n      const transform = new TransformStream<LanguageModelV3StreamPart, LanguageModelV3StreamPart>({\n        transform(chunk, controller) {\n          if (chunk.type === 'finish') {\n            if (chunk.usage?.inputTokens?.total != null) {\n              janitor.feedTokenUsage(chunk.usage.inputTokens.total);\n            } else if (!usageWarned && !options.tokenizer) {\n              usageWarned = true;\n              console.warn(\n                '[context-chef] Stream finish did not include usage.inputTokens.total. ' +\n                  'Token-based compression may not trigger accurately. ' +\n                  'Consider providing a tokenizer for precise token counting.',\n              );\n            }\n          }\n          controller.enqueue(chunk);\n        },\n      });\n\n      return { ...rest, stream: stream.pipeThrough(transform) };\n    },\n  };\n}\n\n/**\n * Adapts an AI SDK LanguageModelV3 into the compressionModel callback\n * that Janitor expects: (messages: Message[]) => Promise<string>\n *\n * Tool messages are converted to user messages describing the tool interaction,\n * since generateText only accepts system/user/assistant roles.\n */\nfunction createCompressionAdapter(model: LanguageModelV3): (messages: Message[]) => Promise<string> {\n  return async (messages: Message[]): Promise<string> => {\n    const formatted = messages.map((m) => {\n      if (m.role === 'tool') {\n        return {\n          role: 'user' as const,\n          content: `[Tool result${m.tool_call_id ? ` (${m.tool_call_id})` : ''}: ${m.content}]`,\n        };\n      }\n      // assistant messages with tool_calls: include tool call info in content\n      if (m.role === 'assistant' && m.tool_calls?.length) {\n        const toolCallsDesc = m.tool_calls\n          .map((tc) => `[Called tool: ${tc.function.name}(${tc.function.arguments})]`)\n          .join('\\n');\n        return {\n          role: 'assistant' as const,\n          content: m.content ? `${m.content}\\n${toolCallsDesc}` : toolCallsDesc,\n        };\n      }\n      return {\n        role: m.role as 'system' | 'user' | 'assistant',\n        content: m.content,\n      };\n    });\n\n    const { text } = await generateText({\n      model,\n      messages: formatted,\n      maxOutputTokens: 2048,\n    });\n\n    return text || '[Compression produced no output]';\n  };\n}\n","import type { LanguageModelV3 } from '@ai-sdk/provider';\nimport { wrapLanguageModel } from 'ai';\n\nimport { createMiddleware } from './middleware';\nimport type { ContextChefOptions } from './types';\n\nexport { fromAISDK, toAISDK } from './adapter';\nexport { createMiddleware } from './middleware';\nexport type { CompressOptions, ContextChefOptions, TruncateOptions } from './types';\n\n/**\n * Wraps an AI SDK language model with context-chef middleware for\n * transparent history compression, tool result truncation, and token budget management.\n *\n * @example\n * ```typescript\n * import { withContextChef } from '@context-chef/ai-sdk-middleware';\n * import { openai } from '@ai-sdk/openai';\n * import { generateText } from 'ai';\n *\n * const model = withContextChef(openai('gpt-4o'), {\n *   contextWindow: 128_000,\n *   compress: { model: openai('gpt-4o-mini') },\n *   truncate: { threshold: 5000, headChars: 500, tailChars: 1000 },\n * });\n *\n * // Use exactly like normal — zero other code changes\n * const result = await generateText({ model, messages, tools });\n * ```\n */\nexport function withContextChef(model: LanguageModelV3, options: ContextChefOptions): LanguageModelV3 {\n  const middleware = createMiddleware(options);\n  return wrapLanguageModel({ model, middleware });\n}\n"],"mappings":";;;;;;;;;;;AAeA,SAAgB,UAAU,QAA0C;CAClE,MAAM,WAAsB,EAAE;AAE9B,MAAK,MAAM,OAAO,QAAQ;AACxB,MAAI,IAAI,SAAS,UAAU;AACzB,YAAS,KAAK;IACZ,MAAM;IACN,SAAS,IAAI;IACb,GAAI,IAAI,kBAAkB,EAAE,kBAAkB,IAAI,iBAAiB,GAAG,EAAE;IACzE,CAAC;AACF;;AAGF,MAAI,IAAI,SAAS,QAAQ;GACvB,MAAM,OAAO,IAAI,QACd,QAAQ,MAAM,EAAE,SAAS,OAAO,CAChC,KAAK,MAAM,EAAE,KAAK,CAClB,KAAK,KAAK;AACb,YAAS,KAAK;IACZ,MAAM;IACN,SAAS;IACT,kBAAkB,IAAI;IACtB,eAAe;IACf,GAAI,IAAI,kBAAkB,EAAE,kBAAkB,IAAI,iBAAiB,GAAG,EAAE;IACzE,CAAC;AACF;;AAGF,MAAI,IAAI,SAAS,aAAa;GAC5B,MAAM,OAAiB,EAAE;GACzB,MAAM,YAAwB,EAAE;GAChC,IAAI;AAEJ,QAAK,MAAM,QAAQ,IAAI,QACrB,KAAI,KAAK,SAAS,OAAQ,MAAK,KAAK,KAAK,KAAK;YACrC,KAAK,SAAS,YACrB,WAAU,KAAK;IACb,IAAI,KAAK;IACT,MAAM;IACN,UAAU;KACR,MAAM,KAAK;KACX,WAAW,OAAO,KAAK,UAAU,WAAW,KAAK,QAAQ,KAAK,UAAU,KAAK,MAAM;KACpF;IACF,CAAC;YACO,KAAK,SAAS,YACvB,YAAW,EAAE,UAAU,KAAK,MAAM;GAItC,MAAM,aAAa,KAAK,KAAK,KAAK;GAClC,MAAM,IAAa;IACjB,MAAM;IACN,SAAS;IACT,kBAAkB,IAAI;IACtB,eAAe;IACf,GAAI,IAAI,kBAAkB,EAAE,kBAAkB,IAAI,iBAAiB,GAAG,EAAE;IACzE;AACD,OAAI,UAAU,SAAS,EAAG,GAAE,aAAa;AACzC,OAAI,SAAU,GAAE,WAAW;AAC3B,YAAS,KAAK,EAAE;AAChB;;AAGF,MAAI,IAAI,SAAS,QACf;QAAK,MAAM,QAAQ,IAAI,QACrB,KAAI,KAAK,SAAS,eAAe;IAC/B,MAAM,OAAO,oBAAoB,KAAK,OAAO;AAC7C,aAAS,KAAK;KACZ,MAAM;KACN,SAAS;KACT,cAAc,KAAK;KACnB,kBAAkB,CAAC,KAAK;KACxB,eAAe;KACf,WAAW,KAAK;KACjB,CAAC;;;;AAMV,QAAO;;;;;;;;;AAUT,SAAgB,QAAQ,UAA4C;CAClE,MAAM,SAAgC,EAAE;CAExC,IAAI,IAAI;AACR,QAAO,IAAI,SAAS,QAAQ;EAC1B,MAAM,MAAM,SAAS;EACrB,MAAM,kBAAkB,IAAI;EAC5B,MAAM,kBAAkB,IAAI,kBAAkB,UAAa,IAAI,kBAAkB,IAAI;AAErF,MAAI,IAAI,SAAS,UAAU;AACzB,UAAO,KAAK;IACV,MAAM;IACN,SAAS,IAAI;IACb,GAAI,kBAAkB,EAAE,iBAAiB,GAAG,EAAE;IAC/C,CAAC;AACF;AACA;;AAGF,MAAI,IAAI,SAAS,QAAQ;GACvB,MAAM,UACJ,CAAC,mBAAmB,MAAM,QAAQ,IAAI,iBAAiB,GAClD,IAAI,mBACL,CAAC;IAAE,MAAM;IAAiB,MAAM,IAAI;IAAS,CAAC;AACpD,UAAO,KAAK;IACV,MAAM;IACN;IACA,GAAI,kBAAkB,EAAE,iBAAiB,GAAG,EAAE;IAC/C,CAAC;AACF;AACA;;AAGF,MAAI,IAAI,SAAS,aAAa;GAC5B,MAAM,UACJ,CAAC,mBAAmB,MAAM,QAAQ,IAAI,iBAAiB,GAClD,IAAI,mBACL,CAAC;IAAE,MAAM;IAAiB,MAAM,IAAI;IAAS,CAAC;AACpD,UAAO,KAAK;IACV,MAAM;IACN;IACA,GAAI,kBAAkB,EAAE,iBAAiB,GAAG,EAAE;IAC/C,CAAC;AACF;AACA;;AAGF,MAAI,IAAI,SAAS,QAAQ;GACvB,MAAM,cAA+C,EAAE;AACvD,UAAO,IAAI,SAAS,UAAU,SAAS,GAAG,SAAS,QAAQ;IACzD,MAAM,UAAU,SAAS;AAIzB,QAAI,EAFF,QAAQ,kBAAkB,UAAa,QAAQ,kBAAkB,QAAQ,YAEtD,QAAQ,iBAC3B,aAAY,KAAK,GAAI,QAAQ,iBAAqD;QAElF,aAAY,KAAK;KACf,MAAM;KACN,YAAY,QAAQ,gBAAgB;KACpC,UAAW,QAAQ,aAAwB;KAC3C,QAAQ;MAAE,MAAM;MAAQ,OAAO,QAAQ;MAAS;KACjD,CAAC;AAEJ;;AAEF,UAAO,KAAK;IAAE,MAAM;IAAQ,SAAS;IAAa,CAAC;AACnD;;AAGF;;AAGF,QAAO;;AAGT,SAAS,oBAAoB,QAAiD;AAC5E,SAAQ,OAAO,MAAf;EACE,KAAK;EACL,KAAK,aACH,QAAO,OAAO;EAChB,KAAK;EACL,KAAK,aACH,QAAO,KAAK,UAAU,OAAO,MAAM;EACrC,KAAK,UACH,QAAO,OAAO,MACX,KAAK,MAAwC,EAAE,SAAS,SAAU,EAAE,QAAQ,KAAM,GAAI,CACtF,OAAO,QAAQ,CACf,KAAK,KAAK;EACf,QACE,QAAO,KAAK,UAAU,OAAO;;;;;;;;;;ACvLnC,eAAsB,oBACpB,QACA,SACgC;CAChC,MAAM,EAAE,WAAW,YAAY,GAAG,YAAY,KAAM,YAAY;CAEhE,MAAM,YAAY,UACd,IAAI,UAAU;EAAE;EAAW,SAAS;EAAS,YAAY;EAAI,CAAC,GAC9D;CAEJ,MAAM,SAAgC,EAAE;AAExC,MAAK,MAAM,OAAO,QAAQ;AACxB,MAAI,IAAI,SAAS,QAAQ;AACvB,UAAO,KAAK,IAAI;AAChB;;EAGF,MAAM,aAAiC,EAAE;AAEzC,OAAK,MAAM,QAAQ,IAAI,SAAS;AAC9B,OAAI,KAAK,SAAS,eAAe;AAC/B,eAAW,KAAK,KAAK;AACrB;;GAGF,MAAM,OAAO,YAAY,KAAK,OAAO;AACrC,OAAI,KAAK,UAAU,aAAa,YAAY,aAAa,KAAK,QAAQ;AACpE,eAAW,KAAK,KAAK;AACrB;;AAIF,OAAI,UACF,KAAI;IACF,MAAM,YAAY,MAAM,UAAU,aAAa,MAAM;KAAE;KAAW;KAAW;KAAW,CAAC;AACzF,eAAW,KAAK;KACd,GAAG;KACH,QAAQ;MAAE,MAAM;MAAQ,OAAO,UAAU;MAAS;KACnD,CAAyC;AAC1C;YACO,OAAO;AACd,YAAQ,KACN,gEAAgE,KAAK,WAAW,+CACjC,iBAAiB,QAAQ,MAAM,UAAU,OAAO,MAAM,GACtG;;GAML,MAAM,OAAO,KAAK,MAAM,GAAG,UAAU;GACrC,MAAM,OAAO,KAAK,MAAM,KAAK,SAAS,UAAU;GAGhD,MAAM,YAAY;IAChB;IACA,oBAJiB,KAAK,MAAM,KAAK,CAAC,OAIH,UAAU,KAAK,OAAO;IACrD;IACD,CACE,OAAO,QAAQ,CACf,KAAK,GAAG,CACR,MAAM;AAET,cAAW,KAAK;IACd,GAAG;IACH,QAAQ;KAAE,MAAM;KAAQ,OAAO;KAAW;IAC3C,CAAyC;;AAG5C,SAAO,KAAK;GAAE,GAAG;GAAK,SAAS;GAAY,CAAC;;AAG9C,QAAO;;AAGT,SAAS,YAAY,QAAiD;AACpE,SAAQ,OAAO,MAAf;EACE,KAAK;EACL,KAAK,aACH,QAAO,OAAO;EAChB,KAAK;EACL,KAAK,aACH,QAAO,KAAK,UAAU,OAAO,MAAM;EACrC,KAAK,UACH,QAAO,OAAO,MACX,KAAK,MAAwC,EAAE,SAAS,SAAU,EAAE,QAAQ,KAAM,GAAI,CACtF,OAAO,QAAQ,CACf,KAAK,KAAK;EACf,QACE,QAAO;;;;;;;;;;;;;ACvFb,SAAgB,iBAAiB,SAAsD;CACrF,IAAI,cAAc;CAElB,MAAM,UAAU,IAAI,QAAQ;EAC1B,eAAe,QAAQ;EACvB,WAAW,QAAQ,aAAa,SAAoB,QAAQ,YAAY,KAAK,IAAI,IAAI;EACrF,eAAe,QAAQ,UAAU,iBAAiB;EAClD,kBAAkB,QAAQ,UAAU,QAChC,yBAAyB,QAAQ,SAAS,MAAM,GAChD;EACJ,YAAY,QAAQ,cACf,SAAS,UAAU,QAAQ,aAAa,QAAQ,SAAS,MAAM,GAChE;EACL,CAAC;AAEF,QAAO;EACL,sBAAsB;EAEtB,iBAAiB,OAAO,EAAE,aAAa;GACrC,IAAI,EAAE,WAAW;AAGjB,OAAI,QAAQ,SACV,UAAS,MAAM,oBAAoB,QAAQ,QAAQ,SAAS;GAI9D,MAAM,aAAa,UAAU,OAAO;GACpC,MAAM,aAAa,MAAM,QAAQ,SAAS,WAAW;AAGrD,OAAI,eAAe,WACjB,UAAS,QAAQ,WAAW;AAG9B,UAAO;IAAE,GAAG;IAAQ;IAAQ;;EAG9B,cAAc,OAAO,EAAE,iBAAiB;GACtC,MAAM,SAAS,MAAM,YAAY;AAEjC,OAAI,OAAO,OAAO,aAAa,SAAS,KACtC,SAAQ,eAAe,OAAO,MAAM,YAAY,MAAM;YAC7C,CAAC,eAAe,CAAC,QAAQ,WAAW;AAC7C,kBAAc;AACd,YAAQ,KACN,wLAGD;;AAGH,UAAO;;EAGT,YAAY,OAAO,EAAE,eAAe;GAClC,MAAM,EAAE,QAAQ,GAAG,SAAS,MAAM,UAAU;GAE5C,MAAM,YAAY,IAAI,gBAAsE,EAC1F,UAAU,OAAO,YAAY;AAC3B,QAAI,MAAM,SAAS,UACjB;SAAI,MAAM,OAAO,aAAa,SAAS,KACrC,SAAQ,eAAe,MAAM,MAAM,YAAY,MAAM;cAC5C,CAAC,eAAe,CAAC,QAAQ,WAAW;AAC7C,oBAAc;AACd,cAAQ,KACN,uLAGD;;;AAGL,eAAW,QAAQ,MAAM;MAE5B,CAAC;AAEF,UAAO;IAAE,GAAG;IAAM,QAAQ,OAAO,YAAY,UAAU;IAAE;;EAE5D;;;;;;;;;AAUH,SAAS,yBAAyB,OAAkE;AAClG,QAAO,OAAO,aAAyC;EAwBrD,MAAM,EAAE,SAAS,MAAM,aAAa;GAClC;GACA,UAzBgB,SAAS,KAAK,MAAM;AACpC,QAAI,EAAE,SAAS,OACb,QAAO;KACL,MAAM;KACN,SAAS,eAAe,EAAE,eAAe,KAAK,EAAE,aAAa,KAAK,GAAG,IAAI,EAAE,QAAQ;KACpF;AAGH,QAAI,EAAE,SAAS,eAAe,EAAE,YAAY,QAAQ;KAClD,MAAM,gBAAgB,EAAE,WACrB,KAAK,OAAO,iBAAiB,GAAG,SAAS,KAAK,GAAG,GAAG,SAAS,UAAU,IAAI,CAC3E,KAAK,KAAK;AACb,YAAO;MACL,MAAM;MACN,SAAS,EAAE,UAAU,GAAG,EAAE,QAAQ,IAAI,kBAAkB;MACzD;;AAEH,WAAO;KACL,MAAM,EAAE;KACR,SAAS,EAAE;KACZ;KACD;GAKA,iBAAiB;GAClB,CAAC;AAEF,SAAO,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;ACxGnB,SAAgB,gBAAgB,OAAwB,SAA8C;AAEpG,QAAO,kBAAkB;EAAE;EAAO,YADf,iBAAiB,QAAQ;EACE,CAAC"}

package/package.json

@@ -0,0 +1,60 @@
+{
+  "name": "@context-chef/ai-sdk-middleware",
+  "version": "0.1.0",
+  "type": "module",
+  "main": "./dist/index.cjs",
+  "module": "./dist/index.mjs",
+  "types": "./dist/index.d.mts",
+  "exports": {
+    ".": {
+      "types": "./dist/index.d.mts",
+      "import": "./dist/index.mjs",
+      "require": "./dist/index.cjs"
+    }
+  },
+  "files": [
+    "dist"
+  ],
+  "scripts": {
+    "build": "tsdown",
+    "test": "vitest run",
+    "typecheck": "tsc -p tsconfig.build.json --noEmit"
+  },
+  "keywords": [
+    "ai-sdk",
+    "vercel-ai",
+    "context-chef",
+    "middleware",
+    "context-engineering",
+    "llm",
+    "compression",
+    "token-management"
+  ],
+  "author": "MyPrototypeWhat",
+  "license": "ISC",
+  "description": "AI SDK middleware for context-chef. Transparent history compression, tool result truncation, and token budget management.",
+  "repository": {
+    "type": "git",
+    "url": "https://github.com/MyPrototypeWhat/context-chef.git",
+    "directory": "packages/ai-sdk-middleware"
+  },
+  "homepage": "https://github.com/MyPrototypeWhat/context-chef#readme",
+  "bugs": {
+    "url": "https://github.com/MyPrototypeWhat/context-chef/issues"
+  },
+  "dependencies": {
+    "@context-chef/core": "workspace:*"
+  },
+  "peerDependencies": {
+    "@ai-sdk/provider": ">=3",
+    "ai": ">=6"
+  },
+  "devDependencies": {
+    "@ai-sdk/provider": "^3.0.8",
+    "@types/node": "^25.3.0",
+    "ai": "^6.0.140",
+    "tsdown": "^0.20.3",
+    "typescript": "^5.9.3",
+    "vitest": "^4.0.18"
+  }
+}